1 | //===--- NVPTX.h - Declare NVPTX target feature support ---------*- C++ -*-===// |
---|---|
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 file declares NVPTX TargetInfo objects. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #ifndef LLVM_CLANG_LIB_BASIC_TARGETS_NVPTX_H |
14 | #define LLVM_CLANG_LIB_BASIC_TARGETS_NVPTX_H |
15 | |
16 | #include "clang/Basic/Cuda.h" |
17 | #include "clang/Basic/TargetInfo.h" |
18 | #include "clang/Basic/TargetOptions.h" |
19 | #include "llvm/Support/Compiler.h" |
20 | #include "llvm/Support/NVPTXAddrSpace.h" |
21 | #include "llvm/TargetParser/Triple.h" |
22 | #include <optional> |
23 | |
24 | namespace clang { |
25 | namespace targets { |
26 | |
27 | static const unsigned NVPTXAddrSpaceMap[] = { |
28 | 0, // Default |
29 | 1, // opencl_global |
30 | 3, // opencl_local |
31 | 4, // opencl_constant |
32 | 0, // opencl_private |
33 | // FIXME: generic has to be added to the target |
34 | 0, // opencl_generic |
35 | 1, // opencl_global_device |
36 | 1, // opencl_global_host |
37 | 1, // cuda_device |
38 | 4, // cuda_constant |
39 | 3, // cuda_shared |
40 | 1, // sycl_global |
41 | 1, // sycl_global_device |
42 | 1, // sycl_global_host |
43 | 3, // sycl_local |
44 | 0, // sycl_private |
45 | 0, // ptr32_sptr |
46 | 0, // ptr32_uptr |
47 | 0, // ptr64 |
48 | 0, // hlsl_groupshared |
49 | 0, // hlsl_constant |
50 | 0, // hlsl_private |
51 | 0, // hlsl_device |
52 | 0, // hlsl_input |
53 | // Wasm address space values for this target are dummy values, |
54 | // as it is only enabled for Wasm targets. |
55 | 20, // wasm_funcref |
56 | }; |
57 | |
58 | /// The DWARF address class. Taken from |
59 | /// https://docs.nvidia.com/cuda/archive/10.0/ptx-writers-guide-to-interoperability/index.html#cuda-specific-dwarf |
60 | static const int NVPTXDWARFAddrSpaceMap[] = { |
61 | -1, // Default, opencl_private or opencl_generic - not defined |
62 | 5, // opencl_global |
63 | -1, |
64 | 8, // opencl_local or cuda_shared |
65 | 4, // opencl_constant or cuda_constant |
66 | }; |
67 | |
68 | class LLVM_LIBRARY_VISIBILITY NVPTXTargetInfo : public TargetInfo { |
69 | static const char *const GCCRegNames[]; |
70 | OffloadArch GPU; |
71 | uint32_t PTXVersion; |
72 | std::unique_ptr<TargetInfo> HostTarget; |
73 | |
74 | public: |
75 | NVPTXTargetInfo(const llvm::Triple &Triple, const TargetOptions &Opts, |
76 | unsigned TargetPointerWidth); |
77 | |
78 | void getTargetDefines(const LangOptions &Opts, |
79 | MacroBuilder &Builder) const override; |
80 | |
81 | llvm::SmallVector<Builtin::InfosShard> getTargetBuiltins() const override; |
82 | |
83 | bool useFP16ConversionIntrinsics() const override { return false; } |
84 | |
85 | bool |
86 | initFeatureMap(llvm::StringMap<bool> &Features, DiagnosticsEngine &Diags, |
87 | StringRef CPU, |
88 | const std::vector<std::string> &FeaturesVec) const override { |
89 | if (GPU != OffloadArch::UNUSED) |
90 | Features[OffloadArchToString(A: GPU)] = true; |
91 | Features["ptx"+ std::to_string(val: PTXVersion)] = true; |
92 | return TargetInfo::initFeatureMap(Features, Diags, CPU, FeatureVec: FeaturesVec); |
93 | } |
94 | |
95 | bool hasFeature(StringRef Feature) const override; |
96 | |
97 | virtual bool isAddressSpaceSupersetOf(LangAS A, LangAS B) const override { |
98 | // The generic address space AS(0) is a superset of all the other address |
99 | // spaces used by the backend target. |
100 | return A == B || |
101 | ((A == LangAS::Default || |
102 | (isTargetAddressSpace(AS: A) && |
103 | toTargetAddressSpace(AS: A) == |
104 | llvm::NVPTXAS::ADDRESS_SPACE_GENERIC)) && |
105 | isTargetAddressSpace(AS: B) && |
106 | toTargetAddressSpace(AS: B) >= llvm::NVPTXAS::ADDRESS_SPACE_GENERIC && |
107 | toTargetAddressSpace(AS: B) <= llvm::NVPTXAS::ADDRESS_SPACE_LOCAL && |
108 | toTargetAddressSpace(AS: B) != 2); |
109 | } |
110 | |
111 | ArrayRef<const char *> getGCCRegNames() const override; |
112 | |
113 | ArrayRef<TargetInfo::GCCRegAlias> getGCCRegAliases() const override { |
114 | // No aliases. |
115 | return {}; |
116 | } |
117 | |
118 | bool validateAsmConstraint(const char *&Name, |
119 | TargetInfo::ConstraintInfo &Info) const override { |
120 | switch (*Name) { |
121 | default: |
122 | return false; |
123 | case 'c': |
124 | case 'h': |
125 | case 'r': |
126 | case 'l': |
127 | case 'f': |
128 | case 'd': |
129 | case 'q': |
130 | Info.setAllowsRegister(); |
131 | return true; |
132 | } |
133 | } |
134 | |
135 | std::string_view getClobbers() const override { |
136 | // FIXME: Is this really right? |
137 | return ""; |
138 | } |
139 | |
140 | BuiltinVaListKind getBuiltinVaListKind() const override { |
141 | return TargetInfo::CharPtrBuiltinVaList; |
142 | } |
143 | |
144 | bool isValidCPUName(StringRef Name) const override { |
145 | return StringToOffloadArch(S: Name) != OffloadArch::UNKNOWN; |
146 | } |
147 | |
148 | void fillValidCPUList(SmallVectorImpl<StringRef> &Values) const override { |
149 | for (int i = static_cast<int>(OffloadArch::SM_20); |
150 | i < static_cast<int>(OffloadArch::Generic); ++i) |
151 | Values.emplace_back(Args: OffloadArchToString(A: static_cast<OffloadArch>(i))); |
152 | } |
153 | |
154 | bool setCPU(const std::string &Name) override { |
155 | GPU = StringToOffloadArch(S: Name); |
156 | return GPU != OffloadArch::UNKNOWN; |
157 | } |
158 | |
159 | void setSupportedOpenCLOpts() override { |
160 | auto &Opts = getSupportedOpenCLOpts(); |
161 | Opts["cl_clang_storage_class_specifiers"] = true; |
162 | Opts["__cl_clang_function_pointers"] = true; |
163 | Opts["__cl_clang_variadic_functions"] = true; |
164 | Opts["__cl_clang_non_portable_kernel_param_types"] = true; |
165 | Opts["__cl_clang_bitfields"] = true; |
166 | |
167 | Opts["cl_khr_fp64"] = true; |
168 | Opts["__opencl_c_fp64"] = true; |
169 | Opts["cl_khr_byte_addressable_store"] = true; |
170 | Opts["cl_khr_global_int32_base_atomics"] = true; |
171 | Opts["cl_khr_global_int32_extended_atomics"] = true; |
172 | Opts["cl_khr_local_int32_base_atomics"] = true; |
173 | Opts["cl_khr_local_int32_extended_atomics"] = true; |
174 | |
175 | Opts["__opencl_c_generic_address_space"] = true; |
176 | } |
177 | |
178 | const llvm::omp::GV &getGridValue() const override { |
179 | return llvm::omp::NVPTXGridValues; |
180 | } |
181 | |
182 | /// \returns If a target requires an address within a target specific address |
183 | /// space \p AddressSpace to be converted in order to be used, then return the |
184 | /// corresponding target specific DWARF address space. |
185 | /// |
186 | /// \returns Otherwise return std::nullopt and no conversion will be emitted |
187 | /// in the DWARF. |
188 | std::optional<unsigned> |
189 | getDWARFAddressSpace(unsigned AddressSpace) const override { |
190 | if (AddressSpace >= std::size(NVPTXDWARFAddrSpaceMap) || |
191 | NVPTXDWARFAddrSpaceMap[AddressSpace] < 0) |
192 | return std::nullopt; |
193 | return NVPTXDWARFAddrSpaceMap[AddressSpace]; |
194 | } |
195 | |
196 | CallingConvCheckResult checkCallingConvention(CallingConv CC) const override { |
197 | // CUDA compilations support all of the host's calling conventions. |
198 | // |
199 | // TODO: We should warn if you apply a non-default CC to anything other than |
200 | // a host function. |
201 | if (HostTarget) |
202 | return HostTarget->checkCallingConvention(CC); |
203 | return CCCR_Warning; |
204 | } |
205 | |
206 | bool hasBitIntType() const override { return true; } |
207 | bool hasBFloat16Type() const override { return true; } |
208 | |
209 | OffloadArch getGPU() const { return GPU; } |
210 | }; |
211 | } // namespace targets |
212 | } // namespace clang |
213 | #endif // LLVM_CLANG_LIB_BASIC_TARGETS_NVPTX_H |
214 |
Definitions
- NVPTXAddrSpaceMap
- NVPTXDWARFAddrSpaceMap
- NVPTXTargetInfo
- useFP16ConversionIntrinsics
- initFeatureMap
- isAddressSpaceSupersetOf
- getGCCRegAliases
- validateAsmConstraint
- getClobbers
- getBuiltinVaListKind
- isValidCPUName
- fillValidCPUList
- setCPU
- setSupportedOpenCLOpts
- getGridValue
- getDWARFAddressSpace
- checkCallingConvention
- hasBitIntType
- hasBFloat16Type
Learn to use CMake with our Intro Training
Find out more