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: 52 AMDGPUPseudoSourceValue(unsigned Kind, const TargetInstrInfo &TII) 53 : PseudoSourceValue(Kind, TII) {} 54 55 public: 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 62 bool isAliased(const MachineFrameInfo *) const override { 63 return true; 64 } 65 66 bool mayAlias(const MachineFrameInfo *) const override { 67 return true; 68 } 69 }; 70 71 class AMDGPUBufferPseudoSourceValue final : public AMDGPUPseudoSourceValue { 72 public: 73 explicit AMDGPUBufferPseudoSourceValue(const TargetInstrInfo &TII) 74 : AMDGPUPseudoSourceValue(PSVBuffer, TII) {} 75 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? 84 explicit AMDGPUImagePseudoSourceValue(const TargetInstrInfo &TII) 85 : AMDGPUPseudoSourceValue(PSVImage, TII) {} 86 87 static bool classof(const PseudoSourceValue *V) { 88 return V->kind() == PSVImage; 89 } 90 }; 91 92 class AMDGPUGWSResourcePseudoSourceValue final : public AMDGPUPseudoSourceValue { 93 public: 94 explicit AMDGPUGWSResourcePseudoSourceValue(const TargetInstrInfo &TII) 95 : AMDGPUPseudoSourceValue(GWSResource, TII) {} 96 97 static bool classof(const PseudoSourceValue *V) { 98 return V->kind() == GWSResource; 99 } 100 101 // These are inaccessible memory from IR. 102 bool isAliased(const MachineFrameInfo *) const override { 103 return false; 104 } 105 106 // These are inaccessible memory from IR. 107 bool mayAlias(const MachineFrameInfo *) const override { 108 return false; 109 } 110 111 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. 127 SIArgument() : IsRegister(false), StackOffset(0) {} 128 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 } 147 ~SIArgument() { 148 if (IsRegister) 149 RegisterName.~StringValue(); 150 } 151 152 // Helper to create a register or stack argument. 153 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. 161 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 FP32InputDenormals = true; 240 bool FP32OutputDenormals = true; 241 bool FP64FP16InputDenormals = true; 242 bool FP64FP16OutputDenormals = true; 243 244 SIMode() = default; 245 246 SIMode(const AMDGPU::SIModeRegisterDefaults &Mode) { 247 IEEE = Mode.IEEE; 248 DX10Clamp = Mode.DX10Clamp; 249 FP32InputDenormals = Mode.FP32InputDenormals; 250 FP32OutputDenormals = Mode.FP32OutputDenormals; 251 FP64FP16InputDenormals = Mode.FP64FP16InputDenormals; 252 FP64FP16OutputDenormals = Mode.FP64FP16OutputDenormals; 253 } 254 255 bool operator ==(const SIMode Other) const { 256 return IEEE == Other.IEEE && 257 DX10Clamp == Other.DX10Clamp && 258 FP32InputDenormals == Other.FP32InputDenormals && 259 FP32OutputDenormals == Other.FP32OutputDenormals && 260 FP64FP16InputDenormals == Other.FP64FP16InputDenormals && 261 FP64FP16OutputDenormals == Other.FP64FP16OutputDenormals; 262 } 263 }; 264 265 template <> struct MappingTraits<SIMode> { 266 static void mapping(IO &YamlIO, SIMode &Mode) { 267 YamlIO.mapOptional("ieee", Mode.IEEE, true); 268 YamlIO.mapOptional("dx10-clamp", Mode.DX10Clamp, true); 269 YamlIO.mapOptional("fp32-input-denormals", Mode.FP32InputDenormals, true); 270 YamlIO.mapOptional("fp32-output-denormals", Mode.FP32OutputDenormals, true); 271 YamlIO.mapOptional("fp64-fp16-input-denormals", Mode.FP64FP16InputDenormals, true); 272 YamlIO.mapOptional("fp64-fp16-output-denormals", Mode.FP64FP16OutputDenormals, true); 273 } 274 }; 275 276 struct SIMachineFunctionInfo final : public yaml::MachineFunctionInfo { 277 uint64_t ExplicitKernArgSize = 0; 278 unsigned MaxKernArgAlign = 0; 279 unsigned LDSSize = 0; 280 bool IsEntryFunction = false; 281 bool NoSignedZerosFPMath = false; 282 bool MemoryBound = false; 283 bool WaveLimiter = false; 284 bool HasSpilledSGPRs = false; 285 bool HasSpilledVGPRs = false; 286 uint32_t HighBitsOf32BitAddress = 0; 287 288 StringValue ScratchRSrcReg = "$private_rsrc_reg"; 289 StringValue FrameOffsetReg = "$fp_reg"; 290 StringValue StackPtrOffsetReg = "$sp_reg"; 291 292 Optional<SIArgumentInfo> ArgInfo; 293 SIMode Mode; 294 295 SIMachineFunctionInfo() = default; 296 SIMachineFunctionInfo(const llvm::SIMachineFunctionInfo &, 297 const TargetRegisterInfo &TRI); 298 299 void mappingImpl(yaml::IO &YamlIO) override; 300 ~SIMachineFunctionInfo() = default; 301 }; 302 303 template <> struct MappingTraits<SIMachineFunctionInfo> { 304 static void mapping(IO &YamlIO, SIMachineFunctionInfo &MFI) { 305 YamlIO.mapOptional("explicitKernArgSize", MFI.ExplicitKernArgSize, 306 UINT64_C(0)); 307 YamlIO.mapOptional("maxKernArgAlign", MFI.MaxKernArgAlign, 0u); 308 YamlIO.mapOptional("ldsSize", MFI.LDSSize, 0u); 309 YamlIO.mapOptional("isEntryFunction", MFI.IsEntryFunction, false); 310 YamlIO.mapOptional("noSignedZerosFPMath", MFI.NoSignedZerosFPMath, false); 311 YamlIO.mapOptional("memoryBound", MFI.MemoryBound, false); 312 YamlIO.mapOptional("waveLimiter", MFI.WaveLimiter, false); 313 YamlIO.mapOptional("hasSpilledSGPRs", MFI.HasSpilledSGPRs, false); 314 YamlIO.mapOptional("hasSpilledVGPRs", MFI.HasSpilledVGPRs, false); 315 YamlIO.mapOptional("scratchRSrcReg", MFI.ScratchRSrcReg, 316 StringValue("$private_rsrc_reg")); 317 YamlIO.mapOptional("frameOffsetReg", MFI.FrameOffsetReg, 318 StringValue("$fp_reg")); 319 YamlIO.mapOptional("stackPtrOffsetReg", MFI.StackPtrOffsetReg, 320 StringValue("$sp_reg")); 321 YamlIO.mapOptional("argumentInfo", MFI.ArgInfo); 322 YamlIO.mapOptional("mode", MFI.Mode, SIMode()); 323 YamlIO.mapOptional("highBitsOf32BitAddress", 324 MFI.HighBitsOf32BitAddress, 0u); 325 } 326 }; 327 328 } // end namespace yaml 329 330 /// This class keeps track of the SPI_SP_INPUT_ADDR config register, which 331 /// tells the hardware which interpolation parameters to load. 332 class SIMachineFunctionInfo final : public AMDGPUMachineFunction { 333 friend class GCNTargetMachine; 334 335 Register TIDReg = AMDGPU::NoRegister; 336 337 // Registers that may be reserved for spilling purposes. These may be the same 338 // as the input registers. 339 Register ScratchRSrcReg = AMDGPU::PRIVATE_RSRC_REG; 340 341 // This is the the unswizzled offset from the current dispatch's scratch wave 342 // base to the beginning of the current function's frame. 343 Register FrameOffsetReg = AMDGPU::FP_REG; 344 345 // This is an ABI register used in the non-entry calling convention to 346 // communicate the unswizzled offset from the current dispatch's scratch wave 347 // base to the beginning of the new function's frame. 348 Register StackPtrOffsetReg = AMDGPU::SP_REG; 349 350 AMDGPUFunctionArgInfo ArgInfo; 351 352 // Graphics info. 353 unsigned PSInputAddr = 0; 354 unsigned PSInputEnable = 0; 355 356 /// Number of bytes of arguments this function has on the stack. If the callee 357 /// is expected to restore the argument stack this should be a multiple of 16, 358 /// all usable during a tail call. 359 /// 360 /// The alternative would forbid tail call optimisation in some cases: if we 361 /// want to transfer control from a function with 8-bytes of stack-argument 362 /// space to a function with 16-bytes then misalignment of this value would 363 /// make a stack adjustment necessary, which could not be undone by the 364 /// callee. 365 unsigned BytesInStackArgArea = 0; 366 367 bool ReturnsVoid = true; 368 369 // A pair of default/requested minimum/maximum flat work group sizes. 370 // Minimum - first, maximum - second. 371 std::pair<unsigned, unsigned> FlatWorkGroupSizes = {0, 0}; 372 373 // A pair of default/requested minimum/maximum number of waves per execution 374 // unit. Minimum - first, maximum - second. 375 std::pair<unsigned, unsigned> WavesPerEU = {0, 0}; 376 377 DenseMap<const Value *, 378 std::unique_ptr<const AMDGPUBufferPseudoSourceValue>> BufferPSVs; 379 DenseMap<const Value *, 380 std::unique_ptr<const AMDGPUImagePseudoSourceValue>> ImagePSVs; 381 std::unique_ptr<const AMDGPUGWSResourcePseudoSourceValue> GWSResourcePSV; 382 383 private: 384 unsigned LDSWaveSpillSize = 0; 385 unsigned NumUserSGPRs = 0; 386 unsigned NumSystemSGPRs = 0; 387 388 bool HasSpilledSGPRs = false; 389 bool HasSpilledVGPRs = false; 390 bool HasNonSpillStackObjects = false; 391 bool IsStackRealigned = false; 392 393 unsigned NumSpilledSGPRs = 0; 394 unsigned NumSpilledVGPRs = 0; 395 396 // Feature bits required for inputs passed in user SGPRs. 397 bool PrivateSegmentBuffer : 1; 398 bool DispatchPtr : 1; 399 bool QueuePtr : 1; 400 bool KernargSegmentPtr : 1; 401 bool DispatchID : 1; 402 bool FlatScratchInit : 1; 403 404 // Feature bits required for inputs passed in system SGPRs. 405 bool WorkGroupIDX : 1; // Always initialized. 406 bool WorkGroupIDY : 1; 407 bool WorkGroupIDZ : 1; 408 bool WorkGroupInfo : 1; 409 bool PrivateSegmentWaveByteOffset : 1; 410 411 bool WorkItemIDX : 1; // Always initialized. 412 bool WorkItemIDY : 1; 413 bool WorkItemIDZ : 1; 414 415 // Private memory buffer 416 // Compute directly in sgpr[0:1] 417 // Other shaders indirect 64-bits at sgpr[0:1] 418 bool ImplicitBufferPtr : 1; 419 420 // Pointer to where the ABI inserts special kernel arguments separate from the 421 // user arguments. This is an offset from the KernargSegmentPtr. 422 bool ImplicitArgPtr : 1; 423 424 // The hard-wired high half of the address of the global information table 425 // for AMDPAL OS type. 0xffffffff represents no hard-wired high half, since 426 // current hardware only allows a 16 bit value. 427 unsigned GITPtrHigh; 428 429 unsigned HighBitsOf32BitAddress; 430 unsigned GDSSize; 431 432 // Current recorded maximum possible occupancy. 433 unsigned Occupancy; 434 435 MCPhysReg getNextUserSGPR() const; 436 437 MCPhysReg getNextSystemSGPR() const; 438 439 public: 440 struct SpilledReg { 441 Register VGPR; 442 int Lane = -1; 443 444 SpilledReg() = default; 445 SpilledReg(Register R, int L) : VGPR (R), Lane (L) {} 446 447 bool hasLane() { return Lane != -1;} 448 bool hasReg() { return VGPR != 0;} 449 }; 450 451 struct SGPRSpillVGPRCSR { 452 // VGPR used for SGPR spills 453 Register VGPR; 454 455 // If the VGPR is a CSR, the stack slot used to save/restore it in the 456 // prolog/epilog. 457 Optional<int> FI; 458 459 SGPRSpillVGPRCSR(Register V, Optional<int> F) : VGPR(V), FI(F) {} 460 }; 461 462 struct VGPRSpillToAGPR { 463 SmallVector<MCPhysReg, 32> Lanes; 464 bool FullyAllocated = false; 465 }; 466 467 SparseBitVector<> WWMReservedRegs; 468 469 void ReserveWWMRegister(Register Reg) { WWMReservedRegs.set(Reg); } 470 471 private: 472 // Track VGPR + wave index for each subregister of the SGPR spilled to 473 // frameindex key. 474 DenseMap<int, std::vector<SpilledReg>> SGPRToVGPRSpills; 475 unsigned NumVGPRSpillLanes = 0; 476 SmallVector<SGPRSpillVGPRCSR, 2> SpillVGPRs; 477 478 DenseMap<int, VGPRSpillToAGPR> VGPRToAGPRSpills; 479 480 // AGPRs used for VGPR spills. 481 SmallVector<MCPhysReg, 32> SpillAGPR; 482 483 // VGPRs used for AGPR spills. 484 SmallVector<MCPhysReg, 32> SpillVGPR; 485 486 public: // FIXME 487 /// If this is set, an SGPR used for save/restore of the register used for the 488 /// frame pointer. 489 Register SGPRForFPSaveRestoreCopy; 490 Optional<int> FramePointerSaveIndex; 491 492 /// If this is set, an SGPR used for save/restore of the register used for the 493 /// base pointer. 494 Register SGPRForBPSaveRestoreCopy; 495 Optional<int> BasePointerSaveIndex; 496 497 Register VGPRReservedForSGPRSpill; 498 bool isCalleeSavedReg(const MCPhysReg *CSRegs, MCPhysReg Reg); 499 500 public: 501 SIMachineFunctionInfo(const MachineFunction &MF); 502 503 bool initializeBaseYamlFields(const yaml::SIMachineFunctionInfo &YamlMFI); 504 505 ArrayRef<SpilledReg> getSGPRToVGPRSpills(int FrameIndex) const { 506 auto I = SGPRToVGPRSpills.find(FrameIndex); 507 return (I == SGPRToVGPRSpills.end()) ? 508 ArrayRef<SpilledReg>() : makeArrayRef(I->second); 509 } 510 511 ArrayRef<SGPRSpillVGPRCSR> getSGPRSpillVGPRs() const { 512 return SpillVGPRs; 513 } 514 515 void setSGPRSpillVGPRs(Register NewVGPR, Optional<int> newFI, int Index) { 516 SpillVGPRs[Index].VGPR = NewVGPR; 517 SpillVGPRs[Index].FI = newFI; 518 VGPRReservedForSGPRSpill = NewVGPR; 519 } 520 521 bool removeVGPRForSGPRSpill(Register ReservedVGPR, MachineFunction &MF); 522 523 ArrayRef<MCPhysReg> getAGPRSpillVGPRs() const { 524 return SpillAGPR; 525 } 526 527 ArrayRef<MCPhysReg> getVGPRSpillAGPRs() const { 528 return SpillVGPR; 529 } 530 531 MCPhysReg getVGPRToAGPRSpill(int FrameIndex, unsigned Lane) const { 532 auto I = VGPRToAGPRSpills.find(FrameIndex); 533 return (I == VGPRToAGPRSpills.end()) ? (MCPhysReg)AMDGPU::NoRegister 534 : I->second.Lanes[Lane]; 535 } 536 537 bool haveFreeLanesForSGPRSpill(const MachineFunction &MF, 538 unsigned NumLane) const; 539 bool allocateSGPRSpillToVGPR(MachineFunction &MF, int FI); 540 bool reserveVGPRforSGPRSpills(MachineFunction &MF); 541 bool allocateVGPRSpillToAGPR(MachineFunction &MF, int FI, bool isAGPRtoVGPR); 542 void removeDeadFrameIndices(MachineFrameInfo &MFI); 543 544 bool hasCalculatedTID() const { return TIDReg != 0; }; 545 Register getTIDReg() const { return TIDReg; }; 546 void setTIDReg(Register Reg) { TIDReg = Reg; } 547 548 unsigned getBytesInStackArgArea() const { 549 return BytesInStackArgArea; 550 } 551 552 void setBytesInStackArgArea(unsigned Bytes) { 553 BytesInStackArgArea = Bytes; 554 } 555 556 // Add user SGPRs. 557 Register addPrivateSegmentBuffer(const SIRegisterInfo &TRI); 558 Register addDispatchPtr(const SIRegisterInfo &TRI); 559 Register addQueuePtr(const SIRegisterInfo &TRI); 560 Register addKernargSegmentPtr(const SIRegisterInfo &TRI); 561 Register addDispatchID(const SIRegisterInfo &TRI); 562 Register addFlatScratchInit(const SIRegisterInfo &TRI); 563 Register addImplicitBufferPtr(const SIRegisterInfo &TRI); 564 565 // Add system SGPRs. 566 Register addWorkGroupIDX() { 567 ArgInfo.WorkGroupIDX = ArgDescriptor::createRegister(getNextSystemSGPR()); 568 NumSystemSGPRs += 1; 569 return ArgInfo.WorkGroupIDX.getRegister(); 570 } 571 572 Register addWorkGroupIDY() { 573 ArgInfo.WorkGroupIDY = ArgDescriptor::createRegister(getNextSystemSGPR()); 574 NumSystemSGPRs += 1; 575 return ArgInfo.WorkGroupIDY.getRegister(); 576 } 577 578 Register addWorkGroupIDZ() { 579 ArgInfo.WorkGroupIDZ = ArgDescriptor::createRegister(getNextSystemSGPR()); 580 NumSystemSGPRs += 1; 581 return ArgInfo.WorkGroupIDZ.getRegister(); 582 } 583 584 Register addWorkGroupInfo() { 585 ArgInfo.WorkGroupInfo = ArgDescriptor::createRegister(getNextSystemSGPR()); 586 NumSystemSGPRs += 1; 587 return ArgInfo.WorkGroupInfo.getRegister(); 588 } 589 590 // Add special VGPR inputs 591 void setWorkItemIDX(ArgDescriptor Arg) { 592 ArgInfo.WorkItemIDX = Arg; 593 } 594 595 void setWorkItemIDY(ArgDescriptor Arg) { 596 ArgInfo.WorkItemIDY = Arg; 597 } 598 599 void setWorkItemIDZ(ArgDescriptor Arg) { 600 ArgInfo.WorkItemIDZ = Arg; 601 } 602 603 Register addPrivateSegmentWaveByteOffset() { 604 ArgInfo.PrivateSegmentWaveByteOffset 605 = ArgDescriptor::createRegister(getNextSystemSGPR()); 606 NumSystemSGPRs += 1; 607 return ArgInfo.PrivateSegmentWaveByteOffset.getRegister(); 608 } 609 610 void setPrivateSegmentWaveByteOffset(Register Reg) { 611 ArgInfo.PrivateSegmentWaveByteOffset = ArgDescriptor::createRegister(Reg); 612 } 613 614 bool hasPrivateSegmentBuffer() const { 615 return PrivateSegmentBuffer; 616 } 617 618 bool hasDispatchPtr() const { 619 return DispatchPtr; 620 } 621 622 bool hasQueuePtr() const { 623 return QueuePtr; 624 } 625 626 bool hasKernargSegmentPtr() const { 627 return KernargSegmentPtr; 628 } 629 630 bool hasDispatchID() const { 631 return DispatchID; 632 } 633 634 bool hasFlatScratchInit() const { 635 return FlatScratchInit; 636 } 637 638 bool hasWorkGroupIDX() const { 639 return WorkGroupIDX; 640 } 641 642 bool hasWorkGroupIDY() const { 643 return WorkGroupIDY; 644 } 645 646 bool hasWorkGroupIDZ() const { 647 return WorkGroupIDZ; 648 } 649 650 bool hasWorkGroupInfo() const { 651 return WorkGroupInfo; 652 } 653 654 bool hasPrivateSegmentWaveByteOffset() const { 655 return PrivateSegmentWaveByteOffset; 656 } 657 658 bool hasWorkItemIDX() const { 659 return WorkItemIDX; 660 } 661 662 bool hasWorkItemIDY() const { 663 return WorkItemIDY; 664 } 665 666 bool hasWorkItemIDZ() const { 667 return WorkItemIDZ; 668 } 669 670 bool hasImplicitArgPtr() const { 671 return ImplicitArgPtr; 672 } 673 674 bool hasImplicitBufferPtr() const { 675 return ImplicitBufferPtr; 676 } 677 678 AMDGPUFunctionArgInfo &getArgInfo() { 679 return ArgInfo; 680 } 681 682 const AMDGPUFunctionArgInfo &getArgInfo() const { 683 return ArgInfo; 684 } 685 686 std::tuple<const ArgDescriptor *, const TargetRegisterClass *, LLT> 687 getPreloadedValue(AMDGPUFunctionArgInfo::PreloadedValue Value) const { 688 return ArgInfo.getPreloadedValue(Value); 689 } 690 691 MCRegister getPreloadedReg(AMDGPUFunctionArgInfo::PreloadedValue Value) const { 692 auto Arg = std::get<0>(ArgInfo.getPreloadedValue(Value)); 693 return Arg ? Arg->getRegister() : MCRegister(); 694 } 695 696 unsigned getGITPtrHigh() const { 697 return GITPtrHigh; 698 } 699 700 Register getGITPtrLoReg(const MachineFunction &MF) const; 701 702 uint32_t get32BitAddressHighBits() const { 703 return HighBitsOf32BitAddress; 704 } 705 706 unsigned getGDSSize() const { 707 return GDSSize; 708 } 709 710 unsigned getNumUserSGPRs() const { 711 return NumUserSGPRs; 712 } 713 714 unsigned getNumPreloadedSGPRs() const { 715 return NumUserSGPRs + NumSystemSGPRs; 716 } 717 718 Register getPrivateSegmentWaveByteOffsetSystemSGPR() const { 719 return ArgInfo.PrivateSegmentWaveByteOffset.getRegister(); 720 } 721 722 /// Returns the physical register reserved for use as the resource 723 /// descriptor for scratch accesses. 724 Register getScratchRSrcReg() const { 725 return ScratchRSrcReg; 726 } 727 728 void setScratchRSrcReg(Register Reg) { 729 assert(Reg != 0 && "Should never be unset"); 730 ScratchRSrcReg = Reg; 731 } 732 733 Register getFrameOffsetReg() const { 734 return FrameOffsetReg; 735 } 736 737 void setFrameOffsetReg(Register Reg) { 738 assert(Reg != 0 && "Should never be unset"); 739 FrameOffsetReg = Reg; 740 } 741 742 void setStackPtrOffsetReg(Register Reg) { 743 assert(Reg != 0 && "Should never be unset"); 744 StackPtrOffsetReg = Reg; 745 } 746 747 // Note the unset value for this is AMDGPU::SP_REG rather than 748 // NoRegister. This is mostly a workaround for MIR tests where state that 749 // can't be directly computed from the function is not preserved in serialized 750 // MIR. 751 Register getStackPtrOffsetReg() const { 752 return StackPtrOffsetReg; 753 } 754 755 Register getQueuePtrUserSGPR() const { 756 return ArgInfo.QueuePtr.getRegister(); 757 } 758 759 Register getImplicitBufferPtrUserSGPR() const { 760 return ArgInfo.ImplicitBufferPtr.getRegister(); 761 } 762 763 bool hasSpilledSGPRs() const { 764 return HasSpilledSGPRs; 765 } 766 767 void setHasSpilledSGPRs(bool Spill = true) { 768 HasSpilledSGPRs = Spill; 769 } 770 771 bool hasSpilledVGPRs() const { 772 return HasSpilledVGPRs; 773 } 774 775 void setHasSpilledVGPRs(bool Spill = true) { 776 HasSpilledVGPRs = Spill; 777 } 778 779 bool hasNonSpillStackObjects() const { 780 return HasNonSpillStackObjects; 781 } 782 783 void setHasNonSpillStackObjects(bool StackObject = true) { 784 HasNonSpillStackObjects = StackObject; 785 } 786 787 bool isStackRealigned() const { 788 return IsStackRealigned; 789 } 790 791 void setIsStackRealigned(bool Realigned = true) { 792 IsStackRealigned = Realigned; 793 } 794 795 unsigned getNumSpilledSGPRs() const { 796 return NumSpilledSGPRs; 797 } 798 799 unsigned getNumSpilledVGPRs() const { 800 return NumSpilledVGPRs; 801 } 802 803 void addToSpilledSGPRs(unsigned num) { 804 NumSpilledSGPRs += num; 805 } 806 807 void addToSpilledVGPRs(unsigned num) { 808 NumSpilledVGPRs += num; 809 } 810 811 unsigned getPSInputAddr() const { 812 return PSInputAddr; 813 } 814 815 unsigned getPSInputEnable() const { 816 return PSInputEnable; 817 } 818 819 bool isPSInputAllocated(unsigned Index) const { 820 return PSInputAddr & (1 << Index); 821 } 822 823 void markPSInputAllocated(unsigned Index) { 824 PSInputAddr |= 1 << Index; 825 } 826 827 void markPSInputEnabled(unsigned Index) { 828 PSInputEnable |= 1 << Index; 829 } 830 831 bool returnsVoid() const { 832 return ReturnsVoid; 833 } 834 835 void setIfReturnsVoid(bool Value) { 836 ReturnsVoid = Value; 837 } 838 839 /// \returns A pair of default/requested minimum/maximum flat work group sizes 840 /// for this function. 841 std::pair<unsigned, unsigned> getFlatWorkGroupSizes() const { 842 return FlatWorkGroupSizes; 843 } 844 845 /// \returns Default/requested minimum flat work group size for this function. 846 unsigned getMinFlatWorkGroupSize() const { 847 return FlatWorkGroupSizes.first; 848 } 849 850 /// \returns Default/requested maximum flat work group size for this function. 851 unsigned getMaxFlatWorkGroupSize() const { 852 return FlatWorkGroupSizes.second; 853 } 854 855 /// \returns A pair of default/requested minimum/maximum number of waves per 856 /// execution unit. 857 std::pair<unsigned, unsigned> getWavesPerEU() const { 858 return WavesPerEU; 859 } 860 861 /// \returns Default/requested minimum number of waves per execution unit. 862 unsigned getMinWavesPerEU() const { 863 return WavesPerEU.first; 864 } 865 866 /// \returns Default/requested maximum number of waves per execution unit. 867 unsigned getMaxWavesPerEU() const { 868 return WavesPerEU.second; 869 } 870 871 /// \returns SGPR used for \p Dim's work group ID. 872 Register getWorkGroupIDSGPR(unsigned Dim) const { 873 switch (Dim) { 874 case 0: 875 assert(hasWorkGroupIDX()); 876 return ArgInfo.WorkGroupIDX.getRegister(); 877 case 1: 878 assert(hasWorkGroupIDY()); 879 return ArgInfo.WorkGroupIDY.getRegister(); 880 case 2: 881 assert(hasWorkGroupIDZ()); 882 return ArgInfo.WorkGroupIDZ.getRegister(); 883 } 884 llvm_unreachable("unexpected dimension"); 885 } 886 887 unsigned getLDSWaveSpillSize() const { 888 return LDSWaveSpillSize; 889 } 890 891 const AMDGPUBufferPseudoSourceValue *getBufferPSV(const SIInstrInfo &TII, 892 const Value *BufferRsrc) { 893 assert(BufferRsrc); 894 auto PSV = BufferPSVs.try_emplace( 895 BufferRsrc, 896 std::make_unique<AMDGPUBufferPseudoSourceValue>(TII)); 897 return PSV.first->second.get(); 898 } 899 900 const AMDGPUImagePseudoSourceValue *getImagePSV(const SIInstrInfo &TII, 901 const Value *ImgRsrc) { 902 assert(ImgRsrc); 903 auto PSV = ImagePSVs.try_emplace( 904 ImgRsrc, 905 std::make_unique<AMDGPUImagePseudoSourceValue>(TII)); 906 return PSV.first->second.get(); 907 } 908 909 const AMDGPUGWSResourcePseudoSourceValue *getGWSPSV(const SIInstrInfo &TII) { 910 if (!GWSResourcePSV) { 911 GWSResourcePSV = 912 std::make_unique<AMDGPUGWSResourcePseudoSourceValue>(TII); 913 } 914 915 return GWSResourcePSV.get(); 916 } 917 918 unsigned getOccupancy() const { 919 return Occupancy; 920 } 921 922 unsigned getMinAllowedOccupancy() const { 923 if (!isMemoryBound() && !needsWaveLimiter()) 924 return Occupancy; 925 return (Occupancy < 4) ? Occupancy : 4; 926 } 927 928 void limitOccupancy(const MachineFunction &MF); 929 930 void limitOccupancy(unsigned Limit) { 931 if (Occupancy > Limit) 932 Occupancy = Limit; 933 } 934 935 void increaseOccupancy(const MachineFunction &MF, unsigned Limit) { 936 if (Occupancy < Limit) 937 Occupancy = Limit; 938 limitOccupancy(MF); 939 } 940 }; 941 942 } // end namespace llvm 943 944 #endif // LLVM_LIB_TARGET_AMDGPU_SIMACHINEFUNCTIONINFO_H 945