1 //===- AMDGPUBaseInfo.cpp - AMDGPU Base encoding information --------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 10 #include "AMDGPUBaseInfo.h" 11 #include "AMDGPUTargetTransformInfo.h" 12 #include "AMDGPU.h" 13 #include "SIDefines.h" 14 #include "llvm/ADT/StringRef.h" 15 #include "llvm/ADT/Triple.h" 16 #include "llvm/BinaryFormat/ELF.h" 17 #include "llvm/CodeGen/MachineMemOperand.h" 18 #include "llvm/IR/Attributes.h" 19 #include "llvm/IR/Constants.h" 20 #include "llvm/IR/Function.h" 21 #include "llvm/IR/GlobalValue.h" 22 #include "llvm/IR/Instruction.h" 23 #include "llvm/IR/LLVMContext.h" 24 #include "llvm/IR/Module.h" 25 #include "llvm/MC/MCContext.h" 26 #include "llvm/MC/MCInstrDesc.h" 27 #include "llvm/MC/MCInstrInfo.h" 28 #include "llvm/MC/MCRegisterInfo.h" 29 #include "llvm/MC/MCSectionELF.h" 30 #include "llvm/MC/MCSubtargetInfo.h" 31 #include "llvm/MC/SubtargetFeature.h" 32 #include "llvm/Support/Casting.h" 33 #include "llvm/Support/ErrorHandling.h" 34 #include "llvm/Support/MathExtras.h" 35 #include <algorithm> 36 #include <cassert> 37 #include <cstdint> 38 #include <cstring> 39 #include <utility> 40 41 #include "MCTargetDesc/AMDGPUMCTargetDesc.h" 42 43 #define GET_INSTRINFO_NAMED_OPS 44 #define GET_INSTRMAP_INFO 45 #include "AMDGPUGenInstrInfo.inc" 46 #undef GET_INSTRMAP_INFO 47 #undef GET_INSTRINFO_NAMED_OPS 48 49 namespace { 50 51 /// \returns Bit mask for given bit \p Shift and bit \p Width. 52 unsigned getBitMask(unsigned Shift, unsigned Width) { 53 return ((1 << Width) - 1) << Shift; 54 } 55 56 /// Packs \p Src into \p Dst for given bit \p Shift and bit \p Width. 57 /// 58 /// \returns Packed \p Dst. 59 unsigned packBits(unsigned Src, unsigned Dst, unsigned Shift, unsigned Width) { 60 Dst &= ~(1 << Shift) & ~getBitMask(Shift, Width); 61 Dst |= (Src << Shift) & getBitMask(Shift, Width); 62 return Dst; 63 } 64 65 /// Unpacks bits from \p Src for given bit \p Shift and bit \p Width. 66 /// 67 /// \returns Unpacked bits. 68 unsigned unpackBits(unsigned Src, unsigned Shift, unsigned Width) { 69 return (Src & getBitMask(Shift, Width)) >> Shift; 70 } 71 72 /// \returns Vmcnt bit shift (lower bits). 73 unsigned getVmcntBitShiftLo() { return 0; } 74 75 /// \returns Vmcnt bit width (lower bits). 76 unsigned getVmcntBitWidthLo() { return 4; } 77 78 /// \returns Expcnt bit shift. 79 unsigned getExpcntBitShift() { return 4; } 80 81 /// \returns Expcnt bit width. 82 unsigned getExpcntBitWidth() { return 3; } 83 84 /// \returns Lgkmcnt bit shift. 85 unsigned getLgkmcntBitShift() { return 8; } 86 87 /// \returns Lgkmcnt bit width. 88 unsigned getLgkmcntBitWidth() { return 4; } 89 90 /// \returns Vmcnt bit shift (higher bits). 91 unsigned getVmcntBitShiftHi() { return 14; } 92 93 /// \returns Vmcnt bit width (higher bits). 94 unsigned getVmcntBitWidthHi() { return 2; } 95 96 } // end namespace anonymous 97 98 namespace llvm { 99 100 namespace AMDGPU { 101 102 LLVM_READNONE 103 static inline Channels indexToChannel(unsigned Channel) { 104 switch (Channel) { 105 case 1: 106 return AMDGPU::Channels_1; 107 case 2: 108 return AMDGPU::Channels_2; 109 case 3: 110 return AMDGPU::Channels_3; 111 case 4: 112 return AMDGPU::Channels_4; 113 default: 114 llvm_unreachable("invalid MIMG channel"); 115 } 116 } 117 118 119 // FIXME: Need to handle d16 images correctly. 120 static unsigned rcToChannels(unsigned RCID) { 121 switch (RCID) { 122 case AMDGPU::VGPR_32RegClassID: 123 return 1; 124 case AMDGPU::VReg_64RegClassID: 125 return 2; 126 case AMDGPU::VReg_96RegClassID: 127 return 3; 128 case AMDGPU::VReg_128RegClassID: 129 return 4; 130 default: 131 llvm_unreachable("invalid MIMG register class"); 132 } 133 } 134 135 int getMaskedMIMGOp(const MCInstrInfo &MII, unsigned Opc, unsigned NewChannels) { 136 AMDGPU::Channels Channel = AMDGPU::indexToChannel(NewChannels); 137 unsigned OrigChannels = rcToChannels(MII.get(Opc).OpInfo[0].RegClass); 138 if (NewChannels == OrigChannels) 139 return Opc; 140 141 switch (OrigChannels) { 142 case 1: 143 return AMDGPU::getMaskedMIMGOp1(Opc, Channel); 144 case 2: 145 return AMDGPU::getMaskedMIMGOp2(Opc, Channel); 146 case 3: 147 return AMDGPU::getMaskedMIMGOp3(Opc, Channel); 148 case 4: 149 return AMDGPU::getMaskedMIMGOp4(Opc, Channel); 150 default: 151 llvm_unreachable("invalid MIMG channel"); 152 } 153 } 154 155 int getMaskedMIMGAtomicOp(const MCInstrInfo &MII, unsigned Opc, unsigned NewChannels) { 156 assert(AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst) != -1); 157 assert(NewChannels == 1 || NewChannels == 2 || NewChannels == 4); 158 159 unsigned OrigChannels = rcToChannels(MII.get(Opc).OpInfo[0].RegClass); 160 assert(OrigChannels == 1 || OrigChannels == 2 || OrigChannels == 4); 161 162 if (NewChannels == OrigChannels) return Opc; 163 164 if (OrigChannels <= 2 && NewChannels <= 2) { 165 // This is an ordinary atomic (not an atomic_cmpswap) 166 return (OrigChannels == 1)? 167 AMDGPU::getMIMGAtomicOp1(Opc) : AMDGPU::getMIMGAtomicOp2(Opc); 168 } else if (OrigChannels >= 2 && NewChannels >= 2) { 169 // This is an atomic_cmpswap 170 return (OrigChannels == 2)? 171 AMDGPU::getMIMGAtomicOp1(Opc) : AMDGPU::getMIMGAtomicOp2(Opc); 172 } else { // invalid OrigChannels/NewChannels value 173 return -1; 174 } 175 } 176 177 // Wrapper for Tablegen'd function. enum Subtarget is not defined in any 178 // header files, so we need to wrap it in a function that takes unsigned 179 // instead. 180 int getMCOpcode(uint16_t Opcode, unsigned Gen) { 181 return getMCOpcodeGen(Opcode, static_cast<Subtarget>(Gen)); 182 } 183 184 namespace IsaInfo { 185 186 IsaVersion getIsaVersion(const FeatureBitset &Features) { 187 // GCN GFX6 (Southern Islands (SI)). 188 if (Features.test(FeatureISAVersion6_0_0)) 189 return {6, 0, 0}; 190 if (Features.test(FeatureISAVersion6_0_1)) 191 return {6, 0, 1}; 192 193 // GCN GFX7 (Sea Islands (CI)). 194 if (Features.test(FeatureISAVersion7_0_0)) 195 return {7, 0, 0}; 196 if (Features.test(FeatureISAVersion7_0_1)) 197 return {7, 0, 1}; 198 if (Features.test(FeatureISAVersion7_0_2)) 199 return {7, 0, 2}; 200 if (Features.test(FeatureISAVersion7_0_3)) 201 return {7, 0, 3}; 202 if (Features.test(FeatureISAVersion7_0_4)) 203 return {7, 0, 4}; 204 if (Features.test(FeatureSeaIslands)) 205 return {7, 0, 0}; 206 207 // GCN GFX8 (Volcanic Islands (VI)). 208 if (Features.test(FeatureISAVersion8_0_1)) 209 return {8, 0, 1}; 210 if (Features.test(FeatureISAVersion8_0_2)) 211 return {8, 0, 2}; 212 if (Features.test(FeatureISAVersion8_0_3)) 213 return {8, 0, 3}; 214 if (Features.test(FeatureISAVersion8_1_0)) 215 return {8, 1, 0}; 216 if (Features.test(FeatureVolcanicIslands)) 217 return {8, 0, 0}; 218 219 // GCN GFX9. 220 if (Features.test(FeatureISAVersion9_0_0)) 221 return {9, 0, 0}; 222 if (Features.test(FeatureISAVersion9_0_2)) 223 return {9, 0, 2}; 224 if (Features.test(FeatureISAVersion9_0_4)) 225 return {9, 0, 4}; 226 if (Features.test(FeatureISAVersion9_0_6)) 227 return {9, 0, 6}; 228 if (Features.test(FeatureGFX9)) 229 return {9, 0, 0}; 230 231 if (!Features.test(FeatureGCN) || Features.test(FeatureSouthernIslands)) 232 return {0, 0, 0}; 233 return {7, 0, 0}; 234 } 235 236 void streamIsaVersion(const MCSubtargetInfo *STI, raw_ostream &Stream) { 237 auto TargetTriple = STI->getTargetTriple(); 238 auto ISAVersion = IsaInfo::getIsaVersion(STI->getFeatureBits()); 239 240 Stream << TargetTriple.getArchName() << '-' 241 << TargetTriple.getVendorName() << '-' 242 << TargetTriple.getOSName() << '-' 243 << TargetTriple.getEnvironmentName() << '-' 244 << "gfx" 245 << ISAVersion.Major 246 << ISAVersion.Minor 247 << ISAVersion.Stepping; 248 Stream.flush(); 249 } 250 251 bool hasCodeObjectV3(const FeatureBitset &Features) { 252 return Features.test(FeatureCodeObjectV3); 253 } 254 255 unsigned getWavefrontSize(const FeatureBitset &Features) { 256 if (Features.test(FeatureWavefrontSize16)) 257 return 16; 258 if (Features.test(FeatureWavefrontSize32)) 259 return 32; 260 261 return 64; 262 } 263 264 unsigned getLocalMemorySize(const FeatureBitset &Features) { 265 if (Features.test(FeatureLocalMemorySize32768)) 266 return 32768; 267 if (Features.test(FeatureLocalMemorySize65536)) 268 return 65536; 269 270 return 0; 271 } 272 273 unsigned getEUsPerCU(const FeatureBitset &Features) { 274 return 4; 275 } 276 277 unsigned getMaxWorkGroupsPerCU(const FeatureBitset &Features, 278 unsigned FlatWorkGroupSize) { 279 if (!Features.test(FeatureGCN)) 280 return 8; 281 unsigned N = getWavesPerWorkGroup(Features, FlatWorkGroupSize); 282 if (N == 1) 283 return 40; 284 N = 40 / N; 285 return std::min(N, 16u); 286 } 287 288 unsigned getMaxWavesPerCU(const FeatureBitset &Features) { 289 return getMaxWavesPerEU(Features) * getEUsPerCU(Features); 290 } 291 292 unsigned getMaxWavesPerCU(const FeatureBitset &Features, 293 unsigned FlatWorkGroupSize) { 294 return getWavesPerWorkGroup(Features, FlatWorkGroupSize); 295 } 296 297 unsigned getMinWavesPerEU(const FeatureBitset &Features) { 298 return 1; 299 } 300 301 unsigned getMaxWavesPerEU(const FeatureBitset &Features) { 302 if (!Features.test(FeatureGCN)) 303 return 8; 304 // FIXME: Need to take scratch memory into account. 305 return 10; 306 } 307 308 unsigned getMaxWavesPerEU(const FeatureBitset &Features, 309 unsigned FlatWorkGroupSize) { 310 return alignTo(getMaxWavesPerCU(Features, FlatWorkGroupSize), 311 getEUsPerCU(Features)) / getEUsPerCU(Features); 312 } 313 314 unsigned getMinFlatWorkGroupSize(const FeatureBitset &Features) { 315 return 1; 316 } 317 318 unsigned getMaxFlatWorkGroupSize(const FeatureBitset &Features) { 319 return 2048; 320 } 321 322 unsigned getWavesPerWorkGroup(const FeatureBitset &Features, 323 unsigned FlatWorkGroupSize) { 324 return alignTo(FlatWorkGroupSize, getWavefrontSize(Features)) / 325 getWavefrontSize(Features); 326 } 327 328 unsigned getSGPRAllocGranule(const FeatureBitset &Features) { 329 IsaVersion Version = getIsaVersion(Features); 330 if (Version.Major >= 8) 331 return 16; 332 return 8; 333 } 334 335 unsigned getSGPREncodingGranule(const FeatureBitset &Features) { 336 return 8; 337 } 338 339 unsigned getTotalNumSGPRs(const FeatureBitset &Features) { 340 IsaVersion Version = getIsaVersion(Features); 341 if (Version.Major >= 8) 342 return 800; 343 return 512; 344 } 345 346 unsigned getAddressableNumSGPRs(const FeatureBitset &Features) { 347 if (Features.test(FeatureSGPRInitBug)) 348 return FIXED_NUM_SGPRS_FOR_INIT_BUG; 349 350 IsaVersion Version = getIsaVersion(Features); 351 if (Version.Major >= 8) 352 return 102; 353 return 104; 354 } 355 356 unsigned getMinNumSGPRs(const FeatureBitset &Features, unsigned WavesPerEU) { 357 assert(WavesPerEU != 0); 358 359 if (WavesPerEU >= getMaxWavesPerEU(Features)) 360 return 0; 361 362 unsigned MinNumSGPRs = getTotalNumSGPRs(Features) / (WavesPerEU + 1); 363 if (Features.test(FeatureTrapHandler)) 364 MinNumSGPRs -= std::min(MinNumSGPRs, (unsigned)TRAP_NUM_SGPRS); 365 MinNumSGPRs = alignDown(MinNumSGPRs, getSGPRAllocGranule(Features)) + 1; 366 return std::min(MinNumSGPRs, getAddressableNumSGPRs(Features)); 367 } 368 369 unsigned getMaxNumSGPRs(const FeatureBitset &Features, unsigned WavesPerEU, 370 bool Addressable) { 371 assert(WavesPerEU != 0); 372 373 IsaVersion Version = getIsaVersion(Features); 374 unsigned AddressableNumSGPRs = getAddressableNumSGPRs(Features); 375 if (Version.Major >= 8 && !Addressable) 376 AddressableNumSGPRs = 112; 377 unsigned MaxNumSGPRs = getTotalNumSGPRs(Features) / WavesPerEU; 378 if (Features.test(FeatureTrapHandler)) 379 MaxNumSGPRs -= std::min(MaxNumSGPRs, (unsigned)TRAP_NUM_SGPRS); 380 MaxNumSGPRs = alignDown(MaxNumSGPRs, getSGPRAllocGranule(Features)); 381 return std::min(MaxNumSGPRs, AddressableNumSGPRs); 382 } 383 384 unsigned getVGPRAllocGranule(const FeatureBitset &Features) { 385 return 4; 386 } 387 388 unsigned getVGPREncodingGranule(const FeatureBitset &Features) { 389 return getVGPRAllocGranule(Features); 390 } 391 392 unsigned getTotalNumVGPRs(const FeatureBitset &Features) { 393 return 256; 394 } 395 396 unsigned getAddressableNumVGPRs(const FeatureBitset &Features) { 397 return getTotalNumVGPRs(Features); 398 } 399 400 unsigned getMinNumVGPRs(const FeatureBitset &Features, unsigned WavesPerEU) { 401 assert(WavesPerEU != 0); 402 403 if (WavesPerEU >= getMaxWavesPerEU(Features)) 404 return 0; 405 unsigned MinNumVGPRs = 406 alignDown(getTotalNumVGPRs(Features) / (WavesPerEU + 1), 407 getVGPRAllocGranule(Features)) + 1; 408 return std::min(MinNumVGPRs, getAddressableNumVGPRs(Features)); 409 } 410 411 unsigned getMaxNumVGPRs(const FeatureBitset &Features, unsigned WavesPerEU) { 412 assert(WavesPerEU != 0); 413 414 unsigned MaxNumVGPRs = alignDown(getTotalNumVGPRs(Features) / WavesPerEU, 415 getVGPRAllocGranule(Features)); 416 unsigned AddressableNumVGPRs = getAddressableNumVGPRs(Features); 417 return std::min(MaxNumVGPRs, AddressableNumVGPRs); 418 } 419 420 } // end namespace IsaInfo 421 422 void initDefaultAMDKernelCodeT(amd_kernel_code_t &Header, 423 const FeatureBitset &Features) { 424 IsaInfo::IsaVersion ISA = IsaInfo::getIsaVersion(Features); 425 426 memset(&Header, 0, sizeof(Header)); 427 428 Header.amd_kernel_code_version_major = 1; 429 Header.amd_kernel_code_version_minor = 2; 430 Header.amd_machine_kind = 1; // AMD_MACHINE_KIND_AMDGPU 431 Header.amd_machine_version_major = ISA.Major; 432 Header.amd_machine_version_minor = ISA.Minor; 433 Header.amd_machine_version_stepping = ISA.Stepping; 434 Header.kernel_code_entry_byte_offset = sizeof(Header); 435 // wavefront_size is specified as a power of 2: 2^6 = 64 threads. 436 Header.wavefront_size = 6; 437 438 // If the code object does not support indirect functions, then the value must 439 // be 0xffffffff. 440 Header.call_convention = -1; 441 442 // These alignment values are specified in powers of two, so alignment = 443 // 2^n. The minimum alignment is 2^4 = 16. 444 Header.kernarg_segment_alignment = 4; 445 Header.group_segment_alignment = 4; 446 Header.private_segment_alignment = 4; 447 } 448 449 bool isGroupSegment(const GlobalValue *GV) { 450 return GV->getType()->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS; 451 } 452 453 bool isGlobalSegment(const GlobalValue *GV) { 454 return GV->getType()->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS; 455 } 456 457 bool isReadOnlySegment(const GlobalValue *GV) { 458 return GV->getType()->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS || 459 GV->getType()->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS_32BIT; 460 } 461 462 bool shouldEmitConstantsToTextSection(const Triple &TT) { 463 return TT.getOS() != Triple::AMDHSA; 464 } 465 466 int getIntegerAttribute(const Function &F, StringRef Name, int Default) { 467 Attribute A = F.getFnAttribute(Name); 468 int Result = Default; 469 470 if (A.isStringAttribute()) { 471 StringRef Str = A.getValueAsString(); 472 if (Str.getAsInteger(0, Result)) { 473 LLVMContext &Ctx = F.getContext(); 474 Ctx.emitError("can't parse integer attribute " + Name); 475 } 476 } 477 478 return Result; 479 } 480 481 std::pair<int, int> getIntegerPairAttribute(const Function &F, 482 StringRef Name, 483 std::pair<int, int> Default, 484 bool OnlyFirstRequired) { 485 Attribute A = F.getFnAttribute(Name); 486 if (!A.isStringAttribute()) 487 return Default; 488 489 LLVMContext &Ctx = F.getContext(); 490 std::pair<int, int> Ints = Default; 491 std::pair<StringRef, StringRef> Strs = A.getValueAsString().split(','); 492 if (Strs.first.trim().getAsInteger(0, Ints.first)) { 493 Ctx.emitError("can't parse first integer attribute " + Name); 494 return Default; 495 } 496 if (Strs.second.trim().getAsInteger(0, Ints.second)) { 497 if (!OnlyFirstRequired || !Strs.second.trim().empty()) { 498 Ctx.emitError("can't parse second integer attribute " + Name); 499 return Default; 500 } 501 } 502 503 return Ints; 504 } 505 506 unsigned getVmcntBitMask(const IsaInfo::IsaVersion &Version) { 507 unsigned VmcntLo = (1 << getVmcntBitWidthLo()) - 1; 508 if (Version.Major < 9) 509 return VmcntLo; 510 511 unsigned VmcntHi = ((1 << getVmcntBitWidthHi()) - 1) << getVmcntBitWidthLo(); 512 return VmcntLo | VmcntHi; 513 } 514 515 unsigned getExpcntBitMask(const IsaInfo::IsaVersion &Version) { 516 return (1 << getExpcntBitWidth()) - 1; 517 } 518 519 unsigned getLgkmcntBitMask(const IsaInfo::IsaVersion &Version) { 520 return (1 << getLgkmcntBitWidth()) - 1; 521 } 522 523 unsigned getWaitcntBitMask(const IsaInfo::IsaVersion &Version) { 524 unsigned VmcntLo = getBitMask(getVmcntBitShiftLo(), getVmcntBitWidthLo()); 525 unsigned Expcnt = getBitMask(getExpcntBitShift(), getExpcntBitWidth()); 526 unsigned Lgkmcnt = getBitMask(getLgkmcntBitShift(), getLgkmcntBitWidth()); 527 unsigned Waitcnt = VmcntLo | Expcnt | Lgkmcnt; 528 if (Version.Major < 9) 529 return Waitcnt; 530 531 unsigned VmcntHi = getBitMask(getVmcntBitShiftHi(), getVmcntBitWidthHi()); 532 return Waitcnt | VmcntHi; 533 } 534 535 unsigned decodeVmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt) { 536 unsigned VmcntLo = 537 unpackBits(Waitcnt, getVmcntBitShiftLo(), getVmcntBitWidthLo()); 538 if (Version.Major < 9) 539 return VmcntLo; 540 541 unsigned VmcntHi = 542 unpackBits(Waitcnt, getVmcntBitShiftHi(), getVmcntBitWidthHi()); 543 VmcntHi <<= getVmcntBitWidthLo(); 544 return VmcntLo | VmcntHi; 545 } 546 547 unsigned decodeExpcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt) { 548 return unpackBits(Waitcnt, getExpcntBitShift(), getExpcntBitWidth()); 549 } 550 551 unsigned decodeLgkmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt) { 552 return unpackBits(Waitcnt, getLgkmcntBitShift(), getLgkmcntBitWidth()); 553 } 554 555 void decodeWaitcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt, 556 unsigned &Vmcnt, unsigned &Expcnt, unsigned &Lgkmcnt) { 557 Vmcnt = decodeVmcnt(Version, Waitcnt); 558 Expcnt = decodeExpcnt(Version, Waitcnt); 559 Lgkmcnt = decodeLgkmcnt(Version, Waitcnt); 560 } 561 562 unsigned encodeVmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt, 563 unsigned Vmcnt) { 564 Waitcnt = 565 packBits(Vmcnt, Waitcnt, getVmcntBitShiftLo(), getVmcntBitWidthLo()); 566 if (Version.Major < 9) 567 return Waitcnt; 568 569 Vmcnt >>= getVmcntBitWidthLo(); 570 return packBits(Vmcnt, Waitcnt, getVmcntBitShiftHi(), getVmcntBitWidthHi()); 571 } 572 573 unsigned encodeExpcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt, 574 unsigned Expcnt) { 575 return packBits(Expcnt, Waitcnt, getExpcntBitShift(), getExpcntBitWidth()); 576 } 577 578 unsigned encodeLgkmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt, 579 unsigned Lgkmcnt) { 580 return packBits(Lgkmcnt, Waitcnt, getLgkmcntBitShift(), getLgkmcntBitWidth()); 581 } 582 583 unsigned encodeWaitcnt(const IsaInfo::IsaVersion &Version, 584 unsigned Vmcnt, unsigned Expcnt, unsigned Lgkmcnt) { 585 unsigned Waitcnt = getWaitcntBitMask(Version); 586 Waitcnt = encodeVmcnt(Version, Waitcnt, Vmcnt); 587 Waitcnt = encodeExpcnt(Version, Waitcnt, Expcnt); 588 Waitcnt = encodeLgkmcnt(Version, Waitcnt, Lgkmcnt); 589 return Waitcnt; 590 } 591 592 unsigned getInitialPSInputAddr(const Function &F) { 593 return getIntegerAttribute(F, "InitialPSInputAddr", 0); 594 } 595 596 bool isShader(CallingConv::ID cc) { 597 switch(cc) { 598 case CallingConv::AMDGPU_VS: 599 case CallingConv::AMDGPU_LS: 600 case CallingConv::AMDGPU_HS: 601 case CallingConv::AMDGPU_ES: 602 case CallingConv::AMDGPU_GS: 603 case CallingConv::AMDGPU_PS: 604 case CallingConv::AMDGPU_CS: 605 return true; 606 default: 607 return false; 608 } 609 } 610 611 bool isCompute(CallingConv::ID cc) { 612 return !isShader(cc) || cc == CallingConv::AMDGPU_CS; 613 } 614 615 bool isEntryFunctionCC(CallingConv::ID CC) { 616 switch (CC) { 617 case CallingConv::AMDGPU_KERNEL: 618 case CallingConv::SPIR_KERNEL: 619 case CallingConv::AMDGPU_VS: 620 case CallingConv::AMDGPU_GS: 621 case CallingConv::AMDGPU_PS: 622 case CallingConv::AMDGPU_CS: 623 case CallingConv::AMDGPU_ES: 624 case CallingConv::AMDGPU_HS: 625 case CallingConv::AMDGPU_LS: 626 return true; 627 default: 628 return false; 629 } 630 } 631 632 bool hasXNACK(const MCSubtargetInfo &STI) { 633 return STI.getFeatureBits()[AMDGPU::FeatureXNACK]; 634 } 635 636 bool hasMIMG_R128(const MCSubtargetInfo &STI) { 637 return STI.getFeatureBits()[AMDGPU::FeatureMIMG_R128]; 638 } 639 640 bool hasPackedD16(const MCSubtargetInfo &STI) { 641 return !STI.getFeatureBits()[AMDGPU::FeatureUnpackedD16VMem]; 642 } 643 644 bool isSI(const MCSubtargetInfo &STI) { 645 return STI.getFeatureBits()[AMDGPU::FeatureSouthernIslands]; 646 } 647 648 bool isCI(const MCSubtargetInfo &STI) { 649 return STI.getFeatureBits()[AMDGPU::FeatureSeaIslands]; 650 } 651 652 bool isVI(const MCSubtargetInfo &STI) { 653 return STI.getFeatureBits()[AMDGPU::FeatureVolcanicIslands]; 654 } 655 656 bool isGFX9(const MCSubtargetInfo &STI) { 657 return STI.getFeatureBits()[AMDGPU::FeatureGFX9]; 658 } 659 660 bool isGCN3Encoding(const MCSubtargetInfo &STI) { 661 return STI.getFeatureBits()[AMDGPU::FeatureGCN3Encoding]; 662 } 663 664 bool isSGPR(unsigned Reg, const MCRegisterInfo* TRI) { 665 const MCRegisterClass SGPRClass = TRI->getRegClass(AMDGPU::SReg_32RegClassID); 666 const unsigned FirstSubReg = TRI->getSubReg(Reg, 1); 667 return SGPRClass.contains(FirstSubReg != 0 ? FirstSubReg : Reg) || 668 Reg == AMDGPU::SCC; 669 } 670 671 bool isRegIntersect(unsigned Reg0, unsigned Reg1, const MCRegisterInfo* TRI) { 672 for (MCRegAliasIterator R(Reg0, TRI, true); R.isValid(); ++R) { 673 if (*R == Reg1) return true; 674 } 675 return false; 676 } 677 678 #define MAP_REG2REG \ 679 using namespace AMDGPU; \ 680 switch(Reg) { \ 681 default: return Reg; \ 682 CASE_CI_VI(FLAT_SCR) \ 683 CASE_CI_VI(FLAT_SCR_LO) \ 684 CASE_CI_VI(FLAT_SCR_HI) \ 685 CASE_VI_GFX9(TTMP0) \ 686 CASE_VI_GFX9(TTMP1) \ 687 CASE_VI_GFX9(TTMP2) \ 688 CASE_VI_GFX9(TTMP3) \ 689 CASE_VI_GFX9(TTMP4) \ 690 CASE_VI_GFX9(TTMP5) \ 691 CASE_VI_GFX9(TTMP6) \ 692 CASE_VI_GFX9(TTMP7) \ 693 CASE_VI_GFX9(TTMP8) \ 694 CASE_VI_GFX9(TTMP9) \ 695 CASE_VI_GFX9(TTMP10) \ 696 CASE_VI_GFX9(TTMP11) \ 697 CASE_VI_GFX9(TTMP12) \ 698 CASE_VI_GFX9(TTMP13) \ 699 CASE_VI_GFX9(TTMP14) \ 700 CASE_VI_GFX9(TTMP15) \ 701 CASE_VI_GFX9(TTMP0_TTMP1) \ 702 CASE_VI_GFX9(TTMP2_TTMP3) \ 703 CASE_VI_GFX9(TTMP4_TTMP5) \ 704 CASE_VI_GFX9(TTMP6_TTMP7) \ 705 CASE_VI_GFX9(TTMP8_TTMP9) \ 706 CASE_VI_GFX9(TTMP10_TTMP11) \ 707 CASE_VI_GFX9(TTMP12_TTMP13) \ 708 CASE_VI_GFX9(TTMP14_TTMP15) \ 709 CASE_VI_GFX9(TTMP0_TTMP1_TTMP2_TTMP3) \ 710 CASE_VI_GFX9(TTMP4_TTMP5_TTMP6_TTMP7) \ 711 CASE_VI_GFX9(TTMP8_TTMP9_TTMP10_TTMP11) \ 712 CASE_VI_GFX9(TTMP12_TTMP13_TTMP14_TTMP15) \ 713 CASE_VI_GFX9(TTMP0_TTMP1_TTMP2_TTMP3_TTMP4_TTMP5_TTMP6_TTMP7) \ 714 CASE_VI_GFX9(TTMP4_TTMP5_TTMP6_TTMP7_TTMP8_TTMP9_TTMP10_TTMP11) \ 715 CASE_VI_GFX9(TTMP8_TTMP9_TTMP10_TTMP11_TTMP12_TTMP13_TTMP14_TTMP15) \ 716 CASE_VI_GFX9(TTMP0_TTMP1_TTMP2_TTMP3_TTMP4_TTMP5_TTMP6_TTMP7_TTMP8_TTMP9_TTMP10_TTMP11_TTMP12_TTMP13_TTMP14_TTMP15) \ 717 } 718 719 #define CASE_CI_VI(node) \ 720 assert(!isSI(STI)); \ 721 case node: return isCI(STI) ? node##_ci : node##_vi; 722 723 #define CASE_VI_GFX9(node) \ 724 case node: return isGFX9(STI) ? node##_gfx9 : node##_vi; 725 726 unsigned getMCReg(unsigned Reg, const MCSubtargetInfo &STI) { 727 MAP_REG2REG 728 } 729 730 #undef CASE_CI_VI 731 #undef CASE_VI_GFX9 732 733 #define CASE_CI_VI(node) case node##_ci: case node##_vi: return node; 734 #define CASE_VI_GFX9(node) case node##_vi: case node##_gfx9: return node; 735 736 unsigned mc2PseudoReg(unsigned Reg) { 737 MAP_REG2REG 738 } 739 740 #undef CASE_CI_VI 741 #undef CASE_VI_GFX9 742 #undef MAP_REG2REG 743 744 bool isSISrcOperand(const MCInstrDesc &Desc, unsigned OpNo) { 745 assert(OpNo < Desc.NumOperands); 746 unsigned OpType = Desc.OpInfo[OpNo].OperandType; 747 return OpType >= AMDGPU::OPERAND_SRC_FIRST && 748 OpType <= AMDGPU::OPERAND_SRC_LAST; 749 } 750 751 bool isSISrcFPOperand(const MCInstrDesc &Desc, unsigned OpNo) { 752 assert(OpNo < Desc.NumOperands); 753 unsigned OpType = Desc.OpInfo[OpNo].OperandType; 754 switch (OpType) { 755 case AMDGPU::OPERAND_REG_IMM_FP32: 756 case AMDGPU::OPERAND_REG_IMM_FP64: 757 case AMDGPU::OPERAND_REG_IMM_FP16: 758 case AMDGPU::OPERAND_REG_INLINE_C_FP32: 759 case AMDGPU::OPERAND_REG_INLINE_C_FP64: 760 case AMDGPU::OPERAND_REG_INLINE_C_FP16: 761 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16: 762 return true; 763 default: 764 return false; 765 } 766 } 767 768 bool isSISrcInlinableOperand(const MCInstrDesc &Desc, unsigned OpNo) { 769 assert(OpNo < Desc.NumOperands); 770 unsigned OpType = Desc.OpInfo[OpNo].OperandType; 771 return OpType >= AMDGPU::OPERAND_REG_INLINE_C_FIRST && 772 OpType <= AMDGPU::OPERAND_REG_INLINE_C_LAST; 773 } 774 775 // Avoid using MCRegisterClass::getSize, since that function will go away 776 // (move from MC* level to Target* level). Return size in bits. 777 unsigned getRegBitWidth(unsigned RCID) { 778 switch (RCID) { 779 case AMDGPU::SGPR_32RegClassID: 780 case AMDGPU::VGPR_32RegClassID: 781 case AMDGPU::VS_32RegClassID: 782 case AMDGPU::SReg_32RegClassID: 783 case AMDGPU::SReg_32_XM0RegClassID: 784 return 32; 785 case AMDGPU::SGPR_64RegClassID: 786 case AMDGPU::VS_64RegClassID: 787 case AMDGPU::SReg_64RegClassID: 788 case AMDGPU::VReg_64RegClassID: 789 return 64; 790 case AMDGPU::VReg_96RegClassID: 791 return 96; 792 case AMDGPU::SGPR_128RegClassID: 793 case AMDGPU::SReg_128RegClassID: 794 case AMDGPU::VReg_128RegClassID: 795 return 128; 796 case AMDGPU::SReg_256RegClassID: 797 case AMDGPU::VReg_256RegClassID: 798 return 256; 799 case AMDGPU::SReg_512RegClassID: 800 case AMDGPU::VReg_512RegClassID: 801 return 512; 802 default: 803 llvm_unreachable("Unexpected register class"); 804 } 805 } 806 807 unsigned getRegBitWidth(const MCRegisterClass &RC) { 808 return getRegBitWidth(RC.getID()); 809 } 810 811 unsigned getRegOperandSize(const MCRegisterInfo *MRI, const MCInstrDesc &Desc, 812 unsigned OpNo) { 813 assert(OpNo < Desc.NumOperands); 814 unsigned RCID = Desc.OpInfo[OpNo].RegClass; 815 return getRegBitWidth(MRI->getRegClass(RCID)) / 8; 816 } 817 818 bool isInlinableLiteral64(int64_t Literal, bool HasInv2Pi) { 819 if (Literal >= -16 && Literal <= 64) 820 return true; 821 822 uint64_t Val = static_cast<uint64_t>(Literal); 823 return (Val == DoubleToBits(0.0)) || 824 (Val == DoubleToBits(1.0)) || 825 (Val == DoubleToBits(-1.0)) || 826 (Val == DoubleToBits(0.5)) || 827 (Val == DoubleToBits(-0.5)) || 828 (Val == DoubleToBits(2.0)) || 829 (Val == DoubleToBits(-2.0)) || 830 (Val == DoubleToBits(4.0)) || 831 (Val == DoubleToBits(-4.0)) || 832 (Val == 0x3fc45f306dc9c882 && HasInv2Pi); 833 } 834 835 bool isInlinableLiteral32(int32_t Literal, bool HasInv2Pi) { 836 if (Literal >= -16 && Literal <= 64) 837 return true; 838 839 // The actual type of the operand does not seem to matter as long 840 // as the bits match one of the inline immediate values. For example: 841 // 842 // -nan has the hexadecimal encoding of 0xfffffffe which is -2 in decimal, 843 // so it is a legal inline immediate. 844 // 845 // 1065353216 has the hexadecimal encoding 0x3f800000 which is 1.0f in 846 // floating-point, so it is a legal inline immediate. 847 848 uint32_t Val = static_cast<uint32_t>(Literal); 849 return (Val == FloatToBits(0.0f)) || 850 (Val == FloatToBits(1.0f)) || 851 (Val == FloatToBits(-1.0f)) || 852 (Val == FloatToBits(0.5f)) || 853 (Val == FloatToBits(-0.5f)) || 854 (Val == FloatToBits(2.0f)) || 855 (Val == FloatToBits(-2.0f)) || 856 (Val == FloatToBits(4.0f)) || 857 (Val == FloatToBits(-4.0f)) || 858 (Val == 0x3e22f983 && HasInv2Pi); 859 } 860 861 bool isInlinableLiteral16(int16_t Literal, bool HasInv2Pi) { 862 if (!HasInv2Pi) 863 return false; 864 865 if (Literal >= -16 && Literal <= 64) 866 return true; 867 868 uint16_t Val = static_cast<uint16_t>(Literal); 869 return Val == 0x3C00 || // 1.0 870 Val == 0xBC00 || // -1.0 871 Val == 0x3800 || // 0.5 872 Val == 0xB800 || // -0.5 873 Val == 0x4000 || // 2.0 874 Val == 0xC000 || // -2.0 875 Val == 0x4400 || // 4.0 876 Val == 0xC400 || // -4.0 877 Val == 0x3118; // 1/2pi 878 } 879 880 bool isInlinableLiteralV216(int32_t Literal, bool HasInv2Pi) { 881 assert(HasInv2Pi); 882 883 int16_t Lo16 = static_cast<int16_t>(Literal); 884 int16_t Hi16 = static_cast<int16_t>(Literal >> 16); 885 return Lo16 == Hi16 && isInlinableLiteral16(Lo16, HasInv2Pi); 886 } 887 888 bool isArgPassedInSGPR(const Argument *A) { 889 const Function *F = A->getParent(); 890 891 // Arguments to compute shaders are never a source of divergence. 892 CallingConv::ID CC = F->getCallingConv(); 893 switch (CC) { 894 case CallingConv::AMDGPU_KERNEL: 895 case CallingConv::SPIR_KERNEL: 896 return true; 897 case CallingConv::AMDGPU_VS: 898 case CallingConv::AMDGPU_LS: 899 case CallingConv::AMDGPU_HS: 900 case CallingConv::AMDGPU_ES: 901 case CallingConv::AMDGPU_GS: 902 case CallingConv::AMDGPU_PS: 903 case CallingConv::AMDGPU_CS: 904 // For non-compute shaders, SGPR inputs are marked with either inreg or byval. 905 // Everything else is in VGPRs. 906 return F->getAttributes().hasParamAttribute(A->getArgNo(), Attribute::InReg) || 907 F->getAttributes().hasParamAttribute(A->getArgNo(), Attribute::ByVal); 908 default: 909 // TODO: Should calls support inreg for SGPR inputs? 910 return false; 911 } 912 } 913 914 int64_t getSMRDEncodedOffset(const MCSubtargetInfo &ST, int64_t ByteOffset) { 915 if (isGCN3Encoding(ST)) 916 return ByteOffset; 917 return ByteOffset >> 2; 918 } 919 920 bool isLegalSMRDImmOffset(const MCSubtargetInfo &ST, int64_t ByteOffset) { 921 int64_t EncodedOffset = getSMRDEncodedOffset(ST, ByteOffset); 922 return isGCN3Encoding(ST) ? 923 isUInt<20>(EncodedOffset) : isUInt<8>(EncodedOffset); 924 } 925 926 } // end namespace AMDGPU 927 928 } // end namespace llvm 929 930 namespace llvm { 931 namespace AMDGPU { 932 933 AMDGPUAS getAMDGPUAS(Triple T) { 934 AMDGPUAS AS; 935 AS.FLAT_ADDRESS = 0; 936 AS.PRIVATE_ADDRESS = 5; 937 AS.REGION_ADDRESS = 2; 938 return AS; 939 } 940 941 AMDGPUAS getAMDGPUAS(const TargetMachine &M) { 942 return getAMDGPUAS(M.getTargetTriple()); 943 } 944 945 AMDGPUAS getAMDGPUAS(const Module &M) { 946 return getAMDGPUAS(Triple(M.getTargetTriple())); 947 } 948 949 namespace { 950 951 struct SourceOfDivergence { 952 unsigned Intr; 953 }; 954 const SourceOfDivergence *lookupSourceOfDivergenceByIntr(unsigned Intr); 955 956 #define GET_SOURCEOFDIVERGENCE_IMPL 957 #include "AMDGPUGenSearchableTables.inc" 958 959 } // end anonymous namespace 960 961 bool isIntrinsicSourceOfDivergence(unsigned IntrID) { 962 return lookupSourceOfDivergenceByIntr(IntrID); 963 } 964 } // namespace AMDGPU 965 } // namespace llvm 966