1 //===- AMDGPUBaseInfo.h - Top level definitions for AMDGPU ------*- C++ -*-===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 9 #ifndef LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUBASEINFO_H 10 #define LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUBASEINFO_H 11 12 #include "SIDefines.h" 13 #include "llvm/IR/CallingConv.h" 14 #include "llvm/Support/Alignment.h" 15 16 struct amd_kernel_code_t; 17 18 namespace llvm { 19 20 struct Align; 21 class Argument; 22 class Function; 23 class GCNSubtarget; 24 class GlobalValue; 25 class MCRegisterClass; 26 class MCRegisterInfo; 27 class MCSubtargetInfo; 28 class StringRef; 29 class Triple; 30 31 namespace amdhsa { 32 struct kernel_descriptor_t; 33 } 34 35 namespace AMDGPU { 36 37 struct IsaVersion; 38 39 /// \returns HSA OS ABI Version identification. 40 Optional<uint8_t> getHsaAbiVersion(const MCSubtargetInfo *STI); 41 /// \returns True if HSA OS ABI Version identification is 2, 42 /// false otherwise. 43 bool isHsaAbiVersion2(const MCSubtargetInfo *STI); 44 /// \returns True if HSA OS ABI Version identification is 3, 45 /// false otherwise. 46 bool isHsaAbiVersion3(const MCSubtargetInfo *STI); 47 /// \returns True if HSA OS ABI Version identification is 4, 48 /// false otherwise. 49 bool isHsaAbiVersion4(const MCSubtargetInfo *STI); 50 /// \returns True if HSA OS ABI Version identification is 3 or 4, 51 /// false otherwise. 52 bool isHsaAbiVersion3Or4(const MCSubtargetInfo *STI); 53 54 struct GcnBufferFormatInfo { 55 unsigned Format; 56 unsigned BitsPerComp; 57 unsigned NumComponents; 58 unsigned NumFormat; 59 unsigned DataFormat; 60 }; 61 62 #define GET_MIMGBaseOpcode_DECL 63 #define GET_MIMGDim_DECL 64 #define GET_MIMGEncoding_DECL 65 #define GET_MIMGLZMapping_DECL 66 #define GET_MIMGMIPMapping_DECL 67 #include "AMDGPUGenSearchableTables.inc" 68 69 namespace IsaInfo { 70 71 enum { 72 // The closed Vulkan driver sets 96, which limits the wave count to 8 but 73 // doesn't spill SGPRs as much as when 80 is set. 74 FIXED_NUM_SGPRS_FOR_INIT_BUG = 96, 75 TRAP_NUM_SGPRS = 16 76 }; 77 78 enum class TargetIDSetting { 79 Unsupported, 80 Any, 81 Off, 82 On 83 }; 84 85 class AMDGPUTargetID { 86 private: 87 const MCSubtargetInfo &STI; 88 TargetIDSetting XnackSetting; 89 TargetIDSetting SramEccSetting; 90 91 public: 92 explicit AMDGPUTargetID(const MCSubtargetInfo &STI); 93 ~AMDGPUTargetID() = default; 94 95 /// \return True if the current xnack setting is not "Unsupported". 96 bool isXnackSupported() const { 97 return XnackSetting != TargetIDSetting::Unsupported; 98 } 99 100 /// \returns True if the current xnack setting is "On" or "Any". 101 bool isXnackOnOrAny() const { 102 return XnackSetting == TargetIDSetting::On || 103 XnackSetting == TargetIDSetting::Any; 104 } 105 106 /// \returns True if current xnack setting is "On" or "Off", 107 /// false otherwise. 108 bool isXnackOnOrOff() const { 109 return getXnackSetting() == TargetIDSetting::On || 110 getXnackSetting() == TargetIDSetting::Off; 111 } 112 113 /// \returns The current xnack TargetIDSetting, possible options are 114 /// "Unsupported", "Any", "Off", and "On". 115 TargetIDSetting getXnackSetting() const { 116 return XnackSetting; 117 } 118 119 /// Sets xnack setting to \p NewXnackSetting. 120 void setXnackSetting(TargetIDSetting NewXnackSetting) { 121 XnackSetting = NewXnackSetting; 122 } 123 124 /// \return True if the current sramecc setting is not "Unsupported". 125 bool isSramEccSupported() const { 126 return SramEccSetting != TargetIDSetting::Unsupported; 127 } 128 129 /// \returns True if the current sramecc setting is "On" or "Any". 130 bool isSramEccOnOrAny() const { 131 return SramEccSetting == TargetIDSetting::On || 132 SramEccSetting == TargetIDSetting::Any; 133 } 134 135 /// \returns True if current sramecc setting is "On" or "Off", 136 /// false otherwise. 137 bool isSramEccOnOrOff() const { 138 return getSramEccSetting() == TargetIDSetting::On || 139 getSramEccSetting() == TargetIDSetting::Off; 140 } 141 142 /// \returns The current sramecc TargetIDSetting, possible options are 143 /// "Unsupported", "Any", "Off", and "On". 144 TargetIDSetting getSramEccSetting() const { 145 return SramEccSetting; 146 } 147 148 /// Sets sramecc setting to \p NewSramEccSetting. 149 void setSramEccSetting(TargetIDSetting NewSramEccSetting) { 150 SramEccSetting = NewSramEccSetting; 151 } 152 153 void setTargetIDFromFeaturesString(StringRef FS); 154 void setTargetIDFromTargetIDStream(StringRef TargetID); 155 156 /// \returns String representation of an object. 157 std::string toString() const; 158 }; 159 160 /// \returns Wavefront size for given subtarget \p STI. 161 unsigned getWavefrontSize(const MCSubtargetInfo *STI); 162 163 /// \returns Local memory size in bytes for given subtarget \p STI. 164 unsigned getLocalMemorySize(const MCSubtargetInfo *STI); 165 166 /// \returns Number of execution units per compute unit for given subtarget \p 167 /// STI. 168 unsigned getEUsPerCU(const MCSubtargetInfo *STI); 169 170 /// \returns Maximum number of work groups per compute unit for given subtarget 171 /// \p STI and limited by given \p FlatWorkGroupSize. 172 unsigned getMaxWorkGroupsPerCU(const MCSubtargetInfo *STI, 173 unsigned FlatWorkGroupSize); 174 175 /// \returns Minimum number of waves per execution unit for given subtarget \p 176 /// STI. 177 unsigned getMinWavesPerEU(const MCSubtargetInfo *STI); 178 179 /// \returns Maximum number of waves per execution unit for given subtarget \p 180 /// STI without any kind of limitation. 181 unsigned getMaxWavesPerEU(const MCSubtargetInfo *STI); 182 183 /// \returns Number of waves per execution unit required to support the given \p 184 /// FlatWorkGroupSize. 185 unsigned getWavesPerEUForWorkGroup(const MCSubtargetInfo *STI, 186 unsigned FlatWorkGroupSize); 187 188 /// \returns Minimum flat work group size for given subtarget \p STI. 189 unsigned getMinFlatWorkGroupSize(const MCSubtargetInfo *STI); 190 191 /// \returns Maximum flat work group size for given subtarget \p STI. 192 unsigned getMaxFlatWorkGroupSize(const MCSubtargetInfo *STI); 193 194 /// \returns Number of waves per work group for given subtarget \p STI and 195 /// \p FlatWorkGroupSize. 196 unsigned getWavesPerWorkGroup(const MCSubtargetInfo *STI, 197 unsigned FlatWorkGroupSize); 198 199 /// \returns SGPR allocation granularity for given subtarget \p STI. 200 unsigned getSGPRAllocGranule(const MCSubtargetInfo *STI); 201 202 /// \returns SGPR encoding granularity for given subtarget \p STI. 203 unsigned getSGPREncodingGranule(const MCSubtargetInfo *STI); 204 205 /// \returns Total number of SGPRs for given subtarget \p STI. 206 unsigned getTotalNumSGPRs(const MCSubtargetInfo *STI); 207 208 /// \returns Addressable number of SGPRs for given subtarget \p STI. 209 unsigned getAddressableNumSGPRs(const MCSubtargetInfo *STI); 210 211 /// \returns Minimum number of SGPRs that meets the given number of waves per 212 /// execution unit requirement for given subtarget \p STI. 213 unsigned getMinNumSGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU); 214 215 /// \returns Maximum number of SGPRs that meets the given number of waves per 216 /// execution unit requirement for given subtarget \p STI. 217 unsigned getMaxNumSGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU, 218 bool Addressable); 219 220 /// \returns Number of extra SGPRs implicitly required by given subtarget \p 221 /// STI when the given special registers are used. 222 unsigned getNumExtraSGPRs(const MCSubtargetInfo *STI, bool VCCUsed, 223 bool FlatScrUsed, bool XNACKUsed); 224 225 /// \returns Number of extra SGPRs implicitly required by given subtarget \p 226 /// STI when the given special registers are used. XNACK is inferred from 227 /// \p STI. 228 unsigned getNumExtraSGPRs(const MCSubtargetInfo *STI, bool VCCUsed, 229 bool FlatScrUsed); 230 231 /// \returns Number of SGPR blocks needed for given subtarget \p STI when 232 /// \p NumSGPRs are used. \p NumSGPRs should already include any special 233 /// register counts. 234 unsigned getNumSGPRBlocks(const MCSubtargetInfo *STI, unsigned NumSGPRs); 235 236 /// \returns VGPR allocation granularity for given subtarget \p STI. 237 /// 238 /// For subtargets which support it, \p EnableWavefrontSize32 should match 239 /// the ENABLE_WAVEFRONT_SIZE32 kernel descriptor field. 240 unsigned getVGPRAllocGranule(const MCSubtargetInfo *STI, 241 Optional<bool> EnableWavefrontSize32 = None); 242 243 /// \returns VGPR encoding granularity for given subtarget \p STI. 244 /// 245 /// For subtargets which support it, \p EnableWavefrontSize32 should match 246 /// the ENABLE_WAVEFRONT_SIZE32 kernel descriptor field. 247 unsigned getVGPREncodingGranule(const MCSubtargetInfo *STI, 248 Optional<bool> EnableWavefrontSize32 = None); 249 250 /// \returns Total number of VGPRs for given subtarget \p STI. 251 unsigned getTotalNumVGPRs(const MCSubtargetInfo *STI); 252 253 /// \returns Addressable number of VGPRs for given subtarget \p STI. 254 unsigned getAddressableNumVGPRs(const MCSubtargetInfo *STI); 255 256 /// \returns Minimum number of VGPRs that meets given number of waves per 257 /// execution unit requirement for given subtarget \p STI. 258 unsigned getMinNumVGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU); 259 260 /// \returns Maximum number of VGPRs that meets given number of waves per 261 /// execution unit requirement for given subtarget \p STI. 262 unsigned getMaxNumVGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU); 263 264 /// \returns Number of VGPR blocks needed for given subtarget \p STI when 265 /// \p NumVGPRs are used. 266 /// 267 /// For subtargets which support it, \p EnableWavefrontSize32 should match the 268 /// ENABLE_WAVEFRONT_SIZE32 kernel descriptor field. 269 unsigned getNumVGPRBlocks(const MCSubtargetInfo *STI, unsigned NumSGPRs, 270 Optional<bool> EnableWavefrontSize32 = None); 271 272 } // end namespace IsaInfo 273 274 LLVM_READONLY 275 int16_t getNamedOperandIdx(uint16_t Opcode, uint16_t NamedIdx); 276 277 LLVM_READONLY 278 int getSOPPWithRelaxation(uint16_t Opcode); 279 280 struct MIMGBaseOpcodeInfo { 281 MIMGBaseOpcode BaseOpcode; 282 bool Store; 283 bool Atomic; 284 bool AtomicX2; 285 bool Sampler; 286 bool Gather4; 287 288 uint8_t NumExtraArgs; 289 bool Gradients; 290 bool G16; 291 bool Coordinates; 292 bool LodOrClampOrMip; 293 bool HasD16; 294 bool MSAA; 295 }; 296 297 LLVM_READONLY 298 const MIMGBaseOpcodeInfo *getMIMGBaseOpcodeInfo(unsigned BaseOpcode); 299 300 struct MIMGDimInfo { 301 MIMGDim Dim; 302 uint8_t NumCoords; 303 uint8_t NumGradients; 304 bool MSAA; 305 bool DA; 306 uint8_t Encoding; 307 const char *AsmSuffix; 308 }; 309 310 LLVM_READONLY 311 const MIMGDimInfo *getMIMGDimInfo(unsigned DimEnum); 312 313 LLVM_READONLY 314 const MIMGDimInfo *getMIMGDimInfoByEncoding(uint8_t DimEnc); 315 316 LLVM_READONLY 317 const MIMGDimInfo *getMIMGDimInfoByAsmSuffix(StringRef AsmSuffix); 318 319 struct MIMGLZMappingInfo { 320 MIMGBaseOpcode L; 321 MIMGBaseOpcode LZ; 322 }; 323 324 struct MIMGMIPMappingInfo { 325 MIMGBaseOpcode MIP; 326 MIMGBaseOpcode NONMIP; 327 }; 328 329 struct MIMGG16MappingInfo { 330 MIMGBaseOpcode G; 331 MIMGBaseOpcode G16; 332 }; 333 334 LLVM_READONLY 335 const MIMGLZMappingInfo *getMIMGLZMappingInfo(unsigned L); 336 337 LLVM_READONLY 338 const MIMGMIPMappingInfo *getMIMGMIPMappingInfo(unsigned MIP); 339 340 LLVM_READONLY 341 const MIMGG16MappingInfo *getMIMGG16MappingInfo(unsigned G); 342 343 LLVM_READONLY 344 int getMIMGOpcode(unsigned BaseOpcode, unsigned MIMGEncoding, 345 unsigned VDataDwords, unsigned VAddrDwords); 346 347 LLVM_READONLY 348 int getMaskedMIMGOp(unsigned Opc, unsigned NewChannels); 349 350 struct MIMGInfo { 351 uint16_t Opcode; 352 uint16_t BaseOpcode; 353 uint8_t MIMGEncoding; 354 uint8_t VDataDwords; 355 uint8_t VAddrDwords; 356 }; 357 358 LLVM_READONLY 359 const MIMGInfo *getMIMGInfo(unsigned Opc); 360 361 LLVM_READONLY 362 int getMTBUFBaseOpcode(unsigned Opc); 363 364 LLVM_READONLY 365 int getMTBUFOpcode(unsigned BaseOpc, unsigned Elements); 366 367 LLVM_READONLY 368 int getMTBUFElements(unsigned Opc); 369 370 LLVM_READONLY 371 bool getMTBUFHasVAddr(unsigned Opc); 372 373 LLVM_READONLY 374 bool getMTBUFHasSrsrc(unsigned Opc); 375 376 LLVM_READONLY 377 bool getMTBUFHasSoffset(unsigned Opc); 378 379 LLVM_READONLY 380 int getMUBUFBaseOpcode(unsigned Opc); 381 382 LLVM_READONLY 383 int getMUBUFOpcode(unsigned BaseOpc, unsigned Elements); 384 385 LLVM_READONLY 386 int getMUBUFElements(unsigned Opc); 387 388 LLVM_READONLY 389 bool getMUBUFHasVAddr(unsigned Opc); 390 391 LLVM_READONLY 392 bool getMUBUFHasSrsrc(unsigned Opc); 393 394 LLVM_READONLY 395 bool getMUBUFHasSoffset(unsigned Opc); 396 397 LLVM_READONLY 398 bool getSMEMIsBuffer(unsigned Opc); 399 400 LLVM_READONLY 401 const GcnBufferFormatInfo *getGcnBufferFormatInfo(uint8_t BitsPerComp, 402 uint8_t NumComponents, 403 uint8_t NumFormat, 404 const MCSubtargetInfo &STI); 405 LLVM_READONLY 406 const GcnBufferFormatInfo *getGcnBufferFormatInfo(uint8_t Format, 407 const MCSubtargetInfo &STI); 408 409 LLVM_READONLY 410 int getMCOpcode(uint16_t Opcode, unsigned Gen); 411 412 void initDefaultAMDKernelCodeT(amd_kernel_code_t &Header, 413 const MCSubtargetInfo *STI); 414 415 amdhsa::kernel_descriptor_t getDefaultAmdhsaKernelDescriptor( 416 const MCSubtargetInfo *STI); 417 418 bool isGroupSegment(const GlobalValue *GV); 419 bool isGlobalSegment(const GlobalValue *GV); 420 bool isReadOnlySegment(const GlobalValue *GV); 421 422 /// \returns True if constants should be emitted to .text section for given 423 /// target triple \p TT, false otherwise. 424 bool shouldEmitConstantsToTextSection(const Triple &TT); 425 426 /// \returns Integer value requested using \p F's \p Name attribute. 427 /// 428 /// \returns \p Default if attribute is not present. 429 /// 430 /// \returns \p Default and emits error if requested value cannot be converted 431 /// to integer. 432 int getIntegerAttribute(const Function &F, StringRef Name, int Default); 433 434 /// \returns A pair of integer values requested using \p F's \p Name attribute 435 /// in "first[,second]" format ("second" is optional unless \p OnlyFirstRequired 436 /// is false). 437 /// 438 /// \returns \p Default if attribute is not present. 439 /// 440 /// \returns \p Default and emits error if one of the requested values cannot be 441 /// converted to integer, or \p OnlyFirstRequired is false and "second" value is 442 /// not present. 443 std::pair<int, int> getIntegerPairAttribute(const Function &F, 444 StringRef Name, 445 std::pair<int, int> Default, 446 bool OnlyFirstRequired = false); 447 448 /// Represents the counter values to wait for in an s_waitcnt instruction. 449 /// 450 /// Large values (including the maximum possible integer) can be used to 451 /// represent "don't care" waits. 452 struct Waitcnt { 453 unsigned VmCnt = ~0u; 454 unsigned ExpCnt = ~0u; 455 unsigned LgkmCnt = ~0u; 456 unsigned VsCnt = ~0u; 457 458 Waitcnt() {} 459 Waitcnt(unsigned VmCnt, unsigned ExpCnt, unsigned LgkmCnt, unsigned VsCnt) 460 : VmCnt(VmCnt), ExpCnt(ExpCnt), LgkmCnt(LgkmCnt), VsCnt(VsCnt) {} 461 462 static Waitcnt allZero(bool HasVscnt) { 463 return Waitcnt(0, 0, 0, HasVscnt ? 0 : ~0u); 464 } 465 static Waitcnt allZeroExceptVsCnt() { return Waitcnt(0, 0, 0, ~0u); } 466 467 bool hasWait() const { 468 return VmCnt != ~0u || ExpCnt != ~0u || LgkmCnt != ~0u || VsCnt != ~0u; 469 } 470 471 bool dominates(const Waitcnt &Other) const { 472 return VmCnt <= Other.VmCnt && ExpCnt <= Other.ExpCnt && 473 LgkmCnt <= Other.LgkmCnt && VsCnt <= Other.VsCnt; 474 } 475 476 Waitcnt combined(const Waitcnt &Other) const { 477 return Waitcnt(std::min(VmCnt, Other.VmCnt), std::min(ExpCnt, Other.ExpCnt), 478 std::min(LgkmCnt, Other.LgkmCnt), 479 std::min(VsCnt, Other.VsCnt)); 480 } 481 }; 482 483 /// \returns Vmcnt bit mask for given isa \p Version. 484 unsigned getVmcntBitMask(const IsaVersion &Version); 485 486 /// \returns Expcnt bit mask for given isa \p Version. 487 unsigned getExpcntBitMask(const IsaVersion &Version); 488 489 /// \returns Lgkmcnt bit mask for given isa \p Version. 490 unsigned getLgkmcntBitMask(const IsaVersion &Version); 491 492 /// \returns Waitcnt bit mask for given isa \p Version. 493 unsigned getWaitcntBitMask(const IsaVersion &Version); 494 495 /// \returns Decoded Vmcnt from given \p Waitcnt for given isa \p Version. 496 unsigned decodeVmcnt(const IsaVersion &Version, unsigned Waitcnt); 497 498 /// \returns Decoded Expcnt from given \p Waitcnt for given isa \p Version. 499 unsigned decodeExpcnt(const IsaVersion &Version, unsigned Waitcnt); 500 501 /// \returns Decoded Lgkmcnt from given \p Waitcnt for given isa \p Version. 502 unsigned decodeLgkmcnt(const IsaVersion &Version, unsigned Waitcnt); 503 504 /// Decodes Vmcnt, Expcnt and Lgkmcnt from given \p Waitcnt for given isa 505 /// \p Version, and writes decoded values into \p Vmcnt, \p Expcnt and 506 /// \p Lgkmcnt respectively. 507 /// 508 /// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are decoded as follows: 509 /// \p Vmcnt = \p Waitcnt[3:0] (pre-gfx9 only) 510 /// \p Vmcnt = \p Waitcnt[3:0] | \p Waitcnt[15:14] (gfx9+ only) 511 /// \p Expcnt = \p Waitcnt[6:4] 512 /// \p Lgkmcnt = \p Waitcnt[11:8] (pre-gfx10 only) 513 /// \p Lgkmcnt = \p Waitcnt[13:8] (gfx10+ only) 514 void decodeWaitcnt(const IsaVersion &Version, unsigned Waitcnt, 515 unsigned &Vmcnt, unsigned &Expcnt, unsigned &Lgkmcnt); 516 517 Waitcnt decodeWaitcnt(const IsaVersion &Version, unsigned Encoded); 518 519 /// \returns \p Waitcnt with encoded \p Vmcnt for given isa \p Version. 520 unsigned encodeVmcnt(const IsaVersion &Version, unsigned Waitcnt, 521 unsigned Vmcnt); 522 523 /// \returns \p Waitcnt with encoded \p Expcnt for given isa \p Version. 524 unsigned encodeExpcnt(const IsaVersion &Version, unsigned Waitcnt, 525 unsigned Expcnt); 526 527 /// \returns \p Waitcnt with encoded \p Lgkmcnt for given isa \p Version. 528 unsigned encodeLgkmcnt(const IsaVersion &Version, unsigned Waitcnt, 529 unsigned Lgkmcnt); 530 531 /// Encodes \p Vmcnt, \p Expcnt and \p Lgkmcnt into Waitcnt for given isa 532 /// \p Version. 533 /// 534 /// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are encoded as follows: 535 /// Waitcnt[3:0] = \p Vmcnt (pre-gfx9 only) 536 /// Waitcnt[3:0] = \p Vmcnt[3:0] (gfx9+ only) 537 /// Waitcnt[6:4] = \p Expcnt 538 /// Waitcnt[11:8] = \p Lgkmcnt (pre-gfx10 only) 539 /// Waitcnt[13:8] = \p Lgkmcnt (gfx10+ only) 540 /// Waitcnt[15:14] = \p Vmcnt[5:4] (gfx9+ only) 541 /// 542 /// \returns Waitcnt with encoded \p Vmcnt, \p Expcnt and \p Lgkmcnt for given 543 /// isa \p Version. 544 unsigned encodeWaitcnt(const IsaVersion &Version, 545 unsigned Vmcnt, unsigned Expcnt, unsigned Lgkmcnt); 546 547 unsigned encodeWaitcnt(const IsaVersion &Version, const Waitcnt &Decoded); 548 549 namespace Hwreg { 550 551 LLVM_READONLY 552 int64_t getHwregId(const StringRef Name); 553 554 LLVM_READNONE 555 bool isValidHwreg(int64_t Id, const MCSubtargetInfo &STI); 556 557 LLVM_READNONE 558 bool isValidHwreg(int64_t Id); 559 560 LLVM_READNONE 561 bool isValidHwregOffset(int64_t Offset); 562 563 LLVM_READNONE 564 bool isValidHwregWidth(int64_t Width); 565 566 LLVM_READNONE 567 uint64_t encodeHwreg(uint64_t Id, uint64_t Offset, uint64_t Width); 568 569 LLVM_READNONE 570 StringRef getHwreg(unsigned Id, const MCSubtargetInfo &STI); 571 572 void decodeHwreg(unsigned Val, unsigned &Id, unsigned &Offset, unsigned &Width); 573 574 } // namespace Hwreg 575 576 namespace Exp { 577 578 bool getTgtName(unsigned Id, StringRef &Name, int &Index); 579 580 LLVM_READONLY 581 unsigned getTgtId(const StringRef Name); 582 583 LLVM_READNONE 584 bool isSupportedTgtId(unsigned Id, const MCSubtargetInfo &STI); 585 586 } // namespace Exp 587 588 namespace MTBUFFormat { 589 590 LLVM_READNONE 591 int64_t encodeDfmtNfmt(unsigned Dfmt, unsigned Nfmt); 592 593 void decodeDfmtNfmt(unsigned Format, unsigned &Dfmt, unsigned &Nfmt); 594 595 int64_t getDfmt(const StringRef Name); 596 597 StringRef getDfmtName(unsigned Id); 598 599 int64_t getNfmt(const StringRef Name, const MCSubtargetInfo &STI); 600 601 StringRef getNfmtName(unsigned Id, const MCSubtargetInfo &STI); 602 603 bool isValidDfmtNfmt(unsigned Val, const MCSubtargetInfo &STI); 604 605 bool isValidNfmt(unsigned Val, const MCSubtargetInfo &STI); 606 607 int64_t getUnifiedFormat(const StringRef Name); 608 609 StringRef getUnifiedFormatName(unsigned Id); 610 611 bool isValidUnifiedFormat(unsigned Val); 612 613 int64_t convertDfmtNfmt2Ufmt(unsigned Dfmt, unsigned Nfmt); 614 615 bool isValidFormatEncoding(unsigned Val, const MCSubtargetInfo &STI); 616 617 unsigned getDefaultFormatEncoding(const MCSubtargetInfo &STI); 618 619 } // namespace MTBUFFormat 620 621 namespace SendMsg { 622 623 LLVM_READONLY 624 int64_t getMsgId(const StringRef Name); 625 626 LLVM_READONLY 627 int64_t getMsgOpId(int64_t MsgId, const StringRef Name); 628 629 LLVM_READNONE 630 StringRef getMsgName(int64_t MsgId); 631 632 LLVM_READNONE 633 StringRef getMsgOpName(int64_t MsgId, int64_t OpId); 634 635 LLVM_READNONE 636 bool isValidMsgId(int64_t MsgId, const MCSubtargetInfo &STI, bool Strict = true); 637 638 LLVM_READNONE 639 bool isValidMsgOp(int64_t MsgId, int64_t OpId, const MCSubtargetInfo &STI, 640 bool Strict = true); 641 642 LLVM_READNONE 643 bool isValidMsgStream(int64_t MsgId, int64_t OpId, int64_t StreamId, 644 const MCSubtargetInfo &STI, bool Strict = true); 645 646 LLVM_READNONE 647 bool msgRequiresOp(int64_t MsgId); 648 649 LLVM_READNONE 650 bool msgSupportsStream(int64_t MsgId, int64_t OpId); 651 652 void decodeMsg(unsigned Val, 653 uint16_t &MsgId, 654 uint16_t &OpId, 655 uint16_t &StreamId); 656 657 LLVM_READNONE 658 uint64_t encodeMsg(uint64_t MsgId, 659 uint64_t OpId, 660 uint64_t StreamId); 661 662 } // namespace SendMsg 663 664 665 unsigned getInitialPSInputAddr(const Function &F); 666 667 LLVM_READNONE 668 bool isShader(CallingConv::ID CC); 669 670 LLVM_READNONE 671 bool isGraphics(CallingConv::ID CC); 672 673 LLVM_READNONE 674 bool isCompute(CallingConv::ID CC); 675 676 LLVM_READNONE 677 bool isEntryFunctionCC(CallingConv::ID CC); 678 679 // These functions are considered entrypoints into the current module, i.e. they 680 // are allowed to be called from outside the current module. This is different 681 // from isEntryFunctionCC, which is only true for functions that are entered by 682 // the hardware. Module entry points include all entry functions but also 683 // include functions that can be called from other functions inside or outside 684 // the current module. Module entry functions are allowed to allocate LDS. 685 LLVM_READNONE 686 bool isModuleEntryFunctionCC(CallingConv::ID CC); 687 688 // FIXME: Remove this when calling conventions cleaned up 689 LLVM_READNONE 690 inline bool isKernel(CallingConv::ID CC) { 691 switch (CC) { 692 case CallingConv::AMDGPU_KERNEL: 693 case CallingConv::SPIR_KERNEL: 694 return true; 695 default: 696 return false; 697 } 698 } 699 700 bool hasXNACK(const MCSubtargetInfo &STI); 701 bool hasSRAMECC(const MCSubtargetInfo &STI); 702 bool hasMIMG_R128(const MCSubtargetInfo &STI); 703 bool hasGFX10A16(const MCSubtargetInfo &STI); 704 bool hasG16(const MCSubtargetInfo &STI); 705 bool hasPackedD16(const MCSubtargetInfo &STI); 706 707 bool isSI(const MCSubtargetInfo &STI); 708 bool isCI(const MCSubtargetInfo &STI); 709 bool isVI(const MCSubtargetInfo &STI); 710 bool isGFX9(const MCSubtargetInfo &STI); 711 bool isGFX9Plus(const MCSubtargetInfo &STI); 712 bool isGFX10(const MCSubtargetInfo &STI); 713 bool isGFX10Plus(const MCSubtargetInfo &STI); 714 bool isGCN3Encoding(const MCSubtargetInfo &STI); 715 bool isGFX10_BEncoding(const MCSubtargetInfo &STI); 716 bool hasGFX10_3Insts(const MCSubtargetInfo &STI); 717 bool isGFX90A(const MCSubtargetInfo &STI); 718 719 /// Is Reg - scalar register 720 bool isSGPR(unsigned Reg, const MCRegisterInfo* TRI); 721 722 /// Is there any intersection between registers 723 bool isRegIntersect(unsigned Reg0, unsigned Reg1, const MCRegisterInfo* TRI); 724 725 /// If \p Reg is a pseudo reg, return the correct hardware register given 726 /// \p STI otherwise return \p Reg. 727 unsigned getMCReg(unsigned Reg, const MCSubtargetInfo &STI); 728 729 /// Convert hardware register \p Reg to a pseudo register 730 LLVM_READNONE 731 unsigned mc2PseudoReg(unsigned Reg); 732 733 /// Can this operand also contain immediate values? 734 bool isSISrcOperand(const MCInstrDesc &Desc, unsigned OpNo); 735 736 /// Is this floating-point operand? 737 bool isSISrcFPOperand(const MCInstrDesc &Desc, unsigned OpNo); 738 739 /// Does this opearnd support only inlinable literals? 740 bool isSISrcInlinableOperand(const MCInstrDesc &Desc, unsigned OpNo); 741 742 /// Get the size in bits of a register from the register class \p RC. 743 unsigned getRegBitWidth(unsigned RCID); 744 745 /// Get the size in bits of a register from the register class \p RC. 746 unsigned getRegBitWidth(const MCRegisterClass &RC); 747 748 /// Get size of register operand 749 unsigned getRegOperandSize(const MCRegisterInfo *MRI, const MCInstrDesc &Desc, 750 unsigned OpNo); 751 752 LLVM_READNONE 753 inline unsigned getOperandSize(const MCOperandInfo &OpInfo) { 754 switch (OpInfo.OperandType) { 755 case AMDGPU::OPERAND_REG_IMM_INT32: 756 case AMDGPU::OPERAND_REG_IMM_FP32: 757 case AMDGPU::OPERAND_REG_INLINE_C_INT32: 758 case AMDGPU::OPERAND_REG_INLINE_C_FP32: 759 case AMDGPU::OPERAND_REG_INLINE_AC_INT32: 760 case AMDGPU::OPERAND_REG_INLINE_AC_FP32: 761 case AMDGPU::OPERAND_REG_IMM_V2INT32: 762 case AMDGPU::OPERAND_REG_IMM_V2FP32: 763 case AMDGPU::OPERAND_REG_INLINE_C_V2INT32: 764 case AMDGPU::OPERAND_REG_INLINE_C_V2FP32: 765 return 4; 766 767 case AMDGPU::OPERAND_REG_IMM_INT64: 768 case AMDGPU::OPERAND_REG_IMM_FP64: 769 case AMDGPU::OPERAND_REG_INLINE_C_INT64: 770 case AMDGPU::OPERAND_REG_INLINE_C_FP64: 771 case AMDGPU::OPERAND_REG_INLINE_AC_FP64: 772 return 8; 773 774 case AMDGPU::OPERAND_REG_IMM_INT16: 775 case AMDGPU::OPERAND_REG_IMM_FP16: 776 case AMDGPU::OPERAND_REG_INLINE_C_INT16: 777 case AMDGPU::OPERAND_REG_INLINE_C_FP16: 778 case AMDGPU::OPERAND_REG_INLINE_C_V2INT16: 779 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16: 780 case AMDGPU::OPERAND_REG_INLINE_AC_INT16: 781 case AMDGPU::OPERAND_REG_INLINE_AC_FP16: 782 case AMDGPU::OPERAND_REG_INLINE_AC_V2INT16: 783 case AMDGPU::OPERAND_REG_INLINE_AC_V2FP16: 784 case AMDGPU::OPERAND_REG_IMM_V2INT16: 785 case AMDGPU::OPERAND_REG_IMM_V2FP16: 786 return 2; 787 788 default: 789 llvm_unreachable("unhandled operand type"); 790 } 791 } 792 793 LLVM_READNONE 794 inline unsigned getOperandSize(const MCInstrDesc &Desc, unsigned OpNo) { 795 return getOperandSize(Desc.OpInfo[OpNo]); 796 } 797 798 /// Is this literal inlinable, and not one of the values intended for floating 799 /// point values. 800 LLVM_READNONE 801 inline bool isInlinableIntLiteral(int64_t Literal) { 802 return Literal >= -16 && Literal <= 64; 803 } 804 805 /// Is this literal inlinable 806 LLVM_READNONE 807 bool isInlinableLiteral64(int64_t Literal, bool HasInv2Pi); 808 809 LLVM_READNONE 810 bool isInlinableLiteral32(int32_t Literal, bool HasInv2Pi); 811 812 LLVM_READNONE 813 bool isInlinableLiteral16(int16_t Literal, bool HasInv2Pi); 814 815 LLVM_READNONE 816 bool isInlinableLiteralV216(int32_t Literal, bool HasInv2Pi); 817 818 LLVM_READNONE 819 bool isInlinableIntLiteralV216(int32_t Literal); 820 821 LLVM_READNONE 822 bool isFoldableLiteralV216(int32_t Literal, bool HasInv2Pi); 823 824 bool isArgPassedInSGPR(const Argument *Arg); 825 826 LLVM_READONLY 827 bool isLegalSMRDEncodedUnsignedOffset(const MCSubtargetInfo &ST, 828 int64_t EncodedOffset); 829 830 LLVM_READONLY 831 bool isLegalSMRDEncodedSignedOffset(const MCSubtargetInfo &ST, 832 int64_t EncodedOffset, 833 bool IsBuffer); 834 835 /// Convert \p ByteOffset to dwords if the subtarget uses dword SMRD immediate 836 /// offsets. 837 uint64_t convertSMRDOffsetUnits(const MCSubtargetInfo &ST, uint64_t ByteOffset); 838 839 /// \returns The encoding that will be used for \p ByteOffset in the 840 /// SMRD offset field, or None if it won't fit. On GFX9 and GFX10 841 /// S_LOAD instructions have a signed offset, on other subtargets it is 842 /// unsigned. S_BUFFER has an unsigned offset for all subtargets. 843 Optional<int64_t> getSMRDEncodedOffset(const MCSubtargetInfo &ST, 844 int64_t ByteOffset, bool IsBuffer); 845 846 /// \return The encoding that can be used for a 32-bit literal offset in an SMRD 847 /// instruction. This is only useful on CI.s 848 Optional<int64_t> getSMRDEncodedLiteralOffset32(const MCSubtargetInfo &ST, 849 int64_t ByteOffset); 850 851 /// For FLAT segment the offset must be positive; 852 /// MSB is ignored and forced to zero. 853 /// 854 /// \return The number of bits available for the offset field in flat 855 /// instructions. 856 unsigned getNumFlatOffsetBits(const MCSubtargetInfo &ST, bool Signed); 857 858 /// \returns true if this offset is small enough to fit in the SMRD 859 /// offset field. \p ByteOffset should be the offset in bytes and 860 /// not the encoded offset. 861 bool isLegalSMRDImmOffset(const MCSubtargetInfo &ST, int64_t ByteOffset); 862 863 bool splitMUBUFOffset(uint32_t Imm, uint32_t &SOffset, uint32_t &ImmOffset, 864 const GCNSubtarget *Subtarget, 865 Align Alignment = Align(4)); 866 867 LLVM_READNONE 868 inline bool isLegal64BitDPPControl(unsigned DC) { 869 return DC >= DPP::ROW_NEWBCAST_FIRST && DC <= DPP::ROW_NEWBCAST_LAST; 870 } 871 872 /// \returns true if the intrinsic is divergent 873 bool isIntrinsicSourceOfDivergence(unsigned IntrID); 874 875 // Track defaults for fields in the MODE registser. 876 struct SIModeRegisterDefaults { 877 /// Floating point opcodes that support exception flag gathering quiet and 878 /// propagate signaling NaN inputs per IEEE 754-2008. Min_dx10 and max_dx10 879 /// become IEEE 754- 2008 compliant due to signaling NaN propagation and 880 /// quieting. 881 bool IEEE : 1; 882 883 /// Used by the vector ALU to force DX10-style treatment of NaNs: when set, 884 /// clamp NaN to zero; otherwise, pass NaN through. 885 bool DX10Clamp : 1; 886 887 /// If this is set, neither input or output denormals are flushed for most f32 888 /// instructions. 889 bool FP32InputDenormals : 1; 890 bool FP32OutputDenormals : 1; 891 892 /// If this is set, neither input or output denormals are flushed for both f64 893 /// and f16/v2f16 instructions. 894 bool FP64FP16InputDenormals : 1; 895 bool FP64FP16OutputDenormals : 1; 896 897 SIModeRegisterDefaults() : 898 IEEE(true), 899 DX10Clamp(true), 900 FP32InputDenormals(true), 901 FP32OutputDenormals(true), 902 FP64FP16InputDenormals(true), 903 FP64FP16OutputDenormals(true) {} 904 905 SIModeRegisterDefaults(const Function &F); 906 907 static SIModeRegisterDefaults getDefaultForCallingConv(CallingConv::ID CC) { 908 SIModeRegisterDefaults Mode; 909 Mode.IEEE = !AMDGPU::isShader(CC); 910 return Mode; 911 } 912 913 bool operator ==(const SIModeRegisterDefaults Other) const { 914 return IEEE == Other.IEEE && DX10Clamp == Other.DX10Clamp && 915 FP32InputDenormals == Other.FP32InputDenormals && 916 FP32OutputDenormals == Other.FP32OutputDenormals && 917 FP64FP16InputDenormals == Other.FP64FP16InputDenormals && 918 FP64FP16OutputDenormals == Other.FP64FP16OutputDenormals; 919 } 920 921 bool allFP32Denormals() const { 922 return FP32InputDenormals && FP32OutputDenormals; 923 } 924 925 bool allFP64FP16Denormals() const { 926 return FP64FP16InputDenormals && FP64FP16OutputDenormals; 927 } 928 929 /// Get the encoding value for the FP_DENORM bits of the mode register for the 930 /// FP32 denormal mode. 931 uint32_t fpDenormModeSPValue() const { 932 if (FP32InputDenormals && FP32OutputDenormals) 933 return FP_DENORM_FLUSH_NONE; 934 if (FP32InputDenormals) 935 return FP_DENORM_FLUSH_OUT; 936 if (FP32OutputDenormals) 937 return FP_DENORM_FLUSH_IN; 938 return FP_DENORM_FLUSH_IN_FLUSH_OUT; 939 } 940 941 /// Get the encoding value for the FP_DENORM bits of the mode register for the 942 /// FP64/FP16 denormal mode. 943 uint32_t fpDenormModeDPValue() const { 944 if (FP64FP16InputDenormals && FP64FP16OutputDenormals) 945 return FP_DENORM_FLUSH_NONE; 946 if (FP64FP16InputDenormals) 947 return FP_DENORM_FLUSH_OUT; 948 if (FP64FP16OutputDenormals) 949 return FP_DENORM_FLUSH_IN; 950 return FP_DENORM_FLUSH_IN_FLUSH_OUT; 951 } 952 953 /// Returns true if a flag is compatible if it's enabled in the callee, but 954 /// disabled in the caller. 955 static bool oneWayCompatible(bool CallerMode, bool CalleeMode) { 956 return CallerMode == CalleeMode || (!CallerMode && CalleeMode); 957 } 958 959 // FIXME: Inlining should be OK for dx10-clamp, since the caller's mode should 960 // be able to override. 961 bool isInlineCompatible(SIModeRegisterDefaults CalleeMode) const { 962 if (DX10Clamp != CalleeMode.DX10Clamp) 963 return false; 964 if (IEEE != CalleeMode.IEEE) 965 return false; 966 967 // Allow inlining denormals enabled into denormals flushed functions. 968 return oneWayCompatible(FP64FP16InputDenormals, CalleeMode.FP64FP16InputDenormals) && 969 oneWayCompatible(FP64FP16OutputDenormals, CalleeMode.FP64FP16OutputDenormals) && 970 oneWayCompatible(FP32InputDenormals, CalleeMode.FP32InputDenormals) && 971 oneWayCompatible(FP32OutputDenormals, CalleeMode.FP32OutputDenormals); 972 } 973 }; 974 975 } // end namespace AMDGPU 976 977 raw_ostream &operator<<(raw_ostream &OS, 978 const AMDGPU::IsaInfo::TargetIDSetting S); 979 980 } // end namespace llvm 981 982 #endif // LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUBASEINFO_H 983