X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FAnalysis%2FScalarEvolutionExpander.cpp;h=7a9efdaa4c242b00ada86b61a76ab0b3f56fe3d4;hb=b8c5cfb13078eb0c6fd3de4a79f642aaf6b6b957;hp=f3e508ab7b173ed3c1eb9711a80d3a1201016850;hpb=4ee451de366474b9c228b4e5fa573795a715216d;p=oota-llvm.git diff --git a/lib/Analysis/ScalarEvolutionExpander.cpp b/lib/Analysis/ScalarEvolutionExpander.cpp index f3e508ab7b1..7a9efdaa4c2 100644 --- a/lib/Analysis/ScalarEvolutionExpander.cpp +++ b/lib/Analysis/ScalarEvolutionExpander.cpp @@ -14,64 +14,145 @@ //===----------------------------------------------------------------------===// #include "llvm/Analysis/ScalarEvolutionExpander.h" +#include "llvm/ADT/SmallSet.h" +#include "llvm/ADT/STLExtras.h" +#include "llvm/Analysis/Dominators.h" #include "llvm/Analysis/LoopInfo.h" +#include "llvm/Analysis/TargetTransformInfo.h" +#include "llvm/IR/DataLayout.h" +#include "llvm/IR/IntrinsicInst.h" +#include "llvm/IR/LLVMContext.h" +#include "llvm/Support/Debug.h" + using namespace llvm; -/// InsertCastOfTo - Insert a cast of V to the specified type, doing what -/// we can to share the casts. -Value *SCEVExpander::InsertCastOfTo(Instruction::CastOps opcode, Value *V, - const Type *Ty) { - // FIXME: keep track of the cast instruction. - if (Constant *C = dyn_cast(V)) - return ConstantExpr::getCast(opcode, C, Ty); - - if (Argument *A = dyn_cast(V)) { - // Check to see if there is already a cast! - for (Value::use_iterator UI = A->use_begin(), E = A->use_end(); - UI != E; ++UI) { - if ((*UI)->getType() == Ty) - if (CastInst *CI = dyn_cast(cast(*UI))) { - // If the cast isn't the first instruction of the function, move it. - if (BasicBlock::iterator(CI) != - A->getParent()->getEntryBlock().begin()) { - CI->moveBefore(A->getParent()->getEntryBlock().begin()); +/// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP, +/// reusing an existing cast if a suitable one exists, moving an existing +/// cast if a suitable one exists but isn't in the right place, or +/// creating a new one. +Value *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty, + Instruction::CastOps Op, + BasicBlock::iterator IP) { + // This function must be called with the builder having a valid insertion + // point. It doesn't need to be the actual IP where the uses of the returned + // cast will be added, but it must dominate such IP. + // We use this precondition to produce a cast that will dominate all its + // uses. In particular, this is crucial for the case where the builder's + // insertion point *is* the point where we were asked to put the cast. + // Since we don't know the builder's insertion point is actually + // where the uses will be added (only that it dominates it), we are + // not allowed to move it. + BasicBlock::iterator BIP = Builder.GetInsertPoint(); + + Instruction *Ret = NULL; + + // Check to see if there is already a cast! + for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); + UI != E; ++UI) { + User *U = *UI; + if (U->getType() == Ty) + if (CastInst *CI = dyn_cast(U)) + if (CI->getOpcode() == Op) { + // If the cast isn't where we want it, create a new cast at IP. + // Likewise, do not reuse a cast at BIP because it must dominate + // instructions that might be inserted before BIP. + if (BasicBlock::iterator(CI) != IP || BIP == IP) { + // Create a new cast, and leave the old cast in place in case + // it is being used as an insert point. Clear its operand + // so that it doesn't hold anything live. + Ret = CastInst::Create(Op, V, Ty, "", IP); + Ret->takeName(CI); + CI->replaceAllUsesWith(Ret); + CI->setOperand(0, UndefValue::get(V->getType())); + break; } - return CI; + Ret = CI; + break; } + } + + // Create a new cast. + if (!Ret) + Ret = CastInst::Create(Op, V, Ty, V->getName(), IP); + + // We assert at the end of the function since IP might point to an + // instruction with different dominance properties than a cast + // (an invoke for example) and not dominate BIP (but the cast does). + assert(SE.DT->dominates(Ret, BIP)); + + rememberInstruction(Ret); + return Ret; +} + +/// InsertNoopCastOfTo - Insert a cast of V to the specified type, +/// which must be possible with a noop cast, doing what we can to share +/// the casts. +Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) { + Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false); + assert((Op == Instruction::BitCast || + Op == Instruction::PtrToInt || + Op == Instruction::IntToPtr) && + "InsertNoopCastOfTo cannot perform non-noop casts!"); + assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) && + "InsertNoopCastOfTo cannot change sizes!"); + + // Short-circuit unnecessary bitcasts. + if (Op == Instruction::BitCast) { + if (V->getType() == Ty) + return V; + if (CastInst *CI = dyn_cast(V)) { + if (CI->getOperand(0)->getType() == Ty) + return CI->getOperand(0); } - return CastInst::create(opcode, V, Ty, V->getName(), - A->getParent()->getEntryBlock().begin()); } - - Instruction *I = cast(V); - - // Check to see if there is already a cast. If there is, use it. - for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); - UI != E; ++UI) { - if ((*UI)->getType() == Ty) - if (CastInst *CI = dyn_cast(cast(*UI))) { - BasicBlock::iterator It = I; ++It; - if (isa(I)) - It = cast(I)->getNormalDest()->begin(); - while (isa(It)) ++It; - if (It != BasicBlock::iterator(CI)) { - // Splice the cast immediately after the operand in question. - CI->moveBefore(It); - } - return CI; - } + // Short-circuit unnecessary inttoptr<->ptrtoint casts. + if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) && + SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) { + if (CastInst *CI = dyn_cast(V)) + if ((CI->getOpcode() == Instruction::PtrToInt || + CI->getOpcode() == Instruction::IntToPtr) && + SE.getTypeSizeInBits(CI->getType()) == + SE.getTypeSizeInBits(CI->getOperand(0)->getType())) + return CI->getOperand(0); + if (ConstantExpr *CE = dyn_cast(V)) + if ((CE->getOpcode() == Instruction::PtrToInt || + CE->getOpcode() == Instruction::IntToPtr) && + SE.getTypeSizeInBits(CE->getType()) == + SE.getTypeSizeInBits(CE->getOperand(0)->getType())) + return CE->getOperand(0); + } + + // Fold a cast of a constant. + if (Constant *C = dyn_cast(V)) + return ConstantExpr::getCast(Op, C, Ty); + + // Cast the argument at the beginning of the entry block, after + // any bitcasts of other arguments. + if (Argument *A = dyn_cast(V)) { + BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin(); + while ((isa(IP) && + isa(cast(IP)->getOperand(0)) && + cast(IP)->getOperand(0) != A) || + isa(IP) || + isa(IP)) + ++IP; + return ReuseOrCreateCast(A, Ty, Op, IP); } + + // Cast the instruction immediately after the instruction. + Instruction *I = cast(V); BasicBlock::iterator IP = I; ++IP; if (InvokeInst *II = dyn_cast(I)) IP = II->getNormalDest()->begin(); - while (isa(IP)) ++IP; - return CastInst::create(opcode, V, Ty, V->getName(), IP); + while (isa(IP) || isa(IP)) + ++IP; + return ReuseOrCreateCast(I, Ty, Op, IP); } /// InsertBinop - Insert the specified binary operator, doing a small amount /// of work to avoid inserting an obviously redundant operation. -Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, Value *LHS, - Value *RHS, Instruction *&InsertPt) { +Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, + Value *LHS, Value *RHS) { // Fold a binop with constant operands. if (Constant *CLHS = dyn_cast(LHS)) if (Constant *CRHS = dyn_cast(RHS)) @@ -79,153 +160,1590 @@ Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, Value *LHS, // Do a quick scan to see if we have this binop nearby. If so, reuse it. unsigned ScanLimit = 6; - for (BasicBlock::iterator IP = InsertPt, E = InsertPt->getParent()->begin(); - ScanLimit; --IP, --ScanLimit) { - if (BinaryOperator *BinOp = dyn_cast(IP)) - if (BinOp->getOpcode() == Opcode && BinOp->getOperand(0) == LHS && - BinOp->getOperand(1) == RHS) { - // If we found the instruction *at* the insert point, insert later - // instructions after it. - if (BinOp == InsertPt) - InsertPt = ++IP; - return BinOp; - } - if (IP == E) break; - } - - // If we don't have - return BinaryOperator::create(Opcode, LHS, RHS, "tmp", InsertPt); -} - -Value *SCEVExpander::visitMulExpr(SCEVMulExpr *S) { - int FirstOp = 0; // Set if we should emit a subtract. - if (SCEVConstant *SC = dyn_cast(S->getOperand(0))) - if (SC->getValue()->isAllOnesValue()) - FirstOp = 1; - - int i = S->getNumOperands()-2; - Value *V = expand(S->getOperand(i+1)); - - // Emit a bunch of multiply instructions - for (; i >= FirstOp; --i) - V = InsertBinop(Instruction::Mul, V, expand(S->getOperand(i)), - InsertPt); - // -1 * ... ---> 0 - ... - if (FirstOp == 1) - V = InsertBinop(Instruction::Sub, Constant::getNullValue(V->getType()), V, - InsertPt); - return V; + BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); + // Scanning starts from the last instruction before the insertion point. + BasicBlock::iterator IP = Builder.GetInsertPoint(); + if (IP != BlockBegin) { + --IP; + for (; ScanLimit; --IP, --ScanLimit) { + // Don't count dbg.value against the ScanLimit, to avoid perturbing the + // generated code. + if (isa(IP)) + ScanLimit++; + if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS && + IP->getOperand(1) == RHS) + return IP; + if (IP == BlockBegin) break; + } + } + + // Save the original insertion point so we can restore it when we're done. + DebugLoc Loc = Builder.GetInsertPoint()->getDebugLoc(); + BuilderType::InsertPointGuard Guard(Builder); + + // Move the insertion point out of as many loops as we can. + while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { + if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break; + BasicBlock *Preheader = L->getLoopPreheader(); + if (!Preheader) break; + + // Ok, move up a level. + Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); + } + + // If we haven't found this binop, insert it. + Instruction *BO = cast(Builder.CreateBinOp(Opcode, LHS, RHS)); + BO->setDebugLoc(Loc); + rememberInstruction(BO); + + return BO; +} + +/// FactorOutConstant - Test if S is divisible by Factor, using signed +/// division. If so, update S with Factor divided out and return true. +/// S need not be evenly divisible if a reasonable remainder can be +/// computed. +/// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made +/// unnecessary; in its place, just signed-divide Ops[i] by the scale and +/// check to see if the divide was folded. +static bool FactorOutConstant(const SCEV *&S, + const SCEV *&Remainder, + const SCEV *Factor, + ScalarEvolution &SE, + const DataLayout *TD) { + // Everything is divisible by one. + if (Factor->isOne()) + return true; + + // x/x == 1. + if (S == Factor) { + S = SE.getConstant(S->getType(), 1); + return true; + } + + // For a Constant, check for a multiple of the given factor. + if (const SCEVConstant *C = dyn_cast(S)) { + // 0/x == 0. + if (C->isZero()) + return true; + // Check for divisibility. + if (const SCEVConstant *FC = dyn_cast(Factor)) { + ConstantInt *CI = + ConstantInt::get(SE.getContext(), + C->getValue()->getValue().sdiv( + FC->getValue()->getValue())); + // If the quotient is zero and the remainder is non-zero, reject + // the value at this scale. It will be considered for subsequent + // smaller scales. + if (!CI->isZero()) { + const SCEV *Div = SE.getConstant(CI); + S = Div; + Remainder = + SE.getAddExpr(Remainder, + SE.getConstant(C->getValue()->getValue().srem( + FC->getValue()->getValue()))); + return true; + } + } + } + + // In a Mul, check if there is a constant operand which is a multiple + // of the given factor. + if (const SCEVMulExpr *M = dyn_cast(S)) { + if (TD) { + // With DataLayout, the size is known. Check if there is a constant + // operand which is a multiple of the given factor. If so, we can + // factor it. + const SCEVConstant *FC = cast(Factor); + if (const SCEVConstant *C = dyn_cast(M->getOperand(0))) + if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) { + SmallVector NewMulOps(M->op_begin(), M->op_end()); + NewMulOps[0] = + SE.getConstant(C->getValue()->getValue().sdiv( + FC->getValue()->getValue())); + S = SE.getMulExpr(NewMulOps); + return true; + } + } else { + // Without DataLayout, check if Factor can be factored out of any of the + // Mul's operands. If so, we can just remove it. + for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) { + const SCEV *SOp = M->getOperand(i); + const SCEV *Remainder = SE.getConstant(SOp->getType(), 0); + if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) && + Remainder->isZero()) { + SmallVector NewMulOps(M->op_begin(), M->op_end()); + NewMulOps[i] = SOp; + S = SE.getMulExpr(NewMulOps); + return true; + } + } + } + } + + // In an AddRec, check if both start and step are divisible. + if (const SCEVAddRecExpr *A = dyn_cast(S)) { + const SCEV *Step = A->getStepRecurrence(SE); + const SCEV *StepRem = SE.getConstant(Step->getType(), 0); + if (!FactorOutConstant(Step, StepRem, Factor, SE, TD)) + return false; + if (!StepRem->isZero()) + return false; + const SCEV *Start = A->getStart(); + if (!FactorOutConstant(Start, Remainder, Factor, SE, TD)) + return false; + S = SE.getAddRecExpr(Start, Step, A->getLoop(), + A->getNoWrapFlags(SCEV::FlagNW)); + return true; + } + + return false; +} + +/// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs +/// is the number of SCEVAddRecExprs present, which are kept at the end of +/// the list. +/// +static void SimplifyAddOperands(SmallVectorImpl &Ops, + Type *Ty, + ScalarEvolution &SE) { + unsigned NumAddRecs = 0; + for (unsigned i = Ops.size(); i > 0 && isa(Ops[i-1]); --i) + ++NumAddRecs; + // Group Ops into non-addrecs and addrecs. + SmallVector NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs); + SmallVector AddRecs(Ops.end() - NumAddRecs, Ops.end()); + // Let ScalarEvolution sort and simplify the non-addrecs list. + const SCEV *Sum = NoAddRecs.empty() ? + SE.getConstant(Ty, 0) : + SE.getAddExpr(NoAddRecs); + // If it returned an add, use the operands. Otherwise it simplified + // the sum into a single value, so just use that. + Ops.clear(); + if (const SCEVAddExpr *Add = dyn_cast(Sum)) + Ops.append(Add->op_begin(), Add->op_end()); + else if (!Sum->isZero()) + Ops.push_back(Sum); + // Then append the addrecs. + Ops.append(AddRecs.begin(), AddRecs.end()); +} + +/// SplitAddRecs - Flatten a list of add operands, moving addrec start values +/// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}. +/// This helps expose more opportunities for folding parts of the expressions +/// into GEP indices. +/// +static void SplitAddRecs(SmallVectorImpl &Ops, + Type *Ty, + ScalarEvolution &SE) { + // Find the addrecs. + SmallVector AddRecs; + for (unsigned i = 0, e = Ops.size(); i != e; ++i) + while (const SCEVAddRecExpr *A = dyn_cast(Ops[i])) { + const SCEV *Start = A->getStart(); + if (Start->isZero()) break; + const SCEV *Zero = SE.getConstant(Ty, 0); + AddRecs.push_back(SE.getAddRecExpr(Zero, + A->getStepRecurrence(SE), + A->getLoop(), + A->getNoWrapFlags(SCEV::FlagNW))); + if (const SCEVAddExpr *Add = dyn_cast(Start)) { + Ops[i] = Zero; + Ops.append(Add->op_begin(), Add->op_end()); + e += Add->getNumOperands(); + } else { + Ops[i] = Start; + } + } + if (!AddRecs.empty()) { + // Add the addrecs onto the end of the list. + Ops.append(AddRecs.begin(), AddRecs.end()); + // Resort the operand list, moving any constants to the front. + SimplifyAddOperands(Ops, Ty, SE); + } +} + +/// expandAddToGEP - Expand an addition expression with a pointer type into +/// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps +/// BasicAliasAnalysis and other passes analyze the result. See the rules +/// for getelementptr vs. inttoptr in +/// http://llvm.org/docs/LangRef.html#pointeraliasing +/// for details. +/// +/// Design note: The correctness of using getelementptr here depends on +/// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as +/// they may introduce pointer arithmetic which may not be safely converted +/// into getelementptr. +/// +/// Design note: It might seem desirable for this function to be more +/// loop-aware. If some of the indices are loop-invariant while others +/// aren't, it might seem desirable to emit multiple GEPs, keeping the +/// loop-invariant portions of the overall computation outside the loop. +/// However, there are a few reasons this is not done here. Hoisting simple +/// arithmetic is a low-level optimization that often isn't very +/// important until late in the optimization process. In fact, passes +/// like InstructionCombining will combine GEPs, even if it means +/// pushing loop-invariant computation down into loops, so even if the +/// GEPs were split here, the work would quickly be undone. The +/// LoopStrengthReduction pass, which is usually run quite late (and +/// after the last InstructionCombining pass), takes care of hoisting +/// loop-invariant portions of expressions, after considering what +/// can be folded using target addressing modes. +/// +Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin, + const SCEV *const *op_end, + PointerType *PTy, + Type *Ty, + Value *V) { + Type *ElTy = PTy->getElementType(); + SmallVector GepIndices; + SmallVector Ops(op_begin, op_end); + bool AnyNonZeroIndices = false; + + // Split AddRecs up into parts as either of the parts may be usable + // without the other. + SplitAddRecs(Ops, Ty, SE); + + Type *IntPtrTy = SE.TD + ? SE.TD->getIntPtrType(PTy) + : Type::getInt64Ty(PTy->getContext()); + + // Descend down the pointer's type and attempt to convert the other + // operands into GEP indices, at each level. The first index in a GEP + // indexes into the array implied by the pointer operand; the rest of + // the indices index into the element or field type selected by the + // preceding index. + for (;;) { + // If the scale size is not 0, attempt to factor out a scale for + // array indexing. + SmallVector ScaledOps; + if (ElTy->isSized()) { + const SCEV *ElSize = SE.getSizeOfExpr(IntPtrTy, ElTy); + if (!ElSize->isZero()) { + SmallVector NewOps; + for (unsigned i = 0, e = Ops.size(); i != e; ++i) { + const SCEV *Op = Ops[i]; + const SCEV *Remainder = SE.getConstant(Ty, 0); + if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) { + // Op now has ElSize factored out. + ScaledOps.push_back(Op); + if (!Remainder->isZero()) + NewOps.push_back(Remainder); + AnyNonZeroIndices = true; + } else { + // The operand was not divisible, so add it to the list of operands + // we'll scan next iteration. + NewOps.push_back(Ops[i]); + } + } + // If we made any changes, update Ops. + if (!ScaledOps.empty()) { + Ops = NewOps; + SimplifyAddOperands(Ops, Ty, SE); + } + } + } + + // Record the scaled array index for this level of the type. If + // we didn't find any operands that could be factored, tentatively + // assume that element zero was selected (since the zero offset + // would obviously be folded away). + Value *Scaled = ScaledOps.empty() ? + Constant::getNullValue(Ty) : + expandCodeFor(SE.getAddExpr(ScaledOps), Ty); + GepIndices.push_back(Scaled); + + // Collect struct field index operands. + while (StructType *STy = dyn_cast(ElTy)) { + bool FoundFieldNo = false; + // An empty struct has no fields. + if (STy->getNumElements() == 0) break; + if (SE.TD) { + // With DataLayout, field offsets are known. See if a constant offset + // falls within any of the struct fields. + if (Ops.empty()) break; + if (const SCEVConstant *C = dyn_cast(Ops[0])) + if (SE.getTypeSizeInBits(C->getType()) <= 64) { + const StructLayout &SL = *SE.TD->getStructLayout(STy); + uint64_t FullOffset = C->getValue()->getZExtValue(); + if (FullOffset < SL.getSizeInBytes()) { + unsigned ElIdx = SL.getElementContainingOffset(FullOffset); + GepIndices.push_back( + ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx)); + ElTy = STy->getTypeAtIndex(ElIdx); + Ops[0] = + SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx)); + AnyNonZeroIndices = true; + FoundFieldNo = true; + } + } + } else { + // Without DataLayout, just check for an offsetof expression of the + // appropriate struct type. + for (unsigned i = 0, e = Ops.size(); i != e; ++i) + if (const SCEVUnknown *U = dyn_cast(Ops[i])) { + Type *CTy; + Constant *FieldNo; + if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) { + GepIndices.push_back(FieldNo); + ElTy = + STy->getTypeAtIndex(cast(FieldNo)->getZExtValue()); + Ops[i] = SE.getConstant(Ty, 0); + AnyNonZeroIndices = true; + FoundFieldNo = true; + break; + } + } + } + // If no struct field offsets were found, tentatively assume that + // field zero was selected (since the zero offset would obviously + // be folded away). + if (!FoundFieldNo) { + ElTy = STy->getTypeAtIndex(0u); + GepIndices.push_back( + Constant::getNullValue(Type::getInt32Ty(Ty->getContext()))); + } + } + + if (ArrayType *ATy = dyn_cast(ElTy)) + ElTy = ATy->getElementType(); + else + break; + } + + // If none of the operands were convertible to proper GEP indices, cast + // the base to i8* and do an ugly getelementptr with that. It's still + // better than ptrtoint+arithmetic+inttoptr at least. + if (!AnyNonZeroIndices) { + // Cast the base to i8*. + V = InsertNoopCastOfTo(V, + Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace())); + + assert(!isa(V) || + SE.DT->dominates(cast(V), Builder.GetInsertPoint())); + + // Expand the operands for a plain byte offset. + Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty); + + // Fold a GEP with constant operands. + if (Constant *CLHS = dyn_cast(V)) + if (Constant *CRHS = dyn_cast(Idx)) + return ConstantExpr::getGetElementPtr(CLHS, CRHS); + + // Do a quick scan to see if we have this GEP nearby. If so, reuse it. + unsigned ScanLimit = 6; + BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); + // Scanning starts from the last instruction before the insertion point. + BasicBlock::iterator IP = Builder.GetInsertPoint(); + if (IP != BlockBegin) { + --IP; + for (; ScanLimit; --IP, --ScanLimit) { + // Don't count dbg.value against the ScanLimit, to avoid perturbing the + // generated code. + if (isa(IP)) + ScanLimit++; + if (IP->getOpcode() == Instruction::GetElementPtr && + IP->getOperand(0) == V && IP->getOperand(1) == Idx) + return IP; + if (IP == BlockBegin) break; + } + } + + // Save the original insertion point so we can restore it when we're done. + BuilderType::InsertPointGuard Guard(Builder); + + // Move the insertion point out of as many loops as we can. + while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { + if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break; + BasicBlock *Preheader = L->getLoopPreheader(); + if (!Preheader) break; + + // Ok, move up a level. + Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); + } + + // Emit a GEP. + Value *GEP = Builder.CreateGEP(V, Idx, "uglygep"); + rememberInstruction(GEP); + + return GEP; + } + + // Save the original insertion point so we can restore it when we're done. + BuilderType::InsertPoint SaveInsertPt = Builder.saveIP(); + + // Move the insertion point out of as many loops as we can. + while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { + if (!L->isLoopInvariant(V)) break; + + bool AnyIndexNotLoopInvariant = false; + for (SmallVectorImpl::const_iterator I = GepIndices.begin(), + E = GepIndices.end(); I != E; ++I) + if (!L->isLoopInvariant(*I)) { + AnyIndexNotLoopInvariant = true; + break; + } + if (AnyIndexNotLoopInvariant) + break; + + BasicBlock *Preheader = L->getLoopPreheader(); + if (!Preheader) break; + + // Ok, move up a level. + Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); + } + + // Insert a pretty getelementptr. Note that this GEP is not marked inbounds, + // because ScalarEvolution may have changed the address arithmetic to + // compute a value which is beyond the end of the allocated object. + Value *Casted = V; + if (V->getType() != PTy) + Casted = InsertNoopCastOfTo(Casted, PTy); + Value *GEP = Builder.CreateGEP(Casted, + GepIndices, + "scevgep"); + Ops.push_back(SE.getUnknown(GEP)); + rememberInstruction(GEP); + + // Restore the original insert point. + Builder.restoreIP(SaveInsertPt); + + return expand(SE.getAddExpr(Ops)); +} + +/// PickMostRelevantLoop - Given two loops pick the one that's most relevant for +/// SCEV expansion. If they are nested, this is the most nested. If they are +/// neighboring, pick the later. +static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B, + DominatorTree &DT) { + if (!A) return B; + if (!B) return A; + if (A->contains(B)) return B; + if (B->contains(A)) return A; + if (DT.dominates(A->getHeader(), B->getHeader())) return B; + if (DT.dominates(B->getHeader(), A->getHeader())) return A; + return A; // Arbitrarily break the tie. +} + +/// getRelevantLoop - Get the most relevant loop associated with the given +/// expression, according to PickMostRelevantLoop. +const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) { + // Test whether we've already computed the most relevant loop for this SCEV. + std::pair::iterator, bool> Pair = + RelevantLoops.insert(std::make_pair(S, static_cast(0))); + if (!Pair.second) + return Pair.first->second; + + if (isa(S)) + // A constant has no relevant loops. + return 0; + if (const SCEVUnknown *U = dyn_cast(S)) { + if (const Instruction *I = dyn_cast(U->getValue())) + return Pair.first->second = SE.LI->getLoopFor(I->getParent()); + // A non-instruction has no relevant loops. + return 0; + } + if (const SCEVNAryExpr *N = dyn_cast(S)) { + const Loop *L = 0; + if (const SCEVAddRecExpr *AR = dyn_cast(S)) + L = AR->getLoop(); + for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end(); + I != E; ++I) + L = PickMostRelevantLoop(L, getRelevantLoop(*I), *SE.DT); + return RelevantLoops[N] = L; + } + if (const SCEVCastExpr *C = dyn_cast(S)) { + const Loop *Result = getRelevantLoop(C->getOperand()); + return RelevantLoops[C] = Result; + } + if (const SCEVUDivExpr *D = dyn_cast(S)) { + const Loop *Result = + PickMostRelevantLoop(getRelevantLoop(D->getLHS()), + getRelevantLoop(D->getRHS()), + *SE.DT); + return RelevantLoops[D] = Result; + } + llvm_unreachable("Unexpected SCEV type!"); +} + +namespace { + +/// LoopCompare - Compare loops by PickMostRelevantLoop. +class LoopCompare { + DominatorTree &DT; +public: + explicit LoopCompare(DominatorTree &dt) : DT(dt) {} + + bool operator()(std::pair LHS, + std::pair RHS) const { + // Keep pointer operands sorted at the end. + if (LHS.second->getType()->isPointerTy() != + RHS.second->getType()->isPointerTy()) + return LHS.second->getType()->isPointerTy(); + + // Compare loops with PickMostRelevantLoop. + if (LHS.first != RHS.first) + return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first; + + // If one operand is a non-constant negative and the other is not, + // put the non-constant negative on the right so that a sub can + // be used instead of a negate and add. + if (LHS.second->isNonConstantNegative()) { + if (!RHS.second->isNonConstantNegative()) + return false; + } else if (RHS.second->isNonConstantNegative()) + return true; + + // Otherwise they are equivalent according to this comparison. + return false; + } +}; + +} + +Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) { + Type *Ty = SE.getEffectiveSCEVType(S->getType()); + + // Collect all the add operands in a loop, along with their associated loops. + // Iterate in reverse so that constants are emitted last, all else equal, and + // so that pointer operands are inserted first, which the code below relies on + // to form more involved GEPs. + SmallVector, 8> OpsAndLoops; + for (std::reverse_iterator I(S->op_end()), + E(S->op_begin()); I != E; ++I) + OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); + + // Sort by loop. Use a stable sort so that constants follow non-constants and + // pointer operands precede non-pointer operands. + std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT)); + + // Emit instructions to add all the operands. Hoist as much as possible + // out of loops, and form meaningful getelementptrs where possible. + Value *Sum = 0; + for (SmallVectorImpl >::iterator + I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) { + const Loop *CurLoop = I->first; + const SCEV *Op = I->second; + if (!Sum) { + // This is the first operand. Just expand it. + Sum = expand(Op); + ++I; + } else if (PointerType *PTy = dyn_cast(Sum->getType())) { + // The running sum expression is a pointer. Try to form a getelementptr + // at this level with that as the base. + SmallVector NewOps; + for (; I != E && I->first == CurLoop; ++I) { + // If the operand is SCEVUnknown and not instructions, peek through + // it, to enable more of it to be folded into the GEP. + const SCEV *X = I->second; + if (const SCEVUnknown *U = dyn_cast(X)) + if (!isa(U->getValue())) + X = SE.getSCEV(U->getValue()); + NewOps.push_back(X); + } + Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum); + } else if (PointerType *PTy = dyn_cast(Op->getType())) { + // The running sum is an integer, and there's a pointer at this level. + // Try to form a getelementptr. If the running sum is instructions, + // use a SCEVUnknown to avoid re-analyzing them. + SmallVector NewOps; + NewOps.push_back(isa(Sum) ? SE.getUnknown(Sum) : + SE.getSCEV(Sum)); + for (++I; I != E && I->first == CurLoop; ++I) + NewOps.push_back(I->second); + Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op)); + } else if (Op->isNonConstantNegative()) { + // Instead of doing a negate and add, just do a subtract. + Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty); + Sum = InsertNoopCastOfTo(Sum, Ty); + Sum = InsertBinop(Instruction::Sub, Sum, W); + ++I; + } else { + // A simple add. + Value *W = expandCodeFor(Op, Ty); + Sum = InsertNoopCastOfTo(Sum, Ty); + // Canonicalize a constant to the RHS. + if (isa(Sum)) std::swap(Sum, W); + Sum = InsertBinop(Instruction::Add, Sum, W); + ++I; + } + } + + return Sum; +} + +Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) { + Type *Ty = SE.getEffectiveSCEVType(S->getType()); + + // Collect all the mul operands in a loop, along with their associated loops. + // Iterate in reverse so that constants are emitted last, all else equal. + SmallVector, 8> OpsAndLoops; + for (std::reverse_iterator I(S->op_end()), + E(S->op_begin()); I != E; ++I) + OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); + + // Sort by loop. Use a stable sort so that constants follow non-constants. + std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT)); + + // Emit instructions to mul all the operands. Hoist as much as possible + // out of loops. + Value *Prod = 0; + for (SmallVectorImpl >::iterator + I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) { + const SCEV *Op = I->second; + if (!Prod) { + // This is the first operand. Just expand it. + Prod = expand(Op); + ++I; + } else if (Op->isAllOnesValue()) { + // Instead of doing a multiply by negative one, just do a negate. + Prod = InsertNoopCastOfTo(Prod, Ty); + Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod); + ++I; + } else { + // A simple mul. + Value *W = expandCodeFor(Op, Ty); + Prod = InsertNoopCastOfTo(Prod, Ty); + // Canonicalize a constant to the RHS. + if (isa(Prod)) std::swap(Prod, W); + Prod = InsertBinop(Instruction::Mul, Prod, W); + ++I; + } + } + + return Prod; +} + +Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) { + Type *Ty = SE.getEffectiveSCEVType(S->getType()); + + Value *LHS = expandCodeFor(S->getLHS(), Ty); + if (const SCEVConstant *SC = dyn_cast(S->getRHS())) { + const APInt &RHS = SC->getValue()->getValue(); + if (RHS.isPowerOf2()) + return InsertBinop(Instruction::LShr, LHS, + ConstantInt::get(Ty, RHS.logBase2())); + } + + Value *RHS = expandCodeFor(S->getRHS(), Ty); + return InsertBinop(Instruction::UDiv, LHS, RHS); +} + +/// Move parts of Base into Rest to leave Base with the minimal +/// expression that provides a pointer operand suitable for a +/// GEP expansion. +static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest, + ScalarEvolution &SE) { + while (const SCEVAddRecExpr *A = dyn_cast(Base)) { + Base = A->getStart(); + Rest = SE.getAddExpr(Rest, + SE.getAddRecExpr(SE.getConstant(A->getType(), 0), + A->getStepRecurrence(SE), + A->getLoop(), + A->getNoWrapFlags(SCEV::FlagNW))); + } + if (const SCEVAddExpr *A = dyn_cast(Base)) { + Base = A->getOperand(A->getNumOperands()-1); + SmallVector NewAddOps(A->op_begin(), A->op_end()); + NewAddOps.back() = Rest; + Rest = SE.getAddExpr(NewAddOps); + ExposePointerBase(Base, Rest, SE); + } +} + +/// Determine if this is a well-behaved chain of instructions leading back to +/// the PHI. If so, it may be reused by expanded expressions. +bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV, + const Loop *L) { + if (IncV->getNumOperands() == 0 || isa(IncV) || + (isa(IncV) && !isa(IncV))) + return false; + // If any of the operands don't dominate the insert position, bail. + // Addrec operands are always loop-invariant, so this can only happen + // if there are instructions which haven't been hoisted. + if (L == IVIncInsertLoop) { + for (User::op_iterator OI = IncV->op_begin()+1, + OE = IncV->op_end(); OI != OE; ++OI) + if (Instruction *OInst = dyn_cast(OI)) + if (!SE.DT->dominates(OInst, IVIncInsertPos)) + return false; + } + // Advance to the next instruction. + IncV = dyn_cast(IncV->getOperand(0)); + if (!IncV) + return false; + + if (IncV->mayHaveSideEffects()) + return false; + + if (IncV != PN) + return true; + + return isNormalAddRecExprPHI(PN, IncV, L); } -Value *SCEVExpander::visitAddRecExpr(SCEVAddRecExpr *S) { - const Type *Ty = S->getType(); +/// getIVIncOperand returns an induction variable increment's induction +/// variable operand. +/// +/// If allowScale is set, any type of GEP is allowed as long as the nonIV +/// operands dominate InsertPos. +/// +/// If allowScale is not set, ensure that a GEP increment conforms to one of the +/// simple patterns generated by getAddRecExprPHILiterally and +/// expandAddtoGEP. If the pattern isn't recognized, return NULL. +Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV, + Instruction *InsertPos, + bool allowScale) { + if (IncV == InsertPos) + return NULL; + + switch (IncV->getOpcode()) { + default: + return NULL; + // Check for a simple Add/Sub or GEP of a loop invariant step. + case Instruction::Add: + case Instruction::Sub: { + Instruction *OInst = dyn_cast(IncV->getOperand(1)); + if (!OInst || SE.DT->dominates(OInst, InsertPos)) + return dyn_cast(IncV->getOperand(0)); + return NULL; + } + case Instruction::BitCast: + return dyn_cast(IncV->getOperand(0)); + case Instruction::GetElementPtr: + for (Instruction::op_iterator I = IncV->op_begin()+1, E = IncV->op_end(); + I != E; ++I) { + if (isa(*I)) + continue; + if (Instruction *OInst = dyn_cast(*I)) { + if (!SE.DT->dominates(OInst, InsertPos)) + return NULL; + } + if (allowScale) { + // allow any kind of GEP as long as it can be hoisted. + continue; + } + // This must be a pointer addition of constants (pretty), which is already + // handled, or some number of address-size elements (ugly). Ugly geps + // have 2 operands. i1* is used by the expander to represent an + // address-size element. + if (IncV->getNumOperands() != 2) + return NULL; + unsigned AS = cast(IncV->getType())->getAddressSpace(); + if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS) + && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS)) + return NULL; + break; + } + return dyn_cast(IncV->getOperand(0)); + } +} + +/// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make +/// it available to other uses in this loop. Recursively hoist any operands, +/// until we reach a value that dominates InsertPos. +bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) { + if (SE.DT->dominates(IncV, InsertPos)) + return true; + + // InsertPos must itself dominate IncV so that IncV's new position satisfies + // its existing users. + if (isa(InsertPos) + || !SE.DT->dominates(InsertPos->getParent(), IncV->getParent())) + return false; + + // Check that the chain of IV operands leading back to Phi can be hoisted. + SmallVector IVIncs; + for(;;) { + Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true); + if (!Oper) + return false; + // IncV is safe to hoist. + IVIncs.push_back(IncV); + IncV = Oper; + if (SE.DT->dominates(IncV, InsertPos)) + break; + } + for (SmallVectorImpl::reverse_iterator I = IVIncs.rbegin(), + E = IVIncs.rend(); I != E; ++I) { + (*I)->moveBefore(InsertPos); + } + return true; +} + +/// Determine if this cyclic phi is in a form that would have been generated by +/// LSR. We don't care if the phi was actually expanded in this pass, as long +/// as it is in a low-cost form, for example, no implied multiplication. This +/// should match any patterns generated by getAddRecExprPHILiterally and +/// expandAddtoGEP. +bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV, + const Loop *L) { + for(Instruction *IVOper = IncV; + (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(), + /*allowScale=*/false));) { + if (IVOper == PN) + return true; + } + return false; +} + +/// expandIVInc - Expand an IV increment at Builder's current InsertPos. +/// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may +/// need to materialize IV increments elsewhere to handle difficult situations. +Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L, + Type *ExpandTy, Type *IntTy, + bool useSubtract) { + Value *IncV; + // If the PHI is a pointer, use a GEP, otherwise use an add or sub. + if (ExpandTy->isPointerTy()) { + PointerType *GEPPtrTy = cast(ExpandTy); + // If the step isn't constant, don't use an implicitly scaled GEP, because + // that would require a multiply inside the loop. + if (!isa(StepV)) + GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()), + GEPPtrTy->getAddressSpace()); + const SCEV *const StepArray[1] = { SE.getSCEV(StepV) }; + IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN); + if (IncV->getType() != PN->getType()) { + IncV = Builder.CreateBitCast(IncV, PN->getType()); + rememberInstruction(IncV); + } + } else { + IncV = useSubtract ? + Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") : + Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next"); + rememberInstruction(IncV); + } + return IncV; +} + +/// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand +/// the base addrec, which is the addrec without any non-loop-dominating +/// values, and return the PHI. +PHINode * +SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized, + const Loop *L, + Type *ExpandTy, + Type *IntTy) { + assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position"); + + // Reuse a previously-inserted PHI, if present. + BasicBlock *LatchBlock = L->getLoopLatch(); + if (LatchBlock) { + for (BasicBlock::iterator I = L->getHeader()->begin(); + PHINode *PN = dyn_cast(I); ++I) { + if (!SE.isSCEVable(PN->getType()) || + (SE.getEffectiveSCEVType(PN->getType()) != + SE.getEffectiveSCEVType(Normalized->getType())) || + SE.getSCEV(PN) != Normalized) + continue; + + Instruction *IncV = + cast(PN->getIncomingValueForBlock(LatchBlock)); + + if (LSRMode) { + if (!isExpandedAddRecExprPHI(PN, IncV, L)) + continue; + if (L == IVIncInsertLoop && !hoistIVInc(IncV, IVIncInsertPos)) + continue; + } + else { + if (!isNormalAddRecExprPHI(PN, IncV, L)) + continue; + if (L == IVIncInsertLoop) + do { + if (SE.DT->dominates(IncV, IVIncInsertPos)) + break; + // Make sure the increment is where we want it. But don't move it + // down past a potential existing post-inc user. + IncV->moveBefore(IVIncInsertPos); + IVIncInsertPos = IncV; + IncV = cast(IncV->getOperand(0)); + } while (IncV != PN); + } + // Ok, the add recurrence looks usable. + // Remember this PHI, even in post-inc mode. + InsertedValues.insert(PN); + // Remember the increment. + rememberInstruction(IncV); + return PN; + } + } + + // Save the original insertion point so we can restore it when we're done. + BuilderType::InsertPointGuard Guard(Builder); + + // Another AddRec may need to be recursively expanded below. For example, if + // this AddRec is quadratic, the StepV may itself be an AddRec in this + // loop. Remove this loop from the PostIncLoops set before expanding such + // AddRecs. Otherwise, we cannot find a valid position for the step + // (i.e. StepV can never dominate its loop header). Ideally, we could do + // SavedIncLoops.swap(PostIncLoops), but we generally have a single element, + // so it's not worth implementing SmallPtrSet::swap. + PostIncLoopSet SavedPostIncLoops = PostIncLoops; + PostIncLoops.clear(); + + // Expand code for the start value. + Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy, + L->getHeader()->begin()); + + // StartV must be hoisted into L's preheader to dominate the new phi. + assert(!isa(StartV) || + SE.DT->properlyDominates(cast(StartV)->getParent(), + L->getHeader())); + + // Expand code for the step value. Do this before creating the PHI so that PHI + // reuse code doesn't see an incomplete PHI. + const SCEV *Step = Normalized->getStepRecurrence(SE); + // If the stride is negative, insert a sub instead of an add for the increment + // (unless it's a constant, because subtracts of constants are canonicalized + // to adds). + bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); + if (useSubtract) + Step = SE.getNegativeSCEV(Step); + // Expand the step somewhere that dominates the loop header. + Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin()); + + // Create the PHI. + BasicBlock *Header = L->getHeader(); + Builder.SetInsertPoint(Header, Header->begin()); + pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); + PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE), + Twine(IVName) + ".iv"); + rememberInstruction(PN); + + // Create the step instructions and populate the PHI. + for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { + BasicBlock *Pred = *HPI; + + // Add a start value. + if (!L->contains(Pred)) { + PN->addIncoming(StartV, Pred); + continue; + } + + // Create a step value and add it to the PHI. + // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the + // instructions at IVIncInsertPos. + Instruction *InsertPos = L == IVIncInsertLoop ? + IVIncInsertPos : Pred->getTerminator(); + Builder.SetInsertPoint(InsertPos); + Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); + if (isa(IncV)) { + if (Normalized->getNoWrapFlags(SCEV::FlagNUW)) + cast(IncV)->setHasNoUnsignedWrap(); + if (Normalized->getNoWrapFlags(SCEV::FlagNSW)) + cast(IncV)->setHasNoSignedWrap(); + } + PN->addIncoming(IncV, Pred); + } + + // After expanding subexpressions, restore the PostIncLoops set so the caller + // can ensure that IVIncrement dominates the current uses. + PostIncLoops = SavedPostIncLoops; + + // Remember this PHI, even in post-inc mode. + InsertedValues.insert(PN); + + return PN; +} + +Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) { + Type *STy = S->getType(); + Type *IntTy = SE.getEffectiveSCEVType(STy); const Loop *L = S->getLoop(); - // We cannot yet do fp recurrences, e.g. the xform of {X,+,F} --> X+{0,+,F} - assert(Ty->isInteger() && "Cannot expand fp recurrences yet!"); + + // Determine a normalized form of this expression, which is the expression + // before any post-inc adjustment is made. + const SCEVAddRecExpr *Normalized = S; + if (PostIncLoops.count(L)) { + PostIncLoopSet Loops; + Loops.insert(L); + Normalized = + cast(TransformForPostIncUse(Normalize, S, 0, 0, + Loops, SE, *SE.DT)); + } + + // Strip off any non-loop-dominating component from the addrec start. + const SCEV *Start = Normalized->getStart(); + const SCEV *PostLoopOffset = 0; + if (!SE.properlyDominates(Start, L->getHeader())) { + PostLoopOffset = Start; + Start = SE.getConstant(Normalized->getType(), 0); + Normalized = cast( + SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE), + Normalized->getLoop(), + Normalized->getNoWrapFlags(SCEV::FlagNW))); + } + + // Strip off any non-loop-dominating component from the addrec step. + const SCEV *Step = Normalized->getStepRecurrence(SE); + const SCEV *PostLoopScale = 0; + if (!SE.dominates(Step, L->getHeader())) { + PostLoopScale = Step; + Step = SE.getConstant(Normalized->getType(), 1); + Normalized = + cast(SE.getAddRecExpr( + Start, Step, Normalized->getLoop(), + Normalized->getNoWrapFlags(SCEV::FlagNW))); + } + + // Expand the core addrec. If we need post-loop scaling, force it to + // expand to an integer type to avoid the need for additional casting. + Type *ExpandTy = PostLoopScale ? IntTy : STy; + PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy); + + // Accommodate post-inc mode, if necessary. + Value *Result; + if (!PostIncLoops.count(L)) + Result = PN; + else { + // In PostInc mode, use the post-incremented value. + BasicBlock *LatchBlock = L->getLoopLatch(); + assert(LatchBlock && "PostInc mode requires a unique loop latch!"); + Result = PN->getIncomingValueForBlock(LatchBlock); + + // For an expansion to use the postinc form, the client must call + // expandCodeFor with an InsertPoint that is either outside the PostIncLoop + // or dominated by IVIncInsertPos. + if (isa(Result) + && !SE.DT->dominates(cast(Result), + Builder.GetInsertPoint())) { + // The induction variable's postinc expansion does not dominate this use. + // IVUsers tries to prevent this case, so it is rare. However, it can + // happen when an IVUser outside the loop is not dominated by the latch + // block. Adjusting IVIncInsertPos before expansion begins cannot handle + // all cases. Consider a phi outide whose operand is replaced during + // expansion with the value of the postinc user. Without fundamentally + // changing the way postinc users are tracked, the only remedy is + // inserting an extra IV increment. StepV might fold into PostLoopOffset, + // but hopefully expandCodeFor handles that. + bool useSubtract = + !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); + if (useSubtract) + Step = SE.getNegativeSCEV(Step); + Value *StepV; + { + // Expand the step somewhere that dominates the loop header. + BuilderType::InsertPointGuard Guard(Builder); + StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin()); + } + Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); + } + } + + // Re-apply any non-loop-dominating scale. + if (PostLoopScale) { + assert(S->isAffine() && "Can't linearly scale non-affine recurrences."); + Result = InsertNoopCastOfTo(Result, IntTy); + Result = Builder.CreateMul(Result, + expandCodeFor(PostLoopScale, IntTy)); + rememberInstruction(Result); + } + + // Re-apply any non-loop-dominating offset. + if (PostLoopOffset) { + if (PointerType *PTy = dyn_cast(ExpandTy)) { + const SCEV *const OffsetArray[1] = { PostLoopOffset }; + Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result); + } else { + Result = InsertNoopCastOfTo(Result, IntTy); + Result = Builder.CreateAdd(Result, + expandCodeFor(PostLoopOffset, IntTy)); + rememberInstruction(Result); + } + } + + return Result; +} + +Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) { + if (!CanonicalMode) return expandAddRecExprLiterally(S); + + Type *Ty = SE.getEffectiveSCEVType(S->getType()); + const Loop *L = S->getLoop(); + + // First check for an existing canonical IV in a suitable type. + PHINode *CanonicalIV = 0; + if (PHINode *PN = L->getCanonicalInductionVariable()) + if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty)) + CanonicalIV = PN; + + // Rewrite an AddRec in terms of the canonical induction variable, if + // its type is more narrow. + if (CanonicalIV && + SE.getTypeSizeInBits(CanonicalIV->getType()) > + SE.getTypeSizeInBits(Ty)) { + SmallVector NewOps(S->getNumOperands()); + for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i) + NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType()); + Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(), + S->getNoWrapFlags(SCEV::FlagNW))); + BasicBlock::iterator NewInsertPt = + llvm::next(BasicBlock::iterator(cast(V))); + BuilderType::InsertPointGuard Guard(Builder); + while (isa(NewInsertPt) || isa(NewInsertPt) || + isa(NewInsertPt)) + ++NewInsertPt; + V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0, + NewInsertPt); + return V; + } // {X,+,F} --> X + {0,+,F} - if (!isa(S->getStart()) || - !cast(S->getStart())->getValue()->isZero()) { - Value *Start = expand(S->getStart()); - std::vector NewOps(S->op_begin(), S->op_end()); - NewOps[0] = SE.getIntegerSCEV(0, Ty); - Value *Rest = expand(SE.getAddRecExpr(NewOps, L)); + if (!S->getStart()->isZero()) { + SmallVector NewOps(S->op_begin(), S->op_end()); + NewOps[0] = SE.getConstant(Ty, 0); + const SCEV *Rest = SE.getAddRecExpr(NewOps, L, + S->getNoWrapFlags(SCEV::FlagNW)); - // FIXME: look for an existing add to use. - return InsertBinop(Instruction::Add, Rest, Start, InsertPt); + // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the + // comments on expandAddToGEP for details. + const SCEV *Base = S->getStart(); + const SCEV *RestArray[1] = { Rest }; + // Dig into the expression to find the pointer base for a GEP. + ExposePointerBase(Base, RestArray[0], SE); + // If we found a pointer, expand the AddRec with a GEP. + if (PointerType *PTy = dyn_cast(Base->getType())) { + // Make sure the Base isn't something exotic, such as a multiplied + // or divided pointer value. In those cases, the result type isn't + // actually a pointer type. + if (!isa(Base) && !isa(Base)) { + Value *StartV = expand(Base); + assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!"); + return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV); + } + } + + // Just do a normal add. Pre-expand the operands to suppress folding. + return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())), + SE.getUnknown(expand(Rest)))); } - // {0,+,1} --> Insert a canonical induction variable into the loop! - if (S->getNumOperands() == 2 && - S->getOperand(1) == SE.getIntegerSCEV(1, Ty)) { + // If we don't yet have a canonical IV, create one. + if (!CanonicalIV) { // Create and insert the PHI node for the induction variable in the // specified loop. BasicBlock *Header = L->getHeader(); - PHINode *PN = new PHINode(Ty, "indvar", Header->begin()); - PN->addIncoming(Constant::getNullValue(Ty), L->getLoopPreheader()); - - pred_iterator HPI = pred_begin(Header); - assert(HPI != pred_end(Header) && "Loop with zero preds???"); - if (!L->contains(*HPI)) ++HPI; - assert(HPI != pred_end(Header) && L->contains(*HPI) && - "No backedge in loop?"); + pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); + CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar", + Header->begin()); + rememberInstruction(CanonicalIV); - // Insert a unit add instruction right before the terminator corresponding - // to the back-edge. + SmallSet PredSeen; Constant *One = ConstantInt::get(Ty, 1); - Instruction *Add = BinaryOperator::createAdd(PN, One, "indvar.next", - (*HPI)->getTerminator()); + for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { + BasicBlock *HP = *HPI; + if (!PredSeen.insert(HP)) + continue; - pred_iterator PI = pred_begin(Header); - if (*PI == L->getLoopPreheader()) - ++PI; - PN->addIncoming(Add, *PI); - return PN; + if (L->contains(HP)) { + // Insert a unit add instruction right before the terminator + // corresponding to the back-edge. + Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One, + "indvar.next", + HP->getTerminator()); + Add->setDebugLoc(HP->getTerminator()->getDebugLoc()); + rememberInstruction(Add); + CanonicalIV->addIncoming(Add, HP); + } else { + CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP); + } + } + } + + // {0,+,1} --> Insert a canonical induction variable into the loop! + if (S->isAffine() && S->getOperand(1)->isOne()) { + assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) && + "IVs with types different from the canonical IV should " + "already have been handled!"); + return CanonicalIV; } - // Get the canonical induction variable I for this loop. - Value *I = getOrInsertCanonicalInductionVariable(L, Ty); + // {0,+,F} --> {0,+,1} * F // If this is a simple linear addrec, emit it now as a special case. - if (S->getNumOperands() == 2) { // {0,+,F} --> i*F - Value *F = expand(S->getOperand(1)); - - // IF the step is by one, just return the inserted IV. - if (ConstantInt *CI = dyn_cast(F)) - if (CI->getValue() == 1) - return I; - - // If the insert point is directly inside of the loop, emit the multiply at - // the insert point. Otherwise, L is a loop that is a parent of the insert - // point loop. If we can, move the multiply to the outer most loop that it - // is safe to be in. - Instruction *MulInsertPt = InsertPt; - Loop *InsertPtLoop = LI.getLoopFor(MulInsertPt->getParent()); - if (InsertPtLoop != L && InsertPtLoop && - L->contains(InsertPtLoop->getHeader())) { - while (InsertPtLoop != L) { - // If we cannot hoist the multiply out of this loop, don't. - if (!InsertPtLoop->isLoopInvariant(F)) break; - - // Otherwise, move the insert point to the preheader of the loop. - MulInsertPt = InsertPtLoop->getLoopPreheader()->getTerminator(); - InsertPtLoop = InsertPtLoop->getParentLoop(); - } - } - - return InsertBinop(Instruction::Mul, I, F, MulInsertPt); - } + if (S->isAffine()) // {0,+,F} --> i*F + return + expand(SE.getTruncateOrNoop( + SE.getMulExpr(SE.getUnknown(CanonicalIV), + SE.getNoopOrAnyExtend(S->getOperand(1), + CanonicalIV->getType())), + Ty)); // If this is a chain of recurrences, turn it into a closed form, using the // folders, then expandCodeFor the closed form. This allows the folders to // simplify the expression without having to build a bunch of special code // into this folder. - SCEVHandle IH = SE.getUnknown(I); // Get I as a "symbolic" SCEV. + const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV. + + // Promote S up to the canonical IV type, if the cast is foldable. + const SCEV *NewS = S; + const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType()); + if (isa(Ext)) + NewS = Ext; - SCEVHandle V = S->evaluateAtIteration(IH, SE); + const SCEV *V = cast(NewS)->evaluateAtIteration(IH, SE); //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n"; - return expand(V); + // Truncate the result down to the original type, if needed. + const SCEV *T = SE.getTruncateOrNoop(V, Ty); + return expand(T); } -Value *SCEVExpander::visitSMaxExpr(SCEVSMaxExpr *S) { - Value *LHS = expand(S->getOperand(0)); - for (unsigned i = 1; i < S->getNumOperands(); ++i) { - Value *RHS = expand(S->getOperand(i)); - Value *ICmp = new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS, "tmp", InsertPt); - LHS = new SelectInst(ICmp, LHS, RHS, "smax", InsertPt); +Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) { + Type *Ty = SE.getEffectiveSCEVType(S->getType()); + Value *V = expandCodeFor(S->getOperand(), + SE.getEffectiveSCEVType(S->getOperand()->getType())); + Value *I = Builder.CreateTrunc(V, Ty); + rememberInstruction(I); + return I; +} + +Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) { + Type *Ty = SE.getEffectiveSCEVType(S->getType()); + Value *V = expandCodeFor(S->getOperand(), + SE.getEffectiveSCEVType(S->getOperand()->getType())); + Value *I = Builder.CreateZExt(V, Ty); + rememberInstruction(I); + return I; +} + +Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) { + Type *Ty = SE.getEffectiveSCEVType(S->getType()); + Value *V = expandCodeFor(S->getOperand(), + SE.getEffectiveSCEVType(S->getOperand()->getType())); + Value *I = Builder.CreateSExt(V, Ty); + rememberInstruction(I); + return I; +} + +Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) { + Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); + Type *Ty = LHS->getType(); + for (int i = S->getNumOperands()-2; i >= 0; --i) { + // In the case of mixed integer and pointer types, do the + // rest of the comparisons as integer. + if (S->getOperand(i)->getType() != Ty) { + Ty = SE.getEffectiveSCEVType(Ty); + LHS = InsertNoopCastOfTo(LHS, Ty); + } + Value *RHS = expandCodeFor(S->getOperand(i), Ty); + Value *ICmp = Builder.CreateICmpSGT(LHS, RHS); + rememberInstruction(ICmp); + Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax"); + rememberInstruction(Sel); + LHS = Sel; + } + // In the case of mixed integer and pointer types, cast the + // final result back to the pointer type. + if (LHS->getType() != S->getType()) + LHS = InsertNoopCastOfTo(LHS, S->getType()); + return LHS; +} + +Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) { + Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); + Type *Ty = LHS->getType(); + for (int i = S->getNumOperands()-2; i >= 0; --i) { + // In the case of mixed integer and pointer types, do the + // rest of the comparisons as integer. + if (S->getOperand(i)->getType() != Ty) { + Ty = SE.getEffectiveSCEVType(Ty); + LHS = InsertNoopCastOfTo(LHS, Ty); + } + Value *RHS = expandCodeFor(S->getOperand(i), Ty); + Value *ICmp = Builder.CreateICmpUGT(LHS, RHS); + rememberInstruction(ICmp); + Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax"); + rememberInstruction(Sel); + LHS = Sel; } + // In the case of mixed integer and pointer types, cast the + // final result back to the pointer type. + if (LHS->getType() != S->getType()) + LHS = InsertNoopCastOfTo(LHS, S->getType()); return LHS; } -Value *SCEVExpander::expand(SCEV *S) { - // Check to see if we already expanded this. - std::map::iterator I = InsertedExpressions.find(S); +Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty, + Instruction *IP) { + Builder.SetInsertPoint(IP->getParent(), IP); + return expandCodeFor(SH, Ty); +} + +Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) { + // Expand the code for this SCEV. + Value *V = expand(SH); + if (Ty) { + assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) && + "non-trivial casts should be done with the SCEVs directly!"); + V = InsertNoopCastOfTo(V, Ty); + } + return V; +} + +Value *SCEVExpander::expand(const SCEV *S) { + // Compute an insertion point for this SCEV object. Hoist the instructions + // as far out in the loop nest as possible. + Instruction *InsertPt = Builder.GetInsertPoint(); + for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ; + L = L->getParentLoop()) + if (SE.isLoopInvariant(S, L)) { + if (!L) break; + if (BasicBlock *Preheader = L->getLoopPreheader()) + InsertPt = Preheader->getTerminator(); + else { + // LSR sets the insertion point for AddRec start/step values to the + // block start to simplify value reuse, even though it's an invalid + // position. SCEVExpander must correct for this in all cases. + InsertPt = L->getHeader()->getFirstInsertionPt(); + } + } else { + // If the SCEV is computable at this level, insert it into the header + // after the PHIs (and after any other instructions that we've inserted + // there) so that it is guaranteed to dominate any user inside the loop. + if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L)) + InsertPt = L->getHeader()->getFirstInsertionPt(); + while (InsertPt != Builder.GetInsertPoint() + && (isInsertedInstruction(InsertPt) + || isa(InsertPt))) { + InsertPt = llvm::next(BasicBlock::iterator(InsertPt)); + } + break; + } + + // Check to see if we already expanded this here. + std::map, TrackingVH >::iterator + I = InsertedExpressions.find(std::make_pair(S, InsertPt)); if (I != InsertedExpressions.end()) return I->second; - + + BuilderType::InsertPointGuard Guard(Builder); + Builder.SetInsertPoint(InsertPt->getParent(), InsertPt); + + // Expand the expression into instructions. Value *V = visit(S); - InsertedExpressions[S] = V; + + // Remember the expanded value for this SCEV at this location. + // + // This is independent of PostIncLoops. The mapped value simply materializes + // the expression at this insertion point. If the mapped value happened to be + // a postinc expansion, it could be reused by a non-postinc user, but only if + // its insertion point was already at the head of the loop. + InsertedExpressions[std::make_pair(S, InsertPt)] = V; return V; } +void SCEVExpander::rememberInstruction(Value *I) { + if (!PostIncLoops.empty()) + InsertedPostIncValues.insert(I); + else + InsertedValues.insert(I); +} + +/// getOrInsertCanonicalInductionVariable - This method returns the +/// canonical induction variable of the specified type for the specified +/// loop (inserting one if there is none). A canonical induction variable +/// starts at zero and steps by one on each iteration. +PHINode * +SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L, + Type *Ty) { + assert(Ty->isIntegerTy() && "Can only insert integer induction variables!"); + + // Build a SCEV for {0,+,1}. + // Conservatively use FlagAnyWrap for now. + const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0), + SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap); + + // Emit code for it. + BuilderType::InsertPointGuard Guard(Builder); + PHINode *V = cast(expandCodeFor(H, 0, L->getHeader()->begin())); + + return V; +} + +/// Sort values by integer width for replaceCongruentIVs. +static bool width_descending(Value *lhs, Value *rhs) { + // Put pointers at the back and make sure pointer < pointer = false. + if (!lhs->getType()->isIntegerTy() || !rhs->getType()->isIntegerTy()) + return rhs->getType()->isIntegerTy() && !lhs->getType()->isIntegerTy(); + return rhs->getType()->getPrimitiveSizeInBits() + < lhs->getType()->getPrimitiveSizeInBits(); +} + +/// replaceCongruentIVs - Check for congruent phis in this loop header and +/// replace them with their most canonical representative. Return the number of +/// phis eliminated. +/// +/// This does not depend on any SCEVExpander state but should be used in +/// the same context that SCEVExpander is used. +unsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT, + SmallVectorImpl &DeadInsts, + const TargetTransformInfo *TTI) { + // Find integer phis in order of increasing width. + SmallVector Phis; + for (BasicBlock::iterator I = L->getHeader()->begin(); + PHINode *Phi = dyn_cast(I); ++I) { + Phis.push_back(Phi); + } + if (TTI) + std::sort(Phis.begin(), Phis.end(), width_descending); + + unsigned NumElim = 0; + DenseMap ExprToIVMap; + // Process phis from wide to narrow. Mapping wide phis to the their truncation + // so narrow phis can reuse them. + for (SmallVectorImpl::const_iterator PIter = Phis.begin(), + PEnd = Phis.end(); PIter != PEnd; ++PIter) { + PHINode *Phi = *PIter; + + // Fold constant phis. They may be congruent to other constant phis and + // would confuse the logic below that expects proper IVs. + if (Value *V = Phi->hasConstantValue()) { + Phi->replaceAllUsesWith(V); + DeadInsts.push_back(Phi); + ++NumElim; + DEBUG_WITH_TYPE(DebugType, dbgs() + << "INDVARS: Eliminated constant iv: " << *Phi << '\n'); + continue; + } + + if (!SE.isSCEVable(Phi->getType())) + continue; + + PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)]; + if (!OrigPhiRef) { + OrigPhiRef = Phi; + if (Phi->getType()->isIntegerTy() && TTI + && TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) { + // This phi can be freely truncated to the narrowest phi type. Map the + // truncated expression to it so it will be reused for narrow types. + const SCEV *TruncExpr = + SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType()); + ExprToIVMap[TruncExpr] = Phi; + } + continue; + } + + // Replacing a pointer phi with an integer phi or vice-versa doesn't make + // sense. + if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy()) + continue; + + if (BasicBlock *LatchBlock = L->getLoopLatch()) { + Instruction *OrigInc = + cast(OrigPhiRef->getIncomingValueForBlock(LatchBlock)); + Instruction *IsomorphicInc = + cast(Phi->getIncomingValueForBlock(LatchBlock)); + + // If this phi has the same width but is more canonical, replace the + // original with it. As part of the "more canonical" determination, + // respect a prior decision to use an IV chain. + if (OrigPhiRef->getType() == Phi->getType() + && !(ChainedPhis.count(Phi) + || isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L)) + && (ChainedPhis.count(Phi) + || isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) { + std::swap(OrigPhiRef, Phi); + std::swap(OrigInc, IsomorphicInc); + } + // Replacing the congruent phi is sufficient because acyclic redundancy + // elimination, CSE/GVN, should handle the rest. However, once SCEV proves + // that a phi is congruent, it's often the head of an IV user cycle that + // is isomorphic with the original phi. It's worth eagerly cleaning up the + // common case of a single IV increment so that DeleteDeadPHIs can remove + // cycles that had postinc uses. + const SCEV *TruncExpr = SE.getTruncateOrNoop(SE.getSCEV(OrigInc), + IsomorphicInc->getType()); + if (OrigInc != IsomorphicInc + && TruncExpr == SE.getSCEV(IsomorphicInc) + && ((isa(OrigInc) && isa(IsomorphicInc)) + || hoistIVInc(OrigInc, IsomorphicInc))) { + DEBUG_WITH_TYPE(DebugType, dbgs() + << "INDVARS: Eliminated congruent iv.inc: " + << *IsomorphicInc << '\n'); + Value *NewInc = OrigInc; + if (OrigInc->getType() != IsomorphicInc->getType()) { + Instruction *IP = isa(OrigInc) + ? (Instruction*)L->getHeader()->getFirstInsertionPt() + : OrigInc->getNextNode(); + IRBuilder<> Builder(IP); + Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc()); + NewInc = Builder. + CreateTruncOrBitCast(OrigInc, IsomorphicInc->getType(), IVName); + } + IsomorphicInc->replaceAllUsesWith(NewInc); + DeadInsts.push_back(IsomorphicInc); + } + } + DEBUG_WITH_TYPE(DebugType, dbgs() + << "INDVARS: Eliminated congruent iv: " << *Phi << '\n'); + ++NumElim; + Value *NewIV = OrigPhiRef; + if (OrigPhiRef->getType() != Phi->getType()) { + IRBuilder<> Builder(L->getHeader()->getFirstInsertionPt()); + Builder.SetCurrentDebugLocation(Phi->getDebugLoc()); + NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName); + } + Phi->replaceAllUsesWith(NewIV); + DeadInsts.push_back(Phi); + } + return NumElim; +} + +namespace { +// Search for a SCEV subexpression that is not safe to expand. Any expression +// that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely +// UDiv expressions. We don't know if the UDiv is derived from an IR divide +// instruction, but the important thing is that we prove the denominator is +// nonzero before expansion. +// +// IVUsers already checks that IV-derived expressions are safe. So this check is +// only needed when the expression includes some subexpression that is not IV +// derived. +// +// Currently, we only allow division by a nonzero constant here. If this is +// inadequate, we could easily allow division by SCEVUnknown by using +// ValueTracking to check isKnownNonZero(). +// +// We cannot generally expand recurrences unless the step dominates the loop +// header. The expander handles the special case of affine recurrences by +// scaling the recurrence outside the loop, but this technique isn't generally +// applicable. Expanding a nested recurrence outside a loop requires computing +// binomial coefficients. This could be done, but the recurrence has to be in a +// perfectly reduced form, which can't be guaranteed. +struct SCEVFindUnsafe { + ScalarEvolution &SE; + bool IsUnsafe; + + SCEVFindUnsafe(ScalarEvolution &se): SE(se), IsUnsafe(false) {} + + bool follow(const SCEV *S) { + if (const SCEVUDivExpr *D = dyn_cast(S)) { + const SCEVConstant *SC = dyn_cast(D->getRHS()); + if (!SC || SC->getValue()->isZero()) { + IsUnsafe = true; + return false; + } + } + if (const SCEVAddRecExpr *AR = dyn_cast(S)) { + const SCEV *Step = AR->getStepRecurrence(SE); + if (!AR->isAffine() && !SE.dominates(Step, AR->getLoop()->getHeader())) { + IsUnsafe = true; + return false; + } + } + return true; + } + bool isDone() const { return IsUnsafe; } +}; +} + +namespace llvm { +bool isSafeToExpand(const SCEV *S, ScalarEvolution &SE) { + SCEVFindUnsafe Search(SE); + visitAll(S, Search); + return !Search.IsUnsafe; +} +}