1 | //===----- CGOpenCLRuntime.cpp - Interface to OpenCL Runtimes -------------===// |
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 | // This provides an abstract class for OpenCL code generation. Concrete |
10 | // subclasses of this implement code generation for specific OpenCL |
11 | // runtime libraries. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "CGOpenCLRuntime.h" |
16 | #include "CodeGenFunction.h" |
17 | #include "TargetInfo.h" |
18 | #include "clang/CodeGen/ConstantInitBuilder.h" |
19 | #include "llvm/IR/DerivedTypes.h" |
20 | #include "llvm/IR/GlobalValue.h" |
21 | #include <assert.h> |
22 | |
23 | using namespace clang; |
24 | using namespace CodeGen; |
25 | |
26 | CGOpenCLRuntime::~CGOpenCLRuntime() {} |
27 | |
28 | void CGOpenCLRuntime::EmitWorkGroupLocalVarDecl(CodeGenFunction &CGF, |
29 | const VarDecl &D) { |
30 | return CGF.EmitStaticVarDecl(D, Linkage: llvm::GlobalValue::InternalLinkage); |
31 | } |
32 | |
33 | llvm::Type *CGOpenCLRuntime::convertOpenCLSpecificType(const Type *T) { |
34 | assert(T->isOpenCLSpecificType() && "Not an OpenCL specific type!" ); |
35 | |
36 | // Check if the target has a specific translation for this type first. |
37 | if (llvm::Type *TransTy = CGM.getTargetCodeGenInfo().getOpenCLType(CGM, T)) |
38 | return TransTy; |
39 | |
40 | if (T->isSamplerT()) |
41 | return getSamplerType(T); |
42 | |
43 | return getPointerType(T); |
44 | } |
45 | |
46 | llvm::PointerType *CGOpenCLRuntime::getPointerType(const Type *T) { |
47 | uint32_t AddrSpc = CGM.getContext().getTargetAddressSpace( |
48 | AS: CGM.getContext().getOpenCLTypeAddrSpace(T)); |
49 | return llvm::PointerType::get(C&: CGM.getLLVMContext(), AddressSpace: AddrSpc); |
50 | } |
51 | |
52 | llvm::Type *CGOpenCLRuntime::getPipeType(const PipeType *T) { |
53 | if (llvm::Type *PipeTy = CGM.getTargetCodeGenInfo().getOpenCLType(CGM, T)) |
54 | return PipeTy; |
55 | |
56 | if (T->isReadOnly()) |
57 | return getPipeType(T, Name: "opencl.pipe_ro_t" , PipeTy&: PipeROTy); |
58 | else |
59 | return getPipeType(T, Name: "opencl.pipe_wo_t" , PipeTy&: PipeWOTy); |
60 | } |
61 | |
62 | llvm::Type *CGOpenCLRuntime::getPipeType(const PipeType *T, StringRef Name, |
63 | llvm::Type *&PipeTy) { |
64 | if (!PipeTy) |
65 | PipeTy = getPointerType(T); |
66 | return PipeTy; |
67 | } |
68 | |
69 | llvm::Type *CGOpenCLRuntime::getSamplerType(const Type *T) { |
70 | if (SamplerTy) |
71 | return SamplerTy; |
72 | |
73 | if (llvm::Type *TransTy = CGM.getTargetCodeGenInfo().getOpenCLType( |
74 | CGM, T: CGM.getContext().OCLSamplerTy.getTypePtr())) |
75 | SamplerTy = TransTy; |
76 | else |
77 | SamplerTy = getPointerType(T); |
78 | return SamplerTy; |
79 | } |
80 | |
81 | llvm::Value *CGOpenCLRuntime::getPipeElemSize(const Expr *PipeArg) { |
82 | const PipeType *PipeTy = PipeArg->getType()->castAs<PipeType>(); |
83 | // The type of the last (implicit) argument to be passed. |
84 | llvm::Type *Int32Ty = llvm::IntegerType::getInt32Ty(C&: CGM.getLLVMContext()); |
85 | unsigned TypeSize = CGM.getContext() |
86 | .getTypeSizeInChars(T: PipeTy->getElementType()) |
87 | .getQuantity(); |
88 | return llvm::ConstantInt::get(Ty: Int32Ty, V: TypeSize, IsSigned: false); |
89 | } |
90 | |
91 | llvm::Value *CGOpenCLRuntime::getPipeElemAlign(const Expr *PipeArg) { |
92 | const PipeType *PipeTy = PipeArg->getType()->castAs<PipeType>(); |
93 | // The type of the last (implicit) argument to be passed. |
94 | llvm::Type *Int32Ty = llvm::IntegerType::getInt32Ty(C&: CGM.getLLVMContext()); |
95 | unsigned TypeSize = CGM.getContext() |
96 | .getTypeAlignInChars(T: PipeTy->getElementType()) |
97 | .getQuantity(); |
98 | return llvm::ConstantInt::get(Ty: Int32Ty, V: TypeSize, IsSigned: false); |
99 | } |
100 | |
101 | llvm::PointerType *CGOpenCLRuntime::getGenericVoidPointerType() { |
102 | assert(CGM.getLangOpts().OpenCL); |
103 | return llvm::PointerType::get( |
104 | C&: CGM.getLLVMContext(), |
105 | AddressSpace: CGM.getContext().getTargetAddressSpace(AS: LangAS::opencl_generic)); |
106 | } |
107 | |
108 | // Get the block literal from an expression derived from the block expression. |
109 | // OpenCL v2.0 s6.12.5: |
110 | // Block variable declarations are implicitly qualified with const. Therefore |
111 | // all block variables must be initialized at declaration time and may not be |
112 | // reassigned. |
113 | static const BlockExpr *getBlockExpr(const Expr *E) { |
114 | const Expr *Prev = nullptr; // to make sure we do not stuck in infinite loop. |
115 | while(!isa<BlockExpr>(Val: E) && E != Prev) { |
116 | Prev = E; |
117 | E = E->IgnoreCasts(); |
118 | if (auto DR = dyn_cast<DeclRefExpr>(Val: E)) { |
119 | E = cast<VarDecl>(Val: DR->getDecl())->getInit(); |
120 | } |
121 | } |
122 | return cast<BlockExpr>(Val: E); |
123 | } |
124 | |
125 | /// Record emitted llvm invoke function and llvm block literal for the |
126 | /// corresponding block expression. |
127 | void CGOpenCLRuntime::recordBlockInfo(const BlockExpr *E, |
128 | llvm::Function *InvokeF, |
129 | llvm::Value *Block, llvm::Type *BlockTy) { |
130 | assert(!EnqueuedBlockMap.contains(E) && "Block expression emitted twice" ); |
131 | assert(isa<llvm::Function>(InvokeF) && "Invalid invoke function" ); |
132 | assert(Block->getType()->isPointerTy() && "Invalid block literal type" ); |
133 | EnqueuedBlockMap[E].InvokeFunc = InvokeF; |
134 | EnqueuedBlockMap[E].BlockArg = Block; |
135 | EnqueuedBlockMap[E].BlockTy = BlockTy; |
136 | EnqueuedBlockMap[E].KernelHandle = nullptr; |
137 | } |
138 | |
139 | llvm::Function *CGOpenCLRuntime::getInvokeFunction(const Expr *E) { |
140 | return EnqueuedBlockMap[getBlockExpr(E)].InvokeFunc; |
141 | } |
142 | |
143 | CGOpenCLRuntime::EnqueuedBlockInfo |
144 | CGOpenCLRuntime::emitOpenCLEnqueuedBlock(CodeGenFunction &CGF, const Expr *E) { |
145 | CGF.EmitScalarExpr(E); |
146 | |
147 | // The block literal may be assigned to a const variable. Chasing down |
148 | // to get the block literal. |
149 | const BlockExpr *Block = getBlockExpr(E); |
150 | |
151 | assert(EnqueuedBlockMap.contains(Block) && "Block expression not emitted" ); |
152 | |
153 | // Do not emit the block wrapper again if it has been emitted. |
154 | if (EnqueuedBlockMap[Block].KernelHandle) { |
155 | return EnqueuedBlockMap[Block]; |
156 | } |
157 | |
158 | auto *F = CGF.getTargetHooks().createEnqueuedBlockKernel( |
159 | CGF, BlockInvokeFunc: EnqueuedBlockMap[Block].InvokeFunc, BlockTy: EnqueuedBlockMap[Block].BlockTy); |
160 | |
161 | // The common part of the post-processing of the kernel goes here. |
162 | EnqueuedBlockMap[Block].KernelHandle = F; |
163 | return EnqueuedBlockMap[Block]; |
164 | } |
165 | |