1 //===- LoopSimplify.cpp - Loop Canonicalization Pass ----------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass performs several transformations to transform natural loops into a
11 // simpler form, which makes subsequent analyses and transformations simpler and
14 // Loop pre-header insertion guarantees that there is a single, non-critical
15 // entry edge from outside of the loop to the loop header. This simplifies a
16 // number of analyses and transformations, such as LICM.
18 // Loop exit-block insertion guarantees that all exit blocks from the loop
19 // (blocks which are outside of the loop that have predecessors inside of the
20 // loop) only have predecessors from inside of the loop (and are thus dominated
21 // by the loop header). This simplifies transformations such as store-sinking
22 // that are built into LICM.
24 // This pass also guarantees that loops will have exactly one backedge.
26 // Note that the simplifycfg pass will clean up blocks which are split out but
27 // end up being unnecessary, so usage of this pass should not pessimize
30 // This pass obviously modifies the CFG, but updates loop information and
31 // dominator information.
33 //===----------------------------------------------------------------------===//
35 #include "llvm/Transforms/Scalar.h"
36 #include "llvm/Constant.h"
37 #include "llvm/Instructions.h"
38 #include "llvm/Function.h"
39 #include "llvm/Type.h"
40 #include "llvm/Analysis/AliasAnalysis.h"
41 #include "llvm/Analysis/Dominators.h"
42 #include "llvm/Analysis/LoopInfo.h"
43 #include "llvm/Support/CFG.h"
44 #include "llvm/Support/Compiler.h"
45 #include "llvm/ADT/SetOperations.h"
46 #include "llvm/ADT/SetVector.h"
47 #include "llvm/ADT/Statistic.h"
48 #include "llvm/ADT/DepthFirstIterator.h"
53 NumInserted("loopsimplify", "Number of pre-header or exit blocks inserted");
55 NumNested("loopsimplify", "Number of nested loops split out");
57 struct VISIBILITY_HIDDEN LoopSimplify : public FunctionPass {
58 // AA - If we have an alias analysis object to update, this is it, otherwise
63 virtual bool runOnFunction(Function &F);
65 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
66 // We need loop information to identify the loops...
67 AU.addRequired<LoopInfo>();
68 AU.addRequired<DominatorSet>();
69 AU.addRequired<DominatorTree>();
71 AU.addPreserved<LoopInfo>();
72 AU.addPreserved<DominatorSet>();
73 AU.addPreserved<ImmediateDominators>();
74 AU.addPreserved<ETForest>();
75 AU.addPreserved<DominatorTree>();
76 AU.addPreserved<DominanceFrontier>();
77 AU.addPreservedID(BreakCriticalEdgesID); // No critical edges added.
80 bool ProcessLoop(Loop *L);
81 BasicBlock *SplitBlockPredecessors(BasicBlock *BB, const char *Suffix,
82 const std::vector<BasicBlock*> &Preds);
83 BasicBlock *RewriteLoopExitBlock(Loop *L, BasicBlock *Exit);
84 void InsertPreheaderForLoop(Loop *L);
85 Loop *SeparateNestedLoop(Loop *L);
86 void InsertUniqueBackedgeBlock(Loop *L);
88 void UpdateDomInfoForRevectoredPreds(BasicBlock *NewBB,
89 std::vector<BasicBlock*> &PredBlocks);
92 RegisterPass<LoopSimplify>
93 X("loopsimplify", "Canonicalize natural loops", true);
96 // Publically exposed interface to pass...
97 const PassInfo *llvm::LoopSimplifyID = X.getPassInfo();
98 FunctionPass *llvm::createLoopSimplifyPass() { return new LoopSimplify(); }
100 /// runOnFunction - Run down all loops in the CFG (recursively, but we could do
101 /// it in any convenient order) inserting preheaders...
103 bool LoopSimplify::runOnFunction(Function &F) {
104 bool Changed = false;
105 LI = &getAnalysis<LoopInfo>();
106 AA = getAnalysisToUpdate<AliasAnalysis>();
108 // Check to see that no blocks (other than the header) in loops have
109 // predecessors that are not in loops. This is not valid for natural loops,
110 // but can occur if the blocks are unreachable. Since they are unreachable we
111 // can just shamelessly destroy their terminators to make them not branch into
113 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
114 // This case can only occur for unreachable blocks. Blocks that are
115 // unreachable can't be in loops, so filter those blocks out.
116 if (LI->getLoopFor(BB)) continue;
118 bool BlockUnreachable = false;
119 TerminatorInst *TI = BB->getTerminator();
121 // Check to see if any successors of this block are non-loop-header loops
122 // that are not the header.
123 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
124 // If this successor is not in a loop, BB is clearly ok.
125 Loop *L = LI->getLoopFor(TI->getSuccessor(i));
128 // If the succ is the loop header, and if L is a top-level loop, then this
129 // is an entrance into a loop through the header, which is also ok.
130 if (L->getHeader() == TI->getSuccessor(i) && L->getParentLoop() == 0)
133 // Otherwise, this is an entrance into a loop from some place invalid.
134 // Either the loop structure is invalid and this is not a natural loop (in
135 // which case the compiler is buggy somewhere else) or BB is unreachable.
136 BlockUnreachable = true;
140 // If this block is ok, check the next one.
141 if (!BlockUnreachable) continue;
143 // Otherwise, this block is dead. To clean up the CFG and to allow later
144 // loop transformations to ignore this case, we delete the edges into the
145 // loop by replacing the terminator.
147 // Remove PHI entries from the successors.
148 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
149 TI->getSuccessor(i)->removePredecessor(BB);
151 // Add a new unreachable instruction.
152 new UnreachableInst(TI);
154 // Delete the dead terminator.
155 if (AA) AA->deleteValue(&BB->back());
156 BB->getInstList().pop_back();
160 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
161 Changed |= ProcessLoop(*I);
166 /// ProcessLoop - Walk the loop structure in depth first order, ensuring that
167 /// all loops have preheaders.
169 bool LoopSimplify::ProcessLoop(Loop *L) {
170 bool Changed = false;
173 // Canonicalize inner loops before outer loops. Inner loop canonicalization
174 // can provide work for the outer loop to canonicalize.
175 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
176 Changed |= ProcessLoop(*I);
178 assert(L->getBlocks()[0] == L->getHeader() &&
179 "Header isn't first block in loop?");
181 // Does the loop already have a preheader? If so, don't insert one.
182 if (L->getLoopPreheader() == 0) {
183 InsertPreheaderForLoop(L);
188 // Next, check to make sure that all exit nodes of the loop only have
189 // predecessors that are inside of the loop. This check guarantees that the
190 // loop preheader/header will dominate the exit blocks. If the exit block has
191 // predecessors from outside of the loop, split the edge now.
192 std::vector<BasicBlock*> ExitBlocks;
193 L->getExitBlocks(ExitBlocks);
195 SetVector<BasicBlock*> ExitBlockSet(ExitBlocks.begin(), ExitBlocks.end());
196 for (SetVector<BasicBlock*>::iterator I = ExitBlockSet.begin(),
197 E = ExitBlockSet.end(); I != E; ++I) {
198 BasicBlock *ExitBlock = *I;
199 for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock);
201 // Must be exactly this loop: no subloops, parent loops, or non-loop preds
203 if (!L->contains(*PI)) {
204 RewriteLoopExitBlock(L, ExitBlock);
211 // If the header has more than two predecessors at this point (from the
212 // preheader and from multiple backedges), we must adjust the loop.
213 unsigned NumBackedges = L->getNumBackEdges();
214 if (NumBackedges != 1) {
215 // If this is really a nested loop, rip it out into a child loop. Don't do
216 // this for loops with a giant number of backedges, just factor them into a
217 // common backedge instead.
218 if (NumBackedges < 8) {
219 if (Loop *NL = SeparateNestedLoop(L)) {
221 // This is a big restructuring change, reprocess the whole loop.
224 // GCC doesn't tail recursion eliminate this.
229 // If we either couldn't, or didn't want to, identify nesting of the loops,
230 // insert a new block that all backedges target, then make it jump to the
232 InsertUniqueBackedgeBlock(L);
237 // Scan over the PHI nodes in the loop header. Since they now have only two
238 // incoming values (the loop is canonicalized), we may have simplified the PHI
239 // down to 'X = phi [X, Y]', which should be replaced with 'Y'.
241 for (BasicBlock::iterator I = L->getHeader()->begin();
242 (PN = dyn_cast<PHINode>(I++)); )
243 if (Value *V = PN->hasConstantValue()) {
244 PN->replaceAllUsesWith(V);
245 PN->eraseFromParent();
251 /// SplitBlockPredecessors - Split the specified block into two blocks. We want
252 /// to move the predecessors specified in the Preds list to point to the new
253 /// block, leaving the remaining predecessors pointing to BB. This method
254 /// updates the SSA PHINode's, but no other analyses.
256 BasicBlock *LoopSimplify::SplitBlockPredecessors(BasicBlock *BB,
258 const std::vector<BasicBlock*> &Preds) {
260 // Create new basic block, insert right before the original block...
261 BasicBlock *NewBB = new BasicBlock(BB->getName()+Suffix, BB->getParent(), BB);
263 // The preheader first gets an unconditional branch to the loop header...
264 BranchInst *BI = new BranchInst(BB, NewBB);
266 // For every PHI node in the block, insert a PHI node into NewBB where the
267 // incoming values from the out of loop edges are moved to NewBB. We have two
268 // possible cases here. If the loop is dead, we just insert dummy entries
269 // into the PHI nodes for the new edge. If the loop is not dead, we move the
270 // incoming edges in BB into new PHI nodes in NewBB.
272 if (!Preds.empty()) { // Is the loop not obviously dead?
273 // Check to see if the values being merged into the new block need PHI
274 // nodes. If so, insert them.
275 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ) {
276 PHINode *PN = cast<PHINode>(I);
279 // Check to see if all of the values coming in are the same. If so, we
280 // don't need to create a new PHI node.
281 Value *InVal = PN->getIncomingValueForBlock(Preds[0]);
282 for (unsigned i = 1, e = Preds.size(); i != e; ++i)
283 if (InVal != PN->getIncomingValueForBlock(Preds[i])) {
288 // If the values coming into the block are not the same, we need a PHI.
290 // Create the new PHI node, insert it into NewBB at the end of the block
291 PHINode *NewPHI = new PHINode(PN->getType(), PN->getName()+".ph", BI);
292 if (AA) AA->copyValue(PN, NewPHI);
294 // Move all of the edges from blocks outside the loop to the new PHI
295 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
296 Value *V = PN->removeIncomingValue(Preds[i], false);
297 NewPHI->addIncoming(V, Preds[i]);
301 // Remove all of the edges coming into the PHI nodes from outside of the
303 for (unsigned i = 0, e = Preds.size(); i != e; ++i)
304 PN->removeIncomingValue(Preds[i], false);
307 // Add an incoming value to the PHI node in the loop for the preheader
309 PN->addIncoming(InVal, NewBB);
311 // Can we eliminate this phi node now?
312 if (Value *V = PN->hasConstantValue(true)) {
313 if (!isa<Instruction>(V) ||
314 getAnalysis<DominatorSet>().dominates(cast<Instruction>(V), PN)) {
315 PN->replaceAllUsesWith(V);
316 if (AA) AA->deleteValue(PN);
317 BB->getInstList().erase(PN);
322 // Now that the PHI nodes are updated, actually move the edges from
323 // Preds to point to NewBB instead of BB.
325 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
326 TerminatorInst *TI = Preds[i]->getTerminator();
327 for (unsigned s = 0, e = TI->getNumSuccessors(); s != e; ++s)
328 if (TI->getSuccessor(s) == BB)
329 TI->setSuccessor(s, NewBB);
332 } else { // Otherwise the loop is dead...
333 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++I) {
334 PHINode *PN = cast<PHINode>(I);
335 // Insert dummy values as the incoming value...
336 PN->addIncoming(Constant::getNullValue(PN->getType()), NewBB);
342 /// InsertPreheaderForLoop - Once we discover that a loop doesn't have a
343 /// preheader, this method is called to insert one. This method has two phases:
344 /// preheader insertion and analysis updating.
346 void LoopSimplify::InsertPreheaderForLoop(Loop *L) {
347 BasicBlock *Header = L->getHeader();
349 // Compute the set of predecessors of the loop that are not in the loop.
350 std::vector<BasicBlock*> OutsideBlocks;
351 for (pred_iterator PI = pred_begin(Header), PE = pred_end(Header);
353 if (!L->contains(*PI)) // Coming in from outside the loop?
354 OutsideBlocks.push_back(*PI); // Keep track of it...
356 // Split out the loop pre-header
358 SplitBlockPredecessors(Header, ".preheader", OutsideBlocks);
360 //===--------------------------------------------------------------------===//
361 // Update analysis results now that we have performed the transformation
364 // We know that we have loop information to update... update it now.
365 if (Loop *Parent = L->getParentLoop())
366 Parent->addBasicBlockToLoop(NewBB, *LI);
368 DominatorSet &DS = getAnalysis<DominatorSet>(); // Update dominator info
369 DominatorTree &DT = getAnalysis<DominatorTree>();
372 // Update the dominator tree information.
373 // The immediate dominator of the preheader is the immediate dominator of
375 DominatorTree::Node *PHDomTreeNode =
376 DT.createNewNode(NewBB, DT.getNode(Header)->getIDom());
377 BasicBlock *oldHeaderIDom = DT.getNode(Header)->getIDom()->getBlock();
379 // Change the header node so that PNHode is the new immediate dominator
380 DT.changeImmediateDominator(DT.getNode(Header), PHDomTreeNode);
383 // The blocks that dominate NewBB are the blocks that dominate Header,
384 // minus Header, plus NewBB.
385 DominatorSet::DomSetType DomSet = DS.getDominators(Header);
386 DomSet.erase(Header); // Header does not dominate us...
387 DS.addBasicBlock(NewBB, DomSet);
389 // The newly created basic block dominates all nodes dominated by Header.
390 for (df_iterator<DominatorTree::Node*> DFI = df_begin(PHDomTreeNode),
391 E = df_end(PHDomTreeNode); DFI != E; ++DFI)
392 DS.addDominator((*DFI)->getBlock(), NewBB);
395 // Update immediate dominator information if we have it...
396 if (ImmediateDominators *ID = getAnalysisToUpdate<ImmediateDominators>()) {
397 // Whatever i-dominated the header node now immediately dominates NewBB
398 ID->addNewBlock(NewBB, ID->get(Header));
400 // The preheader now is the immediate dominator for the header node...
401 ID->setImmediateDominator(Header, NewBB);
404 // Update ET Forest information if we have it...
405 if (ETForest *EF = getAnalysisToUpdate<ETForest>()) {
406 // Whatever i-dominated the header node now immediately dominates NewBB
407 EF->addNewBlock(NewBB, oldHeaderIDom);
409 // The preheader now is the immediate dominator for the header node...
410 EF->setImmediateDominator(Header, NewBB);
413 // Update dominance frontier information...
414 if (DominanceFrontier *DF = getAnalysisToUpdate<DominanceFrontier>()) {
415 // The DF(NewBB) is just (DF(Header)-Header), because NewBB dominates
416 // everything that Header does, and it strictly dominates Header in
418 assert(DF->find(Header) != DF->end() && "Header node doesn't have DF set?");
419 DominanceFrontier::DomSetType NewDFSet = DF->find(Header)->second;
420 NewDFSet.erase(Header);
421 DF->addBasicBlock(NewBB, NewDFSet);
423 // Now we must loop over all of the dominance frontiers in the function,
424 // replacing occurrences of Header with NewBB in some cases. If a block
425 // dominates a (now) predecessor of NewBB, but did not strictly dominate
426 // Header, it will have Header in it's DF set, but should now have NewBB in
428 for (unsigned i = 0, e = OutsideBlocks.size(); i != e; ++i) {
429 // Get all of the dominators of the predecessor...
430 const DominatorSet::DomSetType &PredDoms =
431 DS.getDominators(OutsideBlocks[i]);
432 for (DominatorSet::DomSetType::const_iterator PDI = PredDoms.begin(),
433 PDE = PredDoms.end(); PDI != PDE; ++PDI) {
434 BasicBlock *PredDom = *PDI;
435 // If the loop header is in DF(PredDom), then PredDom didn't dominate
436 // the header but did dominate a predecessor outside of the loop. Now
437 // we change this entry to include the preheader in the DF instead of
439 DominanceFrontier::iterator DFI = DF->find(PredDom);
440 assert(DFI != DF->end() && "No dominance frontier for node?");
441 if (DFI->second.count(Header)) {
442 DF->removeFromFrontier(DFI, Header);
443 DF->addToFrontier(DFI, NewBB);
450 /// RewriteLoopExitBlock - Ensure that the loop preheader dominates all exit
451 /// blocks. This method is used to split exit blocks that have predecessors
452 /// outside of the loop.
453 BasicBlock *LoopSimplify::RewriteLoopExitBlock(Loop *L, BasicBlock *Exit) {
454 std::vector<BasicBlock*> LoopBlocks;
455 for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit); I != E; ++I)
457 LoopBlocks.push_back(*I);
459 assert(!LoopBlocks.empty() && "No edges coming in from outside the loop?");
460 BasicBlock *NewBB = SplitBlockPredecessors(Exit, ".loopexit", LoopBlocks);
462 // Update Loop Information - we know that the new block will be in whichever
463 // loop the Exit block is in. Note that it may not be in that immediate loop,
464 // if the successor is some other loop header. In that case, we continue
465 // walking up the loop tree to find a loop that contains both the successor
466 // block and the predecessor block.
467 Loop *SuccLoop = LI->getLoopFor(Exit);
468 while (SuccLoop && !SuccLoop->contains(L->getHeader()))
469 SuccLoop = SuccLoop->getParentLoop();
471 SuccLoop->addBasicBlockToLoop(NewBB, *LI);
473 // Update dominator information (set, immdom, domtree, and domfrontier)
474 UpdateDomInfoForRevectoredPreds(NewBB, LoopBlocks);
478 /// AddBlockAndPredsToSet - Add the specified block, and all of its
479 /// predecessors, to the specified set, if it's not already in there. Stop
480 /// predecessor traversal when we reach StopBlock.
481 static void AddBlockAndPredsToSet(BasicBlock *BB, BasicBlock *StopBlock,
482 std::set<BasicBlock*> &Blocks) {
483 if (!Blocks.insert(BB).second) return; // already processed.
484 if (BB == StopBlock) return; // Stop here!
486 for (pred_iterator I = pred_begin(BB), E = pred_end(BB); I != E; ++I)
487 AddBlockAndPredsToSet(*I, StopBlock, Blocks);
490 /// FindPHIToPartitionLoops - The first part of loop-nestification is to find a
491 /// PHI node that tells us how to partition the loops.
492 static PHINode *FindPHIToPartitionLoops(Loop *L, DominatorSet &DS,
494 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ) {
495 PHINode *PN = cast<PHINode>(I);
497 if (Value *V = PN->hasConstantValue())
498 if (!isa<Instruction>(V) || DS.dominates(cast<Instruction>(V), PN)) {
499 // This is a degenerate PHI already, don't modify it!
500 PN->replaceAllUsesWith(V);
501 if (AA) AA->deleteValue(PN);
502 PN->eraseFromParent();
506 // Scan this PHI node looking for a use of the PHI node by itself.
507 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
508 if (PN->getIncomingValue(i) == PN &&
509 L->contains(PN->getIncomingBlock(i)))
510 // We found something tasty to remove.
516 /// SeparateNestedLoop - If this loop has multiple backedges, try to pull one of
517 /// them out into a nested loop. This is important for code that looks like
522 /// br cond, Loop, Next
524 /// br cond2, Loop, Out
526 /// To identify this common case, we look at the PHI nodes in the header of the
527 /// loop. PHI nodes with unchanging values on one backedge correspond to values
528 /// that change in the "outer" loop, but not in the "inner" loop.
530 /// If we are able to separate out a loop, return the new outer loop that was
533 Loop *LoopSimplify::SeparateNestedLoop(Loop *L) {
534 PHINode *PN = FindPHIToPartitionLoops(L, getAnalysis<DominatorSet>(), AA);
535 if (PN == 0) return 0; // No known way to partition.
537 // Pull out all predecessors that have varying values in the loop. This
538 // handles the case when a PHI node has multiple instances of itself as
540 std::vector<BasicBlock*> OuterLoopPreds;
541 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
542 if (PN->getIncomingValue(i) != PN ||
543 !L->contains(PN->getIncomingBlock(i)))
544 OuterLoopPreds.push_back(PN->getIncomingBlock(i));
546 BasicBlock *Header = L->getHeader();
547 BasicBlock *NewBB = SplitBlockPredecessors(Header, ".outer", OuterLoopPreds);
549 // Update dominator information (set, immdom, domtree, and domfrontier)
550 UpdateDomInfoForRevectoredPreds(NewBB, OuterLoopPreds);
552 // Create the new outer loop.
553 Loop *NewOuter = new Loop();
555 // Change the parent loop to use the outer loop as its child now.
556 if (Loop *Parent = L->getParentLoop())
557 Parent->replaceChildLoopWith(L, NewOuter);
559 LI->changeTopLevelLoop(L, NewOuter);
561 // This block is going to be our new header block: add it to this loop and all
563 NewOuter->addBasicBlockToLoop(NewBB, *LI);
565 // L is now a subloop of our outer loop.
566 NewOuter->addChildLoop(L);
568 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i)
569 NewOuter->addBlockEntry(L->getBlocks()[i]);
571 // Determine which blocks should stay in L and which should be moved out to
572 // the Outer loop now.
573 DominatorSet &DS = getAnalysis<DominatorSet>();
574 std::set<BasicBlock*> BlocksInL;
575 for (pred_iterator PI = pred_begin(Header), E = pred_end(Header); PI!=E; ++PI)
576 if (DS.dominates(Header, *PI))
577 AddBlockAndPredsToSet(*PI, Header, BlocksInL);
580 // Scan all of the loop children of L, moving them to OuterLoop if they are
581 // not part of the inner loop.
582 for (Loop::iterator I = L->begin(); I != L->end(); )
583 if (BlocksInL.count((*I)->getHeader()))
584 ++I; // Loop remains in L
586 NewOuter->addChildLoop(L->removeChildLoop(I));
588 // Now that we know which blocks are in L and which need to be moved to
589 // OuterLoop, move any blocks that need it.
590 for (unsigned i = 0; i != L->getBlocks().size(); ++i) {
591 BasicBlock *BB = L->getBlocks()[i];
592 if (!BlocksInL.count(BB)) {
593 // Move this block to the parent, updating the exit blocks sets
594 L->removeBlockFromLoop(BB);
596 LI->changeLoopFor(BB, NewOuter);
606 /// InsertUniqueBackedgeBlock - This method is called when the specified loop
607 /// has more than one backedge in it. If this occurs, revector all of these
608 /// backedges to target a new basic block and have that block branch to the loop
609 /// header. This ensures that loops have exactly one backedge.
611 void LoopSimplify::InsertUniqueBackedgeBlock(Loop *L) {
612 assert(L->getNumBackEdges() > 1 && "Must have > 1 backedge!");
614 // Get information about the loop
615 BasicBlock *Preheader = L->getLoopPreheader();
616 BasicBlock *Header = L->getHeader();
617 Function *F = Header->getParent();
619 // Figure out which basic blocks contain back-edges to the loop header.
620 std::vector<BasicBlock*> BackedgeBlocks;
621 for (pred_iterator I = pred_begin(Header), E = pred_end(Header); I != E; ++I)
622 if (*I != Preheader) BackedgeBlocks.push_back(*I);
624 // Create and insert the new backedge block...
625 BasicBlock *BEBlock = new BasicBlock(Header->getName()+".backedge", F);
626 BranchInst *BETerminator = new BranchInst(Header, BEBlock);
628 // Move the new backedge block to right after the last backedge block.
629 Function::iterator InsertPos = BackedgeBlocks.back(); ++InsertPos;
630 F->getBasicBlockList().splice(InsertPos, F->getBasicBlockList(), BEBlock);
632 // Now that the block has been inserted into the function, create PHI nodes in
633 // the backedge block which correspond to any PHI nodes in the header block.
634 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
635 PHINode *PN = cast<PHINode>(I);
636 PHINode *NewPN = new PHINode(PN->getType(), PN->getName()+".be",
638 NewPN->reserveOperandSpace(BackedgeBlocks.size());
639 if (AA) AA->copyValue(PN, NewPN);
641 // Loop over the PHI node, moving all entries except the one for the
642 // preheader over to the new PHI node.
643 unsigned PreheaderIdx = ~0U;
644 bool HasUniqueIncomingValue = true;
645 Value *UniqueValue = 0;
646 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
647 BasicBlock *IBB = PN->getIncomingBlock(i);
648 Value *IV = PN->getIncomingValue(i);
649 if (IBB == Preheader) {
652 NewPN->addIncoming(IV, IBB);
653 if (HasUniqueIncomingValue) {
654 if (UniqueValue == 0)
656 else if (UniqueValue != IV)
657 HasUniqueIncomingValue = false;
662 // Delete all of the incoming values from the old PN except the preheader's
663 assert(PreheaderIdx != ~0U && "PHI has no preheader entry??");
664 if (PreheaderIdx != 0) {
665 PN->setIncomingValue(0, PN->getIncomingValue(PreheaderIdx));
666 PN->setIncomingBlock(0, PN->getIncomingBlock(PreheaderIdx));
668 // Nuke all entries except the zero'th.
669 for (unsigned i = 0, e = PN->getNumIncomingValues()-1; i != e; ++i)
670 PN->removeIncomingValue(e-i, false);
672 // Finally, add the newly constructed PHI node as the entry for the BEBlock.
673 PN->addIncoming(NewPN, BEBlock);
675 // As an optimization, if all incoming values in the new PhiNode (which is a
676 // subset of the incoming values of the old PHI node) have the same value,
677 // eliminate the PHI Node.
678 if (HasUniqueIncomingValue) {
679 NewPN->replaceAllUsesWith(UniqueValue);
680 if (AA) AA->deleteValue(NewPN);
681 BEBlock->getInstList().erase(NewPN);
685 // Now that all of the PHI nodes have been inserted and adjusted, modify the
686 // backedge blocks to just to the BEBlock instead of the header.
687 for (unsigned i = 0, e = BackedgeBlocks.size(); i != e; ++i) {
688 TerminatorInst *TI = BackedgeBlocks[i]->getTerminator();
689 for (unsigned Op = 0, e = TI->getNumSuccessors(); Op != e; ++Op)
690 if (TI->getSuccessor(Op) == Header)
691 TI->setSuccessor(Op, BEBlock);
694 //===--- Update all analyses which we must preserve now -----------------===//
696 // Update Loop Information - we know that this block is now in the current
697 // loop and all parent loops.
698 L->addBasicBlockToLoop(BEBlock, *LI);
700 // Update dominator information (set, immdom, domtree, and domfrontier)
701 UpdateDomInfoForRevectoredPreds(BEBlock, BackedgeBlocks);
704 /// UpdateDomInfoForRevectoredPreds - This method is used to update the four
705 /// different kinds of dominator information (dominator sets, immediate
706 /// dominators, dominator trees, and dominance frontiers) after a new block has
707 /// been added to the CFG.
709 /// This only supports the case when an existing block (known as "NewBBSucc"),
710 /// had some of its predecessors factored into a new basic block. This
711 /// transformation inserts a new basic block ("NewBB"), with a single
712 /// unconditional branch to NewBBSucc, and moves some predecessors of
713 /// "NewBBSucc" to now branch to NewBB. These predecessors are listed in
714 /// PredBlocks, even though they are the same as
715 /// pred_begin(NewBB)/pred_end(NewBB).
717 void LoopSimplify::UpdateDomInfoForRevectoredPreds(BasicBlock *NewBB,
718 std::vector<BasicBlock*> &PredBlocks) {
719 assert(!PredBlocks.empty() && "No predblocks??");
720 assert(succ_begin(NewBB) != succ_end(NewBB) &&
721 ++succ_begin(NewBB) == succ_end(NewBB) &&
722 "NewBB should have a single successor!");
723 BasicBlock *NewBBSucc = *succ_begin(NewBB);
724 DominatorSet &DS = getAnalysis<DominatorSet>();
726 // Update dominator information... The blocks that dominate NewBB are the
727 // intersection of the dominators of predecessors, plus the block itself.
729 DominatorSet::DomSetType NewBBDomSet = DS.getDominators(PredBlocks[0]);
730 for (unsigned i = 1, e = PredBlocks.size(); i != e; ++i)
731 set_intersect(NewBBDomSet, DS.getDominators(PredBlocks[i]));
732 NewBBDomSet.insert(NewBB); // All blocks dominate themselves...
733 DS.addBasicBlock(NewBB, NewBBDomSet);
735 // The newly inserted basic block will dominate existing basic blocks iff the
736 // PredBlocks dominate all of the non-pred blocks. If all predblocks dominate
737 // the non-pred blocks, then they all must be the same block!
739 bool NewBBDominatesNewBBSucc = true;
741 BasicBlock *OnePred = PredBlocks[0];
742 for (unsigned i = 1, e = PredBlocks.size(); i != e; ++i)
743 if (PredBlocks[i] != OnePred) {
744 NewBBDominatesNewBBSucc = false;
748 if (NewBBDominatesNewBBSucc)
749 for (pred_iterator PI = pred_begin(NewBBSucc), E = pred_end(NewBBSucc);
751 if (*PI != NewBB && !DS.dominates(NewBBSucc, *PI)) {
752 NewBBDominatesNewBBSucc = false;
757 // The other scenario where the new block can dominate its successors are when
758 // all predecessors of NewBBSucc that are not NewBB are dominated by NewBBSucc
760 if (!NewBBDominatesNewBBSucc) {
761 NewBBDominatesNewBBSucc = true;
762 for (pred_iterator PI = pred_begin(NewBBSucc), E = pred_end(NewBBSucc);
764 if (*PI != NewBB && !DS.dominates(NewBBSucc, *PI)) {
765 NewBBDominatesNewBBSucc = false;
770 // If NewBB dominates some blocks, then it will dominate all blocks that
772 if (NewBBDominatesNewBBSucc) {
773 BasicBlock *PredBlock = PredBlocks[0];
774 Function *F = NewBB->getParent();
775 for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
776 if (DS.dominates(NewBBSucc, I))
777 DS.addDominator(I, NewBB);
780 // Update immediate dominator information if we have it...
781 BasicBlock *NewBBIDom = 0;
782 if (ImmediateDominators *ID = getAnalysisToUpdate<ImmediateDominators>()) {
783 // To find the immediate dominator of the new exit node, we trace up the
784 // immediate dominators of a predecessor until we find a basic block that
785 // dominates the exit block.
787 BasicBlock *Dom = PredBlocks[0]; // Some random predecessor...
788 while (!NewBBDomSet.count(Dom)) { // Loop until we find a dominator...
789 assert(Dom != 0 && "No shared dominator found???");
793 // Set the immediate dominator now...
794 ID->addNewBlock(NewBB, Dom);
795 NewBBIDom = Dom; // Reuse this if calculating DominatorTree info...
797 // If NewBB strictly dominates other blocks, we need to update their idom's
798 // now. The only block that need adjustment is the NewBBSucc block, whose
799 // idom should currently be set to PredBlocks[0].
800 if (NewBBDominatesNewBBSucc)
801 ID->setImmediateDominator(NewBBSucc, NewBB);
804 // Update DominatorTree information if it is active.
805 if (DominatorTree *DT = getAnalysisToUpdate<DominatorTree>()) {
806 // If we don't have ImmediateDominator info around, calculate the idom as
808 DominatorTree::Node *NewBBIDomNode;
810 NewBBIDomNode = DT->getNode(NewBBIDom);
812 NewBBIDomNode = DT->getNode(PredBlocks[0]); // Random pred
813 while (!NewBBDomSet.count(NewBBIDomNode->getBlock())) {
814 NewBBIDomNode = NewBBIDomNode->getIDom();
815 assert(NewBBIDomNode && "No shared dominator found??");
817 NewBBIDom = NewBBIDomNode->getBlock();
820 // Create the new dominator tree node... and set the idom of NewBB.
821 DominatorTree::Node *NewBBNode = DT->createNewNode(NewBB, NewBBIDomNode);
823 // If NewBB strictly dominates other blocks, then it is now the immediate
824 // dominator of NewBBSucc. Update the dominator tree as appropriate.
825 if (NewBBDominatesNewBBSucc) {
826 DominatorTree::Node *NewBBSuccNode = DT->getNode(NewBBSucc);
827 DT->changeImmediateDominator(NewBBSuccNode, NewBBNode);
831 // Update ET-Forest information if it is active.
832 if (ETForest *EF = getAnalysisToUpdate<ETForest>()) {
833 EF->addNewBlock(NewBB, NewBBIDom);
834 if (NewBBDominatesNewBBSucc)
835 EF->setImmediateDominator(NewBBSucc, NewBB);
838 // Update dominance frontier information...
839 if (DominanceFrontier *DF = getAnalysisToUpdate<DominanceFrontier>()) {
840 // If NewBB dominates NewBBSucc, then DF(NewBB) is now going to be the
841 // DF(PredBlocks[0]) without the stuff that the new block does not dominate
843 if (NewBBDominatesNewBBSucc) {
844 DominanceFrontier::iterator DFI = DF->find(PredBlocks[0]);
845 if (DFI != DF->end()) {
846 DominanceFrontier::DomSetType Set = DFI->second;
847 // Filter out stuff in Set that we do not dominate a predecessor of.
848 for (DominanceFrontier::DomSetType::iterator SetI = Set.begin(),
849 E = Set.end(); SetI != E;) {
850 bool DominatesPred = false;
851 for (pred_iterator PI = pred_begin(*SetI), E = pred_end(*SetI);
853 if (DS.dominates(NewBB, *PI))
854 DominatesPred = true;
861 DF->addBasicBlock(NewBB, Set);
865 // DF(NewBB) is {NewBBSucc} because NewBB does not strictly dominate
866 // NewBBSucc, but it does dominate itself (and there is an edge (NewBB ->
867 // NewBBSucc)). NewBBSucc is the single successor of NewBB.
868 DominanceFrontier::DomSetType NewDFSet;
869 NewDFSet.insert(NewBBSucc);
870 DF->addBasicBlock(NewBB, NewDFSet);
873 // Now we must loop over all of the dominance frontiers in the function,
874 // replacing occurrences of NewBBSucc with NewBB in some cases. All
875 // blocks that dominate a block in PredBlocks and contained NewBBSucc in
876 // their dominance frontier must be updated to contain NewBB instead.
878 for (unsigned i = 0, e = PredBlocks.size(); i != e; ++i) {
879 BasicBlock *Pred = PredBlocks[i];
880 // Get all of the dominators of the predecessor...
881 const DominatorSet::DomSetType &PredDoms = DS.getDominators(Pred);
882 for (DominatorSet::DomSetType::const_iterator PDI = PredDoms.begin(),
883 PDE = PredDoms.end(); PDI != PDE; ++PDI) {
884 BasicBlock *PredDom = *PDI;
886 // If the NewBBSucc node is in DF(PredDom), then PredDom didn't
887 // dominate NewBBSucc but did dominate a predecessor of it. Now we
888 // change this entry to include NewBB in the DF instead of NewBBSucc.
889 DominanceFrontier::iterator DFI = DF->find(PredDom);
890 assert(DFI != DF->end() && "No dominance frontier for node?");
891 if (DFI->second.count(NewBBSucc)) {
892 // If NewBBSucc should not stay in our dominator frontier, remove it.
893 // We remove it unless there is a predecessor of NewBBSucc that we
894 // dominate, but we don't strictly dominate NewBBSucc.
895 bool ShouldRemove = true;
896 if (PredDom == NewBBSucc || !DS.dominates(PredDom, NewBBSucc)) {
897 // Okay, we know that PredDom does not strictly dominate NewBBSucc.
898 // Check to see if it dominates any predecessors of NewBBSucc.
899 for (pred_iterator PI = pred_begin(NewBBSucc),
900 E = pred_end(NewBBSucc); PI != E; ++PI)
901 if (DS.dominates(PredDom, *PI)) {
902 ShouldRemove = false;
908 DF->removeFromFrontier(DFI, NewBBSucc);
909 DF->addToFrontier(DFI, NewBB);