1 //===----------------- LoopRotationUtils.cpp -----------------------------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file provides utilities to convert a loop into a loop with bottom test.
11 //===----------------------------------------------------------------------===//
13 #include "llvm/Transforms/Utils/LoopRotationUtils.h"
14 #include "llvm/ADT/Statistic.h"
15 #include "llvm/Analysis/AssumptionCache.h"
16 #include "llvm/Analysis/BasicAliasAnalysis.h"
17 #include "llvm/Analysis/CodeMetrics.h"
18 #include "llvm/Analysis/DomTreeUpdater.h"
19 #include "llvm/Analysis/GlobalsModRef.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/LoopPass.h"
22 #include "llvm/Analysis/MemorySSA.h"
23 #include "llvm/Analysis/MemorySSAUpdater.h"
24 #include "llvm/Analysis/ScalarEvolution.h"
25 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
26 #include "llvm/Analysis/TargetTransformInfo.h"
27 #include "llvm/Analysis/ValueTracking.h"
28 #include "llvm/IR/CFG.h"
29 #include "llvm/IR/DebugInfoMetadata.h"
30 #include "llvm/IR/Dominators.h"
31 #include "llvm/IR/Function.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/IR/Module.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/raw_ostream.h"
37 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
38 #include "llvm/Transforms/Utils/Cloning.h"
39 #include "llvm/Transforms/Utils/Local.h"
40 #include "llvm/Transforms/Utils/LoopUtils.h"
41 #include "llvm/Transforms/Utils/SSAUpdater.h"
42 #include "llvm/Transforms/Utils/ValueMapper.h"
45 #define DEBUG_TYPE "loop-rotate"
47 STATISTIC(NumNotRotatedDueToHeaderSize,
48 "Number of loops not rotated due to the header size");
49 STATISTIC(NumRotated, "Number of loops rotated");
52 MultiRotate("loop-rotate-multi", cl::init(false), cl::Hidden,
53 cl::desc("Allow loop rotation multiple times in order to reach "
54 "a better latch exit"));
57 /// A simple loop rotation transformation.
59 const unsigned MaxHeaderSize;
61 const TargetTransformInfo *TTI;
65 MemorySSAUpdater *MSSAU;
66 const SimplifyQuery &SQ;
72 LoopRotate(unsigned MaxHeaderSize, LoopInfo *LI,
73 const TargetTransformInfo *TTI, AssumptionCache *AC,
74 DominatorTree *DT, ScalarEvolution *SE, MemorySSAUpdater *MSSAU,
75 const SimplifyQuery &SQ, bool RotationOnly, bool IsUtilMode,
77 : MaxHeaderSize(MaxHeaderSize), LI(LI), TTI(TTI), AC(AC), DT(DT), SE(SE),
78 MSSAU(MSSAU), SQ(SQ), RotationOnly(RotationOnly),
79 IsUtilMode(IsUtilMode), PrepareForLTO(PrepareForLTO) {}
80 bool processLoop(Loop *L);
83 bool rotateLoop(Loop *L, bool SimplifiedLatch);
84 bool simplifyLoopLatch(Loop *L);
86 } // end anonymous namespace
88 /// Insert (K, V) pair into the ValueToValueMap, and verify the key did not
89 /// previously exist in the map, and the value was inserted.
90 static void InsertNewValueIntoMap(ValueToValueMapTy &VM, Value *K, Value *V) {
91 bool Inserted = VM.insert({K, V}).second;
95 /// RewriteUsesOfClonedInstructions - We just cloned the instructions from the
96 /// old header into the preheader. If there were uses of the values produced by
97 /// these instruction that were outside of the loop, we have to insert PHI nodes
98 /// to merge the two values. Do this now.
99 static void RewriteUsesOfClonedInstructions(BasicBlock *OrigHeader,
100 BasicBlock *OrigPreheader,
101 ValueToValueMapTy &ValueMap,
102 SmallVectorImpl<PHINode*> *InsertedPHIs) {
103 // Remove PHI node entries that are no longer live.
104 BasicBlock::iterator I, E = OrigHeader->end();
105 for (I = OrigHeader->begin(); PHINode *PN = dyn_cast<PHINode>(I); ++I)
106 PN->removeIncomingValue(PN->getBasicBlockIndex(OrigPreheader));
108 // Now fix up users of the instructions in OrigHeader, inserting PHI nodes
110 SSAUpdater SSA(InsertedPHIs);
111 for (I = OrigHeader->begin(); I != E; ++I) {
112 Value *OrigHeaderVal = &*I;
114 // If there are no uses of the value (e.g. because it returns void), there
115 // is nothing to rewrite.
116 if (OrigHeaderVal->use_empty())
119 Value *OrigPreHeaderVal = ValueMap.lookup(OrigHeaderVal);
121 // The value now exits in two versions: the initial value in the preheader
122 // and the loop "next" value in the original header.
123 SSA.Initialize(OrigHeaderVal->getType(), OrigHeaderVal->getName());
124 SSA.AddAvailableValue(OrigHeader, OrigHeaderVal);
125 SSA.AddAvailableValue(OrigPreheader, OrigPreHeaderVal);
127 // Visit each use of the OrigHeader instruction.
128 for (Value::use_iterator UI = OrigHeaderVal->use_begin(),
129 UE = OrigHeaderVal->use_end();
131 // Grab the use before incrementing the iterator.
134 // Increment the iterator before removing the use from the list.
137 // SSAUpdater can't handle a non-PHI use in the same block as an
138 // earlier def. We can easily handle those cases manually.
139 Instruction *UserInst = cast<Instruction>(U.getUser());
140 if (!isa<PHINode>(UserInst)) {
141 BasicBlock *UserBB = UserInst->getParent();
143 // The original users in the OrigHeader are already using the
144 // original definitions.
145 if (UserBB == OrigHeader)
148 // Users in the OrigPreHeader need to use the value to which the
149 // original definitions are mapped.
150 if (UserBB == OrigPreheader) {
151 U = OrigPreHeaderVal;
156 // Anything else can be handled by SSAUpdater.
160 // Replace MetadataAsValue(ValueAsMetadata(OrigHeaderVal)) uses in debug
162 SmallVector<DbgValueInst *, 1> DbgValues;
163 llvm::findDbgValues(DbgValues, OrigHeaderVal);
164 for (auto &DbgValue : DbgValues) {
165 // The original users in the OrigHeader are already using the original
167 BasicBlock *UserBB = DbgValue->getParent();
168 if (UserBB == OrigHeader)
171 // Users in the OrigPreHeader need to use the value to which the
172 // original definitions are mapped and anything else can be handled by
173 // the SSAUpdater. To avoid adding PHINodes, check if the value is
174 // available in UserBB, if not substitute undef.
176 if (UserBB == OrigPreheader)
177 NewVal = OrigPreHeaderVal;
178 else if (SSA.HasValueForBlock(UserBB))
179 NewVal = SSA.GetValueInMiddleOfBlock(UserBB);
181 NewVal = UndefValue::get(OrigHeaderVal->getType());
182 DbgValue->setOperand(0,
183 MetadataAsValue::get(OrigHeaderVal->getContext(),
184 ValueAsMetadata::get(NewVal)));
189 // Assuming both header and latch are exiting, look for a phi which is only
190 // used outside the loop (via a LCSSA phi) in the exit from the header.
191 // This means that rotating the loop can remove the phi.
192 static bool profitableToRotateLoopExitingLatch(Loop *L) {
193 BasicBlock *Header = L->getHeader();
194 BranchInst *BI = dyn_cast<BranchInst>(Header->getTerminator());
195 assert(BI && BI->isConditional() && "need header with conditional exit");
196 BasicBlock *HeaderExit = BI->getSuccessor(0);
197 if (L->contains(HeaderExit))
198 HeaderExit = BI->getSuccessor(1);
200 for (auto &Phi : Header->phis()) {
201 // Look for uses of this phi in the loop/via exits other than the header.
202 if (llvm::any_of(Phi.users(), [HeaderExit](const User *U) {
203 return cast<Instruction>(U)->getParent() != HeaderExit;
211 // Check that latch exit is deoptimizing (which means - very unlikely to happen)
212 // and there is another exit from the loop which is non-deoptimizing.
213 // If we rotate latch to that exit our loop has a better chance of being fully
216 // It can give false positives in some rare cases.
217 static bool canRotateDeoptimizingLatchExit(Loop *L) {
218 BasicBlock *Latch = L->getLoopLatch();
219 assert(Latch && "need latch");
220 BranchInst *BI = dyn_cast<BranchInst>(Latch->getTerminator());
221 // Need normal exiting latch.
222 if (!BI || !BI->isConditional())
225 BasicBlock *Exit = BI->getSuccessor(1);
226 if (L->contains(Exit))
227 Exit = BI->getSuccessor(0);
229 // Latch exit is non-deoptimizing, no need to rotate.
230 if (!Exit->getPostdominatingDeoptimizeCall())
233 SmallVector<BasicBlock *, 4> Exits;
234 L->getUniqueExitBlocks(Exits);
235 if (!Exits.empty()) {
236 // There is at least one non-deoptimizing exit.
238 // Note, that BasicBlock::getPostdominatingDeoptimizeCall is not exact,
239 // as it can conservatively return false for deoptimizing exits with
240 // complex enough control flow down to deoptimize call.
242 // That means here we can report success for a case where
243 // all exits are deoptimizing but one of them has complex enough
244 // control flow (e.g. with loops).
246 // That should be a very rare case and false positives for this function
247 // have compile-time effect only.
248 return any_of(Exits, [](const BasicBlock *BB) {
249 return !BB->getPostdominatingDeoptimizeCall();
255 /// Rotate loop LP. Return true if the loop is rotated.
257 /// \param SimplifiedLatch is true if the latch was just folded into the final
258 /// loop exit. In this case we may want to rotate even though the new latch is
259 /// now an exiting branch. This rotation would have happened had the latch not
260 /// been simplified. However, if SimplifiedLatch is false, then we avoid
261 /// rotating loops in which the latch exits to avoid excessive or endless
262 /// rotation. LoopRotate should be repeatable and converge to a canonical
263 /// form. This property is satisfied because simplifying the loop latch can only
264 /// happen once across multiple invocations of the LoopRotate pass.
266 /// If -loop-rotate-multi is enabled we can do multiple rotations in one go
267 /// so to reach a suitable (non-deoptimizing) exit.
268 bool LoopRotate::rotateLoop(Loop *L, bool SimplifiedLatch) {
269 // If the loop has only one block then there is not much to rotate.
270 if (L->getBlocks().size() == 1)
273 bool Rotated = false;
275 BasicBlock *OrigHeader = L->getHeader();
276 BasicBlock *OrigLatch = L->getLoopLatch();
278 BranchInst *BI = dyn_cast<BranchInst>(OrigHeader->getTerminator());
279 if (!BI || BI->isUnconditional())
282 // If the loop header is not one of the loop exiting blocks then
283 // either this loop is already rotated or it is not
284 // suitable for loop rotation transformations.
285 if (!L->isLoopExiting(OrigHeader))
288 // If the loop latch already contains a branch that leaves the loop then the
289 // loop is already rotated.
293 // Rotate if either the loop latch does *not* exit the loop, or if the loop
294 // latch was just simplified. Or if we think it will be profitable.
295 if (L->isLoopExiting(OrigLatch) && !SimplifiedLatch && IsUtilMode == false &&
296 !profitableToRotateLoopExitingLatch(L) &&
297 !canRotateDeoptimizingLatchExit(L))
300 // Check size of original header and reject loop if it is very big or we can't
301 // duplicate blocks inside it.
303 SmallPtrSet<const Value *, 32> EphValues;
304 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
307 Metrics.analyzeBasicBlock(OrigHeader, *TTI, EphValues, PrepareForLTO);
308 if (Metrics.notDuplicatable) {
310 dbgs() << "LoopRotation: NOT rotating - contains non-duplicatable"
311 << " instructions: ";
315 if (Metrics.convergent) {
316 LLVM_DEBUG(dbgs() << "LoopRotation: NOT rotating - contains convergent "
321 if (Metrics.NumInsts > MaxHeaderSize) {
322 LLVM_DEBUG(dbgs() << "LoopRotation: NOT rotating - contains "
324 << " instructions, which is more than the threshold ("
325 << MaxHeaderSize << " instructions): ";
327 ++NumNotRotatedDueToHeaderSize;
331 // When preparing for LTO, avoid rotating loops with calls that could be
332 // inlined during the LTO stage.
333 if (PrepareForLTO && Metrics.NumInlineCandidates > 0)
337 // Now, this loop is suitable for rotation.
338 BasicBlock *OrigPreheader = L->getLoopPreheader();
340 // If the loop could not be converted to canonical form, it must have an
341 // indirectbr in it, just give up.
342 if (!OrigPreheader || !L->hasDedicatedExits())
345 // Anything ScalarEvolution may know about this loop or the PHI nodes
346 // in its header will soon be invalidated. We should also invalidate
347 // all outer loops because insertion and deletion of blocks that happens
348 // during the rotation may violate invariants related to backedge taken
351 SE->forgetTopmostLoop(L);
353 LLVM_DEBUG(dbgs() << "LoopRotation: rotating "; L->dump());
354 if (MSSAU && VerifyMemorySSA)
355 MSSAU->getMemorySSA()->verifyMemorySSA();
357 // Find new Loop header. NewHeader is a Header's one and only successor
358 // that is inside loop. Header's other successor is outside the
359 // loop. Otherwise loop is not suitable for rotation.
360 BasicBlock *Exit = BI->getSuccessor(0);
361 BasicBlock *NewHeader = BI->getSuccessor(1);
362 if (L->contains(Exit))
363 std::swap(Exit, NewHeader);
364 assert(NewHeader && "Unable to determine new loop header");
365 assert(L->contains(NewHeader) && !L->contains(Exit) &&
366 "Unable to determine loop header and exit blocks");
368 // This code assumes that the new header has exactly one predecessor.
369 // Remove any single-entry PHI nodes in it.
370 assert(NewHeader->getSinglePredecessor() &&
371 "New header doesn't have one pred!");
372 FoldSingleEntryPHINodes(NewHeader);
374 // Begin by walking OrigHeader and populating ValueMap with an entry for
376 BasicBlock::iterator I = OrigHeader->begin(), E = OrigHeader->end();
377 ValueToValueMapTy ValueMap, ValueMapMSSA;
379 // For PHI nodes, the value available in OldPreHeader is just the
380 // incoming value from OldPreHeader.
381 for (; PHINode *PN = dyn_cast<PHINode>(I); ++I)
382 InsertNewValueIntoMap(ValueMap, PN,
383 PN->getIncomingValueForBlock(OrigPreheader));
385 // For the rest of the instructions, either hoist to the OrigPreheader if
386 // possible or create a clone in the OldPreHeader if not.
387 Instruction *LoopEntryBranch = OrigPreheader->getTerminator();
389 // Record all debug intrinsics preceding LoopEntryBranch to avoid duplication.
390 using DbgIntrinsicHash =
391 std::pair<std::pair<Value *, DILocalVariable *>, DIExpression *>;
392 auto makeHash = [](DbgVariableIntrinsic *D) -> DbgIntrinsicHash {
393 return {{D->getVariableLocation(), D->getVariable()}, D->getExpression()};
395 SmallDenseSet<DbgIntrinsicHash, 8> DbgIntrinsics;
396 for (auto I = std::next(OrigPreheader->rbegin()), E = OrigPreheader->rend();
398 if (auto *DII = dyn_cast<DbgVariableIntrinsic>(&*I))
399 DbgIntrinsics.insert(makeHash(DII));
404 // Remember the local noalias scope declarations in the header. After the
405 // rotation, they must be duplicated and the scope must be cloned. This
406 // avoids unwanted interaction across iterations.
407 SmallVector<NoAliasScopeDeclInst *, 6> NoAliasDeclInstructions;
408 for (Instruction &I : *OrigHeader)
409 if (auto *Decl = dyn_cast<NoAliasScopeDeclInst>(&I))
410 NoAliasDeclInstructions.push_back(Decl);
413 Instruction *Inst = &*I++;
415 // If the instruction's operands are invariant and it doesn't read or write
416 // memory, then it is safe to hoist. Doing this doesn't change the order of
417 // execution in the preheader, but does prevent the instruction from
418 // executing in each iteration of the loop. This means it is safe to hoist
419 // something that might trap, but isn't safe to hoist something that reads
420 // memory (without proving that the loop doesn't write).
421 if (L->hasLoopInvariantOperands(Inst) && !Inst->mayReadFromMemory() &&
422 !Inst->mayWriteToMemory() && !Inst->isTerminator() &&
423 !isa<DbgInfoIntrinsic>(Inst) && !isa<AllocaInst>(Inst)) {
424 Inst->moveBefore(LoopEntryBranch);
428 // Otherwise, create a duplicate of the instruction.
429 Instruction *C = Inst->clone();
431 // Eagerly remap the operands of the instruction.
432 RemapInstruction(C, ValueMap,
433 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
435 // Avoid inserting the same intrinsic twice.
436 if (auto *DII = dyn_cast<DbgVariableIntrinsic>(C))
437 if (DbgIntrinsics.count(makeHash(DII))) {
442 // With the operands remapped, see if the instruction constant folds or is
443 // otherwise simplifyable. This commonly occurs because the entry from PHI
444 // nodes allows icmps and other instructions to fold.
445 Value *V = SimplifyInstruction(C, SQ);
446 if (V && LI->replacementPreservesLCSSAForm(C, V)) {
447 // If so, then delete the temporary instruction and stick the folded value
449 InsertNewValueIntoMap(ValueMap, Inst, V);
450 if (!C->mayHaveSideEffects()) {
455 InsertNewValueIntoMap(ValueMap, Inst, C);
458 // Otherwise, stick the new instruction into the new block!
459 C->setName(Inst->getName());
460 C->insertBefore(LoopEntryBranch);
462 if (auto *II = dyn_cast<IntrinsicInst>(C))
463 if (II->getIntrinsicID() == Intrinsic::assume)
464 AC->registerAssumption(II);
465 // MemorySSA cares whether the cloned instruction was inserted or not, and
466 // not whether it can be remapped to a simplified value.
468 InsertNewValueIntoMap(ValueMapMSSA, Inst, C);
472 if (!NoAliasDeclInstructions.empty()) {
473 // There are noalias scope declarations:
475 // Original: OrigPre { OrigHeader NewHeader ... Latch }
476 // after: (OrigPre+OrigHeader') { NewHeader ... Latch OrigHeader }
478 // with D: llvm.experimental.noalias.scope.decl,
479 // U: !noalias or !alias.scope depending on D
480 // ... { D U1 U2 } can transform into:
481 // (0) : ... { D U1 U2 } // no relevant rotation for this part
482 // (1) : ... D' { U1 U2 D } // D is part of OrigHeader
483 // (2) : ... D' U1' { U2 D U1 } // D, U1 are part of OrigHeader
485 // We now want to transform:
486 // (1) -> : ... D' { D U1 U2 D'' }
487 // (2) -> : ... D' U1' { D U2 D'' U1'' }
488 // D: original llvm.experimental.noalias.scope.decl
489 // D', U1': duplicate with replaced scopes
490 // D'', U1'': different duplicate with replaced scopes
491 // This ensures a safe fallback to 'may_alias' introduced by the rotate,
492 // as U1'' and U1' scopes will not be compatible wrt to the local restrict
494 // Clone the llvm.experimental.noalias.decl again for the NewHeader.
495 Instruction *NewHeaderInsertionPoint = &(*NewHeader->getFirstNonPHI());
496 for (NoAliasScopeDeclInst *NAD : NoAliasDeclInstructions) {
497 LLVM_DEBUG(dbgs() << " Cloning llvm.experimental.noalias.scope.decl:"
499 Instruction *NewNAD = NAD->clone();
500 NewNAD->insertBefore(NewHeaderInsertionPoint);
503 // Scopes must now be duplicated, once for OrigHeader and once for
506 auto &Context = NewHeader->getContext();
508 SmallVector<MDNode *, 8> NoAliasDeclScopes;
509 for (NoAliasScopeDeclInst *NAD : NoAliasDeclInstructions)
510 NoAliasDeclScopes.push_back(NAD->getScopeList());
512 LLVM_DEBUG(dbgs() << " Updating OrigHeader scopes\n");
513 cloneAndAdaptNoAliasScopes(NoAliasDeclScopes, {OrigHeader}, Context,
515 LLVM_DEBUG(OrigHeader->dump());
517 // Keep the compile time impact low by only adapting the inserted block
518 // of instructions in the OrigPreHeader. This might result in slightly
519 // more aliasing between these instructions and those that were already
520 // present, but it will be much faster when the original PreHeader is
522 LLVM_DEBUG(dbgs() << " Updating part of OrigPreheader scopes\n");
524 cast<Instruction>(ValueMap[*NoAliasDeclInstructions.begin()]);
525 auto *LastInst = &OrigPreheader->back();
526 cloneAndAdaptNoAliasScopes(NoAliasDeclScopes, FirstDecl, LastInst,
528 LLVM_DEBUG(OrigPreheader->dump());
530 LLVM_DEBUG(dbgs() << " Updated NewHeader:\n");
531 LLVM_DEBUG(NewHeader->dump());
535 // Along with all the other instructions, we just cloned OrigHeader's
536 // terminator into OrigPreHeader. Fix up the PHI nodes in each of OrigHeader's
537 // successors by duplicating their incoming values for OrigHeader.
538 for (BasicBlock *SuccBB : successors(OrigHeader))
539 for (BasicBlock::iterator BI = SuccBB->begin();
540 PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
541 PN->addIncoming(PN->getIncomingValueForBlock(OrigHeader), OrigPreheader);
543 // Now that OrigPreHeader has a clone of OrigHeader's terminator, remove
544 // OrigPreHeader's old terminator (the original branch into the loop), and
545 // remove the corresponding incoming values from the PHI nodes in OrigHeader.
546 LoopEntryBranch->eraseFromParent();
548 // Update MemorySSA before the rewrite call below changes the 1:1
549 // instruction:cloned_instruction_or_value mapping.
551 InsertNewValueIntoMap(ValueMapMSSA, OrigHeader, OrigPreheader);
552 MSSAU->updateForClonedBlockIntoPred(OrigHeader, OrigPreheader,
556 SmallVector<PHINode*, 2> InsertedPHIs;
557 // If there were any uses of instructions in the duplicated block outside the
558 // loop, update them, inserting PHI nodes as required
559 RewriteUsesOfClonedInstructions(OrigHeader, OrigPreheader, ValueMap,
562 // Attach dbg.value intrinsics to the new phis if that phi uses a value that
563 // previously had debug metadata attached. This keeps the debug info
564 // up-to-date in the loop body.
565 if (!InsertedPHIs.empty())
566 insertDebugValuesForPHIs(OrigHeader, InsertedPHIs);
568 // NewHeader is now the header of the loop.
569 L->moveToHeader(NewHeader);
570 assert(L->getHeader() == NewHeader && "Latch block is our new header");
572 // Inform DT about changes to the CFG.
574 // The OrigPreheader branches to the NewHeader and Exit now. Then, inform
575 // the DT about the removed edge to the OrigHeader (that got removed).
576 SmallVector<DominatorTree::UpdateType, 3> Updates;
577 Updates.push_back({DominatorTree::Insert, OrigPreheader, Exit});
578 Updates.push_back({DominatorTree::Insert, OrigPreheader, NewHeader});
579 Updates.push_back({DominatorTree::Delete, OrigPreheader, OrigHeader});
582 MSSAU->applyUpdates(Updates, *DT, /*UpdateDT=*/true);
584 MSSAU->getMemorySSA()->verifyMemorySSA();
586 DT->applyUpdates(Updates);
590 // At this point, we've finished our major CFG changes. As part of cloning
591 // the loop into the preheader we've simplified instructions and the
592 // duplicated conditional branch may now be branching on a constant. If it is
593 // branching on a constant and if that constant means that we enter the loop,
594 // then we fold away the cond branch to an uncond branch. This simplifies the
595 // loop in cases important for nested loops, and it also means we don't have
596 // to split as many edges.
597 BranchInst *PHBI = cast<BranchInst>(OrigPreheader->getTerminator());
598 assert(PHBI->isConditional() && "Should be clone of BI condbr!");
599 if (!isa<ConstantInt>(PHBI->getCondition()) ||
600 PHBI->getSuccessor(cast<ConstantInt>(PHBI->getCondition())->isZero()) !=
602 // The conditional branch can't be folded, handle the general case.
603 // Split edges as necessary to preserve LoopSimplify form.
605 // Right now OrigPreHeader has two successors, NewHeader and ExitBlock, and
606 // thus is not a preheader anymore.
607 // Split the edge to form a real preheader.
608 BasicBlock *NewPH = SplitCriticalEdge(
609 OrigPreheader, NewHeader,
610 CriticalEdgeSplittingOptions(DT, LI, MSSAU).setPreserveLCSSA());
611 NewPH->setName(NewHeader->getName() + ".lr.ph");
613 // Preserve canonical loop form, which means that 'Exit' should have only
614 // one predecessor. Note that Exit could be an exit block for multiple
615 // nested loops, causing both of the edges to now be critical and need to
617 SmallVector<BasicBlock *, 4> ExitPreds(pred_begin(Exit), pred_end(Exit));
618 bool SplitLatchEdge = false;
619 for (BasicBlock *ExitPred : ExitPreds) {
620 // We only need to split loop exit edges.
621 Loop *PredLoop = LI->getLoopFor(ExitPred);
622 if (!PredLoop || PredLoop->contains(Exit) ||
623 ExitPred->getTerminator()->isIndirectTerminator())
625 SplitLatchEdge |= L->getLoopLatch() == ExitPred;
626 BasicBlock *ExitSplit = SplitCriticalEdge(
628 CriticalEdgeSplittingOptions(DT, LI, MSSAU).setPreserveLCSSA());
629 ExitSplit->moveBefore(Exit);
631 assert(SplitLatchEdge &&
632 "Despite splitting all preds, failed to split latch exit?");
634 // We can fold the conditional branch in the preheader, this makes things
635 // simpler. The first step is to remove the extra edge to the Exit block.
636 Exit->removePredecessor(OrigPreheader, true /*preserve LCSSA*/);
637 BranchInst *NewBI = BranchInst::Create(NewHeader, PHBI);
638 NewBI->setDebugLoc(PHBI->getDebugLoc());
639 PHBI->eraseFromParent();
641 // With our CFG finalized, update DomTree if it is available.
642 if (DT) DT->deleteEdge(OrigPreheader, Exit);
644 // Update MSSA too, if available.
646 MSSAU->removeEdge(OrigPreheader, Exit);
649 assert(L->getLoopPreheader() && "Invalid loop preheader after loop rotation");
650 assert(L->getLoopLatch() && "Invalid loop latch after loop rotation");
652 if (MSSAU && VerifyMemorySSA)
653 MSSAU->getMemorySSA()->verifyMemorySSA();
655 // Now that the CFG and DomTree are in a consistent state again, try to merge
656 // the OrigHeader block into OrigLatch. This will succeed if they are
657 // connected by an unconditional branch. This is just a cleanup so the
658 // emitted code isn't too gross in this common case.
659 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
660 BasicBlock *PredBB = OrigHeader->getUniquePredecessor();
661 bool DidMerge = MergeBlockIntoPredecessor(OrigHeader, &DTU, LI, MSSAU);
663 RemoveRedundantDbgInstrs(PredBB);
665 if (MSSAU && VerifyMemorySSA)
666 MSSAU->getMemorySSA()->verifyMemorySSA();
668 LLVM_DEBUG(dbgs() << "LoopRotation: into "; L->dump());
673 SimplifiedLatch = false;
675 // Check that new latch is a deoptimizing exit and then repeat rotation if possible.
676 // Deoptimizing latch exit is not a generally typical case, so we just loop over.
677 // TODO: if it becomes a performance bottleneck extend rotation algorithm
678 // to handle multiple rotations in one go.
679 } while (MultiRotate && canRotateDeoptimizingLatchExit(L));
685 /// Determine whether the instructions in this range may be safely and cheaply
686 /// speculated. This is not an important enough situation to develop complex
687 /// heuristics. We handle a single arithmetic instruction along with any type
689 static bool shouldSpeculateInstrs(BasicBlock::iterator Begin,
690 BasicBlock::iterator End, Loop *L) {
691 bool seenIncrement = false;
692 bool MultiExitLoop = false;
694 if (!L->getExitingBlock())
695 MultiExitLoop = true;
697 for (BasicBlock::iterator I = Begin; I != End; ++I) {
699 if (!isSafeToSpeculativelyExecute(&*I))
702 if (isa<DbgInfoIntrinsic>(I))
705 switch (I->getOpcode()) {
708 case Instruction::GetElementPtr:
709 // GEPs are cheap if all indices are constant.
710 if (!cast<GEPOperator>(I)->hasAllConstantIndices())
712 // fall-thru to increment case
714 case Instruction::Add:
715 case Instruction::Sub:
716 case Instruction::And:
717 case Instruction::Or:
718 case Instruction::Xor:
719 case Instruction::Shl:
720 case Instruction::LShr:
721 case Instruction::AShr: {
723 !isa<Constant>(I->getOperand(0))
725 : !isa<Constant>(I->getOperand(1)) ? I->getOperand(1) : nullptr;
729 // If increment operand is used outside of the loop, this speculation
730 // could cause extra live range interference.
732 for (User *UseI : IVOpnd->users()) {
733 auto *UserInst = cast<Instruction>(UseI);
734 if (!L->contains(UserInst))
741 seenIncrement = true;
744 case Instruction::Trunc:
745 case Instruction::ZExt:
746 case Instruction::SExt:
747 // ignore type conversions
754 /// Fold the loop tail into the loop exit by speculating the loop tail
755 /// instructions. Typically, this is a single post-increment. In the case of a
756 /// simple 2-block loop, hoisting the increment can be much better than
757 /// duplicating the entire loop header. In the case of loops with early exits,
758 /// rotation will not work anyway, but simplifyLoopLatch will put the loop in
759 /// canonical form so downstream passes can handle it.
761 /// I don't believe this invalidates SCEV.
762 bool LoopRotate::simplifyLoopLatch(Loop *L) {
763 BasicBlock *Latch = L->getLoopLatch();
764 if (!Latch || Latch->hasAddressTaken())
767 BranchInst *Jmp = dyn_cast<BranchInst>(Latch->getTerminator());
768 if (!Jmp || !Jmp->isUnconditional())
771 BasicBlock *LastExit = Latch->getSinglePredecessor();
772 if (!LastExit || !L->isLoopExiting(LastExit))
775 BranchInst *BI = dyn_cast<BranchInst>(LastExit->getTerminator());
779 if (!shouldSpeculateInstrs(Latch->begin(), Jmp->getIterator(), L))
782 LLVM_DEBUG(dbgs() << "Folding loop latch " << Latch->getName() << " into "
783 << LastExit->getName() << "\n");
785 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
786 MergeBlockIntoPredecessor(Latch, &DTU, LI, MSSAU, nullptr,
787 /*PredecessorWithTwoSuccessors=*/true);
789 if (MSSAU && VerifyMemorySSA)
790 MSSAU->getMemorySSA()->verifyMemorySSA();
795 /// Rotate \c L, and return true if any modification was made.
796 bool LoopRotate::processLoop(Loop *L) {
797 // Save the loop metadata.
798 MDNode *LoopMD = L->getLoopID();
800 bool SimplifiedLatch = false;
802 // Simplify the loop latch before attempting to rotate the header
803 // upward. Rotation may not be needed if the loop tail can be folded into the
806 SimplifiedLatch = simplifyLoopLatch(L);
808 bool MadeChange = rotateLoop(L, SimplifiedLatch);
809 assert((!MadeChange || L->isLoopExiting(L->getLoopLatch())) &&
810 "Loop latch should be exiting after loop-rotate.");
812 // Restore the loop metadata.
813 // NB! We presume LoopRotation DOESN'T ADD its own metadata.
814 if ((MadeChange || SimplifiedLatch) && LoopMD)
815 L->setLoopID(LoopMD);
817 return MadeChange || SimplifiedLatch;
821 /// The utility to convert a loop into a loop with bottom test.
822 bool llvm::LoopRotation(Loop *L, LoopInfo *LI, const TargetTransformInfo *TTI,
823 AssumptionCache *AC, DominatorTree *DT,
824 ScalarEvolution *SE, MemorySSAUpdater *MSSAU,
825 const SimplifyQuery &SQ, bool RotationOnly = true,
826 unsigned Threshold = unsigned(-1),
827 bool IsUtilMode = true, bool PrepareForLTO) {
828 LoopRotate LR(Threshold, LI, TTI, AC, DT, SE, MSSAU, SQ, RotationOnly,
829 IsUtilMode, PrepareForLTO);
830 return LR.processLoop(L);