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