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 "AMDGPU.h" 13 #include "AMDKernelCodeT.h" 14 #include "SIDefines.h" 15 #include "llvm/IR/CallingConv.h" 16 #include "llvm/MC/MCInstrDesc.h" 17 #include "llvm/Support/AMDHSAKernelDescriptor.h" 18 #include "llvm/Support/Compiler.h" 19 #include "llvm/Support/ErrorHandling.h" 20 #include "llvm/Support/TargetParser.h" 21 #include <cstdint> 22 #include <string> 23 #include <utility> 24 25 namespace llvm { 26 27 class Argument; 28 class Function; 29 class GCNSubtarget; 30 class GlobalValue; 31 class MCRegisterClass; 32 class MCRegisterInfo; 33 class MCSubtargetInfo; 34 class StringRef; 35 class Triple; 36 37 namespace AMDGPU { 38 39 struct GcnBufferFormatInfo { 40 unsigned Format; 41 unsigned BitsPerComp; 42 unsigned NumComponents; 43 unsigned NumFormat; 44 unsigned DataFormat; 45 }; 46 47 #define GET_MIMGBaseOpcode_DECL 48 #define GET_MIMGDim_DECL 49 #define GET_MIMGEncoding_DECL 50 #define GET_MIMGLZMapping_DECL 51 #define GET_MIMGMIPMapping_DECL 52 #include "AMDGPUGenSearchableTables.inc" 53 54 namespace IsaInfo { 55 56 enum { 57 // The closed Vulkan driver sets 96, which limits the wave count to 8 but 58 // doesn't spill SGPRs as much as when 80 is set. 59 FIXED_NUM_SGPRS_FOR_INIT_BUG = 96, 60 TRAP_NUM_SGPRS = 16 61 }; 62 63 /// Streams isa version string for given subtarget \p STI into \p Stream. 64 void streamIsaVersion(const MCSubtargetInfo *STI, raw_ostream &Stream); 65 66 /// \returns True if given subtarget \p STI supports code object version 3, 67 /// false otherwise. 68 bool hasCodeObjectV3(const MCSubtargetInfo *STI); 69 70 /// \returns Wavefront size for given subtarget \p STI. 71 unsigned getWavefrontSize(const MCSubtargetInfo *STI); 72 73 /// \returns Local memory size in bytes for given subtarget \p STI. 74 unsigned getLocalMemorySize(const MCSubtargetInfo *STI); 75 76 /// \returns Number of execution units per compute unit for given subtarget \p 77 /// STI. 78 unsigned getEUsPerCU(const MCSubtargetInfo *STI); 79 80 /// \returns Maximum number of work groups per compute unit for given subtarget 81 /// \p STI and limited by given \p FlatWorkGroupSize. 82 unsigned getMaxWorkGroupsPerCU(const MCSubtargetInfo *STI, 83 unsigned FlatWorkGroupSize); 84 85 /// \returns Minimum number of waves per execution unit for given subtarget \p 86 /// STI. 87 unsigned getMinWavesPerEU(const MCSubtargetInfo *STI); 88 89 /// \returns Maximum number of waves per execution unit for given subtarget \p 90 /// STI without any kind of limitation. 91 unsigned getMaxWavesPerEU(const MCSubtargetInfo *STI); 92 93 /// \returns Number of waves per execution unit required to support the given \p 94 /// FlatWorkGroupSize. 95 unsigned getWavesPerEUForWorkGroup(const MCSubtargetInfo *STI, 96 unsigned FlatWorkGroupSize); 97 98 /// \returns Minimum flat work group size for given subtarget \p STI. 99 unsigned getMinFlatWorkGroupSize(const MCSubtargetInfo *STI); 100 101 /// \returns Maximum flat work group size for given subtarget \p STI. 102 unsigned getMaxFlatWorkGroupSize(const MCSubtargetInfo *STI); 103 104 /// \returns Number of waves per work group for given subtarget \p STI and 105 /// \p FlatWorkGroupSize. 106 unsigned getWavesPerWorkGroup(const MCSubtargetInfo *STI, 107 unsigned FlatWorkGroupSize); 108 109 /// \returns SGPR allocation granularity for given subtarget \p STI. 110 unsigned getSGPRAllocGranule(const MCSubtargetInfo *STI); 111 112 /// \returns SGPR encoding granularity for given subtarget \p STI. 113 unsigned getSGPREncodingGranule(const MCSubtargetInfo *STI); 114 115 /// \returns Total number of SGPRs for given subtarget \p STI. 116 unsigned getTotalNumSGPRs(const MCSubtargetInfo *STI); 117 118 /// \returns Addressable number of SGPRs for given subtarget \p STI. 119 unsigned getAddressableNumSGPRs(const MCSubtargetInfo *STI); 120 121 /// \returns Minimum number of SGPRs that meets the given number of waves per 122 /// execution unit requirement for given subtarget \p STI. 123 unsigned getMinNumSGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU); 124 125 /// \returns Maximum number of SGPRs that meets the given number of waves per 126 /// execution unit requirement for given subtarget \p STI. 127 unsigned getMaxNumSGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU, 128 bool Addressable); 129 130 /// \returns Number of extra SGPRs implicitly required by given subtarget \p 131 /// STI when the given special registers are used. 132 unsigned getNumExtraSGPRs(const MCSubtargetInfo *STI, bool VCCUsed, 133 bool FlatScrUsed, bool XNACKUsed); 134 135 /// \returns Number of extra SGPRs implicitly required by given subtarget \p 136 /// STI when the given special registers are used. XNACK is inferred from 137 /// \p STI. 138 unsigned getNumExtraSGPRs(const MCSubtargetInfo *STI, bool VCCUsed, 139 bool FlatScrUsed); 140 141 /// \returns Number of SGPR blocks needed for given subtarget \p STI when 142 /// \p NumSGPRs are used. \p NumSGPRs should already include any special 143 /// register counts. 144 unsigned getNumSGPRBlocks(const MCSubtargetInfo *STI, unsigned NumSGPRs); 145 146 /// \returns VGPR allocation granularity for given subtarget \p STI. 147 /// 148 /// For subtargets which support it, \p EnableWavefrontSize32 should match 149 /// the ENABLE_WAVEFRONT_SIZE32 kernel descriptor field. 150 unsigned getVGPRAllocGranule(const MCSubtargetInfo *STI, 151 Optional<bool> EnableWavefrontSize32 = None); 152 153 /// \returns VGPR encoding granularity for given subtarget \p STI. 154 /// 155 /// For subtargets which support it, \p EnableWavefrontSize32 should match 156 /// the ENABLE_WAVEFRONT_SIZE32 kernel descriptor field. 157 unsigned getVGPREncodingGranule(const MCSubtargetInfo *STI, 158 Optional<bool> EnableWavefrontSize32 = None); 159 160 /// \returns Total number of VGPRs for given subtarget \p STI. 161 unsigned getTotalNumVGPRs(const MCSubtargetInfo *STI); 162 163 /// \returns Addressable number of VGPRs for given subtarget \p STI. 164 unsigned getAddressableNumVGPRs(const MCSubtargetInfo *STI); 165 166 /// \returns Minimum number of VGPRs that meets given number of waves per 167 /// execution unit requirement for given subtarget \p STI. 168 unsigned getMinNumVGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU); 169 170 /// \returns Maximum number of VGPRs that meets given number of waves per 171 /// execution unit requirement for given subtarget \p STI. 172 unsigned getMaxNumVGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU); 173 174 /// \returns Number of VGPR blocks needed for given subtarget \p STI when 175 /// \p NumVGPRs are used. 176 /// 177 /// For subtargets which support it, \p EnableWavefrontSize32 should match the 178 /// ENABLE_WAVEFRONT_SIZE32 kernel descriptor field. 179 unsigned getNumVGPRBlocks(const MCSubtargetInfo *STI, unsigned NumSGPRs, 180 Optional<bool> EnableWavefrontSize32 = None); 181 182 } // end namespace IsaInfo 183 184 LLVM_READONLY 185 int16_t getNamedOperandIdx(uint16_t Opcode, uint16_t NamedIdx); 186 187 LLVM_READONLY 188 int getSOPPWithRelaxation(uint16_t Opcode); 189 190 struct MIMGBaseOpcodeInfo { 191 MIMGBaseOpcode BaseOpcode; 192 bool Store; 193 bool Atomic; 194 bool AtomicX2; 195 bool Sampler; 196 bool Gather4; 197 198 uint8_t NumExtraArgs; 199 bool Gradients; 200 bool Coordinates; 201 bool LodOrClampOrMip; 202 bool HasD16; 203 }; 204 205 LLVM_READONLY 206 const MIMGBaseOpcodeInfo *getMIMGBaseOpcodeInfo(unsigned BaseOpcode); 207 208 struct MIMGDimInfo { 209 MIMGDim Dim; 210 uint8_t NumCoords; 211 uint8_t NumGradients; 212 bool DA; 213 uint8_t Encoding; 214 const char *AsmSuffix; 215 }; 216 217 LLVM_READONLY 218 const MIMGDimInfo *getMIMGDimInfo(unsigned DimEnum); 219 220 LLVM_READONLY 221 const MIMGDimInfo *getMIMGDimInfoByEncoding(uint8_t DimEnc); 222 223 LLVM_READONLY 224 const MIMGDimInfo *getMIMGDimInfoByAsmSuffix(StringRef AsmSuffix); 225 226 struct MIMGLZMappingInfo { 227 MIMGBaseOpcode L; 228 MIMGBaseOpcode LZ; 229 }; 230 231 struct MIMGMIPMappingInfo { 232 MIMGBaseOpcode MIP; 233 MIMGBaseOpcode NONMIP; 234 }; 235 236 LLVM_READONLY 237 const MIMGLZMappingInfo *getMIMGLZMappingInfo(unsigned L); 238 239 LLVM_READONLY 240 const MIMGMIPMappingInfo *getMIMGMIPMappingInfo(unsigned L); 241 242 LLVM_READONLY 243 int getMIMGOpcode(unsigned BaseOpcode, unsigned MIMGEncoding, 244 unsigned VDataDwords, unsigned VAddrDwords); 245 246 LLVM_READONLY 247 int getMaskedMIMGOp(unsigned Opc, unsigned NewChannels); 248 249 struct MIMGInfo { 250 uint16_t Opcode; 251 uint16_t BaseOpcode; 252 uint8_t MIMGEncoding; 253 uint8_t VDataDwords; 254 uint8_t VAddrDwords; 255 }; 256 257 LLVM_READONLY 258 const MIMGInfo *getMIMGInfo(unsigned Opc); 259 260 LLVM_READONLY 261 int getMTBUFBaseOpcode(unsigned Opc); 262 263 LLVM_READONLY 264 int getMTBUFOpcode(unsigned BaseOpc, unsigned Elements); 265 266 LLVM_READONLY 267 int getMTBUFElements(unsigned Opc); 268 269 LLVM_READONLY 270 bool getMTBUFHasVAddr(unsigned Opc); 271 272 LLVM_READONLY 273 bool getMTBUFHasSrsrc(unsigned Opc); 274 275 LLVM_READONLY 276 bool getMTBUFHasSoffset(unsigned Opc); 277 278 LLVM_READONLY 279 int getMUBUFBaseOpcode(unsigned Opc); 280 281 LLVM_READONLY 282 int getMUBUFOpcode(unsigned BaseOpc, unsigned Elements); 283 284 LLVM_READONLY 285 int getMUBUFElements(unsigned Opc); 286 287 LLVM_READONLY 288 bool getMUBUFHasVAddr(unsigned Opc); 289 290 LLVM_READONLY 291 bool getMUBUFHasSrsrc(unsigned Opc); 292 293 LLVM_READONLY 294 bool getMUBUFHasSoffset(unsigned Opc); 295 296 LLVM_READONLY 297 bool getSMEMIsBuffer(unsigned Opc); 298 299 LLVM_READONLY 300 const GcnBufferFormatInfo *getGcnBufferFormatInfo(uint8_t BitsPerComp, 301 uint8_t NumComponents, 302 uint8_t NumFormat, 303 const MCSubtargetInfo &STI); 304 LLVM_READONLY 305 const GcnBufferFormatInfo *getGcnBufferFormatInfo(uint8_t Format, 306 const MCSubtargetInfo &STI); 307 308 LLVM_READONLY 309 int getMCOpcode(uint16_t Opcode, unsigned Gen); 310 311 void initDefaultAMDKernelCodeT(amd_kernel_code_t &Header, 312 const MCSubtargetInfo *STI); 313 314 amdhsa::kernel_descriptor_t getDefaultAmdhsaKernelDescriptor( 315 const MCSubtargetInfo *STI); 316 317 bool isGroupSegment(const GlobalValue *GV); 318 bool isGlobalSegment(const GlobalValue *GV); 319 bool isReadOnlySegment(const GlobalValue *GV); 320 321 /// \returns True if constants should be emitted to .text section for given 322 /// target triple \p TT, false otherwise. 323 bool shouldEmitConstantsToTextSection(const Triple &TT); 324 325 /// \returns Integer value requested using \p F's \p Name attribute. 326 /// 327 /// \returns \p Default if attribute is not present. 328 /// 329 /// \returns \p Default and emits error if requested value cannot be converted 330 /// to integer. 331 int getIntegerAttribute(const Function &F, StringRef Name, int Default); 332 333 /// \returns A pair of integer values requested using \p F's \p Name attribute 334 /// in "first[,second]" format ("second" is optional unless \p OnlyFirstRequired 335 /// is false). 336 /// 337 /// \returns \p Default if attribute is not present. 338 /// 339 /// \returns \p Default and emits error if one of the requested values cannot be 340 /// converted to integer, or \p OnlyFirstRequired is false and "second" value is 341 /// not present. 342 std::pair<int, int> getIntegerPairAttribute(const Function &F, 343 StringRef Name, 344 std::pair<int, int> Default, 345 bool OnlyFirstRequired = false); 346 347 /// Represents the counter values to wait for in an s_waitcnt instruction. 348 /// 349 /// Large values (including the maximum possible integer) can be used to 350 /// represent "don't care" waits. 351 struct Waitcnt { 352 unsigned VmCnt = ~0u; 353 unsigned ExpCnt = ~0u; 354 unsigned LgkmCnt = ~0u; 355 unsigned VsCnt = ~0u; 356 357 Waitcnt() {} 358 Waitcnt(unsigned VmCnt, unsigned ExpCnt, unsigned LgkmCnt, unsigned VsCnt) 359 : VmCnt(VmCnt), ExpCnt(ExpCnt), LgkmCnt(LgkmCnt), VsCnt(VsCnt) {} 360 361 static Waitcnt allZero(const IsaVersion &Version) { 362 return Waitcnt(0, 0, 0, Version.Major >= 10 ? 0 : ~0u); 363 } 364 static Waitcnt allZeroExceptVsCnt() { return Waitcnt(0, 0, 0, ~0u); } 365 366 bool hasWait() const { 367 return VmCnt != ~0u || ExpCnt != ~0u || LgkmCnt != ~0u || VsCnt != ~0u; 368 } 369 370 bool dominates(const Waitcnt &Other) const { 371 return VmCnt <= Other.VmCnt && ExpCnt <= Other.ExpCnt && 372 LgkmCnt <= Other.LgkmCnt && VsCnt <= Other.VsCnt; 373 } 374 375 Waitcnt combined(const Waitcnt &Other) const { 376 return Waitcnt(std::min(VmCnt, Other.VmCnt), std::min(ExpCnt, Other.ExpCnt), 377 std::min(LgkmCnt, Other.LgkmCnt), 378 std::min(VsCnt, Other.VsCnt)); 379 } 380 }; 381 382 /// \returns Vmcnt bit mask for given isa \p Version. 383 unsigned getVmcntBitMask(const IsaVersion &Version); 384 385 /// \returns Expcnt bit mask for given isa \p Version. 386 unsigned getExpcntBitMask(const IsaVersion &Version); 387 388 /// \returns Lgkmcnt bit mask for given isa \p Version. 389 unsigned getLgkmcntBitMask(const IsaVersion &Version); 390 391 /// \returns Waitcnt bit mask for given isa \p Version. 392 unsigned getWaitcntBitMask(const IsaVersion &Version); 393 394 /// \returns Decoded Vmcnt from given \p Waitcnt for given isa \p Version. 395 unsigned decodeVmcnt(const IsaVersion &Version, unsigned Waitcnt); 396 397 /// \returns Decoded Expcnt from given \p Waitcnt for given isa \p Version. 398 unsigned decodeExpcnt(const IsaVersion &Version, unsigned Waitcnt); 399 400 /// \returns Decoded Lgkmcnt from given \p Waitcnt for given isa \p Version. 401 unsigned decodeLgkmcnt(const IsaVersion &Version, unsigned Waitcnt); 402 403 /// Decodes Vmcnt, Expcnt and Lgkmcnt from given \p Waitcnt for given isa 404 /// \p Version, and writes decoded values into \p Vmcnt, \p Expcnt and 405 /// \p Lgkmcnt respectively. 406 /// 407 /// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are decoded as follows: 408 /// \p Vmcnt = \p Waitcnt[3:0] (pre-gfx9 only) 409 /// \p Vmcnt = \p Waitcnt[3:0] | \p Waitcnt[15:14] (gfx9+ only) 410 /// \p Expcnt = \p Waitcnt[6:4] 411 /// \p Lgkmcnt = \p Waitcnt[11:8] (pre-gfx10 only) 412 /// \p Lgkmcnt = \p Waitcnt[13:8] (gfx10+ only) 413 void decodeWaitcnt(const IsaVersion &Version, unsigned Waitcnt, 414 unsigned &Vmcnt, unsigned &Expcnt, unsigned &Lgkmcnt); 415 416 Waitcnt decodeWaitcnt(const IsaVersion &Version, unsigned Encoded); 417 418 /// \returns \p Waitcnt with encoded \p Vmcnt for given isa \p Version. 419 unsigned encodeVmcnt(const IsaVersion &Version, unsigned Waitcnt, 420 unsigned Vmcnt); 421 422 /// \returns \p Waitcnt with encoded \p Expcnt for given isa \p Version. 423 unsigned encodeExpcnt(const IsaVersion &Version, unsigned Waitcnt, 424 unsigned Expcnt); 425 426 /// \returns \p Waitcnt with encoded \p Lgkmcnt for given isa \p Version. 427 unsigned encodeLgkmcnt(const IsaVersion &Version, unsigned Waitcnt, 428 unsigned Lgkmcnt); 429 430 /// Encodes \p Vmcnt, \p Expcnt and \p Lgkmcnt into Waitcnt for given isa 431 /// \p Version. 432 /// 433 /// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are encoded as follows: 434 /// Waitcnt[3:0] = \p Vmcnt (pre-gfx9 only) 435 /// Waitcnt[3:0] = \p Vmcnt[3:0] (gfx9+ only) 436 /// Waitcnt[6:4] = \p Expcnt 437 /// Waitcnt[11:8] = \p Lgkmcnt (pre-gfx10 only) 438 /// Waitcnt[13:8] = \p Lgkmcnt (gfx10+ only) 439 /// Waitcnt[15:14] = \p Vmcnt[5:4] (gfx9+ only) 440 /// 441 /// \returns Waitcnt with encoded \p Vmcnt, \p Expcnt and \p Lgkmcnt for given 442 /// isa \p Version. 443 unsigned encodeWaitcnt(const IsaVersion &Version, 444 unsigned Vmcnt, unsigned Expcnt, unsigned Lgkmcnt); 445 446 unsigned encodeWaitcnt(const IsaVersion &Version, const Waitcnt &Decoded); 447 448 namespace Hwreg { 449 450 LLVM_READONLY 451 int64_t getHwregId(const StringRef Name); 452 453 LLVM_READNONE 454 bool isValidHwreg(int64_t Id, const MCSubtargetInfo &STI); 455 456 LLVM_READNONE 457 bool isValidHwreg(int64_t Id); 458 459 LLVM_READNONE 460 bool isValidHwregOffset(int64_t Offset); 461 462 LLVM_READNONE 463 bool isValidHwregWidth(int64_t Width); 464 465 LLVM_READNONE 466 uint64_t encodeHwreg(uint64_t Id, uint64_t Offset, uint64_t Width); 467 468 LLVM_READNONE 469 StringRef getHwreg(unsigned Id, const MCSubtargetInfo &STI); 470 471 void decodeHwreg(unsigned Val, unsigned &Id, unsigned &Offset, unsigned &Width); 472 473 } // namespace Hwreg 474 475 namespace SendMsg { 476 477 LLVM_READONLY 478 int64_t getMsgId(const StringRef Name); 479 480 LLVM_READONLY 481 int64_t getMsgOpId(int64_t MsgId, const StringRef Name); 482 483 LLVM_READNONE 484 StringRef getMsgName(int64_t MsgId); 485 486 LLVM_READNONE 487 StringRef getMsgOpName(int64_t MsgId, int64_t OpId); 488 489 LLVM_READNONE 490 bool isValidMsgId(int64_t MsgId, const MCSubtargetInfo &STI, bool Strict = true); 491 492 LLVM_READNONE 493 bool isValidMsgOp(int64_t MsgId, int64_t OpId, bool Strict = true); 494 495 LLVM_READNONE 496 bool isValidMsgStream(int64_t MsgId, int64_t OpId, int64_t StreamId, bool Strict = true); 497 498 LLVM_READNONE 499 bool msgRequiresOp(int64_t MsgId); 500 501 LLVM_READNONE 502 bool msgSupportsStream(int64_t MsgId, int64_t OpId); 503 504 void decodeMsg(unsigned Val, 505 uint16_t &MsgId, 506 uint16_t &OpId, 507 uint16_t &StreamId); 508 509 LLVM_READNONE 510 uint64_t encodeMsg(uint64_t MsgId, 511 uint64_t OpId, 512 uint64_t StreamId); 513 514 } // namespace SendMsg 515 516 517 unsigned getInitialPSInputAddr(const Function &F); 518 519 LLVM_READNONE 520 bool isShader(CallingConv::ID CC); 521 522 LLVM_READNONE 523 bool isCompute(CallingConv::ID CC); 524 525 LLVM_READNONE 526 bool isEntryFunctionCC(CallingConv::ID CC); 527 528 // FIXME: Remove this when calling conventions cleaned up 529 LLVM_READNONE 530 inline bool isKernel(CallingConv::ID CC) { 531 switch (CC) { 532 case CallingConv::AMDGPU_KERNEL: 533 case CallingConv::SPIR_KERNEL: 534 return true; 535 default: 536 return false; 537 } 538 } 539 540 bool hasXNACK(const MCSubtargetInfo &STI); 541 bool hasSRAMECC(const MCSubtargetInfo &STI); 542 bool hasMIMG_R128(const MCSubtargetInfo &STI); 543 bool hasGFX10A16(const MCSubtargetInfo &STI); 544 bool hasPackedD16(const MCSubtargetInfo &STI); 545 546 bool isSI(const MCSubtargetInfo &STI); 547 bool isCI(const MCSubtargetInfo &STI); 548 bool isVI(const MCSubtargetInfo &STI); 549 bool isGFX9(const MCSubtargetInfo &STI); 550 bool isGFX10(const MCSubtargetInfo &STI); 551 552 /// Is Reg - scalar register 553 bool isSGPR(unsigned Reg, const MCRegisterInfo* TRI); 554 555 /// Is there any intersection between registers 556 bool isRegIntersect(unsigned Reg0, unsigned Reg1, const MCRegisterInfo* TRI); 557 558 /// If \p Reg is a pseudo reg, return the correct hardware register given 559 /// \p STI otherwise return \p Reg. 560 unsigned getMCReg(unsigned Reg, const MCSubtargetInfo &STI); 561 562 /// Convert hardware register \p Reg to a pseudo register 563 LLVM_READNONE 564 unsigned mc2PseudoReg(unsigned Reg); 565 566 /// Can this operand also contain immediate values? 567 bool isSISrcOperand(const MCInstrDesc &Desc, unsigned OpNo); 568 569 /// Is this floating-point operand? 570 bool isSISrcFPOperand(const MCInstrDesc &Desc, unsigned OpNo); 571 572 /// Does this opearnd support only inlinable literals? 573 bool isSISrcInlinableOperand(const MCInstrDesc &Desc, unsigned OpNo); 574 575 /// Get the size in bits of a register from the register class \p RC. 576 unsigned getRegBitWidth(unsigned RCID); 577 578 /// Get the size in bits of a register from the register class \p RC. 579 unsigned getRegBitWidth(const MCRegisterClass &RC); 580 581 /// Get size of register operand 582 unsigned getRegOperandSize(const MCRegisterInfo *MRI, const MCInstrDesc &Desc, 583 unsigned OpNo); 584 585 LLVM_READNONE 586 inline unsigned getOperandSize(const MCOperandInfo &OpInfo) { 587 switch (OpInfo.OperandType) { 588 case AMDGPU::OPERAND_REG_IMM_INT32: 589 case AMDGPU::OPERAND_REG_IMM_FP32: 590 case AMDGPU::OPERAND_REG_INLINE_C_INT32: 591 case AMDGPU::OPERAND_REG_INLINE_C_FP32: 592 case AMDGPU::OPERAND_REG_INLINE_AC_INT32: 593 case AMDGPU::OPERAND_REG_INLINE_AC_FP32: 594 return 4; 595 596 case AMDGPU::OPERAND_REG_IMM_INT64: 597 case AMDGPU::OPERAND_REG_IMM_FP64: 598 case AMDGPU::OPERAND_REG_INLINE_C_INT64: 599 case AMDGPU::OPERAND_REG_INLINE_C_FP64: 600 return 8; 601 602 case AMDGPU::OPERAND_REG_IMM_INT16: 603 case AMDGPU::OPERAND_REG_IMM_FP16: 604 case AMDGPU::OPERAND_REG_INLINE_C_INT16: 605 case AMDGPU::OPERAND_REG_INLINE_C_FP16: 606 case AMDGPU::OPERAND_REG_INLINE_C_V2INT16: 607 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16: 608 case AMDGPU::OPERAND_REG_INLINE_AC_INT16: 609 case AMDGPU::OPERAND_REG_INLINE_AC_FP16: 610 case AMDGPU::OPERAND_REG_INLINE_AC_V2INT16: 611 case AMDGPU::OPERAND_REG_INLINE_AC_V2FP16: 612 case AMDGPU::OPERAND_REG_IMM_V2INT16: 613 case AMDGPU::OPERAND_REG_IMM_V2FP16: 614 return 2; 615 616 default: 617 llvm_unreachable("unhandled operand type"); 618 } 619 } 620 621 LLVM_READNONE 622 inline unsigned getOperandSize(const MCInstrDesc &Desc, unsigned OpNo) { 623 return getOperandSize(Desc.OpInfo[OpNo]); 624 } 625 626 /// Is this literal inlinable 627 LLVM_READNONE 628 bool isInlinableLiteral64(int64_t Literal, bool HasInv2Pi); 629 630 LLVM_READNONE 631 bool isInlinableLiteral32(int32_t Literal, bool HasInv2Pi); 632 633 LLVM_READNONE 634 bool isInlinableLiteral16(int16_t Literal, bool HasInv2Pi); 635 636 LLVM_READNONE 637 bool isInlinableLiteralV216(int32_t Literal, bool HasInv2Pi); 638 639 bool isArgPassedInSGPR(const Argument *Arg); 640 641 LLVM_READONLY 642 bool isLegalSMRDEncodedUnsignedOffset(const MCSubtargetInfo &ST, 643 int64_t EncodedOffset); 644 645 LLVM_READONLY 646 bool isLegalSMRDEncodedSignedOffset(const MCSubtargetInfo &ST, 647 int64_t EncodedOffset, 648 bool IsBuffer); 649 650 /// Convert \p ByteOffset to dwords if the subtarget uses dword SMRD immediate 651 /// offsets. 652 uint64_t convertSMRDOffsetUnits(const MCSubtargetInfo &ST, uint64_t ByteOffset); 653 654 /// \returns The encoding that will be used for \p ByteOffset in the 655 /// SMRD offset field, or None if it won't fit. On GFX9 and GFX10 656 /// S_LOAD instructions have a signed offset, on other subtargets it is 657 /// unsigned. S_BUFFER has an unsigned offset for all subtargets. 658 Optional<int64_t> getSMRDEncodedOffset(const MCSubtargetInfo &ST, 659 int64_t ByteOffset, bool IsBuffer); 660 661 /// \return The encoding that can be used for a 32-bit literal offset in an SMRD 662 /// instruction. This is only useful on CI.s 663 Optional<int64_t> getSMRDEncodedLiteralOffset32(const MCSubtargetInfo &ST, 664 int64_t ByteOffset); 665 666 /// \returns true if this offset is small enough to fit in the SMRD 667 /// offset field. \p ByteOffset should be the offset in bytes and 668 /// not the encoded offset. 669 bool isLegalSMRDImmOffset(const MCSubtargetInfo &ST, int64_t ByteOffset); 670 671 bool splitMUBUFOffset(uint32_t Imm, uint32_t &SOffset, uint32_t &ImmOffset, 672 const GCNSubtarget *Subtarget, uint32_t Align = 4); 673 674 /// \returns true if the intrinsic is divergent 675 bool isIntrinsicSourceOfDivergence(unsigned IntrID); 676 677 // Track defaults for fields in the MODE registser. 678 struct SIModeRegisterDefaults { 679 /// Floating point opcodes that support exception flag gathering quiet and 680 /// propagate signaling NaN inputs per IEEE 754-2008. Min_dx10 and max_dx10 681 /// become IEEE 754- 2008 compliant due to signaling NaN propagation and 682 /// quieting. 683 bool IEEE : 1; 684 685 /// Used by the vector ALU to force DX10-style treatment of NaNs: when set, 686 /// clamp NaN to zero; otherwise, pass NaN through. 687 bool DX10Clamp : 1; 688 689 /// If this is set, neither input or output denormals are flushed for most f32 690 /// instructions. 691 bool FP32InputDenormals : 1; 692 bool FP32OutputDenormals : 1; 693 694 /// If this is set, neither input or output denormals are flushed for both f64 695 /// and f16/v2f16 instructions. 696 bool FP64FP16InputDenormals : 1; 697 bool FP64FP16OutputDenormals : 1; 698 699 SIModeRegisterDefaults() : 700 IEEE(true), 701 DX10Clamp(true), 702 FP32InputDenormals(true), 703 FP32OutputDenormals(true), 704 FP64FP16InputDenormals(true), 705 FP64FP16OutputDenormals(true) {} 706 707 SIModeRegisterDefaults(const Function &F); 708 709 static SIModeRegisterDefaults getDefaultForCallingConv(CallingConv::ID CC) { 710 const bool IsCompute = AMDGPU::isCompute(CC); 711 712 SIModeRegisterDefaults Mode; 713 Mode.IEEE = IsCompute; 714 return Mode; 715 } 716 717 bool operator ==(const SIModeRegisterDefaults Other) const { 718 return IEEE == Other.IEEE && DX10Clamp == Other.DX10Clamp && 719 FP32InputDenormals == Other.FP32InputDenormals && 720 FP32OutputDenormals == Other.FP32OutputDenormals && 721 FP64FP16InputDenormals == Other.FP64FP16InputDenormals && 722 FP64FP16OutputDenormals == Other.FP64FP16OutputDenormals; 723 } 724 725 bool allFP32Denormals() const { 726 return FP32InputDenormals && FP32OutputDenormals; 727 } 728 729 bool allFP64FP16Denormals() const { 730 return FP64FP16InputDenormals && FP64FP16OutputDenormals; 731 } 732 733 /// Get the encoding value for the FP_DENORM bits of the mode register for the 734 /// FP32 denormal mode. 735 uint32_t fpDenormModeSPValue() const { 736 if (FP32InputDenormals && FP32OutputDenormals) 737 return FP_DENORM_FLUSH_NONE; 738 if (FP32InputDenormals) 739 return FP_DENORM_FLUSH_OUT; 740 if (FP32OutputDenormals) 741 return FP_DENORM_FLUSH_IN; 742 return FP_DENORM_FLUSH_IN_FLUSH_OUT; 743 } 744 745 /// Get the encoding value for the FP_DENORM bits of the mode register for the 746 /// FP64/FP16 denormal mode. 747 uint32_t fpDenormModeDPValue() const { 748 if (FP64FP16InputDenormals && FP64FP16OutputDenormals) 749 return FP_DENORM_FLUSH_NONE; 750 if (FP64FP16InputDenormals) 751 return FP_DENORM_FLUSH_OUT; 752 if (FP64FP16OutputDenormals) 753 return FP_DENORM_FLUSH_IN; 754 return FP_DENORM_FLUSH_IN_FLUSH_OUT; 755 } 756 757 /// Returns true if a flag is compatible if it's enabled in the callee, but 758 /// disabled in the caller. 759 static bool oneWayCompatible(bool CallerMode, bool CalleeMode) { 760 return CallerMode == CalleeMode || (!CallerMode && CalleeMode); 761 } 762 763 // FIXME: Inlining should be OK for dx10-clamp, since the caller's mode should 764 // be able to override. 765 bool isInlineCompatible(SIModeRegisterDefaults CalleeMode) const { 766 if (DX10Clamp != CalleeMode.DX10Clamp) 767 return false; 768 if (IEEE != CalleeMode.IEEE) 769 return false; 770 771 // Allow inlining denormals enabled into denormals flushed functions. 772 return oneWayCompatible(FP64FP16InputDenormals, CalleeMode.FP64FP16InputDenormals) && 773 oneWayCompatible(FP64FP16OutputDenormals, CalleeMode.FP64FP16OutputDenormals) && 774 oneWayCompatible(FP32InputDenormals, CalleeMode.FP32InputDenormals) && 775 oneWayCompatible(FP32OutputDenormals, CalleeMode.FP32OutputDenormals); 776 } 777 }; 778 779 LLVM_READNONE 780 bool isInlinableIntLiteral(int64_t Literal); 781 782 } // end namespace AMDGPU 783 } // end namespace llvm 784 785 #endif // LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUBASEINFO_H 786