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 #define DEBUG_TYPE "loop-rotate"
15 #include "llvm/Transforms/Scalar.h"
16 #include "llvm/ADT/Statistic.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/Dominators.h"
24 #include "llvm/IR/Function.h"
25 #include "llvm/IR/IntrinsicInst.h"
26 #include "llvm/Support/CFG.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
29 #include "llvm/Transforms/Utils/Local.h"
30 #include "llvm/Transforms/Utils/SSAUpdater.h"
31 #include "llvm/Transforms/Utils/ValueMapper.h"
34 #define MAX_HEADER_SIZE 16
36 STATISTIC(NumRotated, "Number of loops rotated");
39 class LoopRotate : public LoopPass {
41 static char ID; // Pass ID, replacement for typeid
42 LoopRotate() : LoopPass(ID) {
43 initializeLoopRotatePass(*PassRegistry::getPassRegistry());
46 // LCSSA form makes instruction renaming easier.
47 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
48 AU.addPreserved<DominatorTreeWrapperPass>();
49 AU.addRequired<LoopInfo>();
50 AU.addPreserved<LoopInfo>();
51 AU.addRequiredID(LoopSimplifyID);
52 AU.addPreservedID(LoopSimplifyID);
53 AU.addRequiredID(LCSSAID);
54 AU.addPreservedID(LCSSAID);
55 AU.addPreserved<ScalarEvolution>();
56 AU.addRequired<TargetTransformInfo>();
59 bool runOnLoop(Loop *L, LPPassManager &LPM);
60 bool simplifyLoopLatch(Loop *L);
61 bool rotateLoop(Loop *L, bool SimplifiedLatch);
65 const TargetTransformInfo *TTI;
69 char LoopRotate::ID = 0;
70 INITIALIZE_PASS_BEGIN(LoopRotate, "loop-rotate", "Rotate Loops", false, false)
71 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
72 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
73 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
74 INITIALIZE_PASS_DEPENDENCY(LCSSA)
75 INITIALIZE_PASS_END(LoopRotate, "loop-rotate", "Rotate Loops", false, false)
77 Pass *llvm::createLoopRotatePass() { return new LoopRotate(); }
79 /// Rotate Loop L as many times as possible. Return true if
80 /// the loop is rotated at least once.
81 bool LoopRotate::runOnLoop(Loop *L, LPPassManager &LPM) {
82 LI = &getAnalysis<LoopInfo>();
83 TTI = &getAnalysis<TargetTransformInfo>();
85 // Simplify the loop latch before attempting to rotate the header
86 // upward. Rotation may not be needed if the loop tail can be folded into the
88 bool SimplifiedLatch = simplifyLoopLatch(L);
90 // One loop can be rotated multiple times.
91 bool MadeChange = false;
92 while (rotateLoop(L, SimplifiedLatch)) {
94 SimplifiedLatch = false;
99 /// RewriteUsesOfClonedInstructions - We just cloned the instructions from the
100 /// old header into the preheader. If there were uses of the values produced by
101 /// these instruction that were outside of the loop, we have to insert PHI nodes
102 /// to merge the two values. Do this now.
103 static void RewriteUsesOfClonedInstructions(BasicBlock *OrigHeader,
104 BasicBlock *OrigPreheader,
105 ValueToValueMapTy &ValueMap) {
106 // Remove PHI node entries that are no longer live.
107 BasicBlock::iterator I, E = OrigHeader->end();
108 for (I = OrigHeader->begin(); PHINode *PN = dyn_cast<PHINode>(I); ++I)
109 PN->removeIncomingValue(PN->getBasicBlockIndex(OrigPreheader));
111 // Now fix up users of the instructions in OrigHeader, inserting PHI nodes
114 for (I = OrigHeader->begin(); I != E; ++I) {
115 Value *OrigHeaderVal = I;
117 // If there are no uses of the value (e.g. because it returns void), there
118 // is nothing to rewrite.
119 if (OrigHeaderVal->use_empty())
122 Value *OrigPreHeaderVal = ValueMap[OrigHeaderVal];
124 // The value now exits in two versions: the initial value in the preheader
125 // and the loop "next" value in the original header.
126 SSA.Initialize(OrigHeaderVal->getType(), OrigHeaderVal->getName());
127 SSA.AddAvailableValue(OrigHeader, OrigHeaderVal);
128 SSA.AddAvailableValue(OrigPreheader, OrigPreHeaderVal);
130 // Visit each use of the OrigHeader instruction.
131 for (Value::use_iterator UI = OrigHeaderVal->use_begin(),
132 UE = OrigHeaderVal->use_end(); UI != UE; ) {
133 // Grab the use before incrementing the iterator.
134 Use &U = UI.getUse();
136 // Increment the iterator before removing the use from the list.
139 // SSAUpdater can't handle a non-PHI use in the same block as an
140 // earlier def. We can easily handle those cases manually.
141 Instruction *UserInst = cast<Instruction>(U.getUser());
142 if (!isa<PHINode>(UserInst)) {
143 BasicBlock *UserBB = UserInst->getParent();
145 // The original users in the OrigHeader are already using the
146 // original definitions.
147 if (UserBB == OrigHeader)
150 // Users in the OrigPreHeader need to use the value to which the
151 // original definitions are mapped.
152 if (UserBB == OrigPreheader) {
153 U = OrigPreHeaderVal;
158 // Anything else can be handled by SSAUpdater.
164 /// Determine whether the instructions in this range my be safely and cheaply
165 /// speculated. This is not an important enough situation to develop complex
166 /// heuristics. We handle a single arithmetic instruction along with any type
168 static bool shouldSpeculateInstrs(BasicBlock::iterator Begin,
169 BasicBlock::iterator End) {
170 bool seenIncrement = false;
171 for (BasicBlock::iterator I = Begin; I != End; ++I) {
173 if (!isSafeToSpeculativelyExecute(I))
176 if (isa<DbgInfoIntrinsic>(I))
179 switch (I->getOpcode()) {
182 case Instruction::GetElementPtr:
183 // GEPs are cheap if all indices are constant.
184 if (!cast<GEPOperator>(I)->hasAllConstantIndices())
186 // fall-thru to increment case
187 case Instruction::Add:
188 case Instruction::Sub:
189 case Instruction::And:
190 case Instruction::Or:
191 case Instruction::Xor:
192 case Instruction::Shl:
193 case Instruction::LShr:
194 case Instruction::AShr:
197 seenIncrement = true;
199 case Instruction::Trunc:
200 case Instruction::ZExt:
201 case Instruction::SExt:
202 // ignore type conversions
209 /// Fold the loop tail into the loop exit by speculating the loop tail
210 /// instructions. Typically, this is a single post-increment. In the case of a
211 /// simple 2-block loop, hoisting the increment can be much better than
212 /// duplicating the entire loop header. In the cast of loops with early exits,
213 /// rotation will not work anyway, but simplifyLoopLatch will put the loop in
214 /// canonical form so downstream passes can handle it.
216 /// I don't believe this invalidates SCEV.
217 bool LoopRotate::simplifyLoopLatch(Loop *L) {
218 BasicBlock *Latch = L->getLoopLatch();
219 if (!Latch || Latch->hasAddressTaken())
222 BranchInst *Jmp = dyn_cast<BranchInst>(Latch->getTerminator());
223 if (!Jmp || !Jmp->isUnconditional())
226 BasicBlock *LastExit = Latch->getSinglePredecessor();
227 if (!LastExit || !L->isLoopExiting(LastExit))
230 BranchInst *BI = dyn_cast<BranchInst>(LastExit->getTerminator());
234 if (!shouldSpeculateInstrs(Latch->begin(), Jmp))
237 DEBUG(dbgs() << "Folding loop latch " << Latch->getName() << " into "
238 << LastExit->getName() << "\n");
240 // Hoist the instructions from Latch into LastExit.
241 LastExit->getInstList().splice(BI, Latch->getInstList(), Latch->begin(), Jmp);
243 unsigned FallThruPath = BI->getSuccessor(0) == Latch ? 0 : 1;
244 BasicBlock *Header = Jmp->getSuccessor(0);
245 assert(Header == L->getHeader() && "expected a backward branch");
247 // Remove Latch from the CFG so that LastExit becomes the new Latch.
248 BI->setSuccessor(FallThruPath, Header);
249 Latch->replaceSuccessorsPhiUsesWith(LastExit);
250 Jmp->eraseFromParent();
252 // Nuke the Latch block.
253 assert(Latch->empty() && "unable to evacuate Latch");
254 LI->removeBlock(Latch);
255 if (DominatorTreeWrapperPass *DTWP =
256 getAnalysisIfAvailable<DominatorTreeWrapperPass>())
257 DTWP->getDomTree().eraseNode(Latch);
258 Latch->eraseFromParent();
262 /// Rotate loop LP. Return true if the loop is rotated.
264 /// \param SimplifiedLatch is true if the latch was just folded into the final
265 /// loop exit. In this case we may want to rotate even though the new latch is
266 /// now an exiting branch. This rotation would have happened had the latch not
267 /// been simplified. However, if SimplifiedLatch is false, then we avoid
268 /// rotating loops in which the latch exits to avoid excessive or endless
269 /// rotation. LoopRotate should be repeatable and converge to a canonical
270 /// form. This property is satisfied because simplifying the loop latch can only
271 /// happen once across multiple invocations of the LoopRotate pass.
272 bool LoopRotate::rotateLoop(Loop *L, bool SimplifiedLatch) {
273 // If the loop has only one block then there is not much to rotate.
274 if (L->getBlocks().size() == 1)
277 BasicBlock *OrigHeader = L->getHeader();
278 BasicBlock *OrigLatch = L->getLoopLatch();
280 BranchInst *BI = dyn_cast<BranchInst>(OrigHeader->getTerminator());
281 if (BI == 0 || BI->isUnconditional())
284 // If the loop header is not one of the loop exiting blocks then
285 // either this loop is already rotated or it is not
286 // suitable for loop rotation transformations.
287 if (!L->isLoopExiting(OrigHeader))
290 // If the loop latch already contains a branch that leaves the loop then the
291 // loop is already rotated.
295 // Rotate if either the loop latch does *not* exit the loop, or if the loop
296 // latch was just simplified.
297 if (L->isLoopExiting(OrigLatch) && !SimplifiedLatch)
300 // Check size of original header and reject loop if it is very big or we can't
301 // duplicate blocks inside it.
304 Metrics.analyzeBasicBlock(OrigHeader, *TTI);
305 if (Metrics.notDuplicatable) {
306 DEBUG(dbgs() << "LoopRotation: NOT rotating - contains non-duplicatable"
307 << " instructions: "; L->dump());
310 if (Metrics.NumInsts > MAX_HEADER_SIZE)
314 // Now, this loop is suitable for rotation.
315 BasicBlock *OrigPreheader = L->getLoopPreheader();
317 // If the loop could not be converted to canonical form, it must have an
318 // indirectbr in it, just give up.
319 if (OrigPreheader == 0)
322 // Anything ScalarEvolution may know about this loop or the PHI nodes
323 // in its header will soon be invalidated.
324 if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>())
327 DEBUG(dbgs() << "LoopRotation: rotating "; L->dump());
329 // Find new Loop header. NewHeader is a Header's one and only successor
330 // that is inside loop. Header's other successor is outside the
331 // loop. Otherwise loop is not suitable for rotation.
332 BasicBlock *Exit = BI->getSuccessor(0);
333 BasicBlock *NewHeader = BI->getSuccessor(1);
334 if (L->contains(Exit))
335 std::swap(Exit, NewHeader);
336 assert(NewHeader && "Unable to determine new loop header");
337 assert(L->contains(NewHeader) && !L->contains(Exit) &&
338 "Unable to determine loop header and exit blocks");
340 // This code assumes that the new header has exactly one predecessor.
341 // Remove any single-entry PHI nodes in it.
342 assert(NewHeader->getSinglePredecessor() &&
343 "New header doesn't have one pred!");
344 FoldSingleEntryPHINodes(NewHeader);
346 // Begin by walking OrigHeader and populating ValueMap with an entry for
348 BasicBlock::iterator I = OrigHeader->begin(), E = OrigHeader->end();
349 ValueToValueMapTy ValueMap;
351 // For PHI nodes, the value available in OldPreHeader is just the
352 // incoming value from OldPreHeader.
353 for (; PHINode *PN = dyn_cast<PHINode>(I); ++I)
354 ValueMap[PN] = PN->getIncomingValueForBlock(OrigPreheader);
356 // For the rest of the instructions, either hoist to the OrigPreheader if
357 // possible or create a clone in the OldPreHeader if not.
358 TerminatorInst *LoopEntryBranch = OrigPreheader->getTerminator();
360 Instruction *Inst = I++;
362 // If the instruction's operands are invariant and it doesn't read or write
363 // memory, then it is safe to hoist. Doing this doesn't change the order of
364 // execution in the preheader, but does prevent the instruction from
365 // executing in each iteration of the loop. This means it is safe to hoist
366 // something that might trap, but isn't safe to hoist something that reads
367 // memory (without proving that the loop doesn't write).
368 if (L->hasLoopInvariantOperands(Inst) &&
369 !Inst->mayReadFromMemory() && !Inst->mayWriteToMemory() &&
370 !isa<TerminatorInst>(Inst) && !isa<DbgInfoIntrinsic>(Inst) &&
371 !isa<AllocaInst>(Inst)) {
372 Inst->moveBefore(LoopEntryBranch);
376 // Otherwise, create a duplicate of the instruction.
377 Instruction *C = Inst->clone();
379 // Eagerly remap the operands of the instruction.
380 RemapInstruction(C, ValueMap,
381 RF_NoModuleLevelChanges|RF_IgnoreMissingEntries);
383 // With the operands remapped, see if the instruction constant folds or is
384 // otherwise simplifyable. This commonly occurs because the entry from PHI
385 // nodes allows icmps and other instructions to fold.
386 Value *V = SimplifyInstruction(C);
387 if (V && LI->replacementPreservesLCSSAForm(C, V)) {
388 // If so, then delete the temporary instruction and stick the folded value
393 // Otherwise, stick the new instruction into the new block!
394 C->setName(Inst->getName());
395 C->insertBefore(LoopEntryBranch);
400 // Along with all the other instructions, we just cloned OrigHeader's
401 // terminator into OrigPreHeader. Fix up the PHI nodes in each of OrigHeader's
402 // successors by duplicating their incoming values for OrigHeader.
403 TerminatorInst *TI = OrigHeader->getTerminator();
404 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
405 for (BasicBlock::iterator BI = TI->getSuccessor(i)->begin();
406 PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
407 PN->addIncoming(PN->getIncomingValueForBlock(OrigHeader), OrigPreheader);
409 // Now that OrigPreHeader has a clone of OrigHeader's terminator, remove
410 // OrigPreHeader's old terminator (the original branch into the loop), and
411 // remove the corresponding incoming values from the PHI nodes in OrigHeader.
412 LoopEntryBranch->eraseFromParent();
414 // If there were any uses of instructions in the duplicated block outside the
415 // loop, update them, inserting PHI nodes as required
416 RewriteUsesOfClonedInstructions(OrigHeader, OrigPreheader, ValueMap);
418 // NewHeader is now the header of the loop.
419 L->moveToHeader(NewHeader);
420 assert(L->getHeader() == NewHeader && "Latch block is our new header");
423 // At this point, we've finished our major CFG changes. As part of cloning
424 // the loop into the preheader we've simplified instructions and the
425 // duplicated conditional branch may now be branching on a constant. If it is
426 // branching on a constant and if that constant means that we enter the loop,
427 // then we fold away the cond branch to an uncond branch. This simplifies the
428 // loop in cases important for nested loops, and it also means we don't have
429 // to split as many edges.
430 BranchInst *PHBI = cast<BranchInst>(OrigPreheader->getTerminator());
431 assert(PHBI->isConditional() && "Should be clone of BI condbr!");
432 if (!isa<ConstantInt>(PHBI->getCondition()) ||
433 PHBI->getSuccessor(cast<ConstantInt>(PHBI->getCondition())->isZero())
435 // The conditional branch can't be folded, handle the general case.
436 // Update DominatorTree to reflect the CFG change we just made. Then split
437 // edges as necessary to preserve LoopSimplify form.
438 if (DominatorTreeWrapperPass *DTWP =
439 getAnalysisIfAvailable<DominatorTreeWrapperPass>()) {
440 DominatorTree &DT = DTWP->getDomTree();
441 // Everything that was dominated by the old loop header is now dominated
442 // by the original loop preheader. Conceptually the header was merged
443 // into the preheader, even though we reuse the actual block as a new
445 DomTreeNode *OrigHeaderNode = DT.getNode(OrigHeader);
446 SmallVector<DomTreeNode *, 8> HeaderChildren(OrigHeaderNode->begin(),
447 OrigHeaderNode->end());
448 DomTreeNode *OrigPreheaderNode = DT.getNode(OrigPreheader);
449 for (unsigned I = 0, E = HeaderChildren.size(); I != E; ++I)
450 DT.changeImmediateDominator(HeaderChildren[I], OrigPreheaderNode);
452 assert(DT.getNode(Exit)->getIDom() == OrigPreheaderNode);
453 assert(DT.getNode(NewHeader)->getIDom() == OrigPreheaderNode);
455 // Update OrigHeader to be dominated by the new header block.
456 DT.changeImmediateDominator(OrigHeader, OrigLatch);
459 // Right now OrigPreHeader has two successors, NewHeader and ExitBlock, and
460 // thus is not a preheader anymore.
461 // Split the edge to form a real preheader.
462 BasicBlock *NewPH = SplitCriticalEdge(OrigPreheader, NewHeader, this);
463 NewPH->setName(NewHeader->getName() + ".lr.ph");
465 // Preserve canonical loop form, which means that 'Exit' should have only
466 // one predecessor. Note that Exit could be an exit block for multiple
467 // nested loops, causing both of the edges to now be critical and need to
469 SmallVector<BasicBlock *, 4> ExitPreds(pred_begin(Exit), pred_end(Exit));
470 bool SplitLatchEdge = false;
471 for (SmallVectorImpl<BasicBlock *>::iterator PI = ExitPreds.begin(),
472 PE = ExitPreds.end();
474 // We only need to split loop exit edges.
475 Loop *PredLoop = LI->getLoopFor(*PI);
476 if (!PredLoop || PredLoop->contains(Exit))
478 SplitLatchEdge |= L->getLoopLatch() == *PI;
479 BasicBlock *ExitSplit = SplitCriticalEdge(*PI, Exit, this);
480 ExitSplit->moveBefore(Exit);
482 assert(SplitLatchEdge &&
483 "Despite splitting all preds, failed to split latch exit?");
485 // We can fold the conditional branch in the preheader, this makes things
486 // simpler. The first step is to remove the extra edge to the Exit block.
487 Exit->removePredecessor(OrigPreheader, true /*preserve LCSSA*/);
488 BranchInst *NewBI = BranchInst::Create(NewHeader, PHBI);
489 NewBI->setDebugLoc(PHBI->getDebugLoc());
490 PHBI->eraseFromParent();
492 // With our CFG finalized, update DomTree if it is available.
493 if (DominatorTreeWrapperPass *DTWP =
494 getAnalysisIfAvailable<DominatorTreeWrapperPass>()) {
495 DominatorTree &DT = DTWP->getDomTree();
496 // Update OrigHeader to be dominated by the new header block.
497 DT.changeImmediateDominator(NewHeader, OrigPreheader);
498 DT.changeImmediateDominator(OrigHeader, OrigLatch);
500 // Brute force incremental dominator tree update. Call
501 // findNearestCommonDominator on all CFG predecessors of each child of the
503 DomTreeNode *OrigHeaderNode = DT.getNode(OrigHeader);
504 SmallVector<DomTreeNode *, 8> HeaderChildren(OrigHeaderNode->begin(),
505 OrigHeaderNode->end());
509 for (unsigned I = 0, E = HeaderChildren.size(); I != E; ++I) {
510 DomTreeNode *Node = HeaderChildren[I];
511 BasicBlock *BB = Node->getBlock();
513 pred_iterator PI = pred_begin(BB);
514 BasicBlock *NearestDom = *PI;
515 for (pred_iterator PE = pred_end(BB); PI != PE; ++PI)
516 NearestDom = DT.findNearestCommonDominator(NearestDom, *PI);
518 // Remember if this changes the DomTree.
519 if (Node->getIDom()->getBlock() != NearestDom) {
520 DT.changeImmediateDominator(BB, NearestDom);
525 // If the dominator changed, this may have an effect on other
526 // predecessors, continue until we reach a fixpoint.
531 assert(L->getLoopPreheader() && "Invalid loop preheader after loop rotation");
532 assert(L->getLoopLatch() && "Invalid loop latch after loop rotation");
534 // Now that the CFG and DomTree are in a consistent state again, try to merge
535 // the OrigHeader block into OrigLatch. This will succeed if they are
536 // connected by an unconditional branch. This is just a cleanup so the
537 // emitted code isn't too gross in this common case.
538 MergeBlockIntoPredecessor(OrigHeader, this);
540 DEBUG(dbgs() << "LoopRotation: into "; L->dump());