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 void initDefaultAMDKernelCodeT(amd_kernel_code_t &Header, 160 const FeatureBitset &Features); 161 162 bool isGroupSegment(const GlobalValue *GV); 163 bool isGlobalSegment(const GlobalValue *GV); 164 bool isReadOnlySegment(const GlobalValue *GV); 165 166 /// \returns True if constants should be emitted to .text section for given 167 /// target triple \p TT, false otherwise. 168 bool shouldEmitConstantsToTextSection(const Triple &TT); 169 170 /// \returns Integer value requested using \p F's \p Name attribute. 171 /// 172 /// \returns \p Default if attribute is not present. 173 /// 174 /// \returns \p Default and emits error if requested value cannot be converted 175 /// to integer. 176 int getIntegerAttribute(const Function &F, StringRef Name, int Default); 177 178 /// \returns A pair of integer values requested using \p F's \p Name attribute 179 /// in "first[,second]" format ("second" is optional unless \p OnlyFirstRequired 180 /// is false). 181 /// 182 /// \returns \p Default if attribute is not present. 183 /// 184 /// \returns \p Default and emits error if one of the requested values cannot be 185 /// converted to integer, or \p OnlyFirstRequired is false and "second" value is 186 /// not present. 187 std::pair<int, int> getIntegerPairAttribute(const Function &F, 188 StringRef Name, 189 std::pair<int, int> Default, 190 bool OnlyFirstRequired = false); 191 192 /// \returns Vmcnt bit mask for given isa \p Version. 193 unsigned getVmcntBitMask(const IsaInfo::IsaVersion &Version); 194 195 /// \returns Expcnt bit mask for given isa \p Version. 196 unsigned getExpcntBitMask(const IsaInfo::IsaVersion &Version); 197 198 /// \returns Lgkmcnt bit mask for given isa \p Version. 199 unsigned getLgkmcntBitMask(const IsaInfo::IsaVersion &Version); 200 201 /// \returns Waitcnt bit mask for given isa \p Version. 202 unsigned getWaitcntBitMask(const IsaInfo::IsaVersion &Version); 203 204 /// \returns Decoded Vmcnt from given \p Waitcnt for given isa \p Version. 205 unsigned decodeVmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt); 206 207 /// \returns Decoded Expcnt from given \p Waitcnt for given isa \p Version. 208 unsigned decodeExpcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt); 209 210 /// \returns Decoded Lgkmcnt from given \p Waitcnt for given isa \p Version. 211 unsigned decodeLgkmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt); 212 213 /// \brief Decodes Vmcnt, Expcnt and Lgkmcnt from given \p Waitcnt for given isa 214 /// \p Version, and writes decoded values into \p Vmcnt, \p Expcnt and 215 /// \p Lgkmcnt respectively. 216 /// 217 /// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are decoded as follows: 218 /// \p Vmcnt = \p Waitcnt[3:0] (pre-gfx9 only) 219 /// \p Vmcnt = \p Waitcnt[3:0] | \p Waitcnt[15:14] (gfx9+ only) 220 /// \p Expcnt = \p Waitcnt[6:4] 221 /// \p Lgkmcnt = \p Waitcnt[11:8] 222 void decodeWaitcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt, 223 unsigned &Vmcnt, unsigned &Expcnt, unsigned &Lgkmcnt); 224 225 /// \returns \p Waitcnt with encoded \p Vmcnt for given isa \p Version. 226 unsigned encodeVmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt, 227 unsigned Vmcnt); 228 229 /// \returns \p Waitcnt with encoded \p Expcnt for given isa \p Version. 230 unsigned encodeExpcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt, 231 unsigned Expcnt); 232 233 /// \returns \p Waitcnt with encoded \p Lgkmcnt for given isa \p Version. 234 unsigned encodeLgkmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt, 235 unsigned Lgkmcnt); 236 237 /// \brief Encodes \p Vmcnt, \p Expcnt and \p Lgkmcnt into Waitcnt for given isa 238 /// \p Version. 239 /// 240 /// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are encoded as follows: 241 /// Waitcnt[3:0] = \p Vmcnt (pre-gfx9 only) 242 /// Waitcnt[3:0] = \p Vmcnt[3:0] (gfx9+ only) 243 /// Waitcnt[6:4] = \p Expcnt 244 /// Waitcnt[11:8] = \p Lgkmcnt 245 /// Waitcnt[15:14] = \p Vmcnt[5:4] (gfx9+ only) 246 /// 247 /// \returns Waitcnt with encoded \p Vmcnt, \p Expcnt and \p Lgkmcnt for given 248 /// isa \p Version. 249 unsigned encodeWaitcnt(const IsaInfo::IsaVersion &Version, 250 unsigned Vmcnt, unsigned Expcnt, unsigned Lgkmcnt); 251 252 unsigned getInitialPSInputAddr(const Function &F); 253 254 LLVM_READNONE 255 bool isShader(CallingConv::ID CC); 256 257 LLVM_READNONE 258 bool isCompute(CallingConv::ID CC); 259 260 LLVM_READNONE 261 bool isEntryFunctionCC(CallingConv::ID CC); 262 263 // FIXME: Remove this when calling conventions cleaned up 264 LLVM_READNONE 265 inline bool isKernel(CallingConv::ID CC) { 266 switch (CC) { 267 case CallingConv::AMDGPU_KERNEL: 268 case CallingConv::SPIR_KERNEL: 269 return true; 270 default: 271 return false; 272 } 273 } 274 275 bool isSI(const MCSubtargetInfo &STI); 276 bool isCI(const MCSubtargetInfo &STI); 277 bool isVI(const MCSubtargetInfo &STI); 278 bool isGFX9(const MCSubtargetInfo &STI); 279 280 /// \brief Is Reg - scalar register 281 bool isSGPR(unsigned Reg, const MCRegisterInfo* TRI); 282 283 /// \brief Is there any intersection between registers 284 bool isRegIntersect(unsigned Reg0, unsigned Reg1, const MCRegisterInfo* TRI); 285 286 /// If \p Reg is a pseudo reg, return the correct hardware register given 287 /// \p STI otherwise return \p Reg. 288 unsigned getMCReg(unsigned Reg, const MCSubtargetInfo &STI); 289 290 /// \brief Convert hardware register \p Reg to a pseudo register 291 LLVM_READNONE 292 unsigned mc2PseudoReg(unsigned Reg); 293 294 /// \brief Can this operand also contain immediate values? 295 bool isSISrcOperand(const MCInstrDesc &Desc, unsigned OpNo); 296 297 /// \brief Is this floating-point operand? 298 bool isSISrcFPOperand(const MCInstrDesc &Desc, unsigned OpNo); 299 300 /// \brief Does this opearnd support only inlinable literals? 301 bool isSISrcInlinableOperand(const MCInstrDesc &Desc, unsigned OpNo); 302 303 /// \brief Get the size in bits of a register from the register class \p RC. 304 unsigned getRegBitWidth(unsigned RCID); 305 306 /// \brief Get the size in bits of a register from the register class \p RC. 307 unsigned getRegBitWidth(const MCRegisterClass &RC); 308 309 /// \brief Get size of register operand 310 unsigned getRegOperandSize(const MCRegisterInfo *MRI, const MCInstrDesc &Desc, 311 unsigned OpNo); 312 313 LLVM_READNONE 314 inline unsigned getOperandSize(const MCOperandInfo &OpInfo) { 315 switch (OpInfo.OperandType) { 316 case AMDGPU::OPERAND_REG_IMM_INT32: 317 case AMDGPU::OPERAND_REG_IMM_FP32: 318 case AMDGPU::OPERAND_REG_INLINE_C_INT32: 319 case AMDGPU::OPERAND_REG_INLINE_C_FP32: 320 return 4; 321 322 case AMDGPU::OPERAND_REG_IMM_INT64: 323 case AMDGPU::OPERAND_REG_IMM_FP64: 324 case AMDGPU::OPERAND_REG_INLINE_C_INT64: 325 case AMDGPU::OPERAND_REG_INLINE_C_FP64: 326 return 8; 327 328 case AMDGPU::OPERAND_REG_IMM_INT16: 329 case AMDGPU::OPERAND_REG_IMM_FP16: 330 case AMDGPU::OPERAND_REG_INLINE_C_INT16: 331 case AMDGPU::OPERAND_REG_INLINE_C_FP16: 332 case AMDGPU::OPERAND_REG_INLINE_C_V2INT16: 333 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16: 334 return 2; 335 336 default: 337 llvm_unreachable("unhandled operand type"); 338 } 339 } 340 341 LLVM_READNONE 342 inline unsigned getOperandSize(const MCInstrDesc &Desc, unsigned OpNo) { 343 return getOperandSize(Desc.OpInfo[OpNo]); 344 } 345 346 /// \brief Is this literal inlinable 347 LLVM_READNONE 348 bool isInlinableLiteral64(int64_t Literal, bool HasInv2Pi); 349 350 LLVM_READNONE 351 bool isInlinableLiteral32(int32_t Literal, bool HasInv2Pi); 352 353 LLVM_READNONE 354 bool isInlinableLiteral16(int16_t Literal, bool HasInv2Pi); 355 356 LLVM_READNONE 357 bool isInlinableLiteralV216(int32_t Literal, bool HasInv2Pi); 358 359 bool isArgPassedInSGPR(const Argument *Arg); 360 bool isUniformMMO(const MachineMemOperand *MMO); 361 362 /// \returns The encoding that will be used for \p ByteOffset in the SMRD 363 /// offset field. 364 int64_t getSMRDEncodedOffset(const MCSubtargetInfo &ST, int64_t ByteOffset); 365 366 /// \returns true if this offset is small enough to fit in the SMRD 367 /// offset field. \p ByteOffset should be the offset in bytes and 368 /// not the encoded offset. 369 bool isLegalSMRDImmOffset(const MCSubtargetInfo &ST, int64_t ByteOffset); 370 371 } // end namespace AMDGPU 372 } // end namespace llvm 373 374 #endif // LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUBASEINFO_H 375