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/Function.h"
17 #include "llvm/Analysis/CodeMetrics.h"
18 #include "llvm/Analysis/LoopPass.h"
19 #include "llvm/Analysis/InstructionSimplify.h"
20 #include "llvm/Analysis/ScalarEvolution.h"
21 #include "llvm/Transforms/Utils/Local.h"
22 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
23 #include "llvm/Transforms/Utils/SSAUpdater.h"
24 #include "llvm/Transforms/Utils/ValueMapper.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/ADT/Statistic.h"
29 #define MAX_HEADER_SIZE 16
31 STATISTIC(NumRotated, "Number of loops rotated");
34 class LoopRotate : public LoopPass {
36 static char ID; // Pass ID, replacement for typeid
37 LoopRotate() : LoopPass(ID) {
38 initializeLoopRotatePass(*PassRegistry::getPassRegistry());
41 // LCSSA form makes instruction renaming easier.
42 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
43 AU.addPreserved<DominatorTree>();
44 AU.addRequired<LoopInfo>();
45 AU.addPreserved<LoopInfo>();
46 AU.addRequiredID(LoopSimplifyID);
47 AU.addPreservedID(LoopSimplifyID);
48 AU.addRequiredID(LCSSAID);
49 AU.addPreservedID(LCSSAID);
50 AU.addPreserved<ScalarEvolution>();
53 bool runOnLoop(Loop *L, LPPassManager &LPM);
54 bool rotateLoop(Loop *L);
61 char LoopRotate::ID = 0;
62 INITIALIZE_PASS_BEGIN(LoopRotate, "loop-rotate", "Rotate Loops", false, false)
63 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
64 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
65 INITIALIZE_PASS_DEPENDENCY(LCSSA)
66 INITIALIZE_PASS_END(LoopRotate, "loop-rotate", "Rotate Loops", false, false)
68 Pass *llvm::createLoopRotatePass() { return new LoopRotate(); }
70 /// Rotate Loop L as many times as possible. Return true if
71 /// the loop is rotated at least once.
72 bool LoopRotate::runOnLoop(Loop *L, LPPassManager &LPM) {
73 LI = &getAnalysis<LoopInfo>();
75 // One loop can be rotated multiple times.
76 bool MadeChange = false;
83 /// RewriteUsesOfClonedInstructions - We just cloned the instructions from the
84 /// old header into the preheader. If there were uses of the values produced by
85 /// these instruction that were outside of the loop, we have to insert PHI nodes
86 /// to merge the two values. Do this now.
87 static void RewriteUsesOfClonedInstructions(BasicBlock *OrigHeader,
88 BasicBlock *OrigPreheader,
89 ValueToValueMapTy &ValueMap) {
90 // Remove PHI node entries that are no longer live.
91 BasicBlock::iterator I, E = OrigHeader->end();
92 for (I = OrigHeader->begin(); PHINode *PN = dyn_cast<PHINode>(I); ++I)
93 PN->removeIncomingValue(PN->getBasicBlockIndex(OrigPreheader));
95 // Now fix up users of the instructions in OrigHeader, inserting PHI nodes
98 for (I = OrigHeader->begin(); I != E; ++I) {
99 Value *OrigHeaderVal = I;
101 // If there are no uses of the value (e.g. because it returns void), there
102 // is nothing to rewrite.
103 if (OrigHeaderVal->use_empty())
106 Value *OrigPreHeaderVal = ValueMap[OrigHeaderVal];
108 // The value now exits in two versions: the initial value in the preheader
109 // and the loop "next" value in the original header.
110 SSA.Initialize(OrigHeaderVal->getType(), OrigHeaderVal->getName());
111 SSA.AddAvailableValue(OrigHeader, OrigHeaderVal);
112 SSA.AddAvailableValue(OrigPreheader, OrigPreHeaderVal);
114 // Visit each use of the OrigHeader instruction.
115 for (Value::use_iterator UI = OrigHeaderVal->use_begin(),
116 UE = OrigHeaderVal->use_end(); UI != UE; ) {
117 // Grab the use before incrementing the iterator.
118 Use &U = UI.getUse();
120 // Increment the iterator before removing the use from the list.
123 // SSAUpdater can't handle a non-PHI use in the same block as an
124 // earlier def. We can easily handle those cases manually.
125 Instruction *UserInst = cast<Instruction>(U.getUser());
126 if (!isa<PHINode>(UserInst)) {
127 BasicBlock *UserBB = UserInst->getParent();
129 // The original users in the OrigHeader are already using the
130 // original definitions.
131 if (UserBB == OrigHeader)
134 // Users in the OrigPreHeader need to use the value to which the
135 // original definitions are mapped.
136 if (UserBB == OrigPreheader) {
137 U = OrigPreHeaderVal;
142 // Anything else can be handled by SSAUpdater.
148 /// Rotate loop LP. Return true if the loop is rotated.
149 bool LoopRotate::rotateLoop(Loop *L) {
150 // If the loop has only one block then there is not much to rotate.
151 if (L->getBlocks().size() == 1)
154 BasicBlock *OrigHeader = L->getHeader();
156 BranchInst *BI = dyn_cast<BranchInst>(OrigHeader->getTerminator());
157 if (BI == 0 || BI->isUnconditional())
160 // If the loop header is not one of the loop exiting blocks then
161 // either this loop is already rotated or it is not
162 // suitable for loop rotation transformations.
163 if (!L->isLoopExiting(OrigHeader))
166 // Updating PHInodes in loops with multiple exits adds complexity.
167 // Keep it simple, and restrict loop rotation to loops with one exit only.
168 // In future, lift this restriction and support for multiple exits if
170 SmallVector<BasicBlock*, 8> ExitBlocks;
171 L->getExitBlocks(ExitBlocks);
172 if (ExitBlocks.size() > 1)
175 // Check size of original header and reject loop if it is very big.
178 Metrics.analyzeBasicBlock(OrigHeader);
179 if (Metrics.NumInsts > MAX_HEADER_SIZE)
183 // Now, this loop is suitable for rotation.
184 BasicBlock *OrigPreheader = L->getLoopPreheader();
185 BasicBlock *OrigLatch = L->getLoopLatch();
186 assert(OrigPreheader && OrigLatch && "Loop not in canonical form?");
188 // Anything ScalarEvolution may know about this loop or the PHI nodes
189 // in its header will soon be invalidated.
190 if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>())
193 // Find new Loop header. NewHeader is a Header's one and only successor
194 // that is inside loop. Header's other successor is outside the
195 // loop. Otherwise loop is not suitable for rotation.
196 BasicBlock *Exit = BI->getSuccessor(0);
197 BasicBlock *NewHeader = BI->getSuccessor(1);
198 if (L->contains(Exit))
199 std::swap(Exit, NewHeader);
200 assert(NewHeader && "Unable to determine new loop header");
201 assert(L->contains(NewHeader) && !L->contains(Exit) &&
202 "Unable to determine loop header and exit blocks");
204 // This code assumes that the new header has exactly one predecessor.
205 // Remove any single-entry PHI nodes in it.
206 assert(NewHeader->getSinglePredecessor() &&
207 "New header doesn't have one pred!");
208 FoldSingleEntryPHINodes(NewHeader);
210 // Begin by walking OrigHeader and populating ValueMap with an entry for
212 BasicBlock::iterator I = OrigHeader->begin(), E = OrigHeader->end();
213 ValueToValueMapTy ValueMap;
215 // For PHI nodes, the value available in OldPreHeader is just the
216 // incoming value from OldPreHeader.
217 for (; PHINode *PN = dyn_cast<PHINode>(I); ++I)
218 ValueMap[PN] = PN->getIncomingValue(PN->getBasicBlockIndex(OrigPreheader));
220 // For the rest of the instructions, either hoist to the OrigPreheader if
221 // possible or create a clone in the OldPreHeader if not.
222 TerminatorInst *LoopEntryBranch = OrigPreheader->getTerminator();
224 Instruction *Inst = I++;
226 // If the instruction's operands are invariant and it doesn't read or write
227 // memory, then it is safe to hoist. Doing this doesn't change the order of
228 // execution in the preheader, but does prevent the instruction from
229 // executing in each iteration of the loop. This means it is safe to hoist
230 // something that might trap, but isn't safe to hoist something that reads
231 // memory (without proving that the loop doesn't write).
232 if (L->hasLoopInvariantOperands(Inst) &&
233 !Inst->mayReadFromMemory() && !Inst->mayWriteToMemory() &&
234 !isa<TerminatorInst>(Inst)) {
235 Inst->moveBefore(LoopEntryBranch);
239 // Otherwise, create a duplicate of the instruction.
240 Instruction *C = Inst->clone();
242 // Eagerly remap the operands of the instruction.
243 RemapInstruction(C, ValueMap,
244 RF_NoModuleLevelChanges|RF_IgnoreMissingEntries);
246 // With the operands remapped, see if the instruction constant folds or is
247 // otherwise simplifyable. This commonly occurs because the entry from PHI
248 // nodes allows icmps and other instructions to fold.
249 Value *V = SimplifyInstruction(C);
250 if (V && LI->replacementPreservesLCSSAForm(C, V)) {
251 // If so, then delete the temporary instruction and stick the folded value
256 // Otherwise, stick the new instruction into the new block!
257 C->setName(Inst->getName());
258 C->insertBefore(LoopEntryBranch);
263 // Along with all the other instructions, we just cloned OrigHeader's
264 // terminator into OrigPreHeader. Fix up the PHI nodes in each of OrigHeader's
265 // successors by duplicating their incoming values for OrigHeader.
266 TerminatorInst *TI = OrigHeader->getTerminator();
267 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
268 for (BasicBlock::iterator BI = TI->getSuccessor(i)->begin();
269 PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
270 PN->addIncoming(PN->getIncomingValueForBlock(OrigHeader), OrigPreheader);
272 // Now that OrigPreHeader has a clone of OrigHeader's terminator, remove
273 // OrigPreHeader's old terminator (the original branch into the loop), and
274 // remove the corresponding incoming values from the PHI nodes in OrigHeader.
275 LoopEntryBranch->eraseFromParent();
277 // If there were any uses of instructions in the duplicated block outside the
278 // loop, update them, inserting PHI nodes as required
279 RewriteUsesOfClonedInstructions(OrigHeader, OrigPreheader, ValueMap);
281 // NewHeader is now the header of the loop.
282 L->moveToHeader(NewHeader);
283 assert(L->getHeader() == NewHeader && "Latch block is our new header");
286 // At this point, we've finished our major CFG changes. As part of cloning
287 // the loop into the preheader we've simplified instructions and the
288 // duplicated conditional branch may now be branching on a constant. If it is
289 // branching on a constant and if that constant means that we enter the loop,
290 // then we fold away the cond branch to an uncond branch. This simplifies the
291 // loop in cases important for nested loops, and it also means we don't have
292 // to split as many edges.
293 BranchInst *PHBI = cast<BranchInst>(OrigPreheader->getTerminator());
294 assert(PHBI->isConditional() && "Should be clone of BI condbr!");
295 if (!isa<ConstantInt>(PHBI->getCondition()) ||
296 PHBI->getSuccessor(cast<ConstantInt>(PHBI->getCondition())->isZero())
298 // The conditional branch can't be folded, handle the general case.
299 // Update DominatorTree to reflect the CFG change we just made. Then split
300 // edges as necessary to preserve LoopSimplify form.
301 if (DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>()) {
302 // Since OrigPreheader now has the conditional branch to Exit block, it is
303 // the dominator of Exit.
304 DT->changeImmediateDominator(Exit, OrigPreheader);
305 DT->changeImmediateDominator(NewHeader, OrigPreheader);
307 // Update OrigHeader to be dominated by the new header block.
308 DT->changeImmediateDominator(OrigHeader, OrigLatch);
311 // Right now OrigPreHeader has two successors, NewHeader and ExitBlock, and
312 // thus is not a preheader anymore. Split the edge to form a real preheader.
313 BasicBlock *NewPH = SplitCriticalEdge(OrigPreheader, NewHeader, this);
314 NewPH->setName(NewHeader->getName() + ".lr.ph");
316 // Preserve canonical loop form, which means that 'Exit' should have only one
318 BasicBlock *ExitSplit = SplitCriticalEdge(L->getLoopLatch(), Exit, this);
319 ExitSplit->moveBefore(Exit);
321 // We can fold the conditional branch in the preheader, this makes things
322 // simpler. The first step is to remove the extra edge to the Exit block.
323 Exit->removePredecessor(OrigPreheader, true /*preserve LCSSA*/);
324 BranchInst::Create(NewHeader, PHBI);
325 PHBI->eraseFromParent();
327 // With our CFG finalized, update DomTree if it is available.
328 if (DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>()) {
329 // Update OrigHeader to be dominated by the new header block.
330 DT->changeImmediateDominator(NewHeader, OrigPreheader);
331 DT->changeImmediateDominator(OrigHeader, OrigLatch);
335 assert(L->getLoopPreheader() && "Invalid loop preheader after loop rotation");
336 assert(L->getLoopLatch() && "Invalid loop latch after loop rotation");
338 // Now that the CFG and DomTree are in a consistent state again, merge the
339 // OrigHeader block into OrigLatch. We know that they are joined by an
340 // unconditional branch. This is just a cleanup so the emitted code isn't
342 bool DidIt = MergeBlockIntoPredecessor(OrigHeader, this);
343 assert(DidIt && "Block merge failed??"); (void)DidIt;