1 //===- AMDGPUBaseInfo.h - Top level definitions for AMDGPU ------*- C++ -*-===// 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 10 #ifndef LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUBASEINFO_H 11 #define LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUBASEINFO_H 12 13 #include "AMDGPU.h" 14 #include "AMDKernelCodeT.h" 15 #include "SIDefines.h" 16 #include "llvm/ADT/StringRef.h" 17 #include "llvm/IR/CallingConv.h" 18 #include "llvm/MC/MCInstrDesc.h" 19 #include "llvm/Support/Compiler.h" 20 #include "llvm/Support/ErrorHandling.h" 21 #include <cstdint> 22 #include <string> 23 #include <utility> 24 25 namespace llvm { 26 27 class Argument; 28 class FeatureBitset; 29 class Function; 30 class GlobalValue; 31 class MachineMemOperand; 32 class MCContext; 33 class MCRegisterClass; 34 class MCRegisterInfo; 35 class MCSection; 36 class MCSubtargetInfo; 37 class Triple; 38 39 namespace AMDGPU { 40 namespace IsaInfo { 41 42 enum { 43 // The closed Vulkan driver sets 96, which limits the wave count to 8 but 44 // doesn't spill SGPRs as much as when 80 is set. 45 FIXED_NUM_SGPRS_FOR_INIT_BUG = 96 46 }; 47 48 /// \brief Instruction set architecture version. 49 struct IsaVersion { 50 unsigned Major; 51 unsigned Minor; 52 unsigned Stepping; 53 }; 54 55 /// \returns Isa version for given subtarget \p Features. 56 IsaVersion getIsaVersion(const FeatureBitset &Features); 57 58 /// \brief Streams isa version string for given subtarget \p STI into \p Stream. 59 void streamIsaVersion(const MCSubtargetInfo *STI, raw_ostream &Stream); 60 61 /// \returns True if given subtarget \p Features support code object version 3, 62 /// false otherwise. 63 bool hasCodeObjectV3(const FeatureBitset &Features); 64 65 /// \returns Wavefront size for given subtarget \p Features. 66 unsigned getWavefrontSize(const FeatureBitset &Features); 67 68 /// \returns Local memory size in bytes for given subtarget \p Features. 69 unsigned getLocalMemorySize(const FeatureBitset &Features); 70 71 /// \returns Number of execution units per compute unit for given subtarget \p 72 /// Features. 73 unsigned getEUsPerCU(const FeatureBitset &Features); 74 75 /// \returns Maximum number of work groups per compute unit for given subtarget 76 /// \p Features and limited by given \p FlatWorkGroupSize. 77 unsigned getMaxWorkGroupsPerCU(const FeatureBitset &Features, 78 unsigned FlatWorkGroupSize); 79 80 /// \returns Maximum number of waves per compute unit for given subtarget \p 81 /// Features without any kind of limitation. 82 unsigned getMaxWavesPerCU(const FeatureBitset &Features); 83 84 /// \returns Maximum number of waves per compute unit for given subtarget \p 85 /// Features and limited by given \p FlatWorkGroupSize. 86 unsigned getMaxWavesPerCU(const FeatureBitset &Features, 87 unsigned FlatWorkGroupSize); 88 89 /// \returns Minimum number of waves per execution unit for given subtarget \p 90 /// Features. 91 unsigned getMinWavesPerEU(const FeatureBitset &Features); 92 93 /// \returns Maximum number of waves per execution unit for given subtarget \p 94 /// Features without any kind of limitation. 95 unsigned getMaxWavesPerEU(const FeatureBitset &Features); 96 97 /// \returns Maximum number of waves per execution unit for given subtarget \p 98 /// Features and limited by given \p FlatWorkGroupSize. 99 unsigned getMaxWavesPerEU(const FeatureBitset &Features, 100 unsigned FlatWorkGroupSize); 101 102 /// \returns Minimum flat work group size for given subtarget \p Features. 103 unsigned getMinFlatWorkGroupSize(const FeatureBitset &Features); 104 105 /// \returns Maximum flat work group size for given subtarget \p Features. 106 unsigned getMaxFlatWorkGroupSize(const FeatureBitset &Features); 107 108 /// \returns Number of waves per work group for given subtarget \p Features and 109 /// limited by given \p FlatWorkGroupSize. 110 unsigned getWavesPerWorkGroup(const FeatureBitset &Features, 111 unsigned FlatWorkGroupSize); 112 113 /// \returns SGPR allocation granularity for given subtarget \p Features. 114 unsigned getSGPRAllocGranule(const FeatureBitset &Features); 115 116 /// \returns SGPR encoding granularity for given subtarget \p Features. 117 unsigned getSGPREncodingGranule(const FeatureBitset &Features); 118 119 /// \returns Total number of SGPRs for given subtarget \p Features. 120 unsigned getTotalNumSGPRs(const FeatureBitset &Features); 121 122 /// \returns Addressable number of SGPRs for given subtarget \p Features. 123 unsigned getAddressableNumSGPRs(const FeatureBitset &Features); 124 125 /// \returns Minimum number of SGPRs that meets the given number of waves per 126 /// execution unit requirement for given subtarget \p Features. 127 unsigned getMinNumSGPRs(const FeatureBitset &Features, unsigned WavesPerEU); 128 129 /// \returns Maximum number of SGPRs that meets the given number of waves per 130 /// execution unit requirement for given subtarget \p Features. 131 unsigned getMaxNumSGPRs(const FeatureBitset &Features, unsigned WavesPerEU, 132 bool Addressable); 133 134 /// \returns VGPR allocation granularity for given subtarget \p Features. 135 unsigned getVGPRAllocGranule(const FeatureBitset &Features); 136 137 /// \returns VGPR encoding granularity for given subtarget \p Features. 138 unsigned getVGPREncodingGranule(const FeatureBitset &Features); 139 140 /// \returns Total number of VGPRs for given subtarget \p Features. 141 unsigned getTotalNumVGPRs(const FeatureBitset &Features); 142 143 /// \returns Addressable number of VGPRs for given subtarget \p Features. 144 unsigned getAddressableNumVGPRs(const FeatureBitset &Features); 145 146 /// \returns Minimum number of VGPRs that meets given number of waves per 147 /// execution unit requirement for given subtarget \p Features. 148 unsigned getMinNumVGPRs(const FeatureBitset &Features, unsigned WavesPerEU); 149 150 /// \returns Maximum number of VGPRs that meets given number of waves per 151 /// execution unit requirement for given subtarget \p Features. 152 unsigned getMaxNumVGPRs(const FeatureBitset &Features, unsigned WavesPerEU); 153 154 } // end namespace IsaInfo 155 156 LLVM_READONLY 157 int16_t getNamedOperandIdx(uint16_t Opcode, uint16_t NamedIdx); 158 159 LLVM_READONLY 160 int getMaskedMIMGOp(const MCInstrInfo &MII, 161 unsigned Opc, unsigned NewChannels); 162 163 LLVM_READONLY 164 int getMaskedMIMGAtomicOp(const MCInstrInfo &MII, 165 unsigned Opc, unsigned NewChannels); 166 167 LLVM_READONLY 168 int getMCOpcode(uint16_t Opcode, unsigned Gen); 169 170 void initDefaultAMDKernelCodeT(amd_kernel_code_t &Header, 171 const FeatureBitset &Features); 172 173 bool isGroupSegment(const GlobalValue *GV); 174 bool isGlobalSegment(const GlobalValue *GV); 175 bool isReadOnlySegment(const GlobalValue *GV); 176 177 /// \returns True if constants should be emitted to .text section for given 178 /// target triple \p TT, false otherwise. 179 bool shouldEmitConstantsToTextSection(const Triple &TT); 180 181 /// \returns Integer value requested using \p F's \p Name attribute. 182 /// 183 /// \returns \p Default if attribute is not present. 184 /// 185 /// \returns \p Default and emits error if requested value cannot be converted 186 /// to integer. 187 int getIntegerAttribute(const Function &F, StringRef Name, int Default); 188 189 /// \returns A pair of integer values requested using \p F's \p Name attribute 190 /// in "first[,second]" format ("second" is optional unless \p OnlyFirstRequired 191 /// is false). 192 /// 193 /// \returns \p Default if attribute is not present. 194 /// 195 /// \returns \p Default and emits error if one of the requested values cannot be 196 /// converted to integer, or \p OnlyFirstRequired is false and "second" value is 197 /// not present. 198 std::pair<int, int> getIntegerPairAttribute(const Function &F, 199 StringRef Name, 200 std::pair<int, int> Default, 201 bool OnlyFirstRequired = false); 202 203 /// \returns Vmcnt bit mask for given isa \p Version. 204 unsigned getVmcntBitMask(const IsaInfo::IsaVersion &Version); 205 206 /// \returns Expcnt bit mask for given isa \p Version. 207 unsigned getExpcntBitMask(const IsaInfo::IsaVersion &Version); 208 209 /// \returns Lgkmcnt bit mask for given isa \p Version. 210 unsigned getLgkmcntBitMask(const IsaInfo::IsaVersion &Version); 211 212 /// \returns Waitcnt bit mask for given isa \p Version. 213 unsigned getWaitcntBitMask(const IsaInfo::IsaVersion &Version); 214 215 /// \returns Decoded Vmcnt from given \p Waitcnt for given isa \p Version. 216 unsigned decodeVmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt); 217 218 /// \returns Decoded Expcnt from given \p Waitcnt for given isa \p Version. 219 unsigned decodeExpcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt); 220 221 /// \returns Decoded Lgkmcnt from given \p Waitcnt for given isa \p Version. 222 unsigned decodeLgkmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt); 223 224 /// \brief Decodes Vmcnt, Expcnt and Lgkmcnt from given \p Waitcnt for given isa 225 /// \p Version, and writes decoded values into \p Vmcnt, \p Expcnt and 226 /// \p Lgkmcnt respectively. 227 /// 228 /// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are decoded as follows: 229 /// \p Vmcnt = \p Waitcnt[3:0] (pre-gfx9 only) 230 /// \p Vmcnt = \p Waitcnt[3:0] | \p Waitcnt[15:14] (gfx9+ only) 231 /// \p Expcnt = \p Waitcnt[6:4] 232 /// \p Lgkmcnt = \p Waitcnt[11:8] 233 void decodeWaitcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt, 234 unsigned &Vmcnt, unsigned &Expcnt, unsigned &Lgkmcnt); 235 236 /// \returns \p Waitcnt with encoded \p Vmcnt for given isa \p Version. 237 unsigned encodeVmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt, 238 unsigned Vmcnt); 239 240 /// \returns \p Waitcnt with encoded \p Expcnt for given isa \p Version. 241 unsigned encodeExpcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt, 242 unsigned Expcnt); 243 244 /// \returns \p Waitcnt with encoded \p Lgkmcnt for given isa \p Version. 245 unsigned encodeLgkmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt, 246 unsigned Lgkmcnt); 247 248 /// \brief Encodes \p Vmcnt, \p Expcnt and \p Lgkmcnt into Waitcnt for given isa 249 /// \p Version. 250 /// 251 /// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are encoded as follows: 252 /// Waitcnt[3:0] = \p Vmcnt (pre-gfx9 only) 253 /// Waitcnt[3:0] = \p Vmcnt[3:0] (gfx9+ only) 254 /// Waitcnt[6:4] = \p Expcnt 255 /// Waitcnt[11:8] = \p Lgkmcnt 256 /// Waitcnt[15:14] = \p Vmcnt[5:4] (gfx9+ only) 257 /// 258 /// \returns Waitcnt with encoded \p Vmcnt, \p Expcnt and \p Lgkmcnt for given 259 /// isa \p Version. 260 unsigned encodeWaitcnt(const IsaInfo::IsaVersion &Version, 261 unsigned Vmcnt, unsigned Expcnt, unsigned Lgkmcnt); 262 263 unsigned getInitialPSInputAddr(const Function &F); 264 265 LLVM_READNONE 266 bool isShader(CallingConv::ID CC); 267 268 LLVM_READNONE 269 bool isCompute(CallingConv::ID CC); 270 271 LLVM_READNONE 272 bool isEntryFunctionCC(CallingConv::ID CC); 273 274 // FIXME: Remove this when calling conventions cleaned up 275 LLVM_READNONE 276 inline bool isKernel(CallingConv::ID CC) { 277 switch (CC) { 278 case CallingConv::AMDGPU_KERNEL: 279 case CallingConv::SPIR_KERNEL: 280 return true; 281 default: 282 return false; 283 } 284 } 285 286 bool hasXNACK(const MCSubtargetInfo &STI); 287 bool hasMIMG_R128(const MCSubtargetInfo &STI); 288 bool hasPackedD16(const MCSubtargetInfo &STI); 289 290 bool isSI(const MCSubtargetInfo &STI); 291 bool isCI(const MCSubtargetInfo &STI); 292 bool isVI(const MCSubtargetInfo &STI); 293 bool isGFX9(const MCSubtargetInfo &STI); 294 295 /// \brief Is Reg - scalar register 296 bool isSGPR(unsigned Reg, const MCRegisterInfo* TRI); 297 298 /// \brief Is there any intersection between registers 299 bool isRegIntersect(unsigned Reg0, unsigned Reg1, const MCRegisterInfo* TRI); 300 301 /// If \p Reg is a pseudo reg, return the correct hardware register given 302 /// \p STI otherwise return \p Reg. 303 unsigned getMCReg(unsigned Reg, const MCSubtargetInfo &STI); 304 305 /// \brief Convert hardware register \p Reg to a pseudo register 306 LLVM_READNONE 307 unsigned mc2PseudoReg(unsigned Reg); 308 309 /// \brief Can this operand also contain immediate values? 310 bool isSISrcOperand(const MCInstrDesc &Desc, unsigned OpNo); 311 312 /// \brief Is this floating-point operand? 313 bool isSISrcFPOperand(const MCInstrDesc &Desc, unsigned OpNo); 314 315 /// \brief Does this opearnd support only inlinable literals? 316 bool isSISrcInlinableOperand(const MCInstrDesc &Desc, unsigned OpNo); 317 318 /// \brief Get the size in bits of a register from the register class \p RC. 319 unsigned getRegBitWidth(unsigned RCID); 320 321 /// \brief Get the size in bits of a register from the register class \p RC. 322 unsigned getRegBitWidth(const MCRegisterClass &RC); 323 324 /// \brief Get size of register operand 325 unsigned getRegOperandSize(const MCRegisterInfo *MRI, const MCInstrDesc &Desc, 326 unsigned OpNo); 327 328 LLVM_READNONE 329 inline unsigned getOperandSize(const MCOperandInfo &OpInfo) { 330 switch (OpInfo.OperandType) { 331 case AMDGPU::OPERAND_REG_IMM_INT32: 332 case AMDGPU::OPERAND_REG_IMM_FP32: 333 case AMDGPU::OPERAND_REG_INLINE_C_INT32: 334 case AMDGPU::OPERAND_REG_INLINE_C_FP32: 335 return 4; 336 337 case AMDGPU::OPERAND_REG_IMM_INT64: 338 case AMDGPU::OPERAND_REG_IMM_FP64: 339 case AMDGPU::OPERAND_REG_INLINE_C_INT64: 340 case AMDGPU::OPERAND_REG_INLINE_C_FP64: 341 return 8; 342 343 case AMDGPU::OPERAND_REG_IMM_INT16: 344 case AMDGPU::OPERAND_REG_IMM_FP16: 345 case AMDGPU::OPERAND_REG_INLINE_C_INT16: 346 case AMDGPU::OPERAND_REG_INLINE_C_FP16: 347 case AMDGPU::OPERAND_REG_INLINE_C_V2INT16: 348 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16: 349 return 2; 350 351 default: 352 llvm_unreachable("unhandled operand type"); 353 } 354 } 355 356 LLVM_READNONE 357 inline unsigned getOperandSize(const MCInstrDesc &Desc, unsigned OpNo) { 358 return getOperandSize(Desc.OpInfo[OpNo]); 359 } 360 361 /// \brief Is this literal inlinable 362 LLVM_READNONE 363 bool isInlinableLiteral64(int64_t Literal, bool HasInv2Pi); 364 365 LLVM_READNONE 366 bool isInlinableLiteral32(int32_t Literal, bool HasInv2Pi); 367 368 LLVM_READNONE 369 bool isInlinableLiteral16(int16_t Literal, bool HasInv2Pi); 370 371 LLVM_READNONE 372 bool isInlinableLiteralV216(int32_t Literal, bool HasInv2Pi); 373 374 bool isArgPassedInSGPR(const Argument *Arg); 375 376 /// \returns The encoding that will be used for \p ByteOffset in the SMRD 377 /// offset field. 378 int64_t getSMRDEncodedOffset(const MCSubtargetInfo &ST, int64_t ByteOffset); 379 380 /// \returns true if this offset is small enough to fit in the SMRD 381 /// offset field. \p ByteOffset should be the offset in bytes and 382 /// not the encoded offset. 383 bool isLegalSMRDImmOffset(const MCSubtargetInfo &ST, int64_t ByteOffset); 384 385 } // end namespace AMDGPU 386 } // end namespace llvm 387 388 #endif // LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUBASEINFO_H 389