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/Function.h"
24 #include "llvm/IR/IntrinsicInst.h"
25 #include "llvm/Support/CFG.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
28 #include "llvm/Transforms/Utils/Local.h"
29 #include "llvm/Transforms/Utils/SSAUpdater.h"
30 #include "llvm/Transforms/Utils/ValueMapper.h"
33 #define MAX_HEADER_SIZE 16
35 STATISTIC(NumRotated, "Number of loops rotated");
38 class LoopRotate : public LoopPass {
40 static char ID; // Pass ID, replacement for typeid
41 LoopRotate() : LoopPass(ID) {
42 initializeLoopRotatePass(*PassRegistry::getPassRegistry());
45 // LCSSA form makes instruction renaming easier.
46 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
47 AU.addPreserved<DominatorTree>();
48 AU.addRequired<LoopInfo>();
49 AU.addPreserved<LoopInfo>();
50 AU.addRequiredID(LoopSimplifyID);
51 AU.addPreservedID(LoopSimplifyID);
52 AU.addRequiredID(LCSSAID);
53 AU.addPreservedID(LCSSAID);
54 AU.addPreserved<ScalarEvolution>();
55 AU.addRequired<TargetTransformInfo>();
58 bool runOnLoop(Loop *L, LPPassManager &LPM);
59 bool simplifyLoopLatch(Loop *L);
60 bool rotateLoop(Loop *L, bool SimplifiedLatch);
64 const TargetTransformInfo *TTI;
68 char LoopRotate::ID = 0;
69 INITIALIZE_PASS_BEGIN(LoopRotate, "loop-rotate", "Rotate Loops", false, false)
70 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
71 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
72 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
73 INITIALIZE_PASS_DEPENDENCY(LCSSA)
74 INITIALIZE_PASS_END(LoopRotate, "loop-rotate", "Rotate Loops", false, false)
76 Pass *llvm::createLoopRotatePass() { return new LoopRotate(); }
78 /// Rotate Loop L as many times as possible. Return true if
79 /// the loop is rotated at least once.
80 bool LoopRotate::runOnLoop(Loop *L, LPPassManager &LPM) {
81 LI = &getAnalysis<LoopInfo>();
82 TTI = &getAnalysis<TargetTransformInfo>();
84 // Simplify the loop latch before attempting to rotate the header
85 // upward. Rotation may not be needed if the loop tail can be folded into the
87 bool SimplifiedLatch = simplifyLoopLatch(L);
89 // One loop can be rotated multiple times.
90 bool MadeChange = false;
91 while (rotateLoop(L, SimplifiedLatch)) {
93 SimplifiedLatch = false;
98 /// RewriteUsesOfClonedInstructions - We just cloned the instructions from the
99 /// old header into the preheader. If there were uses of the values produced by
100 /// these instruction that were outside of the loop, we have to insert PHI nodes
101 /// to merge the two values. Do this now.
102 static void RewriteUsesOfClonedInstructions(BasicBlock *OrigHeader,
103 BasicBlock *OrigPreheader,
104 ValueToValueMapTy &ValueMap) {
105 // Remove PHI node entries that are no longer live.
106 BasicBlock::iterator I, E = OrigHeader->end();
107 for (I = OrigHeader->begin(); PHINode *PN = dyn_cast<PHINode>(I); ++I)
108 PN->removeIncomingValue(PN->getBasicBlockIndex(OrigPreheader));
110 // Now fix up users of the instructions in OrigHeader, inserting PHI nodes
113 for (I = OrigHeader->begin(); I != E; ++I) {
114 Value *OrigHeaderVal = I;
116 // If there are no uses of the value (e.g. because it returns void), there
117 // is nothing to rewrite.
118 if (OrigHeaderVal->use_empty())
121 Value *OrigPreHeaderVal = ValueMap[OrigHeaderVal];
123 // The value now exits in two versions: the initial value in the preheader
124 // and the loop "next" value in the original header.
125 SSA.Initialize(OrigHeaderVal->getType(), OrigHeaderVal->getName());
126 SSA.AddAvailableValue(OrigHeader, OrigHeaderVal);
127 SSA.AddAvailableValue(OrigPreheader, OrigPreHeaderVal);
129 // Visit each use of the OrigHeader instruction.
130 for (Value::use_iterator UI = OrigHeaderVal->use_begin(),
131 UE = OrigHeaderVal->use_end(); UI != UE; ) {
132 // Grab the use before incrementing the iterator.
133 Use &U = UI.getUse();
135 // Increment the iterator before removing the use from the list.
138 // SSAUpdater can't handle a non-PHI use in the same block as an
139 // earlier def. We can easily handle those cases manually.
140 Instruction *UserInst = cast<Instruction>(U.getUser());
141 if (!isa<PHINode>(UserInst)) {
142 BasicBlock *UserBB = UserInst->getParent();
144 // The original users in the OrigHeader are already using the
145 // original definitions.
146 if (UserBB == OrigHeader)
149 // Users in the OrigPreHeader need to use the value to which the
150 // original definitions are mapped.
151 if (UserBB == OrigPreheader) {
152 U = OrigPreHeaderVal;
157 // Anything else can be handled by SSAUpdater.
163 /// Determine whether the instructions in this range my be safely and cheaply
164 /// speculated. This is not an important enough situation to develop complex
165 /// heuristics. We handle a single arithmetic instruction along with any type
167 static bool shouldSpeculateInstrs(BasicBlock::iterator Begin,
168 BasicBlock::iterator End) {
169 bool seenIncrement = false;
170 for (BasicBlock::iterator I = Begin; I != End; ++I) {
172 if (!isSafeToSpeculativelyExecute(I))
175 if (isa<DbgInfoIntrinsic>(I))
178 switch (I->getOpcode()) {
181 case Instruction::GetElementPtr:
182 // GEPs are cheap if all indices are constant.
183 if (!cast<GEPOperator>(I)->hasAllConstantIndices())
185 // fall-thru to increment case
186 case Instruction::Add:
187 case Instruction::Sub:
188 case Instruction::And:
189 case Instruction::Or:
190 case Instruction::Xor:
191 case Instruction::Shl:
192 case Instruction::LShr:
193 case Instruction::AShr:
196 seenIncrement = true;
198 case Instruction::Trunc:
199 case Instruction::ZExt:
200 case Instruction::SExt:
201 // ignore type conversions
208 /// Fold the loop tail into the loop exit by speculating the loop tail
209 /// instructions. Typically, this is a single post-increment. In the case of a
210 /// simple 2-block loop, hoisting the increment can be much better than
211 /// duplicating the entire loop header. In the cast of loops with early exits,
212 /// rotation will not work anyway, but simplifyLoopLatch will put the loop in
213 /// canonical form so downstream passes can handle it.
215 /// I don't believe this invalidates SCEV.
216 bool LoopRotate::simplifyLoopLatch(Loop *L) {
217 BasicBlock *Latch = L->getLoopLatch();
218 if (!Latch || Latch->hasAddressTaken())
221 BranchInst *Jmp = dyn_cast<BranchInst>(Latch->getTerminator());
222 if (!Jmp || !Jmp->isUnconditional())
225 BasicBlock *LastExit = Latch->getSinglePredecessor();
226 if (!LastExit || !L->isLoopExiting(LastExit))
229 BranchInst *BI = dyn_cast<BranchInst>(LastExit->getTerminator());
233 if (!shouldSpeculateInstrs(Latch->begin(), Jmp))
236 DEBUG(dbgs() << "Folding loop latch " << Latch->getName() << " into "
237 << LastExit->getName() << "\n");
239 // Hoist the instructions from Latch into LastExit.
240 LastExit->getInstList().splice(BI, Latch->getInstList(), Latch->begin(), Jmp);
242 unsigned FallThruPath = BI->getSuccessor(0) == Latch ? 0 : 1;
243 BasicBlock *Header = Jmp->getSuccessor(0);
244 assert(Header == L->getHeader() && "expected a backward branch");
246 // Remove Latch from the CFG so that LastExit becomes the new Latch.
247 BI->setSuccessor(FallThruPath, Header);
248 Latch->replaceSuccessorsPhiUsesWith(LastExit);
249 Jmp->eraseFromParent();
251 // Nuke the Latch block.
252 assert(Latch->empty() && "unable to evacuate Latch");
253 LI->removeBlock(Latch);
254 if (DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>())
255 DT->eraseNode(Latch);
256 Latch->eraseFromParent();
260 /// Rotate loop LP. Return true if the loop is rotated.
262 /// \param SimplifiedLatch is true if the latch was just folded into the final
263 /// loop exit. In this case we may want to rotate even though the new latch is
264 /// now an exiting branch. This rotation would have happened had the latch not
265 /// been simplified. However, if SimplifiedLatch is false, then we avoid
266 /// rotating loops in which the latch exits to avoid excessive or endless
267 /// rotation. LoopRotate should be repeatable and converge to a canonical
268 /// form. This property is satisfied because simplifying the loop latch can only
269 /// happen once across multiple invocations of the LoopRotate pass.
270 bool LoopRotate::rotateLoop(Loop *L, bool SimplifiedLatch) {
271 // If the loop has only one block then there is not much to rotate.
272 if (L->getBlocks().size() == 1)
275 BasicBlock *OrigHeader = L->getHeader();
276 BasicBlock *OrigLatch = L->getLoopLatch();
278 BranchInst *BI = dyn_cast<BranchInst>(OrigHeader->getTerminator());
279 if (BI == 0 || 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.
295 if (L->isLoopExiting(OrigLatch) && !SimplifiedLatch)
298 // Check size of original header and reject loop if it is very big or we can't
299 // duplicate blocks inside it.
302 Metrics.analyzeBasicBlock(OrigHeader, *TTI);
303 if (Metrics.notDuplicatable) {
304 DEBUG(dbgs() << "LoopRotation: NOT rotating - contains non duplicatable"
305 << " instructions: "; L->dump());
308 if (Metrics.NumInsts > MAX_HEADER_SIZE)
312 // Now, this loop is suitable for rotation.
313 BasicBlock *OrigPreheader = L->getLoopPreheader();
315 // If the loop could not be converted to canonical form, it must have an
316 // indirectbr in it, just give up.
317 if (OrigPreheader == 0)
320 // Anything ScalarEvolution may know about this loop or the PHI nodes
321 // in its header will soon be invalidated.
322 if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>())
325 DEBUG(dbgs() << "LoopRotation: rotating "; L->dump());
327 // Find new Loop header. NewHeader is a Header's one and only successor
328 // that is inside loop. Header's other successor is outside the
329 // loop. Otherwise loop is not suitable for rotation.
330 BasicBlock *Exit = BI->getSuccessor(0);
331 BasicBlock *NewHeader = BI->getSuccessor(1);
332 if (L->contains(Exit))
333 std::swap(Exit, NewHeader);
334 assert(NewHeader && "Unable to determine new loop header");
335 assert(L->contains(NewHeader) && !L->contains(Exit) &&
336 "Unable to determine loop header and exit blocks");
338 // This code assumes that the new header has exactly one predecessor.
339 // Remove any single-entry PHI nodes in it.
340 assert(NewHeader->getSinglePredecessor() &&
341 "New header doesn't have one pred!");
342 FoldSingleEntryPHINodes(NewHeader);
344 // Begin by walking OrigHeader and populating ValueMap with an entry for
346 BasicBlock::iterator I = OrigHeader->begin(), E = OrigHeader->end();
347 ValueToValueMapTy ValueMap;
349 // For PHI nodes, the value available in OldPreHeader is just the
350 // incoming value from OldPreHeader.
351 for (; PHINode *PN = dyn_cast<PHINode>(I); ++I)
352 ValueMap[PN] = PN->getIncomingValueForBlock(OrigPreheader);
354 // For the rest of the instructions, either hoist to the OrigPreheader if
355 // possible or create a clone in the OldPreHeader if not.
356 TerminatorInst *LoopEntryBranch = OrigPreheader->getTerminator();
358 Instruction *Inst = I++;
360 // If the instruction's operands are invariant and it doesn't read or write
361 // memory, then it is safe to hoist. Doing this doesn't change the order of
362 // execution in the preheader, but does prevent the instruction from
363 // executing in each iteration of the loop. This means it is safe to hoist
364 // something that might trap, but isn't safe to hoist something that reads
365 // memory (without proving that the loop doesn't write).
366 if (L->hasLoopInvariantOperands(Inst) &&
367 !Inst->mayReadFromMemory() && !Inst->mayWriteToMemory() &&
368 !isa<TerminatorInst>(Inst) && !isa<DbgInfoIntrinsic>(Inst) &&
369 !isa<AllocaInst>(Inst)) {
370 Inst->moveBefore(LoopEntryBranch);
374 // Otherwise, create a duplicate of the instruction.
375 Instruction *C = Inst->clone();
377 // Eagerly remap the operands of the instruction.
378 RemapInstruction(C, ValueMap,
379 RF_NoModuleLevelChanges|RF_IgnoreMissingEntries);
381 // With the operands remapped, see if the instruction constant folds or is
382 // otherwise simplifyable. This commonly occurs because the entry from PHI
383 // nodes allows icmps and other instructions to fold.
384 Value *V = SimplifyInstruction(C);
385 if (V && LI->replacementPreservesLCSSAForm(C, V)) {
386 // If so, then delete the temporary instruction and stick the folded value
391 // Otherwise, stick the new instruction into the new block!
392 C->setName(Inst->getName());
393 C->insertBefore(LoopEntryBranch);
398 // Along with all the other instructions, we just cloned OrigHeader's
399 // terminator into OrigPreHeader. Fix up the PHI nodes in each of OrigHeader's
400 // successors by duplicating their incoming values for OrigHeader.
401 TerminatorInst *TI = OrigHeader->getTerminator();
402 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
403 for (BasicBlock::iterator BI = TI->getSuccessor(i)->begin();
404 PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
405 PN->addIncoming(PN->getIncomingValueForBlock(OrigHeader), OrigPreheader);
407 // Now that OrigPreHeader has a clone of OrigHeader's terminator, remove
408 // OrigPreHeader's old terminator (the original branch into the loop), and
409 // remove the corresponding incoming values from the PHI nodes in OrigHeader.
410 LoopEntryBranch->eraseFromParent();
412 // If there were any uses of instructions in the duplicated block outside the
413 // loop, update them, inserting PHI nodes as required
414 RewriteUsesOfClonedInstructions(OrigHeader, OrigPreheader, ValueMap);
416 // NewHeader is now the header of the loop.
417 L->moveToHeader(NewHeader);
418 assert(L->getHeader() == NewHeader && "Latch block is our new header");
421 // At this point, we've finished our major CFG changes. As part of cloning
422 // the loop into the preheader we've simplified instructions and the
423 // duplicated conditional branch may now be branching on a constant. If it is
424 // branching on a constant and if that constant means that we enter the loop,
425 // then we fold away the cond branch to an uncond branch. This simplifies the
426 // loop in cases important for nested loops, and it also means we don't have
427 // to split as many edges.
428 BranchInst *PHBI = cast<BranchInst>(OrigPreheader->getTerminator());
429 assert(PHBI->isConditional() && "Should be clone of BI condbr!");
430 if (!isa<ConstantInt>(PHBI->getCondition()) ||
431 PHBI->getSuccessor(cast<ConstantInt>(PHBI->getCondition())->isZero())
433 // The conditional branch can't be folded, handle the general case.
434 // Update DominatorTree to reflect the CFG change we just made. Then split
435 // edges as necessary to preserve LoopSimplify form.
436 if (DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>()) {
437 // Everything that was dominated by the old loop header is now dominated
438 // by the original loop preheader. Conceptually the header was merged
439 // into the preheader, even though we reuse the actual block as a new
441 DomTreeNode *OrigHeaderNode = DT->getNode(OrigHeader);
442 SmallVector<DomTreeNode *, 8> HeaderChildren(OrigHeaderNode->begin(),
443 OrigHeaderNode->end());
444 DomTreeNode *OrigPreheaderNode = DT->getNode(OrigPreheader);
445 for (unsigned I = 0, E = HeaderChildren.size(); I != E; ++I)
446 DT->changeImmediateDominator(HeaderChildren[I], OrigPreheaderNode);
448 assert(DT->getNode(Exit)->getIDom() == OrigPreheaderNode);
449 assert(DT->getNode(NewHeader)->getIDom() == OrigPreheaderNode);
451 // Update OrigHeader to be dominated by the new header block.
452 DT->changeImmediateDominator(OrigHeader, OrigLatch);
455 // Right now OrigPreHeader has two successors, NewHeader and ExitBlock, and
456 // thus is not a preheader anymore.
457 // Split the edge to form a real preheader.
458 BasicBlock *NewPH = SplitCriticalEdge(OrigPreheader, NewHeader, this);
459 NewPH->setName(NewHeader->getName() + ".lr.ph");
461 // Preserve canonical loop form, which means that 'Exit' should have only
463 BasicBlock *ExitSplit = SplitCriticalEdge(L->getLoopLatch(), Exit, this);
464 ExitSplit->moveBefore(Exit);
466 // We can fold the conditional branch in the preheader, this makes things
467 // simpler. The first step is to remove the extra edge to the Exit block.
468 Exit->removePredecessor(OrigPreheader, true /*preserve LCSSA*/);
469 BranchInst *NewBI = BranchInst::Create(NewHeader, PHBI);
470 NewBI->setDebugLoc(PHBI->getDebugLoc());
471 PHBI->eraseFromParent();
473 // With our CFG finalized, update DomTree if it is available.
474 if (DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>()) {
475 // Update OrigHeader to be dominated by the new header block.
476 DT->changeImmediateDominator(NewHeader, OrigPreheader);
477 DT->changeImmediateDominator(OrigHeader, OrigLatch);
479 // Brute force incremental dominator tree update. Call
480 // findNearestCommonDominator on all CFG predecessors of each child of the
482 DomTreeNode *OrigHeaderNode = DT->getNode(OrigHeader);
483 SmallVector<DomTreeNode *, 8> HeaderChildren(OrigHeaderNode->begin(),
484 OrigHeaderNode->end());
488 for (unsigned I = 0, E = HeaderChildren.size(); I != E; ++I) {
489 DomTreeNode *Node = HeaderChildren[I];
490 BasicBlock *BB = Node->getBlock();
492 pred_iterator PI = pred_begin(BB);
493 BasicBlock *NearestDom = *PI;
494 for (pred_iterator PE = pred_end(BB); PI != PE; ++PI)
495 NearestDom = DT->findNearestCommonDominator(NearestDom, *PI);
497 // Remember if this changes the DomTree.
498 if (Node->getIDom()->getBlock() != NearestDom) {
499 DT->changeImmediateDominator(BB, NearestDom);
504 // If the dominator changed, this may have an effect on other
505 // predecessors, continue until we reach a fixpoint.
510 assert(L->getLoopPreheader() && "Invalid loop preheader after loop rotation");
511 assert(L->getLoopLatch() && "Invalid loop latch after loop rotation");
513 // Now that the CFG and DomTree are in a consistent state again, try to merge
514 // the OrigHeader block into OrigLatch. This will succeed if they are
515 // connected by an unconditional branch. This is just a cleanup so the
516 // emitted code isn't too gross in this common case.
517 MergeBlockIntoPredecessor(OrigHeader, this);
519 DEBUG(dbgs() << "LoopRotation: into "; L->dump());