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 5, 51 /// false otherwise. 52 bool isHsaAbiVersion5(const MCSubtargetInfo *STI); 53 /// \returns True if HSA OS ABI Version identification is 3 and above, 54 /// false otherwise. 55 bool isHsaAbiVersion3AndAbove(const MCSubtargetInfo *STI); 56 57 /// \returns The offset of the hostcall pointer argument from implicitarg_ptr 58 unsigned getHostcallImplicitArgPosition(); 59 60 /// \returns The offset of the heap ptr argument from implicitarg_ptr 61 unsigned getHeapPtrImplicitArgPosition(); 62 63 /// \returns Code object version. 64 unsigned getAmdhsaCodeObjectVersion(); 65 66 struct GcnBufferFormatInfo { 67 unsigned Format; 68 unsigned BitsPerComp; 69 unsigned NumComponents; 70 unsigned NumFormat; 71 unsigned DataFormat; 72 }; 73 74 #define GET_MIMGBaseOpcode_DECL 75 #define GET_MIMGDim_DECL 76 #define GET_MIMGEncoding_DECL 77 #define GET_MIMGLZMapping_DECL 78 #define GET_MIMGMIPMapping_DECL 79 #define GET_MIMGBiASMapping_DECL 80 #include "AMDGPUGenSearchableTables.inc" 81 82 namespace IsaInfo { 83 84 enum { 85 // The closed Vulkan driver sets 96, which limits the wave count to 8 but 86 // doesn't spill SGPRs as much as when 80 is set. 87 FIXED_NUM_SGPRS_FOR_INIT_BUG = 96, 88 TRAP_NUM_SGPRS = 16 89 }; 90 91 enum class TargetIDSetting { 92 Unsupported, 93 Any, 94 Off, 95 On 96 }; 97 98 class AMDGPUTargetID { 99 private: 100 const MCSubtargetInfo &STI; 101 TargetIDSetting XnackSetting; 102 TargetIDSetting SramEccSetting; 103 104 public: 105 explicit AMDGPUTargetID(const MCSubtargetInfo &STI); 106 ~AMDGPUTargetID() = default; 107 108 /// \return True if the current xnack setting is not "Unsupported". 109 bool isXnackSupported() const { 110 return XnackSetting != TargetIDSetting::Unsupported; 111 } 112 113 /// \returns True if the current xnack setting is "On" or "Any". 114 bool isXnackOnOrAny() const { 115 return XnackSetting == TargetIDSetting::On || 116 XnackSetting == TargetIDSetting::Any; 117 } 118 119 /// \returns True if current xnack setting is "On" or "Off", 120 /// false otherwise. 121 bool isXnackOnOrOff() const { 122 return getXnackSetting() == TargetIDSetting::On || 123 getXnackSetting() == TargetIDSetting::Off; 124 } 125 126 /// \returns The current xnack TargetIDSetting, possible options are 127 /// "Unsupported", "Any", "Off", and "On". 128 TargetIDSetting getXnackSetting() const { 129 return XnackSetting; 130 } 131 132 /// Sets xnack setting to \p NewXnackSetting. 133 void setXnackSetting(TargetIDSetting NewXnackSetting) { 134 XnackSetting = NewXnackSetting; 135 } 136 137 /// \return True if the current sramecc setting is not "Unsupported". 138 bool isSramEccSupported() const { 139 return SramEccSetting != TargetIDSetting::Unsupported; 140 } 141 142 /// \returns True if the current sramecc setting is "On" or "Any". 143 bool isSramEccOnOrAny() const { 144 return SramEccSetting == TargetIDSetting::On || 145 SramEccSetting == TargetIDSetting::Any; 146 } 147 148 /// \returns True if current sramecc setting is "On" or "Off", 149 /// false otherwise. 150 bool isSramEccOnOrOff() const { 151 return getSramEccSetting() == TargetIDSetting::On || 152 getSramEccSetting() == TargetIDSetting::Off; 153 } 154 155 /// \returns The current sramecc TargetIDSetting, possible options are 156 /// "Unsupported", "Any", "Off", and "On". 157 TargetIDSetting getSramEccSetting() const { 158 return SramEccSetting; 159 } 160 161 /// Sets sramecc setting to \p NewSramEccSetting. 162 void setSramEccSetting(TargetIDSetting NewSramEccSetting) { 163 SramEccSetting = NewSramEccSetting; 164 } 165 166 void setTargetIDFromFeaturesString(StringRef FS); 167 void setTargetIDFromTargetIDStream(StringRef TargetID); 168 169 /// \returns String representation of an object. 170 std::string toString() const; 171 }; 172 173 /// \returns Wavefront size for given subtarget \p STI. 174 unsigned getWavefrontSize(const MCSubtargetInfo *STI); 175 176 /// \returns Local memory size in bytes for given subtarget \p STI. 177 unsigned getLocalMemorySize(const MCSubtargetInfo *STI); 178 179 /// \returns Number of execution units per compute unit for given subtarget \p 180 /// STI. 181 unsigned getEUsPerCU(const MCSubtargetInfo *STI); 182 183 /// \returns Maximum number of work groups per compute unit for given subtarget 184 /// \p STI and limited by given \p FlatWorkGroupSize. 185 unsigned getMaxWorkGroupsPerCU(const MCSubtargetInfo *STI, 186 unsigned FlatWorkGroupSize); 187 188 /// \returns Minimum number of waves per execution unit for given subtarget \p 189 /// STI. 190 unsigned getMinWavesPerEU(const MCSubtargetInfo *STI); 191 192 /// \returns Maximum number of waves per execution unit for given subtarget \p 193 /// STI without any kind of limitation. 194 unsigned getMaxWavesPerEU(const MCSubtargetInfo *STI); 195 196 /// \returns Number of waves per execution unit required to support the given \p 197 /// FlatWorkGroupSize. 198 unsigned getWavesPerEUForWorkGroup(const MCSubtargetInfo *STI, 199 unsigned FlatWorkGroupSize); 200 201 /// \returns Minimum flat work group size for given subtarget \p STI. 202 unsigned getMinFlatWorkGroupSize(const MCSubtargetInfo *STI); 203 204 /// \returns Maximum flat work group size for given subtarget \p STI. 205 unsigned getMaxFlatWorkGroupSize(const MCSubtargetInfo *STI); 206 207 /// \returns Number of waves per work group for given subtarget \p STI and 208 /// \p FlatWorkGroupSize. 209 unsigned getWavesPerWorkGroup(const MCSubtargetInfo *STI, 210 unsigned FlatWorkGroupSize); 211 212 /// \returns SGPR allocation granularity for given subtarget \p STI. 213 unsigned getSGPRAllocGranule(const MCSubtargetInfo *STI); 214 215 /// \returns SGPR encoding granularity for given subtarget \p STI. 216 unsigned getSGPREncodingGranule(const MCSubtargetInfo *STI); 217 218 /// \returns Total number of SGPRs for given subtarget \p STI. 219 unsigned getTotalNumSGPRs(const MCSubtargetInfo *STI); 220 221 /// \returns Addressable number of SGPRs for given subtarget \p STI. 222 unsigned getAddressableNumSGPRs(const MCSubtargetInfo *STI); 223 224 /// \returns Minimum number of SGPRs that meets the given number of waves per 225 /// execution unit requirement for given subtarget \p STI. 226 unsigned getMinNumSGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU); 227 228 /// \returns Maximum number of SGPRs that meets the given number of waves per 229 /// execution unit requirement for given subtarget \p STI. 230 unsigned getMaxNumSGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU, 231 bool Addressable); 232 233 /// \returns Number of extra SGPRs implicitly required by given subtarget \p 234 /// STI when the given special registers are used. 235 unsigned getNumExtraSGPRs(const MCSubtargetInfo *STI, bool VCCUsed, 236 bool FlatScrUsed, bool XNACKUsed); 237 238 /// \returns Number of extra SGPRs implicitly required by given subtarget \p 239 /// STI when the given special registers are used. XNACK is inferred from 240 /// \p STI. 241 unsigned getNumExtraSGPRs(const MCSubtargetInfo *STI, bool VCCUsed, 242 bool FlatScrUsed); 243 244 /// \returns Number of SGPR blocks needed for given subtarget \p STI when 245 /// \p NumSGPRs are used. \p NumSGPRs should already include any special 246 /// register counts. 247 unsigned getNumSGPRBlocks(const MCSubtargetInfo *STI, unsigned NumSGPRs); 248 249 /// \returns VGPR allocation granularity for given subtarget \p STI. 250 /// 251 /// For subtargets which support it, \p EnableWavefrontSize32 should match 252 /// the ENABLE_WAVEFRONT_SIZE32 kernel descriptor field. 253 unsigned getVGPRAllocGranule(const MCSubtargetInfo *STI, 254 Optional<bool> EnableWavefrontSize32 = None); 255 256 /// \returns VGPR encoding granularity for given subtarget \p STI. 257 /// 258 /// For subtargets which support it, \p EnableWavefrontSize32 should match 259 /// the ENABLE_WAVEFRONT_SIZE32 kernel descriptor field. 260 unsigned getVGPREncodingGranule(const MCSubtargetInfo *STI, 261 Optional<bool> EnableWavefrontSize32 = None); 262 263 /// \returns Total number of VGPRs for given subtarget \p STI. 264 unsigned getTotalNumVGPRs(const MCSubtargetInfo *STI); 265 266 /// \returns Addressable number of VGPRs for given subtarget \p STI. 267 unsigned getAddressableNumVGPRs(const MCSubtargetInfo *STI); 268 269 /// \returns Minimum number of VGPRs that meets given number of waves per 270 /// execution unit requirement for given subtarget \p STI. 271 unsigned getMinNumVGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU); 272 273 /// \returns Maximum number of VGPRs that meets given number of waves per 274 /// execution unit requirement for given subtarget \p STI. 275 unsigned getMaxNumVGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU); 276 277 /// \returns Number of VGPR blocks needed for given subtarget \p STI when 278 /// \p NumVGPRs are used. 279 /// 280 /// For subtargets which support it, \p EnableWavefrontSize32 should match the 281 /// ENABLE_WAVEFRONT_SIZE32 kernel descriptor field. 282 unsigned getNumVGPRBlocks(const MCSubtargetInfo *STI, unsigned NumSGPRs, 283 Optional<bool> EnableWavefrontSize32 = None); 284 285 } // end namespace IsaInfo 286 287 LLVM_READONLY 288 int16_t getNamedOperandIdx(uint16_t Opcode, uint16_t NamedIdx); 289 290 LLVM_READONLY 291 int getSOPPWithRelaxation(uint16_t Opcode); 292 293 struct MIMGBaseOpcodeInfo { 294 MIMGBaseOpcode BaseOpcode; 295 bool Store; 296 bool Atomic; 297 bool AtomicX2; 298 bool Sampler; 299 bool Gather4; 300 301 uint8_t NumExtraArgs; 302 bool Gradients; 303 bool G16; 304 bool Coordinates; 305 bool LodOrClampOrMip; 306 bool HasD16; 307 bool MSAA; 308 bool BVH; 309 }; 310 311 LLVM_READONLY 312 const MIMGBaseOpcodeInfo *getMIMGBaseOpcode(unsigned Opc); 313 314 LLVM_READONLY 315 const MIMGBaseOpcodeInfo *getMIMGBaseOpcodeInfo(unsigned BaseOpcode); 316 317 struct MIMGDimInfo { 318 MIMGDim Dim; 319 uint8_t NumCoords; 320 uint8_t NumGradients; 321 bool MSAA; 322 bool DA; 323 uint8_t Encoding; 324 const char *AsmSuffix; 325 }; 326 327 LLVM_READONLY 328 const MIMGDimInfo *getMIMGDimInfo(unsigned DimEnum); 329 330 LLVM_READONLY 331 const MIMGDimInfo *getMIMGDimInfoByEncoding(uint8_t DimEnc); 332 333 LLVM_READONLY 334 const MIMGDimInfo *getMIMGDimInfoByAsmSuffix(StringRef AsmSuffix); 335 336 struct MIMGLZMappingInfo { 337 MIMGBaseOpcode L; 338 MIMGBaseOpcode LZ; 339 }; 340 341 struct MIMGMIPMappingInfo { 342 MIMGBaseOpcode MIP; 343 MIMGBaseOpcode NONMIP; 344 }; 345 346 struct MIMGBiasMappingInfo { 347 MIMGBaseOpcode Bias; 348 MIMGBaseOpcode NoBias; 349 }; 350 351 struct MIMGOffsetMappingInfo { 352 MIMGBaseOpcode Offset; 353 MIMGBaseOpcode NoOffset; 354 }; 355 356 struct MIMGG16MappingInfo { 357 MIMGBaseOpcode G; 358 MIMGBaseOpcode G16; 359 }; 360 361 LLVM_READONLY 362 const MIMGLZMappingInfo *getMIMGLZMappingInfo(unsigned L); 363 364 LLVM_READONLY 365 const MIMGMIPMappingInfo *getMIMGMIPMappingInfo(unsigned MIP); 366 367 LLVM_READONLY 368 const MIMGBiasMappingInfo *getMIMGBiasMappingInfo(unsigned Bias); 369 370 LLVM_READONLY 371 const MIMGOffsetMappingInfo *getMIMGOffsetMappingInfo(unsigned Offset); 372 373 LLVM_READONLY 374 const MIMGG16MappingInfo *getMIMGG16MappingInfo(unsigned G); 375 376 LLVM_READONLY 377 int getMIMGOpcode(unsigned BaseOpcode, unsigned MIMGEncoding, 378 unsigned VDataDwords, unsigned VAddrDwords); 379 380 LLVM_READONLY 381 int getMaskedMIMGOp(unsigned Opc, unsigned NewChannels); 382 383 LLVM_READONLY 384 unsigned getAddrSizeMIMGOp(const MIMGBaseOpcodeInfo *BaseOpcode, 385 const MIMGDimInfo *Dim, bool IsA16, 386 bool IsG16Supported); 387 388 struct MIMGInfo { 389 uint16_t Opcode; 390 uint16_t BaseOpcode; 391 uint8_t MIMGEncoding; 392 uint8_t VDataDwords; 393 uint8_t VAddrDwords; 394 }; 395 396 LLVM_READONLY 397 const MIMGInfo *getMIMGInfo(unsigned Opc); 398 399 LLVM_READONLY 400 int getMTBUFBaseOpcode(unsigned Opc); 401 402 LLVM_READONLY 403 int getMTBUFOpcode(unsigned BaseOpc, unsigned Elements); 404 405 LLVM_READONLY 406 int getMTBUFElements(unsigned Opc); 407 408 LLVM_READONLY 409 bool getMTBUFHasVAddr(unsigned Opc); 410 411 LLVM_READONLY 412 bool getMTBUFHasSrsrc(unsigned Opc); 413 414 LLVM_READONLY 415 bool getMTBUFHasSoffset(unsigned Opc); 416 417 LLVM_READONLY 418 int getMUBUFBaseOpcode(unsigned Opc); 419 420 LLVM_READONLY 421 int getMUBUFOpcode(unsigned BaseOpc, unsigned Elements); 422 423 LLVM_READONLY 424 int getMUBUFElements(unsigned Opc); 425 426 LLVM_READONLY 427 bool getMUBUFHasVAddr(unsigned Opc); 428 429 LLVM_READONLY 430 bool getMUBUFHasSrsrc(unsigned Opc); 431 432 LLVM_READONLY 433 bool getMUBUFHasSoffset(unsigned Opc); 434 435 LLVM_READONLY 436 bool getMUBUFIsBufferInv(unsigned Opc); 437 438 LLVM_READONLY 439 bool getSMEMIsBuffer(unsigned Opc); 440 441 LLVM_READONLY 442 bool getVOP1IsSingle(unsigned Opc); 443 444 LLVM_READONLY 445 bool getVOP2IsSingle(unsigned Opc); 446 447 LLVM_READONLY 448 bool getVOP3IsSingle(unsigned Opc); 449 450 LLVM_READONLY 451 const GcnBufferFormatInfo *getGcnBufferFormatInfo(uint8_t BitsPerComp, 452 uint8_t NumComponents, 453 uint8_t NumFormat, 454 const MCSubtargetInfo &STI); 455 LLVM_READONLY 456 const GcnBufferFormatInfo *getGcnBufferFormatInfo(uint8_t Format, 457 const MCSubtargetInfo &STI); 458 459 LLVM_READONLY 460 int getMCOpcode(uint16_t Opcode, unsigned Gen); 461 462 void initDefaultAMDKernelCodeT(amd_kernel_code_t &Header, 463 const MCSubtargetInfo *STI); 464 465 amdhsa::kernel_descriptor_t getDefaultAmdhsaKernelDescriptor( 466 const MCSubtargetInfo *STI); 467 468 bool isGroupSegment(const GlobalValue *GV); 469 bool isGlobalSegment(const GlobalValue *GV); 470 bool isReadOnlySegment(const GlobalValue *GV); 471 472 /// \returns True if constants should be emitted to .text section for given 473 /// target triple \p TT, false otherwise. 474 bool shouldEmitConstantsToTextSection(const Triple &TT); 475 476 /// \returns Integer value requested using \p F's \p Name attribute. 477 /// 478 /// \returns \p Default if attribute is not present. 479 /// 480 /// \returns \p Default and emits error if requested value cannot be converted 481 /// to integer. 482 int getIntegerAttribute(const Function &F, StringRef Name, int Default); 483 484 /// \returns A pair of integer values requested using \p F's \p Name attribute 485 /// in "first[,second]" format ("second" is optional unless \p OnlyFirstRequired 486 /// is false). 487 /// 488 /// \returns \p Default if attribute is not present. 489 /// 490 /// \returns \p Default and emits error if one of the requested values cannot be 491 /// converted to integer, or \p OnlyFirstRequired is false and "second" value is 492 /// not present. 493 std::pair<int, int> getIntegerPairAttribute(const Function &F, 494 StringRef Name, 495 std::pair<int, int> Default, 496 bool OnlyFirstRequired = false); 497 498 /// Represents the counter values to wait for in an s_waitcnt instruction. 499 /// 500 /// Large values (including the maximum possible integer) can be used to 501 /// represent "don't care" waits. 502 struct Waitcnt { 503 unsigned VmCnt = ~0u; 504 unsigned ExpCnt = ~0u; 505 unsigned LgkmCnt = ~0u; 506 unsigned VsCnt = ~0u; 507 508 Waitcnt() = default; 509 Waitcnt(unsigned VmCnt, unsigned ExpCnt, unsigned LgkmCnt, unsigned VsCnt) 510 : VmCnt(VmCnt), ExpCnt(ExpCnt), LgkmCnt(LgkmCnt), VsCnt(VsCnt) {} 511 512 static Waitcnt allZero(bool HasVscnt) { 513 return Waitcnt(0, 0, 0, HasVscnt ? 0 : ~0u); 514 } 515 static Waitcnt allZeroExceptVsCnt() { return Waitcnt(0, 0, 0, ~0u); } 516 517 bool hasWait() const { 518 return VmCnt != ~0u || ExpCnt != ~0u || LgkmCnt != ~0u || VsCnt != ~0u; 519 } 520 521 bool hasWaitExceptVsCnt() const { 522 return VmCnt != ~0u || ExpCnt != ~0u || LgkmCnt != ~0u; 523 } 524 525 bool hasWaitVsCnt() const { 526 return VsCnt != ~0u; 527 } 528 529 bool dominates(const Waitcnt &Other) const { 530 return VmCnt <= Other.VmCnt && ExpCnt <= Other.ExpCnt && 531 LgkmCnt <= Other.LgkmCnt && VsCnt <= Other.VsCnt; 532 } 533 534 Waitcnt combined(const Waitcnt &Other) const { 535 return Waitcnt(std::min(VmCnt, Other.VmCnt), std::min(ExpCnt, Other.ExpCnt), 536 std::min(LgkmCnt, Other.LgkmCnt), 537 std::min(VsCnt, Other.VsCnt)); 538 } 539 }; 540 541 /// \returns Vmcnt bit mask for given isa \p Version. 542 unsigned getVmcntBitMask(const IsaVersion &Version); 543 544 /// \returns Expcnt bit mask for given isa \p Version. 545 unsigned getExpcntBitMask(const IsaVersion &Version); 546 547 /// \returns Lgkmcnt bit mask for given isa \p Version. 548 unsigned getLgkmcntBitMask(const IsaVersion &Version); 549 550 /// \returns Waitcnt bit mask for given isa \p Version. 551 unsigned getWaitcntBitMask(const IsaVersion &Version); 552 553 /// \returns Decoded Vmcnt from given \p Waitcnt for given isa \p Version. 554 unsigned decodeVmcnt(const IsaVersion &Version, unsigned Waitcnt); 555 556 /// \returns Decoded Expcnt from given \p Waitcnt for given isa \p Version. 557 unsigned decodeExpcnt(const IsaVersion &Version, unsigned Waitcnt); 558 559 /// \returns Decoded Lgkmcnt from given \p Waitcnt for given isa \p Version. 560 unsigned decodeLgkmcnt(const IsaVersion &Version, unsigned Waitcnt); 561 562 /// Decodes Vmcnt, Expcnt and Lgkmcnt from given \p Waitcnt for given isa 563 /// \p Version, and writes decoded values into \p Vmcnt, \p Expcnt and 564 /// \p Lgkmcnt respectively. 565 /// 566 /// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are decoded as follows: 567 /// \p Vmcnt = \p Waitcnt[3:0] (pre-gfx9 only) 568 /// \p Vmcnt = \p Waitcnt[3:0] | \p Waitcnt[15:14] (gfx9+ only) 569 /// \p Expcnt = \p Waitcnt[6:4] 570 /// \p Lgkmcnt = \p Waitcnt[11:8] (pre-gfx10 only) 571 /// \p Lgkmcnt = \p Waitcnt[13:8] (gfx10+ only) 572 void decodeWaitcnt(const IsaVersion &Version, unsigned Waitcnt, 573 unsigned &Vmcnt, unsigned &Expcnt, unsigned &Lgkmcnt); 574 575 Waitcnt decodeWaitcnt(const IsaVersion &Version, unsigned Encoded); 576 577 /// \returns \p Waitcnt with encoded \p Vmcnt for given isa \p Version. 578 unsigned encodeVmcnt(const IsaVersion &Version, unsigned Waitcnt, 579 unsigned Vmcnt); 580 581 /// \returns \p Waitcnt with encoded \p Expcnt for given isa \p Version. 582 unsigned encodeExpcnt(const IsaVersion &Version, unsigned Waitcnt, 583 unsigned Expcnt); 584 585 /// \returns \p Waitcnt with encoded \p Lgkmcnt for given isa \p Version. 586 unsigned encodeLgkmcnt(const IsaVersion &Version, unsigned Waitcnt, 587 unsigned Lgkmcnt); 588 589 /// Encodes \p Vmcnt, \p Expcnt and \p Lgkmcnt into Waitcnt for given isa 590 /// \p Version. 591 /// 592 /// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are encoded as follows: 593 /// Waitcnt[3:0] = \p Vmcnt (pre-gfx9 only) 594 /// Waitcnt[3:0] = \p Vmcnt[3:0] (gfx9+ only) 595 /// Waitcnt[6:4] = \p Expcnt 596 /// Waitcnt[11:8] = \p Lgkmcnt (pre-gfx10 only) 597 /// Waitcnt[13:8] = \p Lgkmcnt (gfx10+ only) 598 /// Waitcnt[15:14] = \p Vmcnt[5:4] (gfx9+ only) 599 /// 600 /// \returns Waitcnt with encoded \p Vmcnt, \p Expcnt and \p Lgkmcnt for given 601 /// isa \p Version. 602 unsigned encodeWaitcnt(const IsaVersion &Version, 603 unsigned Vmcnt, unsigned Expcnt, unsigned Lgkmcnt); 604 605 unsigned encodeWaitcnt(const IsaVersion &Version, const Waitcnt &Decoded); 606 607 namespace Hwreg { 608 609 LLVM_READONLY 610 int64_t getHwregId(const StringRef Name, const MCSubtargetInfo &STI); 611 612 LLVM_READNONE 613 bool isValidHwreg(int64_t Id, const MCSubtargetInfo &STI); 614 615 LLVM_READNONE 616 bool isValidHwreg(int64_t Id); 617 618 LLVM_READNONE 619 bool isValidHwregOffset(int64_t Offset); 620 621 LLVM_READNONE 622 bool isValidHwregWidth(int64_t Width); 623 624 LLVM_READNONE 625 uint64_t encodeHwreg(uint64_t Id, uint64_t Offset, uint64_t Width); 626 627 LLVM_READNONE 628 StringRef getHwreg(unsigned Id, const MCSubtargetInfo &STI); 629 630 void decodeHwreg(unsigned Val, unsigned &Id, unsigned &Offset, unsigned &Width); 631 632 } // namespace Hwreg 633 634 namespace Exp { 635 636 bool getTgtName(unsigned Id, StringRef &Name, int &Index); 637 638 LLVM_READONLY 639 unsigned getTgtId(const StringRef Name); 640 641 LLVM_READNONE 642 bool isSupportedTgtId(unsigned Id, const MCSubtargetInfo &STI); 643 644 } // namespace Exp 645 646 namespace MTBUFFormat { 647 648 LLVM_READNONE 649 int64_t encodeDfmtNfmt(unsigned Dfmt, unsigned Nfmt); 650 651 void decodeDfmtNfmt(unsigned Format, unsigned &Dfmt, unsigned &Nfmt); 652 653 int64_t getDfmt(const StringRef Name); 654 655 StringRef getDfmtName(unsigned Id); 656 657 int64_t getNfmt(const StringRef Name, const MCSubtargetInfo &STI); 658 659 StringRef getNfmtName(unsigned Id, const MCSubtargetInfo &STI); 660 661 bool isValidDfmtNfmt(unsigned Val, const MCSubtargetInfo &STI); 662 663 bool isValidNfmt(unsigned Val, const MCSubtargetInfo &STI); 664 665 int64_t getUnifiedFormat(const StringRef Name); 666 667 StringRef getUnifiedFormatName(unsigned Id); 668 669 bool isValidUnifiedFormat(unsigned Val); 670 671 int64_t convertDfmtNfmt2Ufmt(unsigned Dfmt, unsigned Nfmt); 672 673 bool isValidFormatEncoding(unsigned Val, const MCSubtargetInfo &STI); 674 675 unsigned getDefaultFormatEncoding(const MCSubtargetInfo &STI); 676 677 } // namespace MTBUFFormat 678 679 namespace SendMsg { 680 681 LLVM_READONLY 682 int64_t getMsgId(const StringRef Name); 683 684 LLVM_READONLY 685 int64_t getMsgOpId(int64_t MsgId, const StringRef Name); 686 687 LLVM_READNONE 688 StringRef getMsgName(int64_t MsgId); 689 690 LLVM_READNONE 691 StringRef getMsgOpName(int64_t MsgId, int64_t OpId); 692 693 LLVM_READNONE 694 bool isValidMsgId(int64_t MsgId, const MCSubtargetInfo &STI, bool Strict = true); 695 696 LLVM_READNONE 697 bool isValidMsgOp(int64_t MsgId, int64_t OpId, const MCSubtargetInfo &STI, 698 bool Strict = true); 699 700 LLVM_READNONE 701 bool isValidMsgStream(int64_t MsgId, int64_t OpId, int64_t StreamId, 702 const MCSubtargetInfo &STI, bool Strict = true); 703 704 LLVM_READNONE 705 bool msgRequiresOp(int64_t MsgId); 706 707 LLVM_READNONE 708 bool msgSupportsStream(int64_t MsgId, int64_t OpId); 709 710 void decodeMsg(unsigned Val, 711 uint16_t &MsgId, 712 uint16_t &OpId, 713 uint16_t &StreamId); 714 715 LLVM_READNONE 716 uint64_t encodeMsg(uint64_t MsgId, 717 uint64_t OpId, 718 uint64_t StreamId); 719 720 } // namespace SendMsg 721 722 723 unsigned getInitialPSInputAddr(const Function &F); 724 725 bool getHasColorExport(const Function &F); 726 727 bool getHasDepthExport(const Function &F); 728 729 LLVM_READNONE 730 bool isShader(CallingConv::ID CC); 731 732 LLVM_READNONE 733 bool isGraphics(CallingConv::ID CC); 734 735 LLVM_READNONE 736 bool isCompute(CallingConv::ID CC); 737 738 LLVM_READNONE 739 bool isEntryFunctionCC(CallingConv::ID CC); 740 741 // These functions are considered entrypoints into the current module, i.e. they 742 // are allowed to be called from outside the current module. This is different 743 // from isEntryFunctionCC, which is only true for functions that are entered by 744 // the hardware. Module entry points include all entry functions but also 745 // include functions that can be called from other functions inside or outside 746 // the current module. Module entry functions are allowed to allocate LDS. 747 LLVM_READNONE 748 bool isModuleEntryFunctionCC(CallingConv::ID CC); 749 750 bool isKernelCC(const Function *Func); 751 752 // FIXME: Remove this when calling conventions cleaned up 753 LLVM_READNONE 754 inline bool isKernel(CallingConv::ID CC) { 755 switch (CC) { 756 case CallingConv::AMDGPU_KERNEL: 757 case CallingConv::SPIR_KERNEL: 758 return true; 759 default: 760 return false; 761 } 762 } 763 764 bool hasXNACK(const MCSubtargetInfo &STI); 765 bool hasSRAMECC(const MCSubtargetInfo &STI); 766 bool hasMIMG_R128(const MCSubtargetInfo &STI); 767 bool hasGFX10A16(const MCSubtargetInfo &STI); 768 bool hasG16(const MCSubtargetInfo &STI); 769 bool hasPackedD16(const MCSubtargetInfo &STI); 770 771 bool isSI(const MCSubtargetInfo &STI); 772 bool isCI(const MCSubtargetInfo &STI); 773 bool isVI(const MCSubtargetInfo &STI); 774 bool isGFX9(const MCSubtargetInfo &STI); 775 bool isGFX9Plus(const MCSubtargetInfo &STI); 776 bool isGFX10(const MCSubtargetInfo &STI); 777 bool isGFX10Plus(const MCSubtargetInfo &STI); 778 bool isGCN3Encoding(const MCSubtargetInfo &STI); 779 bool isGFX10_AEncoding(const MCSubtargetInfo &STI); 780 bool isGFX10_BEncoding(const MCSubtargetInfo &STI); 781 bool hasGFX10_3Insts(const MCSubtargetInfo &STI); 782 bool isGFX90A(const MCSubtargetInfo &STI); 783 bool isGFX940(const MCSubtargetInfo &STI); 784 bool hasArchitectedFlatScratch(const MCSubtargetInfo &STI); 785 bool hasMAIInsts(const MCSubtargetInfo &STI); 786 int getTotalNumVGPRs(bool has90AInsts, int32_t ArgNumAGPR, int32_t ArgNumVGPR); 787 788 /// Is Reg - scalar register 789 bool isSGPR(unsigned Reg, const MCRegisterInfo* TRI); 790 791 /// If \p Reg is a pseudo reg, return the correct hardware register given 792 /// \p STI otherwise return \p Reg. 793 unsigned getMCReg(unsigned Reg, const MCSubtargetInfo &STI); 794 795 /// Convert hardware register \p Reg to a pseudo register 796 LLVM_READNONE 797 unsigned mc2PseudoReg(unsigned Reg); 798 799 /// Can this operand also contain immediate values? 800 bool isSISrcOperand(const MCInstrDesc &Desc, unsigned OpNo); 801 802 /// Is this floating-point operand? 803 bool isSISrcFPOperand(const MCInstrDesc &Desc, unsigned OpNo); 804 805 /// Does this operand support only inlinable literals? 806 bool isSISrcInlinableOperand(const MCInstrDesc &Desc, unsigned OpNo); 807 808 /// Get the size in bits of a register from the register class \p RC. 809 unsigned getRegBitWidth(unsigned RCID); 810 811 /// Get the size in bits of a register from the register class \p RC. 812 unsigned getRegBitWidth(const MCRegisterClass &RC); 813 814 /// Get size of register operand 815 unsigned getRegOperandSize(const MCRegisterInfo *MRI, const MCInstrDesc &Desc, 816 unsigned OpNo); 817 818 LLVM_READNONE 819 inline unsigned getOperandSize(const MCOperandInfo &OpInfo) { 820 switch (OpInfo.OperandType) { 821 case AMDGPU::OPERAND_REG_IMM_INT32: 822 case AMDGPU::OPERAND_REG_IMM_FP32: 823 case AMDGPU::OPERAND_REG_IMM_FP32_DEFERRED: 824 case AMDGPU::OPERAND_REG_INLINE_C_INT32: 825 case AMDGPU::OPERAND_REG_INLINE_C_FP32: 826 case AMDGPU::OPERAND_REG_INLINE_AC_INT32: 827 case AMDGPU::OPERAND_REG_INLINE_AC_FP32: 828 case AMDGPU::OPERAND_REG_IMM_V2INT32: 829 case AMDGPU::OPERAND_REG_IMM_V2FP32: 830 case AMDGPU::OPERAND_REG_INLINE_C_V2INT32: 831 case AMDGPU::OPERAND_REG_INLINE_C_V2FP32: 832 case AMDGPU::OPERAND_KIMM32: 833 case AMDGPU::OPERAND_KIMM16: // mandatory literal is always size 4 834 return 4; 835 836 case AMDGPU::OPERAND_REG_IMM_INT64: 837 case AMDGPU::OPERAND_REG_IMM_FP64: 838 case AMDGPU::OPERAND_REG_INLINE_C_INT64: 839 case AMDGPU::OPERAND_REG_INLINE_C_FP64: 840 case AMDGPU::OPERAND_REG_INLINE_AC_FP64: 841 return 8; 842 843 case AMDGPU::OPERAND_REG_IMM_INT16: 844 case AMDGPU::OPERAND_REG_IMM_FP16: 845 case AMDGPU::OPERAND_REG_IMM_FP16_DEFERRED: 846 case AMDGPU::OPERAND_REG_INLINE_C_INT16: 847 case AMDGPU::OPERAND_REG_INLINE_C_FP16: 848 case AMDGPU::OPERAND_REG_INLINE_C_V2INT16: 849 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16: 850 case AMDGPU::OPERAND_REG_INLINE_AC_INT16: 851 case AMDGPU::OPERAND_REG_INLINE_AC_FP16: 852 case AMDGPU::OPERAND_REG_INLINE_AC_V2INT16: 853 case AMDGPU::OPERAND_REG_INLINE_AC_V2FP16: 854 case AMDGPU::OPERAND_REG_IMM_V2INT16: 855 case AMDGPU::OPERAND_REG_IMM_V2FP16: 856 return 2; 857 858 default: 859 llvm_unreachable("unhandled operand type"); 860 } 861 } 862 863 LLVM_READNONE 864 inline unsigned getOperandSize(const MCInstrDesc &Desc, unsigned OpNo) { 865 return getOperandSize(Desc.OpInfo[OpNo]); 866 } 867 868 /// Is this literal inlinable, and not one of the values intended for floating 869 /// point values. 870 LLVM_READNONE 871 inline bool isInlinableIntLiteral(int64_t Literal) { 872 return Literal >= -16 && Literal <= 64; 873 } 874 875 /// Is this literal inlinable 876 LLVM_READNONE 877 bool isInlinableLiteral64(int64_t Literal, bool HasInv2Pi); 878 879 LLVM_READNONE 880 bool isInlinableLiteral32(int32_t Literal, bool HasInv2Pi); 881 882 LLVM_READNONE 883 bool isInlinableLiteral16(int16_t Literal, bool HasInv2Pi); 884 885 LLVM_READNONE 886 bool isInlinableLiteralV216(int32_t Literal, bool HasInv2Pi); 887 888 LLVM_READNONE 889 bool isInlinableIntLiteralV216(int32_t Literal); 890 891 LLVM_READNONE 892 bool isFoldableLiteralV216(int32_t Literal, bool HasInv2Pi); 893 894 bool isArgPassedInSGPR(const Argument *Arg); 895 896 LLVM_READONLY 897 bool isLegalSMRDEncodedUnsignedOffset(const MCSubtargetInfo &ST, 898 int64_t EncodedOffset); 899 900 LLVM_READONLY 901 bool isLegalSMRDEncodedSignedOffset(const MCSubtargetInfo &ST, 902 int64_t EncodedOffset, 903 bool IsBuffer); 904 905 /// Convert \p ByteOffset to dwords if the subtarget uses dword SMRD immediate 906 /// offsets. 907 uint64_t convertSMRDOffsetUnits(const MCSubtargetInfo &ST, uint64_t ByteOffset); 908 909 /// \returns The encoding that will be used for \p ByteOffset in the 910 /// SMRD offset field, or None if it won't fit. On GFX9 and GFX10 911 /// S_LOAD instructions have a signed offset, on other subtargets it is 912 /// unsigned. S_BUFFER has an unsigned offset for all subtargets. 913 Optional<int64_t> getSMRDEncodedOffset(const MCSubtargetInfo &ST, 914 int64_t ByteOffset, bool IsBuffer); 915 916 /// \return The encoding that can be used for a 32-bit literal offset in an SMRD 917 /// instruction. This is only useful on CI.s 918 Optional<int64_t> getSMRDEncodedLiteralOffset32(const MCSubtargetInfo &ST, 919 int64_t ByteOffset); 920 921 /// For FLAT segment the offset must be positive; 922 /// MSB is ignored and forced to zero. 923 /// 924 /// \return The number of bits available for the offset field in flat 925 /// instructions. 926 unsigned getNumFlatOffsetBits(const MCSubtargetInfo &ST, bool Signed); 927 928 /// \returns true if this offset is small enough to fit in the SMRD 929 /// offset field. \p ByteOffset should be the offset in bytes and 930 /// not the encoded offset. 931 bool isLegalSMRDImmOffset(const MCSubtargetInfo &ST, int64_t ByteOffset); 932 933 bool splitMUBUFOffset(uint32_t Imm, uint32_t &SOffset, uint32_t &ImmOffset, 934 const GCNSubtarget *Subtarget, 935 Align Alignment = Align(4)); 936 937 LLVM_READNONE 938 inline bool isLegal64BitDPPControl(unsigned DC) { 939 return DC >= DPP::ROW_NEWBCAST_FIRST && DC <= DPP::ROW_NEWBCAST_LAST; 940 } 941 942 /// \returns true if the intrinsic is divergent 943 bool isIntrinsicSourceOfDivergence(unsigned IntrID); 944 945 // Track defaults for fields in the MODE register. 946 struct SIModeRegisterDefaults { 947 /// Floating point opcodes that support exception flag gathering quiet and 948 /// propagate signaling NaN inputs per IEEE 754-2008. Min_dx10 and max_dx10 949 /// become IEEE 754- 2008 compliant due to signaling NaN propagation and 950 /// quieting. 951 bool IEEE : 1; 952 953 /// Used by the vector ALU to force DX10-style treatment of NaNs: when set, 954 /// clamp NaN to zero; otherwise, pass NaN through. 955 bool DX10Clamp : 1; 956 957 /// If this is set, neither input or output denormals are flushed for most f32 958 /// instructions. 959 bool FP32InputDenormals : 1; 960 bool FP32OutputDenormals : 1; 961 962 /// If this is set, neither input or output denormals are flushed for both f64 963 /// and f16/v2f16 instructions. 964 bool FP64FP16InputDenormals : 1; 965 bool FP64FP16OutputDenormals : 1; 966 967 SIModeRegisterDefaults() : 968 IEEE(true), 969 DX10Clamp(true), 970 FP32InputDenormals(true), 971 FP32OutputDenormals(true), 972 FP64FP16InputDenormals(true), 973 FP64FP16OutputDenormals(true) {} 974 975 SIModeRegisterDefaults(const Function &F); 976 977 static SIModeRegisterDefaults getDefaultForCallingConv(CallingConv::ID CC) { 978 SIModeRegisterDefaults Mode; 979 Mode.IEEE = !AMDGPU::isShader(CC); 980 return Mode; 981 } 982 983 bool operator ==(const SIModeRegisterDefaults Other) const { 984 return IEEE == Other.IEEE && DX10Clamp == Other.DX10Clamp && 985 FP32InputDenormals == Other.FP32InputDenormals && 986 FP32OutputDenormals == Other.FP32OutputDenormals && 987 FP64FP16InputDenormals == Other.FP64FP16InputDenormals && 988 FP64FP16OutputDenormals == Other.FP64FP16OutputDenormals; 989 } 990 991 bool allFP32Denormals() const { 992 return FP32InputDenormals && FP32OutputDenormals; 993 } 994 995 bool allFP64FP16Denormals() const { 996 return FP64FP16InputDenormals && FP64FP16OutputDenormals; 997 } 998 999 /// Get the encoding value for the FP_DENORM bits of the mode register for the 1000 /// FP32 denormal mode. 1001 uint32_t fpDenormModeSPValue() const { 1002 if (FP32InputDenormals && FP32OutputDenormals) 1003 return FP_DENORM_FLUSH_NONE; 1004 if (FP32InputDenormals) 1005 return FP_DENORM_FLUSH_OUT; 1006 if (FP32OutputDenormals) 1007 return FP_DENORM_FLUSH_IN; 1008 return FP_DENORM_FLUSH_IN_FLUSH_OUT; 1009 } 1010 1011 /// Get the encoding value for the FP_DENORM bits of the mode register for the 1012 /// FP64/FP16 denormal mode. 1013 uint32_t fpDenormModeDPValue() const { 1014 if (FP64FP16InputDenormals && FP64FP16OutputDenormals) 1015 return FP_DENORM_FLUSH_NONE; 1016 if (FP64FP16InputDenormals) 1017 return FP_DENORM_FLUSH_OUT; 1018 if (FP64FP16OutputDenormals) 1019 return FP_DENORM_FLUSH_IN; 1020 return FP_DENORM_FLUSH_IN_FLUSH_OUT; 1021 } 1022 1023 /// Returns true if a flag is compatible if it's enabled in the callee, but 1024 /// disabled in the caller. 1025 static bool oneWayCompatible(bool CallerMode, bool CalleeMode) { 1026 return CallerMode == CalleeMode || (!CallerMode && CalleeMode); 1027 } 1028 1029 // FIXME: Inlining should be OK for dx10-clamp, since the caller's mode should 1030 // be able to override. 1031 bool isInlineCompatible(SIModeRegisterDefaults CalleeMode) const { 1032 if (DX10Clamp != CalleeMode.DX10Clamp) 1033 return false; 1034 if (IEEE != CalleeMode.IEEE) 1035 return false; 1036 1037 // Allow inlining denormals enabled into denormals flushed functions. 1038 return oneWayCompatible(FP64FP16InputDenormals, CalleeMode.FP64FP16InputDenormals) && 1039 oneWayCompatible(FP64FP16OutputDenormals, CalleeMode.FP64FP16OutputDenormals) && 1040 oneWayCompatible(FP32InputDenormals, CalleeMode.FP32InputDenormals) && 1041 oneWayCompatible(FP32OutputDenormals, CalleeMode.FP32OutputDenormals); 1042 } 1043 }; 1044 1045 } // end namespace AMDGPU 1046 1047 raw_ostream &operator<<(raw_ostream &OS, 1048 const AMDGPU::IsaInfo::TargetIDSetting S); 1049 1050 } // end namespace llvm 1051 1052 #endif // LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUBASEINFO_H 1053