1 //===-- AMDGPUTargetStreamer.cpp - Mips Target Streamer Methods -----------===// 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 provides AMDGPU specific target streamer methods. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "AMDGPUTargetStreamer.h" 14 #include "AMDGPUPTNote.h" 15 #include "AMDKernelCodeT.h" 16 #include "Utils/AMDGPUBaseInfo.h" 17 #include "Utils/AMDKernelCodeTUtils.h" 18 #include "llvm/BinaryFormat/AMDGPUMetadataVerifier.h" 19 #include "llvm/BinaryFormat/ELF.h" 20 #include "llvm/MC/MCContext.h" 21 #include "llvm/MC/MCELFStreamer.h" 22 #include "llvm/MC/MCSectionELF.h" 23 #include "llvm/Support/AMDGPUMetadata.h" 24 #include "llvm/Support/AMDHSAKernelDescriptor.h" 25 #include "llvm/Support/FormattedStream.h" 26 27 using namespace llvm; 28 using namespace llvm::AMDGPU; 29 30 //===----------------------------------------------------------------------===// 31 // AMDGPUTargetStreamer 32 //===----------------------------------------------------------------------===// 33 34 static void convertIsaVersionV2(uint32_t &Major, uint32_t &Minor, 35 uint32_t &Stepping, bool Sramecc, bool Xnack) { 36 if (Major == 9 && Minor == 0) { 37 switch (Stepping) { 38 case 0: 39 case 2: 40 case 4: 41 case 6: 42 if (Xnack) 43 Stepping++; 44 } 45 } 46 } 47 48 bool AMDGPUTargetStreamer::EmitHSAMetadataV2(StringRef HSAMetadataString) { 49 HSAMD::Metadata HSAMetadata; 50 if (HSAMD::fromString(HSAMetadataString, HSAMetadata)) 51 return false; 52 return EmitHSAMetadata(HSAMetadata); 53 } 54 55 bool AMDGPUTargetStreamer::EmitHSAMetadataV3(StringRef HSAMetadataString) { 56 msgpack::Document HSAMetadataDoc; 57 if (!HSAMetadataDoc.fromYAML(HSAMetadataString)) 58 return false; 59 return EmitHSAMetadata(HSAMetadataDoc, false); 60 } 61 62 StringRef AMDGPUTargetStreamer::getArchNameFromElfMach(unsigned ElfMach) { 63 AMDGPU::GPUKind AK; 64 65 switch (ElfMach) { 66 default: llvm_unreachable("Unhandled ELF::EF_AMDGPU type"); 67 case ELF::EF_AMDGPU_MACH_R600_R600: AK = GK_R600; break; 68 case ELF::EF_AMDGPU_MACH_R600_R630: AK = GK_R630; break; 69 case ELF::EF_AMDGPU_MACH_R600_RS880: AK = GK_RS880; break; 70 case ELF::EF_AMDGPU_MACH_R600_RV670: AK = GK_RV670; break; 71 case ELF::EF_AMDGPU_MACH_R600_RV710: AK = GK_RV710; break; 72 case ELF::EF_AMDGPU_MACH_R600_RV730: AK = GK_RV730; break; 73 case ELF::EF_AMDGPU_MACH_R600_RV770: AK = GK_RV770; break; 74 case ELF::EF_AMDGPU_MACH_R600_CEDAR: AK = GK_CEDAR; break; 75 case ELF::EF_AMDGPU_MACH_R600_CYPRESS: AK = GK_CYPRESS; break; 76 case ELF::EF_AMDGPU_MACH_R600_JUNIPER: AK = GK_JUNIPER; break; 77 case ELF::EF_AMDGPU_MACH_R600_REDWOOD: AK = GK_REDWOOD; break; 78 case ELF::EF_AMDGPU_MACH_R600_SUMO: AK = GK_SUMO; break; 79 case ELF::EF_AMDGPU_MACH_R600_BARTS: AK = GK_BARTS; break; 80 case ELF::EF_AMDGPU_MACH_R600_CAICOS: AK = GK_CAICOS; break; 81 case ELF::EF_AMDGPU_MACH_R600_CAYMAN: AK = GK_CAYMAN; break; 82 case ELF::EF_AMDGPU_MACH_R600_TURKS: AK = GK_TURKS; break; 83 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX600: AK = GK_GFX600; break; 84 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX601: AK = GK_GFX601; break; 85 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX602: AK = GK_GFX602; break; 86 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX700: AK = GK_GFX700; break; 87 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX701: AK = GK_GFX701; break; 88 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX702: AK = GK_GFX702; break; 89 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX703: AK = GK_GFX703; break; 90 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX704: AK = GK_GFX704; break; 91 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX705: AK = GK_GFX705; break; 92 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX801: AK = GK_GFX801; break; 93 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX802: AK = GK_GFX802; break; 94 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX803: AK = GK_GFX803; break; 95 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX805: AK = GK_GFX805; break; 96 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX810: AK = GK_GFX810; break; 97 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX900: AK = GK_GFX900; break; 98 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX902: AK = GK_GFX902; break; 99 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX904: AK = GK_GFX904; break; 100 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX906: AK = GK_GFX906; break; 101 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX908: AK = GK_GFX908; break; 102 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX909: AK = GK_GFX909; break; 103 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX90A: AK = GK_GFX90A; break; 104 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX90C: AK = GK_GFX90C; break; 105 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1010: AK = GK_GFX1010; break; 106 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1011: AK = GK_GFX1011; break; 107 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1012: AK = GK_GFX1012; break; 108 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1030: AK = GK_GFX1030; break; 109 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1031: AK = GK_GFX1031; break; 110 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1032: AK = GK_GFX1032; break; 111 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1033: AK = GK_GFX1033; break; 112 case ELF::EF_AMDGPU_MACH_AMDGCN_GFX1034: AK = GK_GFX1034; break; 113 case ELF::EF_AMDGPU_MACH_NONE: AK = GK_NONE; break; 114 } 115 116 StringRef GPUName = getArchNameAMDGCN(AK); 117 if (GPUName != "") 118 return GPUName; 119 return getArchNameR600(AK); 120 } 121 122 unsigned AMDGPUTargetStreamer::getElfMach(StringRef GPU) { 123 AMDGPU::GPUKind AK = parseArchAMDGCN(GPU); 124 if (AK == AMDGPU::GPUKind::GK_NONE) 125 AK = parseArchR600(GPU); 126 127 switch (AK) { 128 case GK_R600: return ELF::EF_AMDGPU_MACH_R600_R600; 129 case GK_R630: return ELF::EF_AMDGPU_MACH_R600_R630; 130 case GK_RS880: return ELF::EF_AMDGPU_MACH_R600_RS880; 131 case GK_RV670: return ELF::EF_AMDGPU_MACH_R600_RV670; 132 case GK_RV710: return ELF::EF_AMDGPU_MACH_R600_RV710; 133 case GK_RV730: return ELF::EF_AMDGPU_MACH_R600_RV730; 134 case GK_RV770: return ELF::EF_AMDGPU_MACH_R600_RV770; 135 case GK_CEDAR: return ELF::EF_AMDGPU_MACH_R600_CEDAR; 136 case GK_CYPRESS: return ELF::EF_AMDGPU_MACH_R600_CYPRESS; 137 case GK_JUNIPER: return ELF::EF_AMDGPU_MACH_R600_JUNIPER; 138 case GK_REDWOOD: return ELF::EF_AMDGPU_MACH_R600_REDWOOD; 139 case GK_SUMO: return ELF::EF_AMDGPU_MACH_R600_SUMO; 140 case GK_BARTS: return ELF::EF_AMDGPU_MACH_R600_BARTS; 141 case GK_CAICOS: return ELF::EF_AMDGPU_MACH_R600_CAICOS; 142 case GK_CAYMAN: return ELF::EF_AMDGPU_MACH_R600_CAYMAN; 143 case GK_TURKS: return ELF::EF_AMDGPU_MACH_R600_TURKS; 144 case GK_GFX600: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX600; 145 case GK_GFX601: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX601; 146 case GK_GFX602: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX602; 147 case GK_GFX700: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX700; 148 case GK_GFX701: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX701; 149 case GK_GFX702: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX702; 150 case GK_GFX703: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX703; 151 case GK_GFX704: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX704; 152 case GK_GFX705: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX705; 153 case GK_GFX801: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX801; 154 case GK_GFX802: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX802; 155 case GK_GFX803: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX803; 156 case GK_GFX805: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX805; 157 case GK_GFX810: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX810; 158 case GK_GFX900: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX900; 159 case GK_GFX902: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX902; 160 case GK_GFX904: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX904; 161 case GK_GFX906: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX906; 162 case GK_GFX908: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX908; 163 case GK_GFX909: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX909; 164 case GK_GFX90A: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX90A; 165 case GK_GFX90C: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX90C; 166 case GK_GFX1010: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1010; 167 case GK_GFX1011: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1011; 168 case GK_GFX1012: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1012; 169 case GK_GFX1030: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1030; 170 case GK_GFX1031: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1031; 171 case GK_GFX1032: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1032; 172 case GK_GFX1033: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1033; 173 case GK_GFX1034: return ELF::EF_AMDGPU_MACH_AMDGCN_GFX1034; 174 case GK_NONE: return ELF::EF_AMDGPU_MACH_NONE; 175 } 176 177 llvm_unreachable("unknown GPU"); 178 } 179 180 //===----------------------------------------------------------------------===// 181 // AMDGPUTargetAsmStreamer 182 //===----------------------------------------------------------------------===// 183 184 AMDGPUTargetAsmStreamer::AMDGPUTargetAsmStreamer(MCStreamer &S, 185 formatted_raw_ostream &OS) 186 : AMDGPUTargetStreamer(S), OS(OS) { } 187 188 // A hook for emitting stuff at the end. 189 // We use it for emitting the accumulated PAL metadata as directives. 190 // The PAL metadata is reset after it is emitted. 191 void AMDGPUTargetAsmStreamer::finish() { 192 std::string S; 193 getPALMetadata()->toString(S); 194 OS << S; 195 196 // Reset the pal metadata so its data will not affect a compilation that 197 // reuses this object. 198 getPALMetadata()->reset(); 199 } 200 201 void AMDGPUTargetAsmStreamer::EmitDirectiveAMDGCNTarget() { 202 OS << "\t.amdgcn_target \"" << getTargetID()->toString() << "\"\n"; 203 } 204 205 void AMDGPUTargetAsmStreamer::EmitDirectiveHSACodeObjectVersion( 206 uint32_t Major, uint32_t Minor) { 207 OS << "\t.hsa_code_object_version " << 208 Twine(Major) << "," << Twine(Minor) << '\n'; 209 } 210 211 void 212 AMDGPUTargetAsmStreamer::EmitDirectiveHSACodeObjectISAV2(uint32_t Major, 213 uint32_t Minor, 214 uint32_t Stepping, 215 StringRef VendorName, 216 StringRef ArchName) { 217 convertIsaVersionV2(Major, Minor, Stepping, TargetID->isSramEccOnOrAny(), TargetID->isXnackOnOrAny()); 218 OS << "\t.hsa_code_object_isa " << Twine(Major) << "," << Twine(Minor) << "," 219 << Twine(Stepping) << ",\"" << VendorName << "\",\"" << ArchName << "\"\n"; 220 } 221 222 void 223 AMDGPUTargetAsmStreamer::EmitAMDKernelCodeT(const amd_kernel_code_t &Header) { 224 OS << "\t.amd_kernel_code_t\n"; 225 dumpAmdKernelCode(&Header, OS, "\t\t"); 226 OS << "\t.end_amd_kernel_code_t\n"; 227 } 228 229 void AMDGPUTargetAsmStreamer::EmitAMDGPUSymbolType(StringRef SymbolName, 230 unsigned Type) { 231 switch (Type) { 232 default: llvm_unreachable("Invalid AMDGPU symbol type"); 233 case ELF::STT_AMDGPU_HSA_KERNEL: 234 OS << "\t.amdgpu_hsa_kernel " << SymbolName << '\n' ; 235 break; 236 } 237 } 238 239 void AMDGPUTargetAsmStreamer::emitAMDGPULDS(MCSymbol *Symbol, unsigned Size, 240 Align Alignment) { 241 OS << "\t.amdgpu_lds " << Symbol->getName() << ", " << Size << ", " 242 << Alignment.value() << '\n'; 243 } 244 245 bool AMDGPUTargetAsmStreamer::EmitISAVersion() { 246 OS << "\t.amd_amdgpu_isa \"" << getTargetID()->toString() << "\"\n"; 247 return true; 248 } 249 250 bool AMDGPUTargetAsmStreamer::EmitHSAMetadata( 251 const AMDGPU::HSAMD::Metadata &HSAMetadata) { 252 std::string HSAMetadataString; 253 if (HSAMD::toString(HSAMetadata, HSAMetadataString)) 254 return false; 255 256 OS << '\t' << HSAMD::AssemblerDirectiveBegin << '\n'; 257 OS << HSAMetadataString << '\n'; 258 OS << '\t' << HSAMD::AssemblerDirectiveEnd << '\n'; 259 return true; 260 } 261 262 bool AMDGPUTargetAsmStreamer::EmitHSAMetadata( 263 msgpack::Document &HSAMetadataDoc, bool Strict) { 264 HSAMD::V3::MetadataVerifier Verifier(Strict); 265 if (!Verifier.verify(HSAMetadataDoc.getRoot())) 266 return false; 267 268 std::string HSAMetadataString; 269 raw_string_ostream StrOS(HSAMetadataString); 270 HSAMetadataDoc.toYAML(StrOS); 271 272 OS << '\t' << HSAMD::V3::AssemblerDirectiveBegin << '\n'; 273 OS << StrOS.str() << '\n'; 274 OS << '\t' << HSAMD::V3::AssemblerDirectiveEnd << '\n'; 275 return true; 276 } 277 278 bool AMDGPUTargetAsmStreamer::EmitCodeEnd(const MCSubtargetInfo &STI) { 279 const uint32_t Encoded_s_code_end = 0xbf9f0000; 280 const uint32_t Encoded_s_nop = 0xbf800000; 281 uint32_t Encoded_pad = Encoded_s_code_end; 282 283 // Instruction cache line size in bytes. 284 const unsigned Log2CacheLineSize = 6; 285 const unsigned CacheLineSize = 1u << Log2CacheLineSize; 286 287 // Extra padding amount in bytes to support prefetch mode 3. 288 unsigned FillSize = 3 * CacheLineSize; 289 290 if (AMDGPU::isGFX90A(STI)) { 291 Encoded_pad = Encoded_s_nop; 292 FillSize = 16 * CacheLineSize; 293 } 294 295 OS << "\t.p2alignl " << Log2CacheLineSize << ", " << Encoded_pad << '\n'; 296 OS << "\t.fill " << (FillSize / 4) << ", 4, " << Encoded_pad << '\n'; 297 return true; 298 } 299 300 void AMDGPUTargetAsmStreamer::EmitAmdhsaKernelDescriptor( 301 const MCSubtargetInfo &STI, StringRef KernelName, 302 const amdhsa::kernel_descriptor_t &KD, uint64_t NextVGPR, uint64_t NextSGPR, 303 bool ReserveVCC, bool ReserveFlatScr) { 304 IsaVersion IVersion = getIsaVersion(STI.getCPU()); 305 306 OS << "\t.amdhsa_kernel " << KernelName << '\n'; 307 308 #define PRINT_FIELD(STREAM, DIRECTIVE, KERNEL_DESC, MEMBER_NAME, FIELD_NAME) \ 309 STREAM << "\t\t" << DIRECTIVE << " " \ 310 << AMDHSA_BITS_GET(KERNEL_DESC.MEMBER_NAME, FIELD_NAME) << '\n'; 311 312 OS << "\t\t.amdhsa_group_segment_fixed_size " << KD.group_segment_fixed_size 313 << '\n'; 314 OS << "\t\t.amdhsa_private_segment_fixed_size " 315 << KD.private_segment_fixed_size << '\n'; 316 OS << "\t\t.amdhsa_kernarg_size " << KD.kernarg_size << '\n'; 317 318 if (!hasArchitectedFlatScratch(STI)) 319 PRINT_FIELD( 320 OS, ".amdhsa_user_sgpr_private_segment_buffer", KD, 321 kernel_code_properties, 322 amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER); 323 PRINT_FIELD(OS, ".amdhsa_user_sgpr_dispatch_ptr", KD, 324 kernel_code_properties, 325 amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR); 326 PRINT_FIELD(OS, ".amdhsa_user_sgpr_queue_ptr", KD, 327 kernel_code_properties, 328 amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR); 329 PRINT_FIELD(OS, ".amdhsa_user_sgpr_kernarg_segment_ptr", KD, 330 kernel_code_properties, 331 amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR); 332 PRINT_FIELD(OS, ".amdhsa_user_sgpr_dispatch_id", KD, 333 kernel_code_properties, 334 amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID); 335 if (!hasArchitectedFlatScratch(STI)) 336 PRINT_FIELD(OS, ".amdhsa_user_sgpr_flat_scratch_init", KD, 337 kernel_code_properties, 338 amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT); 339 PRINT_FIELD(OS, ".amdhsa_user_sgpr_private_segment_size", KD, 340 kernel_code_properties, 341 amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE); 342 if (IVersion.Major >= 10) 343 PRINT_FIELD(OS, ".amdhsa_wavefront_size32", KD, 344 kernel_code_properties, 345 amdhsa::KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32); 346 PRINT_FIELD(OS, 347 (hasArchitectedFlatScratch(STI) 348 ? ".amdhsa_enable_private_segment" 349 : ".amdhsa_system_sgpr_private_segment_wavefront_offset"), 350 KD, compute_pgm_rsrc2, 351 amdhsa::COMPUTE_PGM_RSRC2_ENABLE_PRIVATE_SEGMENT); 352 PRINT_FIELD(OS, ".amdhsa_system_sgpr_workgroup_id_x", KD, 353 compute_pgm_rsrc2, 354 amdhsa::COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_X); 355 PRINT_FIELD(OS, ".amdhsa_system_sgpr_workgroup_id_y", KD, 356 compute_pgm_rsrc2, 357 amdhsa::COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_Y); 358 PRINT_FIELD(OS, ".amdhsa_system_sgpr_workgroup_id_z", KD, 359 compute_pgm_rsrc2, 360 amdhsa::COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_Z); 361 PRINT_FIELD(OS, ".amdhsa_system_sgpr_workgroup_info", KD, 362 compute_pgm_rsrc2, 363 amdhsa::COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_INFO); 364 PRINT_FIELD(OS, ".amdhsa_system_vgpr_workitem_id", KD, 365 compute_pgm_rsrc2, 366 amdhsa::COMPUTE_PGM_RSRC2_ENABLE_VGPR_WORKITEM_ID); 367 368 // These directives are required. 369 OS << "\t\t.amdhsa_next_free_vgpr " << NextVGPR << '\n'; 370 OS << "\t\t.amdhsa_next_free_sgpr " << NextSGPR << '\n'; 371 372 if (AMDGPU::isGFX90A(STI)) 373 OS << "\t\t.amdhsa_accum_offset " << 374 (AMDHSA_BITS_GET(KD.compute_pgm_rsrc3, 375 amdhsa::COMPUTE_PGM_RSRC3_GFX90A_ACCUM_OFFSET) + 1) * 4 376 << '\n'; 377 378 if (!ReserveVCC) 379 OS << "\t\t.amdhsa_reserve_vcc " << ReserveVCC << '\n'; 380 if (IVersion.Major >= 7 && !ReserveFlatScr && !hasArchitectedFlatScratch(STI)) 381 OS << "\t\t.amdhsa_reserve_flat_scratch " << ReserveFlatScr << '\n'; 382 383 if (Optional<uint8_t> HsaAbiVer = getHsaAbiVersion(&STI)) { 384 switch (*HsaAbiVer) { 385 default: 386 break; 387 case ELF::ELFABIVERSION_AMDGPU_HSA_V2: 388 break; 389 case ELF::ELFABIVERSION_AMDGPU_HSA_V3: 390 case ELF::ELFABIVERSION_AMDGPU_HSA_V4: 391 if (getTargetID()->isXnackSupported()) 392 OS << "\t\t.amdhsa_reserve_xnack_mask " << getTargetID()->isXnackOnOrAny() << '\n'; 393 break; 394 } 395 } 396 397 PRINT_FIELD(OS, ".amdhsa_float_round_mode_32", KD, 398 compute_pgm_rsrc1, 399 amdhsa::COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_32); 400 PRINT_FIELD(OS, ".amdhsa_float_round_mode_16_64", KD, 401 compute_pgm_rsrc1, 402 amdhsa::COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_16_64); 403 PRINT_FIELD(OS, ".amdhsa_float_denorm_mode_32", KD, 404 compute_pgm_rsrc1, 405 amdhsa::COMPUTE_PGM_RSRC1_FLOAT_DENORM_MODE_32); 406 PRINT_FIELD(OS, ".amdhsa_float_denorm_mode_16_64", KD, 407 compute_pgm_rsrc1, 408 amdhsa::COMPUTE_PGM_RSRC1_FLOAT_DENORM_MODE_16_64); 409 PRINT_FIELD(OS, ".amdhsa_dx10_clamp", KD, 410 compute_pgm_rsrc1, 411 amdhsa::COMPUTE_PGM_RSRC1_ENABLE_DX10_CLAMP); 412 PRINT_FIELD(OS, ".amdhsa_ieee_mode", KD, 413 compute_pgm_rsrc1, 414 amdhsa::COMPUTE_PGM_RSRC1_ENABLE_IEEE_MODE); 415 if (IVersion.Major >= 9) 416 PRINT_FIELD(OS, ".amdhsa_fp16_overflow", KD, 417 compute_pgm_rsrc1, 418 amdhsa::COMPUTE_PGM_RSRC1_FP16_OVFL); 419 if (AMDGPU::isGFX90A(STI)) 420 PRINT_FIELD(OS, ".amdhsa_tg_split", KD, 421 compute_pgm_rsrc3, 422 amdhsa::COMPUTE_PGM_RSRC3_GFX90A_TG_SPLIT); 423 if (IVersion.Major >= 10) { 424 PRINT_FIELD(OS, ".amdhsa_workgroup_processor_mode", KD, 425 compute_pgm_rsrc1, 426 amdhsa::COMPUTE_PGM_RSRC1_WGP_MODE); 427 PRINT_FIELD(OS, ".amdhsa_memory_ordered", KD, 428 compute_pgm_rsrc1, 429 amdhsa::COMPUTE_PGM_RSRC1_MEM_ORDERED); 430 PRINT_FIELD(OS, ".amdhsa_forward_progress", KD, 431 compute_pgm_rsrc1, 432 amdhsa::COMPUTE_PGM_RSRC1_FWD_PROGRESS); 433 } 434 PRINT_FIELD( 435 OS, ".amdhsa_exception_fp_ieee_invalid_op", KD, 436 compute_pgm_rsrc2, 437 amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_INVALID_OPERATION); 438 PRINT_FIELD(OS, ".amdhsa_exception_fp_denorm_src", KD, 439 compute_pgm_rsrc2, 440 amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_FP_DENORMAL_SOURCE); 441 PRINT_FIELD( 442 OS, ".amdhsa_exception_fp_ieee_div_zero", KD, 443 compute_pgm_rsrc2, 444 amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_DIVISION_BY_ZERO); 445 PRINT_FIELD(OS, ".amdhsa_exception_fp_ieee_overflow", KD, 446 compute_pgm_rsrc2, 447 amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_OVERFLOW); 448 PRINT_FIELD(OS, ".amdhsa_exception_fp_ieee_underflow", KD, 449 compute_pgm_rsrc2, 450 amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_UNDERFLOW); 451 PRINT_FIELD(OS, ".amdhsa_exception_fp_ieee_inexact", KD, 452 compute_pgm_rsrc2, 453 amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_INEXACT); 454 PRINT_FIELD(OS, ".amdhsa_exception_int_div_zero", KD, 455 compute_pgm_rsrc2, 456 amdhsa::COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_INT_DIVIDE_BY_ZERO); 457 #undef PRINT_FIELD 458 459 OS << "\t.end_amdhsa_kernel\n"; 460 } 461 462 //===----------------------------------------------------------------------===// 463 // AMDGPUTargetELFStreamer 464 //===----------------------------------------------------------------------===// 465 466 AMDGPUTargetELFStreamer::AMDGPUTargetELFStreamer(MCStreamer &S, 467 const MCSubtargetInfo &STI) 468 : AMDGPUTargetStreamer(S), STI(STI), Streamer(S) {} 469 470 MCELFStreamer &AMDGPUTargetELFStreamer::getStreamer() { 471 return static_cast<MCELFStreamer &>(Streamer); 472 } 473 474 // A hook for emitting stuff at the end. 475 // We use it for emitting the accumulated PAL metadata as a .note record. 476 // The PAL metadata is reset after it is emitted. 477 void AMDGPUTargetELFStreamer::finish() { 478 MCAssembler &MCA = getStreamer().getAssembler(); 479 MCA.setELFHeaderEFlags(getEFlags()); 480 481 std::string Blob; 482 const char *Vendor = getPALMetadata()->getVendor(); 483 unsigned Type = getPALMetadata()->getType(); 484 getPALMetadata()->toBlob(Type, Blob); 485 if (Blob.empty()) 486 return; 487 EmitNote(Vendor, MCConstantExpr::create(Blob.size(), getContext()), Type, 488 [&](MCELFStreamer &OS) { OS.emitBytes(Blob); }); 489 490 // Reset the pal metadata so its data will not affect a compilation that 491 // reuses this object. 492 getPALMetadata()->reset(); 493 } 494 495 void AMDGPUTargetELFStreamer::EmitNote( 496 StringRef Name, const MCExpr *DescSZ, unsigned NoteType, 497 function_ref<void(MCELFStreamer &)> EmitDesc) { 498 auto &S = getStreamer(); 499 auto &Context = S.getContext(); 500 501 auto NameSZ = Name.size() + 1; 502 503 unsigned NoteFlags = 0; 504 // TODO Apparently, this is currently needed for OpenCL as mentioned in 505 // https://reviews.llvm.org/D74995 506 if (STI.getTargetTriple().getOS() == Triple::AMDHSA) 507 NoteFlags = ELF::SHF_ALLOC; 508 509 S.PushSection(); 510 S.SwitchSection( 511 Context.getELFSection(ElfNote::SectionName, ELF::SHT_NOTE, NoteFlags)); 512 S.emitInt32(NameSZ); // namesz 513 S.emitValue(DescSZ, 4); // descz 514 S.emitInt32(NoteType); // type 515 S.emitBytes(Name); // name 516 S.emitValueToAlignment(4, 0, 1, 0); // padding 0 517 EmitDesc(S); // desc 518 S.emitValueToAlignment(4, 0, 1, 0); // padding 0 519 S.PopSection(); 520 } 521 522 unsigned AMDGPUTargetELFStreamer::getEFlags() { 523 switch (STI.getTargetTriple().getArch()) { 524 default: 525 llvm_unreachable("Unsupported Arch"); 526 case Triple::r600: 527 return getEFlagsR600(); 528 case Triple::amdgcn: 529 return getEFlagsAMDGCN(); 530 } 531 } 532 533 unsigned AMDGPUTargetELFStreamer::getEFlagsR600() { 534 assert(STI.getTargetTriple().getArch() == Triple::r600); 535 536 return getElfMach(STI.getCPU()); 537 } 538 539 unsigned AMDGPUTargetELFStreamer::getEFlagsAMDGCN() { 540 assert(STI.getTargetTriple().getArch() == Triple::amdgcn); 541 542 switch (STI.getTargetTriple().getOS()) { 543 default: 544 // TODO: Why are some tests have "mingw" listed as OS? 545 // llvm_unreachable("Unsupported OS"); 546 case Triple::UnknownOS: 547 return getEFlagsUnknownOS(); 548 case Triple::AMDHSA: 549 return getEFlagsAMDHSA(); 550 case Triple::AMDPAL: 551 return getEFlagsAMDPAL(); 552 case Triple::Mesa3D: 553 return getEFlagsMesa3D(); 554 } 555 } 556 557 unsigned AMDGPUTargetELFStreamer::getEFlagsUnknownOS() { 558 // TODO: Why are some tests have "mingw" listed as OS? 559 // assert(STI.getTargetTriple().getOS() == Triple::UnknownOS); 560 561 return getEFlagsV3(); 562 } 563 564 unsigned AMDGPUTargetELFStreamer::getEFlagsAMDHSA() { 565 assert(STI.getTargetTriple().getOS() == Triple::AMDHSA); 566 567 if (Optional<uint8_t> HsaAbiVer = getHsaAbiVersion(&STI)) { 568 switch (*HsaAbiVer) { 569 case ELF::ELFABIVERSION_AMDGPU_HSA_V2: 570 case ELF::ELFABIVERSION_AMDGPU_HSA_V3: 571 return getEFlagsV3(); 572 case ELF::ELFABIVERSION_AMDGPU_HSA_V4: 573 return getEFlagsV4(); 574 } 575 } 576 577 llvm_unreachable("HSA OS ABI Version identification must be defined"); 578 } 579 580 unsigned AMDGPUTargetELFStreamer::getEFlagsAMDPAL() { 581 assert(STI.getTargetTriple().getOS() == Triple::AMDPAL); 582 583 return getEFlagsV3(); 584 } 585 586 unsigned AMDGPUTargetELFStreamer::getEFlagsMesa3D() { 587 assert(STI.getTargetTriple().getOS() == Triple::Mesa3D); 588 589 return getEFlagsV3(); 590 } 591 592 unsigned AMDGPUTargetELFStreamer::getEFlagsV3() { 593 unsigned EFlagsV3 = 0; 594 595 // mach. 596 EFlagsV3 |= getElfMach(STI.getCPU()); 597 598 // xnack. 599 if (getTargetID()->isXnackOnOrAny()) 600 EFlagsV3 |= ELF::EF_AMDGPU_FEATURE_XNACK_V3; 601 // sramecc. 602 if (getTargetID()->isSramEccOnOrAny()) 603 EFlagsV3 |= ELF::EF_AMDGPU_FEATURE_SRAMECC_V3; 604 605 return EFlagsV3; 606 } 607 608 unsigned AMDGPUTargetELFStreamer::getEFlagsV4() { 609 unsigned EFlagsV4 = 0; 610 611 // mach. 612 EFlagsV4 |= getElfMach(STI.getCPU()); 613 614 // xnack. 615 switch (getTargetID()->getXnackSetting()) { 616 case AMDGPU::IsaInfo::TargetIDSetting::Unsupported: 617 EFlagsV4 |= ELF::EF_AMDGPU_FEATURE_XNACK_UNSUPPORTED_V4; 618 break; 619 case AMDGPU::IsaInfo::TargetIDSetting::Any: 620 EFlagsV4 |= ELF::EF_AMDGPU_FEATURE_XNACK_ANY_V4; 621 break; 622 case AMDGPU::IsaInfo::TargetIDSetting::Off: 623 EFlagsV4 |= ELF::EF_AMDGPU_FEATURE_XNACK_OFF_V4; 624 break; 625 case AMDGPU::IsaInfo::TargetIDSetting::On: 626 EFlagsV4 |= ELF::EF_AMDGPU_FEATURE_XNACK_ON_V4; 627 break; 628 } 629 // sramecc. 630 switch (getTargetID()->getSramEccSetting()) { 631 case AMDGPU::IsaInfo::TargetIDSetting::Unsupported: 632 EFlagsV4 |= ELF::EF_AMDGPU_FEATURE_SRAMECC_UNSUPPORTED_V4; 633 break; 634 case AMDGPU::IsaInfo::TargetIDSetting::Any: 635 EFlagsV4 |= ELF::EF_AMDGPU_FEATURE_SRAMECC_ANY_V4; 636 break; 637 case AMDGPU::IsaInfo::TargetIDSetting::Off: 638 EFlagsV4 |= ELF::EF_AMDGPU_FEATURE_SRAMECC_OFF_V4; 639 break; 640 case AMDGPU::IsaInfo::TargetIDSetting::On: 641 EFlagsV4 |= ELF::EF_AMDGPU_FEATURE_SRAMECC_ON_V4; 642 break; 643 } 644 645 return EFlagsV4; 646 } 647 648 void AMDGPUTargetELFStreamer::EmitDirectiveAMDGCNTarget() {} 649 650 void AMDGPUTargetELFStreamer::EmitDirectiveHSACodeObjectVersion( 651 uint32_t Major, uint32_t Minor) { 652 653 EmitNote(ElfNote::NoteNameV2, MCConstantExpr::create(8, getContext()), 654 ELF::NT_AMD_HSA_CODE_OBJECT_VERSION, [&](MCELFStreamer &OS) { 655 OS.emitInt32(Major); 656 OS.emitInt32(Minor); 657 }); 658 } 659 660 void 661 AMDGPUTargetELFStreamer::EmitDirectiveHSACodeObjectISAV2(uint32_t Major, 662 uint32_t Minor, 663 uint32_t Stepping, 664 StringRef VendorName, 665 StringRef ArchName) { 666 uint16_t VendorNameSize = VendorName.size() + 1; 667 uint16_t ArchNameSize = ArchName.size() + 1; 668 669 unsigned DescSZ = sizeof(VendorNameSize) + sizeof(ArchNameSize) + 670 sizeof(Major) + sizeof(Minor) + sizeof(Stepping) + 671 VendorNameSize + ArchNameSize; 672 673 convertIsaVersionV2(Major, Minor, Stepping, TargetID->isSramEccOnOrAny(), TargetID->isXnackOnOrAny()); 674 EmitNote(ElfNote::NoteNameV2, MCConstantExpr::create(DescSZ, getContext()), 675 ELF::NT_AMD_HSA_ISA_VERSION, [&](MCELFStreamer &OS) { 676 OS.emitInt16(VendorNameSize); 677 OS.emitInt16(ArchNameSize); 678 OS.emitInt32(Major); 679 OS.emitInt32(Minor); 680 OS.emitInt32(Stepping); 681 OS.emitBytes(VendorName); 682 OS.emitInt8(0); // NULL terminate VendorName 683 OS.emitBytes(ArchName); 684 OS.emitInt8(0); // NULL terminte ArchName 685 }); 686 } 687 688 void 689 AMDGPUTargetELFStreamer::EmitAMDKernelCodeT(const amd_kernel_code_t &Header) { 690 691 MCStreamer &OS = getStreamer(); 692 OS.PushSection(); 693 OS.emitBytes(StringRef((const char*)&Header, sizeof(Header))); 694 OS.PopSection(); 695 } 696 697 void AMDGPUTargetELFStreamer::EmitAMDGPUSymbolType(StringRef SymbolName, 698 unsigned Type) { 699 MCSymbolELF *Symbol = cast<MCSymbolELF>( 700 getStreamer().getContext().getOrCreateSymbol(SymbolName)); 701 Symbol->setType(Type); 702 } 703 704 void AMDGPUTargetELFStreamer::emitAMDGPULDS(MCSymbol *Symbol, unsigned Size, 705 Align Alignment) { 706 MCSymbolELF *SymbolELF = cast<MCSymbolELF>(Symbol); 707 SymbolELF->setType(ELF::STT_OBJECT); 708 709 if (!SymbolELF->isBindingSet()) { 710 SymbolELF->setBinding(ELF::STB_GLOBAL); 711 SymbolELF->setExternal(true); 712 } 713 714 if (SymbolELF->declareCommon(Size, Alignment.value(), true)) { 715 report_fatal_error("Symbol: " + Symbol->getName() + 716 " redeclared as different type"); 717 } 718 719 SymbolELF->setIndex(ELF::SHN_AMDGPU_LDS); 720 SymbolELF->setSize(MCConstantExpr::create(Size, getContext())); 721 } 722 723 bool AMDGPUTargetELFStreamer::EmitISAVersion() { 724 // Create two labels to mark the beginning and end of the desc field 725 // and a MCExpr to calculate the size of the desc field. 726 auto &Context = getContext(); 727 auto *DescBegin = Context.createTempSymbol(); 728 auto *DescEnd = Context.createTempSymbol(); 729 auto *DescSZ = MCBinaryExpr::createSub( 730 MCSymbolRefExpr::create(DescEnd, Context), 731 MCSymbolRefExpr::create(DescBegin, Context), Context); 732 733 EmitNote(ElfNote::NoteNameV2, DescSZ, ELF::NT_AMD_HSA_ISA_NAME, 734 [&](MCELFStreamer &OS) { 735 OS.emitLabel(DescBegin); 736 OS.emitBytes(getTargetID()->toString()); 737 OS.emitLabel(DescEnd); 738 }); 739 return true; 740 } 741 742 bool AMDGPUTargetELFStreamer::EmitHSAMetadata(msgpack::Document &HSAMetadataDoc, 743 bool Strict) { 744 HSAMD::V3::MetadataVerifier Verifier(Strict); 745 if (!Verifier.verify(HSAMetadataDoc.getRoot())) 746 return false; 747 748 std::string HSAMetadataString; 749 HSAMetadataDoc.writeToBlob(HSAMetadataString); 750 751 // Create two labels to mark the beginning and end of the desc field 752 // and a MCExpr to calculate the size of the desc field. 753 auto &Context = getContext(); 754 auto *DescBegin = Context.createTempSymbol(); 755 auto *DescEnd = Context.createTempSymbol(); 756 auto *DescSZ = MCBinaryExpr::createSub( 757 MCSymbolRefExpr::create(DescEnd, Context), 758 MCSymbolRefExpr::create(DescBegin, Context), Context); 759 760 EmitNote(ElfNote::NoteNameV3, DescSZ, ELF::NT_AMDGPU_METADATA, 761 [&](MCELFStreamer &OS) { 762 OS.emitLabel(DescBegin); 763 OS.emitBytes(HSAMetadataString); 764 OS.emitLabel(DescEnd); 765 }); 766 return true; 767 } 768 769 bool AMDGPUTargetELFStreamer::EmitHSAMetadata( 770 const AMDGPU::HSAMD::Metadata &HSAMetadata) { 771 std::string HSAMetadataString; 772 if (HSAMD::toString(HSAMetadata, HSAMetadataString)) 773 return false; 774 775 // Create two labels to mark the beginning and end of the desc field 776 // and a MCExpr to calculate the size of the desc field. 777 auto &Context = getContext(); 778 auto *DescBegin = Context.createTempSymbol(); 779 auto *DescEnd = Context.createTempSymbol(); 780 auto *DescSZ = MCBinaryExpr::createSub( 781 MCSymbolRefExpr::create(DescEnd, Context), 782 MCSymbolRefExpr::create(DescBegin, Context), Context); 783 784 EmitNote(ElfNote::NoteNameV2, DescSZ, ELF::NT_AMD_HSA_METADATA, 785 [&](MCELFStreamer &OS) { 786 OS.emitLabel(DescBegin); 787 OS.emitBytes(HSAMetadataString); 788 OS.emitLabel(DescEnd); 789 }); 790 return true; 791 } 792 793 bool AMDGPUTargetELFStreamer::EmitCodeEnd(const MCSubtargetInfo &STI) { 794 const uint32_t Encoded_s_code_end = 0xbf9f0000; 795 const uint32_t Encoded_s_nop = 0xbf800000; 796 uint32_t Encoded_pad = Encoded_s_code_end; 797 798 // Instruction cache line size in bytes. 799 const unsigned Log2CacheLineSize = 6; 800 const unsigned CacheLineSize = 1u << Log2CacheLineSize; 801 802 // Extra padding amount in bytes to support prefetch mode 3. 803 unsigned FillSize = 3 * CacheLineSize; 804 805 if (AMDGPU::isGFX90A(STI)) { 806 Encoded_pad = Encoded_s_nop; 807 FillSize = 16 * CacheLineSize; 808 } 809 810 MCStreamer &OS = getStreamer(); 811 OS.PushSection(); 812 OS.emitValueToAlignment(CacheLineSize, Encoded_pad, 4); 813 for (unsigned I = 0; I < FillSize; I += 4) 814 OS.emitInt32(Encoded_pad); 815 OS.PopSection(); 816 return true; 817 } 818 819 void AMDGPUTargetELFStreamer::EmitAmdhsaKernelDescriptor( 820 const MCSubtargetInfo &STI, StringRef KernelName, 821 const amdhsa::kernel_descriptor_t &KernelDescriptor, uint64_t NextVGPR, 822 uint64_t NextSGPR, bool ReserveVCC, bool ReserveFlatScr) { 823 auto &Streamer = getStreamer(); 824 auto &Context = Streamer.getContext(); 825 826 MCSymbolELF *KernelCodeSymbol = cast<MCSymbolELF>( 827 Context.getOrCreateSymbol(Twine(KernelName))); 828 MCSymbolELF *KernelDescriptorSymbol = cast<MCSymbolELF>( 829 Context.getOrCreateSymbol(Twine(KernelName) + Twine(".kd"))); 830 831 // Copy kernel descriptor symbol's binding, other and visibility from the 832 // kernel code symbol. 833 KernelDescriptorSymbol->setBinding(KernelCodeSymbol->getBinding()); 834 KernelDescriptorSymbol->setOther(KernelCodeSymbol->getOther()); 835 KernelDescriptorSymbol->setVisibility(KernelCodeSymbol->getVisibility()); 836 // Kernel descriptor symbol's type and size are fixed. 837 KernelDescriptorSymbol->setType(ELF::STT_OBJECT); 838 KernelDescriptorSymbol->setSize( 839 MCConstantExpr::create(sizeof(KernelDescriptor), Context)); 840 841 // The visibility of the kernel code symbol must be protected or less to allow 842 // static relocations from the kernel descriptor to be used. 843 if (KernelCodeSymbol->getVisibility() == ELF::STV_DEFAULT) 844 KernelCodeSymbol->setVisibility(ELF::STV_PROTECTED); 845 846 Streamer.emitLabel(KernelDescriptorSymbol); 847 Streamer.emitInt32(KernelDescriptor.group_segment_fixed_size); 848 Streamer.emitInt32(KernelDescriptor.private_segment_fixed_size); 849 Streamer.emitInt32(KernelDescriptor.kernarg_size); 850 851 for (uint8_t Res : KernelDescriptor.reserved0) 852 Streamer.emitInt8(Res); 853 854 // FIXME: Remove the use of VK_AMDGPU_REL64 in the expression below. The 855 // expression being created is: 856 // (start of kernel code) - (start of kernel descriptor) 857 // It implies R_AMDGPU_REL64, but ends up being R_AMDGPU_ABS64. 858 Streamer.emitValue(MCBinaryExpr::createSub( 859 MCSymbolRefExpr::create( 860 KernelCodeSymbol, MCSymbolRefExpr::VK_AMDGPU_REL64, Context), 861 MCSymbolRefExpr::create( 862 KernelDescriptorSymbol, MCSymbolRefExpr::VK_None, Context), 863 Context), 864 sizeof(KernelDescriptor.kernel_code_entry_byte_offset)); 865 for (uint8_t Res : KernelDescriptor.reserved1) 866 Streamer.emitInt8(Res); 867 Streamer.emitInt32(KernelDescriptor.compute_pgm_rsrc3); 868 Streamer.emitInt32(KernelDescriptor.compute_pgm_rsrc1); 869 Streamer.emitInt32(KernelDescriptor.compute_pgm_rsrc2); 870 Streamer.emitInt16(KernelDescriptor.kernel_code_properties); 871 for (uint8_t Res : KernelDescriptor.reserved2) 872 Streamer.emitInt8(Res); 873 } 874