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