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 // AA - If we have an alias analysis object to update, this is it, otherwise
62 virtual bool runOnFunction(Function &F);
64 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
65 // We need loop information to identify the loops...
66 AU.addRequired<LoopInfo>();
67 AU.addRequired<DominatorTree>();
68 AU.addRequired<ETForest>();
70 AU.addPreserved<LoopInfo>();
71 AU.addPreserved<ETForest>();
72 AU.addPreserved<DominatorTree>();
73 AU.addPreserved<DominanceFrontier>();
74 AU.addPreservedID(BreakCriticalEdgesID); // No critical edges added.
77 bool ProcessLoop(Loop *L);
78 BasicBlock *SplitBlockPredecessors(BasicBlock *BB, const char *Suffix,
79 const std::vector<BasicBlock*> &Preds);
80 BasicBlock *RewriteLoopExitBlock(Loop *L, BasicBlock *Exit);
81 void InsertPreheaderForLoop(Loop *L);
82 Loop *SeparateNestedLoop(Loop *L);
83 void InsertUniqueBackedgeBlock(Loop *L);
84 void PlaceSplitBlockCarefully(BasicBlock *NewBB,
85 std::vector<BasicBlock*> &SplitPreds,
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 Instruction *I = dyn_cast<Instruction>(V);
314 // If I is in NewBB, the ETForest call will fail, because NewBB isn't
315 // registered in ETForest yet. Handle this case explicitly.
316 if (!I || (I->getParent() != NewBB &&
317 getAnalysis<ETForest>().dominates(I, PN))) {
318 PN->replaceAllUsesWith(V);
319 if (AA) AA->deleteValue(PN);
320 BB->getInstList().erase(PN);
325 // Now that the PHI nodes are updated, actually move the edges from
326 // Preds to point to NewBB instead of BB.
328 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
329 TerminatorInst *TI = Preds[i]->getTerminator();
330 for (unsigned s = 0, e = TI->getNumSuccessors(); s != e; ++s)
331 if (TI->getSuccessor(s) == BB)
332 TI->setSuccessor(s, NewBB);
335 } else { // Otherwise the loop is dead...
336 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++I) {
337 PHINode *PN = cast<PHINode>(I);
338 // Insert dummy values as the incoming value...
339 PN->addIncoming(Constant::getNullValue(PN->getType()), NewBB);
345 /// InsertPreheaderForLoop - Once we discover that a loop doesn't have a
346 /// preheader, this method is called to insert one. This method has two phases:
347 /// preheader insertion and analysis updating.
349 void LoopSimplify::InsertPreheaderForLoop(Loop *L) {
350 BasicBlock *Header = L->getHeader();
352 // Compute the set of predecessors of the loop that are not in the loop.
353 std::vector<BasicBlock*> OutsideBlocks;
354 for (pred_iterator PI = pred_begin(Header), PE = pred_end(Header);
356 if (!L->contains(*PI)) // Coming in from outside the loop?
357 OutsideBlocks.push_back(*PI); // Keep track of it...
359 // Split out the loop pre-header.
361 SplitBlockPredecessors(Header, ".preheader", OutsideBlocks);
364 //===--------------------------------------------------------------------===//
365 // Update analysis results now that we have performed the transformation
368 // We know that we have loop information to update... update it now.
369 if (Loop *Parent = L->getParentLoop())
370 Parent->addBasicBlockToLoop(NewBB, *LI);
372 UpdateDomInfoForRevectoredPreds(NewBB, OutsideBlocks);
374 // Make sure that NewBB is put someplace intelligent, which doesn't mess up
375 // code layout too horribly.
376 PlaceSplitBlockCarefully(NewBB, OutsideBlocks, L);
379 /// RewriteLoopExitBlock - Ensure that the loop preheader dominates all exit
380 /// blocks. This method is used to split exit blocks that have predecessors
381 /// outside of the loop.
382 BasicBlock *LoopSimplify::RewriteLoopExitBlock(Loop *L, BasicBlock *Exit) {
383 std::vector<BasicBlock*> LoopBlocks;
384 for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit); I != E; ++I)
386 LoopBlocks.push_back(*I);
388 assert(!LoopBlocks.empty() && "No edges coming in from outside the loop?");
389 BasicBlock *NewBB = SplitBlockPredecessors(Exit, ".loopexit", LoopBlocks);
391 // Update Loop Information - we know that the new block will be in whichever
392 // loop the Exit block is in. Note that it may not be in that immediate loop,
393 // if the successor is some other loop header. In that case, we continue
394 // walking up the loop tree to find a loop that contains both the successor
395 // block and the predecessor block.
396 Loop *SuccLoop = LI->getLoopFor(Exit);
397 while (SuccLoop && !SuccLoop->contains(L->getHeader()))
398 SuccLoop = SuccLoop->getParentLoop();
400 SuccLoop->addBasicBlockToLoop(NewBB, *LI);
402 // Update dominator information (set, immdom, domtree, and domfrontier)
403 UpdateDomInfoForRevectoredPreds(NewBB, LoopBlocks);
407 /// AddBlockAndPredsToSet - Add the specified block, and all of its
408 /// predecessors, to the specified set, if it's not already in there. Stop
409 /// predecessor traversal when we reach StopBlock.
410 static void AddBlockAndPredsToSet(BasicBlock *InputBB, BasicBlock *StopBlock,
411 std::set<BasicBlock*> &Blocks) {
412 std::vector<BasicBlock *> WorkList;
413 WorkList.push_back(InputBB);
415 BasicBlock *BB = WorkList.back(); WorkList.pop_back();
416 if (Blocks.insert(BB).second && BB != StopBlock)
417 // If BB is not already processed and it is not a stop block then
418 // insert its predecessor in the work list
419 for (pred_iterator I = pred_begin(BB), E = pred_end(BB); I != E; ++I) {
420 BasicBlock *WBB = *I;
421 WorkList.push_back(WBB);
423 } while(!WorkList.empty());
426 /// FindPHIToPartitionLoops - The first part of loop-nestification is to find a
427 /// PHI node that tells us how to partition the loops.
428 static PHINode *FindPHIToPartitionLoops(Loop *L, ETForest *EF,
430 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ) {
431 PHINode *PN = cast<PHINode>(I);
433 if (Value *V = PN->hasConstantValue())
434 if (!isa<Instruction>(V) || EF->dominates(cast<Instruction>(V), PN)) {
435 // This is a degenerate PHI already, don't modify it!
436 PN->replaceAllUsesWith(V);
437 if (AA) AA->deleteValue(PN);
438 PN->eraseFromParent();
442 // Scan this PHI node looking for a use of the PHI node by itself.
443 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
444 if (PN->getIncomingValue(i) == PN &&
445 L->contains(PN->getIncomingBlock(i)))
446 // We found something tasty to remove.
452 // PlaceSplitBlockCarefully - If the block isn't already, move the new block to
453 // right after some 'outside block' block. This prevents the preheader from
454 // being placed inside the loop body, e.g. when the loop hasn't been rotated.
455 void LoopSimplify::PlaceSplitBlockCarefully(BasicBlock *NewBB,
456 std::vector<BasicBlock*>&SplitPreds,
458 // Check to see if NewBB is already well placed.
459 Function::iterator BBI = NewBB; --BBI;
460 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) {
461 if (&*BBI == SplitPreds[i])
465 // If it isn't already after an outside block, move it after one. This is
466 // always good as it makes the uncond branch from the outside block into a
469 // Figure out *which* outside block to put this after. Prefer an outside
470 // block that neighbors a BB actually in the loop.
471 BasicBlock *FoundBB = 0;
472 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) {
473 Function::iterator BBI = SplitPreds[i];
474 if (++BBI != NewBB->getParent()->end() &&
476 FoundBB = SplitPreds[i];
481 // If our heuristic for a *good* bb to place this after doesn't find
482 // anything, just pick something. It's likely better than leaving it within
485 FoundBB = SplitPreds[0];
486 NewBB->moveAfter(FoundBB);
490 /// SeparateNestedLoop - If this loop has multiple backedges, try to pull one of
491 /// them out into a nested loop. This is important for code that looks like
496 /// br cond, Loop, Next
498 /// br cond2, Loop, Out
500 /// To identify this common case, we look at the PHI nodes in the header of the
501 /// loop. PHI nodes with unchanging values on one backedge correspond to values
502 /// that change in the "outer" loop, but not in the "inner" loop.
504 /// If we are able to separate out a loop, return the new outer loop that was
507 Loop *LoopSimplify::SeparateNestedLoop(Loop *L) {
508 ETForest *EF = getAnalysisToUpdate<ETForest>();
509 PHINode *PN = FindPHIToPartitionLoops(L, EF, AA);
510 if (PN == 0) return 0; // No known way to partition.
512 // Pull out all predecessors that have varying values in the loop. This
513 // handles the case when a PHI node has multiple instances of itself as
515 std::vector<BasicBlock*> OuterLoopPreds;
516 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
517 if (PN->getIncomingValue(i) != PN ||
518 !L->contains(PN->getIncomingBlock(i)))
519 OuterLoopPreds.push_back(PN->getIncomingBlock(i));
521 BasicBlock *Header = L->getHeader();
522 BasicBlock *NewBB = SplitBlockPredecessors(Header, ".outer", OuterLoopPreds);
524 // Update dominator information (set, immdom, domtree, and domfrontier)
525 UpdateDomInfoForRevectoredPreds(NewBB, OuterLoopPreds);
527 // Make sure that NewBB is put someplace intelligent, which doesn't mess up
528 // code layout too horribly.
529 PlaceSplitBlockCarefully(NewBB, OuterLoopPreds, L);
531 // Create the new outer loop.
532 Loop *NewOuter = new Loop();
534 // Change the parent loop to use the outer loop as its child now.
535 if (Loop *Parent = L->getParentLoop())
536 Parent->replaceChildLoopWith(L, NewOuter);
538 LI->changeTopLevelLoop(L, NewOuter);
540 // This block is going to be our new header block: add it to this loop and all
542 NewOuter->addBasicBlockToLoop(NewBB, *LI);
544 // L is now a subloop of our outer loop.
545 NewOuter->addChildLoop(L);
547 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i)
548 NewOuter->addBlockEntry(L->getBlocks()[i]);
550 // Determine which blocks should stay in L and which should be moved out to
551 // the Outer loop now.
552 std::set<BasicBlock*> BlocksInL;
553 for (pred_iterator PI = pred_begin(Header), E = pred_end(Header); PI!=E; ++PI)
554 if (EF->dominates(Header, *PI))
555 AddBlockAndPredsToSet(*PI, Header, BlocksInL);
558 // Scan all of the loop children of L, moving them to OuterLoop if they are
559 // not part of the inner loop.
560 for (Loop::iterator I = L->begin(); I != L->end(); )
561 if (BlocksInL.count((*I)->getHeader()))
562 ++I; // Loop remains in L
564 NewOuter->addChildLoop(L->removeChildLoop(I));
566 // Now that we know which blocks are in L and which need to be moved to
567 // OuterLoop, move any blocks that need it.
568 for (unsigned i = 0; i != L->getBlocks().size(); ++i) {
569 BasicBlock *BB = L->getBlocks()[i];
570 if (!BlocksInL.count(BB)) {
571 // Move this block to the parent, updating the exit blocks sets
572 L->removeBlockFromLoop(BB);
574 LI->changeLoopFor(BB, NewOuter);
584 /// InsertUniqueBackedgeBlock - This method is called when the specified loop
585 /// has more than one backedge in it. If this occurs, revector all of these
586 /// backedges to target a new basic block and have that block branch to the loop
587 /// header. This ensures that loops have exactly one backedge.
589 void LoopSimplify::InsertUniqueBackedgeBlock(Loop *L) {
590 assert(L->getNumBackEdges() > 1 && "Must have > 1 backedge!");
592 // Get information about the loop
593 BasicBlock *Preheader = L->getLoopPreheader();
594 BasicBlock *Header = L->getHeader();
595 Function *F = Header->getParent();
597 // Figure out which basic blocks contain back-edges to the loop header.
598 std::vector<BasicBlock*> BackedgeBlocks;
599 for (pred_iterator I = pred_begin(Header), E = pred_end(Header); I != E; ++I)
600 if (*I != Preheader) BackedgeBlocks.push_back(*I);
602 // Create and insert the new backedge block...
603 BasicBlock *BEBlock = new BasicBlock(Header->getName()+".backedge", F);
604 BranchInst *BETerminator = new BranchInst(Header, BEBlock);
606 // Move the new backedge block to right after the last backedge block.
607 Function::iterator InsertPos = BackedgeBlocks.back(); ++InsertPos;
608 F->getBasicBlockList().splice(InsertPos, F->getBasicBlockList(), BEBlock);
610 // Now that the block has been inserted into the function, create PHI nodes in
611 // the backedge block which correspond to any PHI nodes in the header block.
612 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
613 PHINode *PN = cast<PHINode>(I);
614 PHINode *NewPN = new PHINode(PN->getType(), PN->getName()+".be",
616 NewPN->reserveOperandSpace(BackedgeBlocks.size());
617 if (AA) AA->copyValue(PN, NewPN);
619 // Loop over the PHI node, moving all entries except the one for the
620 // preheader over to the new PHI node.
621 unsigned PreheaderIdx = ~0U;
622 bool HasUniqueIncomingValue = true;
623 Value *UniqueValue = 0;
624 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
625 BasicBlock *IBB = PN->getIncomingBlock(i);
626 Value *IV = PN->getIncomingValue(i);
627 if (IBB == Preheader) {
630 NewPN->addIncoming(IV, IBB);
631 if (HasUniqueIncomingValue) {
632 if (UniqueValue == 0)
634 else if (UniqueValue != IV)
635 HasUniqueIncomingValue = false;
640 // Delete all of the incoming values from the old PN except the preheader's
641 assert(PreheaderIdx != ~0U && "PHI has no preheader entry??");
642 if (PreheaderIdx != 0) {
643 PN->setIncomingValue(0, PN->getIncomingValue(PreheaderIdx));
644 PN->setIncomingBlock(0, PN->getIncomingBlock(PreheaderIdx));
646 // Nuke all entries except the zero'th.
647 for (unsigned i = 0, e = PN->getNumIncomingValues()-1; i != e; ++i)
648 PN->removeIncomingValue(e-i, false);
650 // Finally, add the newly constructed PHI node as the entry for the BEBlock.
651 PN->addIncoming(NewPN, BEBlock);
653 // As an optimization, if all incoming values in the new PhiNode (which is a
654 // subset of the incoming values of the old PHI node) have the same value,
655 // eliminate the PHI Node.
656 if (HasUniqueIncomingValue) {
657 NewPN->replaceAllUsesWith(UniqueValue);
658 if (AA) AA->deleteValue(NewPN);
659 BEBlock->getInstList().erase(NewPN);
663 // Now that all of the PHI nodes have been inserted and adjusted, modify the
664 // backedge blocks to just to the BEBlock instead of the header.
665 for (unsigned i = 0, e = BackedgeBlocks.size(); i != e; ++i) {
666 TerminatorInst *TI = BackedgeBlocks[i]->getTerminator();
667 for (unsigned Op = 0, e = TI->getNumSuccessors(); Op != e; ++Op)
668 if (TI->getSuccessor(Op) == Header)
669 TI->setSuccessor(Op, BEBlock);
672 //===--- Update all analyses which we must preserve now -----------------===//
674 // Update Loop Information - we know that this block is now in the current
675 // loop and all parent loops.
676 L->addBasicBlockToLoop(BEBlock, *LI);
678 // Update dominator information (set, immdom, domtree, and domfrontier)
679 UpdateDomInfoForRevectoredPreds(BEBlock, BackedgeBlocks);
682 // Returns true if BasicBlock A dominates at least one block in vector B
683 // Helper function for UpdateDomInfoForRevectoredPreds
684 static bool BlockDominatesAny(BasicBlock* A, const std::vector<BasicBlock*>& B,
686 for (std::vector<BasicBlock*>::const_iterator BI = B.begin(), BE = B.end();
688 if (ETF.dominates(A, *BI))
694 /// UpdateDomInfoForRevectoredPreds - This method is used to update the four
695 /// different kinds of dominator information (immediate dominators,
696 /// dominator trees, et-forest and dominance frontiers) after a new block has
697 /// been added to the CFG.
699 /// This only supports the case when an existing block (known as "NewBBSucc"),
700 /// had some of its predecessors factored into a new basic block. This
701 /// transformation inserts a new basic block ("NewBB"), with a single
702 /// unconditional branch to NewBBSucc, and moves some predecessors of
703 /// "NewBBSucc" to now branch to NewBB. These predecessors are listed in
704 /// PredBlocks, even though they are the same as
705 /// pred_begin(NewBB)/pred_end(NewBB).
707 void LoopSimplify::UpdateDomInfoForRevectoredPreds(BasicBlock *NewBB,
708 std::vector<BasicBlock*> &PredBlocks) {
709 assert(!PredBlocks.empty() && "No predblocks??");
710 assert(succ_begin(NewBB) != succ_end(NewBB) &&
711 ++succ_begin(NewBB) == succ_end(NewBB) &&
712 "NewBB should have a single successor!");
713 BasicBlock *NewBBSucc = *succ_begin(NewBB);
714 ETForest& ETF = getAnalysis<ETForest>();
716 // The newly inserted basic block will dominate existing basic blocks iff the
717 // PredBlocks dominate all of the non-pred blocks. If all predblocks dominate
718 // the non-pred blocks, then they all must be the same block!
720 bool NewBBDominatesNewBBSucc = true;
722 BasicBlock *OnePred = PredBlocks[0];
723 unsigned i = 1, e = PredBlocks.size();
724 for (i = 1; !ETF.isReachableFromEntry(OnePred); ++i) {
725 assert(i != e && "Didn't find reachable pred?");
726 OnePred = PredBlocks[i];
730 if (PredBlocks[i] != OnePred && ETF.isReachableFromEntry(OnePred)){
731 NewBBDominatesNewBBSucc = false;
735 if (NewBBDominatesNewBBSucc)
736 for (pred_iterator PI = pred_begin(NewBBSucc), E = pred_end(NewBBSucc);
738 if (*PI != NewBB && !ETF.dominates(NewBBSucc, *PI)) {
739 NewBBDominatesNewBBSucc = false;
744 // The other scenario where the new block can dominate its successors are when
745 // all predecessors of NewBBSucc that are not NewBB are dominated by NewBBSucc
747 if (!NewBBDominatesNewBBSucc) {
748 NewBBDominatesNewBBSucc = true;
749 for (pred_iterator PI = pred_begin(NewBBSucc), E = pred_end(NewBBSucc);
751 if (*PI != NewBB && !ETF.dominates(NewBBSucc, *PI)) {
752 NewBBDominatesNewBBSucc = false;
757 BasicBlock *NewBBIDom = 0;
759 // Update DominatorTree information if it is active.
760 if (DominatorTree *DT = getAnalysisToUpdate<DominatorTree>()) {
761 // If we don't have ImmediateDominator info around, calculate the idom as
765 for (i = 0; i < PredBlocks.size(); ++i)
766 if (ETF.dominates(&PredBlocks[i]->getParent()->getEntryBlock(), PredBlocks[i])) {
767 NewBBIDom = PredBlocks[i];
770 assert(i != PredBlocks.size() && "No reachable preds?");
771 for (i = i + 1; i < PredBlocks.size(); ++i) {
772 if (ETF.dominates(&PredBlocks[i]->getParent()->getEntryBlock(), PredBlocks[i]))
773 NewBBIDom = ETF.nearestCommonDominator(NewBBIDom, PredBlocks[i]);
775 assert(NewBBIDom && "No immediate dominator found??");
777 DominatorTree::Node *NewBBIDomNode = DT->getNode(NewBBIDom);
779 // Create the new dominator tree node... and set the idom of NewBB.
780 DominatorTree::Node *NewBBNode = DT->createNewNode(NewBB, NewBBIDomNode);
782 // If NewBB strictly dominates other blocks, then it is now the immediate
783 // dominator of NewBBSucc. Update the dominator tree as appropriate.
784 if (NewBBDominatesNewBBSucc) {
785 DominatorTree::Node *NewBBSuccNode = DT->getNode(NewBBSucc);
786 DT->changeImmediateDominator(NewBBSuccNode, NewBBNode);
790 // Update ET-Forest information if it is active.
791 if (ETForest *EF = getAnalysisToUpdate<ETForest>()) {
792 EF->addNewBlock(NewBB, NewBBIDom);
793 if (NewBBDominatesNewBBSucc)
794 EF->setImmediateDominator(NewBBSucc, NewBB);
797 // Update dominance frontier information...
798 if (DominanceFrontier *DF = getAnalysisToUpdate<DominanceFrontier>()) {
799 // If NewBB dominates NewBBSucc, then DF(NewBB) is now going to be the
800 // DF(PredBlocks[0]) without the stuff that the new block does not dominate
802 if (NewBBDominatesNewBBSucc) {
803 DominanceFrontier::iterator DFI = DF->find(PredBlocks[0]);
804 if (DFI != DF->end()) {
805 DominanceFrontier::DomSetType Set = DFI->second;
806 // Filter out stuff in Set that we do not dominate a predecessor of.
807 for (DominanceFrontier::DomSetType::iterator SetI = Set.begin(),
808 E = Set.end(); SetI != E;) {
809 bool DominatesPred = false;
810 for (pred_iterator PI = pred_begin(*SetI), E = pred_end(*SetI);
812 if (ETF.dominates(NewBB, *PI))
813 DominatesPred = true;
820 DF->addBasicBlock(NewBB, Set);
824 // DF(NewBB) is {NewBBSucc} because NewBB does not strictly dominate
825 // NewBBSucc, but it does dominate itself (and there is an edge (NewBB ->
826 // NewBBSucc)). NewBBSucc is the single successor of NewBB.
827 DominanceFrontier::DomSetType NewDFSet;
828 NewDFSet.insert(NewBBSucc);
829 DF->addBasicBlock(NewBB, NewDFSet);
832 // Now we must loop over all of the dominance frontiers in the function,
833 // replacing occurrences of NewBBSucc with NewBB in some cases. All
834 // blocks that dominate a block in PredBlocks and contained NewBBSucc in
835 // their dominance frontier must be updated to contain NewBB instead.
837 for (Function::iterator FI = NewBB->getParent()->begin(),
838 FE = NewBB->getParent()->end(); FI != FE; ++FI) {
839 DominanceFrontier::iterator DFI = DF->find(FI);
840 if (DFI == DF->end()) continue; // unreachable block.
842 // Only consider dominators of NewBBSucc
843 if (!DFI->second.count(NewBBSucc)) continue;
845 if (BlockDominatesAny(FI, PredBlocks, ETF)) {
846 // If NewBBSucc should not stay in our dominator frontier, remove it.
847 // We remove it unless there is a predecessor of NewBBSucc that we
848 // dominate, but we don't strictly dominate NewBBSucc.
849 bool ShouldRemove = true;
850 if ((BasicBlock*)FI == NewBBSucc || !ETF.dominates(FI, NewBBSucc)) {
851 // Okay, we know that PredDom does not strictly dominate NewBBSucc.
852 // Check to see if it dominates any predecessors of NewBBSucc.
853 for (pred_iterator PI = pred_begin(NewBBSucc),
854 E = pred_end(NewBBSucc); PI != E; ++PI)
855 if (ETF.dominates(FI, *PI)) {
856 ShouldRemove = false;
861 DF->removeFromFrontier(DFI, NewBBSucc);
862 DF->addToFrontier(DFI, NewBB);