Chris Lattner | e6bb649 | 2010-12-26 19:39:38 +0000 | [diff] [blame] | 1 | //===-- LoopIdiomRecognize.cpp - Loop idiom recognition -------------------===// |
| 2 | // |
| 3 | // The LLVM Compiler Infrastructure |
| 4 | // |
| 5 | // This file is distributed under the University of Illinois Open Source |
| 6 | // License. See LICENSE.TXT for details. |
| 7 | // |
| 8 | //===----------------------------------------------------------------------===// |
| 9 | // |
| 10 | // This pass implements an idiom recognizer that transforms simple loops into a |
| 11 | // non-loop form. In cases that this kicks in, it can be a significant |
| 12 | // performance win. |
| 13 | // |
| 14 | //===----------------------------------------------------------------------===// |
| 15 | |
| 16 | #define DEBUG_TYPE "loop-idiom" |
| 17 | #include "llvm/Transforms/Scalar.h" |
| 18 | #include "llvm/Analysis/LoopPass.h" |
| 19 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
Chris Lattner | a92ff91 | 2010-12-26 23:42:51 +0000 | [diff] [blame^] | 20 | #include "llvm/Analysis/ScalarEvolutionExpander.h" |
Chris Lattner | 22920b5 | 2010-12-26 20:45:45 +0000 | [diff] [blame] | 21 | #include "llvm/Analysis/ValueTracking.h" |
| 22 | #include "llvm/Target/TargetData.h" |
Chris Lattner | e6bb649 | 2010-12-26 19:39:38 +0000 | [diff] [blame] | 23 | #include "llvm/Support/Debug.h" |
Chris Lattner | a92ff91 | 2010-12-26 23:42:51 +0000 | [diff] [blame^] | 24 | #include "llvm/Support/IRBuilder.h" |
Chris Lattner | e6bb649 | 2010-12-26 19:39:38 +0000 | [diff] [blame] | 25 | #include "llvm/Support/raw_ostream.h" |
| 26 | using namespace llvm; |
| 27 | |
| 28 | // TODO: Recognize "N" size array multiplies: replace with call to blas or |
| 29 | // something. |
| 30 | |
| 31 | namespace { |
| 32 | class LoopIdiomRecognize : public LoopPass { |
Chris Lattner | 22920b5 | 2010-12-26 20:45:45 +0000 | [diff] [blame] | 33 | Loop *CurLoop; |
| 34 | const TargetData *TD; |
| 35 | ScalarEvolution *SE; |
Chris Lattner | e6bb649 | 2010-12-26 19:39:38 +0000 | [diff] [blame] | 36 | public: |
| 37 | static char ID; |
| 38 | explicit LoopIdiomRecognize() : LoopPass(ID) { |
| 39 | initializeLoopIdiomRecognizePass(*PassRegistry::getPassRegistry()); |
| 40 | } |
| 41 | |
| 42 | bool runOnLoop(Loop *L, LPPassManager &LPM); |
| 43 | |
Chris Lattner | 22920b5 | 2010-12-26 20:45:45 +0000 | [diff] [blame] | 44 | bool processLoopStore(StoreInst *SI, const SCEV *BECount); |
Chris Lattner | e6bb649 | 2010-12-26 19:39:38 +0000 | [diff] [blame] | 45 | |
Chris Lattner | a92ff91 | 2010-12-26 23:42:51 +0000 | [diff] [blame^] | 46 | bool processLoopStoreOfSplatValue(StoreInst *SI, unsigned StoreSize, |
| 47 | Value *SplatValue, |
| 48 | const SCEVAddRecExpr *Ev, |
| 49 | const SCEV *BECount); |
| 50 | |
Chris Lattner | e6bb649 | 2010-12-26 19:39:38 +0000 | [diff] [blame] | 51 | /// This transformation requires natural loop information & requires that |
| 52 | /// loop preheaders be inserted into the CFG. |
| 53 | /// |
| 54 | virtual void getAnalysisUsage(AnalysisUsage &AU) const { |
| 55 | AU.addRequired<LoopInfo>(); |
| 56 | AU.addPreserved<LoopInfo>(); |
| 57 | AU.addRequiredID(LoopSimplifyID); |
| 58 | AU.addPreservedID(LoopSimplifyID); |
| 59 | AU.addRequiredID(LCSSAID); |
| 60 | AU.addPreservedID(LCSSAID); |
| 61 | AU.addRequired<ScalarEvolution>(); |
| 62 | AU.addPreserved<ScalarEvolution>(); |
| 63 | AU.addPreserved<DominatorTree>(); |
| 64 | } |
| 65 | }; |
| 66 | } |
| 67 | |
| 68 | char LoopIdiomRecognize::ID = 0; |
| 69 | INITIALIZE_PASS_BEGIN(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms", |
| 70 | false, false) |
| 71 | INITIALIZE_PASS_DEPENDENCY(LoopInfo) |
| 72 | INITIALIZE_PASS_DEPENDENCY(LoopSimplify) |
| 73 | INITIALIZE_PASS_DEPENDENCY(LCSSA) |
| 74 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) |
| 75 | INITIALIZE_PASS_END(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms", |
| 76 | false, false) |
| 77 | |
| 78 | Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognize(); } |
| 79 | |
| 80 | bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) { |
Chris Lattner | 22920b5 | 2010-12-26 20:45:45 +0000 | [diff] [blame] | 81 | CurLoop = L; |
| 82 | |
Chris Lattner | e6bb649 | 2010-12-26 19:39:38 +0000 | [diff] [blame] | 83 | // We only look at trivial single basic block loops. |
| 84 | // TODO: eventually support more complex loops, scanning the header. |
| 85 | if (L->getBlocks().size() != 1) |
| 86 | return false; |
| 87 | |
Chris Lattner | 22920b5 | 2010-12-26 20:45:45 +0000 | [diff] [blame] | 88 | // The trip count of the loop must be analyzable. |
| 89 | SE = &getAnalysis<ScalarEvolution>(); |
| 90 | if (!SE->hasLoopInvariantBackedgeTakenCount(L)) |
| 91 | return false; |
| 92 | const SCEV *BECount = SE->getBackedgeTakenCount(L); |
| 93 | if (isa<SCEVCouldNotCompute>(BECount)) return false; |
| 94 | |
| 95 | // We require target data for now. |
| 96 | TD = getAnalysisIfAvailable<TargetData>(); |
| 97 | if (TD == 0) return false; |
| 98 | |
Chris Lattner | e6bb649 | 2010-12-26 19:39:38 +0000 | [diff] [blame] | 99 | BasicBlock *BB = L->getHeader(); |
Chris Lattner | 22920b5 | 2010-12-26 20:45:45 +0000 | [diff] [blame] | 100 | DEBUG(dbgs() << "loop-idiom Scanning: F[" << BB->getParent()->getName() |
Chris Lattner | e6bb649 | 2010-12-26 19:39:38 +0000 | [diff] [blame] | 101 | << "] Loop %" << BB->getName() << "\n"); |
| 102 | |
Chris Lattner | 22920b5 | 2010-12-26 20:45:45 +0000 | [diff] [blame] | 103 | bool MadeChange = false; |
| 104 | for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) { |
| 105 | // Look for store instructions, which may be memsets. |
Chris Lattner | a92ff91 | 2010-12-26 23:42:51 +0000 | [diff] [blame^] | 106 | StoreInst *SI = dyn_cast<StoreInst>(I++); |
| 107 | if (SI == 0 || SI->isVolatile()) continue; |
| 108 | |
| 109 | |
| 110 | MadeChange |= processLoopStore(SI, BECount); |
Chris Lattner | 22920b5 | 2010-12-26 20:45:45 +0000 | [diff] [blame] | 111 | } |
| 112 | |
| 113 | return MadeChange; |
Chris Lattner | e6bb649 | 2010-12-26 19:39:38 +0000 | [diff] [blame] | 114 | } |
| 115 | |
| 116 | /// scanBlock - Look over a block to see if we can promote anything out of it. |
Chris Lattner | 22920b5 | 2010-12-26 20:45:45 +0000 | [diff] [blame] | 117 | bool LoopIdiomRecognize::processLoopStore(StoreInst *SI, const SCEV *BECount) { |
| 118 | Value *StoredVal = SI->getValueOperand(); |
Chris Lattner | a92ff91 | 2010-12-26 23:42:51 +0000 | [diff] [blame^] | 119 | Value *StorePtr = SI->getPointerOperand(); |
Chris Lattner | e6bb649 | 2010-12-26 19:39:38 +0000 | [diff] [blame] | 120 | |
Chris Lattner | 22920b5 | 2010-12-26 20:45:45 +0000 | [diff] [blame] | 121 | // Check to see if the store updates all bits in memory. We don't want to |
| 122 | // process things like a store of i3. We also require that the store be a |
| 123 | // multiple of a byte. |
| 124 | uint64_t SizeInBits = TD->getTypeSizeInBits(StoredVal->getType()); |
| 125 | if ((SizeInBits & 7) || (SizeInBits >> 32) != 0 || |
| 126 | SizeInBits != TD->getTypeStoreSizeInBits(StoredVal->getType())) |
| 127 | return false; |
Chris Lattner | e6bb649 | 2010-12-26 19:39:38 +0000 | [diff] [blame] | 128 | |
Chris Lattner | 22920b5 | 2010-12-26 20:45:45 +0000 | [diff] [blame] | 129 | // See if the pointer expression is an AddRec like {base,+,1} on the current |
| 130 | // loop, which indicates a strided store. If we have something else, it's a |
| 131 | // random store we can't handle. |
Chris Lattner | a92ff91 | 2010-12-26 23:42:51 +0000 | [diff] [blame^] | 132 | const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr)); |
Chris Lattner | 22920b5 | 2010-12-26 20:45:45 +0000 | [diff] [blame] | 133 | if (Ev == 0 || Ev->getLoop() != CurLoop || !Ev->isAffine()) |
| 134 | return false; |
| 135 | |
| 136 | // Check to see if the stride matches the size of the store. If so, then we |
| 137 | // know that every byte is touched in the loop. |
| 138 | unsigned StoreSize = (unsigned)SizeInBits >> 3; |
| 139 | const SCEVConstant *Stride = dyn_cast<SCEVConstant>(Ev->getOperand(1)); |
| 140 | if (Stride == 0 || StoreSize != Stride->getValue()->getValue()) |
| 141 | return false; |
| 142 | |
Chris Lattner | 22920b5 | 2010-12-26 20:45:45 +0000 | [diff] [blame] | 143 | // If the stored value is a byte-wise value (like i32 -1), then it may be |
| 144 | // turned into a memset of i8 -1, assuming that all the consequtive bytes |
| 145 | // are stored. A store of i32 0x01020304 can never be turned into a memset. |
Chris Lattner | a92ff91 | 2010-12-26 23:42:51 +0000 | [diff] [blame^] | 146 | if (Value *SplatValue = isBytewiseValue(StoredVal)) |
| 147 | return processLoopStoreOfSplatValue(SI, StoreSize, SplatValue, Ev, BECount); |
| 148 | |
| 149 | // Handle the memcpy case here. |
| 150 | errs() << "Found strided store: " << *Ev << "\n"; |
Chris Lattner | 22920b5 | 2010-12-26 20:45:45 +0000 | [diff] [blame] | 151 | |
| 152 | |
Chris Lattner | e6bb649 | 2010-12-26 19:39:38 +0000 | [diff] [blame] | 153 | return false; |
| 154 | } |
| 155 | |
Chris Lattner | a92ff91 | 2010-12-26 23:42:51 +0000 | [diff] [blame^] | 156 | /// processLoopStoreOfSplatValue - We see a strided store of a memsetable value. |
| 157 | /// If we can transform this into a memset in the loop preheader, do so. |
| 158 | bool LoopIdiomRecognize:: |
| 159 | processLoopStoreOfSplatValue(StoreInst *SI, unsigned StoreSize, |
| 160 | Value *SplatValue, |
| 161 | const SCEVAddRecExpr *Ev, const SCEV *BECount) { |
| 162 | // Okay, we have a strided store "p[i]" of a splattable value. We can turn |
| 163 | // this into a memset in the loop preheader now if we want. However, this |
| 164 | // would be unsafe to do if there is anything else in the loop that may read |
| 165 | // or write to the aliased location. Check for an alias. |
| 166 | |
| 167 | // FIXME: TODO safety check. |
| 168 | |
| 169 | // Okay, everything looks good, insert the memset. |
| 170 | BasicBlock *Preheader = CurLoop->getLoopPreheader(); |
| 171 | |
| 172 | IRBuilder<> Builder(Preheader->getTerminator()); |
| 173 | |
| 174 | // The trip count of the loop and the base pointer of the addrec SCEV is |
| 175 | // guaranteed to be loop invariant, which means that it should dominate the |
| 176 | // header. Just insert code for it in the preheader. |
| 177 | SCEVExpander Expander(*SE); |
| 178 | |
| 179 | unsigned AddrSpace = SI->getPointerAddressSpace(); |
| 180 | Value *BasePtr = |
| 181 | Expander.expandCodeFor(Ev->getStart(), Builder.getInt8PtrTy(AddrSpace), |
| 182 | Preheader->getTerminator()); |
| 183 | |
| 184 | // The # stored bytes is (BECount+1)*Size. Expand the trip count out to |
| 185 | // pointer size if it isn't already. |
| 186 | const Type *IntPtr = TD->getIntPtrType(SI->getContext()); |
| 187 | unsigned BESize = SE->getTypeSizeInBits(BECount->getType()); |
| 188 | if (BESize < TD->getPointerSizeInBits()) |
| 189 | BECount = SE->getZeroExtendExpr(BECount, IntPtr); |
| 190 | else if (BESize > TD->getPointerSizeInBits()) |
| 191 | BECount = SE->getTruncateExpr(BECount, IntPtr); |
| 192 | |
| 193 | const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1), |
| 194 | true, true /*nooverflow*/); |
| 195 | if (StoreSize != 1) |
| 196 | NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize), |
| 197 | true, true /*nooverflow*/); |
| 198 | |
| 199 | Value *NumBytes = |
| 200 | Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator()); |
| 201 | |
| 202 | Value *NewCall = |
| 203 | Builder.CreateMemSet(BasePtr, SplatValue, NumBytes, SI->getAlignment()); |
| 204 | |
| 205 | DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n" |
| 206 | << " from store to: " << *Ev << " at: " << *SI << "\n"); |
| 207 | |
| 208 | // Okay, the memset has been formed. Zap the original store. |
| 209 | // FIXME: We want to recursively delete dead instructions, but we have to |
| 210 | // update SCEV. |
| 211 | SI->eraseFromParent(); |
| 212 | return true; |
| 213 | } |
| 214 | |