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 #define DEBUG_TYPE "loopsimplify"
36 #include "llvm/Transforms/Scalar.h"
37 #include "llvm/Constant.h"
38 #include "llvm/Instructions.h"
39 #include "llvm/Function.h"
40 #include "llvm/Type.h"
41 #include "llvm/Analysis/AliasAnalysis.h"
42 #include "llvm/Analysis/Dominators.h"
43 #include "llvm/Analysis/LoopInfo.h"
44 #include "llvm/Support/CFG.h"
45 #include "llvm/Support/Compiler.h"
46 #include "llvm/ADT/SetOperations.h"
47 #include "llvm/ADT/SetVector.h"
48 #include "llvm/ADT/Statistic.h"
49 #include "llvm/ADT/DepthFirstIterator.h"
52 STATISTIC(NumInserted, "Number of pre-header or exit blocks inserted");
53 STATISTIC(NumNested , "Number of nested loops split out");
56 struct VISIBILITY_HIDDEN LoopSimplify : public FunctionPass {
57 static char ID; // Pass identification, replacement for typeid
58 LoopSimplify() : FunctionPass((intptr_t)&ID) {}
60 // AA - If we have an alias analysis object to update, this is it, otherwise
65 virtual bool runOnFunction(Function &F);
67 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
68 // We need loop information to identify the loops...
69 AU.addRequired<LoopInfo>();
70 AU.addRequired<DominatorTree>();
71 AU.addRequired<ETForest>();
73 AU.addPreserved<LoopInfo>();
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);
87 void PlaceSplitBlockCarefully(BasicBlock *NewBB,
88 std::vector<BasicBlock*> &SplitPreds,
91 void UpdateDomInfoForRevectoredPreds(BasicBlock *NewBB,
92 std::vector<BasicBlock*> &PredBlocks);
95 char LoopSimplify::ID = 0;
96 RegisterPass<LoopSimplify>
97 X("loopsimplify", "Canonicalize natural loops", true);
100 // Publically exposed interface to pass...
101 const PassInfo *llvm::LoopSimplifyID = X.getPassInfo();
102 FunctionPass *llvm::createLoopSimplifyPass() { return new LoopSimplify(); }
104 /// runOnFunction - Run down all loops in the CFG (recursively, but we could do
105 /// it in any convenient order) inserting preheaders...
107 bool LoopSimplify::runOnFunction(Function &F) {
108 bool Changed = false;
109 LI = &getAnalysis<LoopInfo>();
110 AA = getAnalysisToUpdate<AliasAnalysis>();
112 // Check to see that no blocks (other than the header) in loops have
113 // predecessors that are not in loops. This is not valid for natural loops,
114 // but can occur if the blocks are unreachable. Since they are unreachable we
115 // can just shamelessly destroy their terminators to make them not branch into
117 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
118 // This case can only occur for unreachable blocks. Blocks that are
119 // unreachable can't be in loops, so filter those blocks out.
120 if (LI->getLoopFor(BB)) continue;
122 bool BlockUnreachable = false;
123 TerminatorInst *TI = BB->getTerminator();
125 // Check to see if any successors of this block are non-loop-header loops
126 // that are not the header.
127 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
128 // If this successor is not in a loop, BB is clearly ok.
129 Loop *L = LI->getLoopFor(TI->getSuccessor(i));
132 // If the succ is the loop header, and if L is a top-level loop, then this
133 // is an entrance into a loop through the header, which is also ok.
134 if (L->getHeader() == TI->getSuccessor(i) && L->getParentLoop() == 0)
137 // Otherwise, this is an entrance into a loop from some place invalid.
138 // Either the loop structure is invalid and this is not a natural loop (in
139 // which case the compiler is buggy somewhere else) or BB is unreachable.
140 BlockUnreachable = true;
144 // If this block is ok, check the next one.
145 if (!BlockUnreachable) continue;
147 // Otherwise, this block is dead. To clean up the CFG and to allow later
148 // loop transformations to ignore this case, we delete the edges into the
149 // loop by replacing the terminator.
151 // Remove PHI entries from the successors.
152 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
153 TI->getSuccessor(i)->removePredecessor(BB);
155 // Add a new unreachable instruction.
156 new UnreachableInst(TI);
158 // Delete the dead terminator.
159 if (AA) AA->deleteValue(&BB->back());
160 BB->getInstList().pop_back();
164 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
165 Changed |= ProcessLoop(*I);
170 /// ProcessLoop - Walk the loop structure in depth first order, ensuring that
171 /// all loops have preheaders.
173 bool LoopSimplify::ProcessLoop(Loop *L) {
174 bool Changed = false;
177 // Canonicalize inner loops before outer loops. Inner loop canonicalization
178 // can provide work for the outer loop to canonicalize.
179 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
180 Changed |= ProcessLoop(*I);
182 assert(L->getBlocks()[0] == L->getHeader() &&
183 "Header isn't first block in loop?");
185 // Does the loop already have a preheader? If so, don't insert one.
186 if (L->getLoopPreheader() == 0) {
187 InsertPreheaderForLoop(L);
192 // Next, check to make sure that all exit nodes of the loop only have
193 // predecessors that are inside of the loop. This check guarantees that the
194 // loop preheader/header will dominate the exit blocks. If the exit block has
195 // predecessors from outside of the loop, split the edge now.
196 std::vector<BasicBlock*> ExitBlocks;
197 L->getExitBlocks(ExitBlocks);
199 SetVector<BasicBlock*> ExitBlockSet(ExitBlocks.begin(), ExitBlocks.end());
200 for (SetVector<BasicBlock*>::iterator I = ExitBlockSet.begin(),
201 E = ExitBlockSet.end(); I != E; ++I) {
202 BasicBlock *ExitBlock = *I;
203 for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock);
205 // Must be exactly this loop: no subloops, parent loops, or non-loop preds
207 if (!L->contains(*PI)) {
208 RewriteLoopExitBlock(L, ExitBlock);
215 // If the header has more than two predecessors at this point (from the
216 // preheader and from multiple backedges), we must adjust the loop.
217 unsigned NumBackedges = L->getNumBackEdges();
218 if (NumBackedges != 1) {
219 // If this is really a nested loop, rip it out into a child loop. Don't do
220 // this for loops with a giant number of backedges, just factor them into a
221 // common backedge instead.
222 if (NumBackedges < 8) {
223 if (Loop *NL = SeparateNestedLoop(L)) {
225 // This is a big restructuring change, reprocess the whole loop.
228 // GCC doesn't tail recursion eliminate this.
233 // If we either couldn't, or didn't want to, identify nesting of the loops,
234 // insert a new block that all backedges target, then make it jump to the
236 InsertUniqueBackedgeBlock(L);
241 // Scan over the PHI nodes in the loop header. Since they now have only two
242 // incoming values (the loop is canonicalized), we may have simplified the PHI
243 // down to 'X = phi [X, Y]', which should be replaced with 'Y'.
245 for (BasicBlock::iterator I = L->getHeader()->begin();
246 (PN = dyn_cast<PHINode>(I++)); )
247 if (Value *V = PN->hasConstantValue()) {
248 PN->replaceAllUsesWith(V);
249 PN->eraseFromParent();
255 /// SplitBlockPredecessors - Split the specified block into two blocks. We want
256 /// to move the predecessors specified in the Preds list to point to the new
257 /// block, leaving the remaining predecessors pointing to BB. This method
258 /// updates the SSA PHINode's, but no other analyses.
260 BasicBlock *LoopSimplify::SplitBlockPredecessors(BasicBlock *BB,
262 const std::vector<BasicBlock*> &Preds) {
264 // Create new basic block, insert right before the original block...
265 BasicBlock *NewBB = new BasicBlock(BB->getName()+Suffix, BB->getParent(), BB);
267 // The preheader first gets an unconditional branch to the loop header...
268 BranchInst *BI = new BranchInst(BB, NewBB);
270 // For every PHI node in the block, insert a PHI node into NewBB where the
271 // incoming values from the out of loop edges are moved to NewBB. We have two
272 // possible cases here. If the loop is dead, we just insert dummy entries
273 // into the PHI nodes for the new edge. If the loop is not dead, we move the
274 // incoming edges in BB into new PHI nodes in NewBB.
276 if (!Preds.empty()) { // Is the loop not obviously dead?
277 // Check to see if the values being merged into the new block need PHI
278 // nodes. If so, insert them.
279 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ) {
280 PHINode *PN = cast<PHINode>(I);
283 // Check to see if all of the values coming in are the same. If so, we
284 // don't need to create a new PHI node.
285 Value *InVal = PN->getIncomingValueForBlock(Preds[0]);
286 for (unsigned i = 1, e = Preds.size(); i != e; ++i)
287 if (InVal != PN->getIncomingValueForBlock(Preds[i])) {
292 // If the values coming into the block are not the same, we need a PHI.
294 // Create the new PHI node, insert it into NewBB at the end of the block
295 PHINode *NewPHI = new PHINode(PN->getType(), PN->getName()+".ph", BI);
296 if (AA) AA->copyValue(PN, NewPHI);
298 // Move all of the edges from blocks outside the loop to the new PHI
299 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
300 Value *V = PN->removeIncomingValue(Preds[i], false);
301 NewPHI->addIncoming(V, Preds[i]);
305 // Remove all of the edges coming into the PHI nodes from outside of the
307 for (unsigned i = 0, e = Preds.size(); i != e; ++i)
308 PN->removeIncomingValue(Preds[i], false);
311 // Add an incoming value to the PHI node in the loop for the preheader
313 PN->addIncoming(InVal, NewBB);
315 // Can we eliminate this phi node now?
316 if (Value *V = PN->hasConstantValue(true)) {
317 Instruction *I = dyn_cast<Instruction>(V);
318 // If I is in NewBB, the ETForest call will fail, because NewBB isn't
319 // registered in ETForest yet. Handle this case explicitly.
320 if (!I || (I->getParent() != NewBB &&
321 getAnalysis<ETForest>().dominates(I, PN))) {
322 PN->replaceAllUsesWith(V);
323 if (AA) AA->deleteValue(PN);
324 BB->getInstList().erase(PN);
329 // Now that the PHI nodes are updated, actually move the edges from
330 // Preds to point to NewBB instead of BB.
332 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
333 TerminatorInst *TI = Preds[i]->getTerminator();
334 for (unsigned s = 0, e = TI->getNumSuccessors(); s != e; ++s)
335 if (TI->getSuccessor(s) == BB)
336 TI->setSuccessor(s, NewBB);
339 } else { // Otherwise the loop is dead...
340 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++I) {
341 PHINode *PN = cast<PHINode>(I);
342 // Insert dummy values as the incoming value...
343 PN->addIncoming(Constant::getNullValue(PN->getType()), NewBB);
349 /// InsertPreheaderForLoop - Once we discover that a loop doesn't have a
350 /// preheader, this method is called to insert one. This method has two phases:
351 /// preheader insertion and analysis updating.
353 void LoopSimplify::InsertPreheaderForLoop(Loop *L) {
354 BasicBlock *Header = L->getHeader();
356 // Compute the set of predecessors of the loop that are not in the loop.
357 std::vector<BasicBlock*> OutsideBlocks;
358 for (pred_iterator PI = pred_begin(Header), PE = pred_end(Header);
360 if (!L->contains(*PI)) // Coming in from outside the loop?
361 OutsideBlocks.push_back(*PI); // Keep track of it...
363 // Split out the loop pre-header.
365 SplitBlockPredecessors(Header, ".preheader", OutsideBlocks);
368 //===--------------------------------------------------------------------===//
369 // Update analysis results now that we have performed the transformation
372 // We know that we have loop information to update... update it now.
373 if (Loop *Parent = L->getParentLoop())
374 Parent->addBasicBlockToLoop(NewBB, *LI);
376 UpdateDomInfoForRevectoredPreds(NewBB, OutsideBlocks);
378 // Make sure that NewBB is put someplace intelligent, which doesn't mess up
379 // code layout too horribly.
380 PlaceSplitBlockCarefully(NewBB, OutsideBlocks, L);
383 /// RewriteLoopExitBlock - Ensure that the loop preheader dominates all exit
384 /// blocks. This method is used to split exit blocks that have predecessors
385 /// outside of the loop.
386 BasicBlock *LoopSimplify::RewriteLoopExitBlock(Loop *L, BasicBlock *Exit) {
387 std::vector<BasicBlock*> LoopBlocks;
388 for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit); I != E; ++I)
390 LoopBlocks.push_back(*I);
392 assert(!LoopBlocks.empty() && "No edges coming in from outside the loop?");
393 BasicBlock *NewBB = SplitBlockPredecessors(Exit, ".loopexit", LoopBlocks);
395 // Update Loop Information - we know that the new block will be in whichever
396 // loop the Exit block is in. Note that it may not be in that immediate loop,
397 // if the successor is some other loop header. In that case, we continue
398 // walking up the loop tree to find a loop that contains both the successor
399 // block and the predecessor block.
400 Loop *SuccLoop = LI->getLoopFor(Exit);
401 while (SuccLoop && !SuccLoop->contains(L->getHeader()))
402 SuccLoop = SuccLoop->getParentLoop();
404 SuccLoop->addBasicBlockToLoop(NewBB, *LI);
406 // Update dominator information (set, immdom, domtree, and domfrontier)
407 UpdateDomInfoForRevectoredPreds(NewBB, LoopBlocks);
411 /// AddBlockAndPredsToSet - Add the specified block, and all of its
412 /// predecessors, to the specified set, if it's not already in there. Stop
413 /// predecessor traversal when we reach StopBlock.
414 static void AddBlockAndPredsToSet(BasicBlock *InputBB, BasicBlock *StopBlock,
415 std::set<BasicBlock*> &Blocks) {
416 std::vector<BasicBlock *> WorkList;
417 WorkList.push_back(InputBB);
419 BasicBlock *BB = WorkList.back(); WorkList.pop_back();
420 if (Blocks.insert(BB).second && BB != StopBlock)
421 // If BB is not already processed and it is not a stop block then
422 // insert its predecessor in the work list
423 for (pred_iterator I = pred_begin(BB), E = pred_end(BB); I != E; ++I) {
424 BasicBlock *WBB = *I;
425 WorkList.push_back(WBB);
427 } while(!WorkList.empty());
430 /// FindPHIToPartitionLoops - The first part of loop-nestification is to find a
431 /// PHI node that tells us how to partition the loops.
432 static PHINode *FindPHIToPartitionLoops(Loop *L, ETForest *EF,
434 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ) {
435 PHINode *PN = cast<PHINode>(I);
437 if (Value *V = PN->hasConstantValue())
438 if (!isa<Instruction>(V) || EF->dominates(cast<Instruction>(V), PN)) {
439 // This is a degenerate PHI already, don't modify it!
440 PN->replaceAllUsesWith(V);
441 if (AA) AA->deleteValue(PN);
442 PN->eraseFromParent();
446 // Scan this PHI node looking for a use of the PHI node by itself.
447 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
448 if (PN->getIncomingValue(i) == PN &&
449 L->contains(PN->getIncomingBlock(i)))
450 // We found something tasty to remove.
456 // PlaceSplitBlockCarefully - If the block isn't already, move the new block to
457 // right after some 'outside block' block. This prevents the preheader from
458 // being placed inside the loop body, e.g. when the loop hasn't been rotated.
459 void LoopSimplify::PlaceSplitBlockCarefully(BasicBlock *NewBB,
460 std::vector<BasicBlock*>&SplitPreds,
462 // Check to see if NewBB is already well placed.
463 Function::iterator BBI = NewBB; --BBI;
464 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) {
465 if (&*BBI == SplitPreds[i])
469 // If it isn't already after an outside block, move it after one. This is
470 // always good as it makes the uncond branch from the outside block into a
473 // Figure out *which* outside block to put this after. Prefer an outside
474 // block that neighbors a BB actually in the loop.
475 BasicBlock *FoundBB = 0;
476 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) {
477 Function::iterator BBI = SplitPreds[i];
478 if (++BBI != NewBB->getParent()->end() &&
480 FoundBB = SplitPreds[i];
485 // If our heuristic for a *good* bb to place this after doesn't find
486 // anything, just pick something. It's likely better than leaving it within
489 FoundBB = SplitPreds[0];
490 NewBB->moveAfter(FoundBB);
494 /// SeparateNestedLoop - If this loop has multiple backedges, try to pull one of
495 /// them out into a nested loop. This is important for code that looks like
500 /// br cond, Loop, Next
502 /// br cond2, Loop, Out
504 /// To identify this common case, we look at the PHI nodes in the header of the
505 /// loop. PHI nodes with unchanging values on one backedge correspond to values
506 /// that change in the "outer" loop, but not in the "inner" loop.
508 /// If we are able to separate out a loop, return the new outer loop that was
511 Loop *LoopSimplify::SeparateNestedLoop(Loop *L) {
512 ETForest *EF = getAnalysisToUpdate<ETForest>();
513 PHINode *PN = FindPHIToPartitionLoops(L, EF, AA);
514 if (PN == 0) return 0; // No known way to partition.
516 // Pull out all predecessors that have varying values in the loop. This
517 // handles the case when a PHI node has multiple instances of itself as
519 std::vector<BasicBlock*> OuterLoopPreds;
520 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
521 if (PN->getIncomingValue(i) != PN ||
522 !L->contains(PN->getIncomingBlock(i)))
523 OuterLoopPreds.push_back(PN->getIncomingBlock(i));
525 BasicBlock *Header = L->getHeader();
526 BasicBlock *NewBB = SplitBlockPredecessors(Header, ".outer", OuterLoopPreds);
528 // Update dominator information (set, immdom, domtree, and domfrontier)
529 UpdateDomInfoForRevectoredPreds(NewBB, OuterLoopPreds);
531 // Make sure that NewBB is put someplace intelligent, which doesn't mess up
532 // code layout too horribly.
533 PlaceSplitBlockCarefully(NewBB, OuterLoopPreds, L);
535 // Create the new outer loop.
536 Loop *NewOuter = new Loop();
538 // Change the parent loop to use the outer loop as its child now.
539 if (Loop *Parent = L->getParentLoop())
540 Parent->replaceChildLoopWith(L, NewOuter);
542 LI->changeTopLevelLoop(L, NewOuter);
544 // This block is going to be our new header block: add it to this loop and all
546 NewOuter->addBasicBlockToLoop(NewBB, *LI);
548 // L is now a subloop of our outer loop.
549 NewOuter->addChildLoop(L);
551 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i)
552 NewOuter->addBlockEntry(L->getBlocks()[i]);
554 // Determine which blocks should stay in L and which should be moved out to
555 // the Outer loop now.
556 std::set<BasicBlock*> BlocksInL;
557 for (pred_iterator PI = pred_begin(Header), E = pred_end(Header); PI!=E; ++PI)
558 if (EF->dominates(Header, *PI))
559 AddBlockAndPredsToSet(*PI, Header, BlocksInL);
562 // Scan all of the loop children of L, moving them to OuterLoop if they are
563 // not part of the inner loop.
564 for (Loop::iterator I = L->begin(); I != L->end(); )
565 if (BlocksInL.count((*I)->getHeader()))
566 ++I; // Loop remains in L
568 NewOuter->addChildLoop(L->removeChildLoop(I));
570 // Now that we know which blocks are in L and which need to be moved to
571 // OuterLoop, move any blocks that need it.
572 for (unsigned i = 0; i != L->getBlocks().size(); ++i) {
573 BasicBlock *BB = L->getBlocks()[i];
574 if (!BlocksInL.count(BB)) {
575 // Move this block to the parent, updating the exit blocks sets
576 L->removeBlockFromLoop(BB);
578 LI->changeLoopFor(BB, NewOuter);
588 /// InsertUniqueBackedgeBlock - This method is called when the specified loop
589 /// has more than one backedge in it. If this occurs, revector all of these
590 /// backedges to target a new basic block and have that block branch to the loop
591 /// header. This ensures that loops have exactly one backedge.
593 void LoopSimplify::InsertUniqueBackedgeBlock(Loop *L) {
594 assert(L->getNumBackEdges() > 1 && "Must have > 1 backedge!");
596 // Get information about the loop
597 BasicBlock *Preheader = L->getLoopPreheader();
598 BasicBlock *Header = L->getHeader();
599 Function *F = Header->getParent();
601 // Figure out which basic blocks contain back-edges to the loop header.
602 std::vector<BasicBlock*> BackedgeBlocks;
603 for (pred_iterator I = pred_begin(Header), E = pred_end(Header); I != E; ++I)
604 if (*I != Preheader) BackedgeBlocks.push_back(*I);
606 // Create and insert the new backedge block...
607 BasicBlock *BEBlock = new BasicBlock(Header->getName()+".backedge", F);
608 BranchInst *BETerminator = new BranchInst(Header, BEBlock);
610 // Move the new backedge block to right after the last backedge block.
611 Function::iterator InsertPos = BackedgeBlocks.back(); ++InsertPos;
612 F->getBasicBlockList().splice(InsertPos, F->getBasicBlockList(), BEBlock);
614 // Now that the block has been inserted into the function, create PHI nodes in
615 // the backedge block which correspond to any PHI nodes in the header block.
616 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
617 PHINode *PN = cast<PHINode>(I);
618 PHINode *NewPN = new PHINode(PN->getType(), PN->getName()+".be",
620 NewPN->reserveOperandSpace(BackedgeBlocks.size());
621 if (AA) AA->copyValue(PN, NewPN);
623 // Loop over the PHI node, moving all entries except the one for the
624 // preheader over to the new PHI node.
625 unsigned PreheaderIdx = ~0U;
626 bool HasUniqueIncomingValue = true;
627 Value *UniqueValue = 0;
628 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
629 BasicBlock *IBB = PN->getIncomingBlock(i);
630 Value *IV = PN->getIncomingValue(i);
631 if (IBB == Preheader) {
634 NewPN->addIncoming(IV, IBB);
635 if (HasUniqueIncomingValue) {
636 if (UniqueValue == 0)
638 else if (UniqueValue != IV)
639 HasUniqueIncomingValue = false;
644 // Delete all of the incoming values from the old PN except the preheader's
645 assert(PreheaderIdx != ~0U && "PHI has no preheader entry??");
646 if (PreheaderIdx != 0) {
647 PN->setIncomingValue(0, PN->getIncomingValue(PreheaderIdx));
648 PN->setIncomingBlock(0, PN->getIncomingBlock(PreheaderIdx));
650 // Nuke all entries except the zero'th.
651 for (unsigned i = 0, e = PN->getNumIncomingValues()-1; i != e; ++i)
652 PN->removeIncomingValue(e-i, false);
654 // Finally, add the newly constructed PHI node as the entry for the BEBlock.
655 PN->addIncoming(NewPN, BEBlock);
657 // As an optimization, if all incoming values in the new PhiNode (which is a
658 // subset of the incoming values of the old PHI node) have the same value,
659 // eliminate the PHI Node.
660 if (HasUniqueIncomingValue) {
661 NewPN->replaceAllUsesWith(UniqueValue);
662 if (AA) AA->deleteValue(NewPN);
663 BEBlock->getInstList().erase(NewPN);
667 // Now that all of the PHI nodes have been inserted and adjusted, modify the
668 // backedge blocks to just to the BEBlock instead of the header.
669 for (unsigned i = 0, e = BackedgeBlocks.size(); i != e; ++i) {
670 TerminatorInst *TI = BackedgeBlocks[i]->getTerminator();
671 for (unsigned Op = 0, e = TI->getNumSuccessors(); Op != e; ++Op)
672 if (TI->getSuccessor(Op) == Header)
673 TI->setSuccessor(Op, BEBlock);
676 //===--- Update all analyses which we must preserve now -----------------===//
678 // Update Loop Information - we know that this block is now in the current
679 // loop and all parent loops.
680 L->addBasicBlockToLoop(BEBlock, *LI);
682 // Update dominator information (set, immdom, domtree, and domfrontier)
683 UpdateDomInfoForRevectoredPreds(BEBlock, BackedgeBlocks);
686 // Returns true if BasicBlock A dominates at least one block in vector B
687 // Helper function for UpdateDomInfoForRevectoredPreds
688 static bool BlockDominatesAny(BasicBlock* A, const std::vector<BasicBlock*>& B,
690 for (std::vector<BasicBlock*>::const_iterator BI = B.begin(), BE = B.end();
692 if (ETF.dominates(A, *BI))
698 /// UpdateDomInfoForRevectoredPreds - This method is used to update the four
699 /// different kinds of dominator information (immediate dominators,
700 /// dominator trees, et-forest and dominance frontiers) after a new block has
701 /// been added to the CFG.
703 /// This only supports the case when an existing block (known as "NewBBSucc"),
704 /// had some of its predecessors factored into a new basic block. This
705 /// transformation inserts a new basic block ("NewBB"), with a single
706 /// unconditional branch to NewBBSucc, and moves some predecessors of
707 /// "NewBBSucc" to now branch to NewBB. These predecessors are listed in
708 /// PredBlocks, even though they are the same as
709 /// pred_begin(NewBB)/pred_end(NewBB).
711 void LoopSimplify::UpdateDomInfoForRevectoredPreds(BasicBlock *NewBB,
712 std::vector<BasicBlock*> &PredBlocks) {
713 assert(!PredBlocks.empty() && "No predblocks??");
714 assert(succ_begin(NewBB) != succ_end(NewBB) &&
715 ++succ_begin(NewBB) == succ_end(NewBB) &&
716 "NewBB should have a single successor!");
717 BasicBlock *NewBBSucc = *succ_begin(NewBB);
718 ETForest& ETF = getAnalysis<ETForest>();
720 // The newly inserted basic block will dominate existing basic blocks iff the
721 // PredBlocks dominate all of the non-pred blocks. If all predblocks dominate
722 // the non-pred blocks, then they all must be the same block!
724 bool NewBBDominatesNewBBSucc = true;
726 BasicBlock *OnePred = PredBlocks[0];
727 unsigned i = 1, e = PredBlocks.size();
728 for (i = 1; !ETF.isReachableFromEntry(OnePred); ++i) {
729 assert(i != e && "Didn't find reachable pred?");
730 OnePred = PredBlocks[i];
734 if (PredBlocks[i] != OnePred && ETF.isReachableFromEntry(OnePred)){
735 NewBBDominatesNewBBSucc = false;
739 if (NewBBDominatesNewBBSucc)
740 for (pred_iterator PI = pred_begin(NewBBSucc), E = pred_end(NewBBSucc);
742 if (*PI != NewBB && !ETF.dominates(NewBBSucc, *PI)) {
743 NewBBDominatesNewBBSucc = false;
748 // The other scenario where the new block can dominate its successors are when
749 // all predecessors of NewBBSucc that are not NewBB are dominated by NewBBSucc
751 if (!NewBBDominatesNewBBSucc) {
752 NewBBDominatesNewBBSucc = true;
753 for (pred_iterator PI = pred_begin(NewBBSucc), E = pred_end(NewBBSucc);
755 if (*PI != NewBB && !ETF.dominates(NewBBSucc, *PI)) {
756 NewBBDominatesNewBBSucc = false;
761 BasicBlock *NewBBIDom = 0;
763 // Update DominatorTree information if it is active.
764 if (DominatorTree *DT = getAnalysisToUpdate<DominatorTree>()) {
765 // If we don't have ImmediateDominator info around, calculate the idom as
769 for (i = 0; i < PredBlocks.size(); ++i)
770 if (ETF.dominates(&PredBlocks[i]->getParent()->getEntryBlock(), PredBlocks[i])) {
771 NewBBIDom = PredBlocks[i];
774 assert(i != PredBlocks.size() && "No reachable preds?");
775 for (i = i + 1; i < PredBlocks.size(); ++i) {
776 if (ETF.dominates(&PredBlocks[i]->getParent()->getEntryBlock(), PredBlocks[i]))
777 NewBBIDom = ETF.nearestCommonDominator(NewBBIDom, PredBlocks[i]);
779 assert(NewBBIDom && "No immediate dominator found??");
781 DomTreeNode *NewBBIDomNode = DT->getNode(NewBBIDom);
783 // Create the new dominator tree node... and set the idom of NewBB.
784 DomTreeNode *NewBBNode = DT->createNewNode(NewBB, NewBBIDomNode);
786 // If NewBB strictly dominates other blocks, then it is now the immediate
787 // dominator of NewBBSucc. Update the dominator tree as appropriate.
788 if (NewBBDominatesNewBBSucc) {
789 DomTreeNode *NewBBSuccNode = DT->getNode(NewBBSucc);
790 DT->changeImmediateDominator(NewBBSuccNode, NewBBNode);
794 // Update ET-Forest information if it is active.
795 if (ETForest *EF = getAnalysisToUpdate<ETForest>()) {
796 EF->addNewBlock(NewBB, NewBBIDom);
797 if (NewBBDominatesNewBBSucc)
798 EF->setImmediateDominator(NewBBSucc, NewBB);
801 // Update dominance frontier information...
802 if (DominanceFrontier *DF = getAnalysisToUpdate<DominanceFrontier>()) {
803 // If NewBB dominates NewBBSucc, then DF(NewBB) is now going to be the
804 // DF(PredBlocks[0]) without the stuff that the new block does not dominate
806 if (NewBBDominatesNewBBSucc) {
807 DominanceFrontier::iterator DFI = DF->find(PredBlocks[0]);
808 if (DFI != DF->end()) {
809 DominanceFrontier::DomSetType Set = DFI->second;
810 // Filter out stuff in Set that we do not dominate a predecessor of.
811 for (DominanceFrontier::DomSetType::iterator SetI = Set.begin(),
812 E = Set.end(); SetI != E;) {
813 bool DominatesPred = false;
814 for (pred_iterator PI = pred_begin(*SetI), E = pred_end(*SetI);
816 if (ETF.dominates(NewBB, *PI))
817 DominatesPred = true;
824 DF->addBasicBlock(NewBB, Set);
828 // DF(NewBB) is {NewBBSucc} because NewBB does not strictly dominate
829 // NewBBSucc, but it does dominate itself (and there is an edge (NewBB ->
830 // NewBBSucc)). NewBBSucc is the single successor of NewBB.
831 DominanceFrontier::DomSetType NewDFSet;
832 NewDFSet.insert(NewBBSucc);
833 DF->addBasicBlock(NewBB, NewDFSet);
836 // Now we must loop over all of the dominance frontiers in the function,
837 // replacing occurrences of NewBBSucc with NewBB in some cases. All
838 // blocks that dominate a block in PredBlocks and contained NewBBSucc in
839 // their dominance frontier must be updated to contain NewBB instead.
841 for (Function::iterator FI = NewBB->getParent()->begin(),
842 FE = NewBB->getParent()->end(); FI != FE; ++FI) {
843 DominanceFrontier::iterator DFI = DF->find(FI);
844 if (DFI == DF->end()) continue; // unreachable block.
846 // Only consider dominators of NewBBSucc
847 if (!DFI->second.count(NewBBSucc)) continue;
849 if (BlockDominatesAny(FI, PredBlocks, ETF)) {
850 // If NewBBSucc should not stay in our dominator frontier, remove it.
851 // We remove it unless there is a predecessor of NewBBSucc that we
852 // dominate, but we don't strictly dominate NewBBSucc.
853 bool ShouldRemove = true;
854 if ((BasicBlock*)FI == NewBBSucc || !ETF.dominates(FI, NewBBSucc)) {
855 // Okay, we know that PredDom does not strictly dominate NewBBSucc.
856 // Check to see if it dominates any predecessors of NewBBSucc.
857 for (pred_iterator PI = pred_begin(NewBBSucc),
858 E = pred_end(NewBBSucc); PI != E; ++PI)
859 if (ETF.dominates(FI, *PI)) {
860 ShouldRemove = false;
865 DF->removeFromFrontier(DFI, NewBBSucc);
866 DF->addToFrontier(DFI, NewBB);