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/GlobalsModRef.h"
22 #include "llvm/Analysis/CFG.h"
23 #include "llvm/Analysis/ConstantFolding.h"
24 #include "llvm/Analysis/InstructionSimplify.h"
25 #include "llvm/Analysis/LazyValueInfo.h"
26 #include "llvm/Analysis/Loads.h"
27 #include "llvm/Analysis/TargetLibraryInfo.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/IntrinsicInst.h"
30 #include "llvm/IR/LLVMContext.h"
31 #include "llvm/IR/Metadata.h"
32 #include "llvm/IR/ValueHandle.h"
33 #include "llvm/Pass.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/raw_ostream.h"
37 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
38 #include "llvm/Transforms/Utils/Local.h"
39 #include "llvm/Transforms/Utils/SSAUpdater.h"
42 #define DEBUG_TYPE "jump-threading"
44 STATISTIC(NumThreads, "Number of jumps threaded");
45 STATISTIC(NumFolds, "Number of terminators folded");
46 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
48 static cl::opt<unsigned>
49 BBDuplicateThreshold("jump-threading-threshold",
50 cl::desc("Max block size to duplicate for jump threading"),
51 cl::init(6), cl::Hidden);
54 // These are at global scope so static functions can use them too.
55 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo;
56 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy;
58 // This is used to keep track of what kind of constant we're currently hoping
60 enum ConstantPreference {
65 /// This pass performs 'jump threading', which looks at blocks that have
66 /// multiple predecessors and multiple successors. If one or more of the
67 /// predecessors of the block can be proven to always jump to one of the
68 /// successors, we forward the edge from the predecessor to the successor by
69 /// duplicating the contents of this block.
71 /// An example of when this can occur is code like this:
78 /// In this case, the unconditional branch at the end of the first if can be
79 /// revectored to the false side of the second if.
81 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.addPreserved<GlobalsAAWrapperPass>();
119 AU.addRequired<TargetLibraryInfoWrapperPass>();
122 void FindLoopHeaders(Function &F);
123 bool ProcessBlock(BasicBlock *BB);
124 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
126 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
127 const SmallVectorImpl<BasicBlock *> &PredBBs);
129 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
130 PredValueInfo &Result,
131 ConstantPreference Preference,
132 Instruction *CxtI = nullptr);
133 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
134 ConstantPreference Preference,
135 Instruction *CxtI = nullptr);
137 bool ProcessBranchOnPHI(PHINode *PN);
138 bool ProcessBranchOnXOR(BinaryOperator *BO);
140 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
141 bool TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB);
145 char JumpThreading::ID = 0;
146 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
147 "Jump Threading", false, false)
148 INITIALIZE_PASS_DEPENDENCY(LazyValueInfo)
149 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
150 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
151 "Jump Threading", false, false)
153 // Public interface to the Jump Threading pass
154 FunctionPass *llvm::createJumpThreadingPass(int Threshold) { return new JumpThreading(Threshold); }
156 /// runOnFunction - Top level algorithm.
158 bool JumpThreading::runOnFunction(Function &F) {
159 if (skipOptnoneFunction(F))
162 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
163 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
164 LVI = &getAnalysis<LazyValueInfo>();
166 // Remove unreachable blocks from function as they may result in infinite
167 // loop. We do threading if we found something profitable. Jump threading a
168 // branch can create other opportunities. If these opportunities form a cycle
169 // i.e. if any jump treading is undoing previous threading in the path, then
170 // we will loop forever. We take care of this issue by not jump threading for
171 // back edges. This works for normal cases but not for unreachable blocks as
172 // they may have cycle with no back edge.
173 removeUnreachableBlocks(F);
177 bool Changed, EverChanged = false;
180 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
182 // Thread all of the branches we can over this block.
183 while (ProcessBlock(BB))
188 // If the block is trivially dead, zap it. This eliminates the successor
189 // edges which simplifies the CFG.
190 if (pred_empty(BB) &&
191 BB != &BB->getParent()->getEntryBlock()) {
192 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
193 << "' with terminator: " << *BB->getTerminator() << '\n');
194 LoopHeaders.erase(BB);
201 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
203 // Can't thread an unconditional jump, but if the block is "almost
204 // empty", we can replace uses of it with uses of the successor and make
206 if (BI && BI->isUnconditional() &&
207 BB != &BB->getParent()->getEntryBlock() &&
208 // If the terminator is the only non-phi instruction, try to nuke it.
209 BB->getFirstNonPHIOrDbg()->isTerminator()) {
210 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
211 // block, we have to make sure it isn't in the LoopHeaders set. We
212 // reinsert afterward if needed.
213 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
214 BasicBlock *Succ = BI->getSuccessor(0);
216 // FIXME: It is always conservatively correct to drop the info
217 // for a block even if it doesn't get erased. This isn't totally
218 // awesome, but it allows us to use AssertingVH to prevent nasty
219 // dangling pointer issues within LazyValueInfo.
221 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
223 // If we deleted BB and BB was the header of a loop, then the
224 // successor is now the header of the loop.
228 if (ErasedFromLoopHeaders)
229 LoopHeaders.insert(BB);
232 EverChanged |= Changed;
239 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
240 /// thread across it. Stop scanning the block when passing the threshold.
241 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB,
242 unsigned Threshold) {
243 /// Ignore PHI nodes, these will be flattened when duplication happens.
244 BasicBlock::const_iterator I = BB->getFirstNonPHI();
246 // FIXME: THREADING will delete values that are just used to compute the
247 // branch, so they shouldn't count against the duplication cost.
249 // Sum up the cost of each instruction until we get to the terminator. Don't
250 // include the terminator because the copy won't include it.
252 for (; !isa<TerminatorInst>(I); ++I) {
254 // Stop scanning the block if we've reached the threshold.
255 if (Size > Threshold)
258 // Debugger intrinsics don't incur code size.
259 if (isa<DbgInfoIntrinsic>(I)) continue;
261 // If this is a pointer->pointer bitcast, it is free.
262 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
265 // Bail out if this instruction gives back a token type, it is not possible
266 // to duplicate it if it used outside this BB.
267 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
270 // All other instructions count for at least one unit.
273 // Calls are more expensive. If they are non-intrinsic calls, we model them
274 // as having cost of 4. If they are a non-vector intrinsic, we model them
275 // as having cost of 2 total, and if they are a vector intrinsic, we model
276 // them as having cost 1.
277 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
278 if (CI->cannotDuplicate() || CI->isConvergent())
279 // Blocks with NoDuplicate are modelled as having infinite cost, so they
280 // are never duplicated.
282 else if (!isa<IntrinsicInst>(CI))
284 else if (!CI->getType()->isVectorTy())
289 // Threading through a switch statement is particularly profitable. If this
290 // block ends in a switch, decrease its cost to make it more likely to happen.
291 if (isa<SwitchInst>(I))
292 Size = Size > 6 ? Size-6 : 0;
294 // The same holds for indirect branches, but slightly more so.
295 if (isa<IndirectBrInst>(I))
296 Size = Size > 8 ? Size-8 : 0;
301 /// FindLoopHeaders - We do not want jump threading to turn proper loop
302 /// structures into irreducible loops. Doing this breaks up the loop nesting
303 /// hierarchy and pessimizes later transformations. To prevent this from
304 /// happening, we first have to find the loop headers. Here we approximate this
305 /// by finding targets of backedges in the CFG.
307 /// Note that there definitely are cases when we want to allow threading of
308 /// edges across a loop header. For example, threading a jump from outside the
309 /// loop (the preheader) to an exit block of the loop is definitely profitable.
310 /// It is also almost always profitable to thread backedges from within the loop
311 /// to exit blocks, and is often profitable to thread backedges to other blocks
312 /// within the loop (forming a nested loop). This simple analysis is not rich
313 /// enough to track all of these properties and keep it up-to-date as the CFG
314 /// mutates, so we don't allow any of these transformations.
316 void JumpThreading::FindLoopHeaders(Function &F) {
317 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
318 FindFunctionBackedges(F, Edges);
320 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
321 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
324 /// getKnownConstant - Helper method to determine if we can thread over a
325 /// terminator with the given value as its condition, and if so what value to
326 /// use for that. What kind of value this is depends on whether we want an
327 /// integer or a block address, but an undef is always accepted.
328 /// Returns null if Val is null or not an appropriate constant.
329 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
333 // Undef is "known" enough.
334 if (UndefValue *U = dyn_cast<UndefValue>(Val))
337 if (Preference == WantBlockAddress)
338 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
340 return dyn_cast<ConstantInt>(Val);
343 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
344 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
345 /// in any of our predecessors. If so, return the known list of value and pred
346 /// BB in the result vector.
348 /// This returns true if there were any known values.
351 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
352 ConstantPreference Preference,
354 // This method walks up use-def chains recursively. Because of this, we could
355 // get into an infinite loop going around loops in the use-def chain. To
356 // prevent this, keep track of what (value, block) pairs we've already visited
357 // and terminate the search if we loop back to them
358 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
361 // An RAII help to remove this pair from the recursion set once the recursion
362 // stack pops back out again.
363 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
365 // If V is a constant, then it is known in all predecessors.
366 if (Constant *KC = getKnownConstant(V, Preference)) {
367 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
368 Result.push_back(std::make_pair(KC, *PI));
373 // If V is a non-instruction value, or an instruction in a different block,
374 // then it can't be derived from a PHI.
375 Instruction *I = dyn_cast<Instruction>(V);
376 if (!I || I->getParent() != BB) {
378 // Okay, if this is a live-in value, see if it has a known value at the end
379 // of any of our predecessors.
381 // FIXME: This should be an edge property, not a block end property.
382 /// TODO: Per PR2563, we could infer value range information about a
383 /// predecessor based on its terminator.
385 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
386 // "I" is a non-local compare-with-a-constant instruction. This would be
387 // able to handle value inequalities better, for example if the compare is
388 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
389 // Perhaps getConstantOnEdge should be smart enough to do this?
391 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
393 // If the value is known by LazyValueInfo to be a constant in a
394 // predecessor, use that information to try to thread this block.
395 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
396 if (Constant *KC = getKnownConstant(PredCst, Preference))
397 Result.push_back(std::make_pair(KC, P));
400 return !Result.empty();
403 /// If I is a PHI node, then we know the incoming values for any constants.
404 if (PHINode *PN = dyn_cast<PHINode>(I)) {
405 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
406 Value *InVal = PN->getIncomingValue(i);
407 if (Constant *KC = getKnownConstant(InVal, Preference)) {
408 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
410 Constant *CI = LVI->getConstantOnEdge(InVal,
411 PN->getIncomingBlock(i),
413 if (Constant *KC = getKnownConstant(CI, Preference))
414 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
418 return !Result.empty();
421 PredValueInfoTy LHSVals, RHSVals;
423 // Handle some boolean conditions.
424 if (I->getType()->getPrimitiveSizeInBits() == 1) {
425 assert(Preference == WantInteger && "One-bit non-integer type?");
427 // X & false -> false
428 if (I->getOpcode() == Instruction::Or ||
429 I->getOpcode() == Instruction::And) {
430 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
432 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
435 if (LHSVals.empty() && RHSVals.empty())
438 ConstantInt *InterestingVal;
439 if (I->getOpcode() == Instruction::Or)
440 InterestingVal = ConstantInt::getTrue(I->getContext());
442 InterestingVal = ConstantInt::getFalse(I->getContext());
444 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
446 // Scan for the sentinel. If we find an undef, force it to the
447 // interesting value: x|undef -> true and x&undef -> false.
448 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
449 if (LHSVals[i].first == InterestingVal ||
450 isa<UndefValue>(LHSVals[i].first)) {
451 Result.push_back(LHSVals[i]);
452 Result.back().first = InterestingVal;
453 LHSKnownBBs.insert(LHSVals[i].second);
455 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
456 if (RHSVals[i].first == InterestingVal ||
457 isa<UndefValue>(RHSVals[i].first)) {
458 // If we already inferred a value for this block on the LHS, don't
460 if (!LHSKnownBBs.count(RHSVals[i].second)) {
461 Result.push_back(RHSVals[i]);
462 Result.back().first = InterestingVal;
466 return !Result.empty();
469 // Handle the NOT form of XOR.
470 if (I->getOpcode() == Instruction::Xor &&
471 isa<ConstantInt>(I->getOperand(1)) &&
472 cast<ConstantInt>(I->getOperand(1))->isOne()) {
473 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
478 // Invert the known values.
479 for (unsigned i = 0, e = Result.size(); i != e; ++i)
480 Result[i].first = ConstantExpr::getNot(Result[i].first);
485 // Try to simplify some other binary operator values.
486 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
487 assert(Preference != WantBlockAddress
488 && "A binary operator creating a block address?");
489 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
490 PredValueInfoTy LHSVals;
491 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
494 // Try to use constant folding to simplify the binary operator.
495 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
496 Constant *V = LHSVals[i].first;
497 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
499 if (Constant *KC = getKnownConstant(Folded, WantInteger))
500 Result.push_back(std::make_pair(KC, LHSVals[i].second));
504 return !Result.empty();
507 // Handle compare with phi operand, where the PHI is defined in this block.
508 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
509 assert(Preference == WantInteger && "Compares only produce integers");
510 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
511 if (PN && PN->getParent() == BB) {
512 const DataLayout &DL = PN->getModule()->getDataLayout();
513 // We can do this simplification if any comparisons fold to true or false.
515 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
516 BasicBlock *PredBB = PN->getIncomingBlock(i);
517 Value *LHS = PN->getIncomingValue(i);
518 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
520 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL);
522 if (!isa<Constant>(RHS))
525 LazyValueInfo::Tristate
526 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
527 cast<Constant>(RHS), PredBB, BB,
529 if (ResT == LazyValueInfo::Unknown)
531 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
534 if (Constant *KC = getKnownConstant(Res, WantInteger))
535 Result.push_back(std::make_pair(KC, PredBB));
538 return !Result.empty();
541 // If comparing a live-in value against a constant, see if we know the
542 // live-in value on any predecessors.
543 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
544 if (!isa<Instruction>(Cmp->getOperand(0)) ||
545 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
546 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
548 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
550 // If the value is known by LazyValueInfo to be a constant in a
551 // predecessor, use that information to try to thread this block.
552 LazyValueInfo::Tristate Res =
553 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
554 RHSCst, P, BB, CxtI ? CxtI : Cmp);
555 if (Res == LazyValueInfo::Unknown)
558 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
559 Result.push_back(std::make_pair(ResC, P));
562 return !Result.empty();
565 // Try to find a constant value for the LHS of a comparison,
566 // and evaluate it statically if we can.
567 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
568 PredValueInfoTy LHSVals;
569 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
572 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
573 Constant *V = LHSVals[i].first;
574 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
576 if (Constant *KC = getKnownConstant(Folded, WantInteger))
577 Result.push_back(std::make_pair(KC, LHSVals[i].second));
580 return !Result.empty();
585 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
586 // Handle select instructions where at least one operand is a known constant
587 // and we can figure out the condition value for any predecessor block.
588 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
589 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
590 PredValueInfoTy Conds;
591 if ((TrueVal || FalseVal) &&
592 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
593 WantInteger, CxtI)) {
594 for (unsigned i = 0, e = Conds.size(); i != e; ++i) {
595 Constant *Cond = Conds[i].first;
597 // Figure out what value to use for the condition.
599 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
601 KnownCond = CI->isOne();
603 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
604 // Either operand will do, so be sure to pick the one that's a known
606 // FIXME: Do this more cleverly if both values are known constants?
607 KnownCond = (TrueVal != nullptr);
610 // See if the select has a known constant value for this predecessor.
611 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
612 Result.push_back(std::make_pair(Val, Conds[i].second));
615 return !Result.empty();
619 // If all else fails, see if LVI can figure out a constant value for us.
620 Constant *CI = LVI->getConstant(V, BB, CxtI);
621 if (Constant *KC = getKnownConstant(CI, Preference)) {
622 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
623 Result.push_back(std::make_pair(KC, *PI));
626 return !Result.empty();
631 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
632 /// in an undefined jump, decide which block is best to revector to.
634 /// Since we can pick an arbitrary destination, we pick the successor with the
635 /// fewest predecessors. This should reduce the in-degree of the others.
637 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
638 TerminatorInst *BBTerm = BB->getTerminator();
639 unsigned MinSucc = 0;
640 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
641 // Compute the successor with the minimum number of predecessors.
642 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
643 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
644 TestBB = BBTerm->getSuccessor(i);
645 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
646 if (NumPreds < MinNumPreds) {
648 MinNumPreds = NumPreds;
655 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
656 if (!BB->hasAddressTaken()) return false;
658 // If the block has its address taken, it may be a tree of dead constants
659 // hanging off of it. These shouldn't keep the block alive.
660 BlockAddress *BA = BlockAddress::get(BB);
661 BA->removeDeadConstantUsers();
662 return !BA->use_empty();
665 /// ProcessBlock - If there are any predecessors whose control can be threaded
666 /// through to a successor, transform them now.
667 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
668 // If the block is trivially dead, just return and let the caller nuke it.
669 // This simplifies other transformations.
670 if (pred_empty(BB) &&
671 BB != &BB->getParent()->getEntryBlock())
674 // If this block has a single predecessor, and if that pred has a single
675 // successor, merge the blocks. This encourages recursive jump threading
676 // because now the condition in this block can be threaded through
677 // predecessors of our predecessor block.
678 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
679 const TerminatorInst *TI = SinglePred->getTerminator();
680 if (!TI->isExceptional() && TI->getNumSuccessors() == 1 &&
681 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
682 // If SinglePred was a loop header, BB becomes one.
683 if (LoopHeaders.erase(SinglePred))
684 LoopHeaders.insert(BB);
686 LVI->eraseBlock(SinglePred);
687 MergeBasicBlockIntoOnlyPred(BB);
693 // What kind of constant we're looking for.
694 ConstantPreference Preference = WantInteger;
696 // Look to see if the terminator is a conditional branch, switch or indirect
697 // branch, if not we can't thread it.
699 Instruction *Terminator = BB->getTerminator();
700 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
701 // Can't thread an unconditional jump.
702 if (BI->isUnconditional()) return false;
703 Condition = BI->getCondition();
704 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
705 Condition = SI->getCondition();
706 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
707 // Can't thread indirect branch with no successors.
708 if (IB->getNumSuccessors() == 0) return false;
709 Condition = IB->getAddress()->stripPointerCasts();
710 Preference = WantBlockAddress;
712 return false; // Must be an invoke.
715 // Run constant folding to see if we can reduce the condition to a simple
717 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
719 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
721 I->replaceAllUsesWith(SimpleVal);
722 I->eraseFromParent();
723 Condition = SimpleVal;
727 // If the terminator is branching on an undef, we can pick any of the
728 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
729 if (isa<UndefValue>(Condition)) {
730 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
732 // Fold the branch/switch.
733 TerminatorInst *BBTerm = BB->getTerminator();
734 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
735 if (i == BestSucc) continue;
736 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
739 DEBUG(dbgs() << " In block '" << BB->getName()
740 << "' folding undef terminator: " << *BBTerm << '\n');
741 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
742 BBTerm->eraseFromParent();
746 // If the terminator of this block is branching on a constant, simplify the
747 // terminator to an unconditional branch. This can occur due to threading in
749 if (getKnownConstant(Condition, Preference)) {
750 DEBUG(dbgs() << " In block '" << BB->getName()
751 << "' folding terminator: " << *BB->getTerminator() << '\n');
753 ConstantFoldTerminator(BB, true);
757 Instruction *CondInst = dyn_cast<Instruction>(Condition);
759 // All the rest of our checks depend on the condition being an instruction.
761 // FIXME: Unify this with code below.
762 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
768 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
769 // If we're branching on a conditional, LVI might be able to determine
770 // it's value at the branch instruction. We only handle comparisons
771 // against a constant at this time.
772 // TODO: This should be extended to handle switches as well.
773 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
774 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
775 if (CondBr && CondConst && CondBr->isConditional()) {
776 LazyValueInfo::Tristate Ret =
777 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
779 if (Ret != LazyValueInfo::Unknown) {
780 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
781 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
782 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
783 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
784 CondBr->eraseFromParent();
785 if (CondCmp->use_empty())
786 CondCmp->eraseFromParent();
787 else if (CondCmp->getParent() == BB) {
788 // If the fact we just learned is true for all uses of the
789 // condition, replace it with a constant value
790 auto *CI = Ret == LazyValueInfo::True ?
791 ConstantInt::getTrue(CondCmp->getType()) :
792 ConstantInt::getFalse(CondCmp->getType());
793 CondCmp->replaceAllUsesWith(CI);
794 CondCmp->eraseFromParent();
800 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB))
804 // Check for some cases that are worth simplifying. Right now we want to look
805 // for loads that are used by a switch or by the condition for the branch. If
806 // we see one, check to see if it's partially redundant. If so, insert a PHI
807 // which can then be used to thread the values.
809 Value *SimplifyValue = CondInst;
810 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
811 if (isa<Constant>(CondCmp->getOperand(1)))
812 SimplifyValue = CondCmp->getOperand(0);
814 // TODO: There are other places where load PRE would be profitable, such as
815 // more complex comparisons.
816 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
817 if (SimplifyPartiallyRedundantLoad(LI))
821 // Handle a variety of cases where we are branching on something derived from
822 // a PHI node in the current block. If we can prove that any predecessors
823 // compute a predictable value based on a PHI node, thread those predecessors.
825 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
828 // If this is an otherwise-unfoldable branch on a phi node in the current
829 // block, see if we can simplify.
830 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
831 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
832 return ProcessBranchOnPHI(PN);
835 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
836 if (CondInst->getOpcode() == Instruction::Xor &&
837 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
838 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
841 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
842 // "(X == 4)", thread through this block.
847 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
848 /// load instruction, eliminate it by replacing it with a PHI node. This is an
849 /// important optimization that encourages jump threading, and needs to be run
850 /// interlaced with other jump threading tasks.
851 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
852 // Don't hack volatile/atomic loads.
853 if (!LI->isSimple()) return false;
855 // If the load is defined in a block with exactly one predecessor, it can't be
856 // partially redundant.
857 BasicBlock *LoadBB = LI->getParent();
858 if (LoadBB->getSinglePredecessor())
861 // If the load is defined in an EH pad, it can't be partially redundant,
862 // because the edges between the invoke and the EH pad cannot have other
863 // instructions between them.
864 if (LoadBB->isEHPad())
867 Value *LoadedPtr = LI->getOperand(0);
869 // If the loaded operand is defined in the LoadBB, it can't be available.
870 // TODO: Could do simple PHI translation, that would be fun :)
871 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
872 if (PtrOp->getParent() == LoadBB)
875 // Scan a few instructions up from the load, to see if it is obviously live at
876 // the entry to its block.
877 BasicBlock::iterator BBIt = LI;
879 if (Value *AvailableVal =
880 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
881 // If the value if the load is locally available within the block, just use
882 // it. This frequently occurs for reg2mem'd allocas.
883 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
885 // If the returned value is the load itself, replace with an undef. This can
886 // only happen in dead loops.
887 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
888 if (AvailableVal->getType() != LI->getType())
890 CastInst::CreateBitOrPointerCast(AvailableVal, LI->getType(), "", LI);
891 LI->replaceAllUsesWith(AvailableVal);
892 LI->eraseFromParent();
896 // Otherwise, if we scanned the whole block and got to the top of the block,
897 // we know the block is locally transparent to the load. If not, something
898 // might clobber its value.
899 if (BBIt != LoadBB->begin())
902 // If all of the loads and stores that feed the value have the same AA tags,
903 // then we can propagate them onto any newly inserted loads.
905 LI->getAAMetadata(AATags);
907 SmallPtrSet<BasicBlock*, 8> PredsScanned;
908 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
909 AvailablePredsTy AvailablePreds;
910 BasicBlock *OneUnavailablePred = nullptr;
912 // If we got here, the loaded value is transparent through to the start of the
913 // block. Check to see if it is available in any of the predecessor blocks.
914 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
916 BasicBlock *PredBB = *PI;
918 // If we already scanned this predecessor, skip it.
919 if (!PredsScanned.insert(PredBB).second)
922 // Scan the predecessor to see if the value is available in the pred.
923 BBIt = PredBB->end();
924 AAMDNodes ThisAATags;
925 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6,
926 nullptr, &ThisAATags);
927 if (!PredAvailable) {
928 OneUnavailablePred = PredBB;
932 // If AA tags disagree or are not present, forget about them.
933 if (AATags != ThisAATags) AATags = AAMDNodes();
935 // If so, this load is partially redundant. Remember this info so that we
936 // can create a PHI node.
937 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
940 // If the loaded value isn't available in any predecessor, it isn't partially
942 if (AvailablePreds.empty()) return false;
944 // Okay, the loaded value is available in at least one (and maybe all!)
945 // predecessors. If the value is unavailable in more than one unique
946 // predecessor, we want to insert a merge block for those common predecessors.
947 // This ensures that we only have to insert one reload, thus not increasing
949 BasicBlock *UnavailablePred = nullptr;
951 // If there is exactly one predecessor where the value is unavailable, the
952 // already computed 'OneUnavailablePred' block is it. If it ends in an
953 // unconditional branch, we know that it isn't a critical edge.
954 if (PredsScanned.size() == AvailablePreds.size()+1 &&
955 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
956 UnavailablePred = OneUnavailablePred;
957 } else if (PredsScanned.size() != AvailablePreds.size()) {
958 // Otherwise, we had multiple unavailable predecessors or we had a critical
959 // edge from the one.
960 SmallVector<BasicBlock*, 8> PredsToSplit;
961 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
963 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
964 AvailablePredSet.insert(AvailablePreds[i].first);
966 // Add all the unavailable predecessors to the PredsToSplit list.
967 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
970 // If the predecessor is an indirect goto, we can't split the edge.
971 if (isa<IndirectBrInst>(P->getTerminator()))
974 if (!AvailablePredSet.count(P))
975 PredsToSplit.push_back(P);
978 // Split them out to their own block.
980 SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split");
983 // If the value isn't available in all predecessors, then there will be
984 // exactly one where it isn't available. Insert a load on that edge and add
985 // it to the AvailablePreds list.
986 if (UnavailablePred) {
987 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
988 "Can't handle critical edge here!");
989 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
991 UnavailablePred->getTerminator());
992 NewVal->setDebugLoc(LI->getDebugLoc());
994 NewVal->setAAMetadata(AATags);
996 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
999 // Now we know that each predecessor of this block has a value in
1000 // AvailablePreds, sort them for efficient access as we're walking the preds.
1001 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1003 // Create a PHI node at the start of the block for the PRE'd load value.
1004 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1005 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
1008 PN->setDebugLoc(LI->getDebugLoc());
1010 // Insert new entries into the PHI for each predecessor. A single block may
1011 // have multiple entries here.
1012 for (pred_iterator PI = PB; PI != PE; ++PI) {
1013 BasicBlock *P = *PI;
1014 AvailablePredsTy::iterator I =
1015 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1016 std::make_pair(P, (Value*)nullptr));
1018 assert(I != AvailablePreds.end() && I->first == P &&
1019 "Didn't find entry for predecessor!");
1021 // If we have an available predecessor but it requires casting, insert the
1022 // cast in the predecessor and use the cast. Note that we have to update the
1023 // AvailablePreds vector as we go so that all of the PHI entries for this
1024 // predecessor use the same bitcast.
1025 Value *&PredV = I->second;
1026 if (PredV->getType() != LI->getType())
1027 PredV = CastInst::CreateBitOrPointerCast(PredV, LI->getType(), "",
1028 P->getTerminator());
1030 PN->addIncoming(PredV, I->first);
1033 //cerr << "PRE: " << *LI << *PN << "\n";
1035 LI->replaceAllUsesWith(PN);
1036 LI->eraseFromParent();
1041 /// FindMostPopularDest - The specified list contains multiple possible
1042 /// threadable destinations. Pick the one that occurs the most frequently in
1045 FindMostPopularDest(BasicBlock *BB,
1046 const SmallVectorImpl<std::pair<BasicBlock*,
1047 BasicBlock*> > &PredToDestList) {
1048 assert(!PredToDestList.empty());
1050 // Determine popularity. If there are multiple possible destinations, we
1051 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1052 // blocks with known and real destinations to threading undef. We'll handle
1053 // them later if interesting.
1054 DenseMap<BasicBlock*, unsigned> DestPopularity;
1055 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1056 if (PredToDestList[i].second)
1057 DestPopularity[PredToDestList[i].second]++;
1059 // Find the most popular dest.
1060 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1061 BasicBlock *MostPopularDest = DPI->first;
1062 unsigned Popularity = DPI->second;
1063 SmallVector<BasicBlock*, 4> SamePopularity;
1065 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1066 // If the popularity of this entry isn't higher than the popularity we've
1067 // seen so far, ignore it.
1068 if (DPI->second < Popularity)
1070 else if (DPI->second == Popularity) {
1071 // If it is the same as what we've seen so far, keep track of it.
1072 SamePopularity.push_back(DPI->first);
1074 // If it is more popular, remember it.
1075 SamePopularity.clear();
1076 MostPopularDest = DPI->first;
1077 Popularity = DPI->second;
1081 // Okay, now we know the most popular destination. If there is more than one
1082 // destination, we need to determine one. This is arbitrary, but we need
1083 // to make a deterministic decision. Pick the first one that appears in the
1085 if (!SamePopularity.empty()) {
1086 SamePopularity.push_back(MostPopularDest);
1087 TerminatorInst *TI = BB->getTerminator();
1088 for (unsigned i = 0; ; ++i) {
1089 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1091 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1092 TI->getSuccessor(i)) == SamePopularity.end())
1095 MostPopularDest = TI->getSuccessor(i);
1100 // Okay, we have finally picked the most popular destination.
1101 return MostPopularDest;
1104 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1105 ConstantPreference Preference,
1106 Instruction *CxtI) {
1107 // If threading this would thread across a loop header, don't even try to
1109 if (LoopHeaders.count(BB))
1112 PredValueInfoTy PredValues;
1113 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1116 assert(!PredValues.empty() &&
1117 "ComputeValueKnownInPredecessors returned true with no values");
1119 DEBUG(dbgs() << "IN BB: " << *BB;
1120 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1121 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1122 << *PredValues[i].first
1123 << " for pred '" << PredValues[i].second->getName() << "'.\n";
1126 // Decide what we want to thread through. Convert our list of known values to
1127 // a list of known destinations for each pred. This also discards duplicate
1128 // predecessors and keeps track of the undefined inputs (which are represented
1129 // as a null dest in the PredToDestList).
1130 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1131 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1133 BasicBlock *OnlyDest = nullptr;
1134 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1136 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1137 BasicBlock *Pred = PredValues[i].second;
1138 if (!SeenPreds.insert(Pred).second)
1139 continue; // Duplicate predecessor entry.
1141 // If the predecessor ends with an indirect goto, we can't change its
1143 if (isa<IndirectBrInst>(Pred->getTerminator()))
1146 Constant *Val = PredValues[i].first;
1149 if (isa<UndefValue>(Val))
1151 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1152 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1153 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1154 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor();
1156 assert(isa<IndirectBrInst>(BB->getTerminator())
1157 && "Unexpected terminator");
1158 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1161 // If we have exactly one destination, remember it for efficiency below.
1162 if (PredToDestList.empty())
1164 else if (OnlyDest != DestBB)
1165 OnlyDest = MultipleDestSentinel;
1167 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1170 // If all edges were unthreadable, we fail.
1171 if (PredToDestList.empty())
1174 // Determine which is the most common successor. If we have many inputs and
1175 // this block is a switch, we want to start by threading the batch that goes
1176 // to the most popular destination first. If we only know about one
1177 // threadable destination (the common case) we can avoid this.
1178 BasicBlock *MostPopularDest = OnlyDest;
1180 if (MostPopularDest == MultipleDestSentinel)
1181 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1183 // Now that we know what the most popular destination is, factor all
1184 // predecessors that will jump to it into a single predecessor.
1185 SmallVector<BasicBlock*, 16> PredsToFactor;
1186 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1187 if (PredToDestList[i].second == MostPopularDest) {
1188 BasicBlock *Pred = PredToDestList[i].first;
1190 // This predecessor may be a switch or something else that has multiple
1191 // edges to the block. Factor each of these edges by listing them
1192 // according to # occurrences in PredsToFactor.
1193 TerminatorInst *PredTI = Pred->getTerminator();
1194 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1195 if (PredTI->getSuccessor(i) == BB)
1196 PredsToFactor.push_back(Pred);
1199 // If the threadable edges are branching on an undefined value, we get to pick
1200 // the destination that these predecessors should get to.
1201 if (!MostPopularDest)
1202 MostPopularDest = BB->getTerminator()->
1203 getSuccessor(GetBestDestForJumpOnUndef(BB));
1205 // Ok, try to thread it!
1206 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1209 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1210 /// a PHI node in the current block. See if there are any simplifications we
1211 /// can do based on inputs to the phi node.
1213 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1214 BasicBlock *BB = PN->getParent();
1216 // TODO: We could make use of this to do it once for blocks with common PHI
1218 SmallVector<BasicBlock*, 1> PredBBs;
1221 // If any of the predecessor blocks end in an unconditional branch, we can
1222 // *duplicate* the conditional branch into that block in order to further
1223 // encourage jump threading and to eliminate cases where we have branch on a
1224 // phi of an icmp (branch on icmp is much better).
1225 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1226 BasicBlock *PredBB = PN->getIncomingBlock(i);
1227 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1228 if (PredBr->isUnconditional()) {
1229 PredBBs[0] = PredBB;
1230 // Try to duplicate BB into PredBB.
1231 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1239 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1240 /// a xor instruction in the current block. See if there are any
1241 /// simplifications we can do based on inputs to the xor.
1243 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1244 BasicBlock *BB = BO->getParent();
1246 // If either the LHS or RHS of the xor is a constant, don't do this
1248 if (isa<ConstantInt>(BO->getOperand(0)) ||
1249 isa<ConstantInt>(BO->getOperand(1)))
1252 // If the first instruction in BB isn't a phi, we won't be able to infer
1253 // anything special about any particular predecessor.
1254 if (!isa<PHINode>(BB->front()))
1257 // If we have a xor as the branch input to this block, and we know that the
1258 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1259 // the condition into the predecessor and fix that value to true, saving some
1260 // logical ops on that path and encouraging other paths to simplify.
1262 // This copies something like this:
1265 // %X = phi i1 [1], [%X']
1266 // %Y = icmp eq i32 %A, %B
1267 // %Z = xor i1 %X, %Y
1272 // %Y = icmp ne i32 %A, %B
1275 PredValueInfoTy XorOpValues;
1277 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1279 assert(XorOpValues.empty());
1280 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1286 assert(!XorOpValues.empty() &&
1287 "ComputeValueKnownInPredecessors returned true with no values");
1289 // Scan the information to see which is most popular: true or false. The
1290 // predecessors can be of the set true, false, or undef.
1291 unsigned NumTrue = 0, NumFalse = 0;
1292 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1293 if (isa<UndefValue>(XorOpValues[i].first))
1294 // Ignore undefs for the count.
1296 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1302 // Determine which value to split on, true, false, or undef if neither.
1303 ConstantInt *SplitVal = nullptr;
1304 if (NumTrue > NumFalse)
1305 SplitVal = ConstantInt::getTrue(BB->getContext());
1306 else if (NumTrue != 0 || NumFalse != 0)
1307 SplitVal = ConstantInt::getFalse(BB->getContext());
1309 // Collect all of the blocks that this can be folded into so that we can
1310 // factor this once and clone it once.
1311 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1312 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1313 if (XorOpValues[i].first != SplitVal &&
1314 !isa<UndefValue>(XorOpValues[i].first))
1317 BlocksToFoldInto.push_back(XorOpValues[i].second);
1320 // If we inferred a value for all of the predecessors, then duplication won't
1321 // help us. However, we can just replace the LHS or RHS with the constant.
1322 if (BlocksToFoldInto.size() ==
1323 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1325 // If all preds provide undef, just nuke the xor, because it is undef too.
1326 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1327 BO->eraseFromParent();
1328 } else if (SplitVal->isZero()) {
1329 // If all preds provide 0, replace the xor with the other input.
1330 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1331 BO->eraseFromParent();
1333 // If all preds provide 1, set the computed value to 1.
1334 BO->setOperand(!isLHS, SplitVal);
1340 // Try to duplicate BB into PredBB.
1341 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1345 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1346 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1347 /// NewPred using the entries from OldPred (suitably mapped).
1348 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1349 BasicBlock *OldPred,
1350 BasicBlock *NewPred,
1351 DenseMap<Instruction*, Value*> &ValueMap) {
1352 for (BasicBlock::iterator PNI = PHIBB->begin();
1353 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1354 // Ok, we have a PHI node. Figure out what the incoming value was for the
1356 Value *IV = PN->getIncomingValueForBlock(OldPred);
1358 // Remap the value if necessary.
1359 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1360 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1361 if (I != ValueMap.end())
1365 PN->addIncoming(IV, NewPred);
1369 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1370 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1371 /// across BB. Transform the IR to reflect this change.
1372 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1373 const SmallVectorImpl<BasicBlock*> &PredBBs,
1374 BasicBlock *SuccBB) {
1375 // If threading to the same block as we come from, we would infinite loop.
1377 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1378 << "' - would thread to self!\n");
1382 // If threading this would thread across a loop header, don't thread the edge.
1383 // See the comments above FindLoopHeaders for justifications and caveats.
1384 if (LoopHeaders.count(BB)) {
1385 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1386 << "' to dest BB '" << SuccBB->getName()
1387 << "' - it might create an irreducible loop!\n");
1391 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, BBDupThreshold);
1392 if (JumpThreadCost > BBDupThreshold) {
1393 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1394 << "' - Cost is too high: " << JumpThreadCost << "\n");
1398 // And finally, do it! Start by factoring the predecessors is needed.
1400 if (PredBBs.size() == 1)
1401 PredBB = PredBBs[0];
1403 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1404 << " common predecessors.\n");
1405 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm");
1408 // And finally, do it!
1409 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1410 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1411 << ", across block:\n "
1414 LVI->threadEdge(PredBB, BB, SuccBB);
1416 // We are going to have to map operands from the original BB block to the new
1417 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1418 // account for entry from PredBB.
1419 DenseMap<Instruction*, Value*> ValueMapping;
1421 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1422 BB->getName()+".thread",
1423 BB->getParent(), BB);
1424 NewBB->moveAfter(PredBB);
1426 BasicBlock::iterator BI = BB->begin();
1427 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1428 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1430 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1431 // mapping and using it to remap operands in the cloned instructions.
1432 for (; !isa<TerminatorInst>(BI); ++BI) {
1433 Instruction *New = BI->clone();
1434 New->setName(BI->getName());
1435 NewBB->getInstList().push_back(New);
1436 ValueMapping[BI] = New;
1438 // Remap operands to patch up intra-block references.
1439 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1440 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1441 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1442 if (I != ValueMapping.end())
1443 New->setOperand(i, I->second);
1447 // We didn't copy the terminator from BB over to NewBB, because there is now
1448 // an unconditional jump to SuccBB. Insert the unconditional jump.
1449 BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB);
1450 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1452 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1453 // PHI nodes for NewBB now.
1454 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1456 // If there were values defined in BB that are used outside the block, then we
1457 // now have to update all uses of the value to use either the original value,
1458 // the cloned value, or some PHI derived value. This can require arbitrary
1459 // PHI insertion, of which we are prepared to do, clean these up now.
1460 SSAUpdater SSAUpdate;
1461 SmallVector<Use*, 16> UsesToRename;
1462 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1463 // Scan all uses of this instruction to see if it is used outside of its
1464 // block, and if so, record them in UsesToRename.
1465 for (Use &U : I->uses()) {
1466 Instruction *User = cast<Instruction>(U.getUser());
1467 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1468 if (UserPN->getIncomingBlock(U) == BB)
1470 } else if (User->getParent() == BB)
1473 UsesToRename.push_back(&U);
1476 // If there are no uses outside the block, we're done with this instruction.
1477 if (UsesToRename.empty())
1480 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1482 // We found a use of I outside of BB. Rename all uses of I that are outside
1483 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1484 // with the two values we know.
1485 SSAUpdate.Initialize(I->getType(), I->getName());
1486 SSAUpdate.AddAvailableValue(BB, I);
1487 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1489 while (!UsesToRename.empty())
1490 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1491 DEBUG(dbgs() << "\n");
1495 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1496 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1497 // us to simplify any PHI nodes in BB.
1498 TerminatorInst *PredTerm = PredBB->getTerminator();
1499 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1500 if (PredTerm->getSuccessor(i) == BB) {
1501 BB->removePredecessor(PredBB, true);
1502 PredTerm->setSuccessor(i, NewBB);
1505 // At this point, the IR is fully up to date and consistent. Do a quick scan
1506 // over the new instructions and zap any that are constants or dead. This
1507 // frequently happens because of phi translation.
1508 SimplifyInstructionsInBlock(NewBB, TLI);
1510 // Threaded an edge!
1515 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1516 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1517 /// If we can duplicate the contents of BB up into PredBB do so now, this
1518 /// improves the odds that the branch will be on an analyzable instruction like
1520 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1521 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1522 assert(!PredBBs.empty() && "Can't handle an empty set");
1524 // If BB is a loop header, then duplicating this block outside the loop would
1525 // cause us to transform this into an irreducible loop, don't do this.
1526 // See the comments above FindLoopHeaders for justifications and caveats.
1527 if (LoopHeaders.count(BB)) {
1528 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1529 << "' into predecessor block '" << PredBBs[0]->getName()
1530 << "' - it might create an irreducible loop!\n");
1534 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, BBDupThreshold);
1535 if (DuplicationCost > BBDupThreshold) {
1536 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1537 << "' - Cost is too high: " << DuplicationCost << "\n");
1541 // And finally, do it! Start by factoring the predecessors is needed.
1543 if (PredBBs.size() == 1)
1544 PredBB = PredBBs[0];
1546 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1547 << " common predecessors.\n");
1548 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm");
1551 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1553 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1554 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1555 << DuplicationCost << " block is:" << *BB << "\n");
1557 // Unless PredBB ends with an unconditional branch, split the edge so that we
1558 // can just clone the bits from BB into the end of the new PredBB.
1559 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1561 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
1562 PredBB = SplitEdge(PredBB, BB);
1563 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1566 // We are going to have to map operands from the original BB block into the
1567 // PredBB block. Evaluate PHI nodes in BB.
1568 DenseMap<Instruction*, Value*> ValueMapping;
1570 BasicBlock::iterator BI = BB->begin();
1571 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1572 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1573 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1574 // mapping and using it to remap operands in the cloned instructions.
1575 for (; BI != BB->end(); ++BI) {
1576 Instruction *New = BI->clone();
1578 // Remap operands to patch up intra-block references.
1579 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1580 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1581 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1582 if (I != ValueMapping.end())
1583 New->setOperand(i, I->second);
1586 // If this instruction can be simplified after the operands are updated,
1587 // just use the simplified value instead. This frequently happens due to
1590 SimplifyInstruction(New, BB->getModule()->getDataLayout())) {
1592 ValueMapping[BI] = IV;
1594 // Otherwise, insert the new instruction into the block.
1595 New->setName(BI->getName());
1596 PredBB->getInstList().insert(OldPredBranch, New);
1597 ValueMapping[BI] = New;
1601 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1602 // add entries to the PHI nodes for branch from PredBB now.
1603 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1604 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1606 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1609 // If there were values defined in BB that are used outside the block, then we
1610 // now have to update all uses of the value to use either the original value,
1611 // the cloned value, or some PHI derived value. This can require arbitrary
1612 // PHI insertion, of which we are prepared to do, clean these up now.
1613 SSAUpdater SSAUpdate;
1614 SmallVector<Use*, 16> UsesToRename;
1615 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1616 // Scan all uses of this instruction to see if it is used outside of its
1617 // block, and if so, record them in UsesToRename.
1618 for (Use &U : I->uses()) {
1619 Instruction *User = cast<Instruction>(U.getUser());
1620 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1621 if (UserPN->getIncomingBlock(U) == BB)
1623 } else if (User->getParent() == BB)
1626 UsesToRename.push_back(&U);
1629 // If there are no uses outside the block, we're done with this instruction.
1630 if (UsesToRename.empty())
1633 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1635 // We found a use of I outside of BB. Rename all uses of I that are outside
1636 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1637 // with the two values we know.
1638 SSAUpdate.Initialize(I->getType(), I->getName());
1639 SSAUpdate.AddAvailableValue(BB, I);
1640 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1642 while (!UsesToRename.empty())
1643 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1644 DEBUG(dbgs() << "\n");
1647 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1649 BB->removePredecessor(PredBB, true);
1651 // Remove the unconditional branch at the end of the PredBB block.
1652 OldPredBranch->eraseFromParent();
1658 /// TryToUnfoldSelect - Look for blocks of the form
1664 /// %p = phi [%a, %bb] ...
1668 /// And expand the select into a branch structure if one of its arms allows %c
1669 /// to be folded. This later enables threading from bb1 over bb2.
1670 bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
1671 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1672 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
1673 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
1675 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
1676 CondLHS->getParent() != BB)
1679 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
1680 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
1681 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
1683 // Look if one of the incoming values is a select in the corresponding
1685 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
1688 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
1689 if (!PredTerm || !PredTerm->isUnconditional())
1692 // Now check if one of the select values would allow us to constant fold the
1693 // terminator in BB. We don't do the transform if both sides fold, those
1694 // cases will be threaded in any case.
1695 LazyValueInfo::Tristate LHSFolds =
1696 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
1697 CondRHS, Pred, BB, CondCmp);
1698 LazyValueInfo::Tristate RHSFolds =
1699 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
1700 CondRHS, Pred, BB, CondCmp);
1701 if ((LHSFolds != LazyValueInfo::Unknown ||
1702 RHSFolds != LazyValueInfo::Unknown) &&
1703 LHSFolds != RHSFolds) {
1704 // Expand the select.
1713 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
1714 BB->getParent(), BB);
1715 // Move the unconditional branch to NewBB.
1716 PredTerm->removeFromParent();
1717 NewBB->getInstList().insert(NewBB->end(), PredTerm);
1718 // Create a conditional branch and update PHI nodes.
1719 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
1720 CondLHS->setIncomingValue(I, SI->getFalseValue());
1721 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
1722 // The select is now dead.
1723 SI->eraseFromParent();
1725 // Update any other PHI nodes in BB.
1726 for (BasicBlock::iterator BI = BB->begin();
1727 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
1729 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);