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