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/AMDHSAKernelDescriptor.h" 20 #include "llvm/Support/Compiler.h" 21 #include "llvm/Support/ErrorHandling.h" 22 #include "llvm/Support/TargetParser.h" 23 #include <cstdint> 24 #include <string> 25 #include <utility> 26 27 namespace llvm { 28 29 class Argument; 30 class AMDGPUSubtarget; 31 class FeatureBitset; 32 class Function; 33 class GCNSubtarget; 34 class GlobalValue; 35 class MCContext; 36 class MCRegisterClass; 37 class MCRegisterInfo; 38 class MCSection; 39 class MCSubtargetInfo; 40 class MachineMemOperand; 41 class Triple; 42 43 namespace AMDGPU { 44 45 #define GET_MIMGBaseOpcode_DECL 46 #define GET_MIMGDim_DECL 47 #define GET_MIMGEncoding_DECL 48 #define GET_MIMGLZMapping_DECL 49 #include "AMDGPUGenSearchableTables.inc" 50 51 namespace IsaInfo { 52 53 enum { 54 // The closed Vulkan driver sets 96, which limits the wave count to 8 but 55 // doesn't spill SGPRs as much as when 80 is set. 56 FIXED_NUM_SGPRS_FOR_INIT_BUG = 96, 57 TRAP_NUM_SGPRS = 16 58 }; 59 60 /// Streams isa version string for given subtarget \p STI into \p Stream. 61 void streamIsaVersion(const MCSubtargetInfo *STI, raw_ostream &Stream); 62 63 /// \returns True if given subtarget \p STI supports code object version 3, 64 /// false otherwise. 65 bool hasCodeObjectV3(const MCSubtargetInfo *STI); 66 67 /// \returns Wavefront size for given subtarget \p STI. 68 unsigned getWavefrontSize(const MCSubtargetInfo *STI); 69 70 /// \returns Local memory size in bytes for given subtarget \p STI. 71 unsigned getLocalMemorySize(const MCSubtargetInfo *STI); 72 73 /// \returns Number of execution units per compute unit for given subtarget \p 74 /// STI. 75 unsigned getEUsPerCU(const MCSubtargetInfo *STI); 76 77 /// \returns Maximum number of work groups per compute unit for given subtarget 78 /// \p STI and limited by given \p FlatWorkGroupSize. 79 unsigned getMaxWorkGroupsPerCU(const MCSubtargetInfo *STI, 80 unsigned FlatWorkGroupSize); 81 82 /// \returns Maximum number of waves per compute unit for given subtarget \p 83 /// STI without any kind of limitation. 84 unsigned getMaxWavesPerCU(const MCSubtargetInfo *STI); 85 86 /// \returns Maximum number of waves per compute unit for given subtarget \p 87 /// STI and limited by given \p FlatWorkGroupSize. 88 unsigned getMaxWavesPerCU(const MCSubtargetInfo *STI, 89 unsigned FlatWorkGroupSize); 90 91 /// \returns Minimum number of waves per execution unit for given subtarget \p 92 /// STI. 93 unsigned getMinWavesPerEU(const MCSubtargetInfo *STI); 94 95 /// \returns Maximum number of waves per execution unit for given subtarget \p 96 /// STI without any kind of limitation. 97 unsigned getMaxWavesPerEU(); 98 99 /// \returns Maximum number of waves per execution unit for given subtarget \p 100 /// STI and limited by given \p FlatWorkGroupSize. 101 unsigned getMaxWavesPerEU(const MCSubtargetInfo *STI, 102 unsigned FlatWorkGroupSize); 103 104 /// \returns Minimum flat work group size for given subtarget \p STI. 105 unsigned getMinFlatWorkGroupSize(const MCSubtargetInfo *STI); 106 107 /// \returns Maximum flat work group size for given subtarget \p STI. 108 unsigned getMaxFlatWorkGroupSize(const MCSubtargetInfo *STI); 109 110 /// \returns Number of waves per work group for given subtarget \p STI and 111 /// limited by given \p FlatWorkGroupSize. 112 unsigned getWavesPerWorkGroup(const MCSubtargetInfo *STI, 113 unsigned FlatWorkGroupSize); 114 115 /// \returns SGPR allocation granularity for given subtarget \p STI. 116 unsigned getSGPRAllocGranule(const MCSubtargetInfo *STI); 117 118 /// \returns SGPR encoding granularity for given subtarget \p STI. 119 unsigned getSGPREncodingGranule(const MCSubtargetInfo *STI); 120 121 /// \returns Total number of SGPRs for given subtarget \p STI. 122 unsigned getTotalNumSGPRs(const MCSubtargetInfo *STI); 123 124 /// \returns Addressable number of SGPRs for given subtarget \p STI. 125 unsigned getAddressableNumSGPRs(const MCSubtargetInfo *STI); 126 127 /// \returns Minimum number of SGPRs that meets the given number of waves per 128 /// execution unit requirement for given subtarget \p STI. 129 unsigned getMinNumSGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU); 130 131 /// \returns Maximum number of SGPRs that meets the given number of waves per 132 /// execution unit requirement for given subtarget \p STI. 133 unsigned getMaxNumSGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU, 134 bool Addressable); 135 136 /// \returns Number of extra SGPRs implicitly required by given subtarget \p 137 /// STI when the given special registers are used. 138 unsigned getNumExtraSGPRs(const MCSubtargetInfo *STI, bool VCCUsed, 139 bool FlatScrUsed, bool XNACKUsed); 140 141 /// \returns Number of extra SGPRs implicitly required by given subtarget \p 142 /// STI when the given special registers are used. XNACK is inferred from 143 /// \p STI. 144 unsigned getNumExtraSGPRs(const MCSubtargetInfo *STI, bool VCCUsed, 145 bool FlatScrUsed); 146 147 /// \returns Number of SGPR blocks needed for given subtarget \p STI when 148 /// \p NumSGPRs are used. \p NumSGPRs should already include any special 149 /// register counts. 150 unsigned getNumSGPRBlocks(const MCSubtargetInfo *STI, unsigned NumSGPRs); 151 152 /// \returns VGPR allocation granularity for given subtarget \p STI. 153 unsigned getVGPRAllocGranule(const MCSubtargetInfo *STI); 154 155 /// \returns VGPR encoding granularity for given subtarget \p STI. 156 unsigned getVGPREncodingGranule(const MCSubtargetInfo *STI); 157 158 /// \returns Total number of VGPRs for given subtarget \p STI. 159 unsigned getTotalNumVGPRs(const MCSubtargetInfo *STI); 160 161 /// \returns Addressable number of VGPRs for given subtarget \p STI. 162 unsigned getAddressableNumVGPRs(const MCSubtargetInfo *STI); 163 164 /// \returns Minimum number of VGPRs that meets given number of waves per 165 /// execution unit requirement for given subtarget \p STI. 166 unsigned getMinNumVGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU); 167 168 /// \returns Maximum number of VGPRs that meets given number of waves per 169 /// execution unit requirement for given subtarget \p STI. 170 unsigned getMaxNumVGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU); 171 172 /// \returns Number of VGPR blocks needed for given subtarget \p STI when 173 /// \p NumVGPRs are used. 174 unsigned getNumVGPRBlocks(const MCSubtargetInfo *STI, unsigned NumSGPRs); 175 176 } // end namespace IsaInfo 177 178 LLVM_READONLY 179 int16_t getNamedOperandIdx(uint16_t Opcode, uint16_t NamedIdx); 180 181 struct MIMGBaseOpcodeInfo { 182 MIMGBaseOpcode BaseOpcode; 183 bool Store; 184 bool Atomic; 185 bool AtomicX2; 186 bool Sampler; 187 188 uint8_t NumExtraArgs; 189 bool Gradients; 190 bool Coordinates; 191 bool LodOrClampOrMip; 192 bool HasD16; 193 }; 194 195 LLVM_READONLY 196 const MIMGBaseOpcodeInfo *getMIMGBaseOpcodeInfo(unsigned BaseOpcode); 197 198 struct MIMGDimInfo { 199 MIMGDim Dim; 200 uint8_t NumCoords; 201 uint8_t NumGradients; 202 bool DA; 203 }; 204 205 LLVM_READONLY 206 const MIMGDimInfo *getMIMGDimInfo(unsigned Dim); 207 208 struct MIMGLZMappingInfo { 209 MIMGBaseOpcode L; 210 MIMGBaseOpcode LZ; 211 }; 212 213 LLVM_READONLY 214 const MIMGLZMappingInfo *getMIMGLZMappingInfo(unsigned L); 215 216 LLVM_READONLY 217 int getMIMGOpcode(unsigned BaseOpcode, unsigned MIMGEncoding, 218 unsigned VDataDwords, unsigned VAddrDwords); 219 220 LLVM_READONLY 221 int getMaskedMIMGOp(unsigned Opc, unsigned NewChannels); 222 223 LLVM_READONLY 224 int getMUBUFBaseOpcode(unsigned Opc); 225 226 LLVM_READONLY 227 int getMUBUFOpcode(unsigned BaseOpc, unsigned Dwords); 228 229 LLVM_READONLY 230 int getMUBUFDwords(unsigned Opc); 231 232 LLVM_READONLY 233 bool getMUBUFHasVAddr(unsigned Opc); 234 235 LLVM_READONLY 236 bool getMUBUFHasSrsrc(unsigned Opc); 237 238 LLVM_READONLY 239 bool getMUBUFHasSoffset(unsigned Opc); 240 241 LLVM_READONLY 242 int getMCOpcode(uint16_t Opcode, unsigned Gen); 243 244 void initDefaultAMDKernelCodeT(amd_kernel_code_t &Header, 245 const MCSubtargetInfo *STI); 246 247 amdhsa::kernel_descriptor_t getDefaultAmdhsaKernelDescriptor(); 248 249 bool isGroupSegment(const GlobalValue *GV); 250 bool isGlobalSegment(const GlobalValue *GV); 251 bool isReadOnlySegment(const GlobalValue *GV); 252 253 /// \returns True if constants should be emitted to .text section for given 254 /// target triple \p TT, false otherwise. 255 bool shouldEmitConstantsToTextSection(const Triple &TT); 256 257 /// \returns Integer value requested using \p F's \p Name attribute. 258 /// 259 /// \returns \p Default if attribute is not present. 260 /// 261 /// \returns \p Default and emits error if requested value cannot be converted 262 /// to integer. 263 int getIntegerAttribute(const Function &F, StringRef Name, int Default); 264 265 /// \returns A pair of integer values requested using \p F's \p Name attribute 266 /// in "first[,second]" format ("second" is optional unless \p OnlyFirstRequired 267 /// is false). 268 /// 269 /// \returns \p Default if attribute is not present. 270 /// 271 /// \returns \p Default and emits error if one of the requested values cannot be 272 /// converted to integer, or \p OnlyFirstRequired is false and "second" value is 273 /// not present. 274 std::pair<int, int> getIntegerPairAttribute(const Function &F, 275 StringRef Name, 276 std::pair<int, int> Default, 277 bool OnlyFirstRequired = false); 278 279 /// Represents the counter values to wait for in an s_waitcnt instruction. 280 /// 281 /// Large values (including the maximum possible integer) can be used to 282 /// represent "don't care" waits. 283 struct Waitcnt { 284 unsigned VmCnt = ~0u; 285 unsigned ExpCnt = ~0u; 286 unsigned LgkmCnt = ~0u; 287 288 Waitcnt() {} 289 Waitcnt(unsigned VmCnt, unsigned ExpCnt, unsigned LgkmCnt) 290 : VmCnt(VmCnt), ExpCnt(ExpCnt), LgkmCnt(LgkmCnt) {} 291 292 static Waitcnt allZero() { return Waitcnt(0, 0, 0); } 293 294 bool dominates(const Waitcnt &Other) const { 295 return VmCnt <= Other.VmCnt && ExpCnt <= Other.ExpCnt && 296 LgkmCnt <= Other.LgkmCnt; 297 } 298 299 Waitcnt combined(const Waitcnt &Other) const { 300 return Waitcnt(std::min(VmCnt, Other.VmCnt), std::min(ExpCnt, Other.ExpCnt), 301 std::min(LgkmCnt, Other.LgkmCnt)); 302 } 303 }; 304 305 /// \returns Vmcnt bit mask for given isa \p Version. 306 unsigned getVmcntBitMask(const IsaVersion &Version); 307 308 /// \returns Expcnt bit mask for given isa \p Version. 309 unsigned getExpcntBitMask(const IsaVersion &Version); 310 311 /// \returns Lgkmcnt bit mask for given isa \p Version. 312 unsigned getLgkmcntBitMask(const IsaVersion &Version); 313 314 /// \returns Waitcnt bit mask for given isa \p Version. 315 unsigned getWaitcntBitMask(const IsaVersion &Version); 316 317 /// \returns Decoded Vmcnt from given \p Waitcnt for given isa \p Version. 318 unsigned decodeVmcnt(const IsaVersion &Version, unsigned Waitcnt); 319 320 /// \returns Decoded Expcnt from given \p Waitcnt for given isa \p Version. 321 unsigned decodeExpcnt(const IsaVersion &Version, unsigned Waitcnt); 322 323 /// \returns Decoded Lgkmcnt from given \p Waitcnt for given isa \p Version. 324 unsigned decodeLgkmcnt(const IsaVersion &Version, unsigned Waitcnt); 325 326 /// Decodes Vmcnt, Expcnt and Lgkmcnt from given \p Waitcnt for given isa 327 /// \p Version, and writes decoded values into \p Vmcnt, \p Expcnt and 328 /// \p Lgkmcnt respectively. 329 /// 330 /// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are decoded as follows: 331 /// \p Vmcnt = \p Waitcnt[3:0] (pre-gfx9 only) 332 /// \p Vmcnt = \p Waitcnt[3:0] | \p Waitcnt[15:14] (gfx9+ only) 333 /// \p Expcnt = \p Waitcnt[6:4] 334 /// \p Lgkmcnt = \p Waitcnt[11:8] 335 void decodeWaitcnt(const IsaVersion &Version, unsigned Waitcnt, 336 unsigned &Vmcnt, unsigned &Expcnt, unsigned &Lgkmcnt); 337 338 Waitcnt decodeWaitcnt(const IsaVersion &Version, unsigned Encoded); 339 340 /// \returns \p Waitcnt with encoded \p Vmcnt for given isa \p Version. 341 unsigned encodeVmcnt(const IsaVersion &Version, unsigned Waitcnt, 342 unsigned Vmcnt); 343 344 /// \returns \p Waitcnt with encoded \p Expcnt for given isa \p Version. 345 unsigned encodeExpcnt(const IsaVersion &Version, unsigned Waitcnt, 346 unsigned Expcnt); 347 348 /// \returns \p Waitcnt with encoded \p Lgkmcnt for given isa \p Version. 349 unsigned encodeLgkmcnt(const IsaVersion &Version, unsigned Waitcnt, 350 unsigned Lgkmcnt); 351 352 /// Encodes \p Vmcnt, \p Expcnt and \p Lgkmcnt into Waitcnt for given isa 353 /// \p Version. 354 /// 355 /// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are encoded as follows: 356 /// Waitcnt[3:0] = \p Vmcnt (pre-gfx9 only) 357 /// Waitcnt[3:0] = \p Vmcnt[3:0] (gfx9+ only) 358 /// Waitcnt[6:4] = \p Expcnt 359 /// Waitcnt[11:8] = \p Lgkmcnt 360 /// Waitcnt[15:14] = \p Vmcnt[5:4] (gfx9+ only) 361 /// 362 /// \returns Waitcnt with encoded \p Vmcnt, \p Expcnt and \p Lgkmcnt for given 363 /// isa \p Version. 364 unsigned encodeWaitcnt(const IsaVersion &Version, 365 unsigned Vmcnt, unsigned Expcnt, unsigned Lgkmcnt); 366 367 unsigned encodeWaitcnt(const IsaVersion &Version, const Waitcnt &Decoded); 368 369 unsigned getInitialPSInputAddr(const Function &F); 370 371 LLVM_READNONE 372 bool isShader(CallingConv::ID CC); 373 374 LLVM_READNONE 375 bool isCompute(CallingConv::ID CC); 376 377 LLVM_READNONE 378 bool isEntryFunctionCC(CallingConv::ID CC); 379 380 // FIXME: Remove this when calling conventions cleaned up 381 LLVM_READNONE 382 inline bool isKernel(CallingConv::ID CC) { 383 switch (CC) { 384 case CallingConv::AMDGPU_KERNEL: 385 case CallingConv::SPIR_KERNEL: 386 return true; 387 default: 388 return false; 389 } 390 } 391 392 bool hasXNACK(const MCSubtargetInfo &STI); 393 bool hasSRAMECC(const MCSubtargetInfo &STI); 394 bool hasMIMG_R128(const MCSubtargetInfo &STI); 395 bool hasPackedD16(const MCSubtargetInfo &STI); 396 397 bool isSI(const MCSubtargetInfo &STI); 398 bool isCI(const MCSubtargetInfo &STI); 399 bool isVI(const MCSubtargetInfo &STI); 400 bool isGFX9(const MCSubtargetInfo &STI); 401 402 /// Is Reg - scalar register 403 bool isSGPR(unsigned Reg, const MCRegisterInfo* TRI); 404 405 /// Is there any intersection between registers 406 bool isRegIntersect(unsigned Reg0, unsigned Reg1, const MCRegisterInfo* TRI); 407 408 /// If \p Reg is a pseudo reg, return the correct hardware register given 409 /// \p STI otherwise return \p Reg. 410 unsigned getMCReg(unsigned Reg, const MCSubtargetInfo &STI); 411 412 /// Convert hardware register \p Reg to a pseudo register 413 LLVM_READNONE 414 unsigned mc2PseudoReg(unsigned Reg); 415 416 /// Can this operand also contain immediate values? 417 bool isSISrcOperand(const MCInstrDesc &Desc, unsigned OpNo); 418 419 /// Is this floating-point operand? 420 bool isSISrcFPOperand(const MCInstrDesc &Desc, unsigned OpNo); 421 422 /// Does this opearnd support only inlinable literals? 423 bool isSISrcInlinableOperand(const MCInstrDesc &Desc, unsigned OpNo); 424 425 /// Get the size in bits of a register from the register class \p RC. 426 unsigned getRegBitWidth(unsigned RCID); 427 428 /// Get the size in bits of a register from the register class \p RC. 429 unsigned getRegBitWidth(const MCRegisterClass &RC); 430 431 /// Get size of register operand 432 unsigned getRegOperandSize(const MCRegisterInfo *MRI, const MCInstrDesc &Desc, 433 unsigned OpNo); 434 435 LLVM_READNONE 436 inline unsigned getOperandSize(const MCOperandInfo &OpInfo) { 437 switch (OpInfo.OperandType) { 438 case AMDGPU::OPERAND_REG_IMM_INT32: 439 case AMDGPU::OPERAND_REG_IMM_FP32: 440 case AMDGPU::OPERAND_REG_INLINE_C_INT32: 441 case AMDGPU::OPERAND_REG_INLINE_C_FP32: 442 return 4; 443 444 case AMDGPU::OPERAND_REG_IMM_INT64: 445 case AMDGPU::OPERAND_REG_IMM_FP64: 446 case AMDGPU::OPERAND_REG_INLINE_C_INT64: 447 case AMDGPU::OPERAND_REG_INLINE_C_FP64: 448 return 8; 449 450 case AMDGPU::OPERAND_REG_IMM_INT16: 451 case AMDGPU::OPERAND_REG_IMM_FP16: 452 case AMDGPU::OPERAND_REG_INLINE_C_INT16: 453 case AMDGPU::OPERAND_REG_INLINE_C_FP16: 454 case AMDGPU::OPERAND_REG_INLINE_C_V2INT16: 455 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16: 456 return 2; 457 458 default: 459 llvm_unreachable("unhandled operand type"); 460 } 461 } 462 463 LLVM_READNONE 464 inline unsigned getOperandSize(const MCInstrDesc &Desc, unsigned OpNo) { 465 return getOperandSize(Desc.OpInfo[OpNo]); 466 } 467 468 /// Is this literal inlinable 469 LLVM_READNONE 470 bool isInlinableLiteral64(int64_t Literal, bool HasInv2Pi); 471 472 LLVM_READNONE 473 bool isInlinableLiteral32(int32_t Literal, bool HasInv2Pi); 474 475 LLVM_READNONE 476 bool isInlinableLiteral16(int16_t Literal, bool HasInv2Pi); 477 478 LLVM_READNONE 479 bool isInlinableLiteralV216(int32_t Literal, bool HasInv2Pi); 480 481 bool isArgPassedInSGPR(const Argument *Arg); 482 483 /// \returns The encoding that will be used for \p ByteOffset in the SMRD 484 /// offset field. 485 int64_t getSMRDEncodedOffset(const MCSubtargetInfo &ST, int64_t ByteOffset); 486 487 /// \returns true if this offset is small enough to fit in the SMRD 488 /// offset field. \p ByteOffset should be the offset in bytes and 489 /// not the encoded offset. 490 bool isLegalSMRDImmOffset(const MCSubtargetInfo &ST, int64_t ByteOffset); 491 492 bool splitMUBUFOffset(uint32_t Imm, uint32_t &SOffset, uint32_t &ImmOffset, 493 const GCNSubtarget *Subtarget, uint32_t Align = 4); 494 495 /// \returns true if the intrinsic is divergent 496 bool isIntrinsicSourceOfDivergence(unsigned IntrID); 497 498 } // end namespace AMDGPU 499 } // end namespace llvm 500 501 #endif // LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUBASEINFO_H 502