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 #define DEBUG_TYPE "jump-threading"
15 #include "llvm/Transforms/Scalar.h"
16 #include "llvm/IntrinsicInst.h"
17 #include "llvm/Pass.h"
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
21 #include "llvm/Transforms/Utils/Local.h"
22 #include "llvm/Support/CommandLine.h"
23 #include "llvm/Support/Compiler.h"
24 #include "llvm/Support/Debug.h"
25 #include "llvm/ADT/SmallPtrSet.h"
28 STATISTIC(NumThreads, "Number of jumps threaded");
29 STATISTIC(NumFolds, "Number of terminators folded");
31 static cl::opt<unsigned>
32 Threshold("jump-threading-threshold",
33 cl::desc("Max block size to duplicate for jump threading"),
34 cl::init(6), cl::Hidden);
37 /// This pass performs 'jump threading', which looks at blocks that have
38 /// multiple predecessors and multiple successors. If one or more of the
39 /// predecessors of the block can be proven to always jump to one of the
40 /// successors, we forward the edge from the predecessor to the successor by
41 /// duplicating the contents of this block.
43 /// An example of when this can occur is code like this:
50 /// In this case, the unconditional branch at the end of the first if can be
51 /// revectored to the false side of the second if.
53 class VISIBILITY_HIDDEN JumpThreading : public FunctionPass {
55 static char ID; // Pass identification
56 JumpThreading() : FunctionPass(&ID) {}
58 bool runOnFunction(Function &F);
59 bool ProcessBlock(BasicBlock *BB);
60 void ThreadEdge(BasicBlock *BB, BasicBlock *PredBB, BasicBlock *SuccBB);
61 BasicBlock *FactorCommonPHIPreds(PHINode *PN, Constant *CstVal);
63 bool ProcessJumpOnPHI(PHINode *PN);
64 bool ProcessBranchOnLogical(Value *V, BasicBlock *BB, bool isAnd);
65 bool ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB);
67 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
71 char JumpThreading::ID = 0;
72 static RegisterPass<JumpThreading>
73 X("jump-threading", "Jump Threading");
75 // Public interface to the Jump Threading pass
76 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
78 /// runOnFunction - Top level algorithm.
80 bool JumpThreading::runOnFunction(Function &F) {
81 DOUT << "Jump threading on function '" << F.getNameStart() << "'\n";
83 bool AnotherIteration = true, EverChanged = false;
84 while (AnotherIteration) {
85 AnotherIteration = false;
87 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
88 while (ProcessBlock(I))
90 AnotherIteration = Changed;
91 EverChanged |= Changed;
96 /// FactorCommonPHIPreds - If there are multiple preds with the same incoming
97 /// value for the PHI, factor them together so we get one block to thread for
99 /// This is important for things like "phi i1 [true, true, false, true, x]"
100 /// where we only need to clone the block for the true blocks once.
102 BasicBlock *JumpThreading::FactorCommonPHIPreds(PHINode *PN, Constant *CstVal) {
103 SmallVector<BasicBlock*, 16> CommonPreds;
104 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
105 if (PN->getIncomingValue(i) == CstVal)
106 CommonPreds.push_back(PN->getIncomingBlock(i));
108 if (CommonPreds.size() == 1)
109 return CommonPreds[0];
111 DOUT << " Factoring out " << CommonPreds.size()
112 << " common predecessors.\n";
113 return SplitBlockPredecessors(PN->getParent(),
114 &CommonPreds[0], CommonPreds.size(),
119 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
120 /// thread across it.
121 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
122 /// Ignore PHI nodes, these will be flattened when duplication happens.
123 BasicBlock::const_iterator I = BB->getFirstNonPHI();
125 // Sum up the cost of each instruction until we get to the terminator. Don't
126 // include the terminator because the copy won't include it.
128 for (; !isa<TerminatorInst>(I); ++I) {
129 // Debugger intrinsics don't incur code size.
130 if (isa<DbgInfoIntrinsic>(I)) continue;
132 // If this is a pointer->pointer bitcast, it is free.
133 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType()))
136 // All other instructions count for at least one unit.
139 // Calls are more expensive. If they are non-intrinsic calls, we model them
140 // as having cost of 4. If they are a non-vector intrinsic, we model them
141 // as having cost of 2 total, and if they are a vector intrinsic, we model
142 // them as having cost 1.
143 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
144 if (!isa<IntrinsicInst>(CI))
146 else if (isa<VectorType>(CI->getType()))
151 // Threading through a switch statement is particularly profitable. If this
152 // block ends in a switch, decrease its cost to make it more likely to happen.
153 if (isa<SwitchInst>(I))
154 Size = Size > 6 ? Size-6 : 0;
159 /// ProcessBlock - If there are any predecessors whose control can be threaded
160 /// through to a successor, transform them now.
161 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
162 // If this block has a single predecessor, and if that pred has a single
163 // successor, merge the blocks. This encourages recursive jump threading
164 // because now the condition in this block can be threaded through
165 // predecessors of our predecessor block.
166 if (BasicBlock *SinglePred = BB->getSinglePredecessor())
167 if (SinglePred->getTerminator()->getNumSuccessors() == 1) {
168 MergeBasicBlockIntoOnlyPred(BB);
172 // See if this block ends with a branch or switch. If so, see if the
173 // condition is a phi node. If so, and if an entry of the phi node is a
174 // constant, we can thread the block.
176 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
177 // Can't thread an unconditional jump.
178 if (BI->isUnconditional()) return false;
179 Condition = BI->getCondition();
180 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
181 Condition = SI->getCondition();
183 return false; // Must be an invoke.
185 // If the terminator of this block is branching on a constant, simplify the
186 // terminator to an unconditional branch. This can occur due to threading in
188 if (isa<ConstantInt>(Condition)) {
189 DOUT << " In block '" << BB->getNameStart()
190 << "' folding terminator: " << *BB->getTerminator();
192 ConstantFoldTerminator(BB);
196 // If there is only a single predecessor of this block, nothing to fold.
197 if (BB->getSinglePredecessor())
200 // See if this is a phi node in the current block.
201 PHINode *PN = dyn_cast<PHINode>(Condition);
202 if (PN && PN->getParent() == BB)
203 return ProcessJumpOnPHI(PN);
205 // If this is a conditional branch whose condition is and/or of a phi, try to
207 if (BinaryOperator *CondI = dyn_cast<BinaryOperator>(Condition)) {
208 if ((CondI->getOpcode() == Instruction::And ||
209 CondI->getOpcode() == Instruction::Or) &&
210 isa<BranchInst>(BB->getTerminator()) &&
211 ProcessBranchOnLogical(CondI, BB,
212 CondI->getOpcode() == Instruction::And))
216 // If we have "br (phi != 42)" and the phi node has any constant values as
217 // operands, we can thread through this block.
218 if (CmpInst *CondCmp = dyn_cast<CmpInst>(Condition))
219 if (isa<PHINode>(CondCmp->getOperand(0)) &&
220 isa<Constant>(CondCmp->getOperand(1)) &&
221 ProcessBranchOnCompare(CondCmp, BB))
224 // Check for some cases that are worth simplifying. Right now we want to look
225 // for loads that are used by a switch or by the condition for the branch. If
226 // we see one, check to see if it's partially redundant. If so, insert a PHI
227 // which can then be used to thread the values.
229 // This is particularly important because reg2mem inserts loads and stores all
230 // over the place, and this blocks jump threading if we don't zap them.
231 Value *SimplifyValue = Condition;
232 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
233 if (isa<Constant>(CondCmp->getOperand(1)))
234 SimplifyValue = CondCmp->getOperand(0);
236 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
237 if (SimplifyPartiallyRedundantLoad(LI))
240 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
241 // "(X == 4)" thread through this block.
247 /// FindAvailableLoadedValue - Scan backwards from ScanFrom checking to see if
248 /// we have the value at the memory address *Ptr locally available within a
249 /// small number of instructions. If the value is available, return it.
251 /// If not, return the iterator for the last validated instruction that the
252 /// value would be live through. If we scanned the entire block, ScanFrom would
253 /// be left at begin().
255 /// FIXME: Move this to transform utils and use from
256 /// InstCombiner::visitLoadInst. It would also be nice to optionally take AA so
257 /// that GVN could do this.
258 static Value *FindAvailableLoadedValue(Value *Ptr,
260 BasicBlock::iterator &ScanFrom) {
262 unsigned NumToScan = 6;
263 while (ScanFrom != ScanBB->begin()) {
264 // Don't scan huge blocks.
265 if (--NumToScan == 0) return 0;
267 Instruction *Inst = --ScanFrom;
269 // If this is a load of Ptr, the loaded value is available.
270 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
271 if (LI->getOperand(0) == Ptr)
274 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
275 // If this is a store through Ptr, the value is available!
276 if (SI->getOperand(1) == Ptr)
277 return SI->getOperand(0);
279 // If Ptr is an alloca and this is a store to a different alloca, ignore
280 // the store. This is a trivial form of alias analysis that is important
281 // for reg2mem'd code.
282 if ((isa<AllocaInst>(Ptr) || isa<GlobalVariable>(Ptr)) &&
283 (isa<AllocaInst>(SI->getOperand(1)) ||
284 isa<GlobalVariable>(SI->getOperand(1))))
287 // Otherwise the store that may or may not alias the pointer, bail out.
293 // If this is some other instruction that may clobber Ptr, bail out.
294 if (Inst->mayWriteToMemory()) {
295 // May modify the pointer, bail out.
301 // Got to the start of the block, we didn't find it, but are done for this
307 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
308 /// load instruction, eliminate it by replacing it with a PHI node. This is an
309 /// important optimization that encourages jump threading, and needs to be run
310 /// interlaced with other jump threading tasks.
311 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
312 // Don't hack volatile loads.
313 if (LI->isVolatile()) return false;
315 // If the load is defined in a block with exactly one predecessor, it can't be
316 // partially redundant.
317 BasicBlock *LoadBB = LI->getParent();
318 if (LoadBB->getSinglePredecessor())
321 Value *LoadedPtr = LI->getOperand(0);
323 // If the loaded operand is defined in the LoadBB, it can't be available.
324 // FIXME: Could do PHI translation, that would be fun :)
325 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
326 if (PtrOp->getParent() == LoadBB)
329 // Scan a few instructions up from the load, to see if it is obviously live at
330 // the entry to its block.
331 BasicBlock::iterator BBIt = LI;
333 if (Value *AvailableVal = FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt)) {
334 // If the value if the load is locally available within the block, just use
335 // it. This frequently occurs for reg2mem'd allocas.
336 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
337 LI->replaceAllUsesWith(AvailableVal);
338 LI->eraseFromParent();
342 // Otherwise, if we scanned the whole block and got to the top of the block,
343 // we know the block is locally transparent to the load. If not, something
344 // might clobber its value.
345 if (BBIt != LoadBB->begin())
349 SmallPtrSet<BasicBlock*, 8> PredsScanned;
350 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
351 AvailablePredsTy AvailablePreds;
352 BasicBlock *OneUnavailablePred = 0;
354 // If we got here, the loaded value is transparent through to the start of the
355 // block. Check to see if it is available in any of the predecessor blocks.
356 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
358 BasicBlock *PredBB = *PI;
360 // If we already scanned this predecessor, skip it.
361 if (!PredsScanned.insert(PredBB))
364 // Scan the predecessor to see if the value is available in the pred.
365 BBIt = PredBB->end();
366 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt);
367 if (!PredAvailable) {
368 OneUnavailablePred = PredBB;
372 // If so, this load is partially redundant. Remember this info so that we
373 // can create a PHI node.
374 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
377 // If the loaded value isn't available in any predecessor, it isn't partially
379 if (AvailablePreds.empty()) return false;
381 // Okay, the loaded value is available in at least one (and maybe all!)
382 // predecessors. If the value is unavailable in more than one unique
383 // predecessor, we want to insert a merge block for those common predecessors.
384 // This ensures that we only have to insert one reload, thus not increasing
386 BasicBlock *UnavailablePred = 0;
388 // If there is exactly one predecessor where the value is unavailable, the
389 // already computed 'OneUnavailablePred' block is it. If it ends in an
390 // unconditional branch, we know that it isn't a critical edge.
391 if (PredsScanned.size() == AvailablePreds.size()+1 &&
392 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
393 UnavailablePred = OneUnavailablePred;
394 } else if (PredsScanned.size() != AvailablePreds.size()) {
395 // Otherwise, we had multiple unavailable predecessors or we had a critical
396 // edge from the one.
397 SmallVector<BasicBlock*, 8> PredsToSplit;
398 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
400 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
401 AvailablePredSet.insert(AvailablePreds[i].first);
403 // Add all the unavailable predecessors to the PredsToSplit list.
404 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
406 if (!AvailablePredSet.count(*PI))
407 PredsToSplit.push_back(*PI);
409 // Split them out to their own block.
411 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
412 "thread-split", this);
415 // If the value isn't available in all predecessors, then there will be
416 // exactly one where it isn't available. Insert a load on that edge and add
417 // it to the AvailablePreds list.
418 if (UnavailablePred) {
419 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
420 "Can't handle critical edge here!");
421 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr",
422 UnavailablePred->getTerminator());
423 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
426 // Now we know that each predecessor of this block has a value in
427 // AvailablePreds, sort them for efficient access as we're walking the preds.
428 std::sort(AvailablePreds.begin(), AvailablePreds.end());
430 // Create a PHI node at the start of the block for the PRE'd load value.
431 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
434 // Insert new entries into the PHI for each predecessor. A single block may
435 // have multiple entries here.
436 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
438 AvailablePredsTy::iterator I =
439 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
440 std::make_pair(*PI, (Value*)0));
442 assert(I != AvailablePreds.end() && I->first == *PI &&
443 "Didn't find entry for predecessor!");
445 PN->addIncoming(I->second, I->first);
448 //cerr << "PRE: " << *LI << *PN << "\n";
450 LI->replaceAllUsesWith(PN);
451 LI->eraseFromParent();
457 /// ProcessJumpOnPHI - We have a conditional branch of switch on a PHI node in
458 /// the current block. See if there are any simplifications we can do based on
459 /// inputs to the phi node.
461 bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) {
462 // See if the phi node has any constant values. If so, we can determine where
463 // the corresponding predecessor will branch.
464 ConstantInt *PredCst = 0;
465 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
466 if ((PredCst = dyn_cast<ConstantInt>(PN->getIncomingValue(i))))
469 // If no incoming value has a constant, we don't know the destination of any
474 // See if the cost of duplicating this block is low enough.
475 BasicBlock *BB = PN->getParent();
476 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
477 if (JumpThreadCost > Threshold) {
478 DOUT << " Not threading BB '" << BB->getNameStart()
479 << "' - Cost is too high: " << JumpThreadCost << "\n";
483 // If so, we can actually do this threading. Merge any common predecessors
484 // that will act the same.
485 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst);
487 // Next, figure out which successor we are threading to.
489 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
490 SuccBB = BI->getSuccessor(PredCst == ConstantInt::getFalse());
492 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
493 SuccBB = SI->getSuccessor(SI->findCaseValue(PredCst));
496 // If threading to the same block as we come from, we would infinite loop.
498 DOUT << " Not threading BB '" << BB->getNameStart()
499 << "' - would thread to self!\n";
503 // And finally, do it!
504 DOUT << " Threading edge from '" << PredBB->getNameStart() << "' to '"
505 << SuccBB->getNameStart() << "' with cost: " << JumpThreadCost
506 << ", across block:\n "
509 ThreadEdge(BB, PredBB, SuccBB);
514 /// ProcessJumpOnLogicalPHI - PN's basic block contains a conditional branch
515 /// whose condition is an AND/OR where one side is PN. If PN has constant
516 /// operands that permit us to evaluate the condition for some operand, thread
517 /// through the block. For example with:
518 /// br (and X, phi(Y, Z, false))
519 /// the predecessor corresponding to the 'false' will always jump to the false
520 /// destination of the branch.
522 bool JumpThreading::ProcessBranchOnLogical(Value *V, BasicBlock *BB,
524 // If this is a binary operator tree of the same AND/OR opcode, check the
526 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(V))
527 if ((isAnd && BO->getOpcode() == Instruction::And) ||
528 (!isAnd && BO->getOpcode() == Instruction::Or)) {
529 if (ProcessBranchOnLogical(BO->getOperand(0), BB, isAnd))
531 if (ProcessBranchOnLogical(BO->getOperand(1), BB, isAnd))
535 // If this isn't a PHI node, we can't handle it.
536 PHINode *PN = dyn_cast<PHINode>(V);
537 if (!PN || PN->getParent() != BB) return false;
539 // We can only do the simplification for phi nodes of 'false' with AND or
540 // 'true' with OR. See if we have any entries in the phi for this.
541 unsigned PredNo = ~0U;
542 ConstantInt *PredCst = ConstantInt::get(Type::Int1Ty, !isAnd);
543 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
544 if (PN->getIncomingValue(i) == PredCst) {
550 // If no match, bail out.
554 // See if the cost of duplicating this block is low enough.
555 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
556 if (JumpThreadCost > Threshold) {
557 DOUT << " Not threading BB '" << BB->getNameStart()
558 << "' - Cost is too high: " << JumpThreadCost << "\n";
562 // If so, we can actually do this threading. Merge any common predecessors
563 // that will act the same.
564 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst);
566 // Next, figure out which successor we are threading to. If this was an AND,
567 // the constant must be FALSE, and we must be targeting the 'false' block.
568 // If this is an OR, the constant must be TRUE, and we must be targeting the
570 BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(isAnd);
572 // If threading to the same block as we come from, we would infinite loop.
574 DOUT << " Not threading BB '" << BB->getNameStart()
575 << "' - would thread to self!\n";
579 // And finally, do it!
580 DOUT << " Threading edge through bool from '" << PredBB->getNameStart()
581 << "' to '" << SuccBB->getNameStart() << "' with cost: "
582 << JumpThreadCost << ", across block:\n "
585 ThreadEdge(BB, PredBB, SuccBB);
590 /// ProcessBranchOnCompare - We found a branch on a comparison between a phi
591 /// node and a constant. If the PHI node contains any constants as inputs, we
592 /// can fold the compare for that edge and thread through it.
593 bool JumpThreading::ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB) {
594 PHINode *PN = cast<PHINode>(Cmp->getOperand(0));
595 Constant *RHS = cast<Constant>(Cmp->getOperand(1));
597 // If the phi isn't in the current block, an incoming edge to this block
598 // doesn't control the destination.
599 if (PN->getParent() != BB)
602 // We can do this simplification if any comparisons fold to true or false.
604 Constant *PredCst = 0;
605 bool TrueDirection = false;
606 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
607 PredCst = dyn_cast<Constant>(PN->getIncomingValue(i));
608 if (PredCst == 0) continue;
611 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Cmp))
612 Res = ConstantExpr::getICmp(ICI->getPredicate(), PredCst, RHS);
614 Res = ConstantExpr::getFCmp(cast<FCmpInst>(Cmp)->getPredicate(),
616 // If this folded to a constant expr, we can't do anything.
617 if (ConstantInt *ResC = dyn_cast<ConstantInt>(Res)) {
618 TrueDirection = ResC->getZExtValue();
621 // If this folded to undef, just go the false way.
622 if (isa<UndefValue>(Res)) {
623 TrueDirection = false;
627 // Otherwise, we can't fold this input.
631 // If no match, bail out.
635 // See if the cost of duplicating this block is low enough.
636 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
637 if (JumpThreadCost > Threshold) {
638 DOUT << " Not threading BB '" << BB->getNameStart()
639 << "' - Cost is too high: " << JumpThreadCost << "\n";
643 // If so, we can actually do this threading. Merge any common predecessors
644 // that will act the same.
645 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst);
647 // Next, get our successor.
648 BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(!TrueDirection);
650 // If threading to the same block as we come from, we would infinite loop.
652 DOUT << " Not threading BB '" << BB->getNameStart()
653 << "' - would thread to self!\n";
658 // And finally, do it!
659 DOUT << " Threading edge through bool from '" << PredBB->getNameStart()
660 << "' to '" << SuccBB->getNameStart() << "' with cost: "
661 << JumpThreadCost << ", across block:\n "
664 ThreadEdge(BB, PredBB, SuccBB);
670 /// ThreadEdge - We have decided that it is safe and profitable to thread an
671 /// edge from PredBB to SuccBB across BB. Transform the IR to reflect this
673 void JumpThreading::ThreadEdge(BasicBlock *BB, BasicBlock *PredBB,
674 BasicBlock *SuccBB) {
676 // Jump Threading can not update SSA properties correctly if the values
677 // defined in the duplicated block are used outside of the block itself. For
678 // this reason, we spill all values that are used outside of BB to the stack.
679 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
680 if (!I->isUsedOutsideOfBlock(BB))
683 // We found a use of I outside of BB. Create a new stack slot to
684 // break this inter-block usage pattern.
685 DemoteRegToStack(*I);
688 // We are going to have to map operands from the original BB block to the new
689 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
690 // account for entry from PredBB.
691 DenseMap<Instruction*, Value*> ValueMapping;
694 BasicBlock::Create(BB->getName()+".thread", BB->getParent(), BB);
695 NewBB->moveAfter(PredBB);
697 BasicBlock::iterator BI = BB->begin();
698 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
699 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
701 // Clone the non-phi instructions of BB into NewBB, keeping track of the
702 // mapping and using it to remap operands in the cloned instructions.
703 for (; !isa<TerminatorInst>(BI); ++BI) {
704 Instruction *New = BI->clone();
705 New->setName(BI->getNameStart());
706 NewBB->getInstList().push_back(New);
707 ValueMapping[BI] = New;
709 // Remap operands to patch up intra-block references.
710 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
711 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i)))
712 if (Value *Remapped = ValueMapping[Inst])
713 New->setOperand(i, Remapped);
716 // We didn't copy the terminator from BB over to NewBB, because there is now
717 // an unconditional jump to SuccBB. Insert the unconditional jump.
718 BranchInst::Create(SuccBB, NewBB);
720 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
721 // PHI nodes for NewBB now.
722 for (BasicBlock::iterator PNI = SuccBB->begin(); isa<PHINode>(PNI); ++PNI) {
723 PHINode *PN = cast<PHINode>(PNI);
724 // Ok, we have a PHI node. Figure out what the incoming value was for the
726 Value *IV = PN->getIncomingValueForBlock(BB);
728 // Remap the value if necessary.
729 if (Instruction *Inst = dyn_cast<Instruction>(IV))
730 if (Value *MappedIV = ValueMapping[Inst])
732 PN->addIncoming(IV, NewBB);
735 // Finally, NewBB is good to go. Update the terminator of PredBB to jump to
736 // NewBB instead of BB. This eliminates predecessors from BB, which requires
737 // us to simplify any PHI nodes in BB.
738 TerminatorInst *PredTerm = PredBB->getTerminator();
739 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
740 if (PredTerm->getSuccessor(i) == BB) {
741 BB->removePredecessor(PredBB);
742 PredTerm->setSuccessor(i, NewBB);