| 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 "Targets.h" |
| 15 |
#include "clang/Basic/Builtins.h" |
| 16 |
#include "clang/Basic/MacroBuilder.h" |
| 17 |
#include "clang/Basic/TargetBuiltins.h" |
| 18 |
#include "llvm/ADT/StringSwitch.h" |
| 19 |
|
| 20 |
using namespace clang; |
| 21 |
using namespace clang::targets; |
| 22 |
|
| 23 |
static constexpr Builtin::Info BuiltinInfo[] = { |
| 24 |
#define BUILTIN(ID, TYPE, ATTRS) \ |
| 25 |
{#ID, TYPE, ATTRS, nullptr, HeaderDesc::NO_HEADER, ALL_LANGUAGES}, |
| 26 |
#define LIBBUILTIN(ID, TYPE, ATTRS, HEADER) \ |
| 27 |
{#ID, TYPE, ATTRS, nullptr, HeaderDesc::HEADER, ALL_LANGUAGES}, |
| 28 |
#define TARGET_BUILTIN(ID, TYPE, ATTRS, FEATURE) \ |
| 29 |
{#ID, TYPE, ATTRS, FEATURE, HeaderDesc::NO_HEADER, ALL_LANGUAGES}, |
| 30 |
#include "clang/Basic/BuiltinsNVPTX.def" |
| 31 |
}; |
| 32 |
|
| 33 |
const char *const NVPTXTargetInfo::GCCRegNames[] = {"r0"}; |
| 34 |
|
| 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."); |
| 41 |
|
| 42 |
PTXVersion = 32; |
| 43 |
for (const StringRef Feature : Opts.FeaturesAsWritten) { |
| 44 |
int PTXV; |
| 45 |
if (!Feature.startswith("+ptx") || |
| 46 |
Feature.drop_front(4).getAsInteger(10, PTXV)) |
| 47 |
continue; |
| 48 |
PTXVersion = PTXV; // TODO: should it be max(PTXVersion, PTXV)? |
| 49 |
} |
| 50 |
|
| 51 |
TLSSupported = false; |
| 52 |
VLASupported = false; |
| 53 |
AddrSpaceMap = &NVPTXAddrSpaceMap; |
| 54 |
UseAddrSpaceMapMangling = true; |
| 55 |
// __bf16 is always available as a load/store only type. |
| 56 |
BFloat16Width = BFloat16Align = 16; |
| 57 |
BFloat16Format = &llvm::APFloat::BFloat(); |
| 58 |
|
| 59 |
// Define available target features |
| 60 |
// These must be defined in sorted order! |
| 61 |
NoAsmVariants = true; |
| 62 |
GPU = CudaArch::SM_20; |
| 63 |
|
| 64 |
if (TargetPointerWidth == 32) |
| 65 |
resetDataLayout("e-p:32:32-i64:64-i128:128-v16:16-v32:32-n16:32:64"); |
| 66 |
else if (Opts.NVPTXUseShortPointers) |
| 67 |
resetDataLayout( |
| 68 |
"e-p3:32:32-p4:32:32-p5:32:32-i64:64-i128:128-v16:16-v32:32-n16:32:64"); |
| 69 |
else |
| 70 |
resetDataLayout("e-i64:64-i128:128-v16:16-v32:32-n16:32:64"); |
| 71 |
|
| 72 |
// If possible, get a TargetInfo for our host triple, so we can match its |
| 73 |
// types. |
| 74 |
llvm::Triple HostTriple(Opts.HostTriple); |
| 75 |
if (!HostTriple.isNVPTX()) |
| 76 |
HostTarget.reset(AllocateTarget(llvm::Triple(Opts.HostTriple), Opts)); |
| 77 |
|
| 78 |
// If no host target, make some guesses about the data layout and return. |
| 79 |
if (!HostTarget) { |
| 80 |
LongWidth = LongAlign = TargetPointerWidth; |
| 81 |
PointerWidth = PointerAlign = TargetPointerWidth; |
| 82 |
switch (TargetPointerWidth) { |
| 83 |
case 32: |
| 84 |
SizeType = TargetInfo::UnsignedInt; |
| 85 |
PtrDiffType = TargetInfo::SignedInt; |
| 86 |
IntPtrType = TargetInfo::SignedInt; |
| 87 |
break; |
| 88 |
case 64: |
| 89 |
SizeType = TargetInfo::UnsignedLong; |
| 90 |
PtrDiffType = TargetInfo::SignedLong; |
| 91 |
IntPtrType = TargetInfo::SignedLong; |
| 92 |
break; |
| 93 |
default: |
| 94 |
llvm_unreachable("TargetPointerWidth must be 32 or 64"); |
| 95 |
} |
| 96 |
return; |
| 97 |
} |
| 98 |
|
| 99 |
// Copy properties from host target. |
| 100 |
PointerWidth = HostTarget->getPointerWidth(LangAS::Default); |
| 101 |
PointerAlign = HostTarget->getPointerAlign(LangAS::Default); |
| 102 |
BoolWidth = HostTarget->getBoolWidth(); |
| 103 |
BoolAlign = HostTarget->getBoolAlign(); |
| 104 |
IntWidth = HostTarget->getIntWidth(); |
| 105 |
IntAlign = HostTarget->getIntAlign(); |
| 106 |
HalfWidth = HostTarget->getHalfWidth(); |
| 107 |
HalfAlign = HostTarget->getHalfAlign(); |
| 108 |
FloatWidth = HostTarget->getFloatWidth(); |
| 109 |
FloatAlign = HostTarget->getFloatAlign(); |
| 110 |
DoubleWidth = HostTarget->getDoubleWidth(); |
| 111 |
DoubleAlign = HostTarget->getDoubleAlign(); |
| 112 |
LongWidth = HostTarget->getLongWidth(); |
| 113 |
LongAlign = HostTarget->getLongAlign(); |
| 114 |
LongLongWidth = HostTarget->getLongLongWidth(); |
| 115 |
LongLongAlign = HostTarget->getLongLongAlign(); |
| 116 |
MinGlobalAlign = HostTarget->getMinGlobalAlign(/* TypeSize = */ 0); |
| 117 |
NewAlign = HostTarget->getNewAlign(); |
| 118 |
DefaultAlignForAttributeAligned = |
| 119 |
HostTarget->getDefaultAlignForAttributeAligned(); |
| 120 |
SizeType = HostTarget->getSizeType(); |
| 121 |
IntMaxType = HostTarget->getIntMaxType(); |
| 122 |
PtrDiffType = HostTarget->getPtrDiffType(LangAS::Default); |
| 123 |
IntPtrType = HostTarget->getIntPtrType(); |
| 124 |
WCharType = HostTarget->getWCharType(); |
| 125 |
WIntType = HostTarget->getWIntType(); |
| 126 |
Char16Type = HostTarget->getChar16Type(); |
| 127 |
Char32Type = HostTarget->getChar32Type(); |
| 128 |
Int64Type = HostTarget->getInt64Type(); |
| 129 |
SigAtomicType = HostTarget->getSigAtomicType(); |
| 130 |
ProcessIDType = HostTarget->getProcessIDType(); |
| 131 |
|
| 132 |
UseBitFieldTypeAlignment = HostTarget->useBitFieldTypeAlignment(); |
| 133 |
UseZeroLengthBitfieldAlignment = HostTarget->useZeroLengthBitfieldAlignment(); |
| 134 |
UseExplicitBitFieldAlignment = HostTarget->useExplicitBitFieldAlignment(); |
| 135 |
ZeroLengthBitfieldBoundary = HostTarget->getZeroLengthBitfieldBoundary(); |
| 136 |
|
| 137 |
// This is a bit of a lie, but it controls __GCC_ATOMIC_XXX_LOCK_FREE, and |
| 138 |
// we need those macros to be identical on host and device, because (among |
| 139 |
// other things) they affect which standard library classes are defined, and |
| 140 |
// we need all classes to be defined on both the host and device. |
| 141 |
MaxAtomicInlineWidth = HostTarget->getMaxAtomicInlineWidth(); |
| 142 |
|
| 143 |
// Properties intentionally not copied from host: |
| 144 |
// - LargeArrayMinWidth, LargeArrayAlign: Not visible across the |
| 145 |
// host/device boundary. |
| 146 |
// - SuitableAlign: Not visible across the host/device boundary, and may |
| 147 |
// correctly be different on host/device, e.g. if host has wider vector |
| 148 |
// types than device. |
| 149 |
// - LongDoubleWidth, LongDoubleAlign: nvptx's long double type is the same |
| 150 |
// as its double type, but that's not necessarily true on the host. |
| 151 |
// TODO: nvcc emits a warning when using long double on device; we should |
| 152 |
// do the same. |
| 153 |
} |
| 154 |
|
| 155 |
ArrayRef<const char *> NVPTXTargetInfo::getGCCRegNames() const { |
| 156 |
return llvm::ArrayRef(GCCRegNames); |
| 157 |
} |
| 158 |
|
| 159 |
bool NVPTXTargetInfo::hasFeature(StringRef Feature) const { |
| 160 |
return llvm::StringSwitch<bool>(Feature) |
| 161 |
.Cases("ptx", "nvptx", true) |
| 162 |
.Default(false); |
| 163 |
} |
| 164 |
|
| 165 |
void NVPTXTargetInfo::getTargetDefines(const LangOptions &Opts, |
| 166 |
MacroBuilder &Builder) const { |
| 167 |
Builder.defineMacro("__PTX__"); |
| 168 |
Builder.defineMacro("__NVPTX__"); |
| 169 |
if (Opts.CUDAIsDevice || Opts.OpenMPIsDevice) { |
| 170 |
// Set __CUDA_ARCH__ for the GPU specified. |
| 171 |
std::string CUDAArchCode = [this] { |
| 172 |
switch (GPU) { |
| 173 |
case CudaArch::GFX600: |
| 174 |
case CudaArch::GFX601: |
| 175 |
case CudaArch::GFX602: |
| 176 |
case CudaArch::GFX700: |
| 177 |
case CudaArch::GFX701: |
| 178 |
case CudaArch::GFX702: |
| 179 |
case CudaArch::GFX703: |
| 180 |
case CudaArch::GFX704: |
| 181 |
case CudaArch::GFX705: |
| 182 |
case CudaArch::GFX801: |
| 183 |
case CudaArch::GFX802: |
| 184 |
case CudaArch::GFX803: |
| 185 |
case CudaArch::GFX805: |
| 186 |
case CudaArch::GFX810: |
| 187 |
case CudaArch::GFX900: |
| 188 |
case CudaArch::GFX902: |
| 189 |
case CudaArch::GFX904: |
| 190 |
case CudaArch::GFX906: |
| 191 |
case CudaArch::GFX908: |
| 192 |
case CudaArch::GFX909: |
| 193 |
case CudaArch::GFX90a: |
| 194 |
case CudaArch::GFX90c: |
| 195 |
case CudaArch::GFX940: |
| 196 |
case CudaArch::GFX1010: |
| 197 |
case CudaArch::GFX1011: |
| 198 |
case CudaArch::GFX1012: |
| 199 |
case CudaArch::GFX1013: |
| 200 |
case CudaArch::GFX1030: |
| 201 |
case CudaArch::GFX1031: |
| 202 |
case CudaArch::GFX1032: |
| 203 |
case CudaArch::GFX1033: |
| 204 |
case CudaArch::GFX1034: |
| 205 |
case CudaArch::GFX1035: |
| 206 |
case CudaArch::GFX1036: |
| 207 |
case CudaArch::GFX1100: |
| 208 |
case CudaArch::GFX1101: |
| 209 |
case CudaArch::GFX1102: |
| 210 |
case CudaArch::GFX1103: |
| 211 |
case CudaArch::Generic: |
| 212 |
case CudaArch::LAST: |
| 213 |
break; |
| 214 |
case CudaArch::UNUSED: |
| 215 |
case CudaArch::UNKNOWN: |
| 216 |
assert(false && "No GPU arch when compiling CUDA device code."); |
| 217 |
return ""; |
| 218 |
case CudaArch::SM_20: |
| 219 |
return "200"; |
| 220 |
case CudaArch::SM_21: |
| 221 |
return "210"; |
| 222 |
case CudaArch::SM_30: |
| 223 |
return "300"; |
| 224 |
case CudaArch::SM_32: |
| 225 |
return "320"; |
| 226 |
case CudaArch::SM_35: |
| 227 |
return "350"; |
| 228 |
case CudaArch::SM_37: |
| 229 |
return "370"; |
| 230 |
case CudaArch::SM_50: |
| 231 |
return "500"; |
| 232 |
case CudaArch::SM_52: |
| 233 |
return "520"; |
| 234 |
case CudaArch::SM_53: |
| 235 |
return "530"; |
| 236 |
case CudaArch::SM_60: |
| 237 |
return "600"; |
| 238 |
case CudaArch::SM_61: |
| 239 |
return "610"; |
| 240 |
case CudaArch::SM_62: |
| 241 |
return "620"; |
| 242 |
case CudaArch::SM_70: |
| 243 |
return "700"; |
| 244 |
case CudaArch::SM_72: |
| 245 |
return "720"; |
| 246 |
case CudaArch::SM_75: |
| 247 |
return "750"; |
| 248 |
case CudaArch::SM_80: |
| 249 |
return "800"; |
| 250 |
case CudaArch::SM_86: |
| 251 |
return "860"; |
| 252 |
case CudaArch::SM_87: |
| 253 |
return "870"; |
| 254 |
case CudaArch::SM_89: |
| 255 |
return "890"; |
| 256 |
case CudaArch::SM_90: |
| 257 |
return "900"; |
| 258 |
} |
| 259 |
llvm_unreachable("unhandled CudaArch"); |
| 260 |
}(); |
| 261 |
Builder.defineMacro("__CUDA_ARCH__", CUDAArchCode); |
| 262 |
} |
| 263 |
} |
| 264 |
|
| 265 |
ArrayRef<Builtin::Info> NVPTXTargetInfo::getTargetBuiltins() const { |
| 266 |
return llvm::ArrayRef(BuiltinInfo, |
| 267 |
clang::NVPTX::LastTSBuiltin - Builtin::FirstTSBuiltin); |
| 268 |
} |