1 //==- SIMachineFunctionInfo.h - SIMachineFunctionInfo interface --*- 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 /// \file 10 // 11 //===----------------------------------------------------------------------===// 12 13 #ifndef LLVM_LIB_TARGET_AMDGPU_SIMACHINEFUNCTIONINFO_H 14 #define LLVM_LIB_TARGET_AMDGPU_SIMACHINEFUNCTIONINFO_H 15 16 #include "AMDGPUArgumentUsageInfo.h" 17 #include "AMDGPUMachineFunction.h" 18 #include "MCTargetDesc/AMDGPUMCTargetDesc.h" 19 #include "SIInstrInfo.h" 20 #include "SIRegisterInfo.h" 21 #include "llvm/ADT/ArrayRef.h" 22 #include "llvm/ADT/DenseMap.h" 23 #include "llvm/ADT/Optional.h" 24 #include "llvm/ADT/STLExtras.h" 25 #include "llvm/ADT/SmallVector.h" 26 #include "llvm/ADT/SparseBitVector.h" 27 #include "llvm/CodeGen/MIRYamlMapping.h" 28 #include "llvm/CodeGen/PseudoSourceValue.h" 29 #include "llvm/CodeGen/TargetInstrInfo.h" 30 #include "llvm/MC/MCRegisterInfo.h" 31 #include "llvm/Support/ErrorHandling.h" 32 #include <array> 33 #include <cassert> 34 #include <utility> 35 #include <vector> 36 37 namespace llvm { 38 39 class MachineFrameInfo; 40 class MachineFunction; 41 class TargetRegisterClass; 42 43 class AMDGPUPseudoSourceValue : public PseudoSourceValue { 44 public: 45 enum AMDGPUPSVKind : unsigned { 46 PSVBuffer = PseudoSourceValue::TargetCustom, 47 PSVImage, 48 GWSResource 49 }; 50 51 protected: AMDGPUPseudoSourceValue(unsigned Kind,const TargetInstrInfo & TII)52 AMDGPUPseudoSourceValue(unsigned Kind, const TargetInstrInfo &TII) 53 : PseudoSourceValue(Kind, TII) {} 54 55 public: isConstant(const MachineFrameInfo *)56 bool isConstant(const MachineFrameInfo *) const override { 57 // This should probably be true for most images, but we will start by being 58 // conservative. 59 return false; 60 } 61 isAliased(const MachineFrameInfo *)62 bool isAliased(const MachineFrameInfo *) const override { 63 return true; 64 } 65 mayAlias(const MachineFrameInfo *)66 bool mayAlias(const MachineFrameInfo *) const override { 67 return true; 68 } 69 }; 70 71 class AMDGPUBufferPseudoSourceValue final : public AMDGPUPseudoSourceValue { 72 public: AMDGPUBufferPseudoSourceValue(const TargetInstrInfo & TII)73 explicit AMDGPUBufferPseudoSourceValue(const TargetInstrInfo &TII) 74 : AMDGPUPseudoSourceValue(PSVBuffer, TII) {} 75 classof(const PseudoSourceValue * V)76 static bool classof(const PseudoSourceValue *V) { 77 return V->kind() == PSVBuffer; 78 } 79 }; 80 81 class AMDGPUImagePseudoSourceValue final : public AMDGPUPseudoSourceValue { 82 public: 83 // TODO: Is the img rsrc useful? AMDGPUImagePseudoSourceValue(const TargetInstrInfo & TII)84 explicit AMDGPUImagePseudoSourceValue(const TargetInstrInfo &TII) 85 : AMDGPUPseudoSourceValue(PSVImage, TII) {} 86 classof(const PseudoSourceValue * V)87 static bool classof(const PseudoSourceValue *V) { 88 return V->kind() == PSVImage; 89 } 90 }; 91 92 class AMDGPUGWSResourcePseudoSourceValue final : public AMDGPUPseudoSourceValue { 93 public: AMDGPUGWSResourcePseudoSourceValue(const TargetInstrInfo & TII)94 explicit AMDGPUGWSResourcePseudoSourceValue(const TargetInstrInfo &TII) 95 : AMDGPUPseudoSourceValue(GWSResource, TII) {} 96 classof(const PseudoSourceValue * V)97 static bool classof(const PseudoSourceValue *V) { 98 return V->kind() == GWSResource; 99 } 100 101 // These are inaccessible memory from IR. isAliased(const MachineFrameInfo *)102 bool isAliased(const MachineFrameInfo *) const override { 103 return false; 104 } 105 106 // These are inaccessible memory from IR. mayAlias(const MachineFrameInfo *)107 bool mayAlias(const MachineFrameInfo *) const override { 108 return false; 109 } 110 printCustom(raw_ostream & OS)111 void printCustom(raw_ostream &OS) const override { 112 OS << "GWSResource"; 113 } 114 }; 115 116 namespace yaml { 117 118 struct SIArgument { 119 bool IsRegister; 120 union { 121 StringValue RegisterName; 122 unsigned StackOffset; 123 }; 124 Optional<unsigned> Mask; 125 126 // Default constructor, which creates a stack argument. SIArgumentSIArgument127 SIArgument() : IsRegister(false), StackOffset(0) {} SIArgumentSIArgument128 SIArgument(const SIArgument &Other) { 129 IsRegister = Other.IsRegister; 130 if (IsRegister) { 131 ::new ((void *)std::addressof(RegisterName)) 132 StringValue(Other.RegisterName); 133 } else 134 StackOffset = Other.StackOffset; 135 Mask = Other.Mask; 136 } 137 SIArgument &operator=(const SIArgument &Other) { 138 IsRegister = Other.IsRegister; 139 if (IsRegister) { 140 ::new ((void *)std::addressof(RegisterName)) 141 StringValue(Other.RegisterName); 142 } else 143 StackOffset = Other.StackOffset; 144 Mask = Other.Mask; 145 return *this; 146 } ~SIArgumentSIArgument147 ~SIArgument() { 148 if (IsRegister) 149 RegisterName.~StringValue(); 150 } 151 152 // Helper to create a register or stack argument. createArgumentSIArgument153 static inline SIArgument createArgument(bool IsReg) { 154 if (IsReg) 155 return SIArgument(IsReg); 156 return SIArgument(); 157 } 158 159 private: 160 // Construct a register argument. SIArgumentSIArgument161 SIArgument(bool) : IsRegister(true), RegisterName() {} 162 }; 163 164 template <> struct MappingTraits<SIArgument> { 165 static void mapping(IO &YamlIO, SIArgument &A) { 166 if (YamlIO.outputting()) { 167 if (A.IsRegister) 168 YamlIO.mapRequired("reg", A.RegisterName); 169 else 170 YamlIO.mapRequired("offset", A.StackOffset); 171 } else { 172 auto Keys = YamlIO.keys(); 173 if (is_contained(Keys, "reg")) { 174 A = SIArgument::createArgument(true); 175 YamlIO.mapRequired("reg", A.RegisterName); 176 } else if (is_contained(Keys, "offset")) 177 YamlIO.mapRequired("offset", A.StackOffset); 178 else 179 YamlIO.setError("missing required key 'reg' or 'offset'"); 180 } 181 YamlIO.mapOptional("mask", A.Mask); 182 } 183 static const bool flow = true; 184 }; 185 186 struct SIArgumentInfo { 187 Optional<SIArgument> PrivateSegmentBuffer; 188 Optional<SIArgument> DispatchPtr; 189 Optional<SIArgument> QueuePtr; 190 Optional<SIArgument> KernargSegmentPtr; 191 Optional<SIArgument> DispatchID; 192 Optional<SIArgument> FlatScratchInit; 193 Optional<SIArgument> PrivateSegmentSize; 194 195 Optional<SIArgument> WorkGroupIDX; 196 Optional<SIArgument> WorkGroupIDY; 197 Optional<SIArgument> WorkGroupIDZ; 198 Optional<SIArgument> WorkGroupInfo; 199 Optional<SIArgument> PrivateSegmentWaveByteOffset; 200 201 Optional<SIArgument> ImplicitArgPtr; 202 Optional<SIArgument> ImplicitBufferPtr; 203 204 Optional<SIArgument> WorkItemIDX; 205 Optional<SIArgument> WorkItemIDY; 206 Optional<SIArgument> WorkItemIDZ; 207 }; 208 209 template <> struct MappingTraits<SIArgumentInfo> { 210 static void mapping(IO &YamlIO, SIArgumentInfo &AI) { 211 YamlIO.mapOptional("privateSegmentBuffer", AI.PrivateSegmentBuffer); 212 YamlIO.mapOptional("dispatchPtr", AI.DispatchPtr); 213 YamlIO.mapOptional("queuePtr", AI.QueuePtr); 214 YamlIO.mapOptional("kernargSegmentPtr", AI.KernargSegmentPtr); 215 YamlIO.mapOptional("dispatchID", AI.DispatchID); 216 YamlIO.mapOptional("flatScratchInit", AI.FlatScratchInit); 217 YamlIO.mapOptional("privateSegmentSize", AI.PrivateSegmentSize); 218 219 YamlIO.mapOptional("workGroupIDX", AI.WorkGroupIDX); 220 YamlIO.mapOptional("workGroupIDY", AI.WorkGroupIDY); 221 YamlIO.mapOptional("workGroupIDZ", AI.WorkGroupIDZ); 222 YamlIO.mapOptional("workGroupInfo", AI.WorkGroupInfo); 223 YamlIO.mapOptional("privateSegmentWaveByteOffset", 224 AI.PrivateSegmentWaveByteOffset); 225 226 YamlIO.mapOptional("implicitArgPtr", AI.ImplicitArgPtr); 227 YamlIO.mapOptional("implicitBufferPtr", AI.ImplicitBufferPtr); 228 229 YamlIO.mapOptional("workItemIDX", AI.WorkItemIDX); 230 YamlIO.mapOptional("workItemIDY", AI.WorkItemIDY); 231 YamlIO.mapOptional("workItemIDZ", AI.WorkItemIDZ); 232 } 233 }; 234 235 // Default to default mode for default calling convention. 236 struct SIMode { 237 bool IEEE = true; 238 bool DX10Clamp = true; 239 bool FP32Denormals = true; 240 bool FP64FP16Denormals = true; 241 242 SIMode() = default; 243 244 SIMode(const AMDGPU::SIModeRegisterDefaults &Mode) { 245 IEEE = Mode.IEEE; 246 DX10Clamp = Mode.DX10Clamp; 247 FP32Denormals = Mode.FP32Denormals; 248 FP64FP16Denormals = Mode.FP64FP16Denormals; 249 } 250 251 bool operator ==(const SIMode Other) const { 252 return IEEE == Other.IEEE && 253 DX10Clamp == Other.DX10Clamp && 254 FP32Denormals == Other.FP32Denormals && 255 FP64FP16Denormals == Other.FP64FP16Denormals; 256 } 257 }; 258 259 template <> struct MappingTraits<SIMode> { 260 static void mapping(IO &YamlIO, SIMode &Mode) { 261 YamlIO.mapOptional("ieee", Mode.IEEE, true); 262 YamlIO.mapOptional("dx10-clamp", Mode.DX10Clamp, true); 263 YamlIO.mapOptional("fp32-denormals", Mode.FP32Denormals, true); 264 YamlIO.mapOptional("fp64-fp16-denormals", Mode.FP64FP16Denormals, true); 265 } 266 }; 267 268 struct SIMachineFunctionInfo final : public yaml::MachineFunctionInfo { 269 uint64_t ExplicitKernArgSize = 0; 270 unsigned MaxKernArgAlign = 0; 271 unsigned LDSSize = 0; 272 bool IsEntryFunction = false; 273 bool NoSignedZerosFPMath = false; 274 bool MemoryBound = false; 275 bool WaveLimiter = false; 276 uint32_t HighBitsOf32BitAddress = 0; 277 278 StringValue ScratchRSrcReg = "$private_rsrc_reg"; 279 StringValue ScratchWaveOffsetReg = "$scratch_wave_offset_reg"; 280 StringValue FrameOffsetReg = "$fp_reg"; 281 StringValue StackPtrOffsetReg = "$sp_reg"; 282 283 Optional<SIArgumentInfo> ArgInfo; 284 SIMode Mode; 285 286 SIMachineFunctionInfo() = default; 287 SIMachineFunctionInfo(const llvm::SIMachineFunctionInfo &, 288 const TargetRegisterInfo &TRI); 289 290 void mappingImpl(yaml::IO &YamlIO) override; 291 ~SIMachineFunctionInfo() = default; 292 }; 293 294 template <> struct MappingTraits<SIMachineFunctionInfo> { 295 static void mapping(IO &YamlIO, SIMachineFunctionInfo &MFI) { 296 YamlIO.mapOptional("explicitKernArgSize", MFI.ExplicitKernArgSize, 297 UINT64_C(0)); 298 YamlIO.mapOptional("maxKernArgAlign", MFI.MaxKernArgAlign, 0u); 299 YamlIO.mapOptional("ldsSize", MFI.LDSSize, 0u); 300 YamlIO.mapOptional("isEntryFunction", MFI.IsEntryFunction, false); 301 YamlIO.mapOptional("noSignedZerosFPMath", MFI.NoSignedZerosFPMath, false); 302 YamlIO.mapOptional("memoryBound", MFI.MemoryBound, false); 303 YamlIO.mapOptional("waveLimiter", MFI.WaveLimiter, false); 304 YamlIO.mapOptional("scratchRSrcReg", MFI.ScratchRSrcReg, 305 StringValue("$private_rsrc_reg")); 306 YamlIO.mapOptional("scratchWaveOffsetReg", MFI.ScratchWaveOffsetReg, 307 StringValue("$scratch_wave_offset_reg")); 308 YamlIO.mapOptional("frameOffsetReg", MFI.FrameOffsetReg, 309 StringValue("$fp_reg")); 310 YamlIO.mapOptional("stackPtrOffsetReg", MFI.StackPtrOffsetReg, 311 StringValue("$sp_reg")); 312 YamlIO.mapOptional("argumentInfo", MFI.ArgInfo); 313 YamlIO.mapOptional("mode", MFI.Mode, SIMode()); 314 YamlIO.mapOptional("highBitsOf32BitAddress", 315 MFI.HighBitsOf32BitAddress, 0u); 316 } 317 }; 318 319 } // end namespace yaml 320 321 /// This class keeps track of the SPI_SP_INPUT_ADDR config register, which 322 /// tells the hardware which interpolation parameters to load. 323 class SIMachineFunctionInfo final : public AMDGPUMachineFunction { 324 friend class GCNTargetMachine; 325 326 unsigned TIDReg = AMDGPU::NoRegister; 327 328 // Registers that may be reserved for spilling purposes. These may be the same 329 // as the input registers. 330 unsigned ScratchRSrcReg = AMDGPU::PRIVATE_RSRC_REG; 331 unsigned ScratchWaveOffsetReg = AMDGPU::SCRATCH_WAVE_OFFSET_REG; 332 333 // This is the current function's incremented size from the kernel's scratch 334 // wave offset register. For an entry function, this is exactly the same as 335 // the ScratchWaveOffsetReg. 336 unsigned FrameOffsetReg = AMDGPU::FP_REG; 337 338 // Top of the stack SGPR offset derived from the ScratchWaveOffsetReg. 339 unsigned StackPtrOffsetReg = AMDGPU::SP_REG; 340 341 AMDGPUFunctionArgInfo ArgInfo; 342 343 // Graphics info. 344 unsigned PSInputAddr = 0; 345 unsigned PSInputEnable = 0; 346 347 /// Number of bytes of arguments this function has on the stack. If the callee 348 /// is expected to restore the argument stack this should be a multiple of 16, 349 /// all usable during a tail call. 350 /// 351 /// The alternative would forbid tail call optimisation in some cases: if we 352 /// want to transfer control from a function with 8-bytes of stack-argument 353 /// space to a function with 16-bytes then misalignment of this value would 354 /// make a stack adjustment necessary, which could not be undone by the 355 /// callee. 356 unsigned BytesInStackArgArea = 0; 357 358 bool ReturnsVoid = true; 359 360 // A pair of default/requested minimum/maximum flat work group sizes. 361 // Minimum - first, maximum - second. 362 std::pair<unsigned, unsigned> FlatWorkGroupSizes = {0, 0}; 363 364 // A pair of default/requested minimum/maximum number of waves per execution 365 // unit. Minimum - first, maximum - second. 366 std::pair<unsigned, unsigned> WavesPerEU = {0, 0}; 367 368 DenseMap<const Value *, 369 std::unique_ptr<const AMDGPUBufferPseudoSourceValue>> BufferPSVs; 370 DenseMap<const Value *, 371 std::unique_ptr<const AMDGPUImagePseudoSourceValue>> ImagePSVs; 372 std::unique_ptr<const AMDGPUGWSResourcePseudoSourceValue> GWSResourcePSV; 373 374 private: 375 unsigned LDSWaveSpillSize = 0; 376 unsigned NumUserSGPRs = 0; 377 unsigned NumSystemSGPRs = 0; 378 379 bool HasSpilledSGPRs = false; 380 bool HasSpilledVGPRs = false; 381 bool HasNonSpillStackObjects = false; 382 bool IsStackRealigned = false; 383 384 unsigned NumSpilledSGPRs = 0; 385 unsigned NumSpilledVGPRs = 0; 386 387 // Feature bits required for inputs passed in user SGPRs. 388 bool PrivateSegmentBuffer : 1; 389 bool DispatchPtr : 1; 390 bool QueuePtr : 1; 391 bool KernargSegmentPtr : 1; 392 bool DispatchID : 1; 393 bool FlatScratchInit : 1; 394 395 // Feature bits required for inputs passed in system SGPRs. 396 bool WorkGroupIDX : 1; // Always initialized. 397 bool WorkGroupIDY : 1; 398 bool WorkGroupIDZ : 1; 399 bool WorkGroupInfo : 1; 400 bool PrivateSegmentWaveByteOffset : 1; 401 402 bool WorkItemIDX : 1; // Always initialized. 403 bool WorkItemIDY : 1; 404 bool WorkItemIDZ : 1; 405 406 // Private memory buffer 407 // Compute directly in sgpr[0:1] 408 // Other shaders indirect 64-bits at sgpr[0:1] 409 bool ImplicitBufferPtr : 1; 410 411 // Pointer to where the ABI inserts special kernel arguments separate from the 412 // user arguments. This is an offset from the KernargSegmentPtr. 413 bool ImplicitArgPtr : 1; 414 415 // The hard-wired high half of the address of the global information table 416 // for AMDPAL OS type. 0xffffffff represents no hard-wired high half, since 417 // current hardware only allows a 16 bit value. 418 unsigned GITPtrHigh; 419 420 unsigned HighBitsOf32BitAddress; 421 unsigned GDSSize; 422 423 // Current recorded maximum possible occupancy. 424 unsigned Occupancy; 425 426 MCPhysReg getNextUserSGPR() const; 427 428 MCPhysReg getNextSystemSGPR() const; 429 430 public: 431 struct SpilledReg { 432 unsigned VGPR = 0; 433 int Lane = -1; 434 435 SpilledReg() = default; 436 SpilledReg(unsigned R, int L) : VGPR (R), Lane (L) {} 437 438 bool hasLane() { return Lane != -1;} 439 bool hasReg() { return VGPR != 0;} 440 }; 441 442 struct SGPRSpillVGPRCSR { 443 // VGPR used for SGPR spills 444 unsigned VGPR; 445 446 // If the VGPR is a CSR, the stack slot used to save/restore it in the 447 // prolog/epilog. 448 Optional<int> FI; 449 450 SGPRSpillVGPRCSR(unsigned V, Optional<int> F) : VGPR(V), FI(F) {} 451 }; 452 453 struct VGPRSpillToAGPR { 454 SmallVector<MCPhysReg, 32> Lanes; 455 bool FullyAllocated = false; 456 }; 457 458 SparseBitVector<> WWMReservedRegs; 459 460 void ReserveWWMRegister(unsigned reg) { WWMReservedRegs.set(reg); } 461 462 private: 463 // SGPR->VGPR spilling support. 464 using SpillRegMask = std::pair<unsigned, unsigned>; 465 466 // Track VGPR + wave index for each subregister of the SGPR spilled to 467 // frameindex key. 468 DenseMap<int, std::vector<SpilledReg>> SGPRToVGPRSpills; 469 unsigned NumVGPRSpillLanes = 0; 470 SmallVector<SGPRSpillVGPRCSR, 2> SpillVGPRs; 471 472 DenseMap<int, VGPRSpillToAGPR> VGPRToAGPRSpills; 473 474 // AGPRs used for VGPR spills. 475 SmallVector<MCPhysReg, 32> SpillAGPR; 476 477 // VGPRs used for AGPR spills. 478 SmallVector<MCPhysReg, 32> SpillVGPR; 479 480 public: // FIXME 481 /// If this is set, an SGPR used for save/restore of the register used for the 482 /// frame pointer. 483 unsigned SGPRForFPSaveRestoreCopy = 0; 484 Optional<int> FramePointerSaveIndex; 485 486 public: 487 SIMachineFunctionInfo(const MachineFunction &MF); 488 489 bool initializeBaseYamlFields(const yaml::SIMachineFunctionInfo &YamlMFI); 490 491 ArrayRef<SpilledReg> getSGPRToVGPRSpills(int FrameIndex) const { 492 auto I = SGPRToVGPRSpills.find(FrameIndex); 493 return (I == SGPRToVGPRSpills.end()) ? 494 ArrayRef<SpilledReg>() : makeArrayRef(I->second); 495 } 496 497 ArrayRef<SGPRSpillVGPRCSR> getSGPRSpillVGPRs() const { 498 return SpillVGPRs; 499 } 500 501 ArrayRef<MCPhysReg> getAGPRSpillVGPRs() const { 502 return SpillAGPR; 503 } 504 505 ArrayRef<MCPhysReg> getVGPRSpillAGPRs() const { 506 return SpillVGPR; 507 } 508 509 MCPhysReg getVGPRToAGPRSpill(int FrameIndex, unsigned Lane) const { 510 auto I = VGPRToAGPRSpills.find(FrameIndex); 511 return (I == VGPRToAGPRSpills.end()) ? (MCPhysReg)AMDGPU::NoRegister 512 : I->second.Lanes[Lane]; 513 } 514 515 bool haveFreeLanesForSGPRSpill(const MachineFunction &MF, 516 unsigned NumLane) const; 517 bool allocateSGPRSpillToVGPR(MachineFunction &MF, int FI); 518 bool allocateVGPRSpillToAGPR(MachineFunction &MF, int FI, bool isAGPRtoVGPR); 519 void removeDeadFrameIndices(MachineFrameInfo &MFI); 520 521 bool hasCalculatedTID() const { return TIDReg != 0; }; 522 unsigned getTIDReg() const { return TIDReg; }; 523 void setTIDReg(unsigned Reg) { TIDReg = Reg; } 524 525 unsigned getBytesInStackArgArea() const { 526 return BytesInStackArgArea; 527 } 528 529 void setBytesInStackArgArea(unsigned Bytes) { 530 BytesInStackArgArea = Bytes; 531 } 532 533 // Add user SGPRs. 534 unsigned addPrivateSegmentBuffer(const SIRegisterInfo &TRI); 535 unsigned addDispatchPtr(const SIRegisterInfo &TRI); 536 unsigned addQueuePtr(const SIRegisterInfo &TRI); 537 unsigned addKernargSegmentPtr(const SIRegisterInfo &TRI); 538 unsigned addDispatchID(const SIRegisterInfo &TRI); 539 unsigned addFlatScratchInit(const SIRegisterInfo &TRI); 540 unsigned addImplicitBufferPtr(const SIRegisterInfo &TRI); 541 542 // Add system SGPRs. 543 unsigned addWorkGroupIDX() { 544 ArgInfo.WorkGroupIDX = ArgDescriptor::createRegister(getNextSystemSGPR()); 545 NumSystemSGPRs += 1; 546 return ArgInfo.WorkGroupIDX.getRegister(); 547 } 548 549 unsigned addWorkGroupIDY() { 550 ArgInfo.WorkGroupIDY = ArgDescriptor::createRegister(getNextSystemSGPR()); 551 NumSystemSGPRs += 1; 552 return ArgInfo.WorkGroupIDY.getRegister(); 553 } 554 555 unsigned addWorkGroupIDZ() { 556 ArgInfo.WorkGroupIDZ = ArgDescriptor::createRegister(getNextSystemSGPR()); 557 NumSystemSGPRs += 1; 558 return ArgInfo.WorkGroupIDZ.getRegister(); 559 } 560 561 unsigned addWorkGroupInfo() { 562 ArgInfo.WorkGroupInfo = ArgDescriptor::createRegister(getNextSystemSGPR()); 563 NumSystemSGPRs += 1; 564 return ArgInfo.WorkGroupInfo.getRegister(); 565 } 566 567 // Add special VGPR inputs 568 void setWorkItemIDX(ArgDescriptor Arg) { 569 ArgInfo.WorkItemIDX = Arg; 570 } 571 572 void setWorkItemIDY(ArgDescriptor Arg) { 573 ArgInfo.WorkItemIDY = Arg; 574 } 575 576 void setWorkItemIDZ(ArgDescriptor Arg) { 577 ArgInfo.WorkItemIDZ = Arg; 578 } 579 580 unsigned addPrivateSegmentWaveByteOffset() { 581 ArgInfo.PrivateSegmentWaveByteOffset 582 = ArgDescriptor::createRegister(getNextSystemSGPR()); 583 NumSystemSGPRs += 1; 584 return ArgInfo.PrivateSegmentWaveByteOffset.getRegister(); 585 } 586 587 void setPrivateSegmentWaveByteOffset(unsigned Reg) { 588 ArgInfo.PrivateSegmentWaveByteOffset = ArgDescriptor::createRegister(Reg); 589 } 590 591 bool hasPrivateSegmentBuffer() const { 592 return PrivateSegmentBuffer; 593 } 594 595 bool hasDispatchPtr() const { 596 return DispatchPtr; 597 } 598 599 bool hasQueuePtr() const { 600 return QueuePtr; 601 } 602 603 bool hasKernargSegmentPtr() const { 604 return KernargSegmentPtr; 605 } 606 607 bool hasDispatchID() const { 608 return DispatchID; 609 } 610 611 bool hasFlatScratchInit() const { 612 return FlatScratchInit; 613 } 614 615 bool hasWorkGroupIDX() const { 616 return WorkGroupIDX; 617 } 618 619 bool hasWorkGroupIDY() const { 620 return WorkGroupIDY; 621 } 622 623 bool hasWorkGroupIDZ() const { 624 return WorkGroupIDZ; 625 } 626 627 bool hasWorkGroupInfo() const { 628 return WorkGroupInfo; 629 } 630 631 bool hasPrivateSegmentWaveByteOffset() const { 632 return PrivateSegmentWaveByteOffset; 633 } 634 635 bool hasWorkItemIDX() const { 636 return WorkItemIDX; 637 } 638 639 bool hasWorkItemIDY() const { 640 return WorkItemIDY; 641 } 642 643 bool hasWorkItemIDZ() const { 644 return WorkItemIDZ; 645 } 646 647 bool hasImplicitArgPtr() const { 648 return ImplicitArgPtr; 649 } 650 651 bool hasImplicitBufferPtr() const { 652 return ImplicitBufferPtr; 653 } 654 655 AMDGPUFunctionArgInfo &getArgInfo() { 656 return ArgInfo; 657 } 658 659 const AMDGPUFunctionArgInfo &getArgInfo() const { 660 return ArgInfo; 661 } 662 663 std::pair<const ArgDescriptor *, const TargetRegisterClass *> 664 getPreloadedValue(AMDGPUFunctionArgInfo::PreloadedValue Value) const { 665 return ArgInfo.getPreloadedValue(Value); 666 } 667 668 Register getPreloadedReg(AMDGPUFunctionArgInfo::PreloadedValue Value) const { 669 auto Arg = ArgInfo.getPreloadedValue(Value).first; 670 return Arg ? Arg->getRegister() : Register(); 671 } 672 673 unsigned getGITPtrHigh() const { 674 return GITPtrHigh; 675 } 676 677 uint32_t get32BitAddressHighBits() const { 678 return HighBitsOf32BitAddress; 679 } 680 681 unsigned getGDSSize() const { 682 return GDSSize; 683 } 684 685 unsigned getNumUserSGPRs() const { 686 return NumUserSGPRs; 687 } 688 689 unsigned getNumPreloadedSGPRs() const { 690 return NumUserSGPRs + NumSystemSGPRs; 691 } 692 693 unsigned getPrivateSegmentWaveByteOffsetSystemSGPR() const { 694 return ArgInfo.PrivateSegmentWaveByteOffset.getRegister(); 695 } 696 697 /// Returns the physical register reserved for use as the resource 698 /// descriptor for scratch accesses. 699 unsigned getScratchRSrcReg() const { 700 return ScratchRSrcReg; 701 } 702 703 void setScratchRSrcReg(unsigned Reg) { 704 assert(Reg != 0 && "Should never be unset"); 705 ScratchRSrcReg = Reg; 706 } 707 708 unsigned getScratchWaveOffsetReg() const { 709 return ScratchWaveOffsetReg; 710 } 711 712 unsigned getFrameOffsetReg() const { 713 return FrameOffsetReg; 714 } 715 716 void setFrameOffsetReg(unsigned Reg) { 717 assert(Reg != 0 && "Should never be unset"); 718 FrameOffsetReg = Reg; 719 } 720 721 void setStackPtrOffsetReg(unsigned Reg) { 722 assert(Reg != 0 && "Should never be unset"); 723 StackPtrOffsetReg = Reg; 724 } 725 726 // Note the unset value for this is AMDGPU::SP_REG rather than 727 // NoRegister. This is mostly a workaround for MIR tests where state that 728 // can't be directly computed from the function is not preserved in serialized 729 // MIR. 730 unsigned getStackPtrOffsetReg() const { 731 return StackPtrOffsetReg; 732 } 733 734 void setScratchWaveOffsetReg(unsigned Reg) { 735 assert(Reg != 0 && "Should never be unset"); 736 ScratchWaveOffsetReg = Reg; 737 } 738 739 unsigned getQueuePtrUserSGPR() const { 740 return ArgInfo.QueuePtr.getRegister(); 741 } 742 743 unsigned getImplicitBufferPtrUserSGPR() const { 744 return ArgInfo.ImplicitBufferPtr.getRegister(); 745 } 746 747 bool hasSpilledSGPRs() const { 748 return HasSpilledSGPRs; 749 } 750 751 void setHasSpilledSGPRs(bool Spill = true) { 752 HasSpilledSGPRs = Spill; 753 } 754 755 bool hasSpilledVGPRs() const { 756 return HasSpilledVGPRs; 757 } 758 759 void setHasSpilledVGPRs(bool Spill = true) { 760 HasSpilledVGPRs = Spill; 761 } 762 763 bool hasNonSpillStackObjects() const { 764 return HasNonSpillStackObjects; 765 } 766 767 void setHasNonSpillStackObjects(bool StackObject = true) { 768 HasNonSpillStackObjects = StackObject; 769 } 770 771 bool isStackRealigned() const { 772 return IsStackRealigned; 773 } 774 775 void setIsStackRealigned(bool Realigned = true) { 776 IsStackRealigned = Realigned; 777 } 778 779 unsigned getNumSpilledSGPRs() const { 780 return NumSpilledSGPRs; 781 } 782 783 unsigned getNumSpilledVGPRs() const { 784 return NumSpilledVGPRs; 785 } 786 787 void addToSpilledSGPRs(unsigned num) { 788 NumSpilledSGPRs += num; 789 } 790 791 void addToSpilledVGPRs(unsigned num) { 792 NumSpilledVGPRs += num; 793 } 794 795 unsigned getPSInputAddr() const { 796 return PSInputAddr; 797 } 798 799 unsigned getPSInputEnable() const { 800 return PSInputEnable; 801 } 802 803 bool isPSInputAllocated(unsigned Index) const { 804 return PSInputAddr & (1 << Index); 805 } 806 807 void markPSInputAllocated(unsigned Index) { 808 PSInputAddr |= 1 << Index; 809 } 810 811 void markPSInputEnabled(unsigned Index) { 812 PSInputEnable |= 1 << Index; 813 } 814 815 bool returnsVoid() const { 816 return ReturnsVoid; 817 } 818 819 void setIfReturnsVoid(bool Value) { 820 ReturnsVoid = Value; 821 } 822 823 /// \returns A pair of default/requested minimum/maximum flat work group sizes 824 /// for this function. 825 std::pair<unsigned, unsigned> getFlatWorkGroupSizes() const { 826 return FlatWorkGroupSizes; 827 } 828 829 /// \returns Default/requested minimum flat work group size for this function. 830 unsigned getMinFlatWorkGroupSize() const { 831 return FlatWorkGroupSizes.first; 832 } 833 834 /// \returns Default/requested maximum flat work group size for this function. 835 unsigned getMaxFlatWorkGroupSize() const { 836 return FlatWorkGroupSizes.second; 837 } 838 839 /// \returns A pair of default/requested minimum/maximum number of waves per 840 /// execution unit. 841 std::pair<unsigned, unsigned> getWavesPerEU() const { 842 return WavesPerEU; 843 } 844 845 /// \returns Default/requested minimum number of waves per execution unit. 846 unsigned getMinWavesPerEU() const { 847 return WavesPerEU.first; 848 } 849 850 /// \returns Default/requested maximum number of waves per execution unit. 851 unsigned getMaxWavesPerEU() const { 852 return WavesPerEU.second; 853 } 854 855 /// \returns SGPR used for \p Dim's work group ID. 856 unsigned getWorkGroupIDSGPR(unsigned Dim) const { 857 switch (Dim) { 858 case 0: 859 assert(hasWorkGroupIDX()); 860 return ArgInfo.WorkGroupIDX.getRegister(); 861 case 1: 862 assert(hasWorkGroupIDY()); 863 return ArgInfo.WorkGroupIDY.getRegister(); 864 case 2: 865 assert(hasWorkGroupIDZ()); 866 return ArgInfo.WorkGroupIDZ.getRegister(); 867 } 868 llvm_unreachable("unexpected dimension"); 869 } 870 871 unsigned getLDSWaveSpillSize() const { 872 return LDSWaveSpillSize; 873 } 874 875 const AMDGPUBufferPseudoSourceValue *getBufferPSV(const SIInstrInfo &TII, 876 const Value *BufferRsrc) { 877 assert(BufferRsrc); 878 auto PSV = BufferPSVs.try_emplace( 879 BufferRsrc, 880 std::make_unique<AMDGPUBufferPseudoSourceValue>(TII)); 881 return PSV.first->second.get(); 882 } 883 884 const AMDGPUImagePseudoSourceValue *getImagePSV(const SIInstrInfo &TII, 885 const Value *ImgRsrc) { 886 assert(ImgRsrc); 887 auto PSV = ImagePSVs.try_emplace( 888 ImgRsrc, 889 std::make_unique<AMDGPUImagePseudoSourceValue>(TII)); 890 return PSV.first->second.get(); 891 } 892 893 const AMDGPUGWSResourcePseudoSourceValue *getGWSPSV(const SIInstrInfo &TII) { 894 if (!GWSResourcePSV) { 895 GWSResourcePSV = 896 std::make_unique<AMDGPUGWSResourcePseudoSourceValue>(TII); 897 } 898 899 return GWSResourcePSV.get(); 900 } 901 902 unsigned getOccupancy() const { 903 return Occupancy; 904 } 905 906 unsigned getMinAllowedOccupancy() const { 907 if (!isMemoryBound() && !needsWaveLimiter()) 908 return Occupancy; 909 return (Occupancy < 4) ? Occupancy : 4; 910 } 911 912 void limitOccupancy(const MachineFunction &MF); 913 914 void limitOccupancy(unsigned Limit) { 915 if (Occupancy > Limit) 916 Occupancy = Limit; 917 } 918 919 void increaseOccupancy(const MachineFunction &MF, unsigned Limit) { 920 if (Occupancy < Limit) 921 Occupancy = Limit; 922 limitOccupancy(MF); 923 } 924 }; 925 926 } // end namespace llvm 927 928 #endif // LLVM_LIB_TARGET_AMDGPU_SIMACHINEFUNCTIONINFO_H 929