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/Analysis/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 BBDuplicateThreshold("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 unsigned BBDupThreshold;
93 // RAII helper for updating the recursion stack.
94 struct RecursionSetRemover {
95 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet;
96 std::pair<Value*, BasicBlock*> ThePair;
98 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S,
99 std::pair<Value*, BasicBlock*> P)
100 : TheSet(S), ThePair(P) { }
102 ~RecursionSetRemover() {
103 TheSet.erase(ThePair);
107 static char ID; // Pass identification
108 JumpThreading(int T = -1) : FunctionPass(ID) {
109 BBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
110 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
113 bool runOnFunction(Function &F) override;
115 void getAnalysisUsage(AnalysisUsage &AU) const override {
116 AU.addRequired<LazyValueInfo>();
117 AU.addPreserved<LazyValueInfo>();
118 AU.addRequired<TargetLibraryInfoWrapperPass>();
121 void FindLoopHeaders(Function &F);
122 bool ProcessBlock(BasicBlock *BB);
123 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
125 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
126 const SmallVectorImpl<BasicBlock *> &PredBBs);
128 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
129 PredValueInfo &Result,
130 ConstantPreference Preference,
131 Instruction *CxtI = nullptr);
132 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
133 ConstantPreference Preference,
134 Instruction *CxtI = nullptr);
136 bool ProcessBranchOnPHI(PHINode *PN);
137 bool ProcessBranchOnXOR(BinaryOperator *BO);
139 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
140 bool TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB);
144 char JumpThreading::ID = 0;
145 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
146 "Jump Threading", false, false)
147 INITIALIZE_PASS_DEPENDENCY(LazyValueInfo)
148 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
149 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
150 "Jump Threading", false, false)
152 // Public interface to the Jump Threading pass
153 FunctionPass *llvm::createJumpThreadingPass(int Threshold) { return new JumpThreading(Threshold); }
155 /// runOnFunction - Top level algorithm.
157 bool JumpThreading::runOnFunction(Function &F) {
158 if (skipOptnoneFunction(F))
161 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
162 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
163 DL = DLP ? &DLP->getDataLayout() : nullptr;
164 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
165 LVI = &getAnalysis<LazyValueInfo>();
167 // Remove unreachable blocks from function as they may result in infinite
168 // loop. We do threading if we found something profitable. Jump threading a
169 // branch can create other opportunities. If these opportunities form a cycle
170 // i.e. if any jump treading is undoing previous threading in the path, then
171 // we will loop forever. We take care of this issue by not jump threading for
172 // back edges. This works for normal cases but not for unreachable blocks as
173 // they may have cycle with no back edge.
174 removeUnreachableBlocks(F);
178 bool Changed, EverChanged = false;
181 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
183 // Thread all of the branches we can over this block.
184 while (ProcessBlock(BB))
189 // If the block is trivially dead, zap it. This eliminates the successor
190 // edges which simplifies the CFG.
191 if (pred_empty(BB) &&
192 BB != &BB->getParent()->getEntryBlock()) {
193 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
194 << "' with terminator: " << *BB->getTerminator() << '\n');
195 LoopHeaders.erase(BB);
202 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
204 // Can't thread an unconditional jump, but if the block is "almost
205 // empty", we can replace uses of it with uses of the successor and make
207 if (BI && BI->isUnconditional() &&
208 BB != &BB->getParent()->getEntryBlock() &&
209 // If the terminator is the only non-phi instruction, try to nuke it.
210 BB->getFirstNonPHIOrDbg()->isTerminator()) {
211 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
212 // block, we have to make sure it isn't in the LoopHeaders set. We
213 // reinsert afterward if needed.
214 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
215 BasicBlock *Succ = BI->getSuccessor(0);
217 // FIXME: It is always conservatively correct to drop the info
218 // for a block even if it doesn't get erased. This isn't totally
219 // awesome, but it allows us to use AssertingVH to prevent nasty
220 // dangling pointer issues within LazyValueInfo.
222 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
224 // If we deleted BB and BB was the header of a loop, then the
225 // successor is now the header of the loop.
229 if (ErasedFromLoopHeaders)
230 LoopHeaders.insert(BB);
233 EverChanged |= Changed;
240 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
241 /// thread across it. Stop scanning the block when passing the threshold.
242 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB,
243 unsigned Threshold) {
244 /// Ignore PHI nodes, these will be flattened when duplication happens.
245 BasicBlock::const_iterator I = BB->getFirstNonPHI();
247 // FIXME: THREADING will delete values that are just used to compute the
248 // branch, so they shouldn't count against the duplication cost.
250 // Sum up the cost of each instruction until we get to the terminator. Don't
251 // include the terminator because the copy won't include it.
253 for (; !isa<TerminatorInst>(I); ++I) {
255 // Stop scanning the block if we've reached the threshold.
256 if (Size > Threshold)
259 // Debugger intrinsics don't incur code size.
260 if (isa<DbgInfoIntrinsic>(I)) continue;
262 // If this is a pointer->pointer bitcast, it is free.
263 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
266 // All other instructions count for at least one unit.
269 // Calls are more expensive. If they are non-intrinsic calls, we model them
270 // as having cost of 4. If they are a non-vector intrinsic, we model them
271 // as having cost of 2 total, and if they are a vector intrinsic, we model
272 // them as having cost 1.
273 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
274 if (CI->cannotDuplicate())
275 // Blocks with NoDuplicate are modelled as having infinite cost, so they
276 // are never duplicated.
278 else if (!isa<IntrinsicInst>(CI))
280 else if (!CI->getType()->isVectorTy())
285 // Threading through a switch statement is particularly profitable. If this
286 // block ends in a switch, decrease its cost to make it more likely to happen.
287 if (isa<SwitchInst>(I))
288 Size = Size > 6 ? Size-6 : 0;
290 // The same holds for indirect branches, but slightly more so.
291 if (isa<IndirectBrInst>(I))
292 Size = Size > 8 ? Size-8 : 0;
297 /// FindLoopHeaders - We do not want jump threading to turn proper loop
298 /// structures into irreducible loops. Doing this breaks up the loop nesting
299 /// hierarchy and pessimizes later transformations. To prevent this from
300 /// happening, we first have to find the loop headers. Here we approximate this
301 /// by finding targets of backedges in the CFG.
303 /// Note that there definitely are cases when we want to allow threading of
304 /// edges across a loop header. For example, threading a jump from outside the
305 /// loop (the preheader) to an exit block of the loop is definitely profitable.
306 /// It is also almost always profitable to thread backedges from within the loop
307 /// to exit blocks, and is often profitable to thread backedges to other blocks
308 /// within the loop (forming a nested loop). This simple analysis is not rich
309 /// enough to track all of these properties and keep it up-to-date as the CFG
310 /// mutates, so we don't allow any of these transformations.
312 void JumpThreading::FindLoopHeaders(Function &F) {
313 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
314 FindFunctionBackedges(F, Edges);
316 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
317 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
320 /// getKnownConstant - Helper method to determine if we can thread over a
321 /// terminator with the given value as its condition, and if so what value to
322 /// use for that. What kind of value this is depends on whether we want an
323 /// integer or a block address, but an undef is always accepted.
324 /// Returns null if Val is null or not an appropriate constant.
325 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
329 // Undef is "known" enough.
330 if (UndefValue *U = dyn_cast<UndefValue>(Val))
333 if (Preference == WantBlockAddress)
334 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
336 return dyn_cast<ConstantInt>(Val);
339 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
340 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
341 /// in any of our predecessors. If so, return the known list of value and pred
342 /// BB in the result vector.
344 /// This returns true if there were any known values.
347 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
348 ConstantPreference Preference,
350 // This method walks up use-def chains recursively. Because of this, we could
351 // get into an infinite loop going around loops in the use-def chain. To
352 // prevent this, keep track of what (value, block) pairs we've already visited
353 // and terminate the search if we loop back to them
354 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
357 // An RAII help to remove this pair from the recursion set once the recursion
358 // stack pops back out again.
359 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
361 // If V is a constant, then it is known in all predecessors.
362 if (Constant *KC = getKnownConstant(V, Preference)) {
363 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
364 Result.push_back(std::make_pair(KC, *PI));
369 // If V is a non-instruction value, or an instruction in a different block,
370 // then it can't be derived from a PHI.
371 Instruction *I = dyn_cast<Instruction>(V);
372 if (!I || I->getParent() != BB) {
374 // Okay, if this is a live-in value, see if it has a known value at the end
375 // of any of our predecessors.
377 // FIXME: This should be an edge property, not a block end property.
378 /// TODO: Per PR2563, we could infer value range information about a
379 /// predecessor based on its terminator.
381 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
382 // "I" is a non-local compare-with-a-constant instruction. This would be
383 // able to handle value inequalities better, for example if the compare is
384 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
385 // Perhaps getConstantOnEdge should be smart enough to do this?
387 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
389 // If the value is known by LazyValueInfo to be a constant in a
390 // predecessor, use that information to try to thread this block.
391 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
392 if (Constant *KC = getKnownConstant(PredCst, Preference))
393 Result.push_back(std::make_pair(KC, P));
396 return !Result.empty();
399 /// If I is a PHI node, then we know the incoming values for any constants.
400 if (PHINode *PN = dyn_cast<PHINode>(I)) {
401 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
402 Value *InVal = PN->getIncomingValue(i);
403 if (Constant *KC = getKnownConstant(InVal, Preference)) {
404 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
406 Constant *CI = LVI->getConstantOnEdge(InVal,
407 PN->getIncomingBlock(i),
409 if (Constant *KC = getKnownConstant(CI, Preference))
410 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
414 return !Result.empty();
417 PredValueInfoTy LHSVals, RHSVals;
419 // Handle some boolean conditions.
420 if (I->getType()->getPrimitiveSizeInBits() == 1) {
421 assert(Preference == WantInteger && "One-bit non-integer type?");
423 // X & false -> false
424 if (I->getOpcode() == Instruction::Or ||
425 I->getOpcode() == Instruction::And) {
426 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
428 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
431 if (LHSVals.empty() && RHSVals.empty())
434 ConstantInt *InterestingVal;
435 if (I->getOpcode() == Instruction::Or)
436 InterestingVal = ConstantInt::getTrue(I->getContext());
438 InterestingVal = ConstantInt::getFalse(I->getContext());
440 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
442 // Scan for the sentinel. If we find an undef, force it to the
443 // interesting value: x|undef -> true and x&undef -> false.
444 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
445 if (LHSVals[i].first == InterestingVal ||
446 isa<UndefValue>(LHSVals[i].first)) {
447 Result.push_back(LHSVals[i]);
448 Result.back().first = InterestingVal;
449 LHSKnownBBs.insert(LHSVals[i].second);
451 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
452 if (RHSVals[i].first == InterestingVal ||
453 isa<UndefValue>(RHSVals[i].first)) {
454 // If we already inferred a value for this block on the LHS, don't
456 if (!LHSKnownBBs.count(RHSVals[i].second)) {
457 Result.push_back(RHSVals[i]);
458 Result.back().first = InterestingVal;
462 return !Result.empty();
465 // Handle the NOT form of XOR.
466 if (I->getOpcode() == Instruction::Xor &&
467 isa<ConstantInt>(I->getOperand(1)) &&
468 cast<ConstantInt>(I->getOperand(1))->isOne()) {
469 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
474 // Invert the known values.
475 for (unsigned i = 0, e = Result.size(); i != e; ++i)
476 Result[i].first = ConstantExpr::getNot(Result[i].first);
481 // Try to simplify some other binary operator values.
482 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
483 assert(Preference != WantBlockAddress
484 && "A binary operator creating a block address?");
485 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
486 PredValueInfoTy LHSVals;
487 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
490 // Try to use constant folding to simplify the binary operator.
491 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
492 Constant *V = LHSVals[i].first;
493 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
495 if (Constant *KC = getKnownConstant(Folded, WantInteger))
496 Result.push_back(std::make_pair(KC, LHSVals[i].second));
500 return !Result.empty();
503 // Handle compare with phi operand, where the PHI is defined in this block.
504 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
505 assert(Preference == WantInteger && "Compares only produce integers");
506 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
507 if (PN && PN->getParent() == BB) {
508 // We can do this simplification if any comparisons fold to true or false.
510 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
511 BasicBlock *PredBB = PN->getIncomingBlock(i);
512 Value *LHS = PN->getIncomingValue(i);
513 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
515 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL);
517 if (!isa<Constant>(RHS))
520 LazyValueInfo::Tristate
521 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
522 cast<Constant>(RHS), PredBB, BB,
524 if (ResT == LazyValueInfo::Unknown)
526 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
529 if (Constant *KC = getKnownConstant(Res, WantInteger))
530 Result.push_back(std::make_pair(KC, PredBB));
533 return !Result.empty();
536 // If comparing a live-in value against a constant, see if we know the
537 // live-in value on any predecessors.
538 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
539 if (!isa<Instruction>(Cmp->getOperand(0)) ||
540 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
541 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
543 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
545 // If the value is known by LazyValueInfo to be a constant in a
546 // predecessor, use that information to try to thread this block.
547 LazyValueInfo::Tristate Res =
548 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
549 RHSCst, P, BB, CxtI ? CxtI : Cmp);
550 if (Res == LazyValueInfo::Unknown)
553 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
554 Result.push_back(std::make_pair(ResC, P));
557 return !Result.empty();
560 // Try to find a constant value for the LHS of a comparison,
561 // and evaluate it statically if we can.
562 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
563 PredValueInfoTy LHSVals;
564 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
567 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
568 Constant *V = LHSVals[i].first;
569 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
571 if (Constant *KC = getKnownConstant(Folded, WantInteger))
572 Result.push_back(std::make_pair(KC, LHSVals[i].second));
575 return !Result.empty();
580 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
581 // Handle select instructions where at least one operand is a known constant
582 // and we can figure out the condition value for any predecessor block.
583 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
584 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
585 PredValueInfoTy Conds;
586 if ((TrueVal || FalseVal) &&
587 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
588 WantInteger, CxtI)) {
589 for (unsigned i = 0, e = Conds.size(); i != e; ++i) {
590 Constant *Cond = Conds[i].first;
592 // Figure out what value to use for the condition.
594 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
596 KnownCond = CI->isOne();
598 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
599 // Either operand will do, so be sure to pick the one that's a known
601 // FIXME: Do this more cleverly if both values are known constants?
602 KnownCond = (TrueVal != nullptr);
605 // See if the select has a known constant value for this predecessor.
606 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
607 Result.push_back(std::make_pair(Val, Conds[i].second));
610 return !Result.empty();
614 // If all else fails, see if LVI can figure out a constant value for us.
615 Constant *CI = LVI->getConstant(V, BB, CxtI);
616 if (Constant *KC = getKnownConstant(CI, Preference)) {
617 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
618 Result.push_back(std::make_pair(KC, *PI));
621 return !Result.empty();
626 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
627 /// in an undefined jump, decide which block is best to revector to.
629 /// Since we can pick an arbitrary destination, we pick the successor with the
630 /// fewest predecessors. This should reduce the in-degree of the others.
632 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
633 TerminatorInst *BBTerm = BB->getTerminator();
634 unsigned MinSucc = 0;
635 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
636 // Compute the successor with the minimum number of predecessors.
637 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
638 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
639 TestBB = BBTerm->getSuccessor(i);
640 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
641 if (NumPreds < MinNumPreds) {
643 MinNumPreds = NumPreds;
650 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
651 if (!BB->hasAddressTaken()) return false;
653 // If the block has its address taken, it may be a tree of dead constants
654 // hanging off of it. These shouldn't keep the block alive.
655 BlockAddress *BA = BlockAddress::get(BB);
656 BA->removeDeadConstantUsers();
657 return !BA->use_empty();
660 /// ProcessBlock - If there are any predecessors whose control can be threaded
661 /// through to a successor, transform them now.
662 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
663 // If the block is trivially dead, just return and let the caller nuke it.
664 // This simplifies other transformations.
665 if (pred_empty(BB) &&
666 BB != &BB->getParent()->getEntryBlock())
669 // If this block has a single predecessor, and if that pred has a single
670 // successor, merge the blocks. This encourages recursive jump threading
671 // because now the condition in this block can be threaded through
672 // predecessors of our predecessor block.
673 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
674 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
675 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
676 // If SinglePred was a loop header, BB becomes one.
677 if (LoopHeaders.erase(SinglePred))
678 LoopHeaders.insert(BB);
680 LVI->eraseBlock(SinglePred);
681 MergeBasicBlockIntoOnlyPred(BB);
687 // What kind of constant we're looking for.
688 ConstantPreference Preference = WantInteger;
690 // Look to see if the terminator is a conditional branch, switch or indirect
691 // branch, if not we can't thread it.
693 Instruction *Terminator = BB->getTerminator();
694 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
695 // Can't thread an unconditional jump.
696 if (BI->isUnconditional()) return false;
697 Condition = BI->getCondition();
698 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
699 Condition = SI->getCondition();
700 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
701 // Can't thread indirect branch with no successors.
702 if (IB->getNumSuccessors() == 0) return false;
703 Condition = IB->getAddress()->stripPointerCasts();
704 Preference = WantBlockAddress;
706 return false; // Must be an invoke.
709 // Run constant folding to see if we can reduce the condition to a simple
711 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
712 Value *SimpleVal = ConstantFoldInstruction(I, DL, TLI);
714 I->replaceAllUsesWith(SimpleVal);
715 I->eraseFromParent();
716 Condition = SimpleVal;
720 // If the terminator is branching on an undef, we can pick any of the
721 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
722 if (isa<UndefValue>(Condition)) {
723 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
725 // Fold the branch/switch.
726 TerminatorInst *BBTerm = BB->getTerminator();
727 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
728 if (i == BestSucc) continue;
729 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
732 DEBUG(dbgs() << " In block '" << BB->getName()
733 << "' folding undef terminator: " << *BBTerm << '\n');
734 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
735 BBTerm->eraseFromParent();
739 // If the terminator of this block is branching on a constant, simplify the
740 // terminator to an unconditional branch. This can occur due to threading in
742 if (getKnownConstant(Condition, Preference)) {
743 DEBUG(dbgs() << " In block '" << BB->getName()
744 << "' folding terminator: " << *BB->getTerminator() << '\n');
746 ConstantFoldTerminator(BB, true);
750 Instruction *CondInst = dyn_cast<Instruction>(Condition);
752 // All the rest of our checks depend on the condition being an instruction.
754 // FIXME: Unify this with code below.
755 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
761 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
762 // For a comparison where the LHS is outside this block, it's possible
763 // that we've branched on it before. Used LVI to see if we can simplify
764 // the branch based on that.
765 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
766 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
767 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
768 if (CondBr && CondConst && CondBr->isConditional() && PI != PE &&
769 (!isa<Instruction>(CondCmp->getOperand(0)) ||
770 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
771 // For predecessor edge, determine if the comparison is true or false
772 // on that edge. If they're all true or all false, we can simplify the
774 // FIXME: We could handle mixed true/false by duplicating code.
775 LazyValueInfo::Tristate Baseline =
776 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0),
777 CondConst, *PI, BB, CondCmp);
778 if (Baseline != LazyValueInfo::Unknown) {
779 // Check that all remaining incoming values match the first one.
781 LazyValueInfo::Tristate Ret =
782 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
783 CondCmp->getOperand(0), CondConst, *PI, BB,
785 if (Ret != Baseline) break;
788 // If we terminated early, then one of the values didn't match.
790 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0;
791 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1;
792 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
793 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
794 CondBr->eraseFromParent();
799 } else if (CondBr && CondConst && CondBr->isConditional()) {
800 // There might be an invariant in the same block with the conditional
801 // that can determine the predicate.
803 LazyValueInfo::Tristate Ret =
804 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
806 if (Ret != LazyValueInfo::Unknown) {
807 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
808 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
809 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
810 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
811 CondBr->eraseFromParent();
816 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB))
820 // Check for some cases that are worth simplifying. Right now we want to look
821 // for loads that are used by a switch or by the condition for the branch. If
822 // we see one, check to see if it's partially redundant. If so, insert a PHI
823 // which can then be used to thread the values.
825 Value *SimplifyValue = CondInst;
826 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
827 if (isa<Constant>(CondCmp->getOperand(1)))
828 SimplifyValue = CondCmp->getOperand(0);
830 // TODO: There are other places where load PRE would be profitable, such as
831 // more complex comparisons.
832 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
833 if (SimplifyPartiallyRedundantLoad(LI))
837 // Handle a variety of cases where we are branching on something derived from
838 // a PHI node in the current block. If we can prove that any predecessors
839 // compute a predictable value based on a PHI node, thread those predecessors.
841 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
844 // If this is an otherwise-unfoldable branch on a phi node in the current
845 // block, see if we can simplify.
846 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
847 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
848 return ProcessBranchOnPHI(PN);
851 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
852 if (CondInst->getOpcode() == Instruction::Xor &&
853 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
854 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
857 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
858 // "(X == 4)", thread through this block.
863 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
864 /// load instruction, eliminate it by replacing it with a PHI node. This is an
865 /// important optimization that encourages jump threading, and needs to be run
866 /// interlaced with other jump threading tasks.
867 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
868 // Don't hack volatile/atomic loads.
869 if (!LI->isSimple()) return false;
871 // If the load is defined in a block with exactly one predecessor, it can't be
872 // partially redundant.
873 BasicBlock *LoadBB = LI->getParent();
874 if (LoadBB->getSinglePredecessor())
877 // If the load is defined in a landing pad, it can't be partially redundant,
878 // because the edges between the invoke and the landing pad cannot have other
879 // instructions between them.
880 if (LoadBB->isLandingPad())
883 Value *LoadedPtr = LI->getOperand(0);
885 // If the loaded operand is defined in the LoadBB, it can't be available.
886 // TODO: Could do simple PHI translation, that would be fun :)
887 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
888 if (PtrOp->getParent() == LoadBB)
891 // Scan a few instructions up from the load, to see if it is obviously live at
892 // the entry to its block.
893 BasicBlock::iterator BBIt = LI;
895 if (Value *AvailableVal =
896 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
897 // If the value if the load is locally available within the block, just use
898 // it. This frequently occurs for reg2mem'd allocas.
899 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
901 // If the returned value is the load itself, replace with an undef. This can
902 // only happen in dead loops.
903 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
904 if (AvailableVal->getType() != LI->getType())
906 CastInst::CreateBitOrPointerCast(AvailableVal, LI->getType(), "", LI);
907 LI->replaceAllUsesWith(AvailableVal);
908 LI->eraseFromParent();
912 // Otherwise, if we scanned the whole block and got to the top of the block,
913 // we know the block is locally transparent to the load. If not, something
914 // might clobber its value.
915 if (BBIt != LoadBB->begin())
918 // If all of the loads and stores that feed the value have the same AA tags,
919 // then we can propagate them onto any newly inserted loads.
921 LI->getAAMetadata(AATags);
923 SmallPtrSet<BasicBlock*, 8> PredsScanned;
924 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
925 AvailablePredsTy AvailablePreds;
926 BasicBlock *OneUnavailablePred = nullptr;
928 // If we got here, the loaded value is transparent through to the start of the
929 // block. Check to see if it is available in any of the predecessor blocks.
930 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
932 BasicBlock *PredBB = *PI;
934 // If we already scanned this predecessor, skip it.
935 if (!PredsScanned.insert(PredBB).second)
938 // Scan the predecessor to see if the value is available in the pred.
939 BBIt = PredBB->end();
940 AAMDNodes ThisAATags;
941 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6,
942 nullptr, &ThisAATags);
943 if (!PredAvailable) {
944 OneUnavailablePred = PredBB;
948 // If AA tags disagree or are not present, forget about them.
949 if (AATags != ThisAATags) AATags = AAMDNodes();
951 // If so, this load is partially redundant. Remember this info so that we
952 // can create a PHI node.
953 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
956 // If the loaded value isn't available in any predecessor, it isn't partially
958 if (AvailablePreds.empty()) return false;
960 // Okay, the loaded value is available in at least one (and maybe all!)
961 // predecessors. If the value is unavailable in more than one unique
962 // predecessor, we want to insert a merge block for those common predecessors.
963 // This ensures that we only have to insert one reload, thus not increasing
965 BasicBlock *UnavailablePred = nullptr;
967 // If there is exactly one predecessor where the value is unavailable, the
968 // already computed 'OneUnavailablePred' block is it. If it ends in an
969 // unconditional branch, we know that it isn't a critical edge.
970 if (PredsScanned.size() == AvailablePreds.size()+1 &&
971 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
972 UnavailablePred = OneUnavailablePred;
973 } else if (PredsScanned.size() != AvailablePreds.size()) {
974 // Otherwise, we had multiple unavailable predecessors or we had a critical
975 // edge from the one.
976 SmallVector<BasicBlock*, 8> PredsToSplit;
977 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
979 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
980 AvailablePredSet.insert(AvailablePreds[i].first);
982 // Add all the unavailable predecessors to the PredsToSplit list.
983 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
986 // If the predecessor is an indirect goto, we can't split the edge.
987 if (isa<IndirectBrInst>(P->getTerminator()))
990 if (!AvailablePredSet.count(P))
991 PredsToSplit.push_back(P);
994 // Split them out to their own block.
996 SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split");
999 // If the value isn't available in all predecessors, then there will be
1000 // exactly one where it isn't available. Insert a load on that edge and add
1001 // it to the AvailablePreds list.
1002 if (UnavailablePred) {
1003 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1004 "Can't handle critical edge here!");
1005 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
1007 UnavailablePred->getTerminator());
1008 NewVal->setDebugLoc(LI->getDebugLoc());
1010 NewVal->setAAMetadata(AATags);
1012 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1015 // Now we know that each predecessor of this block has a value in
1016 // AvailablePreds, sort them for efficient access as we're walking the preds.
1017 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1019 // Create a PHI node at the start of the block for the PRE'd load value.
1020 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1021 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
1024 PN->setDebugLoc(LI->getDebugLoc());
1026 // Insert new entries into the PHI for each predecessor. A single block may
1027 // have multiple entries here.
1028 for (pred_iterator PI = PB; PI != PE; ++PI) {
1029 BasicBlock *P = *PI;
1030 AvailablePredsTy::iterator I =
1031 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1032 std::make_pair(P, (Value*)nullptr));
1034 assert(I != AvailablePreds.end() && I->first == P &&
1035 "Didn't find entry for predecessor!");
1037 // If we have an available predecessor but it requires casting, insert the
1038 // cast in the predecessor and use the cast. Note that we have to update the
1039 // AvailablePreds vector as we go so that all of the PHI entries for this
1040 // predecessor use the same bitcast.
1041 Value *&PredV = I->second;
1042 if (PredV->getType() != LI->getType())
1043 PredV = CastInst::CreateBitOrPointerCast(PredV, LI->getType(), "",
1044 P->getTerminator());
1046 PN->addIncoming(PredV, I->first);
1049 //cerr << "PRE: " << *LI << *PN << "\n";
1051 LI->replaceAllUsesWith(PN);
1052 LI->eraseFromParent();
1057 /// FindMostPopularDest - The specified list contains multiple possible
1058 /// threadable destinations. Pick the one that occurs the most frequently in
1061 FindMostPopularDest(BasicBlock *BB,
1062 const SmallVectorImpl<std::pair<BasicBlock*,
1063 BasicBlock*> > &PredToDestList) {
1064 assert(!PredToDestList.empty());
1066 // Determine popularity. If there are multiple possible destinations, we
1067 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1068 // blocks with known and real destinations to threading undef. We'll handle
1069 // them later if interesting.
1070 DenseMap<BasicBlock*, unsigned> DestPopularity;
1071 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1072 if (PredToDestList[i].second)
1073 DestPopularity[PredToDestList[i].second]++;
1075 // Find the most popular dest.
1076 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1077 BasicBlock *MostPopularDest = DPI->first;
1078 unsigned Popularity = DPI->second;
1079 SmallVector<BasicBlock*, 4> SamePopularity;
1081 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1082 // If the popularity of this entry isn't higher than the popularity we've
1083 // seen so far, ignore it.
1084 if (DPI->second < Popularity)
1086 else if (DPI->second == Popularity) {
1087 // If it is the same as what we've seen so far, keep track of it.
1088 SamePopularity.push_back(DPI->first);
1090 // If it is more popular, remember it.
1091 SamePopularity.clear();
1092 MostPopularDest = DPI->first;
1093 Popularity = DPI->second;
1097 // Okay, now we know the most popular destination. If there is more than one
1098 // destination, we need to determine one. This is arbitrary, but we need
1099 // to make a deterministic decision. Pick the first one that appears in the
1101 if (!SamePopularity.empty()) {
1102 SamePopularity.push_back(MostPopularDest);
1103 TerminatorInst *TI = BB->getTerminator();
1104 for (unsigned i = 0; ; ++i) {
1105 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1107 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1108 TI->getSuccessor(i)) == SamePopularity.end())
1111 MostPopularDest = TI->getSuccessor(i);
1116 // Okay, we have finally picked the most popular destination.
1117 return MostPopularDest;
1120 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1121 ConstantPreference Preference,
1122 Instruction *CxtI) {
1123 // If threading this would thread across a loop header, don't even try to
1125 if (LoopHeaders.count(BB))
1128 PredValueInfoTy PredValues;
1129 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1132 assert(!PredValues.empty() &&
1133 "ComputeValueKnownInPredecessors returned true with no values");
1135 DEBUG(dbgs() << "IN BB: " << *BB;
1136 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1137 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1138 << *PredValues[i].first
1139 << " for pred '" << PredValues[i].second->getName() << "'.\n";
1142 // Decide what we want to thread through. Convert our list of known values to
1143 // a list of known destinations for each pred. This also discards duplicate
1144 // predecessors and keeps track of the undefined inputs (which are represented
1145 // as a null dest in the PredToDestList).
1146 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1147 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1149 BasicBlock *OnlyDest = nullptr;
1150 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1152 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1153 BasicBlock *Pred = PredValues[i].second;
1154 if (!SeenPreds.insert(Pred).second)
1155 continue; // Duplicate predecessor entry.
1157 // If the predecessor ends with an indirect goto, we can't change its
1159 if (isa<IndirectBrInst>(Pred->getTerminator()))
1162 Constant *Val = PredValues[i].first;
1165 if (isa<UndefValue>(Val))
1167 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1168 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1169 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1170 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor();
1172 assert(isa<IndirectBrInst>(BB->getTerminator())
1173 && "Unexpected terminator");
1174 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1177 // If we have exactly one destination, remember it for efficiency below.
1178 if (PredToDestList.empty())
1180 else if (OnlyDest != DestBB)
1181 OnlyDest = MultipleDestSentinel;
1183 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1186 // If all edges were unthreadable, we fail.
1187 if (PredToDestList.empty())
1190 // Determine which is the most common successor. If we have many inputs and
1191 // this block is a switch, we want to start by threading the batch that goes
1192 // to the most popular destination first. If we only know about one
1193 // threadable destination (the common case) we can avoid this.
1194 BasicBlock *MostPopularDest = OnlyDest;
1196 if (MostPopularDest == MultipleDestSentinel)
1197 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1199 // Now that we know what the most popular destination is, factor all
1200 // predecessors that will jump to it into a single predecessor.
1201 SmallVector<BasicBlock*, 16> PredsToFactor;
1202 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1203 if (PredToDestList[i].second == MostPopularDest) {
1204 BasicBlock *Pred = PredToDestList[i].first;
1206 // This predecessor may be a switch or something else that has multiple
1207 // edges to the block. Factor each of these edges by listing them
1208 // according to # occurrences in PredsToFactor.
1209 TerminatorInst *PredTI = Pred->getTerminator();
1210 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1211 if (PredTI->getSuccessor(i) == BB)
1212 PredsToFactor.push_back(Pred);
1215 // If the threadable edges are branching on an undefined value, we get to pick
1216 // the destination that these predecessors should get to.
1217 if (!MostPopularDest)
1218 MostPopularDest = BB->getTerminator()->
1219 getSuccessor(GetBestDestForJumpOnUndef(BB));
1221 // Ok, try to thread it!
1222 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1225 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1226 /// a PHI node in the current block. See if there are any simplifications we
1227 /// can do based on inputs to the phi node.
1229 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1230 BasicBlock *BB = PN->getParent();
1232 // TODO: We could make use of this to do it once for blocks with common PHI
1234 SmallVector<BasicBlock*, 1> PredBBs;
1237 // If any of the predecessor blocks end in an unconditional branch, we can
1238 // *duplicate* the conditional branch into that block in order to further
1239 // encourage jump threading and to eliminate cases where we have branch on a
1240 // phi of an icmp (branch on icmp is much better).
1241 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1242 BasicBlock *PredBB = PN->getIncomingBlock(i);
1243 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1244 if (PredBr->isUnconditional()) {
1245 PredBBs[0] = PredBB;
1246 // Try to duplicate BB into PredBB.
1247 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1255 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1256 /// a xor instruction in the current block. See if there are any
1257 /// simplifications we can do based on inputs to the xor.
1259 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1260 BasicBlock *BB = BO->getParent();
1262 // If either the LHS or RHS of the xor is a constant, don't do this
1264 if (isa<ConstantInt>(BO->getOperand(0)) ||
1265 isa<ConstantInt>(BO->getOperand(1)))
1268 // If the first instruction in BB isn't a phi, we won't be able to infer
1269 // anything special about any particular predecessor.
1270 if (!isa<PHINode>(BB->front()))
1273 // If we have a xor as the branch input to this block, and we know that the
1274 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1275 // the condition into the predecessor and fix that value to true, saving some
1276 // logical ops on that path and encouraging other paths to simplify.
1278 // This copies something like this:
1281 // %X = phi i1 [1], [%X']
1282 // %Y = icmp eq i32 %A, %B
1283 // %Z = xor i1 %X, %Y
1288 // %Y = icmp ne i32 %A, %B
1291 PredValueInfoTy XorOpValues;
1293 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1295 assert(XorOpValues.empty());
1296 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1302 assert(!XorOpValues.empty() &&
1303 "ComputeValueKnownInPredecessors returned true with no values");
1305 // Scan the information to see which is most popular: true or false. The
1306 // predecessors can be of the set true, false, or undef.
1307 unsigned NumTrue = 0, NumFalse = 0;
1308 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1309 if (isa<UndefValue>(XorOpValues[i].first))
1310 // Ignore undefs for the count.
1312 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1318 // Determine which value to split on, true, false, or undef if neither.
1319 ConstantInt *SplitVal = nullptr;
1320 if (NumTrue > NumFalse)
1321 SplitVal = ConstantInt::getTrue(BB->getContext());
1322 else if (NumTrue != 0 || NumFalse != 0)
1323 SplitVal = ConstantInt::getFalse(BB->getContext());
1325 // Collect all of the blocks that this can be folded into so that we can
1326 // factor this once and clone it once.
1327 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1328 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1329 if (XorOpValues[i].first != SplitVal &&
1330 !isa<UndefValue>(XorOpValues[i].first))
1333 BlocksToFoldInto.push_back(XorOpValues[i].second);
1336 // If we inferred a value for all of the predecessors, then duplication won't
1337 // help us. However, we can just replace the LHS or RHS with the constant.
1338 if (BlocksToFoldInto.size() ==
1339 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1341 // If all preds provide undef, just nuke the xor, because it is undef too.
1342 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1343 BO->eraseFromParent();
1344 } else if (SplitVal->isZero()) {
1345 // If all preds provide 0, replace the xor with the other input.
1346 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1347 BO->eraseFromParent();
1349 // If all preds provide 1, set the computed value to 1.
1350 BO->setOperand(!isLHS, SplitVal);
1356 // Try to duplicate BB into PredBB.
1357 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1361 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1362 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1363 /// NewPred using the entries from OldPred (suitably mapped).
1364 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1365 BasicBlock *OldPred,
1366 BasicBlock *NewPred,
1367 DenseMap<Instruction*, Value*> &ValueMap) {
1368 for (BasicBlock::iterator PNI = PHIBB->begin();
1369 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1370 // Ok, we have a PHI node. Figure out what the incoming value was for the
1372 Value *IV = PN->getIncomingValueForBlock(OldPred);
1374 // Remap the value if necessary.
1375 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1376 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1377 if (I != ValueMap.end())
1381 PN->addIncoming(IV, NewPred);
1385 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1386 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1387 /// across BB. Transform the IR to reflect this change.
1388 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1389 const SmallVectorImpl<BasicBlock*> &PredBBs,
1390 BasicBlock *SuccBB) {
1391 // If threading to the same block as we come from, we would infinite loop.
1393 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1394 << "' - would thread to self!\n");
1398 // If threading this would thread across a loop header, don't thread the edge.
1399 // See the comments above FindLoopHeaders for justifications and caveats.
1400 if (LoopHeaders.count(BB)) {
1401 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1402 << "' to dest BB '" << SuccBB->getName()
1403 << "' - it might create an irreducible loop!\n");
1407 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, BBDupThreshold);
1408 if (JumpThreadCost > BBDupThreshold) {
1409 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1410 << "' - Cost is too high: " << JumpThreadCost << "\n");
1414 // And finally, do it! Start by factoring the predecessors is needed.
1416 if (PredBBs.size() == 1)
1417 PredBB = PredBBs[0];
1419 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1420 << " common predecessors.\n");
1421 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm");
1424 // And finally, do it!
1425 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1426 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1427 << ", across block:\n "
1430 LVI->threadEdge(PredBB, BB, SuccBB);
1432 // We are going to have to map operands from the original BB block to the new
1433 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1434 // account for entry from PredBB.
1435 DenseMap<Instruction*, Value*> ValueMapping;
1437 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1438 BB->getName()+".thread",
1439 BB->getParent(), BB);
1440 NewBB->moveAfter(PredBB);
1442 BasicBlock::iterator BI = BB->begin();
1443 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1444 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1446 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1447 // mapping and using it to remap operands in the cloned instructions.
1448 for (; !isa<TerminatorInst>(BI); ++BI) {
1449 Instruction *New = BI->clone();
1450 New->setName(BI->getName());
1451 NewBB->getInstList().push_back(New);
1452 ValueMapping[BI] = New;
1454 // Remap operands to patch up intra-block references.
1455 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1456 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1457 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1458 if (I != ValueMapping.end())
1459 New->setOperand(i, I->second);
1463 // We didn't copy the terminator from BB over to NewBB, because there is now
1464 // an unconditional jump to SuccBB. Insert the unconditional jump.
1465 BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB);
1466 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1468 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1469 // PHI nodes for NewBB now.
1470 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1472 // If there were values defined in BB that are used outside the block, then we
1473 // now have to update all uses of the value to use either the original value,
1474 // the cloned value, or some PHI derived value. This can require arbitrary
1475 // PHI insertion, of which we are prepared to do, clean these up now.
1476 SSAUpdater SSAUpdate;
1477 SmallVector<Use*, 16> UsesToRename;
1478 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1479 // Scan all uses of this instruction to see if it is used outside of its
1480 // block, and if so, record them in UsesToRename.
1481 for (Use &U : I->uses()) {
1482 Instruction *User = cast<Instruction>(U.getUser());
1483 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1484 if (UserPN->getIncomingBlock(U) == BB)
1486 } else if (User->getParent() == BB)
1489 UsesToRename.push_back(&U);
1492 // If there are no uses outside the block, we're done with this instruction.
1493 if (UsesToRename.empty())
1496 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1498 // We found a use of I outside of BB. Rename all uses of I that are outside
1499 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1500 // with the two values we know.
1501 SSAUpdate.Initialize(I->getType(), I->getName());
1502 SSAUpdate.AddAvailableValue(BB, I);
1503 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1505 while (!UsesToRename.empty())
1506 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1507 DEBUG(dbgs() << "\n");
1511 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1512 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1513 // us to simplify any PHI nodes in BB.
1514 TerminatorInst *PredTerm = PredBB->getTerminator();
1515 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1516 if (PredTerm->getSuccessor(i) == BB) {
1517 BB->removePredecessor(PredBB, true);
1518 PredTerm->setSuccessor(i, NewBB);
1521 // At this point, the IR is fully up to date and consistent. Do a quick scan
1522 // over the new instructions and zap any that are constants or dead. This
1523 // frequently happens because of phi translation.
1524 SimplifyInstructionsInBlock(NewBB, DL, TLI);
1526 // Threaded an edge!
1531 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1532 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1533 /// If we can duplicate the contents of BB up into PredBB do so now, this
1534 /// improves the odds that the branch will be on an analyzable instruction like
1536 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1537 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1538 assert(!PredBBs.empty() && "Can't handle an empty set");
1540 // If BB is a loop header, then duplicating this block outside the loop would
1541 // cause us to transform this into an irreducible loop, don't do this.
1542 // See the comments above FindLoopHeaders for justifications and caveats.
1543 if (LoopHeaders.count(BB)) {
1544 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1545 << "' into predecessor block '" << PredBBs[0]->getName()
1546 << "' - it might create an irreducible loop!\n");
1550 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, BBDupThreshold);
1551 if (DuplicationCost > BBDupThreshold) {
1552 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1553 << "' - Cost is too high: " << DuplicationCost << "\n");
1557 // And finally, do it! Start by factoring the predecessors is needed.
1559 if (PredBBs.size() == 1)
1560 PredBB = PredBBs[0];
1562 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1563 << " common predecessors.\n");
1564 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm");
1567 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1569 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1570 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1571 << DuplicationCost << " block is:" << *BB << "\n");
1573 // Unless PredBB ends with an unconditional branch, split the edge so that we
1574 // can just clone the bits from BB into the end of the new PredBB.
1575 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1577 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
1578 PredBB = SplitEdge(PredBB, BB);
1579 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1582 // We are going to have to map operands from the original BB block into the
1583 // PredBB block. Evaluate PHI nodes in BB.
1584 DenseMap<Instruction*, Value*> ValueMapping;
1586 BasicBlock::iterator BI = BB->begin();
1587 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1588 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1590 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1591 // mapping and using it to remap operands in the cloned instructions.
1592 for (; BI != BB->end(); ++BI) {
1593 Instruction *New = BI->clone();
1595 // Remap operands to patch up intra-block references.
1596 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1597 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1598 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1599 if (I != ValueMapping.end())
1600 New->setOperand(i, I->second);
1603 // If this instruction can be simplified after the operands are updated,
1604 // just use the simplified value instead. This frequently happens due to
1606 if (Value *IV = SimplifyInstruction(New, DL)) {
1608 ValueMapping[BI] = IV;
1610 // Otherwise, insert the new instruction into the block.
1611 New->setName(BI->getName());
1612 PredBB->getInstList().insert(OldPredBranch, New);
1613 ValueMapping[BI] = New;
1617 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1618 // add entries to the PHI nodes for branch from PredBB now.
1619 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1620 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1622 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1625 // If there were values defined in BB that are used outside the block, then we
1626 // now have to update all uses of the value to use either the original value,
1627 // the cloned value, or some PHI derived value. This can require arbitrary
1628 // PHI insertion, of which we are prepared to do, clean these up now.
1629 SSAUpdater SSAUpdate;
1630 SmallVector<Use*, 16> UsesToRename;
1631 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1632 // Scan all uses of this instruction to see if it is used outside of its
1633 // block, and if so, record them in UsesToRename.
1634 for (Use &U : I->uses()) {
1635 Instruction *User = cast<Instruction>(U.getUser());
1636 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1637 if (UserPN->getIncomingBlock(U) == BB)
1639 } else if (User->getParent() == BB)
1642 UsesToRename.push_back(&U);
1645 // If there are no uses outside the block, we're done with this instruction.
1646 if (UsesToRename.empty())
1649 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1651 // We found a use of I outside of BB. Rename all uses of I that are outside
1652 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1653 // with the two values we know.
1654 SSAUpdate.Initialize(I->getType(), I->getName());
1655 SSAUpdate.AddAvailableValue(BB, I);
1656 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1658 while (!UsesToRename.empty())
1659 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1660 DEBUG(dbgs() << "\n");
1663 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1665 BB->removePredecessor(PredBB, true);
1667 // Remove the unconditional branch at the end of the PredBB block.
1668 OldPredBranch->eraseFromParent();
1674 /// TryToUnfoldSelect - Look for blocks of the form
1680 /// %p = phi [%a, %bb] ...
1684 /// And expand the select into a branch structure if one of its arms allows %c
1685 /// to be folded. This later enables threading from bb1 over bb2.
1686 bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
1687 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1688 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
1689 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
1691 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
1692 CondLHS->getParent() != BB)
1695 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
1696 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
1697 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
1699 // Look if one of the incoming values is a select in the corresponding
1701 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
1704 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
1705 if (!PredTerm || !PredTerm->isUnconditional())
1708 // Now check if one of the select values would allow us to constant fold the
1709 // terminator in BB. We don't do the transform if both sides fold, those
1710 // cases will be threaded in any case.
1711 LazyValueInfo::Tristate LHSFolds =
1712 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
1713 CondRHS, Pred, BB, CondCmp);
1714 LazyValueInfo::Tristate RHSFolds =
1715 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
1716 CondRHS, Pred, BB, CondCmp);
1717 if ((LHSFolds != LazyValueInfo::Unknown ||
1718 RHSFolds != LazyValueInfo::Unknown) &&
1719 LHSFolds != RHSFolds) {
1720 // Expand the select.
1729 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
1730 BB->getParent(), BB);
1731 // Move the unconditional branch to NewBB.
1732 PredTerm->removeFromParent();
1733 NewBB->getInstList().insert(NewBB->end(), PredTerm);
1734 // Create a conditional branch and update PHI nodes.
1735 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
1736 CondLHS->setIncomingValue(I, SI->getFalseValue());
1737 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
1738 // The select is now dead.
1739 SI->eraseFromParent();
1741 // Update any other PHI nodes in BB.
1742 for (BasicBlock::iterator BI = BB->begin();
1743 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
1745 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);