X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;ds=sidebyside;f=lib%2FIR%2FConstantFold.cpp;h=d3caf04489d47dbb04f7c8402d4ebc0b6962ac7c;hb=1b44f585f970eda9098e9c386d956ce2f962d6c7;hp=4cb22bc97a7d8c83b517ef57576aec909fcac250;hpb=8eef439528e2113443fd18a3a2c39cf885bf7aec;p=oota-llvm.git diff --git a/lib/IR/ConstantFold.cpp b/lib/IR/ConstantFold.cpp index 4cb22bc97a7..d3caf04489d 100644 --- a/lib/IR/ConstantFold.cpp +++ b/lib/IR/ConstantFold.cpp @@ -27,12 +27,14 @@ #include "llvm/IR/GlobalVariable.h" #include "llvm/IR/Instructions.h" #include "llvm/IR/Operator.h" +#include "llvm/IR/PatternMatch.h" #include "llvm/Support/Compiler.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/ManagedStatic.h" #include "llvm/Support/MathExtras.h" #include using namespace llvm; +using namespace llvm::PatternMatch; //===----------------------------------------------------------------------===// // ConstantFold*Instruction Implementations @@ -130,7 +132,8 @@ static Constant *FoldBitCast(Constant *V, Type *DestTy) { if (ElTy == DPTy->getElementType()) // This GEP is inbounds because all indices are zero. - return ConstantExpr::getInBoundsGetElementPtr(V, IdxList); + return ConstantExpr::getInBoundsGetElementPtr(PTy->getElementType(), + V, IdxList); } // Handle casts from one vector constant to another. We know that the src @@ -167,7 +170,8 @@ static Constant *FoldBitCast(Constant *V, Type *DestTy) { // be the same. Consequently, we just fold to V. return V; - if (DestTy->isFloatingPointTy()) + // See note below regarding the PPC_FP128 restriction. + if (DestTy->isFloatingPointTy() && !DestTy->isPPC_FP128Ty()) return ConstantFP::get(DestTy->getContext(), APFloat(DestTy->getFltSemantics(), CI->getValue())); @@ -177,9 +181,19 @@ static Constant *FoldBitCast(Constant *V, Type *DestTy) { } // Handle ConstantFP input: FP -> Integral. - if (ConstantFP *FP = dyn_cast(V)) + if (ConstantFP *FP = dyn_cast(V)) { + // PPC_FP128 is really the sum of two consecutive doubles, where the first + // double is always stored first in memory, regardless of the target + // endianness. The memory layout of i128, however, depends on the target + // endianness, and so we can't fold this without target endianness + // information. This should instead be handled by + // Analysis/ConstantFolding.cpp + if (FP->getType()->isPPC_FP128Ty()) + return nullptr; + return ConstantInt::get(FP->getContext(), FP->getValueAPF().bitcastToAPInt()); + } return nullptr; } @@ -913,49 +927,70 @@ Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, return C1; return Constant::getNullValue(C1->getType()); // undef & X -> 0 case Instruction::Mul: { - ConstantInt *CI; - // X * undef -> undef if X is odd or undef - if (((CI = dyn_cast(C1)) && CI->getValue()[0]) || - ((CI = dyn_cast(C2)) && CI->getValue()[0]) || - (isa(C1) && isa(C2))) - return UndefValue::get(C1->getType()); + // undef * undef -> undef + if (isa(C1) && isa(C2)) + return C1; + const APInt *CV; + // X * undef -> undef if X is odd + if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV))) + if ((*CV)[0]) + return UndefValue::get(C1->getType()); // X * undef -> 0 otherwise return Constant::getNullValue(C1->getType()); } - case Instruction::UDiv: case Instruction::SDiv: + case Instruction::UDiv: + // X / undef -> undef + if (match(C1, m_Zero())) + return C2; + // undef / 0 -> undef // undef / 1 -> undef - if (Opcode == Instruction::UDiv || Opcode == Instruction::SDiv) - if (ConstantInt *CI2 = dyn_cast(C2)) - if (CI2->isOne()) - return C1; - // FALL THROUGH + if (match(C2, m_Zero()) || match(C2, m_One())) + return C1; + // undef / X -> 0 otherwise + return Constant::getNullValue(C1->getType()); case Instruction::URem: case Instruction::SRem: - if (!isa(C2)) // undef / X -> 0 - return Constant::getNullValue(C1->getType()); - return C2; // X / undef -> undef + // X % undef -> undef + if (match(C2, m_Undef())) + return C2; + // undef % 0 -> undef + if (match(C2, m_Zero())) + return C1; + // undef % X -> 0 otherwise + return Constant::getNullValue(C1->getType()); case Instruction::Or: // X | undef -> -1 if (isa(C1) && isa(C2)) // undef | undef -> undef return C1; return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0 case Instruction::LShr: - if (isa(C2) && isa(C1)) - return C1; // undef lshr undef -> undef - return Constant::getNullValue(C1->getType()); // X lshr undef -> 0 - // undef lshr X -> 0 + // X >>l undef -> undef + if (isa(C2)) + return C2; + // undef >>l 0 -> undef + if (match(C2, m_Zero())) + return C1; + // undef >>l X -> 0 + return Constant::getNullValue(C1->getType()); case Instruction::AShr: - if (!isa(C2)) // undef ashr X --> all ones - return Constant::getAllOnesValue(C1->getType()); - else if (isa(C1)) - return C1; // undef ashr undef -> undef - else - return C1; // X ashr undef --> X + // X >>a undef -> undef + if (isa(C2)) + return C2; + // undef >>a 0 -> undef + if (match(C2, m_Zero())) + return C1; + // TODO: undef >>a X -> undef if the shift is exact + // undef >>a X -> 0 + return Constant::getNullValue(C1->getType()); case Instruction::Shl: - if (isa(C2) && isa(C1)) - return C1; // undef shl undef -> undef - // undef << X -> 0 or X << undef -> 0 + // X << undef -> undef + if (isa(C2)) + return C2; + // undef << 0 -> undef + if (match(C2, m_Zero())) + return C1; + // undef << X -> 0 return Constant::getNullValue(C1->getType()); } } @@ -1097,27 +1132,18 @@ Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, return ConstantInt::get(CI1->getContext(), C1V | C2V); case Instruction::Xor: return ConstantInt::get(CI1->getContext(), C1V ^ C2V); - case Instruction::Shl: { - uint32_t shiftAmt = C2V.getZExtValue(); - if (shiftAmt < C1V.getBitWidth()) - return ConstantInt::get(CI1->getContext(), C1V.shl(shiftAmt)); - else - return UndefValue::get(C1->getType()); // too big shift is undef - } - case Instruction::LShr: { - uint32_t shiftAmt = C2V.getZExtValue(); - if (shiftAmt < C1V.getBitWidth()) - return ConstantInt::get(CI1->getContext(), C1V.lshr(shiftAmt)); - else - return UndefValue::get(C1->getType()); // too big shift is undef - } - case Instruction::AShr: { - uint32_t shiftAmt = C2V.getZExtValue(); - if (shiftAmt < C1V.getBitWidth()) - return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt)); - else - return UndefValue::get(C1->getType()); // too big shift is undef - } + case Instruction::Shl: + if (C2V.ult(C1V.getBitWidth())) + return ConstantInt::get(CI1->getContext(), C1V.shl(C2V)); + return UndefValue::get(C1->getType()); // too big shift is undef + case Instruction::LShr: + if (C2V.ult(C1V.getBitWidth())) + return ConstantInt::get(CI1->getContext(), C1V.lshr(C2V)); + return UndefValue::get(C1->getType()); // too big shift is undef + case Instruction::AShr: + if (C2V.ult(C1V.getBitWidth())) + return ConstantInt::get(CI1->getContext(), C1V.ashr(C2V)); + return UndefValue::get(C1->getType()); // too big shift is undef } } @@ -1259,15 +1285,17 @@ static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) { if (!isa(C1) || !isa(C2)) return -2; // don't know! - // Ok, we have two differing integer indices. Sign extend them to be the same - // type. Long is always big enough, so we use it. - if (!C1->getType()->isIntegerTy(64)) - C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext())); + // We cannot compare the indices if they don't fit in an int64_t. + if (cast(C1)->getValue().getActiveBits() > 64 || + cast(C2)->getValue().getActiveBits() > 64) + return -2; // don't know! - if (!C2->getType()->isIntegerTy(64)) - C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext())); + // Ok, we have two differing integer indices. Sign extend them to be the same + // type. + int64_t C1Val = cast(C1)->getSExtValue(); + int64_t C2Val = cast(C2)->getSExtValue(); - if (C1 == C2) return 0; // They are equal + if (C1Val == C2Val) return 0; // They are equal // If the type being indexed over is really just a zero sized type, there is // no pointer difference being made here. @@ -1276,8 +1304,7 @@ static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) { // If they are really different, now that they are the same type, then we // found a difference! - if (cast(C1)->getSExtValue() < - cast(C2)->getSExtValue()) + if (C1Val < C2Val) return -1; else return 1; @@ -1303,7 +1330,7 @@ static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) { if (!isa(V1)) { if (!isa(V2)) { - // We distilled thisUse the standard constant folder for a few cases + // Simple case, use the standard constant folder. ConstantInt *R = nullptr; R = dyn_cast( ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2)); @@ -1641,15 +1668,22 @@ Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, // Handle some degenerate cases first if (isa(C1) || isa(C2)) { + CmpInst::Predicate Predicate = CmpInst::Predicate(pred); + bool isIntegerPredicate = ICmpInst::isIntPredicate(Predicate); // For EQ and NE, we can always pick a value for the undef to make the // predicate pass or fail, so we can return undef. - // Also, if both operands are undef, we can return undef. - if (ICmpInst::isEquality(ICmpInst::Predicate(pred)) || - (isa(C1) && isa(C2))) + // Also, if both operands are undef, we can return undef for int comparison. + if (ICmpInst::isEquality(Predicate) || (isIntegerPredicate && C1 == C2)) return UndefValue::get(ResultTy); - // Otherwise, pick the same value as the non-undef operand, and fold - // it to true or false. - return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred)); + + // Otherwise, for integer compare, pick the same value as the non-undef + // operand, and fold it to true or false. + if (isIntegerPredicate) + return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred)); + + // Choosing NaN for the undef will always make unordered comparison succeed + // and ordered comparison fails. + return ConstantInt::get(ResultTy, CmpInst::isUnordered(Predicate)); } // icmp eq/ne(null,GV) -> false/true @@ -1765,7 +1799,10 @@ Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, return ConstantVector::get(ResElts); } - if (C1->getType()->isFloatingPointTy()) { + if (C1->getType()->isFloatingPointTy() && + // Only call evaluateFCmpRelation if we have a constant expr to avoid + // infinite recursive loop + (isa(C1) || isa(C2))) { int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. switch (evaluateFCmpRelation(C1, C2)) { default: llvm_unreachable("Unknown relation!"); @@ -1995,7 +2032,8 @@ static Constant *ConstantFoldGetElementPtrImpl(Constant *C, if (isa(C)) { PointerType *Ptr = cast(C->getType()); - Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs); + Type *Ty = GetElementPtrInst::getIndexedType( + cast(Ptr->getScalarType())->getElementType(), Idxs); assert(Ty && "Invalid indices for GEP!"); return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace())); } @@ -2009,7 +2047,8 @@ static Constant *ConstantFoldGetElementPtrImpl(Constant *C, } if (isNull) { PointerType *Ptr = cast(C->getType()); - Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs); + Type *Ty = GetElementPtrInst::getIndexedType( + cast(Ptr->getScalarType())->getElementType(), Idxs); assert(Ty && "Invalid indices for GEP!"); return ConstantPointerNull::get(PointerType::get(Ty, Ptr->getAddressSpace())); @@ -2055,8 +2094,7 @@ static Constant *ConstantFoldGetElementPtrImpl(Constant *C, if (PerformFold) { SmallVector NewIndices; NewIndices.reserve(Idxs.size() + CE->getNumOperands()); - for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i) - NewIndices.push_back(CE->getOperand(i)); + NewIndices.append(CE->op_begin() + 1, CE->op_end() - 1); // Add the last index of the source with the first index of the new GEP. // Make sure to handle the case when they are actually different types. @@ -2065,9 +2103,15 @@ static Constant *ConstantFoldGetElementPtrImpl(Constant *C, if (!Idx0->isNullValue()) { Type *IdxTy = Combined->getType(); if (IdxTy != Idx0->getType()) { - Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext()); - Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty); - Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty); + unsigned CommonExtendedWidth = + std::max(IdxTy->getIntegerBitWidth(), + Idx0->getType()->getIntegerBitWidth()); + CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); + + Type *CommonTy = + Type::getIntNTy(IdxTy->getContext(), CommonExtendedWidth); + Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, CommonTy); + Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, CommonTy); Combined = ConstantExpr::get(Instruction::Add, C1, C2); } else { Combined = @@ -2077,10 +2121,9 @@ static Constant *ConstantFoldGetElementPtrImpl(Constant *C, NewIndices.push_back(Combined); NewIndices.append(Idxs.begin() + 1, Idxs.end()); - return - ConstantExpr::getGetElementPtr(CE->getOperand(0), NewIndices, - inBounds && - cast(CE)->isInBounds()); + return ConstantExpr::getGetElementPtr( + cast(CE)->getSourceElementType(), CE->getOperand(0), + NewIndices, inBounds && cast(CE)->isInBounds()); } } @@ -2105,8 +2148,8 @@ static Constant *ConstantFoldGetElementPtrImpl(Constant *C, if (SrcArrayTy && DstArrayTy && SrcArrayTy->getElementType() == DstArrayTy->getElementType() && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace()) - return ConstantExpr::getGetElementPtr((Constant*)CE->getOperand(0), - Idxs, inBounds); + return ConstantExpr::getGetElementPtr( + SrcArrayTy, (Constant *)CE->getOperand(0), Idxs, inBounds); } } } @@ -2140,14 +2183,20 @@ static Constant *ConstantFoldGetElementPtrImpl(Constant *C, Constant *PrevIdx = cast(Idxs[i-1]); Constant *Div = ConstantExpr::getSDiv(CI, Factor); + unsigned CommonExtendedWidth = + std::max(PrevIdx->getType()->getIntegerBitWidth(), + Div->getType()->getIntegerBitWidth()); + CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); + // Before adding, extend both operands to i64 to avoid // overflow trouble. - if (!PrevIdx->getType()->isIntegerTy(64)) - PrevIdx = ConstantExpr::getSExt(PrevIdx, - Type::getInt64Ty(Div->getContext())); - if (!Div->getType()->isIntegerTy(64)) - Div = ConstantExpr::getSExt(Div, - Type::getInt64Ty(Div->getContext())); + if (!PrevIdx->getType()->isIntegerTy(CommonExtendedWidth)) + PrevIdx = ConstantExpr::getSExt( + PrevIdx, + Type::getIntNTy(Div->getContext(), CommonExtendedWidth)); + if (!Div->getType()->isIntegerTy(CommonExtendedWidth)) + Div = ConstantExpr::getSExt( + Div, Type::getIntNTy(Div->getContext(), CommonExtendedWidth)); NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div); } else { @@ -2166,7 +2215,7 @@ static Constant *ConstantFoldGetElementPtrImpl(Constant *C, if (!NewIdxs.empty()) { for (unsigned i = 0, e = Idxs.size(); i != e; ++i) if (!NewIdxs[i]) NewIdxs[i] = cast(Idxs[i]); - return ConstantExpr::getGetElementPtr(C, NewIdxs, inBounds); + return ConstantExpr::getGetElementPtr(nullptr, C, NewIdxs, inBounds); } // If all indices are known integers and normalized, we can do a simple @@ -2174,7 +2223,7 @@ static Constant *ConstantFoldGetElementPtrImpl(Constant *C, if (!Unknown && !inBounds) if (auto *GV = dyn_cast(C)) if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs)) - return ConstantExpr::getInBoundsGetElementPtr(C, Idxs); + return ConstantExpr::getInBoundsGetElementPtr(nullptr, C, Idxs); return nullptr; }