llvm/lib/TargetParser/TargetParser.cpp

   1 //===-- TargetParser - Parser for target features ---------------*- 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 implements a target parser to recognise hardware features such as
  10 // FPU/CPU/ARCH names as well as specific support such as HDIV, etc.
  11 //
  12 //===----------------------------------------------------------------------===//
  13
  14 #include "llvm/TargetParser/TargetParser.h"
  15 #include "llvm/ADT/ArrayRef.h"
  16 #include "llvm/TargetParser/Triple.h"
  17
  18 using namespace llvm;
  19 using namespace AMDGPU;
  20
  21 namespace {
  22
  23 struct GPUInfo {
  24   StringLiteral Name;
  25   StringLiteral CanonicalName;
  26   AMDGPU::GPUKind Kind;
  27   unsigned Features;
  28 };
  29
  30 constexpr GPUInfo R600GPUs[] = {
  31   // Name       Canonical    Kind        Features
  32   //            Name
  33   {{"r600"},    {"r600"},    GK_R600,    FEATURE_NONE },
  34   {{"rv630"},   {"r600"},    GK_R600,    FEATURE_NONE },
  35   {{"rv635"},   {"r600"},    GK_R600,    FEATURE_NONE },
  36   {{"r630"},    {"r630"},    GK_R630,    FEATURE_NONE },
  37   {{"rs780"},   {"rs880"},   GK_RS880,   FEATURE_NONE },
  38   {{"rs880"},   {"rs880"},   GK_RS880,   FEATURE_NONE },
  39   {{"rv610"},   {"rs880"},   GK_RS880,   FEATURE_NONE },
  40   {{"rv620"},   {"rs880"},   GK_RS880,   FEATURE_NONE },
  41   {{"rv670"},   {"rv670"},   GK_RV670,   FEATURE_NONE },
  42   {{"rv710"},   {"rv710"},   GK_RV710,   FEATURE_NONE },
  43   {{"rv730"},   {"rv730"},   GK_RV730,   FEATURE_NONE },
  44   {{"rv740"},   {"rv770"},   GK_RV770,   FEATURE_NONE },
  45   {{"rv770"},   {"rv770"},   GK_RV770,   FEATURE_NONE },
  46   {{"cedar"},   {"cedar"},   GK_CEDAR,   FEATURE_NONE },
  47   {{"palm"},    {"cedar"},   GK_CEDAR,   FEATURE_NONE },
  48   {{"cypress"}, {"cypress"}, GK_CYPRESS, FEATURE_FMA  },
  49   {{"hemlock"}, {"cypress"}, GK_CYPRESS, FEATURE_FMA  },
  50   {{"juniper"}, {"juniper"}, GK_JUNIPER, FEATURE_NONE },
  51   {{"redwood"}, {"redwood"}, GK_REDWOOD, FEATURE_NONE },
  52   {{"sumo"},    {"sumo"},    GK_SUMO,    FEATURE_NONE },
  53   {{"sumo2"},   {"sumo"},    GK_SUMO,    FEATURE_NONE },
  54   {{"barts"},   {"barts"},   GK_BARTS,   FEATURE_NONE },
  55   {{"caicos"},  {"caicos"},  GK_CAICOS,  FEATURE_NONE },
  56   {{"aruba"},   {"cayman"},  GK_CAYMAN,  FEATURE_FMA  },
  57   {{"cayman"},  {"cayman"},  GK_CAYMAN,  FEATURE_FMA  },
  58   {{"turks"},   {"turks"},   GK_TURKS,   FEATURE_NONE }
  59 };
  60
  61 // This table should be sorted by the value of GPUKind
  62 // Don't bother listing the implicitly true features
  63 constexpr GPUInfo AMDGCNGPUs[] = {
  64   // Name         Canonical    Kind        Features
  65   //              Name
  66   {{"gfx600"},    {"gfx600"},  GK_GFX600,  FEATURE_FAST_FMA_F32},
  67   {{"tahiti"},    {"gfx600"},  GK_GFX600,  FEATURE_FAST_FMA_F32},
  68   {{"gfx601"},    {"gfx601"},  GK_GFX601,  FEATURE_NONE},
  69   {{"pitcairn"},  {"gfx601"},  GK_GFX601,  FEATURE_NONE},
  70   {{"verde"},     {"gfx601"},  GK_GFX601,  FEATURE_NONE},
  71   {{"gfx602"},    {"gfx602"},  GK_GFX602,  FEATURE_NONE},
  72   {{"hainan"},    {"gfx602"},  GK_GFX602,  FEATURE_NONE},
  73   {{"oland"},     {"gfx602"},  GK_GFX602,  FEATURE_NONE},
  74   {{"gfx700"},    {"gfx700"},  GK_GFX700,  FEATURE_NONE},
  75   {{"kaveri"},    {"gfx700"},  GK_GFX700,  FEATURE_NONE},
  76   {{"gfx701"},    {"gfx701"},  GK_GFX701,  FEATURE_FAST_FMA_F32},
  77   {{"hawaii"},    {"gfx701"},  GK_GFX701,  FEATURE_FAST_FMA_F32},
  78   {{"gfx702"},    {"gfx702"},  GK_GFX702,  FEATURE_FAST_FMA_F32},
  79   {{"gfx703"},    {"gfx703"},  GK_GFX703,  FEATURE_NONE},
  80   {{"kabini"},    {"gfx703"},  GK_GFX703,  FEATURE_NONE},
  81   {{"mullins"},   {"gfx703"},  GK_GFX703,  FEATURE_NONE},
  82   {{"gfx704"},    {"gfx704"},  GK_GFX704,  FEATURE_NONE},
  83   {{"bonaire"},   {"gfx704"},  GK_GFX704,  FEATURE_NONE},
  84   {{"gfx705"},    {"gfx705"},  GK_GFX705,  FEATURE_NONE},
  85   {{"gfx801"},    {"gfx801"},  GK_GFX801,  FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK},
  86   {{"carrizo"},   {"gfx801"},  GK_GFX801,  FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK},
  87   {{"gfx802"},    {"gfx802"},  GK_GFX802,  FEATURE_FAST_DENORMAL_F32},
  88   {{"iceland"},   {"gfx802"},  GK_GFX802,  FEATURE_FAST_DENORMAL_F32},
  89   {{"tonga"},     {"gfx802"},  GK_GFX802,  FEATURE_FAST_DENORMAL_F32},
  90   {{"gfx803"},    {"gfx803"},  GK_GFX803,  FEATURE_FAST_DENORMAL_F32},
  91   {{"fiji"},      {"gfx803"},  GK_GFX803,  FEATURE_FAST_DENORMAL_F32},
  92   {{"polaris10"}, {"gfx803"},  GK_GFX803,  FEATURE_FAST_DENORMAL_F32},
  93   {{"polaris11"}, {"gfx803"},  GK_GFX803,  FEATURE_FAST_DENORMAL_F32},
  94   {{"gfx805"},    {"gfx805"},  GK_GFX805,  FEATURE_FAST_DENORMAL_F32},
  95   {{"tongapro"},  {"gfx805"},  GK_GFX805,  FEATURE_FAST_DENORMAL_F32},
  96   {{"gfx810"},    {"gfx810"},  GK_GFX810,  FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK},
  97   {{"stoney"},    {"gfx810"},  GK_GFX810,  FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK},
  98   {{"gfx900"},    {"gfx900"},  GK_GFX900,  FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK},
  99   {{"gfx902"},    {"gfx902"},  GK_GFX902,  FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK},
 100   {{"gfx904"},    {"gfx904"},  GK_GFX904,  FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK},
 101   {{"gfx906"},    {"gfx906"},  GK_GFX906,  FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK|FEATURE_SRAMECC},
 102   {{"gfx908"},    {"gfx908"},  GK_GFX908,  FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK|FEATURE_SRAMECC},
 103   {{"gfx909"},    {"gfx909"},  GK_GFX909,  FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK},
 104   {{"gfx90a"},    {"gfx90a"},  GK_GFX90A,  FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK|FEATURE_SRAMECC},
 105   {{"gfx90c"},    {"gfx90c"},  GK_GFX90C,  FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK},
 106   {{"gfx940"},    {"gfx940"},  GK_GFX940,  FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK|FEATURE_SRAMECC},
 107   {{"gfx941"},    {"gfx941"},  GK_GFX941,  FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK|FEATURE_SRAMECC},
 108   {{"gfx942"},    {"gfx942"},  GK_GFX942,  FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK|FEATURE_SRAMECC},
 109   {{"gfx1010"},   {"gfx1010"}, GK_GFX1010, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_XNACK|FEATURE_WGP},
 110   {{"gfx1011"},   {"gfx1011"}, GK_GFX1011, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_XNACK|FEATURE_WGP},
 111   {{"gfx1012"},   {"gfx1012"}, GK_GFX1012, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_XNACK|FEATURE_WGP},
 112   {{"gfx1013"},   {"gfx1013"}, GK_GFX1013, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_XNACK|FEATURE_WGP},
 113   {{"gfx1030"},   {"gfx1030"}, GK_GFX1030, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP},
 114   {{"gfx1031"},   {"gfx1031"}, GK_GFX1031, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP},
 115   {{"gfx1032"},   {"gfx1032"}, GK_GFX1032, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP},
 116   {{"gfx1033"},   {"gfx1033"}, GK_GFX1033, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP},
 117   {{"gfx1034"},   {"gfx1034"}, GK_GFX1034, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP},
 118   {{"gfx1035"},   {"gfx1035"}, GK_GFX1035, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP},
 119   {{"gfx1036"},   {"gfx1036"}, GK_GFX1036, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP},
 120   {{"gfx1100"},   {"gfx1100"}, GK_GFX1100, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP},
 121   {{"gfx1101"},   {"gfx1101"}, GK_GFX1101, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP},
 122   {{"gfx1102"},   {"gfx1102"}, GK_GFX1102, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP},
 123   {{"gfx1103"},   {"gfx1103"}, GK_GFX1103, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP},
 124   {{"gfx1150"},   {"gfx1150"}, GK_GFX1150, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP},
 125   {{"gfx1151"},   {"gfx1151"}, GK_GFX1151, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP},
 126 };
 127
 128 const GPUInfo *getArchEntry(AMDGPU::GPUKind AK, ArrayRef<GPUInfo> Table) {
 129   GPUInfo Search = { {""}, {""}, AK, AMDGPU::FEATURE_NONE };
 130
 131   auto I =
 132       llvm::lower_bound(Table, Search, [](const GPUInfo &A, const GPUInfo &B) {
 133         return A.Kind < B.Kind;
 134       });
 135
 136   if (I == Table.end() || I->Kind != Search.Kind)
 137     return nullptr;
 138   return I;
 139 }
 140
 141 } // namespace
 142
 143 StringRef llvm::AMDGPU::getArchNameAMDGCN(GPUKind AK) {
 144   if (const auto *Entry = getArchEntry(AK, AMDGCNGPUs))
 145     return Entry->CanonicalName;
 146   return "";
 147 }
 148
 149 StringRef llvm::AMDGPU::getArchNameR600(GPUKind AK) {
 150   if (const auto *Entry = getArchEntry(AK, R600GPUs))
 151     return Entry->CanonicalName;
 152   return "";
 153 }
 154
 155 AMDGPU::GPUKind llvm::AMDGPU::parseArchAMDGCN(StringRef CPU) {
 156   for (const auto &C : AMDGCNGPUs) {
 157     if (CPU == C.Name)
 158       return C.Kind;
 159   }
 160
 161   return AMDGPU::GPUKind::GK_NONE;
 162 }
 163
 164 AMDGPU::GPUKind llvm::AMDGPU::parseArchR600(StringRef CPU) {
 165   for (const auto &C : R600GPUs) {
 166     if (CPU == C.Name)
 167       return C.Kind;
 168   }
 169
 170   return AMDGPU::GPUKind::GK_NONE;
 171 }
 172
 173 unsigned AMDGPU::getArchAttrAMDGCN(GPUKind AK) {
 174   if (const auto *Entry = getArchEntry(AK, AMDGCNGPUs))
 175     return Entry->Features;
 176   return FEATURE_NONE;
 177 }
 178
 179 unsigned AMDGPU::getArchAttrR600(GPUKind AK) {
 180   if (const auto *Entry = getArchEntry(AK, R600GPUs))
 181     return Entry->Features;
 182   return FEATURE_NONE;
 183 }
 184
 185 void AMDGPU::fillValidArchListAMDGCN(SmallVectorImpl<StringRef> &Values) {
 186   // XXX: Should this only report unique canonical names?
 187   for (const auto &C : AMDGCNGPUs)
 188     Values.push_back(C.Name);
 189 }
 190
 191 void AMDGPU::fillValidArchListR600(SmallVectorImpl<StringRef> &Values) {
 192   for (const auto &C : R600GPUs)
 193     Values.push_back(C.Name);
 194 }
 195
 196 AMDGPU::IsaVersion AMDGPU::getIsaVersion(StringRef GPU) {
 197   AMDGPU::GPUKind AK = parseArchAMDGCN(GPU);
 198   if (AK == AMDGPU::GPUKind::GK_NONE) {
 199     if (GPU == "generic-hsa")
 200       return {7, 0, 0};
 201     if (GPU == "generic")
 202       return {6, 0, 0};
 203     return {0, 0, 0};
 204   }
 205
 206   switch (AK) {
 207   case GK_GFX600:  return {6, 0, 0};
 208   case GK_GFX601:  return {6, 0, 1};
 209   case GK_GFX602:  return {6, 0, 2};
 210   case GK_GFX700:  return {7, 0, 0};
 211   case GK_GFX701:  return {7, 0, 1};
 212   case GK_GFX702:  return {7, 0, 2};
 213   case GK_GFX703:  return {7, 0, 3};
 214   case GK_GFX704:  return {7, 0, 4};
 215   case GK_GFX705:  return {7, 0, 5};
 216   case GK_GFX801:  return {8, 0, 1};
 217   case GK_GFX802:  return {8, 0, 2};
 218   case GK_GFX803:  return {8, 0, 3};
 219   case GK_GFX805:  return {8, 0, 5};
 220   case GK_GFX810:  return {8, 1, 0};
 221   case GK_GFX900:  return {9, 0, 0};
 222   case GK_GFX902:  return {9, 0, 2};
 223   case GK_GFX904:  return {9, 0, 4};
 224   case GK_GFX906:  return {9, 0, 6};
 225   case GK_GFX908:  return {9, 0, 8};
 226   case GK_GFX909:  return {9, 0, 9};
 227   case GK_GFX90A:  return {9, 0, 10};
 228   case GK_GFX90C:  return {9, 0, 12};
 229   case GK_GFX940:  return {9, 4, 0};
 230   case GK_GFX941:  return {9, 4, 1};
 231   case GK_GFX942:  return {9, 4, 2};
 232   case GK_GFX1010: return {10, 1, 0};
 233   case GK_GFX1011: return {10, 1, 1};
 234   case GK_GFX1012: return {10, 1, 2};
 235   case GK_GFX1013: return {10, 1, 3};
 236   case GK_GFX1030: return {10, 3, 0};
 237   case GK_GFX1031: return {10, 3, 1};
 238   case GK_GFX1032: return {10, 3, 2};
 239   case GK_GFX1033: return {10, 3, 3};
 240   case GK_GFX1034: return {10, 3, 4};
 241   case GK_GFX1035: return {10, 3, 5};
 242   case GK_GFX1036: return {10, 3, 6};
 243   case GK_GFX1100: return {11, 0, 0};
 244   case GK_GFX1101: return {11, 0, 1};
 245   case GK_GFX1102: return {11, 0, 2};
 246   case GK_GFX1103: return {11, 0, 3};
 247   case GK_GFX1150: return {11, 5, 0};
 248   case GK_GFX1151: return {11, 5, 1};
 249   default:         return {0, 0, 0};
 250   }
 251 }
 252
 253 StringRef AMDGPU::getCanonicalArchName(const Triple &T, StringRef Arch) {
 254   assert(T.isAMDGPU());
 255   auto ProcKind = T.isAMDGCN() ? parseArchAMDGCN(Arch) : parseArchR600(Arch);
 256   if (ProcKind == GK_NONE)
 257     return StringRef();
 258
 259   return T.isAMDGCN() ? getArchNameAMDGCN(ProcKind) : getArchNameR600(ProcKind);
 260 }
 261
 262 void AMDGPU::fillAMDGPUFeatureMap(StringRef GPU, const Triple &T,
 263                                   StringMap<bool> &Features) {
 264   // XXX - What does the member GPU mean if device name string passed here?
 265   if (T.isAMDGCN()) {
 266     switch (parseArchAMDGCN(GPU)) {
 267     case GK_GFX1151:
 268     case GK_GFX1150:
 269     case GK_GFX1103:
 270     case GK_GFX1102:
 271     case GK_GFX1101:
 272     case GK_GFX1100:
 273       Features["ci-insts"] = true;
 274       Features["dot5-insts"] = true;
 275       Features["dot7-insts"] = true;
 276       Features["dot8-insts"] = true;
 277       Features["dot9-insts"] = true;
 278       Features["dot10-insts"] = true;
 279       Features["dl-insts"] = true;
 280       Features["16-bit-insts"] = true;
 281       Features["dpp"] = true;
 282       Features["gfx8-insts"] = true;
 283       Features["gfx9-insts"] = true;
 284       Features["gfx10-insts"] = true;
 285       Features["gfx10-3-insts"] = true;
 286       Features["gfx11-insts"] = true;
 287       Features["atomic-fadd-rtn-insts"] = true;
 288       Features["image-insts"] = true;
 289       Features["gws"] = true;
 290       break;
 291     case GK_GFX1036:
 292     case GK_GFX1035:
 293     case GK_GFX1034:
 294     case GK_GFX1033:
 295     case GK_GFX1032:
 296     case GK_GFX1031:
 297     case GK_GFX1030:
 298       Features["ci-insts"] = true;
 299       Features["dot1-insts"] = true;
 300       Features["dot2-insts"] = true;
 301       Features["dot5-insts"] = true;
 302       Features["dot6-insts"] = true;
 303       Features["dot7-insts"] = true;
 304       Features["dot10-insts"] = true;
 305       Features["dl-insts"] = true;
 306       Features["16-bit-insts"] = true;
 307       Features["dpp"] = true;
 308       Features["gfx8-insts"] = true;
 309       Features["gfx9-insts"] = true;
 310       Features["gfx10-insts"] = true;
 311       Features["gfx10-3-insts"] = true;
 312       Features["image-insts"] = true;
 313       Features["s-memrealtime"] = true;
 314       Features["s-memtime-inst"] = true;
 315       Features["gws"] = true;
 316       break;
 317     case GK_GFX1012:
 318     case GK_GFX1011:
 319       Features["dot1-insts"] = true;
 320       Features["dot2-insts"] = true;
 321       Features["dot5-insts"] = true;
 322       Features["dot6-insts"] = true;
 323       Features["dot7-insts"] = true;
 324       Features["dot10-insts"] = true;
 325       [[fallthrough]];
 326     case GK_GFX1013:
 327     case GK_GFX1010:
 328       Features["dl-insts"] = true;
 329       Features["ci-insts"] = true;
 330       Features["16-bit-insts"] = true;
 331       Features["dpp"] = true;
 332       Features["gfx8-insts"] = true;
 333       Features["gfx9-insts"] = true;
 334       Features["gfx10-insts"] = true;
 335       Features["image-insts"] = true;
 336       Features["s-memrealtime"] = true;
 337       Features["s-memtime-inst"] = true;
 338       Features["gws"] = true;
 339       break;
 340     case GK_GFX942:
 341     case GK_GFX941:
 342     case GK_GFX940:
 343       Features["gfx940-insts"] = true;
 344       Features["fp8-insts"] = true;
 345       Features["atomic-ds-pk-add-16-insts"] = true;
 346       Features["atomic-flat-pk-add-16-insts"] = true;
 347       Features["atomic-global-pk-add-bf16-inst"] = true;
 348       Features["gfx90a-insts"] = true;
 349       Features["atomic-buffer-global-pk-add-f16-insts"] = true;
 350       Features["atomic-fadd-rtn-insts"] = true;
 351       Features["dot3-insts"] = true;
 352       Features["dot4-insts"] = true;
 353       Features["dot5-insts"] = true;
 354       Features["dot6-insts"] = true;
 355       Features["mai-insts"] = true;
 356       Features["dl-insts"] = true;
 357       Features["dot1-insts"] = true;
 358       Features["dot2-insts"] = true;
 359       Features["dot7-insts"] = true;
 360       Features["dot10-insts"] = true;
 361       Features["gfx9-insts"] = true;
 362       Features["gfx8-insts"] = true;
 363       Features["16-bit-insts"] = true;
 364       Features["dpp"] = true;
 365       Features["s-memrealtime"] = true;
 366       Features["ci-insts"] = true;
 367       Features["s-memtime-inst"] = true;
 368       Features["gws"] = true;
 369       break;
 370     case GK_GFX90A:
 371       Features["gfx90a-insts"] = true;
 372       Features["atomic-buffer-global-pk-add-f16-insts"] = true;
 373       Features["atomic-fadd-rtn-insts"] = true;
 374       [[fallthrough]];
 375     case GK_GFX908:
 376       Features["dot3-insts"] = true;
 377       Features["dot4-insts"] = true;
 378       Features["dot5-insts"] = true;
 379       Features["dot6-insts"] = true;
 380       Features["mai-insts"] = true;
 381       [[fallthrough]];
 382     case GK_GFX906:
 383       Features["dl-insts"] = true;
 384       Features["dot1-insts"] = true;
 385       Features["dot2-insts"] = true;
 386       Features["dot7-insts"] = true;
 387       Features["dot10-insts"] = true;
 388       [[fallthrough]];
 389     case GK_GFX90C:
 390     case GK_GFX909:
 391     case GK_GFX904:
 392     case GK_GFX902:
 393     case GK_GFX900:
 394       Features["gfx9-insts"] = true;
 395       [[fallthrough]];
 396     case GK_GFX810:
 397     case GK_GFX805:
 398     case GK_GFX803:
 399     case GK_GFX802:
 400     case GK_GFX801:
 401       Features["gfx8-insts"] = true;
 402       Features["16-bit-insts"] = true;
 403       Features["dpp"] = true;
 404       Features["s-memrealtime"] = true;
 405       [[fallthrough]];
 406     case GK_GFX705:
 407     case GK_GFX704:
 408     case GK_GFX703:
 409     case GK_GFX702:
 410     case GK_GFX701:
 411     case GK_GFX700:
 412       Features["ci-insts"] = true;
 413       [[fallthrough]];
 414     case GK_GFX602:
 415     case GK_GFX601:
 416     case GK_GFX600:
 417       Features["image-insts"] = true;
 418       Features["s-memtime-inst"] = true;
 419       Features["gws"] = true;
 420       break;
 421     case GK_NONE:
 422       break;
 423     default:
 424       llvm_unreachable("Unhandled GPU!");
 425     }
 426   } else {
 427     if (GPU.empty())
 428       GPU = "r600";
 429
 430     switch (llvm::AMDGPU::parseArchR600(GPU)) {
 431     case GK_CAYMAN:
 432     case GK_CYPRESS:
 433     case GK_RV770:
 434     case GK_RV670:
 435       // TODO: Add fp64 when implemented.
 436       break;
 437     case GK_TURKS:
 438     case GK_CAICOS:
 439     case GK_BARTS:
 440     case GK_SUMO:
 441     case GK_REDWOOD:
 442     case GK_JUNIPER:
 443     case GK_CEDAR:
 444     case GK_RV730:
 445     case GK_RV710:
 446     case GK_RS880:
 447     case GK_R630:
 448     case GK_R600:
 449       break;
 450     default:
 451       llvm_unreachable("Unhandled GPU!");
 452     }
 453   }
 454 }
 455
 456 static bool isWave32Capable(StringRef GPU, const Triple &T) {
 457   bool IsWave32Capable = false;
 458   // XXX - What does the member GPU mean if device name string passed here?
 459   if (T.isAMDGCN()) {
 460     switch (parseArchAMDGCN(GPU)) {
 461     case GK_GFX1151:
 462     case GK_GFX1150:
 463     case GK_GFX1103:
 464     case GK_GFX1102:
 465     case GK_GFX1101:
 466     case GK_GFX1100:
 467     case GK_GFX1036:
 468     case GK_GFX1035:
 469     case GK_GFX1034:
 470     case GK_GFX1033:
 471     case GK_GFX1032:
 472     case GK_GFX1031:
 473     case GK_GFX1030:
 474     case GK_GFX1012:
 475     case GK_GFX1011:
 476     case GK_GFX1013:
 477     case GK_GFX1010:
 478       IsWave32Capable = true;
 479       break;
 480     default:
 481       break;
 482     }
 483   }
 484   return IsWave32Capable;
 485 }
 486
 487 bool AMDGPU::insertWaveSizeFeature(StringRef GPU, const Triple &T,
 488                                    StringMap<bool> &Features,
 489                                    std::string &ErrorMsg) {
 490   bool IsWave32Capable = isWave32Capable(GPU, T);
 491   const bool IsNullGPU = GPU.empty();
 492   // FIXME: Not diagnosing wavefrontsize32 on wave64 only targets.
 493   const bool HaveWave32 =
 494       (IsWave32Capable || IsNullGPU) && Features.count("wavefrontsize32");
 495   const bool HaveWave64 = Features.count("wavefrontsize64");
 496   if (HaveWave32 && HaveWave64) {
 497     ErrorMsg = "'wavefrontsize32' and 'wavefrontsize64' are mutually exclusive";
 498     return false;
 499   }
 500   // Don't assume any wavesize with an unknown subtarget.
 501   if (!IsNullGPU) {
 502     // Default to wave32 if available, or wave64 if not
 503     if (!HaveWave32 && !HaveWave64) {
 504       StringRef DefaultWaveSizeFeature =
 505           IsWave32Capable ? "wavefrontsize32" : "wavefrontsize64";
 506       Features.insert(std::make_pair(DefaultWaveSizeFeature, true));
 507     }
 508   }
 509   return true;
 510 }