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 "llvm/ADT/MapVector.h" 21 #include "llvm/CodeGen/MIRYamlMapping.h" 22 #include "llvm/CodeGen/PseudoSourceValue.h" 23 #include "llvm/Support/raw_ostream.h" 24 25 namespace llvm { 26 27 class MachineFrameInfo; 28 class MachineFunction; 29 class SIMachineFunctionInfo; 30 class SIRegisterInfo; 31 class TargetRegisterClass; 32 33 class AMDGPUPseudoSourceValue : public PseudoSourceValue { 34 public: 35 enum AMDGPUPSVKind : unsigned { 36 PSVBuffer = PseudoSourceValue::TargetCustom, 37 PSVImage, 38 GWSResource 39 }; 40 41 protected: 42 AMDGPUPseudoSourceValue(unsigned Kind, const TargetInstrInfo &TII) 43 : PseudoSourceValue(Kind, TII) {} 44 45 public: 46 bool isConstant(const MachineFrameInfo *) const override { 47 // This should probably be true for most images, but we will start by being 48 // conservative. 49 return false; 50 } 51 52 bool isAliased(const MachineFrameInfo *) const override { 53 return true; 54 } 55 56 bool mayAlias(const MachineFrameInfo *) const override { 57 return true; 58 } 59 }; 60 61 class AMDGPUBufferPseudoSourceValue final : public AMDGPUPseudoSourceValue { 62 public: 63 explicit AMDGPUBufferPseudoSourceValue(const TargetInstrInfo &TII) 64 : AMDGPUPseudoSourceValue(PSVBuffer, TII) {} 65 66 static bool classof(const PseudoSourceValue *V) { 67 return V->kind() == PSVBuffer; 68 } 69 70 void printCustom(raw_ostream &OS) const override { OS << "BufferResource"; } 71 }; 72 73 class AMDGPUImagePseudoSourceValue final : public AMDGPUPseudoSourceValue { 74 public: 75 // TODO: Is the img rsrc useful? 76 explicit AMDGPUImagePseudoSourceValue(const TargetInstrInfo &TII) 77 : AMDGPUPseudoSourceValue(PSVImage, TII) {} 78 79 static bool classof(const PseudoSourceValue *V) { 80 return V->kind() == PSVImage; 81 } 82 83 void printCustom(raw_ostream &OS) const override { OS << "ImageResource"; } 84 }; 85 86 class AMDGPUGWSResourcePseudoSourceValue final : public AMDGPUPseudoSourceValue { 87 public: 88 explicit AMDGPUGWSResourcePseudoSourceValue(const TargetInstrInfo &TII) 89 : AMDGPUPseudoSourceValue(GWSResource, TII) {} 90 91 static bool classof(const PseudoSourceValue *V) { 92 return V->kind() == GWSResource; 93 } 94 95 // These are inaccessible memory from IR. 96 bool isAliased(const MachineFrameInfo *) const override { 97 return false; 98 } 99 100 // These are inaccessible memory from IR. 101 bool mayAlias(const MachineFrameInfo *) const override { 102 return false; 103 } 104 105 void printCustom(raw_ostream &OS) const override { 106 OS << "GWSResource"; 107 } 108 }; 109 110 namespace yaml { 111 112 struct SIArgument { 113 bool IsRegister; 114 union { 115 StringValue RegisterName; 116 unsigned StackOffset; 117 }; 118 Optional<unsigned> Mask; 119 120 // Default constructor, which creates a stack argument. 121 SIArgument() : IsRegister(false), StackOffset(0) {} 122 SIArgument(const SIArgument &Other) { 123 IsRegister = Other.IsRegister; 124 if (IsRegister) { 125 ::new ((void *)std::addressof(RegisterName)) 126 StringValue(Other.RegisterName); 127 } else 128 StackOffset = Other.StackOffset; 129 Mask = Other.Mask; 130 } 131 SIArgument &operator=(const SIArgument &Other) { 132 IsRegister = Other.IsRegister; 133 if (IsRegister) { 134 ::new ((void *)std::addressof(RegisterName)) 135 StringValue(Other.RegisterName); 136 } else 137 StackOffset = Other.StackOffset; 138 Mask = Other.Mask; 139 return *this; 140 } 141 ~SIArgument() { 142 if (IsRegister) 143 RegisterName.~StringValue(); 144 } 145 146 // Helper to create a register or stack argument. 147 static inline SIArgument createArgument(bool IsReg) { 148 if (IsReg) 149 return SIArgument(IsReg); 150 return SIArgument(); 151 } 152 153 private: 154 // Construct a register argument. 155 SIArgument(bool) : IsRegister(true), RegisterName() {} 156 }; 157 158 template <> struct MappingTraits<SIArgument> { 159 static void mapping(IO &YamlIO, SIArgument &A) { 160 if (YamlIO.outputting()) { 161 if (A.IsRegister) 162 YamlIO.mapRequired("reg", A.RegisterName); 163 else 164 YamlIO.mapRequired("offset", A.StackOffset); 165 } else { 166 auto Keys = YamlIO.keys(); 167 if (is_contained(Keys, "reg")) { 168 A = SIArgument::createArgument(true); 169 YamlIO.mapRequired("reg", A.RegisterName); 170 } else if (is_contained(Keys, "offset")) 171 YamlIO.mapRequired("offset", A.StackOffset); 172 else 173 YamlIO.setError("missing required key 'reg' or 'offset'"); 174 } 175 YamlIO.mapOptional("mask", A.Mask); 176 } 177 static const bool flow = true; 178 }; 179 180 struct SIArgumentInfo { 181 Optional<SIArgument> PrivateSegmentBuffer; 182 Optional<SIArgument> DispatchPtr; 183 Optional<SIArgument> QueuePtr; 184 Optional<SIArgument> KernargSegmentPtr; 185 Optional<SIArgument> DispatchID; 186 Optional<SIArgument> FlatScratchInit; 187 Optional<SIArgument> PrivateSegmentSize; 188 189 Optional<SIArgument> WorkGroupIDX; 190 Optional<SIArgument> WorkGroupIDY; 191 Optional<SIArgument> WorkGroupIDZ; 192 Optional<SIArgument> WorkGroupInfo; 193 Optional<SIArgument> PrivateSegmentWaveByteOffset; 194 195 Optional<SIArgument> ImplicitArgPtr; 196 Optional<SIArgument> ImplicitBufferPtr; 197 198 Optional<SIArgument> WorkItemIDX; 199 Optional<SIArgument> WorkItemIDY; 200 Optional<SIArgument> WorkItemIDZ; 201 }; 202 203 template <> struct MappingTraits<SIArgumentInfo> { 204 static void mapping(IO &YamlIO, SIArgumentInfo &AI) { 205 YamlIO.mapOptional("privateSegmentBuffer", AI.PrivateSegmentBuffer); 206 YamlIO.mapOptional("dispatchPtr", AI.DispatchPtr); 207 YamlIO.mapOptional("queuePtr", AI.QueuePtr); 208 YamlIO.mapOptional("kernargSegmentPtr", AI.KernargSegmentPtr); 209 YamlIO.mapOptional("dispatchID", AI.DispatchID); 210 YamlIO.mapOptional("flatScratchInit", AI.FlatScratchInit); 211 YamlIO.mapOptional("privateSegmentSize", AI.PrivateSegmentSize); 212 213 YamlIO.mapOptional("workGroupIDX", AI.WorkGroupIDX); 214 YamlIO.mapOptional("workGroupIDY", AI.WorkGroupIDY); 215 YamlIO.mapOptional("workGroupIDZ", AI.WorkGroupIDZ); 216 YamlIO.mapOptional("workGroupInfo", AI.WorkGroupInfo); 217 YamlIO.mapOptional("privateSegmentWaveByteOffset", 218 AI.PrivateSegmentWaveByteOffset); 219 220 YamlIO.mapOptional("implicitArgPtr", AI.ImplicitArgPtr); 221 YamlIO.mapOptional("implicitBufferPtr", AI.ImplicitBufferPtr); 222 223 YamlIO.mapOptional("workItemIDX", AI.WorkItemIDX); 224 YamlIO.mapOptional("workItemIDY", AI.WorkItemIDY); 225 YamlIO.mapOptional("workItemIDZ", AI.WorkItemIDZ); 226 } 227 }; 228 229 // Default to default mode for default calling convention. 230 struct SIMode { 231 bool IEEE = true; 232 bool DX10Clamp = true; 233 bool FP32InputDenormals = true; 234 bool FP32OutputDenormals = true; 235 bool FP64FP16InputDenormals = true; 236 bool FP64FP16OutputDenormals = true; 237 238 SIMode() = default; 239 240 SIMode(const AMDGPU::SIModeRegisterDefaults &Mode) { 241 IEEE = Mode.IEEE; 242 DX10Clamp = Mode.DX10Clamp; 243 FP32InputDenormals = Mode.FP32InputDenormals; 244 FP32OutputDenormals = Mode.FP32OutputDenormals; 245 FP64FP16InputDenormals = Mode.FP64FP16InputDenormals; 246 FP64FP16OutputDenormals = Mode.FP64FP16OutputDenormals; 247 } 248 249 bool operator ==(const SIMode Other) const { 250 return IEEE == Other.IEEE && 251 DX10Clamp == Other.DX10Clamp && 252 FP32InputDenormals == Other.FP32InputDenormals && 253 FP32OutputDenormals == Other.FP32OutputDenormals && 254 FP64FP16InputDenormals == Other.FP64FP16InputDenormals && 255 FP64FP16OutputDenormals == Other.FP64FP16OutputDenormals; 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-input-denormals", Mode.FP32InputDenormals, true); 264 YamlIO.mapOptional("fp32-output-denormals", Mode.FP32OutputDenormals, true); 265 YamlIO.mapOptional("fp64-fp16-input-denormals", Mode.FP64FP16InputDenormals, true); 266 YamlIO.mapOptional("fp64-fp16-output-denormals", Mode.FP64FP16OutputDenormals, true); 267 } 268 }; 269 270 struct SIMachineFunctionInfo final : public yaml::MachineFunctionInfo { 271 uint64_t ExplicitKernArgSize = 0; 272 unsigned MaxKernArgAlign = 0; 273 unsigned LDSSize = 0; 274 Align DynLDSAlign; 275 bool IsEntryFunction = false; 276 bool NoSignedZerosFPMath = false; 277 bool MemoryBound = false; 278 bool WaveLimiter = false; 279 bool HasSpilledSGPRs = false; 280 bool HasSpilledVGPRs = false; 281 uint32_t HighBitsOf32BitAddress = 0; 282 283 // TODO: 10 may be a better default since it's the maximum. 284 unsigned Occupancy = 0; 285 286 StringValue ScratchRSrcReg = "$private_rsrc_reg"; 287 StringValue FrameOffsetReg = "$fp_reg"; 288 StringValue StackPtrOffsetReg = "$sp_reg"; 289 290 Optional<SIArgumentInfo> ArgInfo; 291 SIMode Mode; 292 Optional<FrameIndex> ScavengeFI; 293 294 SIMachineFunctionInfo() = default; 295 SIMachineFunctionInfo(const llvm::SIMachineFunctionInfo &, 296 const TargetRegisterInfo &TRI, 297 const llvm::MachineFunction &MF); 298 299 void mappingImpl(yaml::IO &YamlIO) override; 300 ~SIMachineFunctionInfo() = default; 301 }; 302 303 template <> struct MappingTraits<SIMachineFunctionInfo> { 304 static void mapping(IO &YamlIO, SIMachineFunctionInfo &MFI) { 305 YamlIO.mapOptional("explicitKernArgSize", MFI.ExplicitKernArgSize, 306 UINT64_C(0)); 307 YamlIO.mapOptional("maxKernArgAlign", MFI.MaxKernArgAlign, 0u); 308 YamlIO.mapOptional("ldsSize", MFI.LDSSize, 0u); 309 YamlIO.mapOptional("dynLDSAlign", MFI.DynLDSAlign, Align()); 310 YamlIO.mapOptional("isEntryFunction", MFI.IsEntryFunction, false); 311 YamlIO.mapOptional("noSignedZerosFPMath", MFI.NoSignedZerosFPMath, false); 312 YamlIO.mapOptional("memoryBound", MFI.MemoryBound, false); 313 YamlIO.mapOptional("waveLimiter", MFI.WaveLimiter, false); 314 YamlIO.mapOptional("hasSpilledSGPRs", MFI.HasSpilledSGPRs, false); 315 YamlIO.mapOptional("hasSpilledVGPRs", MFI.HasSpilledVGPRs, false); 316 YamlIO.mapOptional("scratchRSrcReg", MFI.ScratchRSrcReg, 317 StringValue("$private_rsrc_reg")); 318 YamlIO.mapOptional("frameOffsetReg", MFI.FrameOffsetReg, 319 StringValue("$fp_reg")); 320 YamlIO.mapOptional("stackPtrOffsetReg", MFI.StackPtrOffsetReg, 321 StringValue("$sp_reg")); 322 YamlIO.mapOptional("argumentInfo", MFI.ArgInfo); 323 YamlIO.mapOptional("mode", MFI.Mode, SIMode()); 324 YamlIO.mapOptional("highBitsOf32BitAddress", 325 MFI.HighBitsOf32BitAddress, 0u); 326 YamlIO.mapOptional("occupancy", MFI.Occupancy, 0); 327 YamlIO.mapOptional("scavengeFI", MFI.ScavengeFI); 328 } 329 }; 330 331 } // end namespace yaml 332 333 /// This class keeps track of the SPI_SP_INPUT_ADDR config register, which 334 /// tells the hardware which interpolation parameters to load. 335 class SIMachineFunctionInfo final : public AMDGPUMachineFunction { 336 friend class GCNTargetMachine; 337 338 Register TIDReg = AMDGPU::NoRegister; 339 340 // Registers that may be reserved for spilling purposes. These may be the same 341 // as the input registers. 342 Register ScratchRSrcReg = AMDGPU::PRIVATE_RSRC_REG; 343 344 // This is the the unswizzled offset from the current dispatch's scratch wave 345 // base to the beginning of the current function's frame. 346 Register FrameOffsetReg = AMDGPU::FP_REG; 347 348 // This is an ABI register used in the non-entry calling convention to 349 // communicate the unswizzled offset from the current dispatch's scratch wave 350 // base to the beginning of the new function's frame. 351 Register StackPtrOffsetReg = AMDGPU::SP_REG; 352 353 AMDGPUFunctionArgInfo ArgInfo; 354 355 // Graphics info. 356 unsigned PSInputAddr = 0; 357 unsigned PSInputEnable = 0; 358 359 /// Number of bytes of arguments this function has on the stack. If the callee 360 /// is expected to restore the argument stack this should be a multiple of 16, 361 /// all usable during a tail call. 362 /// 363 /// The alternative would forbid tail call optimisation in some cases: if we 364 /// want to transfer control from a function with 8-bytes of stack-argument 365 /// space to a function with 16-bytes then misalignment of this value would 366 /// make a stack adjustment necessary, which could not be undone by the 367 /// callee. 368 unsigned BytesInStackArgArea = 0; 369 370 bool ReturnsVoid = true; 371 372 // A pair of default/requested minimum/maximum flat work group sizes. 373 // Minimum - first, maximum - second. 374 std::pair<unsigned, unsigned> FlatWorkGroupSizes = {0, 0}; 375 376 // A pair of default/requested minimum/maximum number of waves per execution 377 // unit. Minimum - first, maximum - second. 378 std::pair<unsigned, unsigned> WavesPerEU = {0, 0}; 379 380 std::unique_ptr<const AMDGPUBufferPseudoSourceValue> BufferPSV; 381 std::unique_ptr<const AMDGPUImagePseudoSourceValue> ImagePSV; 382 std::unique_ptr<const AMDGPUGWSResourcePseudoSourceValue> GWSResourcePSV; 383 384 private: 385 unsigned LDSWaveSpillSize = 0; 386 unsigned NumUserSGPRs = 0; 387 unsigned NumSystemSGPRs = 0; 388 389 bool HasSpilledSGPRs = false; 390 bool HasSpilledVGPRs = false; 391 bool HasNonSpillStackObjects = false; 392 bool IsStackRealigned = false; 393 394 unsigned NumSpilledSGPRs = 0; 395 unsigned NumSpilledVGPRs = 0; 396 397 // Feature bits required for inputs passed in user SGPRs. 398 bool PrivateSegmentBuffer : 1; 399 bool DispatchPtr : 1; 400 bool QueuePtr : 1; 401 bool KernargSegmentPtr : 1; 402 bool DispatchID : 1; 403 bool FlatScratchInit : 1; 404 405 // Feature bits required for inputs passed in system SGPRs. 406 bool WorkGroupIDX : 1; // Always initialized. 407 bool WorkGroupIDY : 1; 408 bool WorkGroupIDZ : 1; 409 bool WorkGroupInfo : 1; 410 bool PrivateSegmentWaveByteOffset : 1; 411 412 bool WorkItemIDX : 1; // Always initialized. 413 bool WorkItemIDY : 1; 414 bool WorkItemIDZ : 1; 415 416 // Private memory buffer 417 // Compute directly in sgpr[0:1] 418 // Other shaders indirect 64-bits at sgpr[0:1] 419 bool ImplicitBufferPtr : 1; 420 421 // Pointer to where the ABI inserts special kernel arguments separate from the 422 // user arguments. This is an offset from the KernargSegmentPtr. 423 bool ImplicitArgPtr : 1; 424 425 // The hard-wired high half of the address of the global information table 426 // for AMDPAL OS type. 0xffffffff represents no hard-wired high half, since 427 // current hardware only allows a 16 bit value. 428 unsigned GITPtrHigh; 429 430 unsigned HighBitsOf32BitAddress; 431 unsigned GDSSize; 432 433 // Current recorded maximum possible occupancy. 434 unsigned Occupancy; 435 436 mutable Optional<bool> UsesAGPRs; 437 438 MCPhysReg getNextUserSGPR() const; 439 440 MCPhysReg getNextSystemSGPR() const; 441 442 public: 443 struct SpilledReg { 444 Register VGPR; 445 int Lane = -1; 446 447 SpilledReg() = default; 448 SpilledReg(Register R, int L) : VGPR (R), Lane (L) {} 449 450 bool hasLane() { return Lane != -1;} 451 bool hasReg() { return VGPR != 0;} 452 }; 453 454 struct SGPRSpillVGPR { 455 // VGPR used for SGPR spills 456 Register VGPR; 457 458 // If the VGPR is is used for SGPR spills in a non-entrypoint function, the 459 // stack slot used to save/restore it in the prolog/epilog. 460 Optional<int> FI; 461 462 SGPRSpillVGPR(Register V, Optional<int> F) : VGPR(V), FI(F) {} 463 }; 464 465 struct VGPRSpillToAGPR { 466 SmallVector<MCPhysReg, 32> Lanes; 467 bool FullyAllocated = false; 468 bool IsDead = false; 469 }; 470 471 // Map WWM VGPR to a stack slot that is used to save/restore it in the 472 // prolog/epilog. 473 MapVector<Register, Optional<int>> WWMReservedRegs; 474 475 private: 476 // Track VGPR + wave index for each subregister of the SGPR spilled to 477 // frameindex key. 478 DenseMap<int, std::vector<SpilledReg>> SGPRToVGPRSpills; 479 unsigned NumVGPRSpillLanes = 0; 480 SmallVector<SGPRSpillVGPR, 2> SpillVGPRs; 481 482 DenseMap<int, VGPRSpillToAGPR> VGPRToAGPRSpills; 483 484 // AGPRs used for VGPR spills. 485 SmallVector<MCPhysReg, 32> SpillAGPR; 486 487 // VGPRs used for AGPR spills. 488 SmallVector<MCPhysReg, 32> SpillVGPR; 489 490 // Emergency stack slot. Sometimes, we create this before finalizing the stack 491 // frame, so save it here and add it to the RegScavenger later. 492 Optional<int> ScavengeFI; 493 494 public: // FIXME 495 /// If this is set, an SGPR used for save/restore of the register used for the 496 /// frame pointer. 497 Register SGPRForFPSaveRestoreCopy; 498 Optional<int> FramePointerSaveIndex; 499 500 /// If this is set, an SGPR used for save/restore of the register used for the 501 /// base pointer. 502 Register SGPRForBPSaveRestoreCopy; 503 Optional<int> BasePointerSaveIndex; 504 505 Register VGPRReservedForSGPRSpill; 506 bool isCalleeSavedReg(const MCPhysReg *CSRegs, MCPhysReg Reg); 507 508 public: 509 SIMachineFunctionInfo(const MachineFunction &MF); 510 511 bool initializeBaseYamlFields(const yaml::SIMachineFunctionInfo &YamlMFI, 512 const MachineFunction &MF, 513 PerFunctionMIParsingState &PFS, 514 SMDiagnostic &Error, SMRange &SourceRange); 515 516 void reserveWWMRegister(Register Reg, Optional<int> FI) { 517 WWMReservedRegs.insert(std::make_pair(Reg, FI)); 518 } 519 520 ArrayRef<SpilledReg> getSGPRToVGPRSpills(int FrameIndex) const { 521 auto I = SGPRToVGPRSpills.find(FrameIndex); 522 return (I == SGPRToVGPRSpills.end()) ? 523 ArrayRef<SpilledReg>() : makeArrayRef(I->second); 524 } 525 526 ArrayRef<SGPRSpillVGPR> getSGPRSpillVGPRs() const { return SpillVGPRs; } 527 528 void setSGPRSpillVGPRs(Register NewVGPR, Optional<int> newFI, int Index) { 529 SpillVGPRs[Index].VGPR = NewVGPR; 530 SpillVGPRs[Index].FI = newFI; 531 VGPRReservedForSGPRSpill = NewVGPR; 532 } 533 534 bool removeVGPRForSGPRSpill(Register ReservedVGPR, MachineFunction &MF); 535 536 ArrayRef<MCPhysReg> getAGPRSpillVGPRs() const { 537 return SpillAGPR; 538 } 539 540 ArrayRef<MCPhysReg> getVGPRSpillAGPRs() const { 541 return SpillVGPR; 542 } 543 544 MCPhysReg getVGPRToAGPRSpill(int FrameIndex, unsigned Lane) const { 545 auto I = VGPRToAGPRSpills.find(FrameIndex); 546 return (I == VGPRToAGPRSpills.end()) ? (MCPhysReg)AMDGPU::NoRegister 547 : I->second.Lanes[Lane]; 548 } 549 550 void setVGPRToAGPRSpillDead(int FrameIndex) { 551 auto I = VGPRToAGPRSpills.find(FrameIndex); 552 if (I != VGPRToAGPRSpills.end()) 553 I->second.IsDead = true; 554 } 555 556 bool haveFreeLanesForSGPRSpill(const MachineFunction &MF, 557 unsigned NumLane) const; 558 bool allocateSGPRSpillToVGPR(MachineFunction &MF, int FI); 559 bool reserveVGPRforSGPRSpills(MachineFunction &MF); 560 bool allocateVGPRSpillToAGPR(MachineFunction &MF, int FI, bool isAGPRtoVGPR); 561 void removeDeadFrameIndices(MachineFrameInfo &MFI); 562 563 int getScavengeFI(MachineFrameInfo &MFI, const SIRegisterInfo &TRI); 564 Optional<int> getOptionalScavengeFI() const { return ScavengeFI; } 565 566 bool hasCalculatedTID() const { return TIDReg != 0; }; 567 Register getTIDReg() const { return TIDReg; }; 568 void setTIDReg(Register Reg) { TIDReg = Reg; } 569 570 unsigned getBytesInStackArgArea() const { 571 return BytesInStackArgArea; 572 } 573 574 void setBytesInStackArgArea(unsigned Bytes) { 575 BytesInStackArgArea = Bytes; 576 } 577 578 // Add user SGPRs. 579 Register addPrivateSegmentBuffer(const SIRegisterInfo &TRI); 580 Register addDispatchPtr(const SIRegisterInfo &TRI); 581 Register addQueuePtr(const SIRegisterInfo &TRI); 582 Register addKernargSegmentPtr(const SIRegisterInfo &TRI); 583 Register addDispatchID(const SIRegisterInfo &TRI); 584 Register addFlatScratchInit(const SIRegisterInfo &TRI); 585 Register addImplicitBufferPtr(const SIRegisterInfo &TRI); 586 587 // Add system SGPRs. 588 Register addWorkGroupIDX() { 589 ArgInfo.WorkGroupIDX = ArgDescriptor::createRegister(getNextSystemSGPR()); 590 NumSystemSGPRs += 1; 591 return ArgInfo.WorkGroupIDX.getRegister(); 592 } 593 594 Register addWorkGroupIDY() { 595 ArgInfo.WorkGroupIDY = ArgDescriptor::createRegister(getNextSystemSGPR()); 596 NumSystemSGPRs += 1; 597 return ArgInfo.WorkGroupIDY.getRegister(); 598 } 599 600 Register addWorkGroupIDZ() { 601 ArgInfo.WorkGroupIDZ = ArgDescriptor::createRegister(getNextSystemSGPR()); 602 NumSystemSGPRs += 1; 603 return ArgInfo.WorkGroupIDZ.getRegister(); 604 } 605 606 Register addWorkGroupInfo() { 607 ArgInfo.WorkGroupInfo = ArgDescriptor::createRegister(getNextSystemSGPR()); 608 NumSystemSGPRs += 1; 609 return ArgInfo.WorkGroupInfo.getRegister(); 610 } 611 612 // Add special VGPR inputs 613 void setWorkItemIDX(ArgDescriptor Arg) { 614 ArgInfo.WorkItemIDX = Arg; 615 } 616 617 void setWorkItemIDY(ArgDescriptor Arg) { 618 ArgInfo.WorkItemIDY = Arg; 619 } 620 621 void setWorkItemIDZ(ArgDescriptor Arg) { 622 ArgInfo.WorkItemIDZ = Arg; 623 } 624 625 Register addPrivateSegmentWaveByteOffset() { 626 ArgInfo.PrivateSegmentWaveByteOffset 627 = ArgDescriptor::createRegister(getNextSystemSGPR()); 628 NumSystemSGPRs += 1; 629 return ArgInfo.PrivateSegmentWaveByteOffset.getRegister(); 630 } 631 632 void setPrivateSegmentWaveByteOffset(Register Reg) { 633 ArgInfo.PrivateSegmentWaveByteOffset = ArgDescriptor::createRegister(Reg); 634 } 635 636 bool hasPrivateSegmentBuffer() const { 637 return PrivateSegmentBuffer; 638 } 639 640 bool hasDispatchPtr() const { 641 return DispatchPtr; 642 } 643 644 bool hasQueuePtr() const { 645 return QueuePtr; 646 } 647 648 bool hasKernargSegmentPtr() const { 649 return KernargSegmentPtr; 650 } 651 652 bool hasDispatchID() const { 653 return DispatchID; 654 } 655 656 bool hasFlatScratchInit() const { 657 return FlatScratchInit; 658 } 659 660 bool hasWorkGroupIDX() const { 661 return WorkGroupIDX; 662 } 663 664 bool hasWorkGroupIDY() const { 665 return WorkGroupIDY; 666 } 667 668 bool hasWorkGroupIDZ() const { 669 return WorkGroupIDZ; 670 } 671 672 bool hasWorkGroupInfo() const { 673 return WorkGroupInfo; 674 } 675 676 bool hasPrivateSegmentWaveByteOffset() const { 677 return PrivateSegmentWaveByteOffset; 678 } 679 680 bool hasWorkItemIDX() const { 681 return WorkItemIDX; 682 } 683 684 bool hasWorkItemIDY() const { 685 return WorkItemIDY; 686 } 687 688 bool hasWorkItemIDZ() const { 689 return WorkItemIDZ; 690 } 691 692 bool hasImplicitArgPtr() const { 693 return ImplicitArgPtr; 694 } 695 696 bool hasImplicitBufferPtr() const { 697 return ImplicitBufferPtr; 698 } 699 700 AMDGPUFunctionArgInfo &getArgInfo() { 701 return ArgInfo; 702 } 703 704 const AMDGPUFunctionArgInfo &getArgInfo() const { 705 return ArgInfo; 706 } 707 708 std::tuple<const ArgDescriptor *, const TargetRegisterClass *, LLT> 709 getPreloadedValue(AMDGPUFunctionArgInfo::PreloadedValue Value) const { 710 return ArgInfo.getPreloadedValue(Value); 711 } 712 713 MCRegister getPreloadedReg(AMDGPUFunctionArgInfo::PreloadedValue Value) const { 714 auto Arg = std::get<0>(ArgInfo.getPreloadedValue(Value)); 715 return Arg ? Arg->getRegister() : MCRegister(); 716 } 717 718 unsigned getGITPtrHigh() const { 719 return GITPtrHigh; 720 } 721 722 Register getGITPtrLoReg(const MachineFunction &MF) const; 723 724 uint32_t get32BitAddressHighBits() const { 725 return HighBitsOf32BitAddress; 726 } 727 728 unsigned getGDSSize() const { 729 return GDSSize; 730 } 731 732 unsigned getNumUserSGPRs() const { 733 return NumUserSGPRs; 734 } 735 736 unsigned getNumPreloadedSGPRs() const { 737 return NumUserSGPRs + NumSystemSGPRs; 738 } 739 740 Register getPrivateSegmentWaveByteOffsetSystemSGPR() const { 741 return ArgInfo.PrivateSegmentWaveByteOffset.getRegister(); 742 } 743 744 /// Returns the physical register reserved for use as the resource 745 /// descriptor for scratch accesses. 746 Register getScratchRSrcReg() const { 747 return ScratchRSrcReg; 748 } 749 750 void setScratchRSrcReg(Register Reg) { 751 assert(Reg != 0 && "Should never be unset"); 752 ScratchRSrcReg = Reg; 753 } 754 755 Register getFrameOffsetReg() const { 756 return FrameOffsetReg; 757 } 758 759 void setFrameOffsetReg(Register Reg) { 760 assert(Reg != 0 && "Should never be unset"); 761 FrameOffsetReg = Reg; 762 } 763 764 void setStackPtrOffsetReg(Register Reg) { 765 assert(Reg != 0 && "Should never be unset"); 766 StackPtrOffsetReg = Reg; 767 } 768 769 // Note the unset value for this is AMDGPU::SP_REG rather than 770 // NoRegister. This is mostly a workaround for MIR tests where state that 771 // can't be directly computed from the function is not preserved in serialized 772 // MIR. 773 Register getStackPtrOffsetReg() const { 774 return StackPtrOffsetReg; 775 } 776 777 Register getQueuePtrUserSGPR() const { 778 return ArgInfo.QueuePtr.getRegister(); 779 } 780 781 Register getImplicitBufferPtrUserSGPR() const { 782 return ArgInfo.ImplicitBufferPtr.getRegister(); 783 } 784 785 bool hasSpilledSGPRs() const { 786 return HasSpilledSGPRs; 787 } 788 789 void setHasSpilledSGPRs(bool Spill = true) { 790 HasSpilledSGPRs = Spill; 791 } 792 793 bool hasSpilledVGPRs() const { 794 return HasSpilledVGPRs; 795 } 796 797 void setHasSpilledVGPRs(bool Spill = true) { 798 HasSpilledVGPRs = Spill; 799 } 800 801 bool hasNonSpillStackObjects() const { 802 return HasNonSpillStackObjects; 803 } 804 805 void setHasNonSpillStackObjects(bool StackObject = true) { 806 HasNonSpillStackObjects = StackObject; 807 } 808 809 bool isStackRealigned() const { 810 return IsStackRealigned; 811 } 812 813 void setIsStackRealigned(bool Realigned = true) { 814 IsStackRealigned = Realigned; 815 } 816 817 unsigned getNumSpilledSGPRs() const { 818 return NumSpilledSGPRs; 819 } 820 821 unsigned getNumSpilledVGPRs() const { 822 return NumSpilledVGPRs; 823 } 824 825 void addToSpilledSGPRs(unsigned num) { 826 NumSpilledSGPRs += num; 827 } 828 829 void addToSpilledVGPRs(unsigned num) { 830 NumSpilledVGPRs += num; 831 } 832 833 unsigned getPSInputAddr() const { 834 return PSInputAddr; 835 } 836 837 unsigned getPSInputEnable() const { 838 return PSInputEnable; 839 } 840 841 bool isPSInputAllocated(unsigned Index) const { 842 return PSInputAddr & (1 << Index); 843 } 844 845 void markPSInputAllocated(unsigned Index) { 846 PSInputAddr |= 1 << Index; 847 } 848 849 void markPSInputEnabled(unsigned Index) { 850 PSInputEnable |= 1 << Index; 851 } 852 853 bool returnsVoid() const { 854 return ReturnsVoid; 855 } 856 857 void setIfReturnsVoid(bool Value) { 858 ReturnsVoid = Value; 859 } 860 861 /// \returns A pair of default/requested minimum/maximum flat work group sizes 862 /// for this function. 863 std::pair<unsigned, unsigned> getFlatWorkGroupSizes() const { 864 return FlatWorkGroupSizes; 865 } 866 867 /// \returns Default/requested minimum flat work group size for this function. 868 unsigned getMinFlatWorkGroupSize() const { 869 return FlatWorkGroupSizes.first; 870 } 871 872 /// \returns Default/requested maximum flat work group size for this function. 873 unsigned getMaxFlatWorkGroupSize() const { 874 return FlatWorkGroupSizes.second; 875 } 876 877 /// \returns A pair of default/requested minimum/maximum number of waves per 878 /// execution unit. 879 std::pair<unsigned, unsigned> getWavesPerEU() const { 880 return WavesPerEU; 881 } 882 883 /// \returns Default/requested minimum number of waves per execution unit. 884 unsigned getMinWavesPerEU() const { 885 return WavesPerEU.first; 886 } 887 888 /// \returns Default/requested maximum number of waves per execution unit. 889 unsigned getMaxWavesPerEU() const { 890 return WavesPerEU.second; 891 } 892 893 /// \returns SGPR used for \p Dim's work group ID. 894 Register getWorkGroupIDSGPR(unsigned Dim) const { 895 switch (Dim) { 896 case 0: 897 assert(hasWorkGroupIDX()); 898 return ArgInfo.WorkGroupIDX.getRegister(); 899 case 1: 900 assert(hasWorkGroupIDY()); 901 return ArgInfo.WorkGroupIDY.getRegister(); 902 case 2: 903 assert(hasWorkGroupIDZ()); 904 return ArgInfo.WorkGroupIDZ.getRegister(); 905 } 906 llvm_unreachable("unexpected dimension"); 907 } 908 909 unsigned getLDSWaveSpillSize() const { 910 return LDSWaveSpillSize; 911 } 912 913 const AMDGPUBufferPseudoSourceValue *getBufferPSV(const SIInstrInfo &TII) { 914 if (!BufferPSV) 915 BufferPSV = std::make_unique<AMDGPUBufferPseudoSourceValue>(TII); 916 917 return BufferPSV.get(); 918 } 919 920 const AMDGPUImagePseudoSourceValue *getImagePSV(const SIInstrInfo &TII) { 921 if (!ImagePSV) 922 ImagePSV = std::make_unique<AMDGPUImagePseudoSourceValue>(TII); 923 924 return ImagePSV.get(); 925 } 926 927 const AMDGPUGWSResourcePseudoSourceValue *getGWSPSV(const SIInstrInfo &TII) { 928 if (!GWSResourcePSV) { 929 GWSResourcePSV = 930 std::make_unique<AMDGPUGWSResourcePseudoSourceValue>(TII); 931 } 932 933 return GWSResourcePSV.get(); 934 } 935 936 unsigned getOccupancy() const { 937 return Occupancy; 938 } 939 940 unsigned getMinAllowedOccupancy() const { 941 if (!isMemoryBound() && !needsWaveLimiter()) 942 return Occupancy; 943 return (Occupancy < 4) ? Occupancy : 4; 944 } 945 946 void limitOccupancy(const MachineFunction &MF); 947 948 void limitOccupancy(unsigned Limit) { 949 if (Occupancy > Limit) 950 Occupancy = Limit; 951 } 952 953 void increaseOccupancy(const MachineFunction &MF, unsigned Limit) { 954 if (Occupancy < Limit) 955 Occupancy = Limit; 956 limitOccupancy(MF); 957 } 958 959 // \returns true if a function needs or may need AGPRs. 960 bool usesAGPRs(const MachineFunction &MF) const; 961 }; 962 963 } // end namespace llvm 964 965 #endif // LLVM_LIB_TARGET_AMDGPU_SIMACHINEFUNCTIONINFO_H 966