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);
87 void PlaceSplitBlockCarefully(BasicBlock *NewBB,
88 std::vector<BasicBlock*> &SplitPreds,
91 void UpdateDomInfoForRevectoredPreds(BasicBlock *NewBB,
92 std::vector<BasicBlock*> &PredBlocks);
95 RegisterPass<LoopSimplify>
96 X("loopsimplify", "Canonicalize natural loops", true);
99 // Publically exposed interface to pass...
100 const PassInfo *llvm::LoopSimplifyID = X.getPassInfo();
101 FunctionPass *llvm::createLoopSimplifyPass() { return new LoopSimplify(); }
103 /// runOnFunction - Run down all loops in the CFG (recursively, but we could do
104 /// it in any convenient order) inserting preheaders...
106 bool LoopSimplify::runOnFunction(Function &F) {
107 bool Changed = false;
108 LI = &getAnalysis<LoopInfo>();
109 AA = getAnalysisToUpdate<AliasAnalysis>();
111 // Check to see that no blocks (other than the header) in loops have
112 // predecessors that are not in loops. This is not valid for natural loops,
113 // but can occur if the blocks are unreachable. Since they are unreachable we
114 // can just shamelessly destroy their terminators to make them not branch into
116 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
117 // This case can only occur for unreachable blocks. Blocks that are
118 // unreachable can't be in loops, so filter those blocks out.
119 if (LI->getLoopFor(BB)) continue;
121 bool BlockUnreachable = false;
122 TerminatorInst *TI = BB->getTerminator();
124 // Check to see if any successors of this block are non-loop-header loops
125 // that are not the header.
126 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
127 // If this successor is not in a loop, BB is clearly ok.
128 Loop *L = LI->getLoopFor(TI->getSuccessor(i));
131 // If the succ is the loop header, and if L is a top-level loop, then this
132 // is an entrance into a loop through the header, which is also ok.
133 if (L->getHeader() == TI->getSuccessor(i) && L->getParentLoop() == 0)
136 // Otherwise, this is an entrance into a loop from some place invalid.
137 // Either the loop structure is invalid and this is not a natural loop (in
138 // which case the compiler is buggy somewhere else) or BB is unreachable.
139 BlockUnreachable = true;
143 // If this block is ok, check the next one.
144 if (!BlockUnreachable) continue;
146 // Otherwise, this block is dead. To clean up the CFG and to allow later
147 // loop transformations to ignore this case, we delete the edges into the
148 // loop by replacing the terminator.
150 // Remove PHI entries from the successors.
151 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
152 TI->getSuccessor(i)->removePredecessor(BB);
154 // Add a new unreachable instruction.
155 new UnreachableInst(TI);
157 // Delete the dead terminator.
158 if (AA) AA->deleteValue(&BB->back());
159 BB->getInstList().pop_back();
163 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
164 Changed |= ProcessLoop(*I);
169 /// ProcessLoop - Walk the loop structure in depth first order, ensuring that
170 /// all loops have preheaders.
172 bool LoopSimplify::ProcessLoop(Loop *L) {
173 bool Changed = false;
176 // Canonicalize inner loops before outer loops. Inner loop canonicalization
177 // can provide work for the outer loop to canonicalize.
178 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
179 Changed |= ProcessLoop(*I);
181 assert(L->getBlocks()[0] == L->getHeader() &&
182 "Header isn't first block in loop?");
184 // Does the loop already have a preheader? If so, don't insert one.
185 if (L->getLoopPreheader() == 0) {
186 InsertPreheaderForLoop(L);
191 // Next, check to make sure that all exit nodes of the loop only have
192 // predecessors that are inside of the loop. This check guarantees that the
193 // loop preheader/header will dominate the exit blocks. If the exit block has
194 // predecessors from outside of the loop, split the edge now.
195 std::vector<BasicBlock*> ExitBlocks;
196 L->getExitBlocks(ExitBlocks);
198 SetVector<BasicBlock*> ExitBlockSet(ExitBlocks.begin(), ExitBlocks.end());
199 for (SetVector<BasicBlock*>::iterator I = ExitBlockSet.begin(),
200 E = ExitBlockSet.end(); I != E; ++I) {
201 BasicBlock *ExitBlock = *I;
202 for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock);
204 // Must be exactly this loop: no subloops, parent loops, or non-loop preds
206 if (!L->contains(*PI)) {
207 RewriteLoopExitBlock(L, ExitBlock);
214 // If the header has more than two predecessors at this point (from the
215 // preheader and from multiple backedges), we must adjust the loop.
216 unsigned NumBackedges = L->getNumBackEdges();
217 if (NumBackedges != 1) {
218 // If this is really a nested loop, rip it out into a child loop. Don't do
219 // this for loops with a giant number of backedges, just factor them into a
220 // common backedge instead.
221 if (NumBackedges < 8) {
222 if (Loop *NL = SeparateNestedLoop(L)) {
224 // This is a big restructuring change, reprocess the whole loop.
227 // GCC doesn't tail recursion eliminate this.
232 // If we either couldn't, or didn't want to, identify nesting of the loops,
233 // insert a new block that all backedges target, then make it jump to the
235 InsertUniqueBackedgeBlock(L);
240 // Scan over the PHI nodes in the loop header. Since they now have only two
241 // incoming values (the loop is canonicalized), we may have simplified the PHI
242 // down to 'X = phi [X, Y]', which should be replaced with 'Y'.
244 for (BasicBlock::iterator I = L->getHeader()->begin();
245 (PN = dyn_cast<PHINode>(I++)); )
246 if (Value *V = PN->hasConstantValue()) {
247 PN->replaceAllUsesWith(V);
248 PN->eraseFromParent();
254 /// SplitBlockPredecessors - Split the specified block into two blocks. We want
255 /// to move the predecessors specified in the Preds list to point to the new
256 /// block, leaving the remaining predecessors pointing to BB. This method
257 /// updates the SSA PHINode's, but no other analyses.
259 BasicBlock *LoopSimplify::SplitBlockPredecessors(BasicBlock *BB,
261 const std::vector<BasicBlock*> &Preds) {
263 // Create new basic block, insert right before the original block...
264 BasicBlock *NewBB = new BasicBlock(BB->getName()+Suffix, BB->getParent(), BB);
266 // The preheader first gets an unconditional branch to the loop header...
267 BranchInst *BI = new BranchInst(BB, NewBB);
269 // For every PHI node in the block, insert a PHI node into NewBB where the
270 // incoming values from the out of loop edges are moved to NewBB. We have two
271 // possible cases here. If the loop is dead, we just insert dummy entries
272 // into the PHI nodes for the new edge. If the loop is not dead, we move the
273 // incoming edges in BB into new PHI nodes in NewBB.
275 if (!Preds.empty()) { // Is the loop not obviously dead?
276 // Check to see if the values being merged into the new block need PHI
277 // nodes. If so, insert them.
278 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ) {
279 PHINode *PN = cast<PHINode>(I);
282 // Check to see if all of the values coming in are the same. If so, we
283 // don't need to create a new PHI node.
284 Value *InVal = PN->getIncomingValueForBlock(Preds[0]);
285 for (unsigned i = 1, e = Preds.size(); i != e; ++i)
286 if (InVal != PN->getIncomingValueForBlock(Preds[i])) {
291 // If the values coming into the block are not the same, we need a PHI.
293 // Create the new PHI node, insert it into NewBB at the end of the block
294 PHINode *NewPHI = new PHINode(PN->getType(), PN->getName()+".ph", BI);
295 if (AA) AA->copyValue(PN, NewPHI);
297 // Move all of the edges from blocks outside the loop to the new PHI
298 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
299 Value *V = PN->removeIncomingValue(Preds[i], false);
300 NewPHI->addIncoming(V, Preds[i]);
304 // Remove all of the edges coming into the PHI nodes from outside of the
306 for (unsigned i = 0, e = Preds.size(); i != e; ++i)
307 PN->removeIncomingValue(Preds[i], false);
310 // Add an incoming value to the PHI node in the loop for the preheader
312 PN->addIncoming(InVal, NewBB);
314 // Can we eliminate this phi node now?
315 if (Value *V = PN->hasConstantValue(true)) {
316 if (!isa<Instruction>(V) ||
317 getAnalysis<DominatorSet>().dominates(cast<Instruction>(V), 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 *BB, BasicBlock *StopBlock,
411 std::set<BasicBlock*> &Blocks) {
412 if (!Blocks.insert(BB).second) return; // already processed.
413 if (BB == StopBlock) return; // Stop here!
415 for (pred_iterator I = pred_begin(BB), E = pred_end(BB); I != E; ++I)
416 AddBlockAndPredsToSet(*I, StopBlock, Blocks);
419 /// FindPHIToPartitionLoops - The first part of loop-nestification is to find a
420 /// PHI node that tells us how to partition the loops.
421 static PHINode *FindPHIToPartitionLoops(Loop *L, DominatorSet &DS,
423 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ) {
424 PHINode *PN = cast<PHINode>(I);
426 if (Value *V = PN->hasConstantValue())
427 if (!isa<Instruction>(V) || DS.dominates(cast<Instruction>(V), PN)) {
428 // This is a degenerate PHI already, don't modify it!
429 PN->replaceAllUsesWith(V);
430 if (AA) AA->deleteValue(PN);
431 PN->eraseFromParent();
435 // Scan this PHI node looking for a use of the PHI node by itself.
436 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
437 if (PN->getIncomingValue(i) == PN &&
438 L->contains(PN->getIncomingBlock(i)))
439 // We found something tasty to remove.
445 // PlaceSplitBlockCarefully - If the block isn't already, move the new block to
446 // right after some 'outside block' block. This prevents the preheader from
447 // being placed inside the loop body, e.g. when the loop hasn't been rotated.
448 void LoopSimplify::PlaceSplitBlockCarefully(BasicBlock *NewBB,
449 std::vector<BasicBlock*>&SplitPreds,
451 // Check to see if NewBB is already well placed.
452 Function::iterator BBI = NewBB; --BBI;
453 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) {
454 if (&*BBI == SplitPreds[i])
458 // If it isn't already after an outside block, move it after one. This is
459 // always good as it makes the uncond branch from the outside block into a
462 // Figure out *which* outside block to put this after. Prefer an outside
463 // block that neighbors a BB actually in the loop.
464 BasicBlock *FoundBB = 0;
465 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) {
466 Function::iterator BBI = SplitPreds[i];
467 if (++BBI != NewBB->getParent()->end() &&
469 FoundBB = SplitPreds[i];
474 // If our heuristic for a *good* bb to place this after doesn't find
475 // anything, just pick something. It's likely better than leaving it within
478 FoundBB = SplitPreds[0];
479 NewBB->moveAfter(FoundBB);
483 /// SeparateNestedLoop - If this loop has multiple backedges, try to pull one of
484 /// them out into a nested loop. This is important for code that looks like
489 /// br cond, Loop, Next
491 /// br cond2, Loop, Out
493 /// To identify this common case, we look at the PHI nodes in the header of the
494 /// loop. PHI nodes with unchanging values on one backedge correspond to values
495 /// that change in the "outer" loop, but not in the "inner" loop.
497 /// If we are able to separate out a loop, return the new outer loop that was
500 Loop *LoopSimplify::SeparateNestedLoop(Loop *L) {
501 PHINode *PN = FindPHIToPartitionLoops(L, getAnalysis<DominatorSet>(), AA);
502 if (PN == 0) return 0; // No known way to partition.
504 // Pull out all predecessors that have varying values in the loop. This
505 // handles the case when a PHI node has multiple instances of itself as
507 std::vector<BasicBlock*> OuterLoopPreds;
508 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
509 if (PN->getIncomingValue(i) != PN ||
510 !L->contains(PN->getIncomingBlock(i)))
511 OuterLoopPreds.push_back(PN->getIncomingBlock(i));
513 BasicBlock *Header = L->getHeader();
514 BasicBlock *NewBB = SplitBlockPredecessors(Header, ".outer", OuterLoopPreds);
516 // Update dominator information (set, immdom, domtree, and domfrontier)
517 UpdateDomInfoForRevectoredPreds(NewBB, OuterLoopPreds);
519 // Make sure that NewBB is put someplace intelligent, which doesn't mess up
520 // code layout too horribly.
521 PlaceSplitBlockCarefully(NewBB, OuterLoopPreds, L);
523 // Create the new outer loop.
524 Loop *NewOuter = new Loop();
526 // Change the parent loop to use the outer loop as its child now.
527 if (Loop *Parent = L->getParentLoop())
528 Parent->replaceChildLoopWith(L, NewOuter);
530 LI->changeTopLevelLoop(L, NewOuter);
532 // This block is going to be our new header block: add it to this loop and all
534 NewOuter->addBasicBlockToLoop(NewBB, *LI);
536 // L is now a subloop of our outer loop.
537 NewOuter->addChildLoop(L);
539 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i)
540 NewOuter->addBlockEntry(L->getBlocks()[i]);
542 // Determine which blocks should stay in L and which should be moved out to
543 // the Outer loop now.
544 DominatorSet &DS = getAnalysis<DominatorSet>();
545 std::set<BasicBlock*> BlocksInL;
546 for (pred_iterator PI = pred_begin(Header), E = pred_end(Header); PI!=E; ++PI)
547 if (DS.dominates(Header, *PI))
548 AddBlockAndPredsToSet(*PI, Header, BlocksInL);
551 // Scan all of the loop children of L, moving them to OuterLoop if they are
552 // not part of the inner loop.
553 for (Loop::iterator I = L->begin(); I != L->end(); )
554 if (BlocksInL.count((*I)->getHeader()))
555 ++I; // Loop remains in L
557 NewOuter->addChildLoop(L->removeChildLoop(I));
559 // Now that we know which blocks are in L and which need to be moved to
560 // OuterLoop, move any blocks that need it.
561 for (unsigned i = 0; i != L->getBlocks().size(); ++i) {
562 BasicBlock *BB = L->getBlocks()[i];
563 if (!BlocksInL.count(BB)) {
564 // Move this block to the parent, updating the exit blocks sets
565 L->removeBlockFromLoop(BB);
567 LI->changeLoopFor(BB, NewOuter);
577 /// InsertUniqueBackedgeBlock - This method is called when the specified loop
578 /// has more than one backedge in it. If this occurs, revector all of these
579 /// backedges to target a new basic block and have that block branch to the loop
580 /// header. This ensures that loops have exactly one backedge.
582 void LoopSimplify::InsertUniqueBackedgeBlock(Loop *L) {
583 assert(L->getNumBackEdges() > 1 && "Must have > 1 backedge!");
585 // Get information about the loop
586 BasicBlock *Preheader = L->getLoopPreheader();
587 BasicBlock *Header = L->getHeader();
588 Function *F = Header->getParent();
590 // Figure out which basic blocks contain back-edges to the loop header.
591 std::vector<BasicBlock*> BackedgeBlocks;
592 for (pred_iterator I = pred_begin(Header), E = pred_end(Header); I != E; ++I)
593 if (*I != Preheader) BackedgeBlocks.push_back(*I);
595 // Create and insert the new backedge block...
596 BasicBlock *BEBlock = new BasicBlock(Header->getName()+".backedge", F);
597 BranchInst *BETerminator = new BranchInst(Header, BEBlock);
599 // Move the new backedge block to right after the last backedge block.
600 Function::iterator InsertPos = BackedgeBlocks.back(); ++InsertPos;
601 F->getBasicBlockList().splice(InsertPos, F->getBasicBlockList(), BEBlock);
603 // Now that the block has been inserted into the function, create PHI nodes in
604 // the backedge block which correspond to any PHI nodes in the header block.
605 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
606 PHINode *PN = cast<PHINode>(I);
607 PHINode *NewPN = new PHINode(PN->getType(), PN->getName()+".be",
609 NewPN->reserveOperandSpace(BackedgeBlocks.size());
610 if (AA) AA->copyValue(PN, NewPN);
612 // Loop over the PHI node, moving all entries except the one for the
613 // preheader over to the new PHI node.
614 unsigned PreheaderIdx = ~0U;
615 bool HasUniqueIncomingValue = true;
616 Value *UniqueValue = 0;
617 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
618 BasicBlock *IBB = PN->getIncomingBlock(i);
619 Value *IV = PN->getIncomingValue(i);
620 if (IBB == Preheader) {
623 NewPN->addIncoming(IV, IBB);
624 if (HasUniqueIncomingValue) {
625 if (UniqueValue == 0)
627 else if (UniqueValue != IV)
628 HasUniqueIncomingValue = false;
633 // Delete all of the incoming values from the old PN except the preheader's
634 assert(PreheaderIdx != ~0U && "PHI has no preheader entry??");
635 if (PreheaderIdx != 0) {
636 PN->setIncomingValue(0, PN->getIncomingValue(PreheaderIdx));
637 PN->setIncomingBlock(0, PN->getIncomingBlock(PreheaderIdx));
639 // Nuke all entries except the zero'th.
640 for (unsigned i = 0, e = PN->getNumIncomingValues()-1; i != e; ++i)
641 PN->removeIncomingValue(e-i, false);
643 // Finally, add the newly constructed PHI node as the entry for the BEBlock.
644 PN->addIncoming(NewPN, BEBlock);
646 // As an optimization, if all incoming values in the new PhiNode (which is a
647 // subset of the incoming values of the old PHI node) have the same value,
648 // eliminate the PHI Node.
649 if (HasUniqueIncomingValue) {
650 NewPN->replaceAllUsesWith(UniqueValue);
651 if (AA) AA->deleteValue(NewPN);
652 BEBlock->getInstList().erase(NewPN);
656 // Now that all of the PHI nodes have been inserted and adjusted, modify the
657 // backedge blocks to just to the BEBlock instead of the header.
658 for (unsigned i = 0, e = BackedgeBlocks.size(); i != e; ++i) {
659 TerminatorInst *TI = BackedgeBlocks[i]->getTerminator();
660 for (unsigned Op = 0, e = TI->getNumSuccessors(); Op != e; ++Op)
661 if (TI->getSuccessor(Op) == Header)
662 TI->setSuccessor(Op, BEBlock);
665 //===--- Update all analyses which we must preserve now -----------------===//
667 // Update Loop Information - we know that this block is now in the current
668 // loop and all parent loops.
669 L->addBasicBlockToLoop(BEBlock, *LI);
671 // Update dominator information (set, immdom, domtree, and domfrontier)
672 UpdateDomInfoForRevectoredPreds(BEBlock, BackedgeBlocks);
675 /// UpdateDomInfoForRevectoredPreds - This method is used to update the four
676 /// different kinds of dominator information (dominator sets, immediate
677 /// dominators, dominator trees, and dominance frontiers) after a new block has
678 /// been added to the CFG.
680 /// This only supports the case when an existing block (known as "NewBBSucc"),
681 /// had some of its predecessors factored into a new basic block. This
682 /// transformation inserts a new basic block ("NewBB"), with a single
683 /// unconditional branch to NewBBSucc, and moves some predecessors of
684 /// "NewBBSucc" to now branch to NewBB. These predecessors are listed in
685 /// PredBlocks, even though they are the same as
686 /// pred_begin(NewBB)/pred_end(NewBB).
688 void LoopSimplify::UpdateDomInfoForRevectoredPreds(BasicBlock *NewBB,
689 std::vector<BasicBlock*> &PredBlocks) {
690 assert(!PredBlocks.empty() && "No predblocks??");
691 assert(succ_begin(NewBB) != succ_end(NewBB) &&
692 ++succ_begin(NewBB) == succ_end(NewBB) &&
693 "NewBB should have a single successor!");
694 BasicBlock *NewBBSucc = *succ_begin(NewBB);
695 DominatorSet &DS = getAnalysis<DominatorSet>();
697 // Update dominator information... The blocks that dominate NewBB are the
698 // intersection of the dominators of predecessors, plus the block itself.
700 DominatorSet::DomSetType NewBBDomSet = DS.getDominators(PredBlocks[0]);
702 unsigned i, e = PredBlocks.size();
703 // It is possible for some preds to not be reachable, and thus have empty
704 // dominator sets (all blocks must dom themselves, so no domset would
705 // otherwise be empty). If we see any of these, don't intersect with them,
706 // as that would certainly leave the resultant set empty.
707 for (i = 1; NewBBDomSet.empty(); ++i) {
708 assert(i != e && "Didn't find reachable pred?");
709 NewBBDomSet = DS.getDominators(PredBlocks[i]);
712 // Intersect the rest of the non-empty sets.
713 for (; i != e; ++i) {
714 const DominatorSet::DomSetType &PredDS = DS.getDominators(PredBlocks[i]);
716 set_intersect(NewBBDomSet, PredDS);
718 NewBBDomSet.insert(NewBB); // All blocks dominate themselves.
719 DS.addBasicBlock(NewBB, NewBBDomSet);
722 // The newly inserted basic block will dominate existing basic blocks iff the
723 // PredBlocks dominate all of the non-pred blocks. If all predblocks dominate
724 // the non-pred blocks, then they all must be the same block!
726 bool NewBBDominatesNewBBSucc = true;
728 BasicBlock *OnePred = PredBlocks[0];
729 unsigned i, e = PredBlocks.size();
730 for (i = 1; !DS.isReachable(OnePred); ++i) {
731 assert(i != e && "Didn't find reachable pred?");
732 OnePred = PredBlocks[i];
736 if (PredBlocks[i] != OnePred && DS.isReachable(PredBlocks[i])) {
737 NewBBDominatesNewBBSucc = false;
741 if (NewBBDominatesNewBBSucc)
742 for (pred_iterator PI = pred_begin(NewBBSucc), E = pred_end(NewBBSucc);
744 if (*PI != NewBB && !DS.dominates(NewBBSucc, *PI)) {
745 NewBBDominatesNewBBSucc = false;
750 // The other scenario where the new block can dominate its successors are when
751 // all predecessors of NewBBSucc that are not NewBB are dominated by NewBBSucc
753 if (!NewBBDominatesNewBBSucc) {
754 NewBBDominatesNewBBSucc = true;
755 for (pred_iterator PI = pred_begin(NewBBSucc), E = pred_end(NewBBSucc);
757 if (*PI != NewBB && !DS.dominates(NewBBSucc, *PI)) {
758 NewBBDominatesNewBBSucc = false;
763 // If NewBB dominates some blocks, then it will dominate all blocks that
765 if (NewBBDominatesNewBBSucc) {
766 BasicBlock *PredBlock = PredBlocks[0];
767 Function *F = NewBB->getParent();
768 for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
769 if (DS.dominates(NewBBSucc, I))
770 DS.addDominator(I, NewBB);
773 // Update immediate dominator information if we have it.
774 BasicBlock *NewBBIDom = 0;
775 if (ImmediateDominators *ID = getAnalysisToUpdate<ImmediateDominators>()) {
776 // To find the immediate dominator of the new exit node, we trace up the
777 // immediate dominators of a predecessor until we find a basic block that
778 // dominates the exit block.
780 BasicBlock *Dom = PredBlocks[0]; // Some random predecessor.
782 // Find a reachable pred.
783 for (unsigned i = 1; !DS.isReachable(Dom); ++i) {
784 assert(i != PredBlocks.size() && "Didn't find reachable pred!");
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 // Scan all the pred blocks that were pulled out. Any individual one may
813 // actually be unreachable, which would mean it doesn't have dom info.
815 for (unsigned i = 0; !NewBBIDomNode; ++i) {
816 assert(i != PredBlocks.size() && "No reachable preds?");
817 NewBBIDomNode = DT->getNode(PredBlocks[i]);
820 while (!NewBBDomSet.count(NewBBIDomNode->getBlock())) {
821 NewBBIDomNode = NewBBIDomNode->getIDom();
822 assert(NewBBIDomNode && "No shared dominator found??");
824 NewBBIDom = NewBBIDomNode->getBlock();
827 // Create the new dominator tree node... and set the idom of NewBB.
828 DominatorTree::Node *NewBBNode = DT->createNewNode(NewBB, NewBBIDomNode);
830 // If NewBB strictly dominates other blocks, then it is now the immediate
831 // dominator of NewBBSucc. Update the dominator tree as appropriate.
832 if (NewBBDominatesNewBBSucc) {
833 DominatorTree::Node *NewBBSuccNode = DT->getNode(NewBBSucc);
834 DT->changeImmediateDominator(NewBBSuccNode, NewBBNode);
838 // Update ET-Forest information if it is active.
839 if (ETForest *EF = getAnalysisToUpdate<ETForest>()) {
840 EF->addNewBlock(NewBB, NewBBIDom);
841 if (NewBBDominatesNewBBSucc)
842 EF->setImmediateDominator(NewBBSucc, NewBB);
845 // Update dominance frontier information...
846 if (DominanceFrontier *DF = getAnalysisToUpdate<DominanceFrontier>()) {
847 // If NewBB dominates NewBBSucc, then DF(NewBB) is now going to be the
848 // DF(PredBlocks[0]) without the stuff that the new block does not dominate
850 if (NewBBDominatesNewBBSucc) {
851 DominanceFrontier::iterator DFI = DF->find(PredBlocks[0]);
852 if (DFI != DF->end()) {
853 DominanceFrontier::DomSetType Set = DFI->second;
854 // Filter out stuff in Set that we do not dominate a predecessor of.
855 for (DominanceFrontier::DomSetType::iterator SetI = Set.begin(),
856 E = Set.end(); SetI != E;) {
857 bool DominatesPred = false;
858 for (pred_iterator PI = pred_begin(*SetI), E = pred_end(*SetI);
860 if (DS.dominates(NewBB, *PI))
861 DominatesPred = true;
868 DF->addBasicBlock(NewBB, Set);
872 // DF(NewBB) is {NewBBSucc} because NewBB does not strictly dominate
873 // NewBBSucc, but it does dominate itself (and there is an edge (NewBB ->
874 // NewBBSucc)). NewBBSucc is the single successor of NewBB.
875 DominanceFrontier::DomSetType NewDFSet;
876 NewDFSet.insert(NewBBSucc);
877 DF->addBasicBlock(NewBB, NewDFSet);
880 // Now we must loop over all of the dominance frontiers in the function,
881 // replacing occurrences of NewBBSucc with NewBB in some cases. All
882 // blocks that dominate a block in PredBlocks and contained NewBBSucc in
883 // their dominance frontier must be updated to contain NewBB instead.
885 for (unsigned i = 0, e = PredBlocks.size(); i != e; ++i) {
886 BasicBlock *Pred = PredBlocks[i];
887 // Get all of the dominators of the predecessor...
888 const DominatorSet::DomSetType &PredDoms = DS.getDominators(Pred);
889 for (DominatorSet::DomSetType::const_iterator PDI = PredDoms.begin(),
890 PDE = PredDoms.end(); PDI != PDE; ++PDI) {
891 BasicBlock *PredDom = *PDI;
893 // If the NewBBSucc node is in DF(PredDom), then PredDom didn't
894 // dominate NewBBSucc but did dominate a predecessor of it. Now we
895 // change this entry to include NewBB in the DF instead of NewBBSucc.
896 DominanceFrontier::iterator DFI = DF->find(PredDom);
897 assert(DFI != DF->end() && "No dominance frontier for node?");
898 if (DFI->second.count(NewBBSucc)) {
899 // If NewBBSucc should not stay in our dominator frontier, remove it.
900 // We remove it unless there is a predecessor of NewBBSucc that we
901 // dominate, but we don't strictly dominate NewBBSucc.
902 bool ShouldRemove = true;
903 if (PredDom == NewBBSucc || !DS.dominates(PredDom, NewBBSucc)) {
904 // Okay, we know that PredDom does not strictly dominate NewBBSucc.
905 // Check to see if it dominates any predecessors of NewBBSucc.
906 for (pred_iterator PI = pred_begin(NewBBSucc),
907 E = pred_end(NewBBSucc); PI != E; ++PI)
908 if (DS.dominates(PredDom, *PI)) {
909 ShouldRemove = false;
915 DF->removeFromFrontier(DFI, NewBBSucc);
916 DF->addToFrontier(DFI, NewBB);