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 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 // State of MODE register, assumed FP mode. 344 AMDGPU::SIModeRegisterDefaults Mode; 345 346 // Graphics info. 347 unsigned PSInputAddr = 0; 348 unsigned PSInputEnable = 0; 349 350 /// Number of bytes of arguments this function has on the stack. If the callee 351 /// is expected to restore the argument stack this should be a multiple of 16, 352 /// all usable during a tail call. 353 /// 354 /// The alternative would forbid tail call optimisation in some cases: if we 355 /// want to transfer control from a function with 8-bytes of stack-argument 356 /// space to a function with 16-bytes then misalignment of this value would 357 /// make a stack adjustment necessary, which could not be undone by the 358 /// callee. 359 unsigned BytesInStackArgArea = 0; 360 361 bool ReturnsVoid = true; 362 363 // A pair of default/requested minimum/maximum flat work group sizes. 364 // Minimum - first, maximum - second. 365 std::pair<unsigned, unsigned> FlatWorkGroupSizes = {0, 0}; 366 367 // A pair of default/requested minimum/maximum number of waves per execution 368 // unit. Minimum - first, maximum - second. 369 std::pair<unsigned, unsigned> WavesPerEU = {0, 0}; 370 371 DenseMap<const Value *, 372 std::unique_ptr<const AMDGPUBufferPseudoSourceValue>> BufferPSVs; 373 DenseMap<const Value *, 374 std::unique_ptr<const AMDGPUImagePseudoSourceValue>> ImagePSVs; 375 std::unique_ptr<const AMDGPUGWSResourcePseudoSourceValue> GWSResourcePSV; 376 377 private: 378 unsigned LDSWaveSpillSize = 0; 379 unsigned NumUserSGPRs = 0; 380 unsigned NumSystemSGPRs = 0; 381 382 bool HasSpilledSGPRs = false; 383 bool HasSpilledVGPRs = false; 384 bool HasNonSpillStackObjects = false; 385 bool IsStackRealigned = false; 386 387 unsigned NumSpilledSGPRs = 0; 388 unsigned NumSpilledVGPRs = 0; 389 390 // Feature bits required for inputs passed in user SGPRs. 391 bool PrivateSegmentBuffer : 1; 392 bool DispatchPtr : 1; 393 bool QueuePtr : 1; 394 bool KernargSegmentPtr : 1; 395 bool DispatchID : 1; 396 bool FlatScratchInit : 1; 397 398 // Feature bits required for inputs passed in system SGPRs. 399 bool WorkGroupIDX : 1; // Always initialized. 400 bool WorkGroupIDY : 1; 401 bool WorkGroupIDZ : 1; 402 bool WorkGroupInfo : 1; 403 bool PrivateSegmentWaveByteOffset : 1; 404 405 bool WorkItemIDX : 1; // Always initialized. 406 bool WorkItemIDY : 1; 407 bool WorkItemIDZ : 1; 408 409 // Private memory buffer 410 // Compute directly in sgpr[0:1] 411 // Other shaders indirect 64-bits at sgpr[0:1] 412 bool ImplicitBufferPtr : 1; 413 414 // Pointer to where the ABI inserts special kernel arguments separate from the 415 // user arguments. This is an offset from the KernargSegmentPtr. 416 bool ImplicitArgPtr : 1; 417 418 // The hard-wired high half of the address of the global information table 419 // for AMDPAL OS type. 0xffffffff represents no hard-wired high half, since 420 // current hardware only allows a 16 bit value. 421 unsigned GITPtrHigh; 422 423 unsigned HighBitsOf32BitAddress; 424 unsigned GDSSize; 425 426 // Current recorded maximum possible occupancy. 427 unsigned Occupancy; 428 429 MCPhysReg getNextUserSGPR() const; 430 431 MCPhysReg getNextSystemSGPR() const; 432 433 public: 434 struct SpilledReg { 435 unsigned VGPR = 0; 436 int Lane = -1; 437 438 SpilledReg() = default; 439 SpilledReg(unsigned R, int L) : VGPR (R), Lane (L) {} 440 441 bool hasLane() { return Lane != -1;} 442 bool hasReg() { return VGPR != 0;} 443 }; 444 445 struct SGPRSpillVGPRCSR { 446 // VGPR used for SGPR spills 447 unsigned VGPR; 448 449 // If the VGPR is a CSR, the stack slot used to save/restore it in the 450 // prolog/epilog. 451 Optional<int> FI; 452 453 SGPRSpillVGPRCSR(unsigned V, Optional<int> F) : VGPR(V), FI(F) {} 454 }; 455 456 struct VGPRSpillToAGPR { 457 SmallVector<MCPhysReg, 32> Lanes; 458 bool FullyAllocated = false; 459 }; 460 461 SparseBitVector<> WWMReservedRegs; 462 463 void ReserveWWMRegister(unsigned reg) { WWMReservedRegs.set(reg); } 464 465 private: 466 // SGPR->VGPR spilling support. 467 using SpillRegMask = std::pair<unsigned, unsigned>; 468 469 // Track VGPR + wave index for each subregister of the SGPR spilled to 470 // frameindex key. 471 DenseMap<int, std::vector<SpilledReg>> SGPRToVGPRSpills; 472 unsigned NumVGPRSpillLanes = 0; 473 SmallVector<SGPRSpillVGPRCSR, 2> SpillVGPRs; 474 475 DenseMap<int, VGPRSpillToAGPR> VGPRToAGPRSpills; 476 477 // AGPRs used for VGPR spills. 478 SmallVector<MCPhysReg, 32> SpillAGPR; 479 480 // VGPRs used for AGPR spills. 481 SmallVector<MCPhysReg, 32> SpillVGPR; 482 483 public: // FIXME 484 /// If this is set, an SGPR used for save/restore of the register used for the 485 /// frame pointer. 486 unsigned SGPRForFPSaveRestoreCopy = 0; 487 Optional<int> FramePointerSaveIndex; 488 489 public: 490 SIMachineFunctionInfo(const MachineFunction &MF); 491 492 bool initializeBaseYamlFields(const yaml::SIMachineFunctionInfo &YamlMFI); 493 494 ArrayRef<SpilledReg> getSGPRToVGPRSpills(int FrameIndex) const { 495 auto I = SGPRToVGPRSpills.find(FrameIndex); 496 return (I == SGPRToVGPRSpills.end()) ? 497 ArrayRef<SpilledReg>() : makeArrayRef(I->second); 498 } 499 500 ArrayRef<SGPRSpillVGPRCSR> getSGPRSpillVGPRs() const { 501 return SpillVGPRs; 502 } 503 504 ArrayRef<MCPhysReg> getAGPRSpillVGPRs() const { 505 return SpillAGPR; 506 } 507 508 ArrayRef<MCPhysReg> getVGPRSpillAGPRs() const { 509 return SpillVGPR; 510 } 511 512 MCPhysReg getVGPRToAGPRSpill(int FrameIndex, unsigned Lane) const { 513 auto I = VGPRToAGPRSpills.find(FrameIndex); 514 return (I == VGPRToAGPRSpills.end()) ? (MCPhysReg)AMDGPU::NoRegister 515 : I->second.Lanes[Lane]; 516 } 517 518 AMDGPU::SIModeRegisterDefaults getMode() const { 519 return Mode; 520 } 521 522 bool haveFreeLanesForSGPRSpill(const MachineFunction &MF, 523 unsigned NumLane) const; 524 bool allocateSGPRSpillToVGPR(MachineFunction &MF, int FI); 525 bool allocateVGPRSpillToAGPR(MachineFunction &MF, int FI, bool isAGPRtoVGPR); 526 void removeDeadFrameIndices(MachineFrameInfo &MFI); 527 528 bool hasCalculatedTID() const { return TIDReg != 0; }; 529 unsigned getTIDReg() const { return TIDReg; }; 530 void setTIDReg(unsigned Reg) { TIDReg = Reg; } 531 532 unsigned getBytesInStackArgArea() const { 533 return BytesInStackArgArea; 534 } 535 536 void setBytesInStackArgArea(unsigned Bytes) { 537 BytesInStackArgArea = Bytes; 538 } 539 540 // Add user SGPRs. 541 unsigned addPrivateSegmentBuffer(const SIRegisterInfo &TRI); 542 unsigned addDispatchPtr(const SIRegisterInfo &TRI); 543 unsigned addQueuePtr(const SIRegisterInfo &TRI); 544 unsigned addKernargSegmentPtr(const SIRegisterInfo &TRI); 545 unsigned addDispatchID(const SIRegisterInfo &TRI); 546 unsigned addFlatScratchInit(const SIRegisterInfo &TRI); 547 unsigned addImplicitBufferPtr(const SIRegisterInfo &TRI); 548 549 // Add system SGPRs. 550 unsigned addWorkGroupIDX() { 551 ArgInfo.WorkGroupIDX = ArgDescriptor::createRegister(getNextSystemSGPR()); 552 NumSystemSGPRs += 1; 553 return ArgInfo.WorkGroupIDX.getRegister(); 554 } 555 556 unsigned addWorkGroupIDY() { 557 ArgInfo.WorkGroupIDY = ArgDescriptor::createRegister(getNextSystemSGPR()); 558 NumSystemSGPRs += 1; 559 return ArgInfo.WorkGroupIDY.getRegister(); 560 } 561 562 unsigned addWorkGroupIDZ() { 563 ArgInfo.WorkGroupIDZ = ArgDescriptor::createRegister(getNextSystemSGPR()); 564 NumSystemSGPRs += 1; 565 return ArgInfo.WorkGroupIDZ.getRegister(); 566 } 567 568 unsigned addWorkGroupInfo() { 569 ArgInfo.WorkGroupInfo = ArgDescriptor::createRegister(getNextSystemSGPR()); 570 NumSystemSGPRs += 1; 571 return ArgInfo.WorkGroupInfo.getRegister(); 572 } 573 574 // Add special VGPR inputs 575 void setWorkItemIDX(ArgDescriptor Arg) { 576 ArgInfo.WorkItemIDX = Arg; 577 } 578 579 void setWorkItemIDY(ArgDescriptor Arg) { 580 ArgInfo.WorkItemIDY = Arg; 581 } 582 583 void setWorkItemIDZ(ArgDescriptor Arg) { 584 ArgInfo.WorkItemIDZ = Arg; 585 } 586 587 unsigned addPrivateSegmentWaveByteOffset() { 588 ArgInfo.PrivateSegmentWaveByteOffset 589 = ArgDescriptor::createRegister(getNextSystemSGPR()); 590 NumSystemSGPRs += 1; 591 return ArgInfo.PrivateSegmentWaveByteOffset.getRegister(); 592 } 593 594 void setPrivateSegmentWaveByteOffset(unsigned Reg) { 595 ArgInfo.PrivateSegmentWaveByteOffset = ArgDescriptor::createRegister(Reg); 596 } 597 598 bool hasPrivateSegmentBuffer() const { 599 return PrivateSegmentBuffer; 600 } 601 602 bool hasDispatchPtr() const { 603 return DispatchPtr; 604 } 605 606 bool hasQueuePtr() const { 607 return QueuePtr; 608 } 609 610 bool hasKernargSegmentPtr() const { 611 return KernargSegmentPtr; 612 } 613 614 bool hasDispatchID() const { 615 return DispatchID; 616 } 617 618 bool hasFlatScratchInit() const { 619 return FlatScratchInit; 620 } 621 622 bool hasWorkGroupIDX() const { 623 return WorkGroupIDX; 624 } 625 626 bool hasWorkGroupIDY() const { 627 return WorkGroupIDY; 628 } 629 630 bool hasWorkGroupIDZ() const { 631 return WorkGroupIDZ; 632 } 633 634 bool hasWorkGroupInfo() const { 635 return WorkGroupInfo; 636 } 637 638 bool hasPrivateSegmentWaveByteOffset() const { 639 return PrivateSegmentWaveByteOffset; 640 } 641 642 bool hasWorkItemIDX() const { 643 return WorkItemIDX; 644 } 645 646 bool hasWorkItemIDY() const { 647 return WorkItemIDY; 648 } 649 650 bool hasWorkItemIDZ() const { 651 return WorkItemIDZ; 652 } 653 654 bool hasImplicitArgPtr() const { 655 return ImplicitArgPtr; 656 } 657 658 bool hasImplicitBufferPtr() const { 659 return ImplicitBufferPtr; 660 } 661 662 AMDGPUFunctionArgInfo &getArgInfo() { 663 return ArgInfo; 664 } 665 666 const AMDGPUFunctionArgInfo &getArgInfo() const { 667 return ArgInfo; 668 } 669 670 std::pair<const ArgDescriptor *, const TargetRegisterClass *> 671 getPreloadedValue(AMDGPUFunctionArgInfo::PreloadedValue Value) const { 672 return ArgInfo.getPreloadedValue(Value); 673 } 674 675 Register getPreloadedReg(AMDGPUFunctionArgInfo::PreloadedValue Value) const { 676 auto Arg = ArgInfo.getPreloadedValue(Value).first; 677 return Arg ? Arg->getRegister() : Register(); 678 } 679 680 unsigned getGITPtrHigh() const { 681 return GITPtrHigh; 682 } 683 684 uint32_t get32BitAddressHighBits() const { 685 return HighBitsOf32BitAddress; 686 } 687 688 unsigned getGDSSize() const { 689 return GDSSize; 690 } 691 692 unsigned getNumUserSGPRs() const { 693 return NumUserSGPRs; 694 } 695 696 unsigned getNumPreloadedSGPRs() const { 697 return NumUserSGPRs + NumSystemSGPRs; 698 } 699 700 unsigned getPrivateSegmentWaveByteOffsetSystemSGPR() const { 701 return ArgInfo.PrivateSegmentWaveByteOffset.getRegister(); 702 } 703 704 /// Returns the physical register reserved for use as the resource 705 /// descriptor for scratch accesses. 706 unsigned getScratchRSrcReg() const { 707 return ScratchRSrcReg; 708 } 709 710 void setScratchRSrcReg(unsigned Reg) { 711 assert(Reg != 0 && "Should never be unset"); 712 ScratchRSrcReg = Reg; 713 } 714 715 unsigned getScratchWaveOffsetReg() const { 716 return ScratchWaveOffsetReg; 717 } 718 719 unsigned getFrameOffsetReg() const { 720 return FrameOffsetReg; 721 } 722 723 void setFrameOffsetReg(unsigned Reg) { 724 assert(Reg != 0 && "Should never be unset"); 725 FrameOffsetReg = Reg; 726 } 727 728 void setStackPtrOffsetReg(unsigned Reg) { 729 assert(Reg != 0 && "Should never be unset"); 730 StackPtrOffsetReg = Reg; 731 } 732 733 // Note the unset value for this is AMDGPU::SP_REG rather than 734 // NoRegister. This is mostly a workaround for MIR tests where state that 735 // can't be directly computed from the function is not preserved in serialized 736 // MIR. 737 unsigned getStackPtrOffsetReg() const { 738 return StackPtrOffsetReg; 739 } 740 741 void setScratchWaveOffsetReg(unsigned Reg) { 742 assert(Reg != 0 && "Should never be unset"); 743 ScratchWaveOffsetReg = Reg; 744 } 745 746 unsigned getQueuePtrUserSGPR() const { 747 return ArgInfo.QueuePtr.getRegister(); 748 } 749 750 unsigned getImplicitBufferPtrUserSGPR() const { 751 return ArgInfo.ImplicitBufferPtr.getRegister(); 752 } 753 754 bool hasSpilledSGPRs() const { 755 return HasSpilledSGPRs; 756 } 757 758 void setHasSpilledSGPRs(bool Spill = true) { 759 HasSpilledSGPRs = Spill; 760 } 761 762 bool hasSpilledVGPRs() const { 763 return HasSpilledVGPRs; 764 } 765 766 void setHasSpilledVGPRs(bool Spill = true) { 767 HasSpilledVGPRs = Spill; 768 } 769 770 bool hasNonSpillStackObjects() const { 771 return HasNonSpillStackObjects; 772 } 773 774 void setHasNonSpillStackObjects(bool StackObject = true) { 775 HasNonSpillStackObjects = StackObject; 776 } 777 778 bool isStackRealigned() const { 779 return IsStackRealigned; 780 } 781 782 void setIsStackRealigned(bool Realigned = true) { 783 IsStackRealigned = Realigned; 784 } 785 786 unsigned getNumSpilledSGPRs() const { 787 return NumSpilledSGPRs; 788 } 789 790 unsigned getNumSpilledVGPRs() const { 791 return NumSpilledVGPRs; 792 } 793 794 void addToSpilledSGPRs(unsigned num) { 795 NumSpilledSGPRs += num; 796 } 797 798 void addToSpilledVGPRs(unsigned num) { 799 NumSpilledVGPRs += num; 800 } 801 802 unsigned getPSInputAddr() const { 803 return PSInputAddr; 804 } 805 806 unsigned getPSInputEnable() const { 807 return PSInputEnable; 808 } 809 810 bool isPSInputAllocated(unsigned Index) const { 811 return PSInputAddr & (1 << Index); 812 } 813 814 void markPSInputAllocated(unsigned Index) { 815 PSInputAddr |= 1 << Index; 816 } 817 818 void markPSInputEnabled(unsigned Index) { 819 PSInputEnable |= 1 << Index; 820 } 821 822 bool returnsVoid() const { 823 return ReturnsVoid; 824 } 825 826 void setIfReturnsVoid(bool Value) { 827 ReturnsVoid = Value; 828 } 829 830 /// \returns A pair of default/requested minimum/maximum flat work group sizes 831 /// for this function. 832 std::pair<unsigned, unsigned> getFlatWorkGroupSizes() const { 833 return FlatWorkGroupSizes; 834 } 835 836 /// \returns Default/requested minimum flat work group size for this function. 837 unsigned getMinFlatWorkGroupSize() const { 838 return FlatWorkGroupSizes.first; 839 } 840 841 /// \returns Default/requested maximum flat work group size for this function. 842 unsigned getMaxFlatWorkGroupSize() const { 843 return FlatWorkGroupSizes.second; 844 } 845 846 /// \returns A pair of default/requested minimum/maximum number of waves per 847 /// execution unit. 848 std::pair<unsigned, unsigned> getWavesPerEU() const { 849 return WavesPerEU; 850 } 851 852 /// \returns Default/requested minimum number of waves per execution unit. 853 unsigned getMinWavesPerEU() const { 854 return WavesPerEU.first; 855 } 856 857 /// \returns Default/requested maximum number of waves per execution unit. 858 unsigned getMaxWavesPerEU() const { 859 return WavesPerEU.second; 860 } 861 862 /// \returns SGPR used for \p Dim's work group ID. 863 unsigned getWorkGroupIDSGPR(unsigned Dim) const { 864 switch (Dim) { 865 case 0: 866 assert(hasWorkGroupIDX()); 867 return ArgInfo.WorkGroupIDX.getRegister(); 868 case 1: 869 assert(hasWorkGroupIDY()); 870 return ArgInfo.WorkGroupIDY.getRegister(); 871 case 2: 872 assert(hasWorkGroupIDZ()); 873 return ArgInfo.WorkGroupIDZ.getRegister(); 874 } 875 llvm_unreachable("unexpected dimension"); 876 } 877 878 unsigned getLDSWaveSpillSize() const { 879 return LDSWaveSpillSize; 880 } 881 882 const AMDGPUBufferPseudoSourceValue *getBufferPSV(const SIInstrInfo &TII, 883 const Value *BufferRsrc) { 884 assert(BufferRsrc); 885 auto PSV = BufferPSVs.try_emplace( 886 BufferRsrc, 887 std::make_unique<AMDGPUBufferPseudoSourceValue>(TII)); 888 return PSV.first->second.get(); 889 } 890 891 const AMDGPUImagePseudoSourceValue *getImagePSV(const SIInstrInfo &TII, 892 const Value *ImgRsrc) { 893 assert(ImgRsrc); 894 auto PSV = ImagePSVs.try_emplace( 895 ImgRsrc, 896 std::make_unique<AMDGPUImagePseudoSourceValue>(TII)); 897 return PSV.first->second.get(); 898 } 899 900 const AMDGPUGWSResourcePseudoSourceValue *getGWSPSV(const SIInstrInfo &TII) { 901 if (!GWSResourcePSV) { 902 GWSResourcePSV = 903 std::make_unique<AMDGPUGWSResourcePseudoSourceValue>(TII); 904 } 905 906 return GWSResourcePSV.get(); 907 } 908 909 unsigned getOccupancy() const { 910 return Occupancy; 911 } 912 913 unsigned getMinAllowedOccupancy() const { 914 if (!isMemoryBound() && !needsWaveLimiter()) 915 return Occupancy; 916 return (Occupancy < 4) ? Occupancy : 4; 917 } 918 919 void limitOccupancy(const MachineFunction &MF); 920 921 void limitOccupancy(unsigned Limit) { 922 if (Occupancy > Limit) 923 Occupancy = Limit; 924 } 925 926 void increaseOccupancy(const MachineFunction &MF, unsigned Limit) { 927 if (Occupancy < Limit) 928 Occupancy = Limit; 929 limitOccupancy(MF); 930 } 931 }; 932 933 } // end namespace llvm 934 935 #endif // LLVM_LIB_TARGET_AMDGPU_SIMACHINEFUNCTIONINFO_H 936