1*7330f729Sjoerg //===--- PatternInit.cpp - Pattern Initialization -------------------------===// 2*7330f729Sjoerg // 3*7330f729Sjoerg // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4*7330f729Sjoerg // See https://llvm.org/LICENSE.txt for license information. 5*7330f729Sjoerg // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6*7330f729Sjoerg // 7*7330f729Sjoerg //===----------------------------------------------------------------------===// 8*7330f729Sjoerg 9*7330f729Sjoerg #include "PatternInit.h" 10*7330f729Sjoerg #include "CodeGenModule.h" 11*7330f729Sjoerg #include "llvm/IR/Constant.h" 12*7330f729Sjoerg #include "llvm/IR/Type.h" 13*7330f729Sjoerg 14*7330f729Sjoerg llvm::Constant *clang::CodeGen::initializationPatternFor(CodeGenModule &CGM, 15*7330f729Sjoerg llvm::Type *Ty) { 16*7330f729Sjoerg // The following value is a guaranteed unmappable pointer value and has a 17*7330f729Sjoerg // repeated byte-pattern which makes it easier to synthesize. We use it for 18*7330f729Sjoerg // pointers as well as integers so that aggregates are likely to be 19*7330f729Sjoerg // initialized with this repeated value. 20*7330f729Sjoerg // For 32-bit platforms it's a bit trickier because, across systems, only the 21*7330f729Sjoerg // zero page can reasonably be expected to be unmapped. We use max 0xFFFFFFFF 22*7330f729Sjoerg // assuming that memory access will overlap into zero page. 23*7330f729Sjoerg const uint64_t IntValue = 24*7330f729Sjoerg CGM.getContext().getTargetInfo().getMaxPointerWidth() < 64 25*7330f729Sjoerg ? 0xFFFFFFFFFFFFFFFFull 26*7330f729Sjoerg : 0xAAAAAAAAAAAAAAAAull; 27*7330f729Sjoerg // Floating-point values are initialized as NaNs because they propagate. Using 28*7330f729Sjoerg // a repeated byte pattern means that it will be easier to initialize 29*7330f729Sjoerg // all-floating-point aggregates and arrays with memset. Further, aggregates 30*7330f729Sjoerg // which mix integral and a few floats might also initialize with memset 31*7330f729Sjoerg // followed by a handful of stores for the floats. Using fairly unique NaNs 32*7330f729Sjoerg // also means they'll be easier to distinguish in a crash. 33*7330f729Sjoerg constexpr bool NegativeNaN = true; 34*7330f729Sjoerg constexpr uint64_t NaNPayload = 0xFFFFFFFFFFFFFFFFull; 35*7330f729Sjoerg if (Ty->isIntOrIntVectorTy()) { 36*7330f729Sjoerg unsigned BitWidth = cast<llvm::IntegerType>( 37*7330f729Sjoerg Ty->isVectorTy() ? Ty->getVectorElementType() : Ty) 38*7330f729Sjoerg ->getBitWidth(); 39*7330f729Sjoerg if (BitWidth <= 64) 40*7330f729Sjoerg return llvm::ConstantInt::get(Ty, IntValue); 41*7330f729Sjoerg return llvm::ConstantInt::get( 42*7330f729Sjoerg Ty, llvm::APInt::getSplat(BitWidth, llvm::APInt(64, IntValue))); 43*7330f729Sjoerg } 44*7330f729Sjoerg if (Ty->isPtrOrPtrVectorTy()) { 45*7330f729Sjoerg auto *PtrTy = cast<llvm::PointerType>( 46*7330f729Sjoerg Ty->isVectorTy() ? Ty->getVectorElementType() : Ty); 47*7330f729Sjoerg unsigned PtrWidth = CGM.getContext().getTargetInfo().getPointerWidth( 48*7330f729Sjoerg PtrTy->getAddressSpace()); 49*7330f729Sjoerg if (PtrWidth > 64) 50*7330f729Sjoerg llvm_unreachable("pattern initialization of unsupported pointer width"); 51*7330f729Sjoerg llvm::Type *IntTy = llvm::IntegerType::get(CGM.getLLVMContext(), PtrWidth); 52*7330f729Sjoerg auto *Int = llvm::ConstantInt::get(IntTy, IntValue); 53*7330f729Sjoerg return llvm::ConstantExpr::getIntToPtr(Int, PtrTy); 54*7330f729Sjoerg } 55*7330f729Sjoerg if (Ty->isFPOrFPVectorTy()) { 56*7330f729Sjoerg unsigned BitWidth = llvm::APFloat::semanticsSizeInBits( 57*7330f729Sjoerg (Ty->isVectorTy() ? Ty->getVectorElementType() : Ty) 58*7330f729Sjoerg ->getFltSemantics()); 59*7330f729Sjoerg llvm::APInt Payload(64, NaNPayload); 60*7330f729Sjoerg if (BitWidth >= 64) 61*7330f729Sjoerg Payload = llvm::APInt::getSplat(BitWidth, Payload); 62*7330f729Sjoerg return llvm::ConstantFP::getQNaN(Ty, NegativeNaN, &Payload); 63*7330f729Sjoerg } 64*7330f729Sjoerg if (Ty->isArrayTy()) { 65*7330f729Sjoerg // Note: this doesn't touch tail padding (at the end of an object, before 66*7330f729Sjoerg // the next array object). It is instead handled by replaceUndef. 67*7330f729Sjoerg auto *ArrTy = cast<llvm::ArrayType>(Ty); 68*7330f729Sjoerg llvm::SmallVector<llvm::Constant *, 8> Element( 69*7330f729Sjoerg ArrTy->getNumElements(), 70*7330f729Sjoerg initializationPatternFor(CGM, ArrTy->getElementType())); 71*7330f729Sjoerg return llvm::ConstantArray::get(ArrTy, Element); 72*7330f729Sjoerg } 73*7330f729Sjoerg 74*7330f729Sjoerg // Note: this doesn't touch struct padding. It will initialize as much union 75*7330f729Sjoerg // padding as is required for the largest type in the union. Padding is 76*7330f729Sjoerg // instead handled by replaceUndef. Stores to structs with volatile members 77*7330f729Sjoerg // don't have a volatile qualifier when initialized according to C++. This is 78*7330f729Sjoerg // fine because stack-based volatiles don't really have volatile semantics 79*7330f729Sjoerg // anyways, and the initialization shouldn't be observable. 80*7330f729Sjoerg auto *StructTy = cast<llvm::StructType>(Ty); 81*7330f729Sjoerg llvm::SmallVector<llvm::Constant *, 8> Struct(StructTy->getNumElements()); 82*7330f729Sjoerg for (unsigned El = 0; El != Struct.size(); ++El) 83*7330f729Sjoerg Struct[El] = initializationPatternFor(CGM, StructTy->getElementType(El)); 84*7330f729Sjoerg return llvm::ConstantStruct::get(StructTy, Struct); 85*7330f729Sjoerg } 86