| 1 | //===--- NVPTX.cpp - Implement NVPTX target feature support ---------------===// |
| 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 implements NVPTX TargetInfo objects. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "NVPTX.h" |
| 14 | #include "clang/Basic/Builtins.h" |
| 15 | #include "clang/Basic/MacroBuilder.h" |
| 16 | #include "clang/Basic/TargetBuiltins.h" |
| 17 | #include "llvm/ADT/StringSwitch.h" |
| 18 | |
| 19 | using namespace clang; |
| 20 | using namespace clang::targets; |
| 21 | |
| 22 | static constexpr int NumBuiltins = |
| 23 | clang::NVPTX::LastTSBuiltin - Builtin::FirstTSBuiltin; |
| 24 | |
| 25 | #define GET_BUILTIN_STR_TABLE |
| 26 | #include "clang/Basic/BuiltinsNVPTX.inc" |
| 27 | #undef GET_BUILTIN_STR_TABLE |
| 28 | |
| 29 | static constexpr Builtin::Info BuiltinInfos[] = { |
| 30 | #define GET_BUILTIN_INFOS |
| 31 | #include "clang/Basic/BuiltinsNVPTX.inc" |
| 32 | #undef GET_BUILTIN_INFOS |
| 33 | }; |
| 34 | static_assert(std::size(BuiltinInfos) == NumBuiltins); |
| 35 | |
| 36 | const char *const NVPTXTargetInfo::GCCRegNames[] = {"r0" }; |
| 37 | |
| 38 | NVPTXTargetInfo::NVPTXTargetInfo(const llvm::Triple &Triple, |
| 39 | const TargetOptions &Opts, |
| 40 | unsigned TargetPointerWidth) |
| 41 | : TargetInfo(Triple) { |
| 42 | assert((TargetPointerWidth == 32 || TargetPointerWidth == 64) && |
| 43 | "NVPTX only supports 32- and 64-bit modes." ); |
| 44 | |
| 45 | PTXVersion = 32; |
| 46 | for (const StringRef Feature : Opts.FeaturesAsWritten) { |
| 47 | int PTXV; |
| 48 | if (!Feature.starts_with(Prefix: "+ptx" ) || |
| 49 | Feature.drop_front(N: 4).getAsInteger(Radix: 10, Result&: PTXV)) |
| 50 | continue; |
| 51 | PTXVersion = PTXV; // TODO: should it be max(PTXVersion, PTXV)? |
| 52 | } |
| 53 | |
| 54 | TLSSupported = false; |
| 55 | VLASupported = false; |
| 56 | AddrSpaceMap = &NVPTXAddrSpaceMap; |
| 57 | UseAddrSpaceMapMangling = true; |
| 58 | // __bf16 is always available as a load/store only type. |
| 59 | BFloat16Width = BFloat16Align = 16; |
| 60 | BFloat16Format = &llvm::APFloat::BFloat(); |
| 61 | |
| 62 | // Define available target features |
| 63 | // These must be defined in sorted order! |
| 64 | NoAsmVariants = true; |
| 65 | GPU = OffloadArch::UNUSED; |
| 66 | |
| 67 | // PTX supports f16 as a fundamental type. |
| 68 | HasFastHalfType = true; |
| 69 | HasFloat16 = true; |
| 70 | |
| 71 | // TODO: Make shortptr a proper ABI? |
| 72 | DataLayoutString = |
| 73 | Triple.computeDataLayout(ABIName: Opts.NVPTXUseShortPointers ? "shortptr" : "" ); |
| 74 | |
| 75 | // If possible, get a TargetInfo for our host triple, so we can match its |
| 76 | // types. |
| 77 | llvm::Triple HostTriple(Opts.HostTriple); |
| 78 | if (!HostTriple.isNVPTX()) |
| 79 | HostTarget = AllocateTarget(Triple: llvm::Triple(Opts.HostTriple), Opts); |
| 80 | |
| 81 | // If no host target, make some guesses about the data layout and return. |
| 82 | if (!HostTarget) { |
| 83 | LongWidth = LongAlign = TargetPointerWidth; |
| 84 | PointerWidth = PointerAlign = TargetPointerWidth; |
| 85 | switch (TargetPointerWidth) { |
| 86 | case 32: |
| 87 | SizeType = TargetInfo::UnsignedInt; |
| 88 | PtrDiffType = TargetInfo::SignedInt; |
| 89 | IntPtrType = TargetInfo::SignedInt; |
| 90 | break; |
| 91 | case 64: |
| 92 | SizeType = TargetInfo::UnsignedLong; |
| 93 | PtrDiffType = TargetInfo::SignedLong; |
| 94 | IntPtrType = TargetInfo::SignedLong; |
| 95 | break; |
| 96 | default: |
| 97 | llvm_unreachable("TargetPointerWidth must be 32 or 64" ); |
| 98 | } |
| 99 | |
| 100 | MaxAtomicInlineWidth = TargetPointerWidth; |
| 101 | return; |
| 102 | } |
| 103 | |
| 104 | // Copy properties from host target. |
| 105 | PointerWidth = HostTarget->getPointerWidth(AddrSpace: LangAS::Default); |
| 106 | PointerAlign = HostTarget->getPointerAlign(AddrSpace: LangAS::Default); |
| 107 | BoolWidth = HostTarget->getBoolWidth(); |
| 108 | BoolAlign = HostTarget->getBoolAlign(); |
| 109 | IntWidth = HostTarget->getIntWidth(); |
| 110 | IntAlign = HostTarget->getIntAlign(); |
| 111 | HalfWidth = HostTarget->getHalfWidth(); |
| 112 | HalfAlign = HostTarget->getHalfAlign(); |
| 113 | FloatWidth = HostTarget->getFloatWidth(); |
| 114 | FloatAlign = HostTarget->getFloatAlign(); |
| 115 | DoubleWidth = HostTarget->getDoubleWidth(); |
| 116 | DoubleAlign = HostTarget->getDoubleAlign(); |
| 117 | LongWidth = HostTarget->getLongWidth(); |
| 118 | LongAlign = HostTarget->getLongAlign(); |
| 119 | LongLongWidth = HostTarget->getLongLongWidth(); |
| 120 | LongLongAlign = HostTarget->getLongLongAlign(); |
| 121 | MinGlobalAlign = HostTarget->getMinGlobalAlign(/* TypeSize = */ Size: 0, |
| 122 | /* HasNonWeakDef = */ true); |
| 123 | NewAlign = HostTarget->getNewAlign(); |
| 124 | DefaultAlignForAttributeAligned = |
| 125 | HostTarget->getDefaultAlignForAttributeAligned(); |
| 126 | SizeType = HostTarget->getSizeType(); |
| 127 | IntMaxType = HostTarget->getIntMaxType(); |
| 128 | PtrDiffType = HostTarget->getPtrDiffType(AddrSpace: LangAS::Default); |
| 129 | IntPtrType = HostTarget->getIntPtrType(); |
| 130 | WCharType = HostTarget->getWCharType(); |
| 131 | WIntType = HostTarget->getWIntType(); |
| 132 | Char16Type = HostTarget->getChar16Type(); |
| 133 | Char32Type = HostTarget->getChar32Type(); |
| 134 | Int64Type = HostTarget->getInt64Type(); |
| 135 | SigAtomicType = HostTarget->getSigAtomicType(); |
| 136 | ProcessIDType = HostTarget->getProcessIDType(); |
| 137 | |
| 138 | UseBitFieldTypeAlignment = HostTarget->useBitFieldTypeAlignment(); |
| 139 | UseZeroLengthBitfieldAlignment = HostTarget->useZeroLengthBitfieldAlignment(); |
| 140 | UseExplicitBitFieldAlignment = HostTarget->useExplicitBitFieldAlignment(); |
| 141 | ZeroLengthBitfieldBoundary = HostTarget->getZeroLengthBitfieldBoundary(); |
| 142 | |
| 143 | // This is a bit of a lie, but it controls __GCC_ATOMIC_XXX_LOCK_FREE, and |
| 144 | // we need those macros to be identical on host and device, because (among |
| 145 | // other things) they affect which standard library classes are defined, and |
| 146 | // we need all classes to be defined on both the host and device. |
| 147 | MaxAtomicInlineWidth = HostTarget->getMaxAtomicInlineWidth(); |
| 148 | |
| 149 | // Properties intentionally not copied from host: |
| 150 | // - LargeArrayMinWidth, LargeArrayAlign: Not visible across the |
| 151 | // host/device boundary. |
| 152 | // - SuitableAlign: Not visible across the host/device boundary, and may |
| 153 | // correctly be different on host/device, e.g. if host has wider vector |
| 154 | // types than device. |
| 155 | // - LongDoubleWidth, LongDoubleAlign: nvptx's long double type is the same |
| 156 | // as its double type, but that's not necessarily true on the host. |
| 157 | // TODO: nvcc emits a warning when using long double on device; we should |
| 158 | // do the same. |
| 159 | } |
| 160 | |
| 161 | ArrayRef<const char *> NVPTXTargetInfo::getGCCRegNames() const { |
| 162 | return llvm::ArrayRef(GCCRegNames); |
| 163 | } |
| 164 | |
| 165 | bool NVPTXTargetInfo::hasFeature(StringRef Feature) const { |
| 166 | return llvm::StringSwitch<bool>(Feature) |
| 167 | .Cases(CaseStrings: {"ptx" , "nvptx" }, Value: true) |
| 168 | .Default(Value: false); |
| 169 | } |
| 170 | |
| 171 | void NVPTXTargetInfo::getTargetDefines(const LangOptions &Opts, |
| 172 | MacroBuilder &Builder) const { |
| 173 | Builder.defineMacro(Name: "__PTX__" ); |
| 174 | Builder.defineMacro(Name: "__NVPTX__" ); |
| 175 | |
| 176 | // Skip setting architecture dependent macros if undefined. |
| 177 | if (!IsNVIDIAOffloadArch(A: GPU)) |
| 178 | return; |
| 179 | |
| 180 | if (Opts.CUDAIsDevice || Opts.OpenMPIsTargetDevice || !HostTarget) { |
| 181 | // Set __CUDA_ARCH__ for the GPU specified. |
| 182 | unsigned ArchID = CudaArchToID(Arch: GPU); |
| 183 | Builder.defineMacro(Name: "__CUDA_ARCH__" , Value: llvm::Twine(ArchID)); |
| 184 | |
| 185 | if (IsNVIDIAAcceleratedOffloadArch(Arch: GPU)) |
| 186 | Builder.defineMacro( |
| 187 | Name: "__CUDA_ARCH_FEAT_SM" + llvm::Twine(ArchID / 10) + "_ALL" , Value: "1" ); |
| 188 | } |
| 189 | } |
| 190 | |
| 191 | llvm::SmallVector<Builtin::InfosShard> |
| 192 | NVPTXTargetInfo::getTargetBuiltins() const { |
| 193 | return {{.Strings: &BuiltinStrings, .Infos: BuiltinInfos}}; |
| 194 | } |
| 195 | |