1 //===- SIMachineFunctionInfo.cpp - SI Machine Function Info ---------------===// 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 #include "SIMachineFunctionInfo.h" 10 #include "AMDGPUTargetMachine.h" 11 #include "AMDGPUSubtarget.h" 12 13 #define MAX_LANES 64 14 15 using namespace llvm; 16 17 SIMachineFunctionInfo::SIMachineFunctionInfo(const MachineFunction &MF) 18 : AMDGPUMachineFunction(MF), 19 PrivateSegmentBuffer(false), 20 DispatchPtr(false), 21 QueuePtr(false), 22 KernargSegmentPtr(false), 23 DispatchID(false), 24 FlatScratchInit(false), 25 WorkGroupIDX(false), 26 WorkGroupIDY(false), 27 WorkGroupIDZ(false), 28 WorkGroupInfo(false), 29 PrivateSegmentWaveByteOffset(false), 30 WorkItemIDX(false), 31 WorkItemIDY(false), 32 WorkItemIDZ(false), 33 ImplicitBufferPtr(false), 34 ImplicitArgPtr(false), 35 GITPtrHigh(0xffffffff), 36 HighBitsOf32BitAddress(0), 37 GDSSize(0) { 38 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); 39 const Function &F = MF.getFunction(); 40 FlatWorkGroupSizes = ST.getFlatWorkGroupSizes(F); 41 WavesPerEU = ST.getWavesPerEU(F); 42 43 Occupancy = ST.computeOccupancy(F, getLDSSize()); 44 CallingConv::ID CC = F.getCallingConv(); 45 46 // FIXME: Should have analysis or something rather than attribute to detect 47 // calls. 48 const bool HasCalls = F.hasFnAttribute("amdgpu-calls"); 49 50 // Enable all kernel inputs if we have the fixed ABI. Don't bother if we don't 51 // have any calls. 52 const bool UseFixedABI = AMDGPUTargetMachine::EnableFixedFunctionABI && 53 (!isEntryFunction() || HasCalls); 54 55 if (CC == CallingConv::AMDGPU_KERNEL || CC == CallingConv::SPIR_KERNEL) { 56 if (!F.arg_empty()) 57 KernargSegmentPtr = true; 58 WorkGroupIDX = true; 59 WorkItemIDX = true; 60 } else if (CC == CallingConv::AMDGPU_PS) { 61 PSInputAddr = AMDGPU::getInitialPSInputAddr(F); 62 } 63 64 if (!isEntryFunction()) { 65 // TODO: Pick a high register, and shift down, similar to a kernel. 66 FrameOffsetReg = AMDGPU::SGPR33; 67 StackPtrOffsetReg = AMDGPU::SGPR32; 68 69 if (!ST.enableFlatScratch()) { 70 // Non-entry functions have no special inputs for now, other registers 71 // required for scratch access. 72 ScratchRSrcReg = AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3; 73 74 ArgInfo.PrivateSegmentBuffer = 75 ArgDescriptor::createRegister(ScratchRSrcReg); 76 } 77 78 if (F.hasFnAttribute("amdgpu-implicitarg-ptr")) 79 ImplicitArgPtr = true; 80 } else { 81 if (F.hasFnAttribute("amdgpu-implicitarg-ptr")) { 82 KernargSegmentPtr = true; 83 MaxKernArgAlign = std::max(ST.getAlignmentForImplicitArgPtr(), 84 MaxKernArgAlign); 85 } 86 } 87 88 if (UseFixedABI) { 89 WorkGroupIDX = true; 90 WorkGroupIDY = true; 91 WorkGroupIDZ = true; 92 WorkItemIDX = true; 93 WorkItemIDY = true; 94 WorkItemIDZ = true; 95 ImplicitArgPtr = true; 96 } else { 97 if (F.hasFnAttribute("amdgpu-work-group-id-x")) 98 WorkGroupIDX = true; 99 100 if (F.hasFnAttribute("amdgpu-work-group-id-y")) 101 WorkGroupIDY = true; 102 103 if (F.hasFnAttribute("amdgpu-work-group-id-z")) 104 WorkGroupIDZ = true; 105 106 if (F.hasFnAttribute("amdgpu-work-item-id-x")) 107 WorkItemIDX = true; 108 109 if (F.hasFnAttribute("amdgpu-work-item-id-y")) 110 WorkItemIDY = true; 111 112 if (F.hasFnAttribute("amdgpu-work-item-id-z")) 113 WorkItemIDZ = true; 114 } 115 116 bool HasStackObjects = F.hasFnAttribute("amdgpu-stack-objects"); 117 if (isEntryFunction()) { 118 // X, XY, and XYZ are the only supported combinations, so make sure Y is 119 // enabled if Z is. 120 if (WorkItemIDZ) 121 WorkItemIDY = true; 122 123 PrivateSegmentWaveByteOffset = true; 124 125 // HS and GS always have the scratch wave offset in SGPR5 on GFX9. 126 if (ST.getGeneration() >= AMDGPUSubtarget::GFX9 && 127 (CC == CallingConv::AMDGPU_HS || CC == CallingConv::AMDGPU_GS)) 128 ArgInfo.PrivateSegmentWaveByteOffset = 129 ArgDescriptor::createRegister(AMDGPU::SGPR5); 130 } 131 132 bool isAmdHsaOrMesa = ST.isAmdHsaOrMesa(F); 133 if (isAmdHsaOrMesa) { 134 if (!ST.enableFlatScratch()) 135 PrivateSegmentBuffer = true; 136 137 if (UseFixedABI) { 138 DispatchPtr = true; 139 QueuePtr = true; 140 141 // FIXME: We don't need this? 142 DispatchID = true; 143 } else { 144 if (F.hasFnAttribute("amdgpu-dispatch-ptr")) 145 DispatchPtr = true; 146 147 if (F.hasFnAttribute("amdgpu-queue-ptr")) 148 QueuePtr = true; 149 150 if (F.hasFnAttribute("amdgpu-dispatch-id")) 151 DispatchID = true; 152 } 153 } else if (ST.isMesaGfxShader(F)) { 154 ImplicitBufferPtr = true; 155 } 156 157 if (UseFixedABI || F.hasFnAttribute("amdgpu-kernarg-segment-ptr")) 158 KernargSegmentPtr = true; 159 160 if (ST.hasFlatAddressSpace() && isEntryFunction() && 161 (isAmdHsaOrMesa || ST.enableFlatScratch())) { 162 // TODO: This could be refined a lot. The attribute is a poor way of 163 // detecting calls or stack objects that may require it before argument 164 // lowering. 165 if (HasCalls || HasStackObjects || ST.enableFlatScratch()) 166 FlatScratchInit = true; 167 } 168 169 Attribute A = F.getFnAttribute("amdgpu-git-ptr-high"); 170 StringRef S = A.getValueAsString(); 171 if (!S.empty()) 172 S.consumeInteger(0, GITPtrHigh); 173 174 A = F.getFnAttribute("amdgpu-32bit-address-high-bits"); 175 S = A.getValueAsString(); 176 if (!S.empty()) 177 S.consumeInteger(0, HighBitsOf32BitAddress); 178 179 S = F.getFnAttribute("amdgpu-gds-size").getValueAsString(); 180 if (!S.empty()) 181 S.consumeInteger(0, GDSSize); 182 } 183 184 void SIMachineFunctionInfo::limitOccupancy(const MachineFunction &MF) { 185 limitOccupancy(getMaxWavesPerEU()); 186 const GCNSubtarget& ST = MF.getSubtarget<GCNSubtarget>(); 187 limitOccupancy(ST.getOccupancyWithLocalMemSize(getLDSSize(), 188 MF.getFunction())); 189 } 190 191 Register SIMachineFunctionInfo::addPrivateSegmentBuffer( 192 const SIRegisterInfo &TRI) { 193 ArgInfo.PrivateSegmentBuffer = 194 ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 195 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SGPR_128RegClass)); 196 NumUserSGPRs += 4; 197 return ArgInfo.PrivateSegmentBuffer.getRegister(); 198 } 199 200 Register SIMachineFunctionInfo::addDispatchPtr(const SIRegisterInfo &TRI) { 201 ArgInfo.DispatchPtr = ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 202 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass)); 203 NumUserSGPRs += 2; 204 return ArgInfo.DispatchPtr.getRegister(); 205 } 206 207 Register SIMachineFunctionInfo::addQueuePtr(const SIRegisterInfo &TRI) { 208 ArgInfo.QueuePtr = ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 209 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass)); 210 NumUserSGPRs += 2; 211 return ArgInfo.QueuePtr.getRegister(); 212 } 213 214 Register SIMachineFunctionInfo::addKernargSegmentPtr(const SIRegisterInfo &TRI) { 215 ArgInfo.KernargSegmentPtr 216 = ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 217 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass)); 218 NumUserSGPRs += 2; 219 return ArgInfo.KernargSegmentPtr.getRegister(); 220 } 221 222 Register SIMachineFunctionInfo::addDispatchID(const SIRegisterInfo &TRI) { 223 ArgInfo.DispatchID = ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 224 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass)); 225 NumUserSGPRs += 2; 226 return ArgInfo.DispatchID.getRegister(); 227 } 228 229 Register SIMachineFunctionInfo::addFlatScratchInit(const SIRegisterInfo &TRI) { 230 ArgInfo.FlatScratchInit = ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 231 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass)); 232 NumUserSGPRs += 2; 233 return ArgInfo.FlatScratchInit.getRegister(); 234 } 235 236 Register SIMachineFunctionInfo::addImplicitBufferPtr(const SIRegisterInfo &TRI) { 237 ArgInfo.ImplicitBufferPtr = ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 238 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass)); 239 NumUserSGPRs += 2; 240 return ArgInfo.ImplicitBufferPtr.getRegister(); 241 } 242 243 bool SIMachineFunctionInfo::isCalleeSavedReg(const MCPhysReg *CSRegs, 244 MCPhysReg Reg) { 245 for (unsigned I = 0; CSRegs[I]; ++I) { 246 if (CSRegs[I] == Reg) 247 return true; 248 } 249 250 return false; 251 } 252 253 /// \p returns true if \p NumLanes slots are available in VGPRs already used for 254 /// SGPR spilling. 255 // 256 // FIXME: This only works after processFunctionBeforeFrameFinalized 257 bool SIMachineFunctionInfo::haveFreeLanesForSGPRSpill(const MachineFunction &MF, 258 unsigned NumNeed) const { 259 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); 260 unsigned WaveSize = ST.getWavefrontSize(); 261 return NumVGPRSpillLanes + NumNeed <= WaveSize * SpillVGPRs.size(); 262 } 263 264 /// Reserve a slice of a VGPR to support spilling for FrameIndex \p FI. 265 bool SIMachineFunctionInfo::allocateSGPRSpillToVGPR(MachineFunction &MF, 266 int FI) { 267 std::vector<SpilledReg> &SpillLanes = SGPRToVGPRSpills[FI]; 268 269 // This has already been allocated. 270 if (!SpillLanes.empty()) 271 return true; 272 273 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); 274 const SIRegisterInfo *TRI = ST.getRegisterInfo(); 275 MachineFrameInfo &FrameInfo = MF.getFrameInfo(); 276 MachineRegisterInfo &MRI = MF.getRegInfo(); 277 unsigned WaveSize = ST.getWavefrontSize(); 278 SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>(); 279 280 unsigned Size = FrameInfo.getObjectSize(FI); 281 unsigned NumLanes = Size / 4; 282 283 if (NumLanes > WaveSize) 284 return false; 285 286 assert(Size >= 4 && "invalid sgpr spill size"); 287 assert(TRI->spillSGPRToVGPR() && "not spilling SGPRs to VGPRs"); 288 289 const MCPhysReg *CSRegs = MRI.getCalleeSavedRegs(); 290 291 // Make sure to handle the case where a wide SGPR spill may span between two 292 // VGPRs. 293 for (unsigned I = 0; I < NumLanes; ++I, ++NumVGPRSpillLanes) { 294 Register LaneVGPR; 295 unsigned VGPRIndex = (NumVGPRSpillLanes % WaveSize); 296 297 // Reserve a VGPR (when NumVGPRSpillLanes = 0, WaveSize, 2*WaveSize, ..) and 298 // when one of the two conditions is true: 299 // 1. One reserved VGPR being tracked by VGPRReservedForSGPRSpill is not yet 300 // reserved. 301 // 2. All spill lanes of reserved VGPR(s) are full and another spill lane is 302 // required. 303 if (FuncInfo->VGPRReservedForSGPRSpill && NumVGPRSpillLanes < WaveSize) { 304 assert(FuncInfo->VGPRReservedForSGPRSpill == SpillVGPRs.back().VGPR); 305 LaneVGPR = FuncInfo->VGPRReservedForSGPRSpill; 306 } else if (VGPRIndex == 0) { 307 LaneVGPR = TRI->findUnusedRegister(MRI, &AMDGPU::VGPR_32RegClass, MF); 308 if (LaneVGPR == AMDGPU::NoRegister) { 309 // We have no VGPRs left for spilling SGPRs. Reset because we will not 310 // partially spill the SGPR to VGPRs. 311 SGPRToVGPRSpills.erase(FI); 312 NumVGPRSpillLanes -= I; 313 return false; 314 } 315 316 Optional<int> CSRSpillFI; 317 if ((FrameInfo.hasCalls() || !isEntryFunction()) && CSRegs && 318 isCalleeSavedReg(CSRegs, LaneVGPR)) { 319 CSRSpillFI = FrameInfo.CreateSpillStackObject(4, Align(4)); 320 } 321 322 SpillVGPRs.push_back(SGPRSpillVGPRCSR(LaneVGPR, CSRSpillFI)); 323 324 // Add this register as live-in to all blocks to avoid machine verifer 325 // complaining about use of an undefined physical register. 326 for (MachineBasicBlock &BB : MF) 327 BB.addLiveIn(LaneVGPR); 328 } else { 329 LaneVGPR = SpillVGPRs.back().VGPR; 330 } 331 332 SpillLanes.push_back(SpilledReg(LaneVGPR, VGPRIndex)); 333 } 334 335 return true; 336 } 337 338 /// Reserve a VGPR for spilling of SGPRs 339 bool SIMachineFunctionInfo::reserveVGPRforSGPRSpills(MachineFunction &MF) { 340 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); 341 const SIRegisterInfo *TRI = ST.getRegisterInfo(); 342 SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>(); 343 344 Register LaneVGPR = TRI->findUnusedRegister( 345 MF.getRegInfo(), &AMDGPU::VGPR_32RegClass, MF, true); 346 if (LaneVGPR == Register()) 347 return false; 348 SpillVGPRs.push_back(SGPRSpillVGPRCSR(LaneVGPR, None)); 349 FuncInfo->VGPRReservedForSGPRSpill = LaneVGPR; 350 return true; 351 } 352 353 /// Reserve AGPRs or VGPRs to support spilling for FrameIndex \p FI. 354 /// Either AGPR is spilled to VGPR to vice versa. 355 /// Returns true if a \p FI can be eliminated completely. 356 bool SIMachineFunctionInfo::allocateVGPRSpillToAGPR(MachineFunction &MF, 357 int FI, 358 bool isAGPRtoVGPR) { 359 MachineRegisterInfo &MRI = MF.getRegInfo(); 360 MachineFrameInfo &FrameInfo = MF.getFrameInfo(); 361 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); 362 363 assert(ST.hasMAIInsts() && FrameInfo.isSpillSlotObjectIndex(FI)); 364 365 auto &Spill = VGPRToAGPRSpills[FI]; 366 367 // This has already been allocated. 368 if (!Spill.Lanes.empty()) 369 return Spill.FullyAllocated; 370 371 unsigned Size = FrameInfo.getObjectSize(FI); 372 unsigned NumLanes = Size / 4; 373 Spill.Lanes.resize(NumLanes, AMDGPU::NoRegister); 374 375 const TargetRegisterClass &RC = 376 isAGPRtoVGPR ? AMDGPU::VGPR_32RegClass : AMDGPU::AGPR_32RegClass; 377 auto Regs = RC.getRegisters(); 378 379 auto &SpillRegs = isAGPRtoVGPR ? SpillAGPR : SpillVGPR; 380 const SIRegisterInfo *TRI = ST.getRegisterInfo(); 381 Spill.FullyAllocated = true; 382 383 // FIXME: Move allocation logic out of MachineFunctionInfo and initialize 384 // once. 385 BitVector OtherUsedRegs; 386 OtherUsedRegs.resize(TRI->getNumRegs()); 387 388 const uint32_t *CSRMask = 389 TRI->getCallPreservedMask(MF, MF.getFunction().getCallingConv()); 390 if (CSRMask) 391 OtherUsedRegs.setBitsInMask(CSRMask); 392 393 // TODO: Should include register tuples, but doesn't matter with current 394 // usage. 395 for (MCPhysReg Reg : SpillAGPR) 396 OtherUsedRegs.set(Reg); 397 for (MCPhysReg Reg : SpillVGPR) 398 OtherUsedRegs.set(Reg); 399 400 SmallVectorImpl<MCPhysReg>::const_iterator NextSpillReg = Regs.begin(); 401 for (unsigned I = 0; I < NumLanes; ++I) { 402 NextSpillReg = std::find_if( 403 NextSpillReg, Regs.end(), [&MRI, &OtherUsedRegs](MCPhysReg Reg) { 404 return MRI.isAllocatable(Reg) && !MRI.isPhysRegUsed(Reg) && 405 !OtherUsedRegs[Reg]; 406 }); 407 408 if (NextSpillReg == Regs.end()) { // Registers exhausted 409 Spill.FullyAllocated = false; 410 break; 411 } 412 413 OtherUsedRegs.set(*NextSpillReg); 414 SpillRegs.push_back(*NextSpillReg); 415 Spill.Lanes[I] = *NextSpillReg++; 416 } 417 418 return Spill.FullyAllocated; 419 } 420 421 void SIMachineFunctionInfo::removeDeadFrameIndices(MachineFrameInfo &MFI) { 422 // The FP & BP spills haven't been inserted yet, so keep them around. 423 for (auto &R : SGPRToVGPRSpills) { 424 if (R.first != FramePointerSaveIndex && R.first != BasePointerSaveIndex) 425 MFI.RemoveStackObject(R.first); 426 } 427 428 // All other SPGRs must be allocated on the default stack, so reset the stack 429 // ID. 430 for (int i = MFI.getObjectIndexBegin(), e = MFI.getObjectIndexEnd(); i != e; 431 ++i) 432 if (i != FramePointerSaveIndex && i != BasePointerSaveIndex) 433 MFI.setStackID(i, TargetStackID::Default); 434 435 for (auto &R : VGPRToAGPRSpills) { 436 if (R.second.FullyAllocated) 437 MFI.RemoveStackObject(R.first); 438 } 439 } 440 441 MCPhysReg SIMachineFunctionInfo::getNextUserSGPR() const { 442 assert(NumSystemSGPRs == 0 && "System SGPRs must be added after user SGPRs"); 443 return AMDGPU::SGPR0 + NumUserSGPRs; 444 } 445 446 MCPhysReg SIMachineFunctionInfo::getNextSystemSGPR() const { 447 return AMDGPU::SGPR0 + NumUserSGPRs + NumSystemSGPRs; 448 } 449 450 Register 451 SIMachineFunctionInfo::getGITPtrLoReg(const MachineFunction &MF) const { 452 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); 453 if (!ST.isAmdPalOS()) 454 return Register(); 455 Register GitPtrLo = AMDGPU::SGPR0; // Low GIT address passed in 456 if (ST.hasMergedShaders()) { 457 switch (MF.getFunction().getCallingConv()) { 458 case CallingConv::AMDGPU_HS: 459 case CallingConv::AMDGPU_GS: 460 // Low GIT address is passed in s8 rather than s0 for an LS+HS or 461 // ES+GS merged shader on gfx9+. 462 GitPtrLo = AMDGPU::SGPR8; 463 return GitPtrLo; 464 default: 465 return GitPtrLo; 466 } 467 } 468 return GitPtrLo; 469 } 470 471 static yaml::StringValue regToString(Register Reg, 472 const TargetRegisterInfo &TRI) { 473 yaml::StringValue Dest; 474 { 475 raw_string_ostream OS(Dest.Value); 476 OS << printReg(Reg, &TRI); 477 } 478 return Dest; 479 } 480 481 static Optional<yaml::SIArgumentInfo> 482 convertArgumentInfo(const AMDGPUFunctionArgInfo &ArgInfo, 483 const TargetRegisterInfo &TRI) { 484 yaml::SIArgumentInfo AI; 485 486 auto convertArg = [&](Optional<yaml::SIArgument> &A, 487 const ArgDescriptor &Arg) { 488 if (!Arg) 489 return false; 490 491 // Create a register or stack argument. 492 yaml::SIArgument SA = yaml::SIArgument::createArgument(Arg.isRegister()); 493 if (Arg.isRegister()) { 494 raw_string_ostream OS(SA.RegisterName.Value); 495 OS << printReg(Arg.getRegister(), &TRI); 496 } else 497 SA.StackOffset = Arg.getStackOffset(); 498 // Check and update the optional mask. 499 if (Arg.isMasked()) 500 SA.Mask = Arg.getMask(); 501 502 A = SA; 503 return true; 504 }; 505 506 bool Any = false; 507 Any |= convertArg(AI.PrivateSegmentBuffer, ArgInfo.PrivateSegmentBuffer); 508 Any |= convertArg(AI.DispatchPtr, ArgInfo.DispatchPtr); 509 Any |= convertArg(AI.QueuePtr, ArgInfo.QueuePtr); 510 Any |= convertArg(AI.KernargSegmentPtr, ArgInfo.KernargSegmentPtr); 511 Any |= convertArg(AI.DispatchID, ArgInfo.DispatchID); 512 Any |= convertArg(AI.FlatScratchInit, ArgInfo.FlatScratchInit); 513 Any |= convertArg(AI.PrivateSegmentSize, ArgInfo.PrivateSegmentSize); 514 Any |= convertArg(AI.WorkGroupIDX, ArgInfo.WorkGroupIDX); 515 Any |= convertArg(AI.WorkGroupIDY, ArgInfo.WorkGroupIDY); 516 Any |= convertArg(AI.WorkGroupIDZ, ArgInfo.WorkGroupIDZ); 517 Any |= convertArg(AI.WorkGroupInfo, ArgInfo.WorkGroupInfo); 518 Any |= convertArg(AI.PrivateSegmentWaveByteOffset, 519 ArgInfo.PrivateSegmentWaveByteOffset); 520 Any |= convertArg(AI.ImplicitArgPtr, ArgInfo.ImplicitArgPtr); 521 Any |= convertArg(AI.ImplicitBufferPtr, ArgInfo.ImplicitBufferPtr); 522 Any |= convertArg(AI.WorkItemIDX, ArgInfo.WorkItemIDX); 523 Any |= convertArg(AI.WorkItemIDY, ArgInfo.WorkItemIDY); 524 Any |= convertArg(AI.WorkItemIDZ, ArgInfo.WorkItemIDZ); 525 526 if (Any) 527 return AI; 528 529 return None; 530 } 531 532 yaml::SIMachineFunctionInfo::SIMachineFunctionInfo( 533 const llvm::SIMachineFunctionInfo &MFI, const TargetRegisterInfo &TRI) 534 : ExplicitKernArgSize(MFI.getExplicitKernArgSize()), 535 MaxKernArgAlign(MFI.getMaxKernArgAlign()), LDSSize(MFI.getLDSSize()), 536 DynLDSAlign(MFI.getDynLDSAlign()), IsEntryFunction(MFI.isEntryFunction()), 537 NoSignedZerosFPMath(MFI.hasNoSignedZerosFPMath()), 538 MemoryBound(MFI.isMemoryBound()), WaveLimiter(MFI.needsWaveLimiter()), 539 HasSpilledSGPRs(MFI.hasSpilledSGPRs()), 540 HasSpilledVGPRs(MFI.hasSpilledVGPRs()), 541 HighBitsOf32BitAddress(MFI.get32BitAddressHighBits()), 542 ScratchRSrcReg(regToString(MFI.getScratchRSrcReg(), TRI)), 543 FrameOffsetReg(regToString(MFI.getFrameOffsetReg(), TRI)), 544 StackPtrOffsetReg(regToString(MFI.getStackPtrOffsetReg(), TRI)), 545 ArgInfo(convertArgumentInfo(MFI.getArgInfo(), TRI)), Mode(MFI.getMode()) { 546 } 547 548 void yaml::SIMachineFunctionInfo::mappingImpl(yaml::IO &YamlIO) { 549 MappingTraits<SIMachineFunctionInfo>::mapping(YamlIO, *this); 550 } 551 552 bool SIMachineFunctionInfo::initializeBaseYamlFields( 553 const yaml::SIMachineFunctionInfo &YamlMFI) { 554 ExplicitKernArgSize = YamlMFI.ExplicitKernArgSize; 555 MaxKernArgAlign = assumeAligned(YamlMFI.MaxKernArgAlign); 556 LDSSize = YamlMFI.LDSSize; 557 DynLDSAlign = YamlMFI.DynLDSAlign; 558 HighBitsOf32BitAddress = YamlMFI.HighBitsOf32BitAddress; 559 IsEntryFunction = YamlMFI.IsEntryFunction; 560 NoSignedZerosFPMath = YamlMFI.NoSignedZerosFPMath; 561 MemoryBound = YamlMFI.MemoryBound; 562 WaveLimiter = YamlMFI.WaveLimiter; 563 HasSpilledSGPRs = YamlMFI.HasSpilledSGPRs; 564 HasSpilledVGPRs = YamlMFI.HasSpilledVGPRs; 565 return false; 566 } 567 568 // Remove VGPR which was reserved for SGPR spills if there are no spilled SGPRs 569 bool SIMachineFunctionInfo::removeVGPRForSGPRSpill(Register ReservedVGPR, 570 MachineFunction &MF) { 571 for (auto *i = SpillVGPRs.begin(); i < SpillVGPRs.end(); i++) { 572 if (i->VGPR == ReservedVGPR) { 573 SpillVGPRs.erase(i); 574 575 for (MachineBasicBlock &MBB : MF) { 576 MBB.removeLiveIn(ReservedVGPR); 577 MBB.sortUniqueLiveIns(); 578 } 579 this->VGPRReservedForSGPRSpill = AMDGPU::NoRegister; 580 return true; 581 } 582 } 583 return false; 584 } 585