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 #define GET_MIMGBiASMapping_DECL 68 #include "AMDGPUGenSearchableTables.inc" 69 70 namespace IsaInfo { 71 72 enum { 73 // The closed Vulkan driver sets 96, which limits the wave count to 8 but 74 // doesn't spill SGPRs as much as when 80 is set. 75 FIXED_NUM_SGPRS_FOR_INIT_BUG = 96, 76 TRAP_NUM_SGPRS = 16 77 }; 78 79 enum class TargetIDSetting { 80 Unsupported, 81 Any, 82 Off, 83 On 84 }; 85 86 class AMDGPUTargetID { 87 private: 88 const MCSubtargetInfo &STI; 89 TargetIDSetting XnackSetting; 90 TargetIDSetting SramEccSetting; 91 92 public: 93 explicit AMDGPUTargetID(const MCSubtargetInfo &STI); 94 ~AMDGPUTargetID() = default; 95 96 /// \return True if the current xnack setting is not "Unsupported". 97 bool isXnackSupported() const { 98 return XnackSetting != TargetIDSetting::Unsupported; 99 } 100 101 /// \returns True if the current xnack setting is "On" or "Any". 102 bool isXnackOnOrAny() const { 103 return XnackSetting == TargetIDSetting::On || 104 XnackSetting == TargetIDSetting::Any; 105 } 106 107 /// \returns True if current xnack setting is "On" or "Off", 108 /// false otherwise. 109 bool isXnackOnOrOff() const { 110 return getXnackSetting() == TargetIDSetting::On || 111 getXnackSetting() == TargetIDSetting::Off; 112 } 113 114 /// \returns The current xnack TargetIDSetting, possible options are 115 /// "Unsupported", "Any", "Off", and "On". 116 TargetIDSetting getXnackSetting() const { 117 return XnackSetting; 118 } 119 120 /// Sets xnack setting to \p NewXnackSetting. 121 void setXnackSetting(TargetIDSetting NewXnackSetting) { 122 XnackSetting = NewXnackSetting; 123 } 124 125 /// \return True if the current sramecc setting is not "Unsupported". 126 bool isSramEccSupported() const { 127 return SramEccSetting != TargetIDSetting::Unsupported; 128 } 129 130 /// \returns True if the current sramecc setting is "On" or "Any". 131 bool isSramEccOnOrAny() const { 132 return SramEccSetting == TargetIDSetting::On || 133 SramEccSetting == TargetIDSetting::Any; 134 } 135 136 /// \returns True if current sramecc setting is "On" or "Off", 137 /// false otherwise. 138 bool isSramEccOnOrOff() const { 139 return getSramEccSetting() == TargetIDSetting::On || 140 getSramEccSetting() == TargetIDSetting::Off; 141 } 142 143 /// \returns The current sramecc TargetIDSetting, possible options are 144 /// "Unsupported", "Any", "Off", and "On". 145 TargetIDSetting getSramEccSetting() const { 146 return SramEccSetting; 147 } 148 149 /// Sets sramecc setting to \p NewSramEccSetting. 150 void setSramEccSetting(TargetIDSetting NewSramEccSetting) { 151 SramEccSetting = NewSramEccSetting; 152 } 153 154 void setTargetIDFromFeaturesString(StringRef FS); 155 void setTargetIDFromTargetIDStream(StringRef TargetID); 156 157 /// \returns String representation of an object. 158 std::string toString() const; 159 }; 160 161 /// \returns Wavefront size for given subtarget \p STI. 162 unsigned getWavefrontSize(const MCSubtargetInfo *STI); 163 164 /// \returns Local memory size in bytes for given subtarget \p STI. 165 unsigned getLocalMemorySize(const MCSubtargetInfo *STI); 166 167 /// \returns Number of execution units per compute unit for given subtarget \p 168 /// STI. 169 unsigned getEUsPerCU(const MCSubtargetInfo *STI); 170 171 /// \returns Maximum number of work groups per compute unit for given subtarget 172 /// \p STI and limited by given \p FlatWorkGroupSize. 173 unsigned getMaxWorkGroupsPerCU(const MCSubtargetInfo *STI, 174 unsigned FlatWorkGroupSize); 175 176 /// \returns Minimum number of waves per execution unit for given subtarget \p 177 /// STI. 178 unsigned getMinWavesPerEU(const MCSubtargetInfo *STI); 179 180 /// \returns Maximum number of waves per execution unit for given subtarget \p 181 /// STI without any kind of limitation. 182 unsigned getMaxWavesPerEU(const MCSubtargetInfo *STI); 183 184 /// \returns Number of waves per execution unit required to support the given \p 185 /// FlatWorkGroupSize. 186 unsigned getWavesPerEUForWorkGroup(const MCSubtargetInfo *STI, 187 unsigned FlatWorkGroupSize); 188 189 /// \returns Minimum flat work group size for given subtarget \p STI. 190 unsigned getMinFlatWorkGroupSize(const MCSubtargetInfo *STI); 191 192 /// \returns Maximum flat work group size for given subtarget \p STI. 193 unsigned getMaxFlatWorkGroupSize(const MCSubtargetInfo *STI); 194 195 /// \returns Number of waves per work group for given subtarget \p STI and 196 /// \p FlatWorkGroupSize. 197 unsigned getWavesPerWorkGroup(const MCSubtargetInfo *STI, 198 unsigned FlatWorkGroupSize); 199 200 /// \returns SGPR allocation granularity for given subtarget \p STI. 201 unsigned getSGPRAllocGranule(const MCSubtargetInfo *STI); 202 203 /// \returns SGPR encoding granularity for given subtarget \p STI. 204 unsigned getSGPREncodingGranule(const MCSubtargetInfo *STI); 205 206 /// \returns Total number of SGPRs for given subtarget \p STI. 207 unsigned getTotalNumSGPRs(const MCSubtargetInfo *STI); 208 209 /// \returns Addressable number of SGPRs for given subtarget \p STI. 210 unsigned getAddressableNumSGPRs(const MCSubtargetInfo *STI); 211 212 /// \returns Minimum number of SGPRs that meets the given number of waves per 213 /// execution unit requirement for given subtarget \p STI. 214 unsigned getMinNumSGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU); 215 216 /// \returns Maximum number of SGPRs that meets the given number of waves per 217 /// execution unit requirement for given subtarget \p STI. 218 unsigned getMaxNumSGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU, 219 bool Addressable); 220 221 /// \returns Number of extra SGPRs implicitly required by given subtarget \p 222 /// STI when the given special registers are used. 223 unsigned getNumExtraSGPRs(const MCSubtargetInfo *STI, bool VCCUsed, 224 bool FlatScrUsed, bool XNACKUsed); 225 226 /// \returns Number of extra SGPRs implicitly required by given subtarget \p 227 /// STI when the given special registers are used. XNACK is inferred from 228 /// \p STI. 229 unsigned getNumExtraSGPRs(const MCSubtargetInfo *STI, bool VCCUsed, 230 bool FlatScrUsed); 231 232 /// \returns Number of SGPR blocks needed for given subtarget \p STI when 233 /// \p NumSGPRs are used. \p NumSGPRs should already include any special 234 /// register counts. 235 unsigned getNumSGPRBlocks(const MCSubtargetInfo *STI, unsigned NumSGPRs); 236 237 /// \returns VGPR allocation granularity for given subtarget \p STI. 238 /// 239 /// For subtargets which support it, \p EnableWavefrontSize32 should match 240 /// the ENABLE_WAVEFRONT_SIZE32 kernel descriptor field. 241 unsigned getVGPRAllocGranule(const MCSubtargetInfo *STI, 242 Optional<bool> EnableWavefrontSize32 = None); 243 244 /// \returns VGPR encoding granularity for given subtarget \p STI. 245 /// 246 /// For subtargets which support it, \p EnableWavefrontSize32 should match 247 /// the ENABLE_WAVEFRONT_SIZE32 kernel descriptor field. 248 unsigned getVGPREncodingGranule(const MCSubtargetInfo *STI, 249 Optional<bool> EnableWavefrontSize32 = None); 250 251 /// \returns Total number of VGPRs for given subtarget \p STI. 252 unsigned getTotalNumVGPRs(const MCSubtargetInfo *STI); 253 254 /// \returns Addressable number of VGPRs for given subtarget \p STI. 255 unsigned getAddressableNumVGPRs(const MCSubtargetInfo *STI); 256 257 /// \returns Minimum number of VGPRs that meets given number of waves per 258 /// execution unit requirement for given subtarget \p STI. 259 unsigned getMinNumVGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU); 260 261 /// \returns Maximum number of VGPRs that meets given number of waves per 262 /// execution unit requirement for given subtarget \p STI. 263 unsigned getMaxNumVGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU); 264 265 /// \returns Number of VGPR blocks needed for given subtarget \p STI when 266 /// \p NumVGPRs are used. 267 /// 268 /// For subtargets which support it, \p EnableWavefrontSize32 should match the 269 /// ENABLE_WAVEFRONT_SIZE32 kernel descriptor field. 270 unsigned getNumVGPRBlocks(const MCSubtargetInfo *STI, unsigned NumSGPRs, 271 Optional<bool> EnableWavefrontSize32 = None); 272 273 } // end namespace IsaInfo 274 275 LLVM_READONLY 276 int16_t getNamedOperandIdx(uint16_t Opcode, uint16_t NamedIdx); 277 278 LLVM_READONLY 279 int getSOPPWithRelaxation(uint16_t Opcode); 280 281 struct MIMGBaseOpcodeInfo { 282 MIMGBaseOpcode BaseOpcode; 283 bool Store; 284 bool Atomic; 285 bool AtomicX2; 286 bool Sampler; 287 bool Gather4; 288 289 uint8_t NumExtraArgs; 290 bool Gradients; 291 bool G16; 292 bool Coordinates; 293 bool LodOrClampOrMip; 294 bool HasD16; 295 bool MSAA; 296 bool BVH; 297 }; 298 299 LLVM_READONLY 300 const MIMGBaseOpcodeInfo *getMIMGBaseOpcode(unsigned Opc); 301 302 LLVM_READONLY 303 const MIMGBaseOpcodeInfo *getMIMGBaseOpcodeInfo(unsigned BaseOpcode); 304 305 struct MIMGDimInfo { 306 MIMGDim Dim; 307 uint8_t NumCoords; 308 uint8_t NumGradients; 309 bool MSAA; 310 bool DA; 311 uint8_t Encoding; 312 const char *AsmSuffix; 313 }; 314 315 LLVM_READONLY 316 const MIMGDimInfo *getMIMGDimInfo(unsigned DimEnum); 317 318 LLVM_READONLY 319 const MIMGDimInfo *getMIMGDimInfoByEncoding(uint8_t DimEnc); 320 321 LLVM_READONLY 322 const MIMGDimInfo *getMIMGDimInfoByAsmSuffix(StringRef AsmSuffix); 323 324 struct MIMGLZMappingInfo { 325 MIMGBaseOpcode L; 326 MIMGBaseOpcode LZ; 327 }; 328 329 struct MIMGMIPMappingInfo { 330 MIMGBaseOpcode MIP; 331 MIMGBaseOpcode NONMIP; 332 }; 333 334 struct MIMGBiasMappingInfo { 335 MIMGBaseOpcode Bias; 336 MIMGBaseOpcode NoBias; 337 }; 338 339 struct MIMGG16MappingInfo { 340 MIMGBaseOpcode G; 341 MIMGBaseOpcode G16; 342 }; 343 344 LLVM_READONLY 345 const MIMGLZMappingInfo *getMIMGLZMappingInfo(unsigned L); 346 347 LLVM_READONLY 348 const MIMGMIPMappingInfo *getMIMGMIPMappingInfo(unsigned MIP); 349 350 LLVM_READONLY 351 const MIMGBiasMappingInfo *getMIMGBiasMappingInfo(unsigned Bias); 352 353 LLVM_READONLY 354 const MIMGG16MappingInfo *getMIMGG16MappingInfo(unsigned G); 355 356 LLVM_READONLY 357 int getMIMGOpcode(unsigned BaseOpcode, unsigned MIMGEncoding, 358 unsigned VDataDwords, unsigned VAddrDwords); 359 360 LLVM_READONLY 361 int getMaskedMIMGOp(unsigned Opc, unsigned NewChannels); 362 363 LLVM_READONLY 364 unsigned getAddrSizeMIMGOp(const MIMGBaseOpcodeInfo *BaseOpcode, 365 const MIMGDimInfo *Dim, bool IsA16, 366 bool IsG16Supported); 367 368 struct MIMGInfo { 369 uint16_t Opcode; 370 uint16_t BaseOpcode; 371 uint8_t MIMGEncoding; 372 uint8_t VDataDwords; 373 uint8_t VAddrDwords; 374 }; 375 376 LLVM_READONLY 377 const MIMGInfo *getMIMGInfo(unsigned Opc); 378 379 LLVM_READONLY 380 int getMTBUFBaseOpcode(unsigned Opc); 381 382 LLVM_READONLY 383 int getMTBUFOpcode(unsigned BaseOpc, unsigned Elements); 384 385 LLVM_READONLY 386 int getMTBUFElements(unsigned Opc); 387 388 LLVM_READONLY 389 bool getMTBUFHasVAddr(unsigned Opc); 390 391 LLVM_READONLY 392 bool getMTBUFHasSrsrc(unsigned Opc); 393 394 LLVM_READONLY 395 bool getMTBUFHasSoffset(unsigned Opc); 396 397 LLVM_READONLY 398 int getMUBUFBaseOpcode(unsigned Opc); 399 400 LLVM_READONLY 401 int getMUBUFOpcode(unsigned BaseOpc, unsigned Elements); 402 403 LLVM_READONLY 404 int getMUBUFElements(unsigned Opc); 405 406 LLVM_READONLY 407 bool getMUBUFHasVAddr(unsigned Opc); 408 409 LLVM_READONLY 410 bool getMUBUFHasSrsrc(unsigned Opc); 411 412 LLVM_READONLY 413 bool getMUBUFHasSoffset(unsigned Opc); 414 415 LLVM_READONLY 416 bool getMUBUFIsBufferInv(unsigned Opc); 417 418 LLVM_READONLY 419 bool getSMEMIsBuffer(unsigned Opc); 420 421 LLVM_READONLY 422 bool getVOP1IsSingle(unsigned Opc); 423 424 LLVM_READONLY 425 bool getVOP2IsSingle(unsigned Opc); 426 427 LLVM_READONLY 428 bool getVOP3IsSingle(unsigned Opc); 429 430 LLVM_READONLY 431 const GcnBufferFormatInfo *getGcnBufferFormatInfo(uint8_t BitsPerComp, 432 uint8_t NumComponents, 433 uint8_t NumFormat, 434 const MCSubtargetInfo &STI); 435 LLVM_READONLY 436 const GcnBufferFormatInfo *getGcnBufferFormatInfo(uint8_t Format, 437 const MCSubtargetInfo &STI); 438 439 LLVM_READONLY 440 int getMCOpcode(uint16_t Opcode, unsigned Gen); 441 442 void initDefaultAMDKernelCodeT(amd_kernel_code_t &Header, 443 const MCSubtargetInfo *STI); 444 445 amdhsa::kernel_descriptor_t getDefaultAmdhsaKernelDescriptor( 446 const MCSubtargetInfo *STI); 447 448 bool isGroupSegment(const GlobalValue *GV); 449 bool isGlobalSegment(const GlobalValue *GV); 450 bool isReadOnlySegment(const GlobalValue *GV); 451 452 /// \returns True if constants should be emitted to .text section for given 453 /// target triple \p TT, false otherwise. 454 bool shouldEmitConstantsToTextSection(const Triple &TT); 455 456 /// \returns Integer value requested using \p F's \p Name attribute. 457 /// 458 /// \returns \p Default if attribute is not present. 459 /// 460 /// \returns \p Default and emits error if requested value cannot be converted 461 /// to integer. 462 int getIntegerAttribute(const Function &F, StringRef Name, int Default); 463 464 /// \returns A pair of integer values requested using \p F's \p Name attribute 465 /// in "first[,second]" format ("second" is optional unless \p OnlyFirstRequired 466 /// is false). 467 /// 468 /// \returns \p Default if attribute is not present. 469 /// 470 /// \returns \p Default and emits error if one of the requested values cannot be 471 /// converted to integer, or \p OnlyFirstRequired is false and "second" value is 472 /// not present. 473 std::pair<int, int> getIntegerPairAttribute(const Function &F, 474 StringRef Name, 475 std::pair<int, int> Default, 476 bool OnlyFirstRequired = false); 477 478 /// Represents the counter values to wait for in an s_waitcnt instruction. 479 /// 480 /// Large values (including the maximum possible integer) can be used to 481 /// represent "don't care" waits. 482 struct Waitcnt { 483 unsigned VmCnt = ~0u; 484 unsigned ExpCnt = ~0u; 485 unsigned LgkmCnt = ~0u; 486 unsigned VsCnt = ~0u; 487 488 Waitcnt() {} 489 Waitcnt(unsigned VmCnt, unsigned ExpCnt, unsigned LgkmCnt, unsigned VsCnt) 490 : VmCnt(VmCnt), ExpCnt(ExpCnt), LgkmCnt(LgkmCnt), VsCnt(VsCnt) {} 491 492 static Waitcnt allZero(bool HasVscnt) { 493 return Waitcnt(0, 0, 0, HasVscnt ? 0 : ~0u); 494 } 495 static Waitcnt allZeroExceptVsCnt() { return Waitcnt(0, 0, 0, ~0u); } 496 497 bool hasWait() const { 498 return VmCnt != ~0u || ExpCnt != ~0u || LgkmCnt != ~0u || VsCnt != ~0u; 499 } 500 501 bool hasWaitExceptVsCnt() const { 502 return VmCnt != ~0u || ExpCnt != ~0u || LgkmCnt != ~0u; 503 } 504 505 bool hasWaitVsCnt() const { 506 return VsCnt != ~0u; 507 } 508 509 bool dominates(const Waitcnt &Other) const { 510 return VmCnt <= Other.VmCnt && ExpCnt <= Other.ExpCnt && 511 LgkmCnt <= Other.LgkmCnt && VsCnt <= Other.VsCnt; 512 } 513 514 Waitcnt combined(const Waitcnt &Other) const { 515 return Waitcnt(std::min(VmCnt, Other.VmCnt), std::min(ExpCnt, Other.ExpCnt), 516 std::min(LgkmCnt, Other.LgkmCnt), 517 std::min(VsCnt, Other.VsCnt)); 518 } 519 }; 520 521 /// \returns Vmcnt bit mask for given isa \p Version. 522 unsigned getVmcntBitMask(const IsaVersion &Version); 523 524 /// \returns Expcnt bit mask for given isa \p Version. 525 unsigned getExpcntBitMask(const IsaVersion &Version); 526 527 /// \returns Lgkmcnt bit mask for given isa \p Version. 528 unsigned getLgkmcntBitMask(const IsaVersion &Version); 529 530 /// \returns Waitcnt bit mask for given isa \p Version. 531 unsigned getWaitcntBitMask(const IsaVersion &Version); 532 533 /// \returns Decoded Vmcnt from given \p Waitcnt for given isa \p Version. 534 unsigned decodeVmcnt(const IsaVersion &Version, unsigned Waitcnt); 535 536 /// \returns Decoded Expcnt from given \p Waitcnt for given isa \p Version. 537 unsigned decodeExpcnt(const IsaVersion &Version, unsigned Waitcnt); 538 539 /// \returns Decoded Lgkmcnt from given \p Waitcnt for given isa \p Version. 540 unsigned decodeLgkmcnt(const IsaVersion &Version, unsigned Waitcnt); 541 542 /// Decodes Vmcnt, Expcnt and Lgkmcnt from given \p Waitcnt for given isa 543 /// \p Version, and writes decoded values into \p Vmcnt, \p Expcnt and 544 /// \p Lgkmcnt respectively. 545 /// 546 /// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are decoded as follows: 547 /// \p Vmcnt = \p Waitcnt[3:0] (pre-gfx9 only) 548 /// \p Vmcnt = \p Waitcnt[3:0] | \p Waitcnt[15:14] (gfx9+ only) 549 /// \p Expcnt = \p Waitcnt[6:4] 550 /// \p Lgkmcnt = \p Waitcnt[11:8] (pre-gfx10 only) 551 /// \p Lgkmcnt = \p Waitcnt[13:8] (gfx10+ only) 552 void decodeWaitcnt(const IsaVersion &Version, unsigned Waitcnt, 553 unsigned &Vmcnt, unsigned &Expcnt, unsigned &Lgkmcnt); 554 555 Waitcnt decodeWaitcnt(const IsaVersion &Version, unsigned Encoded); 556 557 /// \returns \p Waitcnt with encoded \p Vmcnt for given isa \p Version. 558 unsigned encodeVmcnt(const IsaVersion &Version, unsigned Waitcnt, 559 unsigned Vmcnt); 560 561 /// \returns \p Waitcnt with encoded \p Expcnt for given isa \p Version. 562 unsigned encodeExpcnt(const IsaVersion &Version, unsigned Waitcnt, 563 unsigned Expcnt); 564 565 /// \returns \p Waitcnt with encoded \p Lgkmcnt for given isa \p Version. 566 unsigned encodeLgkmcnt(const IsaVersion &Version, unsigned Waitcnt, 567 unsigned Lgkmcnt); 568 569 /// Encodes \p Vmcnt, \p Expcnt and \p Lgkmcnt into Waitcnt for given isa 570 /// \p Version. 571 /// 572 /// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are encoded as follows: 573 /// Waitcnt[3:0] = \p Vmcnt (pre-gfx9 only) 574 /// Waitcnt[3:0] = \p Vmcnt[3:0] (gfx9+ only) 575 /// Waitcnt[6:4] = \p Expcnt 576 /// Waitcnt[11:8] = \p Lgkmcnt (pre-gfx10 only) 577 /// Waitcnt[13:8] = \p Lgkmcnt (gfx10+ only) 578 /// Waitcnt[15:14] = \p Vmcnt[5:4] (gfx9+ only) 579 /// 580 /// \returns Waitcnt with encoded \p Vmcnt, \p Expcnt and \p Lgkmcnt for given 581 /// isa \p Version. 582 unsigned encodeWaitcnt(const IsaVersion &Version, 583 unsigned Vmcnt, unsigned Expcnt, unsigned Lgkmcnt); 584 585 unsigned encodeWaitcnt(const IsaVersion &Version, const Waitcnt &Decoded); 586 587 namespace Hwreg { 588 589 LLVM_READONLY 590 int64_t getHwregId(const StringRef Name); 591 592 LLVM_READNONE 593 bool isValidHwreg(int64_t Id, const MCSubtargetInfo &STI); 594 595 LLVM_READNONE 596 bool isValidHwreg(int64_t Id); 597 598 LLVM_READNONE 599 bool isValidHwregOffset(int64_t Offset); 600 601 LLVM_READNONE 602 bool isValidHwregWidth(int64_t Width); 603 604 LLVM_READNONE 605 uint64_t encodeHwreg(uint64_t Id, uint64_t Offset, uint64_t Width); 606 607 LLVM_READNONE 608 StringRef getHwreg(unsigned Id, const MCSubtargetInfo &STI); 609 610 void decodeHwreg(unsigned Val, unsigned &Id, unsigned &Offset, unsigned &Width); 611 612 } // namespace Hwreg 613 614 namespace Exp { 615 616 bool getTgtName(unsigned Id, StringRef &Name, int &Index); 617 618 LLVM_READONLY 619 unsigned getTgtId(const StringRef Name); 620 621 LLVM_READNONE 622 bool isSupportedTgtId(unsigned Id, const MCSubtargetInfo &STI); 623 624 } // namespace Exp 625 626 namespace MTBUFFormat { 627 628 LLVM_READNONE 629 int64_t encodeDfmtNfmt(unsigned Dfmt, unsigned Nfmt); 630 631 void decodeDfmtNfmt(unsigned Format, unsigned &Dfmt, unsigned &Nfmt); 632 633 int64_t getDfmt(const StringRef Name); 634 635 StringRef getDfmtName(unsigned Id); 636 637 int64_t getNfmt(const StringRef Name, const MCSubtargetInfo &STI); 638 639 StringRef getNfmtName(unsigned Id, const MCSubtargetInfo &STI); 640 641 bool isValidDfmtNfmt(unsigned Val, const MCSubtargetInfo &STI); 642 643 bool isValidNfmt(unsigned Val, const MCSubtargetInfo &STI); 644 645 int64_t getUnifiedFormat(const StringRef Name); 646 647 StringRef getUnifiedFormatName(unsigned Id); 648 649 bool isValidUnifiedFormat(unsigned Val); 650 651 int64_t convertDfmtNfmt2Ufmt(unsigned Dfmt, unsigned Nfmt); 652 653 bool isValidFormatEncoding(unsigned Val, const MCSubtargetInfo &STI); 654 655 unsigned getDefaultFormatEncoding(const MCSubtargetInfo &STI); 656 657 } // namespace MTBUFFormat 658 659 namespace SendMsg { 660 661 LLVM_READONLY 662 int64_t getMsgId(const StringRef Name); 663 664 LLVM_READONLY 665 int64_t getMsgOpId(int64_t MsgId, const StringRef Name); 666 667 LLVM_READNONE 668 StringRef getMsgName(int64_t MsgId); 669 670 LLVM_READNONE 671 StringRef getMsgOpName(int64_t MsgId, int64_t OpId); 672 673 LLVM_READNONE 674 bool isValidMsgId(int64_t MsgId, const MCSubtargetInfo &STI, bool Strict = true); 675 676 LLVM_READNONE 677 bool isValidMsgOp(int64_t MsgId, int64_t OpId, const MCSubtargetInfo &STI, 678 bool Strict = true); 679 680 LLVM_READNONE 681 bool isValidMsgStream(int64_t MsgId, int64_t OpId, int64_t StreamId, 682 const MCSubtargetInfo &STI, bool Strict = true); 683 684 LLVM_READNONE 685 bool msgRequiresOp(int64_t MsgId); 686 687 LLVM_READNONE 688 bool msgSupportsStream(int64_t MsgId, int64_t OpId); 689 690 void decodeMsg(unsigned Val, 691 uint16_t &MsgId, 692 uint16_t &OpId, 693 uint16_t &StreamId); 694 695 LLVM_READNONE 696 uint64_t encodeMsg(uint64_t MsgId, 697 uint64_t OpId, 698 uint64_t StreamId); 699 700 } // namespace SendMsg 701 702 703 unsigned getInitialPSInputAddr(const Function &F); 704 705 bool getHasColorExport(const Function &F); 706 707 bool getHasDepthExport(const Function &F); 708 709 LLVM_READNONE 710 bool isShader(CallingConv::ID CC); 711 712 LLVM_READNONE 713 bool isGraphics(CallingConv::ID CC); 714 715 LLVM_READNONE 716 bool isCompute(CallingConv::ID CC); 717 718 LLVM_READNONE 719 bool isEntryFunctionCC(CallingConv::ID CC); 720 721 // These functions are considered entrypoints into the current module, i.e. they 722 // are allowed to be called from outside the current module. This is different 723 // from isEntryFunctionCC, which is only true for functions that are entered by 724 // the hardware. Module entry points include all entry functions but also 725 // include functions that can be called from other functions inside or outside 726 // the current module. Module entry functions are allowed to allocate LDS. 727 LLVM_READNONE 728 bool isModuleEntryFunctionCC(CallingConv::ID CC); 729 730 // FIXME: Remove this when calling conventions cleaned up 731 LLVM_READNONE 732 inline bool isKernel(CallingConv::ID CC) { 733 switch (CC) { 734 case CallingConv::AMDGPU_KERNEL: 735 case CallingConv::SPIR_KERNEL: 736 return true; 737 default: 738 return false; 739 } 740 } 741 742 bool hasXNACK(const MCSubtargetInfo &STI); 743 bool hasSRAMECC(const MCSubtargetInfo &STI); 744 bool hasMIMG_R128(const MCSubtargetInfo &STI); 745 bool hasGFX10A16(const MCSubtargetInfo &STI); 746 bool hasG16(const MCSubtargetInfo &STI); 747 bool hasPackedD16(const MCSubtargetInfo &STI); 748 749 bool isSI(const MCSubtargetInfo &STI); 750 bool isCI(const MCSubtargetInfo &STI); 751 bool isVI(const MCSubtargetInfo &STI); 752 bool isGFX9(const MCSubtargetInfo &STI); 753 bool isGFX9Plus(const MCSubtargetInfo &STI); 754 bool isGFX10(const MCSubtargetInfo &STI); 755 bool isGFX10Plus(const MCSubtargetInfo &STI); 756 bool isGCN3Encoding(const MCSubtargetInfo &STI); 757 bool isGFX10_AEncoding(const MCSubtargetInfo &STI); 758 bool isGFX10_BEncoding(const MCSubtargetInfo &STI); 759 bool hasGFX10_3Insts(const MCSubtargetInfo &STI); 760 bool isGFX90A(const MCSubtargetInfo &STI); 761 bool hasArchitectedFlatScratch(const MCSubtargetInfo &STI); 762 763 /// Is Reg - scalar register 764 bool isSGPR(unsigned Reg, const MCRegisterInfo* TRI); 765 766 /// Is there any intersection between registers 767 bool isRegIntersect(unsigned Reg0, unsigned Reg1, const MCRegisterInfo* TRI); 768 769 /// If \p Reg is a pseudo reg, return the correct hardware register given 770 /// \p STI otherwise return \p Reg. 771 unsigned getMCReg(unsigned Reg, const MCSubtargetInfo &STI); 772 773 /// Convert hardware register \p Reg to a pseudo register 774 LLVM_READNONE 775 unsigned mc2PseudoReg(unsigned Reg); 776 777 /// Can this operand also contain immediate values? 778 bool isSISrcOperand(const MCInstrDesc &Desc, unsigned OpNo); 779 780 /// Is this floating-point operand? 781 bool isSISrcFPOperand(const MCInstrDesc &Desc, unsigned OpNo); 782 783 /// Does this operand support only inlinable literals? 784 bool isSISrcInlinableOperand(const MCInstrDesc &Desc, unsigned OpNo); 785 786 /// Get the size in bits of a register from the register class \p RC. 787 unsigned getRegBitWidth(unsigned RCID); 788 789 /// Get the size in bits of a register from the register class \p RC. 790 unsigned getRegBitWidth(const MCRegisterClass &RC); 791 792 /// Get size of register operand 793 unsigned getRegOperandSize(const MCRegisterInfo *MRI, const MCInstrDesc &Desc, 794 unsigned OpNo); 795 796 LLVM_READNONE 797 inline unsigned getOperandSize(const MCOperandInfo &OpInfo) { 798 switch (OpInfo.OperandType) { 799 case AMDGPU::OPERAND_REG_IMM_INT32: 800 case AMDGPU::OPERAND_REG_IMM_FP32: 801 case AMDGPU::OPERAND_REG_IMM_FP32_DEFERRED: 802 case AMDGPU::OPERAND_REG_INLINE_C_INT32: 803 case AMDGPU::OPERAND_REG_INLINE_C_FP32: 804 case AMDGPU::OPERAND_REG_INLINE_AC_INT32: 805 case AMDGPU::OPERAND_REG_INLINE_AC_FP32: 806 case AMDGPU::OPERAND_REG_IMM_V2INT32: 807 case AMDGPU::OPERAND_REG_IMM_V2FP32: 808 case AMDGPU::OPERAND_REG_INLINE_C_V2INT32: 809 case AMDGPU::OPERAND_REG_INLINE_C_V2FP32: 810 case AMDGPU::OPERAND_KIMM32: 811 case AMDGPU::OPERAND_KIMM16: // mandatory literal is always size 4 812 return 4; 813 814 case AMDGPU::OPERAND_REG_IMM_INT64: 815 case AMDGPU::OPERAND_REG_IMM_FP64: 816 case AMDGPU::OPERAND_REG_INLINE_C_INT64: 817 case AMDGPU::OPERAND_REG_INLINE_C_FP64: 818 case AMDGPU::OPERAND_REG_INLINE_AC_FP64: 819 return 8; 820 821 case AMDGPU::OPERAND_REG_IMM_INT16: 822 case AMDGPU::OPERAND_REG_IMM_FP16: 823 case AMDGPU::OPERAND_REG_IMM_FP16_DEFERRED: 824 case AMDGPU::OPERAND_REG_INLINE_C_INT16: 825 case AMDGPU::OPERAND_REG_INLINE_C_FP16: 826 case AMDGPU::OPERAND_REG_INLINE_C_V2INT16: 827 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16: 828 case AMDGPU::OPERAND_REG_INLINE_AC_INT16: 829 case AMDGPU::OPERAND_REG_INLINE_AC_FP16: 830 case AMDGPU::OPERAND_REG_INLINE_AC_V2INT16: 831 case AMDGPU::OPERAND_REG_INLINE_AC_V2FP16: 832 case AMDGPU::OPERAND_REG_IMM_V2INT16: 833 case AMDGPU::OPERAND_REG_IMM_V2FP16: 834 return 2; 835 836 default: 837 llvm_unreachable("unhandled operand type"); 838 } 839 } 840 841 LLVM_READNONE 842 inline unsigned getOperandSize(const MCInstrDesc &Desc, unsigned OpNo) { 843 return getOperandSize(Desc.OpInfo[OpNo]); 844 } 845 846 /// Is this literal inlinable, and not one of the values intended for floating 847 /// point values. 848 LLVM_READNONE 849 inline bool isInlinableIntLiteral(int64_t Literal) { 850 return Literal >= -16 && Literal <= 64; 851 } 852 853 /// Is this literal inlinable 854 LLVM_READNONE 855 bool isInlinableLiteral64(int64_t Literal, bool HasInv2Pi); 856 857 LLVM_READNONE 858 bool isInlinableLiteral32(int32_t Literal, bool HasInv2Pi); 859 860 LLVM_READNONE 861 bool isInlinableLiteral16(int16_t Literal, bool HasInv2Pi); 862 863 LLVM_READNONE 864 bool isInlinableLiteralV216(int32_t Literal, bool HasInv2Pi); 865 866 LLVM_READNONE 867 bool isInlinableIntLiteralV216(int32_t Literal); 868 869 LLVM_READNONE 870 bool isFoldableLiteralV216(int32_t Literal, bool HasInv2Pi); 871 872 bool isArgPassedInSGPR(const Argument *Arg); 873 874 LLVM_READONLY 875 bool isLegalSMRDEncodedUnsignedOffset(const MCSubtargetInfo &ST, 876 int64_t EncodedOffset); 877 878 LLVM_READONLY 879 bool isLegalSMRDEncodedSignedOffset(const MCSubtargetInfo &ST, 880 int64_t EncodedOffset, 881 bool IsBuffer); 882 883 /// Convert \p ByteOffset to dwords if the subtarget uses dword SMRD immediate 884 /// offsets. 885 uint64_t convertSMRDOffsetUnits(const MCSubtargetInfo &ST, uint64_t ByteOffset); 886 887 /// \returns The encoding that will be used for \p ByteOffset in the 888 /// SMRD offset field, or None if it won't fit. On GFX9 and GFX10 889 /// S_LOAD instructions have a signed offset, on other subtargets it is 890 /// unsigned. S_BUFFER has an unsigned offset for all subtargets. 891 Optional<int64_t> getSMRDEncodedOffset(const MCSubtargetInfo &ST, 892 int64_t ByteOffset, bool IsBuffer); 893 894 /// \return The encoding that can be used for a 32-bit literal offset in an SMRD 895 /// instruction. This is only useful on CI.s 896 Optional<int64_t> getSMRDEncodedLiteralOffset32(const MCSubtargetInfo &ST, 897 int64_t ByteOffset); 898 899 /// For FLAT segment the offset must be positive; 900 /// MSB is ignored and forced to zero. 901 /// 902 /// \return The number of bits available for the offset field in flat 903 /// instructions. 904 unsigned getNumFlatOffsetBits(const MCSubtargetInfo &ST, bool Signed); 905 906 /// \returns true if this offset is small enough to fit in the SMRD 907 /// offset field. \p ByteOffset should be the offset in bytes and 908 /// not the encoded offset. 909 bool isLegalSMRDImmOffset(const MCSubtargetInfo &ST, int64_t ByteOffset); 910 911 bool splitMUBUFOffset(uint32_t Imm, uint32_t &SOffset, uint32_t &ImmOffset, 912 const GCNSubtarget *Subtarget, 913 Align Alignment = Align(4)); 914 915 LLVM_READNONE 916 inline bool isLegal64BitDPPControl(unsigned DC) { 917 return DC >= DPP::ROW_NEWBCAST_FIRST && DC <= DPP::ROW_NEWBCAST_LAST; 918 } 919 920 /// \returns true if the intrinsic is divergent 921 bool isIntrinsicSourceOfDivergence(unsigned IntrID); 922 923 // Track defaults for fields in the MODE registser. 924 struct SIModeRegisterDefaults { 925 /// Floating point opcodes that support exception flag gathering quiet and 926 /// propagate signaling NaN inputs per IEEE 754-2008. Min_dx10 and max_dx10 927 /// become IEEE 754- 2008 compliant due to signaling NaN propagation and 928 /// quieting. 929 bool IEEE : 1; 930 931 /// Used by the vector ALU to force DX10-style treatment of NaNs: when set, 932 /// clamp NaN to zero; otherwise, pass NaN through. 933 bool DX10Clamp : 1; 934 935 /// If this is set, neither input or output denormals are flushed for most f32 936 /// instructions. 937 bool FP32InputDenormals : 1; 938 bool FP32OutputDenormals : 1; 939 940 /// If this is set, neither input or output denormals are flushed for both f64 941 /// and f16/v2f16 instructions. 942 bool FP64FP16InputDenormals : 1; 943 bool FP64FP16OutputDenormals : 1; 944 945 SIModeRegisterDefaults() : 946 IEEE(true), 947 DX10Clamp(true), 948 FP32InputDenormals(true), 949 FP32OutputDenormals(true), 950 FP64FP16InputDenormals(true), 951 FP64FP16OutputDenormals(true) {} 952 953 SIModeRegisterDefaults(const Function &F); 954 955 static SIModeRegisterDefaults getDefaultForCallingConv(CallingConv::ID CC) { 956 SIModeRegisterDefaults Mode; 957 Mode.IEEE = !AMDGPU::isShader(CC); 958 return Mode; 959 } 960 961 bool operator ==(const SIModeRegisterDefaults Other) const { 962 return IEEE == Other.IEEE && DX10Clamp == Other.DX10Clamp && 963 FP32InputDenormals == Other.FP32InputDenormals && 964 FP32OutputDenormals == Other.FP32OutputDenormals && 965 FP64FP16InputDenormals == Other.FP64FP16InputDenormals && 966 FP64FP16OutputDenormals == Other.FP64FP16OutputDenormals; 967 } 968 969 bool allFP32Denormals() const { 970 return FP32InputDenormals && FP32OutputDenormals; 971 } 972 973 bool allFP64FP16Denormals() const { 974 return FP64FP16InputDenormals && FP64FP16OutputDenormals; 975 } 976 977 /// Get the encoding value for the FP_DENORM bits of the mode register for the 978 /// FP32 denormal mode. 979 uint32_t fpDenormModeSPValue() const { 980 if (FP32InputDenormals && FP32OutputDenormals) 981 return FP_DENORM_FLUSH_NONE; 982 if (FP32InputDenormals) 983 return FP_DENORM_FLUSH_OUT; 984 if (FP32OutputDenormals) 985 return FP_DENORM_FLUSH_IN; 986 return FP_DENORM_FLUSH_IN_FLUSH_OUT; 987 } 988 989 /// Get the encoding value for the FP_DENORM bits of the mode register for the 990 /// FP64/FP16 denormal mode. 991 uint32_t fpDenormModeDPValue() const { 992 if (FP64FP16InputDenormals && FP64FP16OutputDenormals) 993 return FP_DENORM_FLUSH_NONE; 994 if (FP64FP16InputDenormals) 995 return FP_DENORM_FLUSH_OUT; 996 if (FP64FP16OutputDenormals) 997 return FP_DENORM_FLUSH_IN; 998 return FP_DENORM_FLUSH_IN_FLUSH_OUT; 999 } 1000 1001 /// Returns true if a flag is compatible if it's enabled in the callee, but 1002 /// disabled in the caller. 1003 static bool oneWayCompatible(bool CallerMode, bool CalleeMode) { 1004 return CallerMode == CalleeMode || (!CallerMode && CalleeMode); 1005 } 1006 1007 // FIXME: Inlining should be OK for dx10-clamp, since the caller's mode should 1008 // be able to override. 1009 bool isInlineCompatible(SIModeRegisterDefaults CalleeMode) const { 1010 if (DX10Clamp != CalleeMode.DX10Clamp) 1011 return false; 1012 if (IEEE != CalleeMode.IEEE) 1013 return false; 1014 1015 // Allow inlining denormals enabled into denormals flushed functions. 1016 return oneWayCompatible(FP64FP16InputDenormals, CalleeMode.FP64FP16InputDenormals) && 1017 oneWayCompatible(FP64FP16OutputDenormals, CalleeMode.FP64FP16OutputDenormals) && 1018 oneWayCompatible(FP32InputDenormals, CalleeMode.FP32InputDenormals) && 1019 oneWayCompatible(FP32OutputDenormals, CalleeMode.FP32OutputDenormals); 1020 } 1021 }; 1022 1023 } // end namespace AMDGPU 1024 1025 raw_ostream &operator<<(raw_ostream &OS, 1026 const AMDGPU::IsaInfo::TargetIDSetting S); 1027 1028 } // end namespace llvm 1029 1030 #endif // LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUBASEINFO_H 1031