1 //===-- AMDGPUBaseInfo.cpp - AMDGPU Base encoding information--------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 #include "AMDGPUBaseInfo.h" 10 #include "AMDGPU.h" 11 #include "SIDefines.h" 12 #include "llvm/IR/LLVMContext.h" 13 #include "llvm/IR/Function.h" 14 #include "llvm/IR/GlobalValue.h" 15 #include "llvm/MC/MCContext.h" 16 #include "llvm/MC/MCInstrInfo.h" 17 #include "llvm/MC/MCRegisterInfo.h" 18 #include "llvm/MC/MCSectionELF.h" 19 #include "llvm/MC/MCSubtargetInfo.h" 20 #include "llvm/MC/SubtargetFeature.h" 21 22 #define GET_SUBTARGETINFO_ENUM 23 #include "AMDGPUGenSubtargetInfo.inc" 24 #undef GET_SUBTARGETINFO_ENUM 25 26 #define GET_REGINFO_ENUM 27 #include "AMDGPUGenRegisterInfo.inc" 28 #undef GET_REGINFO_ENUM 29 30 #define GET_INSTRINFO_NAMED_OPS 31 #define GET_INSTRINFO_ENUM 32 #include "AMDGPUGenInstrInfo.inc" 33 #undef GET_INSTRINFO_NAMED_OPS 34 #undef GET_INSTRINFO_ENUM 35 36 namespace { 37 38 /// \returns Bit mask for given bit \p Shift and bit \p Width. 39 unsigned getBitMask(unsigned Shift, unsigned Width) { 40 return ((1 << Width) - 1) << Shift; 41 } 42 43 /// \brief Packs \p Src into \p Dst for given bit \p Shift and bit \p Width. 44 /// 45 /// \returns Packed \p Dst. 46 unsigned packBits(unsigned Src, unsigned Dst, unsigned Shift, unsigned Width) { 47 Dst &= ~(1 << Shift) & ~getBitMask(Shift, Width); 48 Dst |= (Src << Shift) & getBitMask(Shift, Width); 49 return Dst; 50 } 51 52 /// \brief Unpacks bits from \p Src for given bit \p Shift and bit \p Width. 53 /// 54 /// \returns Unpacked bits. 55 unsigned unpackBits(unsigned Src, unsigned Shift, unsigned Width) { 56 return (Src & getBitMask(Shift, Width)) >> Shift; 57 } 58 59 /// \returns Vmcnt bit shift. 60 unsigned getVmcntBitShift() { return 0; } 61 62 /// \returns Vmcnt bit width. 63 unsigned getVmcntBitWidth() { return 4; } 64 65 /// \returns Expcnt bit shift. 66 unsigned getExpcntBitShift() { return 4; } 67 68 /// \returns Expcnt bit width. 69 unsigned getExpcntBitWidth() { return 3; } 70 71 /// \returns Lgkmcnt bit shift. 72 unsigned getLgkmcntBitShift() { return 8; } 73 74 /// \returns Lgkmcnt bit width. 75 unsigned getLgkmcntBitWidth() { return 4; } 76 77 } // anonymous namespace 78 79 namespace llvm { 80 namespace AMDGPU { 81 82 IsaVersion getIsaVersion(const FeatureBitset &Features) { 83 84 if (Features.test(FeatureISAVersion7_0_0)) 85 return {7, 0, 0}; 86 87 if (Features.test(FeatureISAVersion7_0_1)) 88 return {7, 0, 1}; 89 90 if (Features.test(FeatureISAVersion8_0_0)) 91 return {8, 0, 0}; 92 93 if (Features.test(FeatureISAVersion8_0_1)) 94 return {8, 0, 1}; 95 96 if (Features.test(FeatureISAVersion8_0_2)) 97 return {8, 0, 2}; 98 99 if (Features.test(FeatureISAVersion8_0_3)) 100 return {8, 0, 3}; 101 102 return {0, 0, 0}; 103 } 104 105 void initDefaultAMDKernelCodeT(amd_kernel_code_t &Header, 106 const FeatureBitset &Features) { 107 108 IsaVersion ISA = getIsaVersion(Features); 109 110 memset(&Header, 0, sizeof(Header)); 111 112 Header.amd_kernel_code_version_major = 1; 113 Header.amd_kernel_code_version_minor = 0; 114 Header.amd_machine_kind = 1; // AMD_MACHINE_KIND_AMDGPU 115 Header.amd_machine_version_major = ISA.Major; 116 Header.amd_machine_version_minor = ISA.Minor; 117 Header.amd_machine_version_stepping = ISA.Stepping; 118 Header.kernel_code_entry_byte_offset = sizeof(Header); 119 // wavefront_size is specified as a power of 2: 2^6 = 64 threads. 120 Header.wavefront_size = 6; 121 // These alignment values are specified in powers of two, so alignment = 122 // 2^n. The minimum alignment is 2^4 = 16. 123 Header.kernarg_segment_alignment = 4; 124 Header.group_segment_alignment = 4; 125 Header.private_segment_alignment = 4; 126 } 127 128 MCSection *getHSATextSection(MCContext &Ctx) { 129 return Ctx.getELFSection(".hsatext", ELF::SHT_PROGBITS, 130 ELF::SHF_ALLOC | ELF::SHF_WRITE | 131 ELF::SHF_EXECINSTR | 132 ELF::SHF_AMDGPU_HSA_AGENT | 133 ELF::SHF_AMDGPU_HSA_CODE); 134 } 135 136 MCSection *getHSADataGlobalAgentSection(MCContext &Ctx) { 137 return Ctx.getELFSection(".hsadata_global_agent", ELF::SHT_PROGBITS, 138 ELF::SHF_ALLOC | ELF::SHF_WRITE | 139 ELF::SHF_AMDGPU_HSA_GLOBAL | 140 ELF::SHF_AMDGPU_HSA_AGENT); 141 } 142 143 MCSection *getHSADataGlobalProgramSection(MCContext &Ctx) { 144 return Ctx.getELFSection(".hsadata_global_program", ELF::SHT_PROGBITS, 145 ELF::SHF_ALLOC | ELF::SHF_WRITE | 146 ELF::SHF_AMDGPU_HSA_GLOBAL); 147 } 148 149 MCSection *getHSARodataReadonlyAgentSection(MCContext &Ctx) { 150 return Ctx.getELFSection(".hsarodata_readonly_agent", ELF::SHT_PROGBITS, 151 ELF::SHF_ALLOC | ELF::SHF_AMDGPU_HSA_READONLY | 152 ELF::SHF_AMDGPU_HSA_AGENT); 153 } 154 155 bool isGroupSegment(const GlobalValue *GV) { 156 return GV->getType()->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS; 157 } 158 159 bool isGlobalSegment(const GlobalValue *GV) { 160 return GV->getType()->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS; 161 } 162 163 bool isReadOnlySegment(const GlobalValue *GV) { 164 return GV->getType()->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS; 165 } 166 167 bool shouldEmitConstantsToTextSection(const Triple &TT) { 168 return TT.getOS() != Triple::AMDHSA; 169 } 170 171 int getIntegerAttribute(const Function &F, StringRef Name, int Default) { 172 Attribute A = F.getFnAttribute(Name); 173 int Result = Default; 174 175 if (A.isStringAttribute()) { 176 StringRef Str = A.getValueAsString(); 177 if (Str.getAsInteger(0, Result)) { 178 LLVMContext &Ctx = F.getContext(); 179 Ctx.emitError("can't parse integer attribute " + Name); 180 } 181 } 182 183 return Result; 184 } 185 186 std::pair<int, int> getIntegerPairAttribute(const Function &F, 187 StringRef Name, 188 std::pair<int, int> Default, 189 bool OnlyFirstRequired) { 190 Attribute A = F.getFnAttribute(Name); 191 if (!A.isStringAttribute()) 192 return Default; 193 194 LLVMContext &Ctx = F.getContext(); 195 std::pair<int, int> Ints = Default; 196 std::pair<StringRef, StringRef> Strs = A.getValueAsString().split(','); 197 if (Strs.first.trim().getAsInteger(0, Ints.first)) { 198 Ctx.emitError("can't parse first integer attribute " + Name); 199 return Default; 200 } 201 if (Strs.second.trim().getAsInteger(0, Ints.second)) { 202 if (!OnlyFirstRequired || Strs.second.trim().size()) { 203 Ctx.emitError("can't parse second integer attribute " + Name); 204 return Default; 205 } 206 } 207 208 return Ints; 209 } 210 211 unsigned getWaitcntBitMask(IsaVersion Version) { 212 unsigned Vmcnt = getBitMask(getVmcntBitShift(), getVmcntBitWidth()); 213 unsigned Expcnt = getBitMask(getExpcntBitShift(), getExpcntBitWidth()); 214 unsigned Lgkmcnt = getBitMask(getLgkmcntBitShift(), getLgkmcntBitWidth()); 215 return Vmcnt | Expcnt | Lgkmcnt; 216 } 217 218 unsigned getVmcntBitMask(IsaVersion Version) { 219 return (1 << getVmcntBitWidth()) - 1; 220 } 221 222 unsigned getExpcntBitMask(IsaVersion Version) { 223 return (1 << getExpcntBitWidth()) - 1; 224 } 225 226 unsigned getLgkmcntBitMask(IsaVersion Version) { 227 return (1 << getLgkmcntBitWidth()) - 1; 228 } 229 230 unsigned decodeVmcnt(IsaVersion Version, unsigned Waitcnt) { 231 return unpackBits(Waitcnt, getVmcntBitShift(), getVmcntBitWidth()); 232 } 233 234 unsigned decodeExpcnt(IsaVersion Version, unsigned Waitcnt) { 235 return unpackBits(Waitcnt, getExpcntBitShift(), getExpcntBitWidth()); 236 } 237 238 unsigned decodeLgkmcnt(IsaVersion Version, unsigned Waitcnt) { 239 return unpackBits(Waitcnt, getLgkmcntBitShift(), getLgkmcntBitWidth()); 240 } 241 242 void decodeWaitcnt(IsaVersion Version, unsigned Waitcnt, 243 unsigned &Vmcnt, unsigned &Expcnt, unsigned &Lgkmcnt) { 244 Vmcnt = decodeVmcnt(Version, Waitcnt); 245 Expcnt = decodeExpcnt(Version, Waitcnt); 246 Lgkmcnt = decodeLgkmcnt(Version, Waitcnt); 247 } 248 249 unsigned encodeVmcnt(IsaVersion Version, unsigned Waitcnt, unsigned Vmcnt) { 250 return packBits(Vmcnt, Waitcnt, getVmcntBitShift(), getVmcntBitWidth()); 251 } 252 253 unsigned encodeExpcnt(IsaVersion Version, unsigned Waitcnt, unsigned Expcnt) { 254 return packBits(Expcnt, Waitcnt, getExpcntBitShift(), getExpcntBitWidth()); 255 } 256 257 unsigned encodeLgkmcnt(IsaVersion Version, unsigned Waitcnt, unsigned Lgkmcnt) { 258 return packBits(Lgkmcnt, Waitcnt, getLgkmcntBitShift(), getLgkmcntBitWidth()); 259 } 260 261 unsigned encodeWaitcnt(IsaVersion Version, 262 unsigned Vmcnt, unsigned Expcnt, unsigned Lgkmcnt) { 263 unsigned Waitcnt = getWaitcntBitMask(Version);; 264 Waitcnt = encodeVmcnt(Version, Waitcnt, Vmcnt); 265 Waitcnt = encodeExpcnt(Version, Waitcnt, Expcnt); 266 Waitcnt = encodeLgkmcnt(Version, Waitcnt, Lgkmcnt); 267 return Waitcnt; 268 } 269 270 unsigned getInitialPSInputAddr(const Function &F) { 271 return getIntegerAttribute(F, "InitialPSInputAddr", 0); 272 } 273 274 bool isShader(CallingConv::ID cc) { 275 switch(cc) { 276 case CallingConv::AMDGPU_VS: 277 case CallingConv::AMDGPU_GS: 278 case CallingConv::AMDGPU_PS: 279 case CallingConv::AMDGPU_CS: 280 return true; 281 default: 282 return false; 283 } 284 } 285 286 bool isCompute(CallingConv::ID cc) { 287 return !isShader(cc) || cc == CallingConv::AMDGPU_CS; 288 } 289 290 bool isSI(const MCSubtargetInfo &STI) { 291 return STI.getFeatureBits()[AMDGPU::FeatureSouthernIslands]; 292 } 293 294 bool isCI(const MCSubtargetInfo &STI) { 295 return STI.getFeatureBits()[AMDGPU::FeatureSeaIslands]; 296 } 297 298 bool isVI(const MCSubtargetInfo &STI) { 299 return STI.getFeatureBits()[AMDGPU::FeatureVolcanicIslands]; 300 } 301 302 unsigned getMCReg(unsigned Reg, const MCSubtargetInfo &STI) { 303 304 switch(Reg) { 305 default: break; 306 case AMDGPU::FLAT_SCR: 307 assert(!isSI(STI)); 308 return isCI(STI) ? AMDGPU::FLAT_SCR_ci : AMDGPU::FLAT_SCR_vi; 309 310 case AMDGPU::FLAT_SCR_LO: 311 assert(!isSI(STI)); 312 return isCI(STI) ? AMDGPU::FLAT_SCR_LO_ci : AMDGPU::FLAT_SCR_LO_vi; 313 314 case AMDGPU::FLAT_SCR_HI: 315 assert(!isSI(STI)); 316 return isCI(STI) ? AMDGPU::FLAT_SCR_HI_ci : AMDGPU::FLAT_SCR_HI_vi; 317 } 318 return Reg; 319 } 320 321 bool isSISrcOperand(const MCInstrDesc &Desc, unsigned OpNo) { 322 unsigned OpType = Desc.OpInfo[OpNo].OperandType; 323 324 return OpType == AMDGPU::OPERAND_REG_IMM32_INT || 325 OpType == AMDGPU::OPERAND_REG_IMM32_FP || 326 OpType == AMDGPU::OPERAND_REG_INLINE_C_INT || 327 OpType == AMDGPU::OPERAND_REG_INLINE_C_FP; 328 } 329 330 bool isSISrcFPOperand(const MCInstrDesc &Desc, unsigned OpNo) { 331 unsigned OpType = Desc.OpInfo[OpNo].OperandType; 332 333 return OpType == AMDGPU::OPERAND_REG_IMM32_FP || 334 OpType == AMDGPU::OPERAND_REG_INLINE_C_FP; 335 } 336 337 bool isSISrcInlinableOperand(const MCInstrDesc &Desc, unsigned OpNo) { 338 unsigned OpType = Desc.OpInfo[OpNo].OperandType; 339 340 return OpType == AMDGPU::OPERAND_REG_INLINE_C_INT || 341 OpType == AMDGPU::OPERAND_REG_INLINE_C_FP; 342 } 343 344 // Avoid using MCRegisterClass::getSize, since that function will go away 345 // (move from MC* level to Target* level). Return size in bits. 346 unsigned getRegBitWidth(const MCRegisterClass &RC) { 347 switch (RC.getID()) { 348 case AMDGPU::SGPR_32RegClassID: 349 case AMDGPU::VGPR_32RegClassID: 350 case AMDGPU::VS_32RegClassID: 351 case AMDGPU::SReg_32RegClassID: 352 case AMDGPU::SReg_32_XM0RegClassID: 353 return 32; 354 case AMDGPU::SGPR_64RegClassID: 355 case AMDGPU::VS_64RegClassID: 356 case AMDGPU::SReg_64RegClassID: 357 case AMDGPU::VReg_64RegClassID: 358 return 64; 359 case AMDGPU::VReg_96RegClassID: 360 return 96; 361 case AMDGPU::SGPR_128RegClassID: 362 case AMDGPU::SReg_128RegClassID: 363 case AMDGPU::VReg_128RegClassID: 364 return 128; 365 case AMDGPU::SReg_256RegClassID: 366 case AMDGPU::VReg_256RegClassID: 367 return 256; 368 case AMDGPU::SReg_512RegClassID: 369 case AMDGPU::VReg_512RegClassID: 370 return 512; 371 default: 372 llvm_unreachable("Unexpected register class"); 373 } 374 } 375 376 unsigned getRegOperandSize(const MCRegisterInfo *MRI, const MCInstrDesc &Desc, 377 unsigned OpNo) { 378 unsigned RCID = Desc.OpInfo[OpNo].RegClass; 379 return getRegBitWidth(MRI->getRegClass(RCID)) / 8; 380 } 381 382 bool isInlinableLiteral64(int64_t Literal, bool IsVI) { 383 if (Literal >= -16 && Literal <= 64) 384 return true; 385 386 double D = BitsToDouble(Literal); 387 388 if (D == 0.5 || D == -0.5 || 389 D == 1.0 || D == -1.0 || 390 D == 2.0 || D == -2.0 || 391 D == 4.0 || D == -4.0) 392 return true; 393 394 if (IsVI && Literal == 0x3fc45f306dc9c882) 395 return true; 396 397 return false; 398 } 399 400 bool isInlinableLiteral32(int32_t Literal, bool IsVI) { 401 if (Literal >= -16 && Literal <= 64) 402 return true; 403 404 float F = BitsToFloat(Literal); 405 406 if (F == 0.5 || F == -0.5 || 407 F == 1.0 || F == -1.0 || 408 F == 2.0 || F == -2.0 || 409 F == 4.0 || F == -4.0) 410 return true; 411 412 if (IsVI && Literal == 0x3e22f983) 413 return true; 414 415 return false; 416 } 417 418 419 } // End namespace AMDGPU 420 } // End namespace llvm 421