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

source code of clang/lib/Basic/Targets/NVPTX.h