1 //===--- NVPTX.cpp - Implement NVPTX target feature support ---------------===//
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
7 //===----------------------------------------------------------------------===//
9 // This file implements NVPTX TargetInfo objects.
11 //===----------------------------------------------------------------------===//
15 #include "clang/Basic/Builtins.h"
16 #include "clang/Basic/MacroBuilder.h"
17 #include "clang/Basic/TargetBuiltins.h"
18 #include "llvm/ADT/StringSwitch.h"
20 using namespace clang;
21 using namespace clang::targets;
23 const Builtin::Info NVPTXTargetInfo::BuiltinInfo[] = {
24 #define BUILTIN(ID, TYPE, ATTRS) \
25 {#ID, TYPE, ATTRS, nullptr, ALL_LANGUAGES, nullptr},
26 #define LIBBUILTIN(ID, TYPE, ATTRS, HEADER) \
27 {#ID, TYPE, ATTRS, HEADER, ALL_LANGUAGES, nullptr},
28 #define TARGET_BUILTIN(ID, TYPE, ATTRS, FEATURE) \
29 {#ID, TYPE, ATTRS, nullptr, ALL_LANGUAGES, FEATURE},
30 #include "clang/Basic/BuiltinsNVPTX.def"
33 const char *const NVPTXTargetInfo::GCCRegNames[] = {"r0"};
35 NVPTXTargetInfo::NVPTXTargetInfo(const llvm::Triple &Triple,
36 const TargetOptions &Opts,
37 unsigned TargetPointerWidth)
38 : TargetInfo(Triple) {
39 assert((TargetPointerWidth == 32 || TargetPointerWidth == 64) &&
40 "NVPTX only supports 32- and 64-bit modes.");
43 for (const StringRef Feature : Opts.FeaturesAsWritten) {
44 if (!Feature.startswith("+ptx"))
46 PTXVersion = llvm::StringSwitch<unsigned>(Feature)
62 AddrSpaceMap = &NVPTXAddrSpaceMap;
63 UseAddrSpaceMapMangling = true;
65 // Define available target features
66 // These must be defined in sorted order!
68 GPU = CudaArch::SM_20;
70 if (TargetPointerWidth == 32)
71 resetDataLayout("e-p:32:32-i64:64-i128:128-v16:16-v32:32-n16:32:64");
72 else if (Opts.NVPTXUseShortPointers)
74 "e-p3:32:32-p4:32:32-p5:32:32-i64:64-i128:128-v16:16-v32:32-n16:32:64");
76 resetDataLayout("e-i64:64-i128:128-v16:16-v32:32-n16:32:64");
78 // If possible, get a TargetInfo for our host triple, so we can match its
80 llvm::Triple HostTriple(Opts.HostTriple);
81 if (!HostTriple.isNVPTX())
82 HostTarget.reset(AllocateTarget(llvm::Triple(Opts.HostTriple), Opts));
84 // If no host target, make some guesses about the data layout and return.
86 LongWidth = LongAlign = TargetPointerWidth;
87 PointerWidth = PointerAlign = TargetPointerWidth;
88 switch (TargetPointerWidth) {
90 SizeType = TargetInfo::UnsignedInt;
91 PtrDiffType = TargetInfo::SignedInt;
92 IntPtrType = TargetInfo::SignedInt;
95 SizeType = TargetInfo::UnsignedLong;
96 PtrDiffType = TargetInfo::SignedLong;
97 IntPtrType = TargetInfo::SignedLong;
100 llvm_unreachable("TargetPointerWidth must be 32 or 64");
105 // Copy properties from host target.
106 PointerWidth = HostTarget->getPointerWidth(/* AddrSpace = */ 0);
107 PointerAlign = HostTarget->getPointerAlign(/* AddrSpace = */ 0);
108 BoolWidth = HostTarget->getBoolWidth();
109 BoolAlign = HostTarget->getBoolAlign();
110 IntWidth = HostTarget->getIntWidth();
111 IntAlign = HostTarget->getIntAlign();
112 HalfWidth = HostTarget->getHalfWidth();
113 HalfAlign = HostTarget->getHalfAlign();
114 FloatWidth = HostTarget->getFloatWidth();
115 FloatAlign = HostTarget->getFloatAlign();
116 DoubleWidth = HostTarget->getDoubleWidth();
117 DoubleAlign = HostTarget->getDoubleAlign();
118 LongWidth = HostTarget->getLongWidth();
119 LongAlign = HostTarget->getLongAlign();
120 LongLongWidth = HostTarget->getLongLongWidth();
121 LongLongAlign = HostTarget->getLongLongAlign();
122 MinGlobalAlign = HostTarget->getMinGlobalAlign(/* TypeSize = */ 0);
123 NewAlign = HostTarget->getNewAlign();
124 DefaultAlignForAttributeAligned =
125 HostTarget->getDefaultAlignForAttributeAligned();
126 SizeType = HostTarget->getSizeType();
127 IntMaxType = HostTarget->getIntMaxType();
128 PtrDiffType = HostTarget->getPtrDiffType(/* AddrSpace = */ 0);
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();
138 UseBitFieldTypeAlignment = HostTarget->useBitFieldTypeAlignment();
139 UseZeroLengthBitfieldAlignment = HostTarget->useZeroLengthBitfieldAlignment();
140 UseExplicitBitFieldAlignment = HostTarget->useExplicitBitFieldAlignment();
141 ZeroLengthBitfieldBoundary = HostTarget->getZeroLengthBitfieldBoundary();
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();
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
161 ArrayRef<const char *> NVPTXTargetInfo::getGCCRegNames() const {
162 return llvm::makeArrayRef(GCCRegNames);
165 bool NVPTXTargetInfo::hasFeature(StringRef Feature) const {
166 return llvm::StringSwitch<bool>(Feature)
167 .Cases("ptx", "nvptx", true)
171 void NVPTXTargetInfo::getTargetDefines(const LangOptions &Opts,
172 MacroBuilder &Builder) const {
173 Builder.defineMacro("__PTX__");
174 Builder.defineMacro("__NVPTX__");
175 if (Opts.CUDAIsDevice) {
176 // Set __CUDA_ARCH__ for the GPU specified.
177 std::string CUDAArchCode = [this] {
179 case CudaArch::GFX600:
180 case CudaArch::GFX601:
181 case CudaArch::GFX700:
182 case CudaArch::GFX701:
183 case CudaArch::GFX702:
184 case CudaArch::GFX703:
185 case CudaArch::GFX704:
186 case CudaArch::GFX801:
187 case CudaArch::GFX802:
188 case CudaArch::GFX803:
189 case CudaArch::GFX810:
190 case CudaArch::GFX900:
191 case CudaArch::GFX902:
192 case CudaArch::GFX904:
193 case CudaArch::GFX906:
194 case CudaArch::GFX908:
195 case CudaArch::GFX909:
196 case CudaArch::GFX1010:
197 case CudaArch::GFX1011:
198 case CudaArch::GFX1012:
201 case CudaArch::UNKNOWN:
202 assert(false && "No GPU arch when compiling CUDA device code.");
204 case CudaArch::SM_20:
206 case CudaArch::SM_21:
208 case CudaArch::SM_30:
210 case CudaArch::SM_32:
212 case CudaArch::SM_35:
214 case CudaArch::SM_37:
216 case CudaArch::SM_50:
218 case CudaArch::SM_52:
220 case CudaArch::SM_53:
222 case CudaArch::SM_60:
224 case CudaArch::SM_61:
226 case CudaArch::SM_62:
228 case CudaArch::SM_70:
230 case CudaArch::SM_72:
232 case CudaArch::SM_75:
235 llvm_unreachable("unhandled CudaArch");
237 Builder.defineMacro("__CUDA_ARCH__", CUDAArchCode);
241 ArrayRef<Builtin::Info> NVPTXTargetInfo::getTargetBuiltins() const {
242 return llvm::makeArrayRef(BuiltinInfo, clang::NVPTX::LastTSBuiltin -
243 Builtin::FirstTSBuiltin);