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 }; 288 289 LLVM_READONLY 290 const MIMGBaseOpcodeInfo *getMIMGBaseOpcodeInfo(unsigned BaseOpcode); 291 292 struct MIMGDimInfo { 293 MIMGDim Dim; 294 uint8_t NumCoords; 295 uint8_t NumGradients; 296 bool DA; 297 uint8_t Encoding; 298 const char *AsmSuffix; 299 }; 300 301 LLVM_READONLY 302 const MIMGDimInfo *getMIMGDimInfo(unsigned DimEnum); 303 304 LLVM_READONLY 305 const MIMGDimInfo *getMIMGDimInfoByEncoding(uint8_t DimEnc); 306 307 LLVM_READONLY 308 const MIMGDimInfo *getMIMGDimInfoByAsmSuffix(StringRef AsmSuffix); 309 310 struct MIMGLZMappingInfo { 311 MIMGBaseOpcode L; 312 MIMGBaseOpcode LZ; 313 }; 314 315 struct MIMGMIPMappingInfo { 316 MIMGBaseOpcode MIP; 317 MIMGBaseOpcode NONMIP; 318 }; 319 320 struct MIMGG16MappingInfo { 321 MIMGBaseOpcode G; 322 MIMGBaseOpcode G16; 323 }; 324 325 LLVM_READONLY 326 const MIMGLZMappingInfo *getMIMGLZMappingInfo(unsigned L); 327 328 LLVM_READONLY 329 const MIMGMIPMappingInfo *getMIMGMIPMappingInfo(unsigned MIP); 330 331 LLVM_READONLY 332 const MIMGG16MappingInfo *getMIMGG16MappingInfo(unsigned G); 333 334 LLVM_READONLY 335 int getMIMGOpcode(unsigned BaseOpcode, unsigned MIMGEncoding, 336 unsigned VDataDwords, unsigned VAddrDwords); 337 338 LLVM_READONLY 339 int getMaskedMIMGOp(unsigned Opc, unsigned NewChannels); 340 341 struct MIMGInfo { 342 uint16_t Opcode; 343 uint16_t BaseOpcode; 344 uint8_t MIMGEncoding; 345 uint8_t VDataDwords; 346 uint8_t VAddrDwords; 347 }; 348 349 LLVM_READONLY 350 const MIMGInfo *getMIMGInfo(unsigned Opc); 351 352 LLVM_READONLY 353 int getMTBUFBaseOpcode(unsigned Opc); 354 355 LLVM_READONLY 356 int getMTBUFOpcode(unsigned BaseOpc, unsigned Elements); 357 358 LLVM_READONLY 359 int getMTBUFElements(unsigned Opc); 360 361 LLVM_READONLY 362 bool getMTBUFHasVAddr(unsigned Opc); 363 364 LLVM_READONLY 365 bool getMTBUFHasSrsrc(unsigned Opc); 366 367 LLVM_READONLY 368 bool getMTBUFHasSoffset(unsigned Opc); 369 370 LLVM_READONLY 371 int getMUBUFBaseOpcode(unsigned Opc); 372 373 LLVM_READONLY 374 int getMUBUFOpcode(unsigned BaseOpc, unsigned Elements); 375 376 LLVM_READONLY 377 int getMUBUFElements(unsigned Opc); 378 379 LLVM_READONLY 380 bool getMUBUFHasVAddr(unsigned Opc); 381 382 LLVM_READONLY 383 bool getMUBUFHasSrsrc(unsigned Opc); 384 385 LLVM_READONLY 386 bool getMUBUFHasSoffset(unsigned Opc); 387 388 LLVM_READONLY 389 bool getSMEMIsBuffer(unsigned Opc); 390 391 LLVM_READONLY 392 const GcnBufferFormatInfo *getGcnBufferFormatInfo(uint8_t BitsPerComp, 393 uint8_t NumComponents, 394 uint8_t NumFormat, 395 const MCSubtargetInfo &STI); 396 LLVM_READONLY 397 const GcnBufferFormatInfo *getGcnBufferFormatInfo(uint8_t Format, 398 const MCSubtargetInfo &STI); 399 400 LLVM_READONLY 401 int getMCOpcode(uint16_t Opcode, unsigned Gen); 402 403 void initDefaultAMDKernelCodeT(amd_kernel_code_t &Header, 404 const MCSubtargetInfo *STI); 405 406 amdhsa::kernel_descriptor_t getDefaultAmdhsaKernelDescriptor( 407 const MCSubtargetInfo *STI); 408 409 bool isGroupSegment(const GlobalValue *GV); 410 bool isGlobalSegment(const GlobalValue *GV); 411 bool isReadOnlySegment(const GlobalValue *GV); 412 413 /// \returns True if constants should be emitted to .text section for given 414 /// target triple \p TT, false otherwise. 415 bool shouldEmitConstantsToTextSection(const Triple &TT); 416 417 /// \returns Integer value requested using \p F's \p Name attribute. 418 /// 419 /// \returns \p Default if attribute is not present. 420 /// 421 /// \returns \p Default and emits error if requested value cannot be converted 422 /// to integer. 423 int getIntegerAttribute(const Function &F, StringRef Name, int Default); 424 425 /// \returns A pair of integer values requested using \p F's \p Name attribute 426 /// in "first[,second]" format ("second" is optional unless \p OnlyFirstRequired 427 /// is false). 428 /// 429 /// \returns \p Default if attribute is not present. 430 /// 431 /// \returns \p Default and emits error if one of the requested values cannot be 432 /// converted to integer, or \p OnlyFirstRequired is false and "second" value is 433 /// not present. 434 std::pair<int, int> getIntegerPairAttribute(const Function &F, 435 StringRef Name, 436 std::pair<int, int> Default, 437 bool OnlyFirstRequired = false); 438 439 /// Represents the counter values to wait for in an s_waitcnt instruction. 440 /// 441 /// Large values (including the maximum possible integer) can be used to 442 /// represent "don't care" waits. 443 struct Waitcnt { 444 unsigned VmCnt = ~0u; 445 unsigned ExpCnt = ~0u; 446 unsigned LgkmCnt = ~0u; 447 unsigned VsCnt = ~0u; 448 449 Waitcnt() {} 450 Waitcnt(unsigned VmCnt, unsigned ExpCnt, unsigned LgkmCnt, unsigned VsCnt) 451 : VmCnt(VmCnt), ExpCnt(ExpCnt), LgkmCnt(LgkmCnt), VsCnt(VsCnt) {} 452 453 static Waitcnt allZero(bool HasVscnt) { 454 return Waitcnt(0, 0, 0, HasVscnt ? 0 : ~0u); 455 } 456 static Waitcnt allZeroExceptVsCnt() { return Waitcnt(0, 0, 0, ~0u); } 457 458 bool hasWait() const { 459 return VmCnt != ~0u || ExpCnt != ~0u || LgkmCnt != ~0u || VsCnt != ~0u; 460 } 461 462 bool dominates(const Waitcnt &Other) const { 463 return VmCnt <= Other.VmCnt && ExpCnt <= Other.ExpCnt && 464 LgkmCnt <= Other.LgkmCnt && VsCnt <= Other.VsCnt; 465 } 466 467 Waitcnt combined(const Waitcnt &Other) const { 468 return Waitcnt(std::min(VmCnt, Other.VmCnt), std::min(ExpCnt, Other.ExpCnt), 469 std::min(LgkmCnt, Other.LgkmCnt), 470 std::min(VsCnt, Other.VsCnt)); 471 } 472 }; 473 474 /// \returns Vmcnt bit mask for given isa \p Version. 475 unsigned getVmcntBitMask(const IsaVersion &Version); 476 477 /// \returns Expcnt bit mask for given isa \p Version. 478 unsigned getExpcntBitMask(const IsaVersion &Version); 479 480 /// \returns Lgkmcnt bit mask for given isa \p Version. 481 unsigned getLgkmcntBitMask(const IsaVersion &Version); 482 483 /// \returns Waitcnt bit mask for given isa \p Version. 484 unsigned getWaitcntBitMask(const IsaVersion &Version); 485 486 /// \returns Decoded Vmcnt from given \p Waitcnt for given isa \p Version. 487 unsigned decodeVmcnt(const IsaVersion &Version, unsigned Waitcnt); 488 489 /// \returns Decoded Expcnt from given \p Waitcnt for given isa \p Version. 490 unsigned decodeExpcnt(const IsaVersion &Version, unsigned Waitcnt); 491 492 /// \returns Decoded Lgkmcnt from given \p Waitcnt for given isa \p Version. 493 unsigned decodeLgkmcnt(const IsaVersion &Version, unsigned Waitcnt); 494 495 /// Decodes Vmcnt, Expcnt and Lgkmcnt from given \p Waitcnt for given isa 496 /// \p Version, and writes decoded values into \p Vmcnt, \p Expcnt and 497 /// \p Lgkmcnt respectively. 498 /// 499 /// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are decoded as follows: 500 /// \p Vmcnt = \p Waitcnt[3:0] (pre-gfx9 only) 501 /// \p Vmcnt = \p Waitcnt[3:0] | \p Waitcnt[15:14] (gfx9+ only) 502 /// \p Expcnt = \p Waitcnt[6:4] 503 /// \p Lgkmcnt = \p Waitcnt[11:8] (pre-gfx10 only) 504 /// \p Lgkmcnt = \p Waitcnt[13:8] (gfx10+ only) 505 void decodeWaitcnt(const IsaVersion &Version, unsigned Waitcnt, 506 unsigned &Vmcnt, unsigned &Expcnt, unsigned &Lgkmcnt); 507 508 Waitcnt decodeWaitcnt(const IsaVersion &Version, unsigned Encoded); 509 510 /// \returns \p Waitcnt with encoded \p Vmcnt for given isa \p Version. 511 unsigned encodeVmcnt(const IsaVersion &Version, unsigned Waitcnt, 512 unsigned Vmcnt); 513 514 /// \returns \p Waitcnt with encoded \p Expcnt for given isa \p Version. 515 unsigned encodeExpcnt(const IsaVersion &Version, unsigned Waitcnt, 516 unsigned Expcnt); 517 518 /// \returns \p Waitcnt with encoded \p Lgkmcnt for given isa \p Version. 519 unsigned encodeLgkmcnt(const IsaVersion &Version, unsigned Waitcnt, 520 unsigned Lgkmcnt); 521 522 /// Encodes \p Vmcnt, \p Expcnt and \p Lgkmcnt into Waitcnt for given isa 523 /// \p Version. 524 /// 525 /// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are encoded as follows: 526 /// Waitcnt[3:0] = \p Vmcnt (pre-gfx9 only) 527 /// Waitcnt[3:0] = \p Vmcnt[3:0] (gfx9+ only) 528 /// Waitcnt[6:4] = \p Expcnt 529 /// Waitcnt[11:8] = \p Lgkmcnt (pre-gfx10 only) 530 /// Waitcnt[13:8] = \p Lgkmcnt (gfx10+ only) 531 /// Waitcnt[15:14] = \p Vmcnt[5:4] (gfx9+ only) 532 /// 533 /// \returns Waitcnt with encoded \p Vmcnt, \p Expcnt and \p Lgkmcnt for given 534 /// isa \p Version. 535 unsigned encodeWaitcnt(const IsaVersion &Version, 536 unsigned Vmcnt, unsigned Expcnt, unsigned Lgkmcnt); 537 538 unsigned encodeWaitcnt(const IsaVersion &Version, const Waitcnt &Decoded); 539 540 namespace Hwreg { 541 542 LLVM_READONLY 543 int64_t getHwregId(const StringRef Name); 544 545 LLVM_READNONE 546 bool isValidHwreg(int64_t Id, const MCSubtargetInfo &STI); 547 548 LLVM_READNONE 549 bool isValidHwreg(int64_t Id); 550 551 LLVM_READNONE 552 bool isValidHwregOffset(int64_t Offset); 553 554 LLVM_READNONE 555 bool isValidHwregWidth(int64_t Width); 556 557 LLVM_READNONE 558 uint64_t encodeHwreg(uint64_t Id, uint64_t Offset, uint64_t Width); 559 560 LLVM_READNONE 561 StringRef getHwreg(unsigned Id, const MCSubtargetInfo &STI); 562 563 void decodeHwreg(unsigned Val, unsigned &Id, unsigned &Offset, unsigned &Width); 564 565 } // namespace Hwreg 566 567 namespace Exp { 568 569 bool getTgtName(unsigned Id, StringRef &Name, int &Index); 570 571 LLVM_READONLY 572 unsigned getTgtId(const StringRef Name); 573 574 LLVM_READNONE 575 bool isSupportedTgtId(unsigned Id, const MCSubtargetInfo &STI); 576 577 } // namespace Exp 578 579 namespace MTBUFFormat { 580 581 LLVM_READNONE 582 int64_t encodeDfmtNfmt(unsigned Dfmt, unsigned Nfmt); 583 584 void decodeDfmtNfmt(unsigned Format, unsigned &Dfmt, unsigned &Nfmt); 585 586 int64_t getDfmt(const StringRef Name); 587 588 StringRef getDfmtName(unsigned Id); 589 590 int64_t getNfmt(const StringRef Name, const MCSubtargetInfo &STI); 591 592 StringRef getNfmtName(unsigned Id, const MCSubtargetInfo &STI); 593 594 bool isValidDfmtNfmt(unsigned Val, const MCSubtargetInfo &STI); 595 596 bool isValidNfmt(unsigned Val, const MCSubtargetInfo &STI); 597 598 int64_t getUnifiedFormat(const StringRef Name); 599 600 StringRef getUnifiedFormatName(unsigned Id); 601 602 bool isValidUnifiedFormat(unsigned Val); 603 604 int64_t convertDfmtNfmt2Ufmt(unsigned Dfmt, unsigned Nfmt); 605 606 bool isValidFormatEncoding(unsigned Val, const MCSubtargetInfo &STI); 607 608 unsigned getDefaultFormatEncoding(const MCSubtargetInfo &STI); 609 610 } // namespace MTBUFFormat 611 612 namespace SendMsg { 613 614 LLVM_READONLY 615 int64_t getMsgId(const StringRef Name); 616 617 LLVM_READONLY 618 int64_t getMsgOpId(int64_t MsgId, const StringRef Name); 619 620 LLVM_READNONE 621 StringRef getMsgName(int64_t MsgId); 622 623 LLVM_READNONE 624 StringRef getMsgOpName(int64_t MsgId, int64_t OpId); 625 626 LLVM_READNONE 627 bool isValidMsgId(int64_t MsgId, const MCSubtargetInfo &STI, bool Strict = true); 628 629 LLVM_READNONE 630 bool isValidMsgOp(int64_t MsgId, int64_t OpId, const MCSubtargetInfo &STI, 631 bool Strict = true); 632 633 LLVM_READNONE 634 bool isValidMsgStream(int64_t MsgId, int64_t OpId, int64_t StreamId, 635 const MCSubtargetInfo &STI, bool Strict = true); 636 637 LLVM_READNONE 638 bool msgRequiresOp(int64_t MsgId); 639 640 LLVM_READNONE 641 bool msgSupportsStream(int64_t MsgId, int64_t OpId); 642 643 void decodeMsg(unsigned Val, 644 uint16_t &MsgId, 645 uint16_t &OpId, 646 uint16_t &StreamId); 647 648 LLVM_READNONE 649 uint64_t encodeMsg(uint64_t MsgId, 650 uint64_t OpId, 651 uint64_t StreamId); 652 653 } // namespace SendMsg 654 655 656 unsigned getInitialPSInputAddr(const Function &F); 657 658 LLVM_READNONE 659 bool isShader(CallingConv::ID CC); 660 661 LLVM_READNONE 662 bool isGraphics(CallingConv::ID CC); 663 664 LLVM_READNONE 665 bool isCompute(CallingConv::ID CC); 666 667 LLVM_READNONE 668 bool isEntryFunctionCC(CallingConv::ID CC); 669 670 // These functions are considered entrypoints into the current module, i.e. they 671 // are allowed to be called from outside the current module. This is different 672 // from isEntryFunctionCC, which is only true for functions that are entered by 673 // the hardware. Module entry points include all entry functions but also 674 // include functions that can be called from other functions inside or outside 675 // the current module. Module entry functions are allowed to allocate LDS. 676 LLVM_READNONE 677 bool isModuleEntryFunctionCC(CallingConv::ID CC); 678 679 // FIXME: Remove this when calling conventions cleaned up 680 LLVM_READNONE 681 inline bool isKernel(CallingConv::ID CC) { 682 switch (CC) { 683 case CallingConv::AMDGPU_KERNEL: 684 case CallingConv::SPIR_KERNEL: 685 return true; 686 default: 687 return false; 688 } 689 } 690 691 bool hasXNACK(const MCSubtargetInfo &STI); 692 bool hasSRAMECC(const MCSubtargetInfo &STI); 693 bool hasMIMG_R128(const MCSubtargetInfo &STI); 694 bool hasGFX10A16(const MCSubtargetInfo &STI); 695 bool hasG16(const MCSubtargetInfo &STI); 696 bool hasPackedD16(const MCSubtargetInfo &STI); 697 698 bool isSI(const MCSubtargetInfo &STI); 699 bool isCI(const MCSubtargetInfo &STI); 700 bool isVI(const MCSubtargetInfo &STI); 701 bool isGFX9(const MCSubtargetInfo &STI); 702 bool isGFX9Plus(const MCSubtargetInfo &STI); 703 bool isGFX10(const MCSubtargetInfo &STI); 704 bool isGFX10Plus(const MCSubtargetInfo &STI); 705 bool isGCN3Encoding(const MCSubtargetInfo &STI); 706 bool isGFX10_BEncoding(const MCSubtargetInfo &STI); 707 bool hasGFX10_3Insts(const MCSubtargetInfo &STI); 708 bool isGFX90A(const MCSubtargetInfo &STI); 709 710 /// Is Reg - scalar register 711 bool isSGPR(unsigned Reg, const MCRegisterInfo* TRI); 712 713 /// Is there any intersection between registers 714 bool isRegIntersect(unsigned Reg0, unsigned Reg1, const MCRegisterInfo* TRI); 715 716 /// If \p Reg is a pseudo reg, return the correct hardware register given 717 /// \p STI otherwise return \p Reg. 718 unsigned getMCReg(unsigned Reg, const MCSubtargetInfo &STI); 719 720 /// Convert hardware register \p Reg to a pseudo register 721 LLVM_READNONE 722 unsigned mc2PseudoReg(unsigned Reg); 723 724 /// Can this operand also contain immediate values? 725 bool isSISrcOperand(const MCInstrDesc &Desc, unsigned OpNo); 726 727 /// Is this floating-point operand? 728 bool isSISrcFPOperand(const MCInstrDesc &Desc, unsigned OpNo); 729 730 /// Does this opearnd support only inlinable literals? 731 bool isSISrcInlinableOperand(const MCInstrDesc &Desc, unsigned OpNo); 732 733 /// Get the size in bits of a register from the register class \p RC. 734 unsigned getRegBitWidth(unsigned RCID); 735 736 /// Get the size in bits of a register from the register class \p RC. 737 unsigned getRegBitWidth(const MCRegisterClass &RC); 738 739 /// Get size of register operand 740 unsigned getRegOperandSize(const MCRegisterInfo *MRI, const MCInstrDesc &Desc, 741 unsigned OpNo); 742 743 LLVM_READNONE 744 inline unsigned getOperandSize(const MCOperandInfo &OpInfo) { 745 switch (OpInfo.OperandType) { 746 case AMDGPU::OPERAND_REG_IMM_INT32: 747 case AMDGPU::OPERAND_REG_IMM_FP32: 748 case AMDGPU::OPERAND_REG_INLINE_C_INT32: 749 case AMDGPU::OPERAND_REG_INLINE_C_FP32: 750 case AMDGPU::OPERAND_REG_INLINE_AC_INT32: 751 case AMDGPU::OPERAND_REG_INLINE_AC_FP32: 752 case AMDGPU::OPERAND_REG_IMM_V2INT32: 753 case AMDGPU::OPERAND_REG_IMM_V2FP32: 754 case AMDGPU::OPERAND_REG_INLINE_C_V2INT32: 755 case AMDGPU::OPERAND_REG_INLINE_C_V2FP32: 756 return 4; 757 758 case AMDGPU::OPERAND_REG_IMM_INT64: 759 case AMDGPU::OPERAND_REG_IMM_FP64: 760 case AMDGPU::OPERAND_REG_INLINE_C_INT64: 761 case AMDGPU::OPERAND_REG_INLINE_C_FP64: 762 case AMDGPU::OPERAND_REG_INLINE_AC_FP64: 763 return 8; 764 765 case AMDGPU::OPERAND_REG_IMM_INT16: 766 case AMDGPU::OPERAND_REG_IMM_FP16: 767 case AMDGPU::OPERAND_REG_INLINE_C_INT16: 768 case AMDGPU::OPERAND_REG_INLINE_C_FP16: 769 case AMDGPU::OPERAND_REG_INLINE_C_V2INT16: 770 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16: 771 case AMDGPU::OPERAND_REG_INLINE_AC_INT16: 772 case AMDGPU::OPERAND_REG_INLINE_AC_FP16: 773 case AMDGPU::OPERAND_REG_INLINE_AC_V2INT16: 774 case AMDGPU::OPERAND_REG_INLINE_AC_V2FP16: 775 case AMDGPU::OPERAND_REG_IMM_V2INT16: 776 case AMDGPU::OPERAND_REG_IMM_V2FP16: 777 return 2; 778 779 default: 780 llvm_unreachable("unhandled operand type"); 781 } 782 } 783 784 LLVM_READNONE 785 inline unsigned getOperandSize(const MCInstrDesc &Desc, unsigned OpNo) { 786 return getOperandSize(Desc.OpInfo[OpNo]); 787 } 788 789 /// Is this literal inlinable, and not one of the values intended for floating 790 /// point values. 791 LLVM_READNONE 792 inline bool isInlinableIntLiteral(int64_t Literal) { 793 return Literal >= -16 && Literal <= 64; 794 } 795 796 /// Is this literal inlinable 797 LLVM_READNONE 798 bool isInlinableLiteral64(int64_t Literal, bool HasInv2Pi); 799 800 LLVM_READNONE 801 bool isInlinableLiteral32(int32_t Literal, bool HasInv2Pi); 802 803 LLVM_READNONE 804 bool isInlinableLiteral16(int16_t Literal, bool HasInv2Pi); 805 806 LLVM_READNONE 807 bool isInlinableLiteralV216(int32_t Literal, bool HasInv2Pi); 808 809 LLVM_READNONE 810 bool isInlinableIntLiteralV216(int32_t Literal); 811 812 LLVM_READNONE 813 bool isFoldableLiteralV216(int32_t Literal, bool HasInv2Pi); 814 815 bool isArgPassedInSGPR(const Argument *Arg); 816 817 LLVM_READONLY 818 bool isLegalSMRDEncodedUnsignedOffset(const MCSubtargetInfo &ST, 819 int64_t EncodedOffset); 820 821 LLVM_READONLY 822 bool isLegalSMRDEncodedSignedOffset(const MCSubtargetInfo &ST, 823 int64_t EncodedOffset, 824 bool IsBuffer); 825 826 /// Convert \p ByteOffset to dwords if the subtarget uses dword SMRD immediate 827 /// offsets. 828 uint64_t convertSMRDOffsetUnits(const MCSubtargetInfo &ST, uint64_t ByteOffset); 829 830 /// \returns The encoding that will be used for \p ByteOffset in the 831 /// SMRD offset field, or None if it won't fit. On GFX9 and GFX10 832 /// S_LOAD instructions have a signed offset, on other subtargets it is 833 /// unsigned. S_BUFFER has an unsigned offset for all subtargets. 834 Optional<int64_t> getSMRDEncodedOffset(const MCSubtargetInfo &ST, 835 int64_t ByteOffset, bool IsBuffer); 836 837 /// \return The encoding that can be used for a 32-bit literal offset in an SMRD 838 /// instruction. This is only useful on CI.s 839 Optional<int64_t> getSMRDEncodedLiteralOffset32(const MCSubtargetInfo &ST, 840 int64_t ByteOffset); 841 842 /// For FLAT segment the offset must be positive; 843 /// MSB is ignored and forced to zero. 844 /// 845 /// \return The number of bits available for the offset field in flat 846 /// instructions. 847 unsigned getNumFlatOffsetBits(const MCSubtargetInfo &ST, bool Signed); 848 849 /// \returns true if this offset is small enough to fit in the SMRD 850 /// offset field. \p ByteOffset should be the offset in bytes and 851 /// not the encoded offset. 852 bool isLegalSMRDImmOffset(const MCSubtargetInfo &ST, int64_t ByteOffset); 853 854 bool splitMUBUFOffset(uint32_t Imm, uint32_t &SOffset, uint32_t &ImmOffset, 855 const GCNSubtarget *Subtarget, 856 Align Alignment = Align(4)); 857 858 LLVM_READNONE 859 inline bool isLegal64BitDPPControl(unsigned DC) { 860 return DC >= DPP::ROW_NEWBCAST_FIRST && DC <= DPP::ROW_NEWBCAST_LAST; 861 } 862 863 /// \returns true if the intrinsic is divergent 864 bool isIntrinsicSourceOfDivergence(unsigned IntrID); 865 866 // Track defaults for fields in the MODE registser. 867 struct SIModeRegisterDefaults { 868 /// Floating point opcodes that support exception flag gathering quiet and 869 /// propagate signaling NaN inputs per IEEE 754-2008. Min_dx10 and max_dx10 870 /// become IEEE 754- 2008 compliant due to signaling NaN propagation and 871 /// quieting. 872 bool IEEE : 1; 873 874 /// Used by the vector ALU to force DX10-style treatment of NaNs: when set, 875 /// clamp NaN to zero; otherwise, pass NaN through. 876 bool DX10Clamp : 1; 877 878 /// If this is set, neither input or output denormals are flushed for most f32 879 /// instructions. 880 bool FP32InputDenormals : 1; 881 bool FP32OutputDenormals : 1; 882 883 /// If this is set, neither input or output denormals are flushed for both f64 884 /// and f16/v2f16 instructions. 885 bool FP64FP16InputDenormals : 1; 886 bool FP64FP16OutputDenormals : 1; 887 888 SIModeRegisterDefaults() : 889 IEEE(true), 890 DX10Clamp(true), 891 FP32InputDenormals(true), 892 FP32OutputDenormals(true), 893 FP64FP16InputDenormals(true), 894 FP64FP16OutputDenormals(true) {} 895 896 SIModeRegisterDefaults(const Function &F); 897 898 static SIModeRegisterDefaults getDefaultForCallingConv(CallingConv::ID CC) { 899 SIModeRegisterDefaults Mode; 900 Mode.IEEE = !AMDGPU::isShader(CC); 901 return Mode; 902 } 903 904 bool operator ==(const SIModeRegisterDefaults Other) const { 905 return IEEE == Other.IEEE && DX10Clamp == Other.DX10Clamp && 906 FP32InputDenormals == Other.FP32InputDenormals && 907 FP32OutputDenormals == Other.FP32OutputDenormals && 908 FP64FP16InputDenormals == Other.FP64FP16InputDenormals && 909 FP64FP16OutputDenormals == Other.FP64FP16OutputDenormals; 910 } 911 912 bool allFP32Denormals() const { 913 return FP32InputDenormals && FP32OutputDenormals; 914 } 915 916 bool allFP64FP16Denormals() const { 917 return FP64FP16InputDenormals && FP64FP16OutputDenormals; 918 } 919 920 /// Get the encoding value for the FP_DENORM bits of the mode register for the 921 /// FP32 denormal mode. 922 uint32_t fpDenormModeSPValue() const { 923 if (FP32InputDenormals && FP32OutputDenormals) 924 return FP_DENORM_FLUSH_NONE; 925 if (FP32InputDenormals) 926 return FP_DENORM_FLUSH_OUT; 927 if (FP32OutputDenormals) 928 return FP_DENORM_FLUSH_IN; 929 return FP_DENORM_FLUSH_IN_FLUSH_OUT; 930 } 931 932 /// Get the encoding value for the FP_DENORM bits of the mode register for the 933 /// FP64/FP16 denormal mode. 934 uint32_t fpDenormModeDPValue() const { 935 if (FP64FP16InputDenormals && FP64FP16OutputDenormals) 936 return FP_DENORM_FLUSH_NONE; 937 if (FP64FP16InputDenormals) 938 return FP_DENORM_FLUSH_OUT; 939 if (FP64FP16OutputDenormals) 940 return FP_DENORM_FLUSH_IN; 941 return FP_DENORM_FLUSH_IN_FLUSH_OUT; 942 } 943 944 /// Returns true if a flag is compatible if it's enabled in the callee, but 945 /// disabled in the caller. 946 static bool oneWayCompatible(bool CallerMode, bool CalleeMode) { 947 return CallerMode == CalleeMode || (!CallerMode && CalleeMode); 948 } 949 950 // FIXME: Inlining should be OK for dx10-clamp, since the caller's mode should 951 // be able to override. 952 bool isInlineCompatible(SIModeRegisterDefaults CalleeMode) const { 953 if (DX10Clamp != CalleeMode.DX10Clamp) 954 return false; 955 if (IEEE != CalleeMode.IEEE) 956 return false; 957 958 // Allow inlining denormals enabled into denormals flushed functions. 959 return oneWayCompatible(FP64FP16InputDenormals, CalleeMode.FP64FP16InputDenormals) && 960 oneWayCompatible(FP64FP16OutputDenormals, CalleeMode.FP64FP16OutputDenormals) && 961 oneWayCompatible(FP32InputDenormals, CalleeMode.FP32InputDenormals) && 962 oneWayCompatible(FP32OutputDenormals, CalleeMode.FP32OutputDenormals); 963 } 964 }; 965 966 } // end namespace AMDGPU 967 968 raw_ostream &operator<<(raw_ostream &OS, 969 const AMDGPU::IsaInfo::TargetIDSetting S); 970 971 } // end namespace llvm 972 973 #endif // LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUBASEINFO_H 974