LLVM API Documentation
00001 //===- ScalarEvolutionExpander.cpp - Scalar Evolution Analysis --*- C++ -*-===// 00002 // 00003 // The LLVM Compiler Infrastructure 00004 // 00005 // This file is distributed under the University of Illinois Open Source 00006 // License. See LICENSE.TXT for details. 00007 // 00008 //===----------------------------------------------------------------------===// 00009 // 00010 // This file contains the implementation of the scalar evolution expander, 00011 // which is used to generate the code corresponding to a given scalar evolution 00012 // expression. 00013 // 00014 //===----------------------------------------------------------------------===// 00015 00016 #include "llvm/Analysis/ScalarEvolutionExpander.h" 00017 #include "llvm/ADT/STLExtras.h" 00018 #include "llvm/Analysis/LoopInfo.h" 00019 #include "llvm/Analysis/TargetTransformInfo.h" 00020 #include "llvm/IR/DataLayout.h" 00021 #include "llvm/IR/IntrinsicInst.h" 00022 #include "llvm/IR/LLVMContext.h" 00023 #include "llvm/Support/Debug.h" 00024 00025 using namespace llvm; 00026 00027 /// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP, 00028 /// reusing an existing cast if a suitable one exists, moving an existing 00029 /// cast if a suitable one exists but isn't in the right place, or 00030 /// creating a new one. 00031 Value *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty, 00032 Instruction::CastOps Op, 00033 BasicBlock::iterator IP) { 00034 // This function must be called with the builder having a valid insertion 00035 // point. It doesn't need to be the actual IP where the uses of the returned 00036 // cast will be added, but it must dominate such IP. 00037 // We use this precondition to produce a cast that will dominate all its 00038 // uses. In particular, this is crucial for the case where the builder's 00039 // insertion point *is* the point where we were asked to put the cast. 00040 // Since we don't know the builder's insertion point is actually 00041 // where the uses will be added (only that it dominates it), we are 00042 // not allowed to move it. 00043 BasicBlock::iterator BIP = Builder.GetInsertPoint(); 00044 00045 Instruction *Ret = NULL; 00046 00047 // Check to see if there is already a cast! 00048 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); 00049 UI != E; ++UI) { 00050 User *U = *UI; 00051 if (U->getType() == Ty) 00052 if (CastInst *CI = dyn_cast<CastInst>(U)) 00053 if (CI->getOpcode() == Op) { 00054 // If the cast isn't where we want it, create a new cast at IP. 00055 // Likewise, do not reuse a cast at BIP because it must dominate 00056 // instructions that might be inserted before BIP. 00057 if (BasicBlock::iterator(CI) != IP || BIP == IP) { 00058 // Create a new cast, and leave the old cast in place in case 00059 // it is being used as an insert point. Clear its operand 00060 // so that it doesn't hold anything live. 00061 Ret = CastInst::Create(Op, V, Ty, "", IP); 00062 Ret->takeName(CI); 00063 CI->replaceAllUsesWith(Ret); 00064 CI->setOperand(0, UndefValue::get(V->getType())); 00065 break; 00066 } 00067 Ret = CI; 00068 break; 00069 } 00070 } 00071 00072 // Create a new cast. 00073 if (!Ret) 00074 Ret = CastInst::Create(Op, V, Ty, V->getName(), IP); 00075 00076 // We assert at the end of the function since IP might point to an 00077 // instruction with different dominance properties than a cast 00078 // (an invoke for example) and not dominate BIP (but the cast does). 00079 assert(SE.DT->dominates(Ret, BIP)); 00080 00081 rememberInstruction(Ret); 00082 return Ret; 00083 } 00084 00085 /// InsertNoopCastOfTo - Insert a cast of V to the specified type, 00086 /// which must be possible with a noop cast, doing what we can to share 00087 /// the casts. 00088 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) { 00089 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false); 00090 assert((Op == Instruction::BitCast || 00091 Op == Instruction::PtrToInt || 00092 Op == Instruction::IntToPtr) && 00093 "InsertNoopCastOfTo cannot perform non-noop casts!"); 00094 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) && 00095 "InsertNoopCastOfTo cannot change sizes!"); 00096 00097 // Short-circuit unnecessary bitcasts. 00098 if (Op == Instruction::BitCast) { 00099 if (V->getType() == Ty) 00100 return V; 00101 if (CastInst *CI = dyn_cast<CastInst>(V)) { 00102 if (CI->getOperand(0)->getType() == Ty) 00103 return CI->getOperand(0); 00104 } 00105 } 00106 // Short-circuit unnecessary inttoptr<->ptrtoint casts. 00107 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) && 00108 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) { 00109 if (CastInst *CI = dyn_cast<CastInst>(V)) 00110 if ((CI->getOpcode() == Instruction::PtrToInt || 00111 CI->getOpcode() == Instruction::IntToPtr) && 00112 SE.getTypeSizeInBits(CI->getType()) == 00113 SE.getTypeSizeInBits(CI->getOperand(0)->getType())) 00114 return CI->getOperand(0); 00115 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 00116 if ((CE->getOpcode() == Instruction::PtrToInt || 00117 CE->getOpcode() == Instruction::IntToPtr) && 00118 SE.getTypeSizeInBits(CE->getType()) == 00119 SE.getTypeSizeInBits(CE->getOperand(0)->getType())) 00120 return CE->getOperand(0); 00121 } 00122 00123 // Fold a cast of a constant. 00124 if (Constant *C = dyn_cast<Constant>(V)) 00125 return ConstantExpr::getCast(Op, C, Ty); 00126 00127 // Cast the argument at the beginning of the entry block, after 00128 // any bitcasts of other arguments. 00129 if (Argument *A = dyn_cast<Argument>(V)) { 00130 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin(); 00131 while ((isa<BitCastInst>(IP) && 00132 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) && 00133 cast<BitCastInst>(IP)->getOperand(0) != A) || 00134 isa<DbgInfoIntrinsic>(IP) || 00135 isa<LandingPadInst>(IP)) 00136 ++IP; 00137 return ReuseOrCreateCast(A, Ty, Op, IP); 00138 } 00139 00140 // Cast the instruction immediately after the instruction. 00141 Instruction *I = cast<Instruction>(V); 00142 BasicBlock::iterator IP = I; ++IP; 00143 if (InvokeInst *II = dyn_cast<InvokeInst>(I)) 00144 IP = II->getNormalDest()->begin(); 00145 while (isa<PHINode>(IP) || isa<LandingPadInst>(IP)) 00146 ++IP; 00147 return ReuseOrCreateCast(I, Ty, Op, IP); 00148 } 00149 00150 /// InsertBinop - Insert the specified binary operator, doing a small amount 00151 /// of work to avoid inserting an obviously redundant operation. 00152 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, 00153 Value *LHS, Value *RHS) { 00154 // Fold a binop with constant operands. 00155 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 00156 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 00157 return ConstantExpr::get(Opcode, CLHS, CRHS); 00158 00159 // Do a quick scan to see if we have this binop nearby. If so, reuse it. 00160 unsigned ScanLimit = 6; 00161 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 00162 // Scanning starts from the last instruction before the insertion point. 00163 BasicBlock::iterator IP = Builder.GetInsertPoint(); 00164 if (IP != BlockBegin) { 00165 --IP; 00166 for (; ScanLimit; --IP, --ScanLimit) { 00167 // Don't count dbg.value against the ScanLimit, to avoid perturbing the 00168 // generated code. 00169 if (isa<DbgInfoIntrinsic>(IP)) 00170 ScanLimit++; 00171 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS && 00172 IP->getOperand(1) == RHS) 00173 return IP; 00174 if (IP == BlockBegin) break; 00175 } 00176 } 00177 00178 // Save the original insertion point so we can restore it when we're done. 00179 BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 00180 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 00181 00182 // Move the insertion point out of as many loops as we can. 00183 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 00184 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break; 00185 BasicBlock *Preheader = L->getLoopPreheader(); 00186 if (!Preheader) break; 00187 00188 // Ok, move up a level. 00189 Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 00190 } 00191 00192 // If we haven't found this binop, insert it. 00193 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS)); 00194 BO->setDebugLoc(SaveInsertPt->getDebugLoc()); 00195 rememberInstruction(BO); 00196 00197 // Restore the original insert point. 00198 if (SaveInsertBB) 00199 restoreInsertPoint(SaveInsertBB, SaveInsertPt); 00200 00201 return BO; 00202 } 00203 00204 /// FactorOutConstant - Test if S is divisible by Factor, using signed 00205 /// division. If so, update S with Factor divided out and return true. 00206 /// S need not be evenly divisible if a reasonable remainder can be 00207 /// computed. 00208 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made 00209 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and 00210 /// check to see if the divide was folded. 00211 static bool FactorOutConstant(const SCEV *&S, 00212 const SCEV *&Remainder, 00213 const SCEV *Factor, 00214 ScalarEvolution &SE, 00215 const DataLayout *TD) { 00216 // Everything is divisible by one. 00217 if (Factor->isOne()) 00218 return true; 00219 00220 // x/x == 1. 00221 if (S == Factor) { 00222 S = SE.getConstant(S->getType(), 1); 00223 return true; 00224 } 00225 00226 // For a Constant, check for a multiple of the given factor. 00227 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { 00228 // 0/x == 0. 00229 if (C->isZero()) 00230 return true; 00231 // Check for divisibility. 00232 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) { 00233 ConstantInt *CI = 00234 ConstantInt::get(SE.getContext(), 00235 C->getValue()->getValue().sdiv( 00236 FC->getValue()->getValue())); 00237 // If the quotient is zero and the remainder is non-zero, reject 00238 // the value at this scale. It will be considered for subsequent 00239 // smaller scales. 00240 if (!CI->isZero()) { 00241 const SCEV *Div = SE.getConstant(CI); 00242 S = Div; 00243 Remainder = 00244 SE.getAddExpr(Remainder, 00245 SE.getConstant(C->getValue()->getValue().srem( 00246 FC->getValue()->getValue()))); 00247 return true; 00248 } 00249 } 00250 } 00251 00252 // In a Mul, check if there is a constant operand which is a multiple 00253 // of the given factor. 00254 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) { 00255 if (TD) { 00256 // With DataLayout, the size is known. Check if there is a constant 00257 // operand which is a multiple of the given factor. If so, we can 00258 // factor it. 00259 const SCEVConstant *FC = cast<SCEVConstant>(Factor); 00260 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0))) 00261 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) { 00262 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end()); 00263 NewMulOps[0] = 00264 SE.getConstant(C->getValue()->getValue().sdiv( 00265 FC->getValue()->getValue())); 00266 S = SE.getMulExpr(NewMulOps); 00267 return true; 00268 } 00269 } else { 00270 // Without DataLayout, check if Factor can be factored out of any of the 00271 // Mul's operands. If so, we can just remove it. 00272 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) { 00273 const SCEV *SOp = M->getOperand(i); 00274 const SCEV *Remainder = SE.getConstant(SOp->getType(), 0); 00275 if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) && 00276 Remainder->isZero()) { 00277 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end()); 00278 NewMulOps[i] = SOp; 00279 S = SE.getMulExpr(NewMulOps); 00280 return true; 00281 } 00282 } 00283 } 00284 } 00285 00286 // In an AddRec, check if both start and step are divisible. 00287 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) { 00288 const SCEV *Step = A->getStepRecurrence(SE); 00289 const SCEV *StepRem = SE.getConstant(Step->getType(), 0); 00290 if (!FactorOutConstant(Step, StepRem, Factor, SE, TD)) 00291 return false; 00292 if (!StepRem->isZero()) 00293 return false; 00294 const SCEV *Start = A->getStart(); 00295 if (!FactorOutConstant(Start, Remainder, Factor, SE, TD)) 00296 return false; 00297 // FIXME: can use A->getNoWrapFlags(FlagNW) 00298 S = SE.getAddRecExpr(Start, Step, A->getLoop(), SCEV::FlagAnyWrap); 00299 return true; 00300 } 00301 00302 return false; 00303 } 00304 00305 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs 00306 /// is the number of SCEVAddRecExprs present, which are kept at the end of 00307 /// the list. 00308 /// 00309 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops, 00310 Type *Ty, 00311 ScalarEvolution &SE) { 00312 unsigned NumAddRecs = 0; 00313 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i) 00314 ++NumAddRecs; 00315 // Group Ops into non-addrecs and addrecs. 00316 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs); 00317 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end()); 00318 // Let ScalarEvolution sort and simplify the non-addrecs list. 00319 const SCEV *Sum = NoAddRecs.empty() ? 00320 SE.getConstant(Ty, 0) : 00321 SE.getAddExpr(NoAddRecs); 00322 // If it returned an add, use the operands. Otherwise it simplified 00323 // the sum into a single value, so just use that. 00324 Ops.clear(); 00325 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum)) 00326 Ops.append(Add->op_begin(), Add->op_end()); 00327 else if (!Sum->isZero()) 00328 Ops.push_back(Sum); 00329 // Then append the addrecs. 00330 Ops.append(AddRecs.begin(), AddRecs.end()); 00331 } 00332 00333 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values 00334 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}. 00335 /// This helps expose more opportunities for folding parts of the expressions 00336 /// into GEP indices. 00337 /// 00338 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops, 00339 Type *Ty, 00340 ScalarEvolution &SE) { 00341 // Find the addrecs. 00342 SmallVector<const SCEV *, 8> AddRecs; 00343 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 00344 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) { 00345 const SCEV *Start = A->getStart(); 00346 if (Start->isZero()) break; 00347 const SCEV *Zero = SE.getConstant(Ty, 0); 00348 AddRecs.push_back(SE.getAddRecExpr(Zero, 00349 A->getStepRecurrence(SE), 00350 A->getLoop(), 00351 // FIXME: A->getNoWrapFlags(FlagNW) 00352 SCEV::FlagAnyWrap)); 00353 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) { 00354 Ops[i] = Zero; 00355 Ops.append(Add->op_begin(), Add->op_end()); 00356 e += Add->getNumOperands(); 00357 } else { 00358 Ops[i] = Start; 00359 } 00360 } 00361 if (!AddRecs.empty()) { 00362 // Add the addrecs onto the end of the list. 00363 Ops.append(AddRecs.begin(), AddRecs.end()); 00364 // Resort the operand list, moving any constants to the front. 00365 SimplifyAddOperands(Ops, Ty, SE); 00366 } 00367 } 00368 00369 /// expandAddToGEP - Expand an addition expression with a pointer type into 00370 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps 00371 /// BasicAliasAnalysis and other passes analyze the result. See the rules 00372 /// for getelementptr vs. inttoptr in 00373 /// http://llvm.org/docs/LangRef.html#pointeraliasing 00374 /// for details. 00375 /// 00376 /// Design note: The correctness of using getelementptr here depends on 00377 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as 00378 /// they may introduce pointer arithmetic which may not be safely converted 00379 /// into getelementptr. 00380 /// 00381 /// Design note: It might seem desirable for this function to be more 00382 /// loop-aware. If some of the indices are loop-invariant while others 00383 /// aren't, it might seem desirable to emit multiple GEPs, keeping the 00384 /// loop-invariant portions of the overall computation outside the loop. 00385 /// However, there are a few reasons this is not done here. Hoisting simple 00386 /// arithmetic is a low-level optimization that often isn't very 00387 /// important until late in the optimization process. In fact, passes 00388 /// like InstructionCombining will combine GEPs, even if it means 00389 /// pushing loop-invariant computation down into loops, so even if the 00390 /// GEPs were split here, the work would quickly be undone. The 00391 /// LoopStrengthReduction pass, which is usually run quite late (and 00392 /// after the last InstructionCombining pass), takes care of hoisting 00393 /// loop-invariant portions of expressions, after considering what 00394 /// can be folded using target addressing modes. 00395 /// 00396 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin, 00397 const SCEV *const *op_end, 00398 PointerType *PTy, 00399 Type *Ty, 00400 Value *V) { 00401 Type *ElTy = PTy->getElementType(); 00402 SmallVector<Value *, 4> GepIndices; 00403 SmallVector<const SCEV *, 8> Ops(op_begin, op_end); 00404 bool AnyNonZeroIndices = false; 00405 00406 // Split AddRecs up into parts as either of the parts may be usable 00407 // without the other. 00408 SplitAddRecs(Ops, Ty, SE); 00409 00410 // Descend down the pointer's type and attempt to convert the other 00411 // operands into GEP indices, at each level. The first index in a GEP 00412 // indexes into the array implied by the pointer operand; the rest of 00413 // the indices index into the element or field type selected by the 00414 // preceding index. 00415 for (;;) { 00416 // If the scale size is not 0, attempt to factor out a scale for 00417 // array indexing. 00418 SmallVector<const SCEV *, 8> ScaledOps; 00419 if (ElTy->isSized()) { 00420 const SCEV *ElSize = SE.getSizeOfExpr(ElTy); 00421 if (!ElSize->isZero()) { 00422 SmallVector<const SCEV *, 8> NewOps; 00423 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 00424 const SCEV *Op = Ops[i]; 00425 const SCEV *Remainder = SE.getConstant(Ty, 0); 00426 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) { 00427 // Op now has ElSize factored out. 00428 ScaledOps.push_back(Op); 00429 if (!Remainder->isZero()) 00430 NewOps.push_back(Remainder); 00431 AnyNonZeroIndices = true; 00432 } else { 00433 // The operand was not divisible, so add it to the list of operands 00434 // we'll scan next iteration. 00435 NewOps.push_back(Ops[i]); 00436 } 00437 } 00438 // If we made any changes, update Ops. 00439 if (!ScaledOps.empty()) { 00440 Ops = NewOps; 00441 SimplifyAddOperands(Ops, Ty, SE); 00442 } 00443 } 00444 } 00445 00446 // Record the scaled array index for this level of the type. If 00447 // we didn't find any operands that could be factored, tentatively 00448 // assume that element zero was selected (since the zero offset 00449 // would obviously be folded away). 00450 Value *Scaled = ScaledOps.empty() ? 00451 Constant::getNullValue(Ty) : 00452 expandCodeFor(SE.getAddExpr(ScaledOps), Ty); 00453 GepIndices.push_back(Scaled); 00454 00455 // Collect struct field index operands. 00456 while (StructType *STy = dyn_cast<StructType>(ElTy)) { 00457 bool FoundFieldNo = false; 00458 // An empty struct has no fields. 00459 if (STy->getNumElements() == 0) break; 00460 if (SE.TD) { 00461 // With DataLayout, field offsets are known. See if a constant offset 00462 // falls within any of the struct fields. 00463 if (Ops.empty()) break; 00464 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0])) 00465 if (SE.getTypeSizeInBits(C->getType()) <= 64) { 00466 const StructLayout &SL = *SE.TD->getStructLayout(STy); 00467 uint64_t FullOffset = C->getValue()->getZExtValue(); 00468 if (FullOffset < SL.getSizeInBytes()) { 00469 unsigned ElIdx = SL.getElementContainingOffset(FullOffset); 00470 GepIndices.push_back( 00471 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx)); 00472 ElTy = STy->getTypeAtIndex(ElIdx); 00473 Ops[0] = 00474 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx)); 00475 AnyNonZeroIndices = true; 00476 FoundFieldNo = true; 00477 } 00478 } 00479 } else { 00480 // Without DataLayout, just check for an offsetof expression of the 00481 // appropriate struct type. 00482 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 00483 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) { 00484 Type *CTy; 00485 Constant *FieldNo; 00486 if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) { 00487 GepIndices.push_back(FieldNo); 00488 ElTy = 00489 STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue()); 00490 Ops[i] = SE.getConstant(Ty, 0); 00491 AnyNonZeroIndices = true; 00492 FoundFieldNo = true; 00493 break; 00494 } 00495 } 00496 } 00497 // If no struct field offsets were found, tentatively assume that 00498 // field zero was selected (since the zero offset would obviously 00499 // be folded away). 00500 if (!FoundFieldNo) { 00501 ElTy = STy->getTypeAtIndex(0u); 00502 GepIndices.push_back( 00503 Constant::getNullValue(Type::getInt32Ty(Ty->getContext()))); 00504 } 00505 } 00506 00507 if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy)) 00508 ElTy = ATy->getElementType(); 00509 else 00510 break; 00511 } 00512 00513 // If none of the operands were convertible to proper GEP indices, cast 00514 // the base to i8* and do an ugly getelementptr with that. It's still 00515 // better than ptrtoint+arithmetic+inttoptr at least. 00516 if (!AnyNonZeroIndices) { 00517 // Cast the base to i8*. 00518 V = InsertNoopCastOfTo(V, 00519 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace())); 00520 00521 assert(!isa<Instruction>(V) || 00522 SE.DT->dominates(cast<Instruction>(V), Builder.GetInsertPoint())); 00523 00524 // Expand the operands for a plain byte offset. 00525 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty); 00526 00527 // Fold a GEP with constant operands. 00528 if (Constant *CLHS = dyn_cast<Constant>(V)) 00529 if (Constant *CRHS = dyn_cast<Constant>(Idx)) 00530 return ConstantExpr::getGetElementPtr(CLHS, CRHS); 00531 00532 // Do a quick scan to see if we have this GEP nearby. If so, reuse it. 00533 unsigned ScanLimit = 6; 00534 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 00535 // Scanning starts from the last instruction before the insertion point. 00536 BasicBlock::iterator IP = Builder.GetInsertPoint(); 00537 if (IP != BlockBegin) { 00538 --IP; 00539 for (; ScanLimit; --IP, --ScanLimit) { 00540 // Don't count dbg.value against the ScanLimit, to avoid perturbing the 00541 // generated code. 00542 if (isa<DbgInfoIntrinsic>(IP)) 00543 ScanLimit++; 00544 if (IP->getOpcode() == Instruction::GetElementPtr && 00545 IP->getOperand(0) == V && IP->getOperand(1) == Idx) 00546 return IP; 00547 if (IP == BlockBegin) break; 00548 } 00549 } 00550 00551 // Save the original insertion point so we can restore it when we're done. 00552 BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 00553 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 00554 00555 // Move the insertion point out of as many loops as we can. 00556 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 00557 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break; 00558 BasicBlock *Preheader = L->getLoopPreheader(); 00559 if (!Preheader) break; 00560 00561 // Ok, move up a level. 00562 Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 00563 } 00564 00565 // Emit a GEP. 00566 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep"); 00567 rememberInstruction(GEP); 00568 00569 // Restore the original insert point. 00570 if (SaveInsertBB) 00571 restoreInsertPoint(SaveInsertBB, SaveInsertPt); 00572 00573 return GEP; 00574 } 00575 00576 // Save the original insertion point so we can restore it when we're done. 00577 BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 00578 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 00579 00580 // Move the insertion point out of as many loops as we can. 00581 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 00582 if (!L->isLoopInvariant(V)) break; 00583 00584 bool AnyIndexNotLoopInvariant = false; 00585 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(), 00586 E = GepIndices.end(); I != E; ++I) 00587 if (!L->isLoopInvariant(*I)) { 00588 AnyIndexNotLoopInvariant = true; 00589 break; 00590 } 00591 if (AnyIndexNotLoopInvariant) 00592 break; 00593 00594 BasicBlock *Preheader = L->getLoopPreheader(); 00595 if (!Preheader) break; 00596 00597 // Ok, move up a level. 00598 Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 00599 } 00600 00601 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds, 00602 // because ScalarEvolution may have changed the address arithmetic to 00603 // compute a value which is beyond the end of the allocated object. 00604 Value *Casted = V; 00605 if (V->getType() != PTy) 00606 Casted = InsertNoopCastOfTo(Casted, PTy); 00607 Value *GEP = Builder.CreateGEP(Casted, 00608 GepIndices, 00609 "scevgep"); 00610 Ops.push_back(SE.getUnknown(GEP)); 00611 rememberInstruction(GEP); 00612 00613 // Restore the original insert point. 00614 if (SaveInsertBB) 00615 restoreInsertPoint(SaveInsertBB, SaveInsertPt); 00616 00617 return expand(SE.getAddExpr(Ops)); 00618 } 00619 00620 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for 00621 /// SCEV expansion. If they are nested, this is the most nested. If they are 00622 /// neighboring, pick the later. 00623 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B, 00624 DominatorTree &DT) { 00625 if (!A) return B; 00626 if (!B) return A; 00627 if (A->contains(B)) return B; 00628 if (B->contains(A)) return A; 00629 if (DT.dominates(A->getHeader(), B->getHeader())) return B; 00630 if (DT.dominates(B->getHeader(), A->getHeader())) return A; 00631 return A; // Arbitrarily break the tie. 00632 } 00633 00634 /// getRelevantLoop - Get the most relevant loop associated with the given 00635 /// expression, according to PickMostRelevantLoop. 00636 const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) { 00637 // Test whether we've already computed the most relevant loop for this SCEV. 00638 std::pair<DenseMap<const SCEV *, const Loop *>::iterator, bool> Pair = 00639 RelevantLoops.insert(std::make_pair(S, static_cast<const Loop *>(0))); 00640 if (!Pair.second) 00641 return Pair.first->second; 00642 00643 if (isa<SCEVConstant>(S)) 00644 // A constant has no relevant loops. 00645 return 0; 00646 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 00647 if (const Instruction *I = dyn_cast<Instruction>(U->getValue())) 00648 return Pair.first->second = SE.LI->getLoopFor(I->getParent()); 00649 // A non-instruction has no relevant loops. 00650 return 0; 00651 } 00652 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) { 00653 const Loop *L = 0; 00654 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) 00655 L = AR->getLoop(); 00656 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end(); 00657 I != E; ++I) 00658 L = PickMostRelevantLoop(L, getRelevantLoop(*I), *SE.DT); 00659 return RelevantLoops[N] = L; 00660 } 00661 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) { 00662 const Loop *Result = getRelevantLoop(C->getOperand()); 00663 return RelevantLoops[C] = Result; 00664 } 00665 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 00666 const Loop *Result = 00667 PickMostRelevantLoop(getRelevantLoop(D->getLHS()), 00668 getRelevantLoop(D->getRHS()), 00669 *SE.DT); 00670 return RelevantLoops[D] = Result; 00671 } 00672 llvm_unreachable("Unexpected SCEV type!"); 00673 } 00674 00675 namespace { 00676 00677 /// LoopCompare - Compare loops by PickMostRelevantLoop. 00678 class LoopCompare { 00679 DominatorTree &DT; 00680 public: 00681 explicit LoopCompare(DominatorTree &dt) : DT(dt) {} 00682 00683 bool operator()(std::pair<const Loop *, const SCEV *> LHS, 00684 std::pair<const Loop *, const SCEV *> RHS) const { 00685 // Keep pointer operands sorted at the end. 00686 if (LHS.second->getType()->isPointerTy() != 00687 RHS.second->getType()->isPointerTy()) 00688 return LHS.second->getType()->isPointerTy(); 00689 00690 // Compare loops with PickMostRelevantLoop. 00691 if (LHS.first != RHS.first) 00692 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first; 00693 00694 // If one operand is a non-constant negative and the other is not, 00695 // put the non-constant negative on the right so that a sub can 00696 // be used instead of a negate and add. 00697 if (LHS.second->isNonConstantNegative()) { 00698 if (!RHS.second->isNonConstantNegative()) 00699 return false; 00700 } else if (RHS.second->isNonConstantNegative()) 00701 return true; 00702 00703 // Otherwise they are equivalent according to this comparison. 00704 return false; 00705 } 00706 }; 00707 00708 } 00709 00710 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) { 00711 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 00712 00713 // Collect all the add operands in a loop, along with their associated loops. 00714 // Iterate in reverse so that constants are emitted last, all else equal, and 00715 // so that pointer operands are inserted first, which the code below relies on 00716 // to form more involved GEPs. 00717 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 00718 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()), 00719 E(S->op_begin()); I != E; ++I) 00720 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 00721 00722 // Sort by loop. Use a stable sort so that constants follow non-constants and 00723 // pointer operands precede non-pointer operands. 00724 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT)); 00725 00726 // Emit instructions to add all the operands. Hoist as much as possible 00727 // out of loops, and form meaningful getelementptrs where possible. 00728 Value *Sum = 0; 00729 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator 00730 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) { 00731 const Loop *CurLoop = I->first; 00732 const SCEV *Op = I->second; 00733 if (!Sum) { 00734 // This is the first operand. Just expand it. 00735 Sum = expand(Op); 00736 ++I; 00737 } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) { 00738 // The running sum expression is a pointer. Try to form a getelementptr 00739 // at this level with that as the base. 00740 SmallVector<const SCEV *, 4> NewOps; 00741 for (; I != E && I->first == CurLoop; ++I) { 00742 // If the operand is SCEVUnknown and not instructions, peek through 00743 // it, to enable more of it to be folded into the GEP. 00744 const SCEV *X = I->second; 00745 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X)) 00746 if (!isa<Instruction>(U->getValue())) 00747 X = SE.getSCEV(U->getValue()); 00748 NewOps.push_back(X); 00749 } 00750 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum); 00751 } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) { 00752 // The running sum is an integer, and there's a pointer at this level. 00753 // Try to form a getelementptr. If the running sum is instructions, 00754 // use a SCEVUnknown to avoid re-analyzing them. 00755 SmallVector<const SCEV *, 4> NewOps; 00756 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) : 00757 SE.getSCEV(Sum)); 00758 for (++I; I != E && I->first == CurLoop; ++I) 00759 NewOps.push_back(I->second); 00760 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op)); 00761 } else if (Op->isNonConstantNegative()) { 00762 // Instead of doing a negate and add, just do a subtract. 00763 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty); 00764 Sum = InsertNoopCastOfTo(Sum, Ty); 00765 Sum = InsertBinop(Instruction::Sub, Sum, W); 00766 ++I; 00767 } else { 00768 // A simple add. 00769 Value *W = expandCodeFor(Op, Ty); 00770 Sum = InsertNoopCastOfTo(Sum, Ty); 00771 // Canonicalize a constant to the RHS. 00772 if (isa<Constant>(Sum)) std::swap(Sum, W); 00773 Sum = InsertBinop(Instruction::Add, Sum, W); 00774 ++I; 00775 } 00776 } 00777 00778 return Sum; 00779 } 00780 00781 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) { 00782 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 00783 00784 // Collect all the mul operands in a loop, along with their associated loops. 00785 // Iterate in reverse so that constants are emitted last, all else equal. 00786 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 00787 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()), 00788 E(S->op_begin()); I != E; ++I) 00789 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 00790 00791 // Sort by loop. Use a stable sort so that constants follow non-constants. 00792 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT)); 00793 00794 // Emit instructions to mul all the operands. Hoist as much as possible 00795 // out of loops. 00796 Value *Prod = 0; 00797 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator 00798 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) { 00799 const SCEV *Op = I->second; 00800 if (!Prod) { 00801 // This is the first operand. Just expand it. 00802 Prod = expand(Op); 00803 ++I; 00804 } else if (Op->isAllOnesValue()) { 00805 // Instead of doing a multiply by negative one, just do a negate. 00806 Prod = InsertNoopCastOfTo(Prod, Ty); 00807 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod); 00808 ++I; 00809 } else { 00810 // A simple mul. 00811 Value *W = expandCodeFor(Op, Ty); 00812 Prod = InsertNoopCastOfTo(Prod, Ty); 00813 // Canonicalize a constant to the RHS. 00814 if (isa<Constant>(Prod)) std::swap(Prod, W); 00815 Prod = InsertBinop(Instruction::Mul, Prod, W); 00816 ++I; 00817 } 00818 } 00819 00820 return Prod; 00821 } 00822 00823 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) { 00824 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 00825 00826 Value *LHS = expandCodeFor(S->getLHS(), Ty); 00827 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) { 00828 const APInt &RHS = SC->getValue()->getValue(); 00829 if (RHS.isPowerOf2()) 00830 return InsertBinop(Instruction::LShr, LHS, 00831 ConstantInt::get(Ty, RHS.logBase2())); 00832 } 00833 00834 Value *RHS = expandCodeFor(S->getRHS(), Ty); 00835 return InsertBinop(Instruction::UDiv, LHS, RHS); 00836 } 00837 00838 /// Move parts of Base into Rest to leave Base with the minimal 00839 /// expression that provides a pointer operand suitable for a 00840 /// GEP expansion. 00841 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest, 00842 ScalarEvolution &SE) { 00843 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) { 00844 Base = A->getStart(); 00845 Rest = SE.getAddExpr(Rest, 00846 SE.getAddRecExpr(SE.getConstant(A->getType(), 0), 00847 A->getStepRecurrence(SE), 00848 A->getLoop(), 00849 // FIXME: A->getNoWrapFlags(FlagNW) 00850 SCEV::FlagAnyWrap)); 00851 } 00852 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) { 00853 Base = A->getOperand(A->getNumOperands()-1); 00854 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end()); 00855 NewAddOps.back() = Rest; 00856 Rest = SE.getAddExpr(NewAddOps); 00857 ExposePointerBase(Base, Rest, SE); 00858 } 00859 } 00860 00861 /// Determine if this is a well-behaved chain of instructions leading back to 00862 /// the PHI. If so, it may be reused by expanded expressions. 00863 bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV, 00864 const Loop *L) { 00865 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) || 00866 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV))) 00867 return false; 00868 // If any of the operands don't dominate the insert position, bail. 00869 // Addrec operands are always loop-invariant, so this can only happen 00870 // if there are instructions which haven't been hoisted. 00871 if (L == IVIncInsertLoop) { 00872 for (User::op_iterator OI = IncV->op_begin()+1, 00873 OE = IncV->op_end(); OI != OE; ++OI) 00874 if (Instruction *OInst = dyn_cast<Instruction>(OI)) 00875 if (!SE.DT->dominates(OInst, IVIncInsertPos)) 00876 return false; 00877 } 00878 // Advance to the next instruction. 00879 IncV = dyn_cast<Instruction>(IncV->getOperand(0)); 00880 if (!IncV) 00881 return false; 00882 00883 if (IncV->mayHaveSideEffects()) 00884 return false; 00885 00886 if (IncV != PN) 00887 return true; 00888 00889 return isNormalAddRecExprPHI(PN, IncV, L); 00890 } 00891 00892 /// getIVIncOperand returns an induction variable increment's induction 00893 /// variable operand. 00894 /// 00895 /// If allowScale is set, any type of GEP is allowed as long as the nonIV 00896 /// operands dominate InsertPos. 00897 /// 00898 /// If allowScale is not set, ensure that a GEP increment conforms to one of the 00899 /// simple patterns generated by getAddRecExprPHILiterally and 00900 /// expandAddtoGEP. If the pattern isn't recognized, return NULL. 00901 Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV, 00902 Instruction *InsertPos, 00903 bool allowScale) { 00904 if (IncV == InsertPos) 00905 return NULL; 00906 00907 switch (IncV->getOpcode()) { 00908 default: 00909 return NULL; 00910 // Check for a simple Add/Sub or GEP of a loop invariant step. 00911 case Instruction::Add: 00912 case Instruction::Sub: { 00913 Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1)); 00914 if (!OInst || SE.DT->dominates(OInst, InsertPos)) 00915 return dyn_cast<Instruction>(IncV->getOperand(0)); 00916 return NULL; 00917 } 00918 case Instruction::BitCast: 00919 return dyn_cast<Instruction>(IncV->getOperand(0)); 00920 case Instruction::GetElementPtr: 00921 for (Instruction::op_iterator I = IncV->op_begin()+1, E = IncV->op_end(); 00922 I != E; ++I) { 00923 if (isa<Constant>(*I)) 00924 continue; 00925 if (Instruction *OInst = dyn_cast<Instruction>(*I)) { 00926 if (!SE.DT->dominates(OInst, InsertPos)) 00927 return NULL; 00928 } 00929 if (allowScale) { 00930 // allow any kind of GEP as long as it can be hoisted. 00931 continue; 00932 } 00933 // This must be a pointer addition of constants (pretty), which is already 00934 // handled, or some number of address-size elements (ugly). Ugly geps 00935 // have 2 operands. i1* is used by the expander to represent an 00936 // address-size element. 00937 if (IncV->getNumOperands() != 2) 00938 return NULL; 00939 unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace(); 00940 if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS) 00941 && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS)) 00942 return NULL; 00943 break; 00944 } 00945 return dyn_cast<Instruction>(IncV->getOperand(0)); 00946 } 00947 } 00948 00949 /// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make 00950 /// it available to other uses in this loop. Recursively hoist any operands, 00951 /// until we reach a value that dominates InsertPos. 00952 bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) { 00953 if (SE.DT->dominates(IncV, InsertPos)) 00954 return true; 00955 00956 // InsertPos must itself dominate IncV so that IncV's new position satisfies 00957 // its existing users. 00958 if (isa<PHINode>(InsertPos) 00959 || !SE.DT->dominates(InsertPos->getParent(), IncV->getParent())) 00960 return false; 00961 00962 // Check that the chain of IV operands leading back to Phi can be hoisted. 00963 SmallVector<Instruction*, 4> IVIncs; 00964 for(;;) { 00965 Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true); 00966 if (!Oper) 00967 return false; 00968 // IncV is safe to hoist. 00969 IVIncs.push_back(IncV); 00970 IncV = Oper; 00971 if (SE.DT->dominates(IncV, InsertPos)) 00972 break; 00973 } 00974 for (SmallVectorImpl<Instruction*>::reverse_iterator I = IVIncs.rbegin(), 00975 E = IVIncs.rend(); I != E; ++I) { 00976 (*I)->moveBefore(InsertPos); 00977 } 00978 return true; 00979 } 00980 00981 /// Determine if this cyclic phi is in a form that would have been generated by 00982 /// LSR. We don't care if the phi was actually expanded in this pass, as long 00983 /// as it is in a low-cost form, for example, no implied multiplication. This 00984 /// should match any patterns generated by getAddRecExprPHILiterally and 00985 /// expandAddtoGEP. 00986 bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV, 00987 const Loop *L) { 00988 for(Instruction *IVOper = IncV; 00989 (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(), 00990 /*allowScale=*/false));) { 00991 if (IVOper == PN) 00992 return true; 00993 } 00994 return false; 00995 } 00996 00997 /// expandIVInc - Expand an IV increment at Builder's current InsertPos. 00998 /// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may 00999 /// need to materialize IV increments elsewhere to handle difficult situations. 01000 Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L, 01001 Type *ExpandTy, Type *IntTy, 01002 bool useSubtract) { 01003 Value *IncV; 01004 // If the PHI is a pointer, use a GEP, otherwise use an add or sub. 01005 if (ExpandTy->isPointerTy()) { 01006 PointerType *GEPPtrTy = cast<PointerType>(ExpandTy); 01007 // If the step isn't constant, don't use an implicitly scaled GEP, because 01008 // that would require a multiply inside the loop. 01009 if (!isa<ConstantInt>(StepV)) 01010 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()), 01011 GEPPtrTy->getAddressSpace()); 01012 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) }; 01013 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN); 01014 if (IncV->getType() != PN->getType()) { 01015 IncV = Builder.CreateBitCast(IncV, PN->getType()); 01016 rememberInstruction(IncV); 01017 } 01018 } else { 01019 IncV = useSubtract ? 01020 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") : 01021 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next"); 01022 rememberInstruction(IncV); 01023 } 01024 return IncV; 01025 } 01026 01027 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand 01028 /// the base addrec, which is the addrec without any non-loop-dominating 01029 /// values, and return the PHI. 01030 PHINode * 01031 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized, 01032 const Loop *L, 01033 Type *ExpandTy, 01034 Type *IntTy) { 01035 assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position"); 01036 01037 // Reuse a previously-inserted PHI, if present. 01038 BasicBlock *LatchBlock = L->getLoopLatch(); 01039 if (LatchBlock) { 01040 for (BasicBlock::iterator I = L->getHeader()->begin(); 01041 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 01042 if (!SE.isSCEVable(PN->getType()) || 01043 (SE.getEffectiveSCEVType(PN->getType()) != 01044 SE.getEffectiveSCEVType(Normalized->getType())) || 01045 SE.getSCEV(PN) != Normalized) 01046 continue; 01047 01048 Instruction *IncV = 01049 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock)); 01050 01051 if (LSRMode) { 01052 if (!isExpandedAddRecExprPHI(PN, IncV, L)) 01053 continue; 01054 if (L == IVIncInsertLoop && !hoistIVInc(IncV, IVIncInsertPos)) 01055 continue; 01056 } 01057 else { 01058 if (!isNormalAddRecExprPHI(PN, IncV, L)) 01059 continue; 01060 if (L == IVIncInsertLoop) 01061 do { 01062 if (SE.DT->dominates(IncV, IVIncInsertPos)) 01063 break; 01064 // Make sure the increment is where we want it. But don't move it 01065 // down past a potential existing post-inc user. 01066 IncV->moveBefore(IVIncInsertPos); 01067 IVIncInsertPos = IncV; 01068 IncV = cast<Instruction>(IncV->getOperand(0)); 01069 } while (IncV != PN); 01070 } 01071 // Ok, the add recurrence looks usable. 01072 // Remember this PHI, even in post-inc mode. 01073 InsertedValues.insert(PN); 01074 // Remember the increment. 01075 rememberInstruction(IncV); 01076 return PN; 01077 } 01078 } 01079 01080 // Save the original insertion point so we can restore it when we're done. 01081 BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 01082 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 01083 01084 // Another AddRec may need to be recursively expanded below. For example, if 01085 // this AddRec is quadratic, the StepV may itself be an AddRec in this 01086 // loop. Remove this loop from the PostIncLoops set before expanding such 01087 // AddRecs. Otherwise, we cannot find a valid position for the step 01088 // (i.e. StepV can never dominate its loop header). Ideally, we could do 01089 // SavedIncLoops.swap(PostIncLoops), but we generally have a single element, 01090 // so it's not worth implementing SmallPtrSet::swap. 01091 PostIncLoopSet SavedPostIncLoops = PostIncLoops; 01092 PostIncLoops.clear(); 01093 01094 // Expand code for the start value. 01095 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy, 01096 L->getHeader()->begin()); 01097 01098 // StartV must be hoisted into L's preheader to dominate the new phi. 01099 assert(!isa<Instruction>(StartV) || 01100 SE.DT->properlyDominates(cast<Instruction>(StartV)->getParent(), 01101 L->getHeader())); 01102 01103 // Expand code for the step value. Do this before creating the PHI so that PHI 01104 // reuse code doesn't see an incomplete PHI. 01105 const SCEV *Step = Normalized->getStepRecurrence(SE); 01106 // If the stride is negative, insert a sub instead of an add for the increment 01107 // (unless it's a constant, because subtracts of constants are canonicalized 01108 // to adds). 01109 bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 01110 if (useSubtract) 01111 Step = SE.getNegativeSCEV(Step); 01112 // Expand the step somewhere that dominates the loop header. 01113 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin()); 01114 01115 // Create the PHI. 01116 BasicBlock *Header = L->getHeader(); 01117 Builder.SetInsertPoint(Header, Header->begin()); 01118 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 01119 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE), 01120 Twine(IVName) + ".iv"); 01121 rememberInstruction(PN); 01122 01123 // Create the step instructions and populate the PHI. 01124 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 01125 BasicBlock *Pred = *HPI; 01126 01127 // Add a start value. 01128 if (!L->contains(Pred)) { 01129 PN->addIncoming(StartV, Pred); 01130 continue; 01131 } 01132 01133 // Create a step value and add it to the PHI. 01134 // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the 01135 // instructions at IVIncInsertPos. 01136 Instruction *InsertPos = L == IVIncInsertLoop ? 01137 IVIncInsertPos : Pred->getTerminator(); 01138 Builder.SetInsertPoint(InsertPos); 01139 Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 01140 01141 PN->addIncoming(IncV, Pred); 01142 } 01143 01144 // Restore the original insert point. 01145 if (SaveInsertBB) 01146 restoreInsertPoint(SaveInsertBB, SaveInsertPt); 01147 01148 // After expanding subexpressions, restore the PostIncLoops set so the caller 01149 // can ensure that IVIncrement dominates the current uses. 01150 PostIncLoops = SavedPostIncLoops; 01151 01152 // Remember this PHI, even in post-inc mode. 01153 InsertedValues.insert(PN); 01154 01155 return PN; 01156 } 01157 01158 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) { 01159 Type *STy = S->getType(); 01160 Type *IntTy = SE.getEffectiveSCEVType(STy); 01161 const Loop *L = S->getLoop(); 01162 01163 // Determine a normalized form of this expression, which is the expression 01164 // before any post-inc adjustment is made. 01165 const SCEVAddRecExpr *Normalized = S; 01166 if (PostIncLoops.count(L)) { 01167 PostIncLoopSet Loops; 01168 Loops.insert(L); 01169 Normalized = 01170 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, 0, 0, 01171 Loops, SE, *SE.DT)); 01172 } 01173 01174 // Strip off any non-loop-dominating component from the addrec start. 01175 const SCEV *Start = Normalized->getStart(); 01176 const SCEV *PostLoopOffset = 0; 01177 if (!SE.properlyDominates(Start, L->getHeader())) { 01178 PostLoopOffset = Start; 01179 Start = SE.getConstant(Normalized->getType(), 0); 01180 Normalized = cast<SCEVAddRecExpr>( 01181 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE), 01182 Normalized->getLoop(), 01183 // FIXME: Normalized->getNoWrapFlags(FlagNW) 01184 SCEV::FlagAnyWrap)); 01185 } 01186 01187 // Strip off any non-loop-dominating component from the addrec step. 01188 const SCEV *Step = Normalized->getStepRecurrence(SE); 01189 const SCEV *PostLoopScale = 0; 01190 if (!SE.dominates(Step, L->getHeader())) { 01191 PostLoopScale = Step; 01192 Step = SE.getConstant(Normalized->getType(), 1); 01193 Normalized = 01194 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, Step, 01195 Normalized->getLoop(), 01196 // FIXME: Normalized 01197 // ->getNoWrapFlags(FlagNW) 01198 SCEV::FlagAnyWrap)); 01199 } 01200 01201 // Expand the core addrec. If we need post-loop scaling, force it to 01202 // expand to an integer type to avoid the need for additional casting. 01203 Type *ExpandTy = PostLoopScale ? IntTy : STy; 01204 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy); 01205 01206 // Accommodate post-inc mode, if necessary. 01207 Value *Result; 01208 if (!PostIncLoops.count(L)) 01209 Result = PN; 01210 else { 01211 // In PostInc mode, use the post-incremented value. 01212 BasicBlock *LatchBlock = L->getLoopLatch(); 01213 assert(LatchBlock && "PostInc mode requires a unique loop latch!"); 01214 Result = PN->getIncomingValueForBlock(LatchBlock); 01215 01216 // For an expansion to use the postinc form, the client must call 01217 // expandCodeFor with an InsertPoint that is either outside the PostIncLoop 01218 // or dominated by IVIncInsertPos. 01219 if (isa<Instruction>(Result) 01220 && !SE.DT->dominates(cast<Instruction>(Result), 01221 Builder.GetInsertPoint())) { 01222 // The induction variable's postinc expansion does not dominate this use. 01223 // IVUsers tries to prevent this case, so it is rare. However, it can 01224 // happen when an IVUser outside the loop is not dominated by the latch 01225 // block. Adjusting IVIncInsertPos before expansion begins cannot handle 01226 // all cases. Consider a phi outide whose operand is replaced during 01227 // expansion with the value of the postinc user. Without fundamentally 01228 // changing the way postinc users are tracked, the only remedy is 01229 // inserting an extra IV increment. StepV might fold into PostLoopOffset, 01230 // but hopefully expandCodeFor handles that. 01231 bool useSubtract = 01232 !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 01233 if (useSubtract) 01234 Step = SE.getNegativeSCEV(Step); 01235 // Expand the step somewhere that dominates the loop header. 01236 BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 01237 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 01238 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin()); 01239 // Restore the insertion point to the place where the caller has 01240 // determined dominates all uses. 01241 restoreInsertPoint(SaveInsertBB, SaveInsertPt); 01242 Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 01243 } 01244 } 01245 01246 // Re-apply any non-loop-dominating scale. 01247 if (PostLoopScale) { 01248 Result = InsertNoopCastOfTo(Result, IntTy); 01249 Result = Builder.CreateMul(Result, 01250 expandCodeFor(PostLoopScale, IntTy)); 01251 rememberInstruction(Result); 01252 } 01253 01254 // Re-apply any non-loop-dominating offset. 01255 if (PostLoopOffset) { 01256 if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) { 01257 const SCEV *const OffsetArray[1] = { PostLoopOffset }; 01258 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result); 01259 } else { 01260 Result = InsertNoopCastOfTo(Result, IntTy); 01261 Result = Builder.CreateAdd(Result, 01262 expandCodeFor(PostLoopOffset, IntTy)); 01263 rememberInstruction(Result); 01264 } 01265 } 01266 01267 return Result; 01268 } 01269 01270 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) { 01271 if (!CanonicalMode) return expandAddRecExprLiterally(S); 01272 01273 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 01274 const Loop *L = S->getLoop(); 01275 01276 // First check for an existing canonical IV in a suitable type. 01277 PHINode *CanonicalIV = 0; 01278 if (PHINode *PN = L->getCanonicalInductionVariable()) 01279 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty)) 01280 CanonicalIV = PN; 01281 01282 // Rewrite an AddRec in terms of the canonical induction variable, if 01283 // its type is more narrow. 01284 if (CanonicalIV && 01285 SE.getTypeSizeInBits(CanonicalIV->getType()) > 01286 SE.getTypeSizeInBits(Ty)) { 01287 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands()); 01288 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i) 01289 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType()); 01290 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(), 01291 // FIXME: S->getNoWrapFlags(FlagNW) 01292 SCEV::FlagAnyWrap)); 01293 BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 01294 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 01295 BasicBlock::iterator NewInsertPt = 01296 llvm::next(BasicBlock::iterator(cast<Instruction>(V))); 01297 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt) || 01298 isa<LandingPadInst>(NewInsertPt)) 01299 ++NewInsertPt; 01300 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0, 01301 NewInsertPt); 01302 restoreInsertPoint(SaveInsertBB, SaveInsertPt); 01303 return V; 01304 } 01305 01306 // {X,+,F} --> X + {0,+,F} 01307 if (!S->getStart()->isZero()) { 01308 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end()); 01309 NewOps[0] = SE.getConstant(Ty, 0); 01310 // FIXME: can use S->getNoWrapFlags() 01311 const SCEV *Rest = SE.getAddRecExpr(NewOps, L, SCEV::FlagAnyWrap); 01312 01313 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the 01314 // comments on expandAddToGEP for details. 01315 const SCEV *Base = S->getStart(); 01316 const SCEV *RestArray[1] = { Rest }; 01317 // Dig into the expression to find the pointer base for a GEP. 01318 ExposePointerBase(Base, RestArray[0], SE); 01319 // If we found a pointer, expand the AddRec with a GEP. 01320 if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) { 01321 // Make sure the Base isn't something exotic, such as a multiplied 01322 // or divided pointer value. In those cases, the result type isn't 01323 // actually a pointer type. 01324 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) { 01325 Value *StartV = expand(Base); 01326 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!"); 01327 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV); 01328 } 01329 } 01330 01331 // Just do a normal add. Pre-expand the operands to suppress folding. 01332 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())), 01333 SE.getUnknown(expand(Rest)))); 01334 } 01335 01336 // If we don't yet have a canonical IV, create one. 01337 if (!CanonicalIV) { 01338 // Create and insert the PHI node for the induction variable in the 01339 // specified loop. 01340 BasicBlock *Header = L->getHeader(); 01341 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 01342 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar", 01343 Header->begin()); 01344 rememberInstruction(CanonicalIV); 01345 01346 Constant *One = ConstantInt::get(Ty, 1); 01347 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 01348 BasicBlock *HP = *HPI; 01349 if (L->contains(HP)) { 01350 // Insert a unit add instruction right before the terminator 01351 // corresponding to the back-edge. 01352 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One, 01353 "indvar.next", 01354 HP->getTerminator()); 01355 Add->setDebugLoc(HP->getTerminator()->getDebugLoc()); 01356 rememberInstruction(Add); 01357 CanonicalIV->addIncoming(Add, HP); 01358 } else { 01359 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP); 01360 } 01361 } 01362 } 01363 01364 // {0,+,1} --> Insert a canonical induction variable into the loop! 01365 if (S->isAffine() && S->getOperand(1)->isOne()) { 01366 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) && 01367 "IVs with types different from the canonical IV should " 01368 "already have been handled!"); 01369 return CanonicalIV; 01370 } 01371 01372 // {0,+,F} --> {0,+,1} * F 01373 01374 // If this is a simple linear addrec, emit it now as a special case. 01375 if (S->isAffine()) // {0,+,F} --> i*F 01376 return 01377 expand(SE.getTruncateOrNoop( 01378 SE.getMulExpr(SE.getUnknown(CanonicalIV), 01379 SE.getNoopOrAnyExtend(S->getOperand(1), 01380 CanonicalIV->getType())), 01381 Ty)); 01382 01383 // If this is a chain of recurrences, turn it into a closed form, using the 01384 // folders, then expandCodeFor the closed form. This allows the folders to 01385 // simplify the expression without having to build a bunch of special code 01386 // into this folder. 01387 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV. 01388 01389 // Promote S up to the canonical IV type, if the cast is foldable. 01390 const SCEV *NewS = S; 01391 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType()); 01392 if (isa<SCEVAddRecExpr>(Ext)) 01393 NewS = Ext; 01394 01395 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE); 01396 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n"; 01397 01398 // Truncate the result down to the original type, if needed. 01399 const SCEV *T = SE.getTruncateOrNoop(V, Ty); 01400 return expand(T); 01401 } 01402 01403 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) { 01404 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 01405 Value *V = expandCodeFor(S->getOperand(), 01406 SE.getEffectiveSCEVType(S->getOperand()->getType())); 01407 Value *I = Builder.CreateTrunc(V, Ty); 01408 rememberInstruction(I); 01409 return I; 01410 } 01411 01412 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) { 01413 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 01414 Value *V = expandCodeFor(S->getOperand(), 01415 SE.getEffectiveSCEVType(S->getOperand()->getType())); 01416 Value *I = Builder.CreateZExt(V, Ty); 01417 rememberInstruction(I); 01418 return I; 01419 } 01420 01421 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) { 01422 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 01423 Value *V = expandCodeFor(S->getOperand(), 01424 SE.getEffectiveSCEVType(S->getOperand()->getType())); 01425 Value *I = Builder.CreateSExt(V, Ty); 01426 rememberInstruction(I); 01427 return I; 01428 } 01429 01430 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) { 01431 Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 01432 Type *Ty = LHS->getType(); 01433 for (int i = S->getNumOperands()-2; i >= 0; --i) { 01434 // In the case of mixed integer and pointer types, do the 01435 // rest of the comparisons as integer. 01436 if (S->getOperand(i)->getType() != Ty) { 01437 Ty = SE.getEffectiveSCEVType(Ty); 01438 LHS = InsertNoopCastOfTo(LHS, Ty); 01439 } 01440 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 01441 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS); 01442 rememberInstruction(ICmp); 01443 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax"); 01444 rememberInstruction(Sel); 01445 LHS = Sel; 01446 } 01447 // In the case of mixed integer and pointer types, cast the 01448 // final result back to the pointer type. 01449 if (LHS->getType() != S->getType()) 01450 LHS = InsertNoopCastOfTo(LHS, S->getType()); 01451 return LHS; 01452 } 01453 01454 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) { 01455 Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 01456 Type *Ty = LHS->getType(); 01457 for (int i = S->getNumOperands()-2; i >= 0; --i) { 01458 // In the case of mixed integer and pointer types, do the 01459 // rest of the comparisons as integer. 01460 if (S->getOperand(i)->getType() != Ty) { 01461 Ty = SE.getEffectiveSCEVType(Ty); 01462 LHS = InsertNoopCastOfTo(LHS, Ty); 01463 } 01464 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 01465 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS); 01466 rememberInstruction(ICmp); 01467 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax"); 01468 rememberInstruction(Sel); 01469 LHS = Sel; 01470 } 01471 // In the case of mixed integer and pointer types, cast the 01472 // final result back to the pointer type. 01473 if (LHS->getType() != S->getType()) 01474 LHS = InsertNoopCastOfTo(LHS, S->getType()); 01475 return LHS; 01476 } 01477 01478 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty, 01479 Instruction *IP) { 01480 Builder.SetInsertPoint(IP->getParent(), IP); 01481 return expandCodeFor(SH, Ty); 01482 } 01483 01484 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) { 01485 // Expand the code for this SCEV. 01486 Value *V = expand(SH); 01487 if (Ty) { 01488 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) && 01489 "non-trivial casts should be done with the SCEVs directly!"); 01490 V = InsertNoopCastOfTo(V, Ty); 01491 } 01492 return V; 01493 } 01494 01495 Value *SCEVExpander::expand(const SCEV *S) { 01496 // Compute an insertion point for this SCEV object. Hoist the instructions 01497 // as far out in the loop nest as possible. 01498 Instruction *InsertPt = Builder.GetInsertPoint(); 01499 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ; 01500 L = L->getParentLoop()) 01501 if (SE.isLoopInvariant(S, L)) { 01502 if (!L) break; 01503 if (BasicBlock *Preheader = L->getLoopPreheader()) 01504 InsertPt = Preheader->getTerminator(); 01505 else { 01506 // LSR sets the insertion point for AddRec start/step values to the 01507 // block start to simplify value reuse, even though it's an invalid 01508 // position. SCEVExpander must correct for this in all cases. 01509 InsertPt = L->getHeader()->getFirstInsertionPt(); 01510 } 01511 } else { 01512 // If the SCEV is computable at this level, insert it into the header 01513 // after the PHIs (and after any other instructions that we've inserted 01514 // there) so that it is guaranteed to dominate any user inside the loop. 01515 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L)) 01516 InsertPt = L->getHeader()->getFirstInsertionPt(); 01517 while (InsertPt != Builder.GetInsertPoint() 01518 && (isInsertedInstruction(InsertPt) 01519 || isa<DbgInfoIntrinsic>(InsertPt))) { 01520 InsertPt = llvm::next(BasicBlock::iterator(InsertPt)); 01521 } 01522 break; 01523 } 01524 01525 // Check to see if we already expanded this here. 01526 std::map<std::pair<const SCEV *, Instruction *>, TrackingVH<Value> >::iterator 01527 I = InsertedExpressions.find(std::make_pair(S, InsertPt)); 01528 if (I != InsertedExpressions.end()) 01529 return I->second; 01530 01531 BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 01532 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 01533 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt); 01534 01535 // Expand the expression into instructions. 01536 Value *V = visit(S); 01537 01538 // Remember the expanded value for this SCEV at this location. 01539 // 01540 // This is independent of PostIncLoops. The mapped value simply materializes 01541 // the expression at this insertion point. If the mapped value happened to be 01542 // a postinc expansion, it could be reused by a non postinc user, but only if 01543 // its insertion point was already at the head of the loop. 01544 InsertedExpressions[std::make_pair(S, InsertPt)] = V; 01545 01546 restoreInsertPoint(SaveInsertBB, SaveInsertPt); 01547 return V; 01548 } 01549 01550 void SCEVExpander::rememberInstruction(Value *I) { 01551 if (!PostIncLoops.empty()) 01552 InsertedPostIncValues.insert(I); 01553 else 01554 InsertedValues.insert(I); 01555 } 01556 01557 void SCEVExpander::restoreInsertPoint(BasicBlock *BB, BasicBlock::iterator I) { 01558 Builder.SetInsertPoint(BB, I); 01559 } 01560 01561 /// getOrInsertCanonicalInductionVariable - This method returns the 01562 /// canonical induction variable of the specified type for the specified 01563 /// loop (inserting one if there is none). A canonical induction variable 01564 /// starts at zero and steps by one on each iteration. 01565 PHINode * 01566 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L, 01567 Type *Ty) { 01568 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!"); 01569 01570 // Build a SCEV for {0,+,1}<L>. 01571 // Conservatively use FlagAnyWrap for now. 01572 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0), 01573 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap); 01574 01575 // Emit code for it. 01576 BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 01577 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 01578 PHINode *V = cast<PHINode>(expandCodeFor(H, 0, L->getHeader()->begin())); 01579 if (SaveInsertBB) 01580 restoreInsertPoint(SaveInsertBB, SaveInsertPt); 01581 01582 return V; 01583 } 01584 01585 /// Sort values by integer width for replaceCongruentIVs. 01586 static bool width_descending(Value *lhs, Value *rhs) { 01587 // Put pointers at the back and make sure pointer < pointer = false. 01588 if (!lhs->getType()->isIntegerTy() || !rhs->getType()->isIntegerTy()) 01589 return rhs->getType()->isIntegerTy() && !lhs->getType()->isIntegerTy(); 01590 return rhs->getType()->getPrimitiveSizeInBits() 01591 < lhs->getType()->getPrimitiveSizeInBits(); 01592 } 01593 01594 /// replaceCongruentIVs - Check for congruent phis in this loop header and 01595 /// replace them with their most canonical representative. Return the number of 01596 /// phis eliminated. 01597 /// 01598 /// This does not depend on any SCEVExpander state but should be used in 01599 /// the same context that SCEVExpander is used. 01600 unsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT, 01601 SmallVectorImpl<WeakVH> &DeadInsts, 01602 const TargetTransformInfo *TTI) { 01603 // Find integer phis in order of increasing width. 01604 SmallVector<PHINode*, 8> Phis; 01605 for (BasicBlock::iterator I = L->getHeader()->begin(); 01606 PHINode *Phi = dyn_cast<PHINode>(I); ++I) { 01607 Phis.push_back(Phi); 01608 } 01609 if (TTI) 01610 std::sort(Phis.begin(), Phis.end(), width_descending); 01611 01612 unsigned NumElim = 0; 01613 DenseMap<const SCEV *, PHINode *> ExprToIVMap; 01614 // Process phis from wide to narrow. Mapping wide phis to the their truncation 01615 // so narrow phis can reuse them. 01616 for (SmallVectorImpl<PHINode*>::const_iterator PIter = Phis.begin(), 01617 PEnd = Phis.end(); PIter != PEnd; ++PIter) { 01618 PHINode *Phi = *PIter; 01619 01620 // Fold constant phis. They may be congruent to other constant phis and 01621 // would confuse the logic below that expects proper IVs. 01622 if (Value *V = Phi->hasConstantValue()) { 01623 Phi->replaceAllUsesWith(V); 01624 DeadInsts.push_back(Phi); 01625 ++NumElim; 01626 DEBUG_WITH_TYPE(DebugType, dbgs() 01627 << "INDVARS: Eliminated constant iv: " << *Phi << '\n'); 01628 continue; 01629 } 01630 01631 if (!SE.isSCEVable(Phi->getType())) 01632 continue; 01633 01634 PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)]; 01635 if (!OrigPhiRef) { 01636 OrigPhiRef = Phi; 01637 if (Phi->getType()->isIntegerTy() && TTI 01638 && TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) { 01639 // This phi can be freely truncated to the narrowest phi type. Map the 01640 // truncated expression to it so it will be reused for narrow types. 01641 const SCEV *TruncExpr = 01642 SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType()); 01643 ExprToIVMap[TruncExpr] = Phi; 01644 } 01645 continue; 01646 } 01647 01648 // Replacing a pointer phi with an integer phi or vice-versa doesn't make 01649 // sense. 01650 if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy()) 01651 continue; 01652 01653 if (BasicBlock *LatchBlock = L->getLoopLatch()) { 01654 Instruction *OrigInc = 01655 cast<Instruction>(OrigPhiRef->getIncomingValueForBlock(LatchBlock)); 01656 Instruction *IsomorphicInc = 01657 cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock)); 01658 01659 // If this phi has the same width but is more canonical, replace the 01660 // original with it. As part of the "more canonical" determination, 01661 // respect a prior decision to use an IV chain. 01662 if (OrigPhiRef->getType() == Phi->getType() 01663 && !(ChainedPhis.count(Phi) 01664 || isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L)) 01665 && (ChainedPhis.count(Phi) 01666 || isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) { 01667 std::swap(OrigPhiRef, Phi); 01668 std::swap(OrigInc, IsomorphicInc); 01669 } 01670 // Replacing the congruent phi is sufficient because acyclic redundancy 01671 // elimination, CSE/GVN, should handle the rest. However, once SCEV proves 01672 // that a phi is congruent, it's often the head of an IV user cycle that 01673 // is isomorphic with the original phi. It's worth eagerly cleaning up the 01674 // common case of a single IV increment so that DeleteDeadPHIs can remove 01675 // cycles that had postinc uses. 01676 const SCEV *TruncExpr = SE.getTruncateOrNoop(SE.getSCEV(OrigInc), 01677 IsomorphicInc->getType()); 01678 if (OrigInc != IsomorphicInc 01679 && TruncExpr == SE.getSCEV(IsomorphicInc) 01680 && ((isa<PHINode>(OrigInc) && isa<PHINode>(IsomorphicInc)) 01681 || hoistIVInc(OrigInc, IsomorphicInc))) { 01682 DEBUG_WITH_TYPE(DebugType, dbgs() 01683 << "INDVARS: Eliminated congruent iv.inc: " 01684 << *IsomorphicInc << '\n'); 01685 Value *NewInc = OrigInc; 01686 if (OrigInc->getType() != IsomorphicInc->getType()) { 01687 Instruction *IP = isa<PHINode>(OrigInc) 01688 ? (Instruction*)L->getHeader()->getFirstInsertionPt() 01689 : OrigInc->getNextNode(); 01690 IRBuilder<> Builder(IP); 01691 Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc()); 01692 NewInc = Builder. 01693 CreateTruncOrBitCast(OrigInc, IsomorphicInc->getType(), IVName); 01694 } 01695 IsomorphicInc->replaceAllUsesWith(NewInc); 01696 DeadInsts.push_back(IsomorphicInc); 01697 } 01698 } 01699 DEBUG_WITH_TYPE(DebugType, dbgs() 01700 << "INDVARS: Eliminated congruent iv: " << *Phi << '\n'); 01701 ++NumElim; 01702 Value *NewIV = OrigPhiRef; 01703 if (OrigPhiRef->getType() != Phi->getType()) { 01704 IRBuilder<> Builder(L->getHeader()->getFirstInsertionPt()); 01705 Builder.SetCurrentDebugLocation(Phi->getDebugLoc()); 01706 NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName); 01707 } 01708 Phi->replaceAllUsesWith(NewIV); 01709 DeadInsts.push_back(Phi); 01710 } 01711 return NumElim; 01712 } 01713 01714 namespace { 01715 // Search for a SCEV subexpression that is not safe to expand. Any expression 01716 // that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely 01717 // UDiv expressions. We don't know if the UDiv is derived from an IR divide 01718 // instruction, but the important thing is that we prove the denominator is 01719 // nonzero before expansion. 01720 // 01721 // IVUsers already checks that IV-derived expressions are safe. So this check is 01722 // only needed when the expression includes some subexpression that is not IV 01723 // derived. 01724 // 01725 // Currently, we only allow division by a nonzero constant here. If this is 01726 // inadequate, we could easily allow division by SCEVUnknown by using 01727 // ValueTracking to check isKnownNonZero(). 01728 struct SCEVFindUnsafe { 01729 bool IsUnsafe; 01730 01731 SCEVFindUnsafe(): IsUnsafe(false) {} 01732 01733 bool follow(const SCEV *S) { 01734 const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S); 01735 if (!D) 01736 return true; 01737 const SCEVConstant *SC = dyn_cast<SCEVConstant>(D->getRHS()); 01738 if (SC && !SC->getValue()->isZero()) 01739 return true; 01740 IsUnsafe = true; 01741 return false; 01742 } 01743 bool isDone() const { return IsUnsafe; } 01744 }; 01745 } 01746 01747 namespace llvm { 01748 bool isSafeToExpand(const SCEV *S) { 01749 SCEVFindUnsafe Search; 01750 visitAll(S, Search); 01751 return !Search.IsUnsafe; 01752 } 01753 }