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 "AMDGPUTargetMachine.h" 19 #include "MCTargetDesc/AMDGPUMCTargetDesc.h" 20 #include "SIInstrInfo.h" 21 #include "llvm/ADT/SetVector.h" 22 #include "llvm/CodeGen/MIRYamlMapping.h" 23 #include "llvm/CodeGen/PseudoSourceValue.h" 24 #include "llvm/Support/raw_ostream.h" 25 26 namespace llvm { 27 28 class MachineFrameInfo; 29 class MachineFunction; 30 class SIMachineFunctionInfo; 31 class SIRegisterInfo; 32 class TargetRegisterClass; 33 34 class AMDGPUPseudoSourceValue : public PseudoSourceValue { 35 public: 36 enum AMDGPUPSVKind : unsigned { 37 PSVBuffer = PseudoSourceValue::TargetCustom, 38 PSVImage, 39 GWSResource 40 }; 41 42 protected: 43 AMDGPUPseudoSourceValue(unsigned Kind, const AMDGPUTargetMachine &TM) 44 : PseudoSourceValue(Kind, TM) {} 45 46 public: 47 bool isConstant(const MachineFrameInfo *) const override { 48 // This should probably be true for most images, but we will start by being 49 // conservative. 50 return false; 51 } 52 53 bool isAliased(const MachineFrameInfo *) const override { 54 return true; 55 } 56 57 bool mayAlias(const MachineFrameInfo *) const override { 58 return true; 59 } 60 }; 61 62 class AMDGPUBufferPseudoSourceValue final : public AMDGPUPseudoSourceValue { 63 public: 64 explicit AMDGPUBufferPseudoSourceValue(const AMDGPUTargetMachine &TM) 65 : AMDGPUPseudoSourceValue(PSVBuffer, TM) {} 66 67 static bool classof(const PseudoSourceValue *V) { 68 return V->kind() == PSVBuffer; 69 } 70 71 void printCustom(raw_ostream &OS) const override { OS << "BufferResource"; } 72 }; 73 74 class AMDGPUImagePseudoSourceValue final : public AMDGPUPseudoSourceValue { 75 public: 76 // TODO: Is the img rsrc useful? 77 explicit AMDGPUImagePseudoSourceValue(const AMDGPUTargetMachine &TM) 78 : AMDGPUPseudoSourceValue(PSVImage, TM) {} 79 80 static bool classof(const PseudoSourceValue *V) { 81 return V->kind() == PSVImage; 82 } 83 84 void printCustom(raw_ostream &OS) const override { OS << "ImageResource"; } 85 }; 86 87 class AMDGPUGWSResourcePseudoSourceValue final : public AMDGPUPseudoSourceValue { 88 public: 89 explicit AMDGPUGWSResourcePseudoSourceValue(const AMDGPUTargetMachine &TM) 90 : AMDGPUPseudoSourceValue(GWSResource, TM) {} 91 92 static bool classof(const PseudoSourceValue *V) { 93 return V->kind() == GWSResource; 94 } 95 96 // These are inaccessible memory from IR. 97 bool isAliased(const MachineFrameInfo *) const override { 98 return false; 99 } 100 101 // These are inaccessible memory from IR. 102 bool mayAlias(const MachineFrameInfo *) const override { 103 return false; 104 } 105 106 void printCustom(raw_ostream &OS) const override { 107 OS << "GWSResource"; 108 } 109 }; 110 111 namespace yaml { 112 113 struct SIArgument { 114 bool IsRegister; 115 union { 116 StringValue RegisterName; 117 unsigned StackOffset; 118 }; 119 Optional<unsigned> Mask; 120 121 // Default constructor, which creates a stack argument. 122 SIArgument() : IsRegister(false), StackOffset(0) {} 123 SIArgument(const SIArgument &Other) { 124 IsRegister = Other.IsRegister; 125 if (IsRegister) { 126 ::new ((void *)std::addressof(RegisterName)) 127 StringValue(Other.RegisterName); 128 } else 129 StackOffset = Other.StackOffset; 130 Mask = Other.Mask; 131 } 132 SIArgument &operator=(const SIArgument &Other) { 133 IsRegister = Other.IsRegister; 134 if (IsRegister) { 135 ::new ((void *)std::addressof(RegisterName)) 136 StringValue(Other.RegisterName); 137 } else 138 StackOffset = Other.StackOffset; 139 Mask = Other.Mask; 140 return *this; 141 } 142 ~SIArgument() { 143 if (IsRegister) 144 RegisterName.~StringValue(); 145 } 146 147 // Helper to create a register or stack argument. 148 static inline SIArgument createArgument(bool IsReg) { 149 if (IsReg) 150 return SIArgument(IsReg); 151 return SIArgument(); 152 } 153 154 private: 155 // Construct a register argument. 156 SIArgument(bool) : IsRegister(true), RegisterName() {} 157 }; 158 159 template <> struct MappingTraits<SIArgument> { 160 static void mapping(IO &YamlIO, SIArgument &A) { 161 if (YamlIO.outputting()) { 162 if (A.IsRegister) 163 YamlIO.mapRequired("reg", A.RegisterName); 164 else 165 YamlIO.mapRequired("offset", A.StackOffset); 166 } else { 167 auto Keys = YamlIO.keys(); 168 if (is_contained(Keys, "reg")) { 169 A = SIArgument::createArgument(true); 170 YamlIO.mapRequired("reg", A.RegisterName); 171 } else if (is_contained(Keys, "offset")) 172 YamlIO.mapRequired("offset", A.StackOffset); 173 else 174 YamlIO.setError("missing required key 'reg' or 'offset'"); 175 } 176 YamlIO.mapOptional("mask", A.Mask); 177 } 178 static const bool flow = true; 179 }; 180 181 struct SIArgumentInfo { 182 Optional<SIArgument> PrivateSegmentBuffer; 183 Optional<SIArgument> DispatchPtr; 184 Optional<SIArgument> QueuePtr; 185 Optional<SIArgument> KernargSegmentPtr; 186 Optional<SIArgument> DispatchID; 187 Optional<SIArgument> FlatScratchInit; 188 Optional<SIArgument> PrivateSegmentSize; 189 190 Optional<SIArgument> WorkGroupIDX; 191 Optional<SIArgument> WorkGroupIDY; 192 Optional<SIArgument> WorkGroupIDZ; 193 Optional<SIArgument> WorkGroupInfo; 194 Optional<SIArgument> PrivateSegmentWaveByteOffset; 195 196 Optional<SIArgument> ImplicitArgPtr; 197 Optional<SIArgument> ImplicitBufferPtr; 198 199 Optional<SIArgument> WorkItemIDX; 200 Optional<SIArgument> WorkItemIDY; 201 Optional<SIArgument> WorkItemIDZ; 202 }; 203 204 template <> struct MappingTraits<SIArgumentInfo> { 205 static void mapping(IO &YamlIO, SIArgumentInfo &AI) { 206 YamlIO.mapOptional("privateSegmentBuffer", AI.PrivateSegmentBuffer); 207 YamlIO.mapOptional("dispatchPtr", AI.DispatchPtr); 208 YamlIO.mapOptional("queuePtr", AI.QueuePtr); 209 YamlIO.mapOptional("kernargSegmentPtr", AI.KernargSegmentPtr); 210 YamlIO.mapOptional("dispatchID", AI.DispatchID); 211 YamlIO.mapOptional("flatScratchInit", AI.FlatScratchInit); 212 YamlIO.mapOptional("privateSegmentSize", AI.PrivateSegmentSize); 213 214 YamlIO.mapOptional("workGroupIDX", AI.WorkGroupIDX); 215 YamlIO.mapOptional("workGroupIDY", AI.WorkGroupIDY); 216 YamlIO.mapOptional("workGroupIDZ", AI.WorkGroupIDZ); 217 YamlIO.mapOptional("workGroupInfo", AI.WorkGroupInfo); 218 YamlIO.mapOptional("privateSegmentWaveByteOffset", 219 AI.PrivateSegmentWaveByteOffset); 220 221 YamlIO.mapOptional("implicitArgPtr", AI.ImplicitArgPtr); 222 YamlIO.mapOptional("implicitBufferPtr", AI.ImplicitBufferPtr); 223 224 YamlIO.mapOptional("workItemIDX", AI.WorkItemIDX); 225 YamlIO.mapOptional("workItemIDY", AI.WorkItemIDY); 226 YamlIO.mapOptional("workItemIDZ", AI.WorkItemIDZ); 227 } 228 }; 229 230 // Default to default mode for default calling convention. 231 struct SIMode { 232 bool IEEE = true; 233 bool DX10Clamp = true; 234 bool FP32InputDenormals = true; 235 bool FP32OutputDenormals = true; 236 bool FP64FP16InputDenormals = true; 237 bool FP64FP16OutputDenormals = true; 238 239 SIMode() = default; 240 241 SIMode(const AMDGPU::SIModeRegisterDefaults &Mode) { 242 IEEE = Mode.IEEE; 243 DX10Clamp = Mode.DX10Clamp; 244 FP32InputDenormals = Mode.FP32InputDenormals; 245 FP32OutputDenormals = Mode.FP32OutputDenormals; 246 FP64FP16InputDenormals = Mode.FP64FP16InputDenormals; 247 FP64FP16OutputDenormals = Mode.FP64FP16OutputDenormals; 248 } 249 250 bool operator ==(const SIMode Other) const { 251 return IEEE == Other.IEEE && 252 DX10Clamp == Other.DX10Clamp && 253 FP32InputDenormals == Other.FP32InputDenormals && 254 FP32OutputDenormals == Other.FP32OutputDenormals && 255 FP64FP16InputDenormals == Other.FP64FP16InputDenormals && 256 FP64FP16OutputDenormals == Other.FP64FP16OutputDenormals; 257 } 258 }; 259 260 template <> struct MappingTraits<SIMode> { 261 static void mapping(IO &YamlIO, SIMode &Mode) { 262 YamlIO.mapOptional("ieee", Mode.IEEE, true); 263 YamlIO.mapOptional("dx10-clamp", Mode.DX10Clamp, true); 264 YamlIO.mapOptional("fp32-input-denormals", Mode.FP32InputDenormals, true); 265 YamlIO.mapOptional("fp32-output-denormals", Mode.FP32OutputDenormals, true); 266 YamlIO.mapOptional("fp64-fp16-input-denormals", Mode.FP64FP16InputDenormals, true); 267 YamlIO.mapOptional("fp64-fp16-output-denormals", Mode.FP64FP16OutputDenormals, true); 268 } 269 }; 270 271 struct SIMachineFunctionInfo final : public yaml::MachineFunctionInfo { 272 uint64_t ExplicitKernArgSize = 0; 273 unsigned MaxKernArgAlign = 0; 274 uint32_t LDSSize = 0; 275 uint32_t GDSSize = 0; 276 Align DynLDSAlign; 277 bool IsEntryFunction = false; 278 bool NoSignedZerosFPMath = false; 279 bool MemoryBound = false; 280 bool WaveLimiter = false; 281 bool HasSpilledSGPRs = false; 282 bool HasSpilledVGPRs = false; 283 uint32_t HighBitsOf32BitAddress = 0; 284 285 // TODO: 10 may be a better default since it's the maximum. 286 unsigned Occupancy = 0; 287 288 SmallVector<StringValue> WWMReservedRegs; 289 290 StringValue ScratchRSrcReg = "$private_rsrc_reg"; 291 StringValue FrameOffsetReg = "$fp_reg"; 292 StringValue StackPtrOffsetReg = "$sp_reg"; 293 294 unsigned BytesInStackArgArea = 0; 295 bool ReturnsVoid = true; 296 297 Optional<SIArgumentInfo> ArgInfo; 298 SIMode Mode; 299 Optional<FrameIndex> ScavengeFI; 300 StringValue VGPRForAGPRCopy; 301 302 SIMachineFunctionInfo() = default; 303 SIMachineFunctionInfo(const llvm::SIMachineFunctionInfo &, 304 const TargetRegisterInfo &TRI, 305 const llvm::MachineFunction &MF); 306 307 void mappingImpl(yaml::IO &YamlIO) override; 308 ~SIMachineFunctionInfo() = default; 309 }; 310 311 template <> struct MappingTraits<SIMachineFunctionInfo> { 312 static void mapping(IO &YamlIO, SIMachineFunctionInfo &MFI) { 313 YamlIO.mapOptional("explicitKernArgSize", MFI.ExplicitKernArgSize, 314 UINT64_C(0)); 315 YamlIO.mapOptional("maxKernArgAlign", MFI.MaxKernArgAlign, 0u); 316 YamlIO.mapOptional("ldsSize", MFI.LDSSize, 0u); 317 YamlIO.mapOptional("gdsSize", MFI.GDSSize, 0u); 318 YamlIO.mapOptional("dynLDSAlign", MFI.DynLDSAlign, Align()); 319 YamlIO.mapOptional("isEntryFunction", MFI.IsEntryFunction, false); 320 YamlIO.mapOptional("noSignedZerosFPMath", MFI.NoSignedZerosFPMath, false); 321 YamlIO.mapOptional("memoryBound", MFI.MemoryBound, false); 322 YamlIO.mapOptional("waveLimiter", MFI.WaveLimiter, false); 323 YamlIO.mapOptional("hasSpilledSGPRs", MFI.HasSpilledSGPRs, false); 324 YamlIO.mapOptional("hasSpilledVGPRs", MFI.HasSpilledVGPRs, false); 325 YamlIO.mapOptional("scratchRSrcReg", MFI.ScratchRSrcReg, 326 StringValue("$private_rsrc_reg")); 327 YamlIO.mapOptional("frameOffsetReg", MFI.FrameOffsetReg, 328 StringValue("$fp_reg")); 329 YamlIO.mapOptional("stackPtrOffsetReg", MFI.StackPtrOffsetReg, 330 StringValue("$sp_reg")); 331 YamlIO.mapOptional("bytesInStackArgArea", MFI.BytesInStackArgArea, 0u); 332 YamlIO.mapOptional("returnsVoid", MFI.ReturnsVoid, true); 333 YamlIO.mapOptional("argumentInfo", MFI.ArgInfo); 334 YamlIO.mapOptional("mode", MFI.Mode, SIMode()); 335 YamlIO.mapOptional("highBitsOf32BitAddress", 336 MFI.HighBitsOf32BitAddress, 0u); 337 YamlIO.mapOptional("occupancy", MFI.Occupancy, 0); 338 YamlIO.mapOptional("wwmReservedRegs", MFI.WWMReservedRegs); 339 YamlIO.mapOptional("scavengeFI", MFI.ScavengeFI); 340 YamlIO.mapOptional("vgprForAGPRCopy", MFI.VGPRForAGPRCopy, 341 StringValue()); // Don't print out when it's empty. 342 } 343 }; 344 345 } // end namespace yaml 346 347 /// This class keeps track of the SPI_SP_INPUT_ADDR config register, which 348 /// tells the hardware which interpolation parameters to load. 349 class SIMachineFunctionInfo final : public AMDGPUMachineFunction { 350 friend class GCNTargetMachine; 351 352 // Registers that may be reserved for spilling purposes. These may be the same 353 // as the input registers. 354 Register ScratchRSrcReg = AMDGPU::PRIVATE_RSRC_REG; 355 356 // This is the the unswizzled offset from the current dispatch's scratch wave 357 // base to the beginning of the current function's frame. 358 Register FrameOffsetReg = AMDGPU::FP_REG; 359 360 // This is an ABI register used in the non-entry calling convention to 361 // communicate the unswizzled offset from the current dispatch's scratch wave 362 // base to the beginning of the new function's frame. 363 Register StackPtrOffsetReg = AMDGPU::SP_REG; 364 365 AMDGPUFunctionArgInfo ArgInfo; 366 367 // Graphics info. 368 unsigned PSInputAddr = 0; 369 unsigned PSInputEnable = 0; 370 371 /// Number of bytes of arguments this function has on the stack. If the callee 372 /// is expected to restore the argument stack this should be a multiple of 16, 373 /// all usable during a tail call. 374 /// 375 /// The alternative would forbid tail call optimisation in some cases: if we 376 /// want to transfer control from a function with 8-bytes of stack-argument 377 /// space to a function with 16-bytes then misalignment of this value would 378 /// make a stack adjustment necessary, which could not be undone by the 379 /// callee. 380 unsigned BytesInStackArgArea = 0; 381 382 bool ReturnsVoid = true; 383 384 // A pair of default/requested minimum/maximum flat work group sizes. 385 // Minimum - first, maximum - second. 386 std::pair<unsigned, unsigned> FlatWorkGroupSizes = {0, 0}; 387 388 // A pair of default/requested minimum/maximum number of waves per execution 389 // unit. Minimum - first, maximum - second. 390 std::pair<unsigned, unsigned> WavesPerEU = {0, 0}; 391 392 std::unique_ptr<const AMDGPUBufferPseudoSourceValue> BufferPSV; 393 std::unique_ptr<const AMDGPUImagePseudoSourceValue> ImagePSV; 394 std::unique_ptr<const AMDGPUGWSResourcePseudoSourceValue> GWSResourcePSV; 395 396 private: 397 unsigned NumUserSGPRs = 0; 398 unsigned NumSystemSGPRs = 0; 399 400 bool HasSpilledSGPRs = false; 401 bool HasSpilledVGPRs = false; 402 bool HasNonSpillStackObjects = false; 403 bool IsStackRealigned = false; 404 405 unsigned NumSpilledSGPRs = 0; 406 unsigned NumSpilledVGPRs = 0; 407 408 // Feature bits required for inputs passed in user SGPRs. 409 bool PrivateSegmentBuffer : 1; 410 bool DispatchPtr : 1; 411 bool QueuePtr : 1; 412 bool KernargSegmentPtr : 1; 413 bool DispatchID : 1; 414 bool FlatScratchInit : 1; 415 416 // Feature bits required for inputs passed in system SGPRs. 417 bool WorkGroupIDX : 1; // Always initialized. 418 bool WorkGroupIDY : 1; 419 bool WorkGroupIDZ : 1; 420 bool WorkGroupInfo : 1; 421 bool PrivateSegmentWaveByteOffset : 1; 422 423 bool WorkItemIDX : 1; // Always initialized. 424 bool WorkItemIDY : 1; 425 bool WorkItemIDZ : 1; 426 427 // Private memory buffer 428 // Compute directly in sgpr[0:1] 429 // Other shaders indirect 64-bits at sgpr[0:1] 430 bool ImplicitBufferPtr : 1; 431 432 // Pointer to where the ABI inserts special kernel arguments separate from the 433 // user arguments. This is an offset from the KernargSegmentPtr. 434 bool ImplicitArgPtr : 1; 435 436 bool MayNeedAGPRs : 1; 437 438 // The hard-wired high half of the address of the global information table 439 // for AMDPAL OS type. 0xffffffff represents no hard-wired high half, since 440 // current hardware only allows a 16 bit value. 441 unsigned GITPtrHigh; 442 443 unsigned HighBitsOf32BitAddress; 444 445 // Current recorded maximum possible occupancy. 446 unsigned Occupancy; 447 448 mutable Optional<bool> UsesAGPRs; 449 450 MCPhysReg getNextUserSGPR() const; 451 452 MCPhysReg getNextSystemSGPR() const; 453 454 public: 455 struct SpilledReg { 456 Register VGPR; 457 int Lane = -1; 458 459 SpilledReg() = default; 460 SpilledReg(Register R, int L) : VGPR (R), Lane (L) {} 461 462 bool hasLane() { return Lane != -1;} 463 bool hasReg() { return VGPR != 0;} 464 }; 465 466 struct SGPRSpillVGPR { 467 // VGPR used for SGPR spills 468 Register VGPR; 469 470 // If the VGPR is is used for SGPR spills in a non-entrypoint function, the 471 // stack slot used to save/restore it in the prolog/epilog. 472 Optional<int> FI; 473 474 SGPRSpillVGPR(Register V, Optional<int> F) : VGPR(V), FI(F) {} 475 }; 476 477 struct VGPRSpillToAGPR { 478 SmallVector<MCPhysReg, 32> Lanes; 479 bool FullyAllocated = false; 480 bool IsDead = false; 481 }; 482 483 // Track VGPRs reserved for WWM. 484 SmallSetVector<Register, 8> WWMReservedRegs; 485 486 /// Track stack slots used for save/restore of reserved WWM VGPRs in the 487 /// prolog/epilog. 488 489 /// FIXME: This is temporary state only needed in PrologEpilogInserter, and 490 /// doesn't really belong here. It does not require serialization 491 SmallVector<int, 8> WWMReservedFrameIndexes; 492 493 void allocateWWMReservedSpillSlots(MachineFrameInfo &MFI, 494 const SIRegisterInfo &TRI); 495 496 auto wwmAllocation() const { 497 assert(WWMReservedRegs.size() == WWMReservedFrameIndexes.size()); 498 return zip(WWMReservedRegs, WWMReservedFrameIndexes); 499 } 500 501 private: 502 // Track VGPR + wave index for each subregister of the SGPR spilled to 503 // frameindex key. 504 DenseMap<int, std::vector<SpilledReg>> SGPRToVGPRSpills; 505 unsigned NumVGPRSpillLanes = 0; 506 SmallVector<SGPRSpillVGPR, 2> SpillVGPRs; 507 508 DenseMap<int, VGPRSpillToAGPR> VGPRToAGPRSpills; 509 510 // AGPRs used for VGPR spills. 511 SmallVector<MCPhysReg, 32> SpillAGPR; 512 513 // VGPRs used for AGPR spills. 514 SmallVector<MCPhysReg, 32> SpillVGPR; 515 516 // Emergency stack slot. Sometimes, we create this before finalizing the stack 517 // frame, so save it here and add it to the RegScavenger later. 518 Optional<int> ScavengeFI; 519 520 private: 521 Register VGPRForAGPRCopy; 522 523 public: 524 Register getVGPRForAGPRCopy() const { 525 return VGPRForAGPRCopy; 526 } 527 528 void setVGPRForAGPRCopy(Register NewVGPRForAGPRCopy) { 529 VGPRForAGPRCopy = NewVGPRForAGPRCopy; 530 } 531 532 public: // FIXME 533 /// If this is set, an SGPR used for save/restore of the register used for the 534 /// frame pointer. 535 Register SGPRForFPSaveRestoreCopy; 536 Optional<int> FramePointerSaveIndex; 537 538 /// If this is set, an SGPR used for save/restore of the register used for the 539 /// base pointer. 540 Register SGPRForBPSaveRestoreCopy; 541 Optional<int> BasePointerSaveIndex; 542 543 bool isCalleeSavedReg(const MCPhysReg *CSRegs, MCPhysReg Reg); 544 545 public: 546 SIMachineFunctionInfo(const MachineFunction &MF); 547 548 bool initializeBaseYamlFields(const yaml::SIMachineFunctionInfo &YamlMFI, 549 const MachineFunction &MF, 550 PerFunctionMIParsingState &PFS, 551 SMDiagnostic &Error, SMRange &SourceRange); 552 553 void reserveWWMRegister(Register Reg) { 554 WWMReservedRegs.insert(Reg); 555 } 556 557 ArrayRef<SpilledReg> getSGPRToVGPRSpills(int FrameIndex) const { 558 auto I = SGPRToVGPRSpills.find(FrameIndex); 559 return (I == SGPRToVGPRSpills.end()) ? 560 ArrayRef<SpilledReg>() : makeArrayRef(I->second); 561 } 562 563 ArrayRef<SGPRSpillVGPR> getSGPRSpillVGPRs() const { return SpillVGPRs; } 564 565 ArrayRef<MCPhysReg> getAGPRSpillVGPRs() const { 566 return SpillAGPR; 567 } 568 569 ArrayRef<MCPhysReg> getVGPRSpillAGPRs() const { 570 return SpillVGPR; 571 } 572 573 MCPhysReg getVGPRToAGPRSpill(int FrameIndex, unsigned Lane) const { 574 auto I = VGPRToAGPRSpills.find(FrameIndex); 575 return (I == VGPRToAGPRSpills.end()) ? (MCPhysReg)AMDGPU::NoRegister 576 : I->second.Lanes[Lane]; 577 } 578 579 void setVGPRToAGPRSpillDead(int FrameIndex) { 580 auto I = VGPRToAGPRSpills.find(FrameIndex); 581 if (I != VGPRToAGPRSpills.end()) 582 I->second.IsDead = true; 583 } 584 585 bool haveFreeLanesForSGPRSpill(const MachineFunction &MF, 586 unsigned NumLane) const; 587 bool allocateSGPRSpillToVGPR(MachineFunction &MF, int FI); 588 bool allocateVGPRSpillToAGPR(MachineFunction &MF, int FI, bool isAGPRtoVGPR); 589 590 /// If \p ResetSGPRSpillStackIDs is true, reset the stack ID from sgpr-spill 591 /// to the default stack. 592 bool removeDeadFrameIndices(MachineFrameInfo &MFI, 593 bool ResetSGPRSpillStackIDs); 594 595 int getScavengeFI(MachineFrameInfo &MFI, const SIRegisterInfo &TRI); 596 Optional<int> getOptionalScavengeFI() const { return ScavengeFI; } 597 598 unsigned getBytesInStackArgArea() const { 599 return BytesInStackArgArea; 600 } 601 602 void setBytesInStackArgArea(unsigned Bytes) { 603 BytesInStackArgArea = Bytes; 604 } 605 606 // Add user SGPRs. 607 Register addPrivateSegmentBuffer(const SIRegisterInfo &TRI); 608 Register addDispatchPtr(const SIRegisterInfo &TRI); 609 Register addQueuePtr(const SIRegisterInfo &TRI); 610 Register addKernargSegmentPtr(const SIRegisterInfo &TRI); 611 Register addDispatchID(const SIRegisterInfo &TRI); 612 Register addFlatScratchInit(const SIRegisterInfo &TRI); 613 Register addImplicitBufferPtr(const SIRegisterInfo &TRI); 614 615 // Add system SGPRs. 616 Register addWorkGroupIDX() { 617 ArgInfo.WorkGroupIDX = ArgDescriptor::createRegister(getNextSystemSGPR()); 618 NumSystemSGPRs += 1; 619 return ArgInfo.WorkGroupIDX.getRegister(); 620 } 621 622 Register addWorkGroupIDY() { 623 ArgInfo.WorkGroupIDY = ArgDescriptor::createRegister(getNextSystemSGPR()); 624 NumSystemSGPRs += 1; 625 return ArgInfo.WorkGroupIDY.getRegister(); 626 } 627 628 Register addWorkGroupIDZ() { 629 ArgInfo.WorkGroupIDZ = ArgDescriptor::createRegister(getNextSystemSGPR()); 630 NumSystemSGPRs += 1; 631 return ArgInfo.WorkGroupIDZ.getRegister(); 632 } 633 634 Register addWorkGroupInfo() { 635 ArgInfo.WorkGroupInfo = ArgDescriptor::createRegister(getNextSystemSGPR()); 636 NumSystemSGPRs += 1; 637 return ArgInfo.WorkGroupInfo.getRegister(); 638 } 639 640 // Add special VGPR inputs 641 void setWorkItemIDX(ArgDescriptor Arg) { 642 ArgInfo.WorkItemIDX = Arg; 643 } 644 645 void setWorkItemIDY(ArgDescriptor Arg) { 646 ArgInfo.WorkItemIDY = Arg; 647 } 648 649 void setWorkItemIDZ(ArgDescriptor Arg) { 650 ArgInfo.WorkItemIDZ = Arg; 651 } 652 653 Register addPrivateSegmentWaveByteOffset() { 654 ArgInfo.PrivateSegmentWaveByteOffset 655 = ArgDescriptor::createRegister(getNextSystemSGPR()); 656 NumSystemSGPRs += 1; 657 return ArgInfo.PrivateSegmentWaveByteOffset.getRegister(); 658 } 659 660 void setPrivateSegmentWaveByteOffset(Register Reg) { 661 ArgInfo.PrivateSegmentWaveByteOffset = ArgDescriptor::createRegister(Reg); 662 } 663 664 bool hasPrivateSegmentBuffer() const { 665 return PrivateSegmentBuffer; 666 } 667 668 bool hasDispatchPtr() const { 669 return DispatchPtr; 670 } 671 672 bool hasQueuePtr() const { 673 return QueuePtr; 674 } 675 676 bool hasKernargSegmentPtr() const { 677 return KernargSegmentPtr; 678 } 679 680 bool hasDispatchID() const { 681 return DispatchID; 682 } 683 684 bool hasFlatScratchInit() const { 685 return FlatScratchInit; 686 } 687 688 bool hasWorkGroupIDX() const { 689 return WorkGroupIDX; 690 } 691 692 bool hasWorkGroupIDY() const { 693 return WorkGroupIDY; 694 } 695 696 bool hasWorkGroupIDZ() const { 697 return WorkGroupIDZ; 698 } 699 700 bool hasWorkGroupInfo() const { 701 return WorkGroupInfo; 702 } 703 704 bool hasPrivateSegmentWaveByteOffset() const { 705 return PrivateSegmentWaveByteOffset; 706 } 707 708 bool hasWorkItemIDX() const { 709 return WorkItemIDX; 710 } 711 712 bool hasWorkItemIDY() const { 713 return WorkItemIDY; 714 } 715 716 bool hasWorkItemIDZ() const { 717 return WorkItemIDZ; 718 } 719 720 bool hasImplicitArgPtr() const { 721 return ImplicitArgPtr; 722 } 723 724 bool hasImplicitBufferPtr() const { 725 return ImplicitBufferPtr; 726 } 727 728 AMDGPUFunctionArgInfo &getArgInfo() { 729 return ArgInfo; 730 } 731 732 const AMDGPUFunctionArgInfo &getArgInfo() const { 733 return ArgInfo; 734 } 735 736 std::tuple<const ArgDescriptor *, const TargetRegisterClass *, LLT> 737 getPreloadedValue(AMDGPUFunctionArgInfo::PreloadedValue Value) const { 738 return ArgInfo.getPreloadedValue(Value); 739 } 740 741 MCRegister getPreloadedReg(AMDGPUFunctionArgInfo::PreloadedValue Value) const { 742 auto Arg = std::get<0>(ArgInfo.getPreloadedValue(Value)); 743 return Arg ? Arg->getRegister() : MCRegister(); 744 } 745 746 unsigned getGITPtrHigh() const { 747 return GITPtrHigh; 748 } 749 750 Register getGITPtrLoReg(const MachineFunction &MF) const; 751 752 uint32_t get32BitAddressHighBits() const { 753 return HighBitsOf32BitAddress; 754 } 755 756 unsigned getNumUserSGPRs() const { 757 return NumUserSGPRs; 758 } 759 760 unsigned getNumPreloadedSGPRs() const { 761 return NumUserSGPRs + NumSystemSGPRs; 762 } 763 764 Register getPrivateSegmentWaveByteOffsetSystemSGPR() const { 765 return ArgInfo.PrivateSegmentWaveByteOffset.getRegister(); 766 } 767 768 /// Returns the physical register reserved for use as the resource 769 /// descriptor for scratch accesses. 770 Register getScratchRSrcReg() const { 771 return ScratchRSrcReg; 772 } 773 774 void setScratchRSrcReg(Register Reg) { 775 assert(Reg != 0 && "Should never be unset"); 776 ScratchRSrcReg = Reg; 777 } 778 779 Register getFrameOffsetReg() const { 780 return FrameOffsetReg; 781 } 782 783 void setFrameOffsetReg(Register Reg) { 784 assert(Reg != 0 && "Should never be unset"); 785 FrameOffsetReg = Reg; 786 } 787 788 void setStackPtrOffsetReg(Register Reg) { 789 assert(Reg != 0 && "Should never be unset"); 790 StackPtrOffsetReg = Reg; 791 } 792 793 // Note the unset value for this is AMDGPU::SP_REG rather than 794 // NoRegister. This is mostly a workaround for MIR tests where state that 795 // can't be directly computed from the function is not preserved in serialized 796 // MIR. 797 Register getStackPtrOffsetReg() const { 798 return StackPtrOffsetReg; 799 } 800 801 Register getQueuePtrUserSGPR() const { 802 return ArgInfo.QueuePtr.getRegister(); 803 } 804 805 Register getImplicitBufferPtrUserSGPR() const { 806 return ArgInfo.ImplicitBufferPtr.getRegister(); 807 } 808 809 bool hasSpilledSGPRs() const { 810 return HasSpilledSGPRs; 811 } 812 813 void setHasSpilledSGPRs(bool Spill = true) { 814 HasSpilledSGPRs = Spill; 815 } 816 817 bool hasSpilledVGPRs() const { 818 return HasSpilledVGPRs; 819 } 820 821 void setHasSpilledVGPRs(bool Spill = true) { 822 HasSpilledVGPRs = Spill; 823 } 824 825 bool hasNonSpillStackObjects() const { 826 return HasNonSpillStackObjects; 827 } 828 829 void setHasNonSpillStackObjects(bool StackObject = true) { 830 HasNonSpillStackObjects = StackObject; 831 } 832 833 bool isStackRealigned() const { 834 return IsStackRealigned; 835 } 836 837 void setIsStackRealigned(bool Realigned = true) { 838 IsStackRealigned = Realigned; 839 } 840 841 unsigned getNumSpilledSGPRs() const { 842 return NumSpilledSGPRs; 843 } 844 845 unsigned getNumSpilledVGPRs() const { 846 return NumSpilledVGPRs; 847 } 848 849 void addToSpilledSGPRs(unsigned num) { 850 NumSpilledSGPRs += num; 851 } 852 853 void addToSpilledVGPRs(unsigned num) { 854 NumSpilledVGPRs += num; 855 } 856 857 unsigned getPSInputAddr() const { 858 return PSInputAddr; 859 } 860 861 unsigned getPSInputEnable() const { 862 return PSInputEnable; 863 } 864 865 bool isPSInputAllocated(unsigned Index) const { 866 return PSInputAddr & (1 << Index); 867 } 868 869 void markPSInputAllocated(unsigned Index) { 870 PSInputAddr |= 1 << Index; 871 } 872 873 void markPSInputEnabled(unsigned Index) { 874 PSInputEnable |= 1 << Index; 875 } 876 877 bool returnsVoid() const { 878 return ReturnsVoid; 879 } 880 881 void setIfReturnsVoid(bool Value) { 882 ReturnsVoid = Value; 883 } 884 885 /// \returns A pair of default/requested minimum/maximum flat work group sizes 886 /// for this function. 887 std::pair<unsigned, unsigned> getFlatWorkGroupSizes() const { 888 return FlatWorkGroupSizes; 889 } 890 891 /// \returns Default/requested minimum flat work group size for this function. 892 unsigned getMinFlatWorkGroupSize() const { 893 return FlatWorkGroupSizes.first; 894 } 895 896 /// \returns Default/requested maximum flat work group size for this function. 897 unsigned getMaxFlatWorkGroupSize() const { 898 return FlatWorkGroupSizes.second; 899 } 900 901 /// \returns A pair of default/requested minimum/maximum number of waves per 902 /// execution unit. 903 std::pair<unsigned, unsigned> getWavesPerEU() const { 904 return WavesPerEU; 905 } 906 907 /// \returns Default/requested minimum number of waves per execution unit. 908 unsigned getMinWavesPerEU() const { 909 return WavesPerEU.first; 910 } 911 912 /// \returns Default/requested maximum number of waves per execution unit. 913 unsigned getMaxWavesPerEU() const { 914 return WavesPerEU.second; 915 } 916 917 /// \returns SGPR used for \p Dim's work group ID. 918 Register getWorkGroupIDSGPR(unsigned Dim) const { 919 switch (Dim) { 920 case 0: 921 assert(hasWorkGroupIDX()); 922 return ArgInfo.WorkGroupIDX.getRegister(); 923 case 1: 924 assert(hasWorkGroupIDY()); 925 return ArgInfo.WorkGroupIDY.getRegister(); 926 case 2: 927 assert(hasWorkGroupIDZ()); 928 return ArgInfo.WorkGroupIDZ.getRegister(); 929 } 930 llvm_unreachable("unexpected dimension"); 931 } 932 933 const AMDGPUBufferPseudoSourceValue * 934 getBufferPSV(const AMDGPUTargetMachine &TM) { 935 if (!BufferPSV) 936 BufferPSV = std::make_unique<AMDGPUBufferPseudoSourceValue>(TM); 937 938 return BufferPSV.get(); 939 } 940 941 const AMDGPUImagePseudoSourceValue * 942 getImagePSV(const AMDGPUTargetMachine &TM) { 943 if (!ImagePSV) 944 ImagePSV = std::make_unique<AMDGPUImagePseudoSourceValue>(TM); 945 946 return ImagePSV.get(); 947 } 948 949 const AMDGPUGWSResourcePseudoSourceValue * 950 getGWSPSV(const AMDGPUTargetMachine &TM) { 951 if (!GWSResourcePSV) { 952 GWSResourcePSV = std::make_unique<AMDGPUGWSResourcePseudoSourceValue>(TM); 953 } 954 955 return GWSResourcePSV.get(); 956 } 957 958 unsigned getOccupancy() const { 959 return Occupancy; 960 } 961 962 unsigned getMinAllowedOccupancy() const { 963 if (!isMemoryBound() && !needsWaveLimiter()) 964 return Occupancy; 965 return (Occupancy < 4) ? Occupancy : 4; 966 } 967 968 void limitOccupancy(const MachineFunction &MF); 969 970 void limitOccupancy(unsigned Limit) { 971 if (Occupancy > Limit) 972 Occupancy = Limit; 973 } 974 975 void increaseOccupancy(const MachineFunction &MF, unsigned Limit) { 976 if (Occupancy < Limit) 977 Occupancy = Limit; 978 limitOccupancy(MF); 979 } 980 981 bool mayNeedAGPRs() const { 982 return MayNeedAGPRs; 983 } 984 985 // \returns true if a function has a use of AGPRs via inline asm or 986 // has a call which may use it. 987 bool mayUseAGPRs(const MachineFunction &MF) const; 988 989 // \returns true if a function needs or may need AGPRs. 990 bool usesAGPRs(const MachineFunction &MF) const; 991 }; 992 993 } // end namespace llvm 994 995 #endif // LLVM_LIB_TARGET_AMDGPU_SIMACHINEFUNCTIONINFO_H 996