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/Module.h"
61 #include "llvm/IR/Type.h"
62 #include "llvm/Support/Debug.h"
63 #include "llvm/Support/raw_ostream.h"
64 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
65 #include "llvm/Transforms/Utils/Local.h"
66 #include "llvm/Transforms/Utils/LoopUtils.h"
69 #define DEBUG_TYPE "loop-simplify"
71 STATISTIC(NumInserted, "Number of pre-header or exit blocks inserted");
72 STATISTIC(NumNested , "Number of nested loops split out");
74 // If the block isn't already, move the new block to right after some 'outside
75 // block' block. This prevents the preheader from being placed inside the loop
76 // body, e.g. when the loop hasn't been rotated.
77 static void placeSplitBlockCarefully(BasicBlock *NewBB,
78 SmallVectorImpl<BasicBlock *> &SplitPreds,
80 // Check to see if NewBB is already well placed.
81 Function::iterator BBI = NewBB; --BBI;
82 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) {
83 if (&*BBI == SplitPreds[i])
87 // If it isn't already after an outside block, move it after one. This is
88 // always good as it makes the uncond branch from the outside block into a
91 // Figure out *which* outside block to put this after. Prefer an outside
92 // block that neighbors a BB actually in the loop.
93 BasicBlock *FoundBB = nullptr;
94 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) {
95 Function::iterator BBI = SplitPreds[i];
96 if (++BBI != NewBB->getParent()->end() &&
98 FoundBB = SplitPreds[i];
103 // If our heuristic for a *good* bb to place this after doesn't find
104 // anything, just pick something. It's likely better than leaving it within
107 FoundBB = SplitPreds[0];
108 NewBB->moveAfter(FoundBB);
111 /// InsertPreheaderForLoop - Once we discover that a loop doesn't have a
112 /// preheader, this method is called to insert one. This method has two phases:
113 /// preheader insertion and analysis updating.
115 BasicBlock *llvm::InsertPreheaderForLoop(Loop *L, Pass *PP) {
116 BasicBlock *Header = L->getHeader();
118 // Get analyses that we try to update.
119 auto *AA = PP->getAnalysisIfAvailable<AliasAnalysis>();
120 auto *DTWP = PP->getAnalysisIfAvailable<DominatorTreeWrapperPass>();
121 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
122 auto *LIWP = PP->getAnalysisIfAvailable<LoopInfoWrapperPass>();
123 auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr;
124 bool PreserveLCSSA = PP->mustPreserveAnalysisID(LCSSAID);
126 // Compute the set of predecessors of the loop that are not in the loop.
127 SmallVector<BasicBlock*, 8> OutsideBlocks;
128 for (pred_iterator PI = pred_begin(Header), PE = pred_end(Header);
131 if (!L->contains(P)) { // Coming in from outside the loop?
132 // If the loop is branched to from an indirect branch, we won't
133 // be able to fully transform the loop, because it prohibits
135 if (isa<IndirectBrInst>(P->getTerminator())) return nullptr;
138 OutsideBlocks.push_back(P);
142 // Split out the loop pre-header.
143 BasicBlock *PreheaderBB;
144 PreheaderBB = SplitBlockPredecessors(Header, OutsideBlocks, ".preheader",
145 AA, DT, LI, PreserveLCSSA);
147 DEBUG(dbgs() << "LoopSimplify: Creating pre-header "
148 << PreheaderBB->getName() << "\n");
150 // Make sure that NewBB is put someplace intelligent, which doesn't mess up
151 // code layout too horribly.
152 placeSplitBlockCarefully(PreheaderBB, OutsideBlocks, L);
157 /// \brief Ensure that the loop preheader dominates all exit blocks.
159 /// This method is used to split exit blocks that have predecessors outside of
161 static BasicBlock *rewriteLoopExitBlock(Loop *L, BasicBlock *Exit,
162 AliasAnalysis *AA, DominatorTree *DT,
163 LoopInfo *LI, Pass *PP) {
164 SmallVector<BasicBlock*, 8> LoopBlocks;
165 for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit); I != E; ++I) {
167 if (L->contains(P)) {
168 // Don't do this if the loop is exited via an indirect branch.
169 if (isa<IndirectBrInst>(P->getTerminator())) return nullptr;
171 LoopBlocks.push_back(P);
175 assert(!LoopBlocks.empty() && "No edges coming in from outside the loop?");
176 BasicBlock *NewExitBB = nullptr;
178 bool PreserveLCSSA = PP->mustPreserveAnalysisID(LCSSAID);
180 NewExitBB = SplitBlockPredecessors(Exit, LoopBlocks, ".loopexit", AA, DT,
183 DEBUG(dbgs() << "LoopSimplify: Creating dedicated exit block "
184 << NewExitBB->getName() << "\n");
188 /// Add the specified block, and all of its predecessors, to the specified set,
189 /// if it's not already in there. Stop predecessor traversal when we reach
191 static void addBlockAndPredsToSet(BasicBlock *InputBB, BasicBlock *StopBlock,
192 std::set<BasicBlock*> &Blocks) {
193 SmallVector<BasicBlock *, 8> Worklist;
194 Worklist.push_back(InputBB);
196 BasicBlock *BB = Worklist.pop_back_val();
197 if (Blocks.insert(BB).second && BB != StopBlock)
198 // If BB is not already processed and it is not a stop block then
199 // insert its predecessor in the work list
200 for (pred_iterator I = pred_begin(BB), E = pred_end(BB); I != E; ++I) {
201 BasicBlock *WBB = *I;
202 Worklist.push_back(WBB);
204 } while (!Worklist.empty());
207 /// \brief The first part of loop-nestification is to find a PHI node that tells
208 /// us how to partition the loops.
209 static PHINode *findPHIToPartitionLoops(Loop *L, AliasAnalysis *AA,
211 AssumptionCache *AC) {
212 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
213 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ) {
214 PHINode *PN = cast<PHINode>(I);
216 if (Value *V = SimplifyInstruction(PN, DL, nullptr, DT, AC)) {
217 // This is a degenerate PHI already, don't modify it!
218 PN->replaceAllUsesWith(V);
219 if (AA) AA->deleteValue(PN);
220 PN->eraseFromParent();
224 // Scan this PHI node looking for a use of the PHI node by itself.
225 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
226 if (PN->getIncomingValue(i) == PN &&
227 L->contains(PN->getIncomingBlock(i)))
228 // We found something tasty to remove.
234 /// \brief If this loop has multiple backedges, try to pull one of them out into
237 /// This is important for code that looks like
242 /// br cond, Loop, Next
244 /// br cond2, Loop, Out
246 /// To identify this common case, we look at the PHI nodes in the header of the
247 /// loop. PHI nodes with unchanging values on one backedge correspond to values
248 /// that change in the "outer" loop, but not in the "inner" loop.
250 /// If we are able to separate out a loop, return the new outer loop that was
253 static Loop *separateNestedLoop(Loop *L, BasicBlock *Preheader,
254 AliasAnalysis *AA, DominatorTree *DT,
255 LoopInfo *LI, ScalarEvolution *SE, Pass *PP,
256 AssumptionCache *AC) {
257 // Don't try to separate loops without a preheader.
261 // The header is not a landing pad; preheader insertion should ensure this.
262 assert(!L->getHeader()->isLandingPad() &&
263 "Can't insert backedge to landing pad");
265 PHINode *PN = findPHIToPartitionLoops(L, AA, DT, AC);
266 if (!PN) return nullptr; // No known way to partition.
268 // Pull out all predecessors that have varying values in the loop. This
269 // handles the case when a PHI node has multiple instances of itself as
271 SmallVector<BasicBlock*, 8> OuterLoopPreds;
272 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
273 if (PN->getIncomingValue(i) != PN ||
274 !L->contains(PN->getIncomingBlock(i))) {
275 // We can't split indirectbr edges.
276 if (isa<IndirectBrInst>(PN->getIncomingBlock(i)->getTerminator()))
278 OuterLoopPreds.push_back(PN->getIncomingBlock(i));
281 DEBUG(dbgs() << "LoopSimplify: Splitting out a new outer loop\n");
283 // If ScalarEvolution is around and knows anything about values in
284 // this loop, tell it to forget them, because we're about to
285 // substantially change it.
289 bool PreserveLCSSA = PP->mustPreserveAnalysisID(LCSSAID);
291 BasicBlock *Header = L->getHeader();
292 BasicBlock *NewBB = SplitBlockPredecessors(Header, OuterLoopPreds, ".outer",
293 AA, DT, LI, PreserveLCSSA);
295 // Make sure that NewBB is put someplace intelligent, which doesn't mess up
296 // code layout too horribly.
297 placeSplitBlockCarefully(NewBB, OuterLoopPreds, L);
299 // Create the new outer loop.
300 Loop *NewOuter = new Loop();
302 // Change the parent loop to use the outer loop as its child now.
303 if (Loop *Parent = L->getParentLoop())
304 Parent->replaceChildLoopWith(L, NewOuter);
306 LI->changeTopLevelLoop(L, NewOuter);
308 // L is now a subloop of our outer loop.
309 NewOuter->addChildLoop(L);
311 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
313 NewOuter->addBlockEntry(*I);
315 // Now reset the header in L, which had been moved by
316 // SplitBlockPredecessors for the outer loop.
317 L->moveToHeader(Header);
319 // Determine which blocks should stay in L and which should be moved out to
320 // the Outer loop now.
321 std::set<BasicBlock*> BlocksInL;
322 for (pred_iterator PI=pred_begin(Header), E = pred_end(Header); PI!=E; ++PI) {
324 if (DT->dominates(Header, P))
325 addBlockAndPredsToSet(P, Header, BlocksInL);
328 // Scan all of the loop children of L, moving them to OuterLoop if they are
329 // not part of the inner loop.
330 const std::vector<Loop*> &SubLoops = L->getSubLoops();
331 for (size_t I = 0; I != SubLoops.size(); )
332 if (BlocksInL.count(SubLoops[I]->getHeader()))
333 ++I; // Loop remains in L
335 NewOuter->addChildLoop(L->removeChildLoop(SubLoops.begin() + I));
337 // Now that we know which blocks are in L and which need to be moved to
338 // OuterLoop, move any blocks that need it.
339 for (unsigned i = 0; i != L->getBlocks().size(); ++i) {
340 BasicBlock *BB = L->getBlocks()[i];
341 if (!BlocksInL.count(BB)) {
342 // Move this block to the parent, updating the exit blocks sets
343 L->removeBlockFromLoop(BB);
345 LI->changeLoopFor(BB, NewOuter);
353 /// \brief This method is called when the specified loop has more than one
356 /// If this occurs, revector all of these backedges to target a new basic block
357 /// and have that block branch to the loop header. This ensures that loops
358 /// have exactly one backedge.
359 static BasicBlock *insertUniqueBackedgeBlock(Loop *L, BasicBlock *Preheader,
361 DominatorTree *DT, LoopInfo *LI) {
362 assert(L->getNumBackEdges() > 1 && "Must have > 1 backedge!");
364 // Get information about the loop
365 BasicBlock *Header = L->getHeader();
366 Function *F = Header->getParent();
368 // Unique backedge insertion currently depends on having a preheader.
372 // The header is not a landing pad; preheader insertion should ensure this.
373 assert(!Header->isLandingPad() && "Can't insert backedge to landing pad");
375 // Figure out which basic blocks contain back-edges to the loop header.
376 std::vector<BasicBlock*> BackedgeBlocks;
377 for (pred_iterator I = pred_begin(Header), E = pred_end(Header); I != E; ++I){
380 // Indirectbr edges cannot be split, so we must fail if we find one.
381 if (isa<IndirectBrInst>(P->getTerminator()))
384 if (P != Preheader) BackedgeBlocks.push_back(P);
387 // Create and insert the new backedge block...
388 BasicBlock *BEBlock = BasicBlock::Create(Header->getContext(),
389 Header->getName()+".backedge", F);
390 BranchInst *BETerminator = BranchInst::Create(Header, BEBlock);
392 DEBUG(dbgs() << "LoopSimplify: Inserting unique backedge block "
393 << BEBlock->getName() << "\n");
395 // Move the new backedge block to right after the last backedge block.
396 Function::iterator InsertPos = BackedgeBlocks.back(); ++InsertPos;
397 F->getBasicBlockList().splice(InsertPos, F->getBasicBlockList(), BEBlock);
399 // Now that the block has been inserted into the function, create PHI nodes in
400 // the backedge block which correspond to any PHI nodes in the header block.
401 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
402 PHINode *PN = cast<PHINode>(I);
403 PHINode *NewPN = PHINode::Create(PN->getType(), BackedgeBlocks.size(),
404 PN->getName()+".be", BETerminator);
405 if (AA) AA->copyValue(PN, NewPN);
407 // Loop over the PHI node, moving all entries except the one for the
408 // preheader over to the new PHI node.
409 unsigned PreheaderIdx = ~0U;
410 bool HasUniqueIncomingValue = true;
411 Value *UniqueValue = nullptr;
412 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
413 BasicBlock *IBB = PN->getIncomingBlock(i);
414 Value *IV = PN->getIncomingValue(i);
415 if (IBB == Preheader) {
418 NewPN->addIncoming(IV, IBB);
419 if (HasUniqueIncomingValue) {
422 else if (UniqueValue != IV)
423 HasUniqueIncomingValue = false;
428 // Delete all of the incoming values from the old PN except the preheader's
429 assert(PreheaderIdx != ~0U && "PHI has no preheader entry??");
430 if (PreheaderIdx != 0) {
431 PN->setIncomingValue(0, PN->getIncomingValue(PreheaderIdx));
432 PN->setIncomingBlock(0, PN->getIncomingBlock(PreheaderIdx));
434 // Nuke all entries except the zero'th.
435 for (unsigned i = 0, e = PN->getNumIncomingValues()-1; i != e; ++i)
436 PN->removeIncomingValue(e-i, false);
438 // Finally, add the newly constructed PHI node as the entry for the BEBlock.
439 PN->addIncoming(NewPN, BEBlock);
441 // As an optimization, if all incoming values in the new PhiNode (which is a
442 // subset of the incoming values of the old PHI node) have the same value,
443 // eliminate the PHI Node.
444 if (HasUniqueIncomingValue) {
445 NewPN->replaceAllUsesWith(UniqueValue);
446 if (AA) AA->deleteValue(NewPN);
447 BEBlock->getInstList().erase(NewPN);
451 // Now that all of the PHI nodes have been inserted and adjusted, modify the
452 // backedge blocks to just to the BEBlock instead of the header.
453 for (unsigned i = 0, e = BackedgeBlocks.size(); i != e; ++i) {
454 TerminatorInst *TI = BackedgeBlocks[i]->getTerminator();
455 for (unsigned Op = 0, e = TI->getNumSuccessors(); Op != e; ++Op)
456 if (TI->getSuccessor(Op) == Header)
457 TI->setSuccessor(Op, BEBlock);
460 //===--- Update all analyses which we must preserve now -----------------===//
462 // Update Loop Information - we know that this block is now in the current
463 // loop and all parent loops.
464 L->addBasicBlockToLoop(BEBlock, *LI);
466 // Update dominator information
467 DT->splitBlock(BEBlock);
472 /// \brief Simplify one loop and queue further loops for simplification.
474 /// FIXME: Currently this accepts both lots of analyses that it uses and a raw
475 /// Pass pointer. The Pass pointer is used by numerous utilities to update
476 /// specific analyses. Rather than a pass it would be much cleaner and more
477 /// explicit if they accepted the analysis directly and then updated it.
478 static bool simplifyOneLoop(Loop *L, SmallVectorImpl<Loop *> &Worklist,
479 AliasAnalysis *AA, DominatorTree *DT, LoopInfo *LI,
480 ScalarEvolution *SE, Pass *PP,
481 AssumptionCache *AC) {
482 bool Changed = false;
485 // Check to see that no blocks (other than the header) in this loop have
486 // predecessors that are not in the loop. This is not valid for natural
487 // loops, but can occur if the blocks are unreachable. Since they are
488 // unreachable we can just shamelessly delete those CFG edges!
489 for (Loop::block_iterator BB = L->block_begin(), E = L->block_end();
491 if (*BB == L->getHeader()) continue;
493 SmallPtrSet<BasicBlock*, 4> BadPreds;
494 for (pred_iterator PI = pred_begin(*BB),
495 PE = pred_end(*BB); PI != PE; ++PI) {
501 // Delete each unique out-of-loop (and thus dead) predecessor.
502 for (BasicBlock *P : BadPreds) {
504 DEBUG(dbgs() << "LoopSimplify: Deleting edge from dead predecessor "
505 << P->getName() << "\n");
507 // Inform each successor of each dead pred.
508 for (succ_iterator SI = succ_begin(P), SE = succ_end(P); SI != SE; ++SI)
509 (*SI)->removePredecessor(P);
510 // Zap the dead pred's terminator and replace it with unreachable.
511 TerminatorInst *TI = P->getTerminator();
512 TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
513 P->getTerminator()->eraseFromParent();
514 new UnreachableInst(P->getContext(), P);
519 // If there are exiting blocks with branches on undef, resolve the undef in
520 // the direction which will exit the loop. This will help simplify loop
521 // trip count computations.
522 SmallVector<BasicBlock*, 8> ExitingBlocks;
523 L->getExitingBlocks(ExitingBlocks);
524 for (SmallVectorImpl<BasicBlock *>::iterator I = ExitingBlocks.begin(),
525 E = ExitingBlocks.end(); I != E; ++I)
526 if (BranchInst *BI = dyn_cast<BranchInst>((*I)->getTerminator()))
527 if (BI->isConditional()) {
528 if (UndefValue *Cond = dyn_cast<UndefValue>(BI->getCondition())) {
530 DEBUG(dbgs() << "LoopSimplify: Resolving \"br i1 undef\" to exit in "
531 << (*I)->getName() << "\n");
533 BI->setCondition(ConstantInt::get(Cond->getType(),
534 !L->contains(BI->getSuccessor(0))));
536 // This may make the loop analyzable, force SCEV recomputation.
544 // Does the loop already have a preheader? If so, don't insert one.
545 BasicBlock *Preheader = L->getLoopPreheader();
547 Preheader = InsertPreheaderForLoop(L, PP);
554 // Next, check to make sure that all exit nodes of the loop only have
555 // predecessors that are inside of the loop. This check guarantees that the
556 // loop preheader/header will dominate the exit blocks. If the exit block has
557 // predecessors from outside of the loop, split the edge now.
558 SmallVector<BasicBlock*, 8> ExitBlocks;
559 L->getExitBlocks(ExitBlocks);
561 SmallSetVector<BasicBlock *, 8> ExitBlockSet(ExitBlocks.begin(),
563 for (SmallSetVector<BasicBlock *, 8>::iterator I = ExitBlockSet.begin(),
564 E = ExitBlockSet.end(); I != E; ++I) {
565 BasicBlock *ExitBlock = *I;
566 for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock);
568 // Must be exactly this loop: no subloops, parent loops, or non-loop preds
570 if (!L->contains(*PI)) {
571 if (rewriteLoopExitBlock(L, ExitBlock, AA, DT, LI, PP)) {
579 // If the header has more than two predecessors at this point (from the
580 // preheader and from multiple backedges), we must adjust the loop.
581 BasicBlock *LoopLatch = L->getLoopLatch();
583 // If this is really a nested loop, rip it out into a child loop. Don't do
584 // this for loops with a giant number of backedges, just factor them into a
585 // common backedge instead.
586 if (L->getNumBackEdges() < 8) {
588 separateNestedLoop(L, Preheader, AA, DT, LI, SE, PP, AC)) {
590 // Enqueue the outer loop as it should be processed next in our
591 // depth-first nest walk.
592 Worklist.push_back(OuterL);
594 // This is a big restructuring change, reprocess the whole loop.
596 // GCC doesn't tail recursion eliminate this.
597 // FIXME: It isn't clear we can't rely on LLVM to TRE this.
602 // If we either couldn't, or didn't want to, identify nesting of the loops,
603 // insert a new block that all backedges target, then make it jump to the
605 LoopLatch = insertUniqueBackedgeBlock(L, Preheader, AA, DT, LI);
612 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
614 // Scan over the PHI nodes in the loop header. Since they now have only two
615 // incoming values (the loop is canonicalized), we may have simplified the PHI
616 // down to 'X = phi [X, Y]', which should be replaced with 'Y'.
618 for (BasicBlock::iterator I = L->getHeader()->begin();
619 (PN = dyn_cast<PHINode>(I++)); )
620 if (Value *V = SimplifyInstruction(PN, DL, nullptr, DT, AC)) {
621 if (AA) AA->deleteValue(PN);
622 if (SE) SE->forgetValue(PN);
623 PN->replaceAllUsesWith(V);
624 PN->eraseFromParent();
627 // If this loop has multiple exits and the exits all go to the same
628 // block, attempt to merge the exits. This helps several passes, such
629 // as LoopRotation, which do not support loops with multiple exits.
630 // SimplifyCFG also does this (and this code uses the same utility
631 // function), however this code is loop-aware, where SimplifyCFG is
632 // not. That gives it the advantage of being able to hoist
633 // loop-invariant instructions out of the way to open up more
634 // opportunities, and the disadvantage of having the responsibility
635 // to preserve dominator information.
636 bool UniqueExit = true;
637 if (!ExitBlocks.empty())
638 for (unsigned i = 1, e = ExitBlocks.size(); i != e; ++i)
639 if (ExitBlocks[i] != ExitBlocks[0]) {
644 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
645 BasicBlock *ExitingBlock = ExitingBlocks[i];
646 if (!ExitingBlock->getSinglePredecessor()) continue;
647 BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
648 if (!BI || !BI->isConditional()) continue;
649 CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition());
650 if (!CI || CI->getParent() != ExitingBlock) continue;
652 // Attempt to hoist out all instructions except for the
653 // comparison and the branch.
654 bool AllInvariant = true;
655 bool AnyInvariant = false;
656 for (BasicBlock::iterator I = ExitingBlock->begin(); &*I != BI; ) {
657 Instruction *Inst = I++;
658 // Skip debug info intrinsics.
659 if (isa<DbgInfoIntrinsic>(Inst))
663 if (!L->makeLoopInvariant(Inst, AnyInvariant,
664 Preheader ? Preheader->getTerminator()
666 AllInvariant = false;
672 // The loop disposition of all SCEV expressions that depend on any
673 // hoisted values have also changed.
675 SE->forgetLoopDispositions(L);
677 if (!AllInvariant) continue;
679 // The block has now been cleared of all instructions except for
680 // a comparison and a conditional branch. SimplifyCFG may be able
682 if (!FoldBranchToCommonDest(BI))
685 // Success. The block is now dead, so remove it from the loop,
686 // update the dominator tree and delete it.
687 DEBUG(dbgs() << "LoopSimplify: Eliminating exiting block "
688 << ExitingBlock->getName() << "\n");
690 // Notify ScalarEvolution before deleting this block. Currently assume the
691 // parent loop doesn't change (spliting edges doesn't count). If blocks,
692 // CFG edges, or other values in the parent loop change, then we need call
693 // to forgetLoop() for the parent instead.
697 assert(pred_begin(ExitingBlock) == pred_end(ExitingBlock));
699 LI->removeBlock(ExitingBlock);
701 DomTreeNode *Node = DT->getNode(ExitingBlock);
702 const std::vector<DomTreeNodeBase<BasicBlock> *> &Children =
704 while (!Children.empty()) {
705 DomTreeNode *Child = Children.front();
706 DT->changeImmediateDominator(Child, Node->getIDom());
708 DT->eraseNode(ExitingBlock);
710 BI->getSuccessor(0)->removePredecessor(ExitingBlock);
711 BI->getSuccessor(1)->removePredecessor(ExitingBlock);
712 ExitingBlock->eraseFromParent();
719 bool llvm::simplifyLoop(Loop *L, DominatorTree *DT, LoopInfo *LI, Pass *PP,
720 AliasAnalysis *AA, ScalarEvolution *SE,
721 AssumptionCache *AC) {
722 bool Changed = false;
724 // Worklist maintains our depth-first queue of loops in this nest to process.
725 SmallVector<Loop *, 4> Worklist;
726 Worklist.push_back(L);
728 // Walk the worklist from front to back, pushing newly found sub loops onto
729 // the back. This will let us process loops from back to front in depth-first
730 // order. We can use this simple process because loops form a tree.
731 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
732 Loop *L2 = Worklist[Idx];
733 Worklist.append(L2->begin(), L2->end());
736 while (!Worklist.empty())
737 Changed |= simplifyOneLoop(Worklist.pop_back_val(), Worklist, AA, DT, LI,
744 struct LoopSimplify : public FunctionPass {
745 static char ID; // Pass identification, replacement for typeid
746 LoopSimplify() : FunctionPass(ID) {
747 initializeLoopSimplifyPass(*PassRegistry::getPassRegistry());
750 // AA - If we have an alias analysis object to update, this is it, otherwise
758 bool runOnFunction(Function &F) override;
760 void getAnalysisUsage(AnalysisUsage &AU) const override {
761 AU.addRequired<AssumptionCacheTracker>();
763 // We need loop information to identify the loops...
764 AU.addRequired<DominatorTreeWrapperPass>();
765 AU.addPreserved<DominatorTreeWrapperPass>();
767 AU.addRequired<LoopInfoWrapperPass>();
768 AU.addPreserved<LoopInfoWrapperPass>();
770 AU.addPreserved<AliasAnalysis>();
771 AU.addPreserved<ScalarEvolution>();
772 AU.addPreserved<DependenceAnalysis>();
773 AU.addPreservedID(BreakCriticalEdgesID); // No critical edges added.
776 /// verifyAnalysis() - Verify LoopSimplifyForm's guarantees.
777 void verifyAnalysis() const override;
781 char LoopSimplify::ID = 0;
782 INITIALIZE_PASS_BEGIN(LoopSimplify, "loop-simplify",
783 "Canonicalize natural loops", false, false)
784 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
785 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
786 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
787 INITIALIZE_PASS_END(LoopSimplify, "loop-simplify",
788 "Canonicalize natural loops", false, false)
790 // Publicly exposed interface to pass...
791 char &llvm::LoopSimplifyID = LoopSimplify::ID;
792 Pass *llvm::createLoopSimplifyPass() { return new LoopSimplify(); }
794 /// runOnFunction - Run down all loops in the CFG (recursively, but we could do
795 /// it in any convenient order) inserting preheaders...
797 bool LoopSimplify::runOnFunction(Function &F) {
798 bool Changed = false;
799 AA = getAnalysisIfAvailable<AliasAnalysis>();
800 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
801 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
802 SE = getAnalysisIfAvailable<ScalarEvolution>();
803 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
805 // Simplify each loop nest in the function.
806 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
807 Changed |= simplifyLoop(*I, DT, LI, this, AA, SE, AC);
812 // FIXME: Restore this code when we re-enable verification in verifyAnalysis
815 static void verifyLoop(Loop *L) {
817 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
820 // It used to be possible to just assert L->isLoopSimplifyForm(), however
821 // with the introduction of indirectbr, there are now cases where it's
822 // not possible to transform a loop as necessary. We can at least check
823 // that there is an indirectbr near any time there's trouble.
825 // Indirectbr can interfere with preheader and unique backedge insertion.
826 if (!L->getLoopPreheader() || !L->getLoopLatch()) {
827 bool HasIndBrPred = false;
828 for (pred_iterator PI = pred_begin(L->getHeader()),
829 PE = pred_end(L->getHeader()); PI != PE; ++PI)
830 if (isa<IndirectBrInst>((*PI)->getTerminator())) {
834 assert(HasIndBrPred &&
835 "LoopSimplify has no excuse for missing loop header info!");
839 // Indirectbr can interfere with exit block canonicalization.
840 if (!L->hasDedicatedExits()) {
841 bool HasIndBrExiting = false;
842 SmallVector<BasicBlock*, 8> ExitingBlocks;
843 L->getExitingBlocks(ExitingBlocks);
844 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
845 if (isa<IndirectBrInst>((ExitingBlocks[i])->getTerminator())) {
846 HasIndBrExiting = true;
851 assert(HasIndBrExiting &&
852 "LoopSimplify has no excuse for missing exit block info!");
853 (void)HasIndBrExiting;
858 void LoopSimplify::verifyAnalysis() const {
859 // FIXME: This routine is being called mid-way through the loop pass manager
860 // as loop passes destroy this analysis. That's actually fine, but we have no
861 // way of expressing that here. Once all of the passes that destroy this are
862 // hoisted out of the loop pass manager we can add back verification here.
864 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)