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 bool getVOP1IsSingle(unsigned Opc); 402 403 LLVM_READONLY 404 bool getVOP2IsSingle(unsigned Opc); 405 406 LLVM_READONLY 407 bool getVOP3IsSingle(unsigned Opc); 408 409 LLVM_READONLY 410 const GcnBufferFormatInfo *getGcnBufferFormatInfo(uint8_t BitsPerComp, 411 uint8_t NumComponents, 412 uint8_t NumFormat, 413 const MCSubtargetInfo &STI); 414 LLVM_READONLY 415 const GcnBufferFormatInfo *getGcnBufferFormatInfo(uint8_t Format, 416 const MCSubtargetInfo &STI); 417 418 LLVM_READONLY 419 int getMCOpcode(uint16_t Opcode, unsigned Gen); 420 421 void initDefaultAMDKernelCodeT(amd_kernel_code_t &Header, 422 const MCSubtargetInfo *STI); 423 424 amdhsa::kernel_descriptor_t getDefaultAmdhsaKernelDescriptor( 425 const MCSubtargetInfo *STI); 426 427 bool isGroupSegment(const GlobalValue *GV); 428 bool isGlobalSegment(const GlobalValue *GV); 429 bool isReadOnlySegment(const GlobalValue *GV); 430 431 /// \returns True if constants should be emitted to .text section for given 432 /// target triple \p TT, false otherwise. 433 bool shouldEmitConstantsToTextSection(const Triple &TT); 434 435 /// \returns Integer value requested using \p F's \p Name attribute. 436 /// 437 /// \returns \p Default if attribute is not present. 438 /// 439 /// \returns \p Default and emits error if requested value cannot be converted 440 /// to integer. 441 int getIntegerAttribute(const Function &F, StringRef Name, int Default); 442 443 /// \returns A pair of integer values requested using \p F's \p Name attribute 444 /// in "first[,second]" format ("second" is optional unless \p OnlyFirstRequired 445 /// is false). 446 /// 447 /// \returns \p Default if attribute is not present. 448 /// 449 /// \returns \p Default and emits error if one of the requested values cannot be 450 /// converted to integer, or \p OnlyFirstRequired is false and "second" value is 451 /// not present. 452 std::pair<int, int> getIntegerPairAttribute(const Function &F, 453 StringRef Name, 454 std::pair<int, int> Default, 455 bool OnlyFirstRequired = false); 456 457 /// Represents the counter values to wait for in an s_waitcnt instruction. 458 /// 459 /// Large values (including the maximum possible integer) can be used to 460 /// represent "don't care" waits. 461 struct Waitcnt { 462 unsigned VmCnt = ~0u; 463 unsigned ExpCnt = ~0u; 464 unsigned LgkmCnt = ~0u; 465 unsigned VsCnt = ~0u; 466 467 Waitcnt() {} 468 Waitcnt(unsigned VmCnt, unsigned ExpCnt, unsigned LgkmCnt, unsigned VsCnt) 469 : VmCnt(VmCnt), ExpCnt(ExpCnt), LgkmCnt(LgkmCnt), VsCnt(VsCnt) {} 470 471 static Waitcnt allZero(bool HasVscnt) { 472 return Waitcnt(0, 0, 0, HasVscnt ? 0 : ~0u); 473 } 474 static Waitcnt allZeroExceptVsCnt() { return Waitcnt(0, 0, 0, ~0u); } 475 476 bool hasWait() const { 477 return VmCnt != ~0u || ExpCnt != ~0u || LgkmCnt != ~0u || VsCnt != ~0u; 478 } 479 480 bool dominates(const Waitcnt &Other) const { 481 return VmCnt <= Other.VmCnt && ExpCnt <= Other.ExpCnt && 482 LgkmCnt <= Other.LgkmCnt && VsCnt <= Other.VsCnt; 483 } 484 485 Waitcnt combined(const Waitcnt &Other) const { 486 return Waitcnt(std::min(VmCnt, Other.VmCnt), std::min(ExpCnt, Other.ExpCnt), 487 std::min(LgkmCnt, Other.LgkmCnt), 488 std::min(VsCnt, Other.VsCnt)); 489 } 490 }; 491 492 /// \returns Vmcnt bit mask for given isa \p Version. 493 unsigned getVmcntBitMask(const IsaVersion &Version); 494 495 /// \returns Expcnt bit mask for given isa \p Version. 496 unsigned getExpcntBitMask(const IsaVersion &Version); 497 498 /// \returns Lgkmcnt bit mask for given isa \p Version. 499 unsigned getLgkmcntBitMask(const IsaVersion &Version); 500 501 /// \returns Waitcnt bit mask for given isa \p Version. 502 unsigned getWaitcntBitMask(const IsaVersion &Version); 503 504 /// \returns Decoded Vmcnt from given \p Waitcnt for given isa \p Version. 505 unsigned decodeVmcnt(const IsaVersion &Version, unsigned Waitcnt); 506 507 /// \returns Decoded Expcnt from given \p Waitcnt for given isa \p Version. 508 unsigned decodeExpcnt(const IsaVersion &Version, unsigned Waitcnt); 509 510 /// \returns Decoded Lgkmcnt from given \p Waitcnt for given isa \p Version. 511 unsigned decodeLgkmcnt(const IsaVersion &Version, unsigned Waitcnt); 512 513 /// Decodes Vmcnt, Expcnt and Lgkmcnt from given \p Waitcnt for given isa 514 /// \p Version, and writes decoded values into \p Vmcnt, \p Expcnt and 515 /// \p Lgkmcnt respectively. 516 /// 517 /// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are decoded as follows: 518 /// \p Vmcnt = \p Waitcnt[3:0] (pre-gfx9 only) 519 /// \p Vmcnt = \p Waitcnt[3:0] | \p Waitcnt[15:14] (gfx9+ only) 520 /// \p Expcnt = \p Waitcnt[6:4] 521 /// \p Lgkmcnt = \p Waitcnt[11:8] (pre-gfx10 only) 522 /// \p Lgkmcnt = \p Waitcnt[13:8] (gfx10+ only) 523 void decodeWaitcnt(const IsaVersion &Version, unsigned Waitcnt, 524 unsigned &Vmcnt, unsigned &Expcnt, unsigned &Lgkmcnt); 525 526 Waitcnt decodeWaitcnt(const IsaVersion &Version, unsigned Encoded); 527 528 /// \returns \p Waitcnt with encoded \p Vmcnt for given isa \p Version. 529 unsigned encodeVmcnt(const IsaVersion &Version, unsigned Waitcnt, 530 unsigned Vmcnt); 531 532 /// \returns \p Waitcnt with encoded \p Expcnt for given isa \p Version. 533 unsigned encodeExpcnt(const IsaVersion &Version, unsigned Waitcnt, 534 unsigned Expcnt); 535 536 /// \returns \p Waitcnt with encoded \p Lgkmcnt for given isa \p Version. 537 unsigned encodeLgkmcnt(const IsaVersion &Version, unsigned Waitcnt, 538 unsigned Lgkmcnt); 539 540 /// Encodes \p Vmcnt, \p Expcnt and \p Lgkmcnt into Waitcnt for given isa 541 /// \p Version. 542 /// 543 /// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are encoded as follows: 544 /// Waitcnt[3:0] = \p Vmcnt (pre-gfx9 only) 545 /// Waitcnt[3:0] = \p Vmcnt[3:0] (gfx9+ only) 546 /// Waitcnt[6:4] = \p Expcnt 547 /// Waitcnt[11:8] = \p Lgkmcnt (pre-gfx10 only) 548 /// Waitcnt[13:8] = \p Lgkmcnt (gfx10+ only) 549 /// Waitcnt[15:14] = \p Vmcnt[5:4] (gfx9+ only) 550 /// 551 /// \returns Waitcnt with encoded \p Vmcnt, \p Expcnt and \p Lgkmcnt for given 552 /// isa \p Version. 553 unsigned encodeWaitcnt(const IsaVersion &Version, 554 unsigned Vmcnt, unsigned Expcnt, unsigned Lgkmcnt); 555 556 unsigned encodeWaitcnt(const IsaVersion &Version, const Waitcnt &Decoded); 557 558 namespace Hwreg { 559 560 LLVM_READONLY 561 int64_t getHwregId(const StringRef Name); 562 563 LLVM_READNONE 564 bool isValidHwreg(int64_t Id, const MCSubtargetInfo &STI); 565 566 LLVM_READNONE 567 bool isValidHwreg(int64_t Id); 568 569 LLVM_READNONE 570 bool isValidHwregOffset(int64_t Offset); 571 572 LLVM_READNONE 573 bool isValidHwregWidth(int64_t Width); 574 575 LLVM_READNONE 576 uint64_t encodeHwreg(uint64_t Id, uint64_t Offset, uint64_t Width); 577 578 LLVM_READNONE 579 StringRef getHwreg(unsigned Id, const MCSubtargetInfo &STI); 580 581 void decodeHwreg(unsigned Val, unsigned &Id, unsigned &Offset, unsigned &Width); 582 583 } // namespace Hwreg 584 585 namespace Exp { 586 587 bool getTgtName(unsigned Id, StringRef &Name, int &Index); 588 589 LLVM_READONLY 590 unsigned getTgtId(const StringRef Name); 591 592 LLVM_READNONE 593 bool isSupportedTgtId(unsigned Id, const MCSubtargetInfo &STI); 594 595 } // namespace Exp 596 597 namespace MTBUFFormat { 598 599 LLVM_READNONE 600 int64_t encodeDfmtNfmt(unsigned Dfmt, unsigned Nfmt); 601 602 void decodeDfmtNfmt(unsigned Format, unsigned &Dfmt, unsigned &Nfmt); 603 604 int64_t getDfmt(const StringRef Name); 605 606 StringRef getDfmtName(unsigned Id); 607 608 int64_t getNfmt(const StringRef Name, const MCSubtargetInfo &STI); 609 610 StringRef getNfmtName(unsigned Id, const MCSubtargetInfo &STI); 611 612 bool isValidDfmtNfmt(unsigned Val, const MCSubtargetInfo &STI); 613 614 bool isValidNfmt(unsigned Val, const MCSubtargetInfo &STI); 615 616 int64_t getUnifiedFormat(const StringRef Name); 617 618 StringRef getUnifiedFormatName(unsigned Id); 619 620 bool isValidUnifiedFormat(unsigned Val); 621 622 int64_t convertDfmtNfmt2Ufmt(unsigned Dfmt, unsigned Nfmt); 623 624 bool isValidFormatEncoding(unsigned Val, const MCSubtargetInfo &STI); 625 626 unsigned getDefaultFormatEncoding(const MCSubtargetInfo &STI); 627 628 } // namespace MTBUFFormat 629 630 namespace SendMsg { 631 632 LLVM_READONLY 633 int64_t getMsgId(const StringRef Name); 634 635 LLVM_READONLY 636 int64_t getMsgOpId(int64_t MsgId, const StringRef Name); 637 638 LLVM_READNONE 639 StringRef getMsgName(int64_t MsgId); 640 641 LLVM_READNONE 642 StringRef getMsgOpName(int64_t MsgId, int64_t OpId); 643 644 LLVM_READNONE 645 bool isValidMsgId(int64_t MsgId, const MCSubtargetInfo &STI, bool Strict = true); 646 647 LLVM_READNONE 648 bool isValidMsgOp(int64_t MsgId, int64_t OpId, const MCSubtargetInfo &STI, 649 bool Strict = true); 650 651 LLVM_READNONE 652 bool isValidMsgStream(int64_t MsgId, int64_t OpId, int64_t StreamId, 653 const MCSubtargetInfo &STI, bool Strict = true); 654 655 LLVM_READNONE 656 bool msgRequiresOp(int64_t MsgId); 657 658 LLVM_READNONE 659 bool msgSupportsStream(int64_t MsgId, int64_t OpId); 660 661 void decodeMsg(unsigned Val, 662 uint16_t &MsgId, 663 uint16_t &OpId, 664 uint16_t &StreamId); 665 666 LLVM_READNONE 667 uint64_t encodeMsg(uint64_t MsgId, 668 uint64_t OpId, 669 uint64_t StreamId); 670 671 } // namespace SendMsg 672 673 674 unsigned getInitialPSInputAddr(const Function &F); 675 676 LLVM_READNONE 677 bool isShader(CallingConv::ID CC); 678 679 LLVM_READNONE 680 bool isGraphics(CallingConv::ID CC); 681 682 LLVM_READNONE 683 bool isCompute(CallingConv::ID CC); 684 685 LLVM_READNONE 686 bool isEntryFunctionCC(CallingConv::ID CC); 687 688 // These functions are considered entrypoints into the current module, i.e. they 689 // are allowed to be called from outside the current module. This is different 690 // from isEntryFunctionCC, which is only true for functions that are entered by 691 // the hardware. Module entry points include all entry functions but also 692 // include functions that can be called from other functions inside or outside 693 // the current module. Module entry functions are allowed to allocate LDS. 694 LLVM_READNONE 695 bool isModuleEntryFunctionCC(CallingConv::ID CC); 696 697 // FIXME: Remove this when calling conventions cleaned up 698 LLVM_READNONE 699 inline bool isKernel(CallingConv::ID CC) { 700 switch (CC) { 701 case CallingConv::AMDGPU_KERNEL: 702 case CallingConv::SPIR_KERNEL: 703 return true; 704 default: 705 return false; 706 } 707 } 708 709 bool hasXNACK(const MCSubtargetInfo &STI); 710 bool hasSRAMECC(const MCSubtargetInfo &STI); 711 bool hasMIMG_R128(const MCSubtargetInfo &STI); 712 bool hasGFX10A16(const MCSubtargetInfo &STI); 713 bool hasG16(const MCSubtargetInfo &STI); 714 bool hasPackedD16(const MCSubtargetInfo &STI); 715 716 bool isSI(const MCSubtargetInfo &STI); 717 bool isCI(const MCSubtargetInfo &STI); 718 bool isVI(const MCSubtargetInfo &STI); 719 bool isGFX9(const MCSubtargetInfo &STI); 720 bool isGFX9Plus(const MCSubtargetInfo &STI); 721 bool isGFX10(const MCSubtargetInfo &STI); 722 bool isGFX10Plus(const MCSubtargetInfo &STI); 723 bool isGCN3Encoding(const MCSubtargetInfo &STI); 724 bool isGFX10_BEncoding(const MCSubtargetInfo &STI); 725 bool hasGFX10_3Insts(const MCSubtargetInfo &STI); 726 bool isGFX90A(const MCSubtargetInfo &STI); 727 728 /// Is Reg - scalar register 729 bool isSGPR(unsigned Reg, const MCRegisterInfo* TRI); 730 731 /// Is there any intersection between registers 732 bool isRegIntersect(unsigned Reg0, unsigned Reg1, const MCRegisterInfo* TRI); 733 734 /// If \p Reg is a pseudo reg, return the correct hardware register given 735 /// \p STI otherwise return \p Reg. 736 unsigned getMCReg(unsigned Reg, const MCSubtargetInfo &STI); 737 738 /// Convert hardware register \p Reg to a pseudo register 739 LLVM_READNONE 740 unsigned mc2PseudoReg(unsigned Reg); 741 742 /// Can this operand also contain immediate values? 743 bool isSISrcOperand(const MCInstrDesc &Desc, unsigned OpNo); 744 745 /// Is this floating-point operand? 746 bool isSISrcFPOperand(const MCInstrDesc &Desc, unsigned OpNo); 747 748 /// Does this opearnd support only inlinable literals? 749 bool isSISrcInlinableOperand(const MCInstrDesc &Desc, unsigned OpNo); 750 751 /// Get the size in bits of a register from the register class \p RC. 752 unsigned getRegBitWidth(unsigned RCID); 753 754 /// Get the size in bits of a register from the register class \p RC. 755 unsigned getRegBitWidth(const MCRegisterClass &RC); 756 757 /// Get size of register operand 758 unsigned getRegOperandSize(const MCRegisterInfo *MRI, const MCInstrDesc &Desc, 759 unsigned OpNo); 760 761 LLVM_READNONE 762 inline unsigned getOperandSize(const MCOperandInfo &OpInfo) { 763 switch (OpInfo.OperandType) { 764 case AMDGPU::OPERAND_REG_IMM_INT32: 765 case AMDGPU::OPERAND_REG_IMM_FP32: 766 case AMDGPU::OPERAND_REG_INLINE_C_INT32: 767 case AMDGPU::OPERAND_REG_INLINE_C_FP32: 768 case AMDGPU::OPERAND_REG_INLINE_AC_INT32: 769 case AMDGPU::OPERAND_REG_INLINE_AC_FP32: 770 case AMDGPU::OPERAND_REG_IMM_V2INT32: 771 case AMDGPU::OPERAND_REG_IMM_V2FP32: 772 case AMDGPU::OPERAND_REG_INLINE_C_V2INT32: 773 case AMDGPU::OPERAND_REG_INLINE_C_V2FP32: 774 return 4; 775 776 case AMDGPU::OPERAND_REG_IMM_INT64: 777 case AMDGPU::OPERAND_REG_IMM_FP64: 778 case AMDGPU::OPERAND_REG_INLINE_C_INT64: 779 case AMDGPU::OPERAND_REG_INLINE_C_FP64: 780 case AMDGPU::OPERAND_REG_INLINE_AC_FP64: 781 return 8; 782 783 case AMDGPU::OPERAND_REG_IMM_INT16: 784 case AMDGPU::OPERAND_REG_IMM_FP16: 785 case AMDGPU::OPERAND_REG_INLINE_C_INT16: 786 case AMDGPU::OPERAND_REG_INLINE_C_FP16: 787 case AMDGPU::OPERAND_REG_INLINE_C_V2INT16: 788 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16: 789 case AMDGPU::OPERAND_REG_INLINE_AC_INT16: 790 case AMDGPU::OPERAND_REG_INLINE_AC_FP16: 791 case AMDGPU::OPERAND_REG_INLINE_AC_V2INT16: 792 case AMDGPU::OPERAND_REG_INLINE_AC_V2FP16: 793 case AMDGPU::OPERAND_REG_IMM_V2INT16: 794 case AMDGPU::OPERAND_REG_IMM_V2FP16: 795 return 2; 796 797 default: 798 llvm_unreachable("unhandled operand type"); 799 } 800 } 801 802 LLVM_READNONE 803 inline unsigned getOperandSize(const MCInstrDesc &Desc, unsigned OpNo) { 804 return getOperandSize(Desc.OpInfo[OpNo]); 805 } 806 807 /// Is this literal inlinable, and not one of the values intended for floating 808 /// point values. 809 LLVM_READNONE 810 inline bool isInlinableIntLiteral(int64_t Literal) { 811 return Literal >= -16 && Literal <= 64; 812 } 813 814 /// Is this literal inlinable 815 LLVM_READNONE 816 bool isInlinableLiteral64(int64_t Literal, bool HasInv2Pi); 817 818 LLVM_READNONE 819 bool isInlinableLiteral32(int32_t Literal, bool HasInv2Pi); 820 821 LLVM_READNONE 822 bool isInlinableLiteral16(int16_t Literal, bool HasInv2Pi); 823 824 LLVM_READNONE 825 bool isInlinableLiteralV216(int32_t Literal, bool HasInv2Pi); 826 827 LLVM_READNONE 828 bool isInlinableIntLiteralV216(int32_t Literal); 829 830 LLVM_READNONE 831 bool isFoldableLiteralV216(int32_t Literal, bool HasInv2Pi); 832 833 bool isArgPassedInSGPR(const Argument *Arg); 834 835 LLVM_READONLY 836 bool isLegalSMRDEncodedUnsignedOffset(const MCSubtargetInfo &ST, 837 int64_t EncodedOffset); 838 839 LLVM_READONLY 840 bool isLegalSMRDEncodedSignedOffset(const MCSubtargetInfo &ST, 841 int64_t EncodedOffset, 842 bool IsBuffer); 843 844 /// Convert \p ByteOffset to dwords if the subtarget uses dword SMRD immediate 845 /// offsets. 846 uint64_t convertSMRDOffsetUnits(const MCSubtargetInfo &ST, uint64_t ByteOffset); 847 848 /// \returns The encoding that will be used for \p ByteOffset in the 849 /// SMRD offset field, or None if it won't fit. On GFX9 and GFX10 850 /// S_LOAD instructions have a signed offset, on other subtargets it is 851 /// unsigned. S_BUFFER has an unsigned offset for all subtargets. 852 Optional<int64_t> getSMRDEncodedOffset(const MCSubtargetInfo &ST, 853 int64_t ByteOffset, bool IsBuffer); 854 855 /// \return The encoding that can be used for a 32-bit literal offset in an SMRD 856 /// instruction. This is only useful on CI.s 857 Optional<int64_t> getSMRDEncodedLiteralOffset32(const MCSubtargetInfo &ST, 858 int64_t ByteOffset); 859 860 /// For FLAT segment the offset must be positive; 861 /// MSB is ignored and forced to zero. 862 /// 863 /// \return The number of bits available for the offset field in flat 864 /// instructions. 865 unsigned getNumFlatOffsetBits(const MCSubtargetInfo &ST, bool Signed); 866 867 /// \returns true if this offset is small enough to fit in the SMRD 868 /// offset field. \p ByteOffset should be the offset in bytes and 869 /// not the encoded offset. 870 bool isLegalSMRDImmOffset(const MCSubtargetInfo &ST, int64_t ByteOffset); 871 872 bool splitMUBUFOffset(uint32_t Imm, uint32_t &SOffset, uint32_t &ImmOffset, 873 const GCNSubtarget *Subtarget, 874 Align Alignment = Align(4)); 875 876 LLVM_READNONE 877 inline bool isLegal64BitDPPControl(unsigned DC) { 878 return DC >= DPP::ROW_NEWBCAST_FIRST && DC <= DPP::ROW_NEWBCAST_LAST; 879 } 880 881 /// \returns true if the intrinsic is divergent 882 bool isIntrinsicSourceOfDivergence(unsigned IntrID); 883 884 // Track defaults for fields in the MODE registser. 885 struct SIModeRegisterDefaults { 886 /// Floating point opcodes that support exception flag gathering quiet and 887 /// propagate signaling NaN inputs per IEEE 754-2008. Min_dx10 and max_dx10 888 /// become IEEE 754- 2008 compliant due to signaling NaN propagation and 889 /// quieting. 890 bool IEEE : 1; 891 892 /// Used by the vector ALU to force DX10-style treatment of NaNs: when set, 893 /// clamp NaN to zero; otherwise, pass NaN through. 894 bool DX10Clamp : 1; 895 896 /// If this is set, neither input or output denormals are flushed for most f32 897 /// instructions. 898 bool FP32InputDenormals : 1; 899 bool FP32OutputDenormals : 1; 900 901 /// If this is set, neither input or output denormals are flushed for both f64 902 /// and f16/v2f16 instructions. 903 bool FP64FP16InputDenormals : 1; 904 bool FP64FP16OutputDenormals : 1; 905 906 SIModeRegisterDefaults() : 907 IEEE(true), 908 DX10Clamp(true), 909 FP32InputDenormals(true), 910 FP32OutputDenormals(true), 911 FP64FP16InputDenormals(true), 912 FP64FP16OutputDenormals(true) {} 913 914 SIModeRegisterDefaults(const Function &F); 915 916 static SIModeRegisterDefaults getDefaultForCallingConv(CallingConv::ID CC) { 917 SIModeRegisterDefaults Mode; 918 Mode.IEEE = !AMDGPU::isShader(CC); 919 return Mode; 920 } 921 922 bool operator ==(const SIModeRegisterDefaults Other) const { 923 return IEEE == Other.IEEE && DX10Clamp == Other.DX10Clamp && 924 FP32InputDenormals == Other.FP32InputDenormals && 925 FP32OutputDenormals == Other.FP32OutputDenormals && 926 FP64FP16InputDenormals == Other.FP64FP16InputDenormals && 927 FP64FP16OutputDenormals == Other.FP64FP16OutputDenormals; 928 } 929 930 bool allFP32Denormals() const { 931 return FP32InputDenormals && FP32OutputDenormals; 932 } 933 934 bool allFP64FP16Denormals() const { 935 return FP64FP16InputDenormals && FP64FP16OutputDenormals; 936 } 937 938 /// Get the encoding value for the FP_DENORM bits of the mode register for the 939 /// FP32 denormal mode. 940 uint32_t fpDenormModeSPValue() const { 941 if (FP32InputDenormals && FP32OutputDenormals) 942 return FP_DENORM_FLUSH_NONE; 943 if (FP32InputDenormals) 944 return FP_DENORM_FLUSH_OUT; 945 if (FP32OutputDenormals) 946 return FP_DENORM_FLUSH_IN; 947 return FP_DENORM_FLUSH_IN_FLUSH_OUT; 948 } 949 950 /// Get the encoding value for the FP_DENORM bits of the mode register for the 951 /// FP64/FP16 denormal mode. 952 uint32_t fpDenormModeDPValue() const { 953 if (FP64FP16InputDenormals && FP64FP16OutputDenormals) 954 return FP_DENORM_FLUSH_NONE; 955 if (FP64FP16InputDenormals) 956 return FP_DENORM_FLUSH_OUT; 957 if (FP64FP16OutputDenormals) 958 return FP_DENORM_FLUSH_IN; 959 return FP_DENORM_FLUSH_IN_FLUSH_OUT; 960 } 961 962 /// Returns true if a flag is compatible if it's enabled in the callee, but 963 /// disabled in the caller. 964 static bool oneWayCompatible(bool CallerMode, bool CalleeMode) { 965 return CallerMode == CalleeMode || (!CallerMode && CalleeMode); 966 } 967 968 // FIXME: Inlining should be OK for dx10-clamp, since the caller's mode should 969 // be able to override. 970 bool isInlineCompatible(SIModeRegisterDefaults CalleeMode) const { 971 if (DX10Clamp != CalleeMode.DX10Clamp) 972 return false; 973 if (IEEE != CalleeMode.IEEE) 974 return false; 975 976 // Allow inlining denormals enabled into denormals flushed functions. 977 return oneWayCompatible(FP64FP16InputDenormals, CalleeMode.FP64FP16InputDenormals) && 978 oneWayCompatible(FP64FP16OutputDenormals, CalleeMode.FP64FP16OutputDenormals) && 979 oneWayCompatible(FP32InputDenormals, CalleeMode.FP32InputDenormals) && 980 oneWayCompatible(FP32OutputDenormals, CalleeMode.FP32OutputDenormals); 981 } 982 }; 983 984 } // end namespace AMDGPU 985 986 raw_ostream &operator<<(raw_ostream &OS, 987 const AMDGPU::IsaInfo::TargetIDSetting S); 988 989 } // end namespace llvm 990 991 #endif // LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUBASEINFO_H 992