1 //===-- LoopUnroll.cpp - Loop unroller pass -------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass implements a simple loop unroller. It works best when loops have
11 // been canonicalized by the -indvars pass, allowing it to determine the trip
12 // counts of loops easily.
14 // This pass will multi-block loops only if they contain no non-unrolled
15 // subloops. The process of unrolling can produce extraneous basic blocks
16 // linked with unconditional branches. This will be corrected in the future.
18 //===----------------------------------------------------------------------===//
20 #define DEBUG_TYPE "loop-unroll"
21 #include "llvm/Transforms/Scalar.h"
22 #include "llvm/Constants.h"
23 #include "llvm/Function.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/Analysis/LoopInfo.h"
26 #include "llvm/Transforms/Utils/Cloning.h"
27 #include "llvm/Transforms/Utils/Local.h"
28 #include "llvm/Support/CFG.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/ADT/STLExtras.h"
33 #include "llvm/IntrinsicInst.h"
41 Statistic<> NumUnrolled("loop-unroll", "Number of loops completely unrolled");
44 UnrollThreshold("unroll-threshold", cl::init(100), cl::Hidden,
45 cl::desc("The cut-off point for loop unrolling"));
47 class LoopUnroll : public FunctionPass {
48 LoopInfo *LI; // The current loop information
50 virtual bool runOnFunction(Function &F);
51 bool visitLoop(Loop *L);
52 BasicBlock* FoldBlockIntoPredecessor(BasicBlock* BB);
54 /// This transformation requires natural loop information & requires that
55 /// loop preheaders be inserted into the CFG...
57 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
58 AU.addRequiredID(LoopSimplifyID);
59 AU.addRequiredID(LCSSAID);
60 AU.addRequired<LoopInfo>();
61 AU.addPreservedID(LCSSAID);
62 AU.addPreserved<LoopInfo>();
65 RegisterPass<LoopUnroll> X("loop-unroll", "Unroll loops");
68 FunctionPass *llvm::createLoopUnrollPass() { return new LoopUnroll(); }
70 bool LoopUnroll::runOnFunction(Function &F) {
72 LI = &getAnalysis<LoopInfo>();
74 // Transform all the top-level loops. Copy the loop list so that the child
75 // can update the loop tree if it needs to delete the loop.
76 std::vector<Loop*> SubLoops(LI->begin(), LI->end());
77 for (unsigned i = 0, e = SubLoops.size(); i != e; ++i)
78 Changed |= visitLoop(SubLoops[i]);
83 /// ApproximateLoopSize - Approximate the size of the loop after it has been
85 static unsigned ApproximateLoopSize(const Loop *L) {
87 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i) {
88 BasicBlock *BB = L->getBlocks()[i];
89 Instruction *Term = BB->getTerminator();
90 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
91 if (isa<PHINode>(I) && BB == L->getHeader()) {
92 // Ignore PHI nodes in the header.
93 } else if (I->hasOneUse() && I->use_back() == Term) {
94 // Ignore instructions only used by the loop terminator.
95 } else if (DbgInfoIntrinsic *DbgI = dyn_cast<DbgInfoIntrinsic>(I)) {
96 // Ignore debug instructions
101 // TODO: Ignore expressions derived from PHI and constants if inval of phi
102 // is a constant, or if operation is associative. This will get induction
110 // RemapInstruction - Convert the instruction operands from referencing the
111 // current values into those specified by ValueMap.
113 static inline void RemapInstruction(Instruction *I,
114 std::map<const Value *, Value*> &ValueMap) {
115 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
116 Value *Op = I->getOperand(op);
117 std::map<const Value *, Value*>::iterator It = ValueMap.find(Op);
118 if (It != ValueMap.end()) Op = It->second;
119 I->setOperand(op, Op);
123 // FoldBlockIntoPredecessor - Folds a basic block into its predecessor if it
124 // only has one predecessor, and that predecessor only has one successor.
125 // Returns the new combined block.
126 BasicBlock* LoopUnroll::FoldBlockIntoPredecessor(BasicBlock* BB) {
127 // Merge basic blocks into their predecessor if there is only one distinct
128 // pred, and if there is only one distinct successor of the predecessor, and
129 // if there are no PHI nodes.
131 BasicBlock *OnlyPred = BB->getSinglePredecessor();
132 if (!OnlyPred) return 0;
134 if (OnlyPred->getTerminator()->getNumSuccessors() != 1)
137 DEBUG(std::cerr << "Merging: " << *BB << "into: " << *OnlyPred);
138 TerminatorInst *Term = OnlyPred->getTerminator();
140 // Resolve any PHI nodes at the start of the block. They are all
141 // guaranteed to have exactly one entry if they exist, unless there are
142 // multiple duplicate (but guaranteed to be equal) entries for the
143 // incoming edges. This occurs when there are multiple edges from
144 // OnlyPred to OnlySucc.
146 while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) {
147 PN->replaceAllUsesWith(PN->getIncomingValue(0));
148 BB->getInstList().pop_front(); // Delete the phi node...
151 // Delete the unconditional branch from the predecessor...
152 OnlyPred->getInstList().pop_back();
154 // Move all definitions in the successor to the predecessor...
155 OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList());
157 // Make all PHI nodes that referred to BB now refer to Pred as their
159 BB->replaceAllUsesWith(OnlyPred);
161 std::string OldName = BB->getName();
163 // Erase basic block from the function...
165 BB->eraseFromParent();
167 // Inherit predecessors name if it exists...
168 if (!OldName.empty() && !OnlyPred->hasName())
169 OnlyPred->setName(OldName);
174 bool LoopUnroll::visitLoop(Loop *L) {
175 bool Changed = false;
177 // Recurse through all subloops before we process this loop. Copy the loop
178 // list so that the child can update the loop tree if it needs to delete the
180 std::vector<Loop*> SubLoops(L->begin(), L->end());
181 for (unsigned i = 0, e = SubLoops.size(); i != e; ++i)
182 Changed |= visitLoop(SubLoops[i]);
184 BasicBlock* Header = L->getHeader();
185 BasicBlock* LatchBlock = L->getLoopLatch();
187 BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator());
188 if (BI == 0) return Changed; // Must end in a conditional branch
190 ConstantInt *TripCountC = dyn_cast_or_null<ConstantInt>(L->getTripCount());
191 if (!TripCountC) return Changed; // Must have constant trip count!
193 uint64_t TripCountFull = TripCountC->getRawValue();
194 if (TripCountFull != TripCountC->getRawValue() || TripCountFull == 0)
195 return Changed; // More than 2^32 iterations???
197 unsigned LoopSize = ApproximateLoopSize(L);
198 DEBUG(std::cerr << "Loop Unroll: F[" << Header->getParent()->getName()
199 << "] Loop %" << Header->getName() << " Loop Size = "
200 << LoopSize << " Trip Count = " << TripCountFull << " - ");
201 uint64_t Size = (uint64_t)LoopSize*TripCountFull;
202 if (Size > UnrollThreshold) {
203 DEBUG(std::cerr << "TOO LARGE: " << Size << ">" << UnrollThreshold << "\n");
206 DEBUG(std::cerr << "UNROLLING!\n");
208 std::vector<BasicBlock*> LoopBlocks = L->getBlocks();
210 unsigned TripCount = (unsigned)TripCountFull;
212 BasicBlock *LoopExit = BI->getSuccessor(L->contains(BI->getSuccessor(0)));
214 // For the first iteration of the loop, we should use the precloned values for
215 // PHI nodes. Insert associations now.
216 std::map<const Value*, Value*> LastValueMap;
217 std::vector<PHINode*> OrigPHINode;
218 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
219 PHINode *PN = cast<PHINode>(I);
220 OrigPHINode.push_back(PN);
222 dyn_cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock)))
223 if (L->contains(I->getParent()))
227 // Remove the exit branch from the loop
228 LatchBlock->getInstList().erase(BI);
230 std::vector<BasicBlock*> Headers;
231 std::vector<BasicBlock*> Latches;
232 Headers.push_back(Header);
233 Latches.push_back(LatchBlock);
235 assert(TripCount != 0 && "Trip count of 0 is impossible!");
236 for (unsigned It = 1; It != TripCount; ++It) {
237 char SuffixBuffer[100];
238 sprintf(SuffixBuffer, ".%d", It);
240 std::vector<BasicBlock*> NewBlocks;
242 for (std::vector<BasicBlock*>::iterator BB = LoopBlocks.begin(),
243 E = LoopBlocks.end(); BB != E; ++BB) {
244 std::map<const Value*, Value*> ValueMap;
245 BasicBlock *New = CloneBasicBlock(*BB, ValueMap, SuffixBuffer);
246 Header->getParent()->getBasicBlockList().push_back(New);
248 // Loop over all of the PHI nodes in the block, changing them to use the
249 // incoming values from the previous block.
251 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
252 PHINode *NewPHI = cast<PHINode>(ValueMap[OrigPHINode[i]]);
253 Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
254 if (Instruction *InValI = dyn_cast<Instruction>(InVal))
255 if (It > 1 && L->contains(InValI->getParent()))
256 InVal = LastValueMap[InValI];
257 ValueMap[OrigPHINode[i]] = InVal;
258 New->getInstList().erase(NewPHI);
261 // Update our running map of newest clones
262 LastValueMap[*BB] = New;
263 for (std::map<const Value*, Value*>::iterator VI = ValueMap.begin(),
264 VE = ValueMap.end(); VI != VE; ++VI)
265 LastValueMap[VI->first] = VI->second;
267 L->addBasicBlockToLoop(New, *LI);
269 // Add phi entries for newly created values to all exit blocks except
270 // the successor of the latch block. The successor of the exit block will
271 // be updated specially after unrolling all the way.
272 if (*BB != LatchBlock)
273 for (Value::use_iterator UI = (*BB)->use_begin(), UE = (*BB)->use_end();
275 Instruction* UseInst = cast<Instruction>(*UI);
276 if (isa<PHINode>(UseInst) && !L->contains(UseInst->getParent())) {
277 PHINode* phi = cast<PHINode>(UseInst);
278 Value* Incoming = phi->getIncomingValueForBlock(*BB);
279 if (isa<Instruction>(Incoming))
280 Incoming = LastValueMap[Incoming];
282 phi->addIncoming(Incoming, New);
286 // Keep track of new headers and latches as we create them, so that
287 // we can insert the proper branches later.
289 Headers.push_back(New);
290 if (*BB == LatchBlock)
291 Latches.push_back(New);
293 NewBlocks.push_back(New);
296 // Remap all instructions in the most recent iteration
297 for (unsigned i = 0; i < NewBlocks.size(); ++i)
298 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
299 E = NewBlocks[i]->end(); I != E; ++I)
300 RemapInstruction(I, LastValueMap);
305 // Update PHI nodes that reference the final latch block
307 std::set<PHINode*> Users;
308 for (Value::use_iterator UI = LatchBlock->use_begin(),
309 UE = LatchBlock->use_end(); UI != UE; ++UI)
310 if (PHINode* phi = dyn_cast<PHINode>(*UI))
313 for (std::set<PHINode*>::iterator SI = Users.begin(), SE = Users.end();
315 Value* InVal = (*SI)->getIncomingValueForBlock(LatchBlock);
316 if (isa<Instruction>(InVal))
317 InVal = LastValueMap[InVal];
318 (*SI)->removeIncomingValue(LatchBlock, false);
320 (*SI)->addIncoming(InVal, cast<BasicBlock>(LastValueMap[LatchBlock]));
324 // Now loop over the PHI nodes in the original block, setting them to their
326 BasicBlock *Preheader = L->getLoopPreheader();
327 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
328 PHINode *PN = OrigPHINode[i];
329 PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
330 Header->getInstList().erase(PN);
333 // Insert the branches that link the different iterations together
334 for (unsigned i = 0; i < Latches.size()-1; ++i) {
335 new BranchInst(Headers[i+1], Latches[i]);
336 if(BasicBlock* Fold = FoldBlockIntoPredecessor(Headers[i+1])) {
337 std::replace(Latches.begin(), Latches.end(), Headers[i+1], Fold);
338 std::replace(Headers.begin(), Headers.end(), Headers[i+1], Fold);
342 // Finally, add an unconditional branch to the block to continue into the exit
344 new BranchInst(LoopExit, Latches[Latches.size()-1]);
345 FoldBlockIntoPredecessor(LoopExit);
347 // At this point, the code is well formed. We now do a quick sweep over the
348 // inserted code, doing constant propagation and dead code elimination as we
350 const std::vector<BasicBlock*> &NewLoopBlocks = L->getBlocks();
351 for (std::vector<BasicBlock*>::const_iterator BB = NewLoopBlocks.begin(),
352 BBE = NewLoopBlocks.end(); BB != BBE; ++BB)
353 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); I != E; ) {
354 Instruction *Inst = I++;
356 if (isInstructionTriviallyDead(Inst))
357 (*BB)->getInstList().erase(Inst);
358 else if (Constant *C = ConstantFoldInstruction(Inst)) {
359 Inst->replaceAllUsesWith(C);
360 (*BB)->getInstList().erase(Inst);
364 // Update the loop information for this loop.
365 Loop *Parent = L->getParentLoop();
367 // Move all of the basic blocks in the loop into the parent loop.
368 for (std::vector<BasicBlock*>::const_iterator BB = NewLoopBlocks.begin(),
369 E = NewLoopBlocks.end(); BB != E; ++BB)
370 LI->changeLoopFor(*BB, Parent);
372 // Remove the loop from the parent.
374 delete Parent->removeChildLoop(std::find(Parent->begin(), Parent->end(),L));
376 delete LI->removeLoop(std::find(LI->begin(), LI->end(), L));