1 //===- LoopRotation.cpp - Loop Rotation 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 file implements Loop Rotation Pass.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Transforms/Scalar.h"
15 #include "llvm/ADT/Statistic.h"
16 #include "llvm/Analysis/AssumptionCache.h"
17 #include "llvm/Analysis/CodeMetrics.h"
18 #include "llvm/Analysis/InstructionSimplify.h"
19 #include "llvm/Analysis/LoopPass.h"
20 #include "llvm/Analysis/ScalarEvolution.h"
21 #include "llvm/Analysis/TargetTransformInfo.h"
22 #include "llvm/Analysis/ValueTracking.h"
23 #include "llvm/IR/CFG.h"
24 #include "llvm/IR/Dominators.h"
25 #include "llvm/IR/Function.h"
26 #include "llvm/IR/IntrinsicInst.h"
27 #include "llvm/Support/CommandLine.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
30 #include "llvm/Transforms/Utils/Local.h"
31 #include "llvm/Transforms/Utils/SSAUpdater.h"
32 #include "llvm/Transforms/Utils/ValueMapper.h"
35 #define DEBUG_TYPE "loop-rotate"
37 static cl::opt<unsigned>
38 DefaultRotationThreshold("rotation-max-header-size", cl::init(16), cl::Hidden,
39 cl::desc("The default maximum header size for automatic loop rotation"));
41 STATISTIC(NumRotated, "Number of loops rotated");
44 class LoopRotate : public LoopPass {
46 static char ID; // Pass ID, replacement for typeid
47 LoopRotate(int SpecifiedMaxHeaderSize = -1) : LoopPass(ID) {
48 initializeLoopRotatePass(*PassRegistry::getPassRegistry());
49 if (SpecifiedMaxHeaderSize == -1)
50 MaxHeaderSize = DefaultRotationThreshold;
52 MaxHeaderSize = unsigned(SpecifiedMaxHeaderSize);
55 // LCSSA form makes instruction renaming easier.
56 void getAnalysisUsage(AnalysisUsage &AU) const override {
57 AU.addRequired<AssumptionCacheTracker>();
58 AU.addPreserved<DominatorTreeWrapperPass>();
59 AU.addRequired<LoopInfoWrapperPass>();
60 AU.addPreserved<LoopInfoWrapperPass>();
61 AU.addRequiredID(LoopSimplifyID);
62 AU.addPreservedID(LoopSimplifyID);
63 AU.addRequiredID(LCSSAID);
64 AU.addPreservedID(LCSSAID);
65 AU.addPreserved<ScalarEvolution>();
66 AU.addRequired<TargetTransformInfo>();
69 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
70 bool simplifyLoopLatch(Loop *L);
71 bool rotateLoop(Loop *L, bool SimplifiedLatch);
74 unsigned MaxHeaderSize;
76 const TargetTransformInfo *TTI;
82 char LoopRotate::ID = 0;
83 INITIALIZE_PASS_BEGIN(LoopRotate, "loop-rotate", "Rotate Loops", false, false)
84 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
85 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
86 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
87 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
88 INITIALIZE_PASS_DEPENDENCY(LCSSA)
89 INITIALIZE_PASS_END(LoopRotate, "loop-rotate", "Rotate Loops", false, false)
91 Pass *llvm::createLoopRotatePass(int MaxHeaderSize) {
92 return new LoopRotate(MaxHeaderSize);
95 /// Rotate Loop L as many times as possible. Return true if
96 /// the loop is rotated at least once.
97 bool LoopRotate::runOnLoop(Loop *L, LPPassManager &LPM) {
98 if (skipOptnoneFunction(L))
101 // Save the loop metadata.
102 MDNode *LoopMD = L->getLoopID();
104 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
105 TTI = &getAnalysis<TargetTransformInfo>();
106 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(
107 *L->getHeader()->getParent());
108 auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>();
109 DT = DTWP ? &DTWP->getDomTree() : nullptr;
111 // Simplify the loop latch before attempting to rotate the header
112 // upward. Rotation may not be needed if the loop tail can be folded into the
114 bool SimplifiedLatch = simplifyLoopLatch(L);
116 // One loop can be rotated multiple times.
117 bool MadeChange = false;
118 while (rotateLoop(L, SimplifiedLatch)) {
120 SimplifiedLatch = false;
123 // Restore the loop metadata.
124 // NB! We presume LoopRotation DOESN'T ADD its own metadata.
125 if ((MadeChange || SimplifiedLatch) && LoopMD)
126 L->setLoopID(LoopMD);
131 /// RewriteUsesOfClonedInstructions - We just cloned the instructions from the
132 /// old header into the preheader. If there were uses of the values produced by
133 /// these instruction that were outside of the loop, we have to insert PHI nodes
134 /// to merge the two values. Do this now.
135 static void RewriteUsesOfClonedInstructions(BasicBlock *OrigHeader,
136 BasicBlock *OrigPreheader,
137 ValueToValueMapTy &ValueMap) {
138 // Remove PHI node entries that are no longer live.
139 BasicBlock::iterator I, E = OrigHeader->end();
140 for (I = OrigHeader->begin(); PHINode *PN = dyn_cast<PHINode>(I); ++I)
141 PN->removeIncomingValue(PN->getBasicBlockIndex(OrigPreheader));
143 // Now fix up users of the instructions in OrigHeader, inserting PHI nodes
146 for (I = OrigHeader->begin(); I != E; ++I) {
147 Value *OrigHeaderVal = I;
149 // If there are no uses of the value (e.g. because it returns void), there
150 // is nothing to rewrite.
151 if (OrigHeaderVal->use_empty())
154 Value *OrigPreHeaderVal = ValueMap[OrigHeaderVal];
156 // The value now exits in two versions: the initial value in the preheader
157 // and the loop "next" value in the original header.
158 SSA.Initialize(OrigHeaderVal->getType(), OrigHeaderVal->getName());
159 SSA.AddAvailableValue(OrigHeader, OrigHeaderVal);
160 SSA.AddAvailableValue(OrigPreheader, OrigPreHeaderVal);
162 // Visit each use of the OrigHeader instruction.
163 for (Value::use_iterator UI = OrigHeaderVal->use_begin(),
164 UE = OrigHeaderVal->use_end(); UI != UE; ) {
165 // Grab the use before incrementing the iterator.
168 // Increment the iterator before removing the use from the list.
171 // SSAUpdater can't handle a non-PHI use in the same block as an
172 // earlier def. We can easily handle those cases manually.
173 Instruction *UserInst = cast<Instruction>(U.getUser());
174 if (!isa<PHINode>(UserInst)) {
175 BasicBlock *UserBB = UserInst->getParent();
177 // The original users in the OrigHeader are already using the
178 // original definitions.
179 if (UserBB == OrigHeader)
182 // Users in the OrigPreHeader need to use the value to which the
183 // original definitions are mapped.
184 if (UserBB == OrigPreheader) {
185 U = OrigPreHeaderVal;
190 // Anything else can be handled by SSAUpdater.
196 /// Determine whether the instructions in this range may be safely and cheaply
197 /// speculated. This is not an important enough situation to develop complex
198 /// heuristics. We handle a single arithmetic instruction along with any type
200 static bool shouldSpeculateInstrs(BasicBlock::iterator Begin,
201 BasicBlock::iterator End, Loop *L) {
202 bool seenIncrement = false;
203 bool MultiExitLoop = false;
205 if (!L->getExitingBlock())
206 MultiExitLoop = true;
208 for (BasicBlock::iterator I = Begin; I != End; ++I) {
210 if (!isSafeToSpeculativelyExecute(I))
213 if (isa<DbgInfoIntrinsic>(I))
216 switch (I->getOpcode()) {
219 case Instruction::GetElementPtr:
220 // GEPs are cheap if all indices are constant.
221 if (!cast<GEPOperator>(I)->hasAllConstantIndices())
223 // fall-thru to increment case
224 case Instruction::Add:
225 case Instruction::Sub:
226 case Instruction::And:
227 case Instruction::Or:
228 case Instruction::Xor:
229 case Instruction::Shl:
230 case Instruction::LShr:
231 case Instruction::AShr: {
232 Value *IVOpnd = nullptr;
233 if (isa<ConstantInt>(I->getOperand(0)))
234 IVOpnd = I->getOperand(1);
236 if (isa<ConstantInt>(I->getOperand(1))) {
240 IVOpnd = I->getOperand(0);
243 // If increment operand is used outside of the loop, this speculation
244 // could cause extra live range interference.
245 if (MultiExitLoop && IVOpnd) {
246 for (User *UseI : IVOpnd->users()) {
247 auto *UserInst = cast<Instruction>(UseI);
248 if (!L->contains(UserInst))
255 seenIncrement = true;
258 case Instruction::Trunc:
259 case Instruction::ZExt:
260 case Instruction::SExt:
261 // ignore type conversions
268 /// Fold the loop tail into the loop exit by speculating the loop tail
269 /// instructions. Typically, this is a single post-increment. In the case of a
270 /// simple 2-block loop, hoisting the increment can be much better than
271 /// duplicating the entire loop header. In the case of loops with early exits,
272 /// rotation will not work anyway, but simplifyLoopLatch will put the loop in
273 /// canonical form so downstream passes can handle it.
275 /// I don't believe this invalidates SCEV.
276 bool LoopRotate::simplifyLoopLatch(Loop *L) {
277 BasicBlock *Latch = L->getLoopLatch();
278 if (!Latch || Latch->hasAddressTaken())
281 BranchInst *Jmp = dyn_cast<BranchInst>(Latch->getTerminator());
282 if (!Jmp || !Jmp->isUnconditional())
285 BasicBlock *LastExit = Latch->getSinglePredecessor();
286 if (!LastExit || !L->isLoopExiting(LastExit))
289 BranchInst *BI = dyn_cast<BranchInst>(LastExit->getTerminator());
293 if (!shouldSpeculateInstrs(Latch->begin(), Jmp, L))
296 DEBUG(dbgs() << "Folding loop latch " << Latch->getName() << " into "
297 << LastExit->getName() << "\n");
299 // Hoist the instructions from Latch into LastExit.
300 LastExit->getInstList().splice(BI, Latch->getInstList(), Latch->begin(), Jmp);
302 unsigned FallThruPath = BI->getSuccessor(0) == Latch ? 0 : 1;
303 BasicBlock *Header = Jmp->getSuccessor(0);
304 assert(Header == L->getHeader() && "expected a backward branch");
306 // Remove Latch from the CFG so that LastExit becomes the new Latch.
307 BI->setSuccessor(FallThruPath, Header);
308 Latch->replaceSuccessorsPhiUsesWith(LastExit);
309 Jmp->eraseFromParent();
311 // Nuke the Latch block.
312 assert(Latch->empty() && "unable to evacuate Latch");
313 LI->removeBlock(Latch);
315 DT->eraseNode(Latch);
316 Latch->eraseFromParent();
320 /// Rotate loop LP. Return true if the loop is rotated.
322 /// \param SimplifiedLatch is true if the latch was just folded into the final
323 /// loop exit. In this case we may want to rotate even though the new latch is
324 /// now an exiting branch. This rotation would have happened had the latch not
325 /// been simplified. However, if SimplifiedLatch is false, then we avoid
326 /// rotating loops in which the latch exits to avoid excessive or endless
327 /// rotation. LoopRotate should be repeatable and converge to a canonical
328 /// form. This property is satisfied because simplifying the loop latch can only
329 /// happen once across multiple invocations of the LoopRotate pass.
330 bool LoopRotate::rotateLoop(Loop *L, bool SimplifiedLatch) {
331 // If the loop has only one block then there is not much to rotate.
332 if (L->getBlocks().size() == 1)
335 BasicBlock *OrigHeader = L->getHeader();
336 BasicBlock *OrigLatch = L->getLoopLatch();
338 BranchInst *BI = dyn_cast<BranchInst>(OrigHeader->getTerminator());
339 if (!BI || BI->isUnconditional())
342 // If the loop header is not one of the loop exiting blocks then
343 // either this loop is already rotated or it is not
344 // suitable for loop rotation transformations.
345 if (!L->isLoopExiting(OrigHeader))
348 // If the loop latch already contains a branch that leaves the loop then the
349 // loop is already rotated.
353 // Rotate if either the loop latch does *not* exit the loop, or if the loop
354 // latch was just simplified.
355 if (L->isLoopExiting(OrigLatch) && !SimplifiedLatch)
358 // Check size of original header and reject loop if it is very big or we can't
359 // duplicate blocks inside it.
361 SmallPtrSet<const Value *, 32> EphValues;
362 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
365 Metrics.analyzeBasicBlock(OrigHeader, *TTI, EphValues);
366 if (Metrics.notDuplicatable) {
367 DEBUG(dbgs() << "LoopRotation: NOT rotating - contains non-duplicatable"
368 << " instructions: "; L->dump());
371 if (Metrics.NumInsts > MaxHeaderSize)
375 // Now, this loop is suitable for rotation.
376 BasicBlock *OrigPreheader = L->getLoopPreheader();
378 // If the loop could not be converted to canonical form, it must have an
379 // indirectbr in it, just give up.
383 // Anything ScalarEvolution may know about this loop or the PHI nodes
384 // in its header will soon be invalidated.
385 if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>())
388 DEBUG(dbgs() << "LoopRotation: rotating "; L->dump());
390 // Find new Loop header. NewHeader is a Header's one and only successor
391 // that is inside loop. Header's other successor is outside the
392 // loop. Otherwise loop is not suitable for rotation.
393 BasicBlock *Exit = BI->getSuccessor(0);
394 BasicBlock *NewHeader = BI->getSuccessor(1);
395 if (L->contains(Exit))
396 std::swap(Exit, NewHeader);
397 assert(NewHeader && "Unable to determine new loop header");
398 assert(L->contains(NewHeader) && !L->contains(Exit) &&
399 "Unable to determine loop header and exit blocks");
401 // This code assumes that the new header has exactly one predecessor.
402 // Remove any single-entry PHI nodes in it.
403 assert(NewHeader->getSinglePredecessor() &&
404 "New header doesn't have one pred!");
405 FoldSingleEntryPHINodes(NewHeader);
407 // Begin by walking OrigHeader and populating ValueMap with an entry for
409 BasicBlock::iterator I = OrigHeader->begin(), E = OrigHeader->end();
410 ValueToValueMapTy ValueMap;
412 // For PHI nodes, the value available in OldPreHeader is just the
413 // incoming value from OldPreHeader.
414 for (; PHINode *PN = dyn_cast<PHINode>(I); ++I)
415 ValueMap[PN] = PN->getIncomingValueForBlock(OrigPreheader);
417 // For the rest of the instructions, either hoist to the OrigPreheader if
418 // possible or create a clone in the OldPreHeader if not.
419 TerminatorInst *LoopEntryBranch = OrigPreheader->getTerminator();
421 Instruction *Inst = I++;
423 // If the instruction's operands are invariant and it doesn't read or write
424 // memory, then it is safe to hoist. Doing this doesn't change the order of
425 // execution in the preheader, but does prevent the instruction from
426 // executing in each iteration of the loop. This means it is safe to hoist
427 // something that might trap, but isn't safe to hoist something that reads
428 // memory (without proving that the loop doesn't write).
429 if (L->hasLoopInvariantOperands(Inst) &&
430 !Inst->mayReadFromMemory() && !Inst->mayWriteToMemory() &&
431 !isa<TerminatorInst>(Inst) && !isa<DbgInfoIntrinsic>(Inst) &&
432 !isa<AllocaInst>(Inst)) {
433 Inst->moveBefore(LoopEntryBranch);
437 // Otherwise, create a duplicate of the instruction.
438 Instruction *C = Inst->clone();
440 // Eagerly remap the operands of the instruction.
441 RemapInstruction(C, ValueMap,
442 RF_NoModuleLevelChanges|RF_IgnoreMissingEntries);
444 // With the operands remapped, see if the instruction constant folds or is
445 // otherwise simplifyable. This commonly occurs because the entry from PHI
446 // nodes allows icmps and other instructions to fold.
447 // FIXME: Provide DL, TLI, DT, AC to SimplifyInstruction.
448 Value *V = SimplifyInstruction(C);
449 if (V && LI->replacementPreservesLCSSAForm(C, V)) {
450 // If so, then delete the temporary instruction and stick the folded value
455 // Otherwise, stick the new instruction into the new block!
456 C->setName(Inst->getName());
457 C->insertBefore(LoopEntryBranch);
462 // Along with all the other instructions, we just cloned OrigHeader's
463 // terminator into OrigPreHeader. Fix up the PHI nodes in each of OrigHeader's
464 // successors by duplicating their incoming values for OrigHeader.
465 TerminatorInst *TI = OrigHeader->getTerminator();
466 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
467 for (BasicBlock::iterator BI = TI->getSuccessor(i)->begin();
468 PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
469 PN->addIncoming(PN->getIncomingValueForBlock(OrigHeader), OrigPreheader);
471 // Now that OrigPreHeader has a clone of OrigHeader's terminator, remove
472 // OrigPreHeader's old terminator (the original branch into the loop), and
473 // remove the corresponding incoming values from the PHI nodes in OrigHeader.
474 LoopEntryBranch->eraseFromParent();
476 // If there were any uses of instructions in the duplicated block outside the
477 // loop, update them, inserting PHI nodes as required
478 RewriteUsesOfClonedInstructions(OrigHeader, OrigPreheader, ValueMap);
480 // NewHeader is now the header of the loop.
481 L->moveToHeader(NewHeader);
482 assert(L->getHeader() == NewHeader && "Latch block is our new header");
485 // At this point, we've finished our major CFG changes. As part of cloning
486 // the loop into the preheader we've simplified instructions and the
487 // duplicated conditional branch may now be branching on a constant. If it is
488 // branching on a constant and if that constant means that we enter the loop,
489 // then we fold away the cond branch to an uncond branch. This simplifies the
490 // loop in cases important for nested loops, and it also means we don't have
491 // to split as many edges.
492 BranchInst *PHBI = cast<BranchInst>(OrigPreheader->getTerminator());
493 assert(PHBI->isConditional() && "Should be clone of BI condbr!");
494 if (!isa<ConstantInt>(PHBI->getCondition()) ||
495 PHBI->getSuccessor(cast<ConstantInt>(PHBI->getCondition())->isZero())
497 // The conditional branch can't be folded, handle the general case.
498 // Update DominatorTree to reflect the CFG change we just made. Then split
499 // edges as necessary to preserve LoopSimplify form.
501 // Everything that was dominated by the old loop header is now dominated
502 // by the original loop preheader. Conceptually the header was merged
503 // into the preheader, even though we reuse the actual block as a new
505 DomTreeNode *OrigHeaderNode = DT->getNode(OrigHeader);
506 SmallVector<DomTreeNode *, 8> HeaderChildren(OrigHeaderNode->begin(),
507 OrigHeaderNode->end());
508 DomTreeNode *OrigPreheaderNode = DT->getNode(OrigPreheader);
509 for (unsigned I = 0, E = HeaderChildren.size(); I != E; ++I)
510 DT->changeImmediateDominator(HeaderChildren[I], OrigPreheaderNode);
512 assert(DT->getNode(Exit)->getIDom() == OrigPreheaderNode);
513 assert(DT->getNode(NewHeader)->getIDom() == OrigPreheaderNode);
515 // Update OrigHeader to be dominated by the new header block.
516 DT->changeImmediateDominator(OrigHeader, OrigLatch);
519 // Right now OrigPreHeader has two successors, NewHeader and ExitBlock, and
520 // thus is not a preheader anymore.
521 // Split the edge to form a real preheader.
522 BasicBlock *NewPH = SplitCriticalEdge(OrigPreheader, NewHeader, this);
523 NewPH->setName(NewHeader->getName() + ".lr.ph");
525 // Preserve canonical loop form, which means that 'Exit' should have only
526 // one predecessor. Note that Exit could be an exit block for multiple
527 // nested loops, causing both of the edges to now be critical and need to
529 SmallVector<BasicBlock *, 4> ExitPreds(pred_begin(Exit), pred_end(Exit));
530 bool SplitLatchEdge = false;
531 for (SmallVectorImpl<BasicBlock *>::iterator PI = ExitPreds.begin(),
532 PE = ExitPreds.end();
534 // We only need to split loop exit edges.
535 Loop *PredLoop = LI->getLoopFor(*PI);
536 if (!PredLoop || PredLoop->contains(Exit))
538 SplitLatchEdge |= L->getLoopLatch() == *PI;
539 BasicBlock *ExitSplit = SplitCriticalEdge(*PI, Exit, this);
540 ExitSplit->moveBefore(Exit);
542 assert(SplitLatchEdge &&
543 "Despite splitting all preds, failed to split latch exit?");
545 // We can fold the conditional branch in the preheader, this makes things
546 // simpler. The first step is to remove the extra edge to the Exit block.
547 Exit->removePredecessor(OrigPreheader, true /*preserve LCSSA*/);
548 BranchInst *NewBI = BranchInst::Create(NewHeader, PHBI);
549 NewBI->setDebugLoc(PHBI->getDebugLoc());
550 PHBI->eraseFromParent();
552 // With our CFG finalized, update DomTree if it is available.
554 // Update OrigHeader to be dominated by the new header block.
555 DT->changeImmediateDominator(NewHeader, OrigPreheader);
556 DT->changeImmediateDominator(OrigHeader, OrigLatch);
558 // Brute force incremental dominator tree update. Call
559 // findNearestCommonDominator on all CFG predecessors of each child of the
561 DomTreeNode *OrigHeaderNode = DT->getNode(OrigHeader);
562 SmallVector<DomTreeNode *, 8> HeaderChildren(OrigHeaderNode->begin(),
563 OrigHeaderNode->end());
567 for (unsigned I = 0, E = HeaderChildren.size(); I != E; ++I) {
568 DomTreeNode *Node = HeaderChildren[I];
569 BasicBlock *BB = Node->getBlock();
571 pred_iterator PI = pred_begin(BB);
572 BasicBlock *NearestDom = *PI;
573 for (pred_iterator PE = pred_end(BB); PI != PE; ++PI)
574 NearestDom = DT->findNearestCommonDominator(NearestDom, *PI);
576 // Remember if this changes the DomTree.
577 if (Node->getIDom()->getBlock() != NearestDom) {
578 DT->changeImmediateDominator(BB, NearestDom);
583 // If the dominator changed, this may have an effect on other
584 // predecessors, continue until we reach a fixpoint.
589 assert(L->getLoopPreheader() && "Invalid loop preheader after loop rotation");
590 assert(L->getLoopLatch() && "Invalid loop latch after loop rotation");
592 // Now that the CFG and DomTree are in a consistent state again, try to merge
593 // the OrigHeader block into OrigLatch. This will succeed if they are
594 // connected by an unconditional branch. This is just a cleanup so the
595 // emitted code isn't too gross in this common case.
596 MergeBlockIntoPredecessor(OrigHeader, this);
598 DEBUG(dbgs() << "LoopRotation: into "; L->dump());