| 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/ADT/Triple.h" |
| 20 |
#include "llvm/Support/Compiler.h" |
| 21 |
#include <optional> |
| 22 |
|
| 23 |
namespace clang { |
| 24 |
namespace targets { |
| 25 |
|
| 26 |
static 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 |
}; |
| 49 |
|
| 50 |
/// The DWARF address class. Taken from |
| 51 |
/// https://docs.nvidia.com/cuda/archive/10.0/ptx-writers-guide-to-interoperability/index.html#cuda-specific-dwarf |
| 52 |
static const int NVPTXDWARFAddrSpaceMap[] = { |
| 53 |
-1, // Default, opencl_private or opencl_generic - not defined |
| 54 |
5, // opencl_global |
| 55 |
-1, |
| 56 |
8, // opencl_local or cuda_shared |
| 57 |
4, // opencl_constant or cuda_constant |
| 58 |
}; |
| 59 |
|
| 60 |
class LLVM_LIBRARY_VISIBILITY NVPTXTargetInfo : public TargetInfo { |
| 61 |
static const char *const GCCRegNames[]; |
| 62 |
CudaArch GPU; |
| 63 |
uint32_t PTXVersion; |
| 64 |
std::unique_ptr<TargetInfo> HostTarget; |
| 65 |
|
| 66 |
public: |
| 67 |
NVPTXTargetInfo(const llvm::Triple &Triple, const TargetOptions &Opts, |
| 68 |
unsigned TargetPointerWidth); |
| 69 |
|
| 70 |
void getTargetDefines(const LangOptions &Opts, |
| 71 |
MacroBuilder &Builder) const override; |
| 72 |
|
| 73 |
ArrayRef<Builtin::Info> getTargetBuiltins() const override; |
| 74 |
|
| 75 |
bool |
| 76 |
initFeatureMap(llvm::StringMap<bool> &Features, DiagnosticsEngine &Diags, |
| 77 |
StringRef CPU, |
| 78 |
const std::vector<std::string> &FeaturesVec) const override { |
| 79 |
Features[CudaArchToString(GPU)] = true; |
| 80 |
Features["ptx" + std::to_string(PTXVersion)] = true; |
| 81 |
return TargetInfo::initFeatureMap(Features, Diags, CPU, FeaturesVec); |
| 82 |
} |
| 83 |
|
| 84 |
bool hasFeature(StringRef Feature) const override; |
| 85 |
|
| 86 |
ArrayRef<const char *> getGCCRegNames() const override; |
| 87 |
|
| 88 |
ArrayRef<TargetInfo::GCCRegAlias> getGCCRegAliases() const override { |
| 89 |
// No aliases. |
| 90 |
return std::nullopt; |
| 91 |
} |
| 92 |
|
| 93 |
bool validateAsmConstraint(const char *&Name, |
| 94 |
TargetInfo::ConstraintInfo &Info) const override { |
| 95 |
switch (*Name) { |
| 96 |
default: |
| 97 |
return false; |
| 98 |
case 'c': |
| 99 |
case 'h': |
| 100 |
case 'r': |
| 101 |
case 'l': |
| 102 |
case 'f': |
| 103 |
case 'd': |
| 104 |
Info.setAllowsRegister(); |
| 105 |
return true; |
| 106 |
} |
| 107 |
} |
| 108 |
|
| 109 |
const char *getClobbers() const override { |
| 110 |
// FIXME: Is this really right? |
| 111 |
return ""; |
| 112 |
} |
| 113 |
|
| 114 |
BuiltinVaListKind getBuiltinVaListKind() const override { |
| 115 |
// FIXME: implement |
| 116 |
return TargetInfo::CharPtrBuiltinVaList; |
| 117 |
} |
| 118 |
|
| 119 |
bool isValidCPUName(StringRef Name) const override { |
| 120 |
return StringToCudaArch(Name) != CudaArch::UNKNOWN; |
| 121 |
} |
| 122 |
|
| 123 |
void fillValidCPUList(SmallVectorImpl<StringRef> &Values) const override { |
| 124 |
for (int i = static_cast<int>(CudaArch::SM_20); |
| 125 |
i < static_cast<int>(CudaArch::Generic); ++i) |
| 126 |
Values.emplace_back(CudaArchToString(static_cast<CudaArch>(i))); |
| 127 |
} |
| 128 |
|
| 129 |
bool setCPU(const std::string &Name) override { |
| 130 |
GPU = StringToCudaArch(Name); |
| 131 |
return GPU != CudaArch::UNKNOWN; |
| 132 |
} |
| 133 |
|
| 134 |
void setSupportedOpenCLOpts() override { |
| 135 |
auto &Opts = getSupportedOpenCLOpts(); |
| 136 |
Opts["cl_clang_storage_class_specifiers"] = true; |
| 137 |
Opts["__cl_clang_function_pointers"] = true; |
| 138 |
Opts["__cl_clang_variadic_functions"] = true; |
| 139 |
Opts["__cl_clang_non_portable_kernel_param_types"] = true; |
| 140 |
Opts["__cl_clang_bitfields"] = true; |
| 141 |
|
| 142 |
Opts["cl_khr_fp64"] = true; |
| 143 |
Opts["__opencl_c_fp64"] = true; |
| 144 |
Opts["cl_khr_byte_addressable_store"] = true; |
| 145 |
Opts["cl_khr_global_int32_base_atomics"] = true; |
| 146 |
Opts["cl_khr_global_int32_extended_atomics"] = true; |
| 147 |
Opts["cl_khr_local_int32_base_atomics"] = true; |
| 148 |
Opts["cl_khr_local_int32_extended_atomics"] = true; |
| 149 |
} |
| 150 |
|
| 151 |
const llvm::omp::GV &getGridValue() const override { |
| 152 |
return llvm::omp::NVPTXGridValues; |
| 153 |
} |
| 154 |
|
| 155 |
/// \returns If a target requires an address within a target specific address |
| 156 |
/// space \p AddressSpace to be converted in order to be used, then return the |
| 157 |
/// corresponding target specific DWARF address space. |
| 158 |
/// |
| 159 |
/// \returns Otherwise return std::nullopt and no conversion will be emitted |
| 160 |
/// in the DWARF. |
| 161 |
std::optional<unsigned> |
| 162 |
getDWARFAddressSpace(unsigned AddressSpace) const override { |
| 163 |
if (AddressSpace >= std::size(NVPTXDWARFAddrSpaceMap) || |
| 164 |
NVPTXDWARFAddrSpaceMap[AddressSpace] < 0) |
| 165 |
return std::nullopt; |
| 166 |
return NVPTXDWARFAddrSpaceMap[AddressSpace]; |
| 167 |
} |
| 168 |
|
| 169 |
CallingConvCheckResult checkCallingConvention(CallingConv CC) const override { |
| 170 |
// CUDA compilations support all of the host's calling conventions. |
| 171 |
// |
| 172 |
// TODO: We should warn if you apply a non-default CC to anything other than |
| 173 |
// a host function. |
| 174 |
if (HostTarget) |
| 175 |
return HostTarget->checkCallingConvention(CC); |
| 176 |
return CCCR_Warning; |
| 177 |
} |
| 178 |
|
| 179 |
bool hasBitIntType() const override { return true; } |
| 180 |
bool hasBFloat16Type() const override { return true; } |
| 181 |
const char *getBFloat16Mangling() const override { return "u6__bf16"; }; |
| 182 |
}; |
| 183 |
} // namespace targets |
| 184 |
} // namespace clang |
| 185 |
#endif // LLVM_CLANG_LIB_BASIC_TARGETS_NVPTX_H |