1 //===--- AMDGPU.cpp - Implement AMDGPU 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 AMDGPU TargetInfo objects. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "AMDGPU.h" 14 #include "clang/Basic/Builtins.h" 15 #include "clang/Basic/CodeGenOptions.h" 16 #include "clang/Basic/LangOptions.h" 17 #include "clang/Basic/MacroBuilder.h" 18 #include "clang/Basic/TargetBuiltins.h" 19 #include "llvm/ADT/StringSwitch.h" 20 21 using namespace clang; 22 using namespace clang::targets; 23 24 namespace clang { 25 namespace targets { 26 27 // If you edit the description strings, make sure you update 28 // getPointerWidthV(). 29 30 static const char *const DataLayoutStringR600 = 31 "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128" 32 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1"; 33 34 static const char *const DataLayoutStringAMDGCN = 35 "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32" 36 "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128" 37 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1" 38 "-ni:7"; 39 40 const LangASMap AMDGPUTargetInfo::AMDGPUDefIsGenMap = { 41 Generic, // Default 42 Global, // opencl_global 43 Local, // opencl_local 44 Constant, // opencl_constant 45 Private, // opencl_private 46 Generic, // opencl_generic 47 Global, // opencl_global_device 48 Global, // opencl_global_host 49 Global, // cuda_device 50 Constant, // cuda_constant 51 Local, // cuda_shared 52 Global, // sycl_global 53 Global, // sycl_global_device 54 Global, // sycl_global_host 55 Local, // sycl_local 56 Private, // sycl_private 57 Generic, // ptr32_sptr 58 Generic, // ptr32_uptr 59 Generic // ptr64 60 }; 61 62 const LangASMap AMDGPUTargetInfo::AMDGPUDefIsPrivMap = { 63 Private, // Default 64 Global, // opencl_global 65 Local, // opencl_local 66 Constant, // opencl_constant 67 Private, // opencl_private 68 Generic, // opencl_generic 69 Global, // opencl_global_device 70 Global, // opencl_global_host 71 Global, // cuda_device 72 Constant, // cuda_constant 73 Local, // cuda_shared 74 // SYCL address space values for this map are dummy 75 Generic, // sycl_global 76 Generic, // sycl_global_device 77 Generic, // sycl_global_host 78 Generic, // sycl_local 79 Generic, // sycl_private 80 Generic, // ptr32_sptr 81 Generic, // ptr32_uptr 82 Generic // ptr64 83 84 }; 85 } // namespace targets 86 } // namespace clang 87 88 const Builtin::Info AMDGPUTargetInfo::BuiltinInfo[] = { 89 #define BUILTIN(ID, TYPE, ATTRS) \ 90 {#ID, TYPE, ATTRS, nullptr, ALL_LANGUAGES, nullptr}, 91 #define TARGET_BUILTIN(ID, TYPE, ATTRS, FEATURE) \ 92 {#ID, TYPE, ATTRS, nullptr, ALL_LANGUAGES, FEATURE}, 93 #include "clang/Basic/BuiltinsAMDGPU.def" 94 }; 95 96 const char *const AMDGPUTargetInfo::GCCRegNames[] = { 97 "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", 98 "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", 99 "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", 100 "v27", "v28", "v29", "v30", "v31", "v32", "v33", "v34", "v35", 101 "v36", "v37", "v38", "v39", "v40", "v41", "v42", "v43", "v44", 102 "v45", "v46", "v47", "v48", "v49", "v50", "v51", "v52", "v53", 103 "v54", "v55", "v56", "v57", "v58", "v59", "v60", "v61", "v62", 104 "v63", "v64", "v65", "v66", "v67", "v68", "v69", "v70", "v71", 105 "v72", "v73", "v74", "v75", "v76", "v77", "v78", "v79", "v80", 106 "v81", "v82", "v83", "v84", "v85", "v86", "v87", "v88", "v89", 107 "v90", "v91", "v92", "v93", "v94", "v95", "v96", "v97", "v98", 108 "v99", "v100", "v101", "v102", "v103", "v104", "v105", "v106", "v107", 109 "v108", "v109", "v110", "v111", "v112", "v113", "v114", "v115", "v116", 110 "v117", "v118", "v119", "v120", "v121", "v122", "v123", "v124", "v125", 111 "v126", "v127", "v128", "v129", "v130", "v131", "v132", "v133", "v134", 112 "v135", "v136", "v137", "v138", "v139", "v140", "v141", "v142", "v143", 113 "v144", "v145", "v146", "v147", "v148", "v149", "v150", "v151", "v152", 114 "v153", "v154", "v155", "v156", "v157", "v158", "v159", "v160", "v161", 115 "v162", "v163", "v164", "v165", "v166", "v167", "v168", "v169", "v170", 116 "v171", "v172", "v173", "v174", "v175", "v176", "v177", "v178", "v179", 117 "v180", "v181", "v182", "v183", "v184", "v185", "v186", "v187", "v188", 118 "v189", "v190", "v191", "v192", "v193", "v194", "v195", "v196", "v197", 119 "v198", "v199", "v200", "v201", "v202", "v203", "v204", "v205", "v206", 120 "v207", "v208", "v209", "v210", "v211", "v212", "v213", "v214", "v215", 121 "v216", "v217", "v218", "v219", "v220", "v221", "v222", "v223", "v224", 122 "v225", "v226", "v227", "v228", "v229", "v230", "v231", "v232", "v233", 123 "v234", "v235", "v236", "v237", "v238", "v239", "v240", "v241", "v242", 124 "v243", "v244", "v245", "v246", "v247", "v248", "v249", "v250", "v251", 125 "v252", "v253", "v254", "v255", "s0", "s1", "s2", "s3", "s4", 126 "s5", "s6", "s7", "s8", "s9", "s10", "s11", "s12", "s13", 127 "s14", "s15", "s16", "s17", "s18", "s19", "s20", "s21", "s22", 128 "s23", "s24", "s25", "s26", "s27", "s28", "s29", "s30", "s31", 129 "s32", "s33", "s34", "s35", "s36", "s37", "s38", "s39", "s40", 130 "s41", "s42", "s43", "s44", "s45", "s46", "s47", "s48", "s49", 131 "s50", "s51", "s52", "s53", "s54", "s55", "s56", "s57", "s58", 132 "s59", "s60", "s61", "s62", "s63", "s64", "s65", "s66", "s67", 133 "s68", "s69", "s70", "s71", "s72", "s73", "s74", "s75", "s76", 134 "s77", "s78", "s79", "s80", "s81", "s82", "s83", "s84", "s85", 135 "s86", "s87", "s88", "s89", "s90", "s91", "s92", "s93", "s94", 136 "s95", "s96", "s97", "s98", "s99", "s100", "s101", "s102", "s103", 137 "s104", "s105", "s106", "s107", "s108", "s109", "s110", "s111", "s112", 138 "s113", "s114", "s115", "s116", "s117", "s118", "s119", "s120", "s121", 139 "s122", "s123", "s124", "s125", "s126", "s127", "exec", "vcc", "scc", 140 "m0", "flat_scratch", "exec_lo", "exec_hi", "vcc_lo", "vcc_hi", 141 "flat_scratch_lo", "flat_scratch_hi", 142 "a0", "a1", "a2", "a3", "a4", "a5", "a6", "a7", "a8", 143 "a9", "a10", "a11", "a12", "a13", "a14", "a15", "a16", "a17", 144 "a18", "a19", "a20", "a21", "a22", "a23", "a24", "a25", "a26", 145 "a27", "a28", "a29", "a30", "a31", "a32", "a33", "a34", "a35", 146 "a36", "a37", "a38", "a39", "a40", "a41", "a42", "a43", "a44", 147 "a45", "a46", "a47", "a48", "a49", "a50", "a51", "a52", "a53", 148 "a54", "a55", "a56", "a57", "a58", "a59", "a60", "a61", "a62", 149 "a63", "a64", "a65", "a66", "a67", "a68", "a69", "a70", "a71", 150 "a72", "a73", "a74", "a75", "a76", "a77", "a78", "a79", "a80", 151 "a81", "a82", "a83", "a84", "a85", "a86", "a87", "a88", "a89", 152 "a90", "a91", "a92", "a93", "a94", "a95", "a96", "a97", "a98", 153 "a99", "a100", "a101", "a102", "a103", "a104", "a105", "a106", "a107", 154 "a108", "a109", "a110", "a111", "a112", "a113", "a114", "a115", "a116", 155 "a117", "a118", "a119", "a120", "a121", "a122", "a123", "a124", "a125", 156 "a126", "a127", "a128", "a129", "a130", "a131", "a132", "a133", "a134", 157 "a135", "a136", "a137", "a138", "a139", "a140", "a141", "a142", "a143", 158 "a144", "a145", "a146", "a147", "a148", "a149", "a150", "a151", "a152", 159 "a153", "a154", "a155", "a156", "a157", "a158", "a159", "a160", "a161", 160 "a162", "a163", "a164", "a165", "a166", "a167", "a168", "a169", "a170", 161 "a171", "a172", "a173", "a174", "a175", "a176", "a177", "a178", "a179", 162 "a180", "a181", "a182", "a183", "a184", "a185", "a186", "a187", "a188", 163 "a189", "a190", "a191", "a192", "a193", "a194", "a195", "a196", "a197", 164 "a198", "a199", "a200", "a201", "a202", "a203", "a204", "a205", "a206", 165 "a207", "a208", "a209", "a210", "a211", "a212", "a213", "a214", "a215", 166 "a216", "a217", "a218", "a219", "a220", "a221", "a222", "a223", "a224", 167 "a225", "a226", "a227", "a228", "a229", "a230", "a231", "a232", "a233", 168 "a234", "a235", "a236", "a237", "a238", "a239", "a240", "a241", "a242", 169 "a243", "a244", "a245", "a246", "a247", "a248", "a249", "a250", "a251", 170 "a252", "a253", "a254", "a255" 171 }; 172 173 ArrayRef<const char *> AMDGPUTargetInfo::getGCCRegNames() const { 174 return llvm::makeArrayRef(GCCRegNames); 175 } 176 177 bool AMDGPUTargetInfo::initFeatureMap( 178 llvm::StringMap<bool> &Features, DiagnosticsEngine &Diags, StringRef CPU, 179 const std::vector<std::string> &FeatureVec) const { 180 181 using namespace llvm::AMDGPU; 182 183 // XXX - What does the member GPU mean if device name string passed here? 184 if (isAMDGCN(getTriple())) { 185 switch (llvm::AMDGPU::parseArchAMDGCN(CPU)) { 186 case GK_GFX1035: 187 case GK_GFX1034: 188 case GK_GFX1033: 189 case GK_GFX1032: 190 case GK_GFX1031: 191 case GK_GFX1030: 192 Features["ci-insts"] = true; 193 Features["dot1-insts"] = true; 194 Features["dot2-insts"] = true; 195 Features["dot5-insts"] = true; 196 Features["dot6-insts"] = true; 197 Features["dot7-insts"] = true; 198 Features["dl-insts"] = true; 199 Features["flat-address-space"] = true; 200 Features["16-bit-insts"] = true; 201 Features["dpp"] = true; 202 Features["gfx8-insts"] = true; 203 Features["gfx9-insts"] = true; 204 Features["gfx10-insts"] = true; 205 Features["gfx10-3-insts"] = true; 206 Features["s-memrealtime"] = true; 207 Features["s-memtime-inst"] = true; 208 break; 209 case GK_GFX1012: 210 case GK_GFX1011: 211 Features["dot1-insts"] = true; 212 Features["dot2-insts"] = true; 213 Features["dot5-insts"] = true; 214 Features["dot6-insts"] = true; 215 Features["dot7-insts"] = true; 216 LLVM_FALLTHROUGH; 217 case GK_GFX1013: 218 case GK_GFX1010: 219 Features["dl-insts"] = true; 220 Features["ci-insts"] = true; 221 Features["flat-address-space"] = true; 222 Features["16-bit-insts"] = true; 223 Features["dpp"] = true; 224 Features["gfx8-insts"] = true; 225 Features["gfx9-insts"] = true; 226 Features["gfx10-insts"] = true; 227 Features["s-memrealtime"] = true; 228 Features["s-memtime-inst"] = true; 229 break; 230 case GK_GFX90A: 231 Features["gfx90a-insts"] = true; 232 LLVM_FALLTHROUGH; 233 case GK_GFX908: 234 Features["dot3-insts"] = true; 235 Features["dot4-insts"] = true; 236 Features["dot5-insts"] = true; 237 Features["dot6-insts"] = true; 238 Features["mai-insts"] = true; 239 LLVM_FALLTHROUGH; 240 case GK_GFX906: 241 Features["dl-insts"] = true; 242 Features["dot1-insts"] = true; 243 Features["dot2-insts"] = true; 244 Features["dot7-insts"] = true; 245 LLVM_FALLTHROUGH; 246 case GK_GFX90C: 247 case GK_GFX909: 248 case GK_GFX904: 249 case GK_GFX902: 250 case GK_GFX900: 251 Features["gfx9-insts"] = true; 252 LLVM_FALLTHROUGH; 253 case GK_GFX810: 254 case GK_GFX805: 255 case GK_GFX803: 256 case GK_GFX802: 257 case GK_GFX801: 258 Features["gfx8-insts"] = true; 259 Features["16-bit-insts"] = true; 260 Features["dpp"] = true; 261 Features["s-memrealtime"] = true; 262 LLVM_FALLTHROUGH; 263 case GK_GFX705: 264 case GK_GFX704: 265 case GK_GFX703: 266 case GK_GFX702: 267 case GK_GFX701: 268 case GK_GFX700: 269 Features["ci-insts"] = true; 270 Features["flat-address-space"] = true; 271 LLVM_FALLTHROUGH; 272 case GK_GFX602: 273 case GK_GFX601: 274 case GK_GFX600: 275 Features["s-memtime-inst"] = true; 276 break; 277 case GK_NONE: 278 break; 279 default: 280 llvm_unreachable("Unhandled GPU!"); 281 } 282 } else { 283 if (CPU.empty()) 284 CPU = "r600"; 285 286 switch (llvm::AMDGPU::parseArchR600(CPU)) { 287 case GK_CAYMAN: 288 case GK_CYPRESS: 289 case GK_RV770: 290 case GK_RV670: 291 // TODO: Add fp64 when implemented. 292 break; 293 case GK_TURKS: 294 case GK_CAICOS: 295 case GK_BARTS: 296 case GK_SUMO: 297 case GK_REDWOOD: 298 case GK_JUNIPER: 299 case GK_CEDAR: 300 case GK_RV730: 301 case GK_RV710: 302 case GK_RS880: 303 case GK_R630: 304 case GK_R600: 305 break; 306 default: 307 llvm_unreachable("Unhandled GPU!"); 308 } 309 } 310 311 return TargetInfo::initFeatureMap(Features, Diags, CPU, FeatureVec); 312 } 313 314 void AMDGPUTargetInfo::fillValidCPUList( 315 SmallVectorImpl<StringRef> &Values) const { 316 if (isAMDGCN(getTriple())) 317 llvm::AMDGPU::fillValidArchListAMDGCN(Values); 318 else 319 llvm::AMDGPU::fillValidArchListR600(Values); 320 } 321 322 void AMDGPUTargetInfo::setAddressSpaceMap(bool DefaultIsPrivate) { 323 AddrSpaceMap = DefaultIsPrivate ? &AMDGPUDefIsPrivMap : &AMDGPUDefIsGenMap; 324 } 325 326 AMDGPUTargetInfo::AMDGPUTargetInfo(const llvm::Triple &Triple, 327 const TargetOptions &Opts) 328 : TargetInfo(Triple), 329 GPUKind(isAMDGCN(Triple) ? 330 llvm::AMDGPU::parseArchAMDGCN(Opts.CPU) : 331 llvm::AMDGPU::parseArchR600(Opts.CPU)), 332 GPUFeatures(isAMDGCN(Triple) ? 333 llvm::AMDGPU::getArchAttrAMDGCN(GPUKind) : 334 llvm::AMDGPU::getArchAttrR600(GPUKind)) { 335 resetDataLayout(isAMDGCN(getTriple()) ? DataLayoutStringAMDGCN 336 : DataLayoutStringR600); 337 338 setAddressSpaceMap(Triple.getOS() == llvm::Triple::Mesa3D || 339 !isAMDGCN(Triple)); 340 UseAddrSpaceMapMangling = true; 341 342 HasLegalHalfType = true; 343 HasFloat16 = true; 344 WavefrontSize = GPUFeatures & llvm::AMDGPU::FEATURE_WAVE32 ? 32 : 64; 345 AllowAMDGPUUnsafeFPAtomics = Opts.AllowAMDGPUUnsafeFPAtomics; 346 347 // Set pointer width and alignment for target address space 0. 348 PointerWidth = PointerAlign = getPointerWidthV(Generic); 349 if (getMaxPointerWidth() == 64) { 350 LongWidth = LongAlign = 64; 351 SizeType = UnsignedLong; 352 PtrDiffType = SignedLong; 353 IntPtrType = SignedLong; 354 } 355 356 MaxAtomicPromoteWidth = MaxAtomicInlineWidth = 64; 357 } 358 359 void AMDGPUTargetInfo::adjust(DiagnosticsEngine &Diags, LangOptions &Opts) { 360 TargetInfo::adjust(Diags, Opts); 361 // ToDo: There are still a few places using default address space as private 362 // address space in OpenCL, which needs to be cleaned up, then Opts.OpenCL 363 // can be removed from the following line. 364 setAddressSpaceMap(/*DefaultIsPrivate=*/Opts.OpenCL || 365 !isAMDGCN(getTriple())); 366 } 367 368 ArrayRef<Builtin::Info> AMDGPUTargetInfo::getTargetBuiltins() const { 369 return llvm::makeArrayRef(BuiltinInfo, clang::AMDGPU::LastTSBuiltin - 370 Builtin::FirstTSBuiltin); 371 } 372 373 void AMDGPUTargetInfo::getTargetDefines(const LangOptions &Opts, 374 MacroBuilder &Builder) const { 375 Builder.defineMacro("__AMD__"); 376 Builder.defineMacro("__AMDGPU__"); 377 378 if (isAMDGCN(getTriple())) 379 Builder.defineMacro("__AMDGCN__"); 380 else 381 Builder.defineMacro("__R600__"); 382 383 if (GPUKind != llvm::AMDGPU::GK_NONE) { 384 StringRef CanonName = isAMDGCN(getTriple()) ? 385 getArchNameAMDGCN(GPUKind) : getArchNameR600(GPUKind); 386 Builder.defineMacro(Twine("__") + Twine(CanonName) + Twine("__")); 387 if (isAMDGCN(getTriple())) { 388 Builder.defineMacro("__amdgcn_processor__", 389 Twine("\"") + Twine(CanonName) + Twine("\"")); 390 Builder.defineMacro("__amdgcn_target_id__", 391 Twine("\"") + Twine(getTargetID().getValue()) + 392 Twine("\"")); 393 for (auto F : getAllPossibleTargetIDFeatures(getTriple(), CanonName)) { 394 auto Loc = OffloadArchFeatures.find(F); 395 if (Loc != OffloadArchFeatures.end()) { 396 std::string NewF = F.str(); 397 std::replace(NewF.begin(), NewF.end(), '-', '_'); 398 Builder.defineMacro(Twine("__amdgcn_feature_") + Twine(NewF) + 399 Twine("__"), 400 Loc->second ? "1" : "0"); 401 } 402 } 403 } 404 } 405 406 // TODO: __HAS_FMAF__, __HAS_LDEXPF__, __HAS_FP64__ are deprecated and will be 407 // removed in the near future. 408 if (hasFMAF()) 409 Builder.defineMacro("__HAS_FMAF__"); 410 if (hasFastFMAF()) 411 Builder.defineMacro("FP_FAST_FMAF"); 412 if (hasLDEXPF()) 413 Builder.defineMacro("__HAS_LDEXPF__"); 414 if (hasFP64()) 415 Builder.defineMacro("__HAS_FP64__"); 416 if (hasFastFMA()) 417 Builder.defineMacro("FP_FAST_FMA"); 418 419 Builder.defineMacro("__AMDGCN_WAVEFRONT_SIZE", Twine(WavefrontSize)); 420 } 421 422 void AMDGPUTargetInfo::setAuxTarget(const TargetInfo *Aux) { 423 assert(HalfFormat == Aux->HalfFormat); 424 assert(FloatFormat == Aux->FloatFormat); 425 assert(DoubleFormat == Aux->DoubleFormat); 426 427 // On x86_64 long double is 80-bit extended precision format, which is 428 // not supported by AMDGPU. 128-bit floating point format is also not 429 // supported by AMDGPU. Therefore keep its own format for these two types. 430 auto SaveLongDoubleFormat = LongDoubleFormat; 431 auto SaveFloat128Format = Float128Format; 432 copyAuxTarget(Aux); 433 LongDoubleFormat = SaveLongDoubleFormat; 434 Float128Format = SaveFloat128Format; 435 // For certain builtin types support on the host target, claim they are 436 // support to pass the compilation of the host code during the device-side 437 // compilation. 438 // FIXME: As the side effect, we also accept `__float128` uses in the device 439 // code. To rejct these builtin types supported in the host target but not in 440 // the device target, one approach would support `device_builtin` attribute 441 // so that we could tell the device builtin types from the host ones. The 442 // also solves the different representations of the same builtin type, such 443 // as `size_t` in the MSVC environment. 444 if (Aux->hasFloat128Type()) { 445 HasFloat128 = true; 446 Float128Format = DoubleFormat; 447 } 448 } 449