1 //===- LoopSimplify.cpp - Loop Canonicalization Pass ----------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This 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 // Indirectbr instructions introduce several complications. If the loop
27 // contains or is entered by an indirectbr instruction, it may not be possible
28 // to transform the loop and make these guarantees. Client code should check
29 // that these conditions are true before relying on them.
31 // Note that the simplifycfg pass will clean up blocks which are split out but
32 // end up being unnecessary, so usage of this pass should not pessimize
35 // This pass obviously modifies the CFG, but updates loop information and
36 // dominator information.
38 //===----------------------------------------------------------------------===//
40 #include "llvm/Transforms/Scalar.h"
41 #include "llvm/ADT/DepthFirstIterator.h"
42 #include "llvm/ADT/SetOperations.h"
43 #include "llvm/ADT/SetVector.h"
44 #include "llvm/ADT/SmallVector.h"
45 #include "llvm/ADT/Statistic.h"
46 #include "llvm/Analysis/AliasAnalysis.h"
47 #include "llvm/Analysis/AssumptionCache.h"
48 #include "llvm/Analysis/DependenceAnalysis.h"
49 #include "llvm/Analysis/InstructionSimplify.h"
50 #include "llvm/Analysis/LoopInfo.h"
51 #include "llvm/Analysis/ScalarEvolution.h"
52 #include "llvm/IR/CFG.h"
53 #include "llvm/IR/Constants.h"
54 #include "llvm/IR/DataLayout.h"
55 #include "llvm/IR/Dominators.h"
56 #include "llvm/IR/Function.h"
57 #include "llvm/IR/Instructions.h"
58 #include "llvm/IR/IntrinsicInst.h"
59 #include "llvm/IR/LLVMContext.h"
60 #include "llvm/IR/Type.h"
61 #include "llvm/Support/Debug.h"
62 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
63 #include "llvm/Transforms/Utils/Local.h"
64 #include "llvm/Transforms/Utils/LoopUtils.h"
67 #define DEBUG_TYPE "loop-simplify"
69 STATISTIC(NumInserted, "Number of pre-header or exit blocks inserted");
70 STATISTIC(NumNested , "Number of nested loops split out");
72 // If the block isn't already, move the new block to right after some 'outside
73 // block' block. This prevents the preheader from being placed inside the loop
74 // body, e.g. when the loop hasn't been rotated.
75 static void placeSplitBlockCarefully(BasicBlock *NewBB,
76 SmallVectorImpl<BasicBlock *> &SplitPreds,
78 // Check to see if NewBB is already well placed.
79 Function::iterator BBI = NewBB; --BBI;
80 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) {
81 if (&*BBI == SplitPreds[i])
85 // If it isn't already after an outside block, move it after one. This is
86 // always good as it makes the uncond branch from the outside block into a
89 // Figure out *which* outside block to put this after. Prefer an outside
90 // block that neighbors a BB actually in the loop.
91 BasicBlock *FoundBB = nullptr;
92 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) {
93 Function::iterator BBI = SplitPreds[i];
94 if (++BBI != NewBB->getParent()->end() &&
96 FoundBB = SplitPreds[i];
101 // If our heuristic for a *good* bb to place this after doesn't find
102 // anything, just pick something. It's likely better than leaving it within
105 FoundBB = SplitPreds[0];
106 NewBB->moveAfter(FoundBB);
109 /// InsertPreheaderForLoop - Once we discover that a loop doesn't have a
110 /// preheader, this method is called to insert one. This method has two phases:
111 /// preheader insertion and analysis updating.
113 BasicBlock *llvm::InsertPreheaderForLoop(Loop *L, Pass *PP) {
114 BasicBlock *Header = L->getHeader();
116 // Get analyses that we try to update.
117 auto *AA = PP->getAnalysisIfAvailable<AliasAnalysis>();
118 auto *DTWP = PP->getAnalysisIfAvailable<DominatorTreeWrapperPass>();
119 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
120 auto *LIWP = PP->getAnalysisIfAvailable<LoopInfoWrapperPass>();
121 auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr;
122 bool PreserveLCSSA = PP->mustPreserveAnalysisID(LCSSAID);
124 // Compute the set of predecessors of the loop that are not in the loop.
125 SmallVector<BasicBlock*, 8> OutsideBlocks;
126 for (pred_iterator PI = pred_begin(Header), PE = pred_end(Header);
129 if (!L->contains(P)) { // Coming in from outside the loop?
130 // If the loop is branched to from an indirect branch, we won't
131 // be able to fully transform the loop, because it prohibits
133 if (isa<IndirectBrInst>(P->getTerminator())) return nullptr;
136 OutsideBlocks.push_back(P);
140 // Split out the loop pre-header.
141 BasicBlock *PreheaderBB;
142 if (!Header->isLandingPad()) {
143 PreheaderBB = SplitBlockPredecessors(Header, OutsideBlocks, ".preheader",
144 AA, DT, LI, PreserveLCSSA);
146 SmallVector<BasicBlock*, 2> NewBBs;
147 SplitLandingPadPredecessors(Header, OutsideBlocks, ".preheader",
148 ".split-lp", NewBBs, AA, DT, LI, PreserveLCSSA);
149 PreheaderBB = NewBBs[0];
152 PreheaderBB->getTerminator()->setDebugLoc(
153 Header->getFirstNonPHI()->getDebugLoc());
154 DEBUG(dbgs() << "LoopSimplify: Creating pre-header "
155 << PreheaderBB->getName() << "\n");
157 // Make sure that NewBB is put someplace intelligent, which doesn't mess up
158 // code layout too horribly.
159 placeSplitBlockCarefully(PreheaderBB, OutsideBlocks, L);
164 /// \brief Ensure that the loop preheader dominates all exit blocks.
166 /// This method is used to split exit blocks that have predecessors outside of
168 static BasicBlock *rewriteLoopExitBlock(Loop *L, BasicBlock *Exit,
169 AliasAnalysis *AA, DominatorTree *DT,
170 LoopInfo *LI, Pass *PP) {
171 SmallVector<BasicBlock*, 8> LoopBlocks;
172 for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit); I != E; ++I) {
174 if (L->contains(P)) {
175 // Don't do this if the loop is exited via an indirect branch.
176 if (isa<IndirectBrInst>(P->getTerminator())) return nullptr;
178 LoopBlocks.push_back(P);
182 assert(!LoopBlocks.empty() && "No edges coming in from outside the loop?");
183 BasicBlock *NewExitBB = nullptr;
185 bool PreserveLCSSA = PP->mustPreserveAnalysisID(LCSSAID);
187 if (Exit->isLandingPad()) {
188 SmallVector<BasicBlock*, 2> NewBBs;
189 SplitLandingPadPredecessors(Exit, LoopBlocks, ".loopexit", ".nonloopexit",
190 NewBBs, AA, DT, LI, PreserveLCSSA);
191 NewExitBB = NewBBs[0];
193 NewExitBB = SplitBlockPredecessors(Exit, LoopBlocks, ".loopexit", AA, DT,
197 DEBUG(dbgs() << "LoopSimplify: Creating dedicated exit block "
198 << NewExitBB->getName() << "\n");
202 /// Add the specified block, and all of its predecessors, to the specified set,
203 /// if it's not already in there. Stop predecessor traversal when we reach
205 static void addBlockAndPredsToSet(BasicBlock *InputBB, BasicBlock *StopBlock,
206 std::set<BasicBlock*> &Blocks) {
207 SmallVector<BasicBlock *, 8> Worklist;
208 Worklist.push_back(InputBB);
210 BasicBlock *BB = Worklist.pop_back_val();
211 if (Blocks.insert(BB).second && BB != StopBlock)
212 // If BB is not already processed and it is not a stop block then
213 // insert its predecessor in the work list
214 for (pred_iterator I = pred_begin(BB), E = pred_end(BB); I != E; ++I) {
215 BasicBlock *WBB = *I;
216 Worklist.push_back(WBB);
218 } while (!Worklist.empty());
221 /// \brief The first part of loop-nestification is to find a PHI node that tells
222 /// us how to partition the loops.
223 static PHINode *findPHIToPartitionLoops(Loop *L, AliasAnalysis *AA,
225 AssumptionCache *AC) {
226 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ) {
227 PHINode *PN = cast<PHINode>(I);
229 if (Value *V = SimplifyInstruction(PN, nullptr, nullptr, DT, AC)) {
230 // This is a degenerate PHI already, don't modify it!
231 PN->replaceAllUsesWith(V);
232 if (AA) AA->deleteValue(PN);
233 PN->eraseFromParent();
237 // Scan this PHI node looking for a use of the PHI node by itself.
238 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
239 if (PN->getIncomingValue(i) == PN &&
240 L->contains(PN->getIncomingBlock(i)))
241 // We found something tasty to remove.
247 /// \brief If this loop has multiple backedges, try to pull one of them out into
250 /// This is important for code that looks like
255 /// br cond, Loop, Next
257 /// br cond2, Loop, Out
259 /// To identify this common case, we look at the PHI nodes in the header of the
260 /// loop. PHI nodes with unchanging values on one backedge correspond to values
261 /// that change in the "outer" loop, but not in the "inner" loop.
263 /// If we are able to separate out a loop, return the new outer loop that was
266 static Loop *separateNestedLoop(Loop *L, BasicBlock *Preheader,
267 AliasAnalysis *AA, DominatorTree *DT,
268 LoopInfo *LI, ScalarEvolution *SE, Pass *PP,
269 AssumptionCache *AC) {
270 // Don't try to separate loops without a preheader.
274 // The header is not a landing pad; preheader insertion should ensure this.
275 assert(!L->getHeader()->isLandingPad() &&
276 "Can't insert backedge to landing pad");
278 PHINode *PN = findPHIToPartitionLoops(L, AA, DT, AC);
279 if (!PN) return nullptr; // No known way to partition.
281 // Pull out all predecessors that have varying values in the loop. This
282 // handles the case when a PHI node has multiple instances of itself as
284 SmallVector<BasicBlock*, 8> OuterLoopPreds;
285 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
286 if (PN->getIncomingValue(i) != PN ||
287 !L->contains(PN->getIncomingBlock(i))) {
288 // We can't split indirectbr edges.
289 if (isa<IndirectBrInst>(PN->getIncomingBlock(i)->getTerminator()))
291 OuterLoopPreds.push_back(PN->getIncomingBlock(i));
294 DEBUG(dbgs() << "LoopSimplify: Splitting out a new outer loop\n");
296 // If ScalarEvolution is around and knows anything about values in
297 // this loop, tell it to forget them, because we're about to
298 // substantially change it.
302 bool PreserveLCSSA = PP->mustPreserveAnalysisID(LCSSAID);
304 BasicBlock *Header = L->getHeader();
305 BasicBlock *NewBB = SplitBlockPredecessors(Header, OuterLoopPreds, ".outer",
306 AA, DT, LI, PreserveLCSSA);
308 // Make sure that NewBB is put someplace intelligent, which doesn't mess up
309 // code layout too horribly.
310 placeSplitBlockCarefully(NewBB, OuterLoopPreds, L);
312 // Create the new outer loop.
313 Loop *NewOuter = new Loop();
315 // Change the parent loop to use the outer loop as its child now.
316 if (Loop *Parent = L->getParentLoop())
317 Parent->replaceChildLoopWith(L, NewOuter);
319 LI->changeTopLevelLoop(L, NewOuter);
321 // L is now a subloop of our outer loop.
322 NewOuter->addChildLoop(L);
324 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
326 NewOuter->addBlockEntry(*I);
328 // Now reset the header in L, which had been moved by
329 // SplitBlockPredecessors for the outer loop.
330 L->moveToHeader(Header);
332 // Determine which blocks should stay in L and which should be moved out to
333 // the Outer loop now.
334 std::set<BasicBlock*> BlocksInL;
335 for (pred_iterator PI=pred_begin(Header), E = pred_end(Header); PI!=E; ++PI) {
337 if (DT->dominates(Header, P))
338 addBlockAndPredsToSet(P, Header, BlocksInL);
341 // Scan all of the loop children of L, moving them to OuterLoop if they are
342 // not part of the inner loop.
343 const std::vector<Loop*> &SubLoops = L->getSubLoops();
344 for (size_t I = 0; I != SubLoops.size(); )
345 if (BlocksInL.count(SubLoops[I]->getHeader()))
346 ++I; // Loop remains in L
348 NewOuter->addChildLoop(L->removeChildLoop(SubLoops.begin() + I));
350 // Now that we know which blocks are in L and which need to be moved to
351 // OuterLoop, move any blocks that need it.
352 for (unsigned i = 0; i != L->getBlocks().size(); ++i) {
353 BasicBlock *BB = L->getBlocks()[i];
354 if (!BlocksInL.count(BB)) {
355 // Move this block to the parent, updating the exit blocks sets
356 L->removeBlockFromLoop(BB);
358 LI->changeLoopFor(BB, NewOuter);
366 /// \brief This method is called when the specified loop has more than one
369 /// If this occurs, revector all of these backedges to target a new basic block
370 /// and have that block branch to the loop header. This ensures that loops
371 /// have exactly one backedge.
372 static BasicBlock *insertUniqueBackedgeBlock(Loop *L, BasicBlock *Preheader,
374 DominatorTree *DT, LoopInfo *LI) {
375 assert(L->getNumBackEdges() > 1 && "Must have > 1 backedge!");
377 // Get information about the loop
378 BasicBlock *Header = L->getHeader();
379 Function *F = Header->getParent();
381 // Unique backedge insertion currently depends on having a preheader.
385 // The header is not a landing pad; preheader insertion should ensure this.
386 assert(!Header->isLandingPad() && "Can't insert backedge to landing pad");
388 // Figure out which basic blocks contain back-edges to the loop header.
389 std::vector<BasicBlock*> BackedgeBlocks;
390 for (pred_iterator I = pred_begin(Header), E = pred_end(Header); I != E; ++I){
393 // Indirectbr edges cannot be split, so we must fail if we find one.
394 if (isa<IndirectBrInst>(P->getTerminator()))
397 if (P != Preheader) BackedgeBlocks.push_back(P);
400 // Create and insert the new backedge block...
401 BasicBlock *BEBlock = BasicBlock::Create(Header->getContext(),
402 Header->getName()+".backedge", F);
403 BranchInst *BETerminator = BranchInst::Create(Header, BEBlock);
405 DEBUG(dbgs() << "LoopSimplify: Inserting unique backedge block "
406 << BEBlock->getName() << "\n");
408 // Move the new backedge block to right after the last backedge block.
409 Function::iterator InsertPos = BackedgeBlocks.back(); ++InsertPos;
410 F->getBasicBlockList().splice(InsertPos, F->getBasicBlockList(), BEBlock);
412 // Now that the block has been inserted into the function, create PHI nodes in
413 // the backedge block which correspond to any PHI nodes in the header block.
414 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
415 PHINode *PN = cast<PHINode>(I);
416 PHINode *NewPN = PHINode::Create(PN->getType(), BackedgeBlocks.size(),
417 PN->getName()+".be", BETerminator);
418 if (AA) AA->copyValue(PN, NewPN);
420 // Loop over the PHI node, moving all entries except the one for the
421 // preheader over to the new PHI node.
422 unsigned PreheaderIdx = ~0U;
423 bool HasUniqueIncomingValue = true;
424 Value *UniqueValue = nullptr;
425 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
426 BasicBlock *IBB = PN->getIncomingBlock(i);
427 Value *IV = PN->getIncomingValue(i);
428 if (IBB == Preheader) {
431 NewPN->addIncoming(IV, IBB);
432 if (HasUniqueIncomingValue) {
435 else if (UniqueValue != IV)
436 HasUniqueIncomingValue = false;
441 // Delete all of the incoming values from the old PN except the preheader's
442 assert(PreheaderIdx != ~0U && "PHI has no preheader entry??");
443 if (PreheaderIdx != 0) {
444 PN->setIncomingValue(0, PN->getIncomingValue(PreheaderIdx));
445 PN->setIncomingBlock(0, PN->getIncomingBlock(PreheaderIdx));
447 // Nuke all entries except the zero'th.
448 for (unsigned i = 0, e = PN->getNumIncomingValues()-1; i != e; ++i)
449 PN->removeIncomingValue(e-i, false);
451 // Finally, add the newly constructed PHI node as the entry for the BEBlock.
452 PN->addIncoming(NewPN, BEBlock);
454 // As an optimization, if all incoming values in the new PhiNode (which is a
455 // subset of the incoming values of the old PHI node) have the same value,
456 // eliminate the PHI Node.
457 if (HasUniqueIncomingValue) {
458 NewPN->replaceAllUsesWith(UniqueValue);
459 if (AA) AA->deleteValue(NewPN);
460 BEBlock->getInstList().erase(NewPN);
464 // Now that all of the PHI nodes have been inserted and adjusted, modify the
465 // backedge blocks to just to the BEBlock instead of the header.
466 for (unsigned i = 0, e = BackedgeBlocks.size(); i != e; ++i) {
467 TerminatorInst *TI = BackedgeBlocks[i]->getTerminator();
468 for (unsigned Op = 0, e = TI->getNumSuccessors(); Op != e; ++Op)
469 if (TI->getSuccessor(Op) == Header)
470 TI->setSuccessor(Op, BEBlock);
473 //===--- Update all analyses which we must preserve now -----------------===//
475 // Update Loop Information - we know that this block is now in the current
476 // loop and all parent loops.
477 L->addBasicBlockToLoop(BEBlock, *LI);
479 // Update dominator information
480 DT->splitBlock(BEBlock);
485 /// \brief Simplify one loop and queue further loops for simplification.
487 /// FIXME: Currently this accepts both lots of analyses that it uses and a raw
488 /// Pass pointer. The Pass pointer is used by numerous utilities to update
489 /// specific analyses. Rather than a pass it would be much cleaner and more
490 /// explicit if they accepted the analysis directly and then updated it.
491 static bool simplifyOneLoop(Loop *L, SmallVectorImpl<Loop *> &Worklist,
492 AliasAnalysis *AA, DominatorTree *DT, LoopInfo *LI,
493 ScalarEvolution *SE, Pass *PP, const DataLayout *DL,
494 AssumptionCache *AC) {
495 bool Changed = false;
498 // Check to see that no blocks (other than the header) in this loop have
499 // predecessors that are not in the loop. This is not valid for natural
500 // loops, but can occur if the blocks are unreachable. Since they are
501 // unreachable we can just shamelessly delete those CFG edges!
502 for (Loop::block_iterator BB = L->block_begin(), E = L->block_end();
504 if (*BB == L->getHeader()) continue;
506 SmallPtrSet<BasicBlock*, 4> BadPreds;
507 for (pred_iterator PI = pred_begin(*BB),
508 PE = pred_end(*BB); PI != PE; ++PI) {
514 // Delete each unique out-of-loop (and thus dead) predecessor.
515 for (BasicBlock *P : BadPreds) {
517 DEBUG(dbgs() << "LoopSimplify: Deleting edge from dead predecessor "
518 << P->getName() << "\n");
520 // Inform each successor of each dead pred.
521 for (succ_iterator SI = succ_begin(P), SE = succ_end(P); SI != SE; ++SI)
522 (*SI)->removePredecessor(P);
523 // Zap the dead pred's terminator and replace it with unreachable.
524 TerminatorInst *TI = P->getTerminator();
525 TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
526 P->getTerminator()->eraseFromParent();
527 new UnreachableInst(P->getContext(), P);
532 // If there are exiting blocks with branches on undef, resolve the undef in
533 // the direction which will exit the loop. This will help simplify loop
534 // trip count computations.
535 SmallVector<BasicBlock*, 8> ExitingBlocks;
536 L->getExitingBlocks(ExitingBlocks);
537 for (SmallVectorImpl<BasicBlock *>::iterator I = ExitingBlocks.begin(),
538 E = ExitingBlocks.end(); I != E; ++I)
539 if (BranchInst *BI = dyn_cast<BranchInst>((*I)->getTerminator()))
540 if (BI->isConditional()) {
541 if (UndefValue *Cond = dyn_cast<UndefValue>(BI->getCondition())) {
543 DEBUG(dbgs() << "LoopSimplify: Resolving \"br i1 undef\" to exit in "
544 << (*I)->getName() << "\n");
546 BI->setCondition(ConstantInt::get(Cond->getType(),
547 !L->contains(BI->getSuccessor(0))));
549 // This may make the loop analyzable, force SCEV recomputation.
557 // Does the loop already have a preheader? If so, don't insert one.
558 BasicBlock *Preheader = L->getLoopPreheader();
560 Preheader = InsertPreheaderForLoop(L, PP);
567 // Next, check to make sure that all exit nodes of the loop only have
568 // predecessors that are inside of the loop. This check guarantees that the
569 // loop preheader/header will dominate the exit blocks. If the exit block has
570 // predecessors from outside of the loop, split the edge now.
571 SmallVector<BasicBlock*, 8> ExitBlocks;
572 L->getExitBlocks(ExitBlocks);
574 SmallSetVector<BasicBlock *, 8> ExitBlockSet(ExitBlocks.begin(),
576 for (SmallSetVector<BasicBlock *, 8>::iterator I = ExitBlockSet.begin(),
577 E = ExitBlockSet.end(); I != E; ++I) {
578 BasicBlock *ExitBlock = *I;
579 for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock);
581 // Must be exactly this loop: no subloops, parent loops, or non-loop preds
583 if (!L->contains(*PI)) {
584 if (rewriteLoopExitBlock(L, ExitBlock, AA, DT, LI, PP)) {
592 // If the header has more than two predecessors at this point (from the
593 // preheader and from multiple backedges), we must adjust the loop.
594 BasicBlock *LoopLatch = L->getLoopLatch();
596 // If this is really a nested loop, rip it out into a child loop. Don't do
597 // this for loops with a giant number of backedges, just factor them into a
598 // common backedge instead.
599 if (L->getNumBackEdges() < 8) {
601 separateNestedLoop(L, Preheader, AA, DT, LI, SE, PP, AC)) {
603 // Enqueue the outer loop as it should be processed next in our
604 // depth-first nest walk.
605 Worklist.push_back(OuterL);
607 // This is a big restructuring change, reprocess the whole loop.
609 // GCC doesn't tail recursion eliminate this.
610 // FIXME: It isn't clear we can't rely on LLVM to TRE this.
615 // If we either couldn't, or didn't want to, identify nesting of the loops,
616 // insert a new block that all backedges target, then make it jump to the
618 LoopLatch = insertUniqueBackedgeBlock(L, Preheader, AA, DT, LI);
625 // Scan over the PHI nodes in the loop header. Since they now have only two
626 // incoming values (the loop is canonicalized), we may have simplified the PHI
627 // down to 'X = phi [X, Y]', which should be replaced with 'Y'.
629 for (BasicBlock::iterator I = L->getHeader()->begin();
630 (PN = dyn_cast<PHINode>(I++)); )
631 if (Value *V = SimplifyInstruction(PN, nullptr, nullptr, DT, AC)) {
632 if (AA) AA->deleteValue(PN);
633 if (SE) SE->forgetValue(PN);
634 PN->replaceAllUsesWith(V);
635 PN->eraseFromParent();
638 // If this loop has multiple exits and the exits all go to the same
639 // block, attempt to merge the exits. This helps several passes, such
640 // as LoopRotation, which do not support loops with multiple exits.
641 // SimplifyCFG also does this (and this code uses the same utility
642 // function), however this code is loop-aware, where SimplifyCFG is
643 // not. That gives it the advantage of being able to hoist
644 // loop-invariant instructions out of the way to open up more
645 // opportunities, and the disadvantage of having the responsibility
646 // to preserve dominator information.
647 bool UniqueExit = true;
648 if (!ExitBlocks.empty())
649 for (unsigned i = 1, e = ExitBlocks.size(); i != e; ++i)
650 if (ExitBlocks[i] != ExitBlocks[0]) {
655 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
656 BasicBlock *ExitingBlock = ExitingBlocks[i];
657 if (!ExitingBlock->getSinglePredecessor()) continue;
658 BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
659 if (!BI || !BI->isConditional()) continue;
660 CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition());
661 if (!CI || CI->getParent() != ExitingBlock) continue;
663 // Attempt to hoist out all instructions except for the
664 // comparison and the branch.
665 bool AllInvariant = true;
666 bool AnyInvariant = false;
667 for (BasicBlock::iterator I = ExitingBlock->begin(); &*I != BI; ) {
668 Instruction *Inst = I++;
669 // Skip debug info intrinsics.
670 if (isa<DbgInfoIntrinsic>(Inst))
674 if (!L->makeLoopInvariant(Inst, AnyInvariant,
675 Preheader ? Preheader->getTerminator()
677 AllInvariant = false;
683 // The loop disposition of all SCEV expressions that depend on any
684 // hoisted values have also changed.
686 SE->forgetLoopDispositions(L);
688 if (!AllInvariant) continue;
690 // The block has now been cleared of all instructions except for
691 // a comparison and a conditional branch. SimplifyCFG may be able
693 if (!FoldBranchToCommonDest(BI, DL)) continue;
695 // Success. The block is now dead, so remove it from the loop,
696 // update the dominator tree and delete it.
697 DEBUG(dbgs() << "LoopSimplify: Eliminating exiting block "
698 << ExitingBlock->getName() << "\n");
700 // Notify ScalarEvolution before deleting this block. Currently assume the
701 // parent loop doesn't change (spliting edges doesn't count). If blocks,
702 // CFG edges, or other values in the parent loop change, then we need call
703 // to forgetLoop() for the parent instead.
707 assert(pred_begin(ExitingBlock) == pred_end(ExitingBlock));
709 LI->removeBlock(ExitingBlock);
711 DomTreeNode *Node = DT->getNode(ExitingBlock);
712 const std::vector<DomTreeNodeBase<BasicBlock> *> &Children =
714 while (!Children.empty()) {
715 DomTreeNode *Child = Children.front();
716 DT->changeImmediateDominator(Child, Node->getIDom());
718 DT->eraseNode(ExitingBlock);
720 BI->getSuccessor(0)->removePredecessor(ExitingBlock);
721 BI->getSuccessor(1)->removePredecessor(ExitingBlock);
722 ExitingBlock->eraseFromParent();
729 bool llvm::simplifyLoop(Loop *L, DominatorTree *DT, LoopInfo *LI, Pass *PP,
730 AliasAnalysis *AA, ScalarEvolution *SE,
731 const DataLayout *DL, AssumptionCache *AC) {
732 bool Changed = false;
734 // Worklist maintains our depth-first queue of loops in this nest to process.
735 SmallVector<Loop *, 4> Worklist;
736 Worklist.push_back(L);
738 // Walk the worklist from front to back, pushing newly found sub loops onto
739 // the back. This will let us process loops from back to front in depth-first
740 // order. We can use this simple process because loops form a tree.
741 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
742 Loop *L2 = Worklist[Idx];
743 for (Loop::iterator I = L2->begin(), E = L2->end(); I != E; ++I)
744 Worklist.push_back(*I);
747 while (!Worklist.empty())
748 Changed |= simplifyOneLoop(Worklist.pop_back_val(), Worklist, AA, DT, LI,
755 struct LoopSimplify : public FunctionPass {
756 static char ID; // Pass identification, replacement for typeid
757 LoopSimplify() : FunctionPass(ID) {
758 initializeLoopSimplifyPass(*PassRegistry::getPassRegistry());
761 // AA - If we have an alias analysis object to update, this is it, otherwise
767 const DataLayout *DL;
770 bool runOnFunction(Function &F) override;
772 void getAnalysisUsage(AnalysisUsage &AU) const override {
773 AU.addRequired<AssumptionCacheTracker>();
775 // We need loop information to identify the loops...
776 AU.addRequired<DominatorTreeWrapperPass>();
777 AU.addPreserved<DominatorTreeWrapperPass>();
779 AU.addRequired<LoopInfoWrapperPass>();
780 AU.addPreserved<LoopInfoWrapperPass>();
782 AU.addPreserved<AliasAnalysis>();
783 AU.addPreserved<ScalarEvolution>();
784 AU.addPreserved<DependenceAnalysis>();
785 AU.addPreservedID(BreakCriticalEdgesID); // No critical edges added.
788 /// verifyAnalysis() - Verify LoopSimplifyForm's guarantees.
789 void verifyAnalysis() const override;
793 char LoopSimplify::ID = 0;
794 INITIALIZE_PASS_BEGIN(LoopSimplify, "loop-simplify",
795 "Canonicalize natural loops", false, false)
796 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
797 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
798 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
799 INITIALIZE_PASS_END(LoopSimplify, "loop-simplify",
800 "Canonicalize natural loops", false, false)
802 // Publicly exposed interface to pass...
803 char &llvm::LoopSimplifyID = LoopSimplify::ID;
804 Pass *llvm::createLoopSimplifyPass() { return new LoopSimplify(); }
806 /// runOnFunction - Run down all loops in the CFG (recursively, but we could do
807 /// it in any convenient order) inserting preheaders...
809 bool LoopSimplify::runOnFunction(Function &F) {
810 bool Changed = false;
811 AA = getAnalysisIfAvailable<AliasAnalysis>();
812 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
813 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
814 SE = getAnalysisIfAvailable<ScalarEvolution>();
815 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
816 DL = DLP ? &DLP->getDataLayout() : nullptr;
817 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
819 // Simplify each loop nest in the function.
820 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
821 Changed |= simplifyLoop(*I, DT, LI, this, AA, SE, DL, AC);
826 // FIXME: Restore this code when we re-enable verification in verifyAnalysis
829 static void verifyLoop(Loop *L) {
831 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
834 // It used to be possible to just assert L->isLoopSimplifyForm(), however
835 // with the introduction of indirectbr, there are now cases where it's
836 // not possible to transform a loop as necessary. We can at least check
837 // that there is an indirectbr near any time there's trouble.
839 // Indirectbr can interfere with preheader and unique backedge insertion.
840 if (!L->getLoopPreheader() || !L->getLoopLatch()) {
841 bool HasIndBrPred = false;
842 for (pred_iterator PI = pred_begin(L->getHeader()),
843 PE = pred_end(L->getHeader()); PI != PE; ++PI)
844 if (isa<IndirectBrInst>((*PI)->getTerminator())) {
848 assert(HasIndBrPred &&
849 "LoopSimplify has no excuse for missing loop header info!");
853 // Indirectbr can interfere with exit block canonicalization.
854 if (!L->hasDedicatedExits()) {
855 bool HasIndBrExiting = false;
856 SmallVector<BasicBlock*, 8> ExitingBlocks;
857 L->getExitingBlocks(ExitingBlocks);
858 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
859 if (isa<IndirectBrInst>((ExitingBlocks[i])->getTerminator())) {
860 HasIndBrExiting = true;
865 assert(HasIndBrExiting &&
866 "LoopSimplify has no excuse for missing exit block info!");
867 (void)HasIndBrExiting;
872 void LoopSimplify::verifyAnalysis() const {
873 // FIXME: This routine is being called mid-way through the loop pass manager
874 // as loop passes destroy this analysis. That's actually fine, but we have no
875 // way of expressing that here. Once all of the passes that destroy this are
876 // hoisted out of the loop pass manager we can add back verification here.
878 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)