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 #include "SIRegisterInfo.h" 13 #include "MCTargetDesc/AMDGPUMCTargetDesc.h" 14 #include "Utils/AMDGPUBaseInfo.h" 15 #include "llvm/ADT/Optional.h" 16 #include "llvm/CodeGen/LiveIntervals.h" 17 #include "llvm/CodeGen/MachineBasicBlock.h" 18 #include "llvm/CodeGen/MachineFrameInfo.h" 19 #include "llvm/CodeGen/MachineFunction.h" 20 #include "llvm/CodeGen/MachineRegisterInfo.h" 21 #include "llvm/CodeGen/MIRParser/MIParser.h" 22 #include "llvm/IR/CallingConv.h" 23 #include "llvm/IR/DiagnosticInfo.h" 24 #include "llvm/IR/Function.h" 25 #include <cassert> 26 #include <vector> 27 28 #define MAX_LANES 64 29 30 using namespace llvm; 31 32 SIMachineFunctionInfo::SIMachineFunctionInfo(const MachineFunction &MF) 33 : AMDGPUMachineFunction(MF), 34 PrivateSegmentBuffer(false), 35 DispatchPtr(false), 36 QueuePtr(false), 37 KernargSegmentPtr(false), 38 DispatchID(false), 39 FlatScratchInit(false), 40 WorkGroupIDX(false), 41 WorkGroupIDY(false), 42 WorkGroupIDZ(false), 43 WorkGroupInfo(false), 44 PrivateSegmentWaveByteOffset(false), 45 WorkItemIDX(false), 46 WorkItemIDY(false), 47 WorkItemIDZ(false), 48 ImplicitBufferPtr(false), 49 ImplicitArgPtr(false), 50 GITPtrHigh(0xffffffff), 51 HighBitsOf32BitAddress(0), 52 GDSSize(0) { 53 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); 54 const Function &F = MF.getFunction(); 55 FlatWorkGroupSizes = ST.getFlatWorkGroupSizes(F); 56 WavesPerEU = ST.getWavesPerEU(F); 57 58 Occupancy = ST.computeOccupancy(F, getLDSSize()); 59 CallingConv::ID CC = F.getCallingConv(); 60 61 // FIXME: Should have analysis or something rather than attribute to detect 62 // calls. 63 const bool HasCalls = F.hasFnAttribute("amdgpu-calls"); 64 65 // Enable all kernel inputs if we have the fixed ABI. Don't bother if we don't 66 // have any calls. 67 const bool UseFixedABI = AMDGPUTargetMachine::EnableFixedFunctionABI && 68 CC != CallingConv::AMDGPU_Gfx && 69 (!isEntryFunction() || HasCalls); 70 71 if (CC == CallingConv::AMDGPU_KERNEL || CC == CallingConv::SPIR_KERNEL) { 72 if (!F.arg_empty() || ST.getImplicitArgNumBytes(F) != 0) 73 KernargSegmentPtr = true; 74 WorkGroupIDX = true; 75 WorkItemIDX = true; 76 } else if (CC == CallingConv::AMDGPU_PS) { 77 PSInputAddr = AMDGPU::getInitialPSInputAddr(F); 78 } 79 80 if (!isEntryFunction()) { 81 if (UseFixedABI) 82 ArgInfo = AMDGPUArgumentUsageInfo::FixedABIFunctionInfo; 83 84 // TODO: Pick a high register, and shift down, similar to a kernel. 85 FrameOffsetReg = AMDGPU::SGPR33; 86 StackPtrOffsetReg = AMDGPU::SGPR32; 87 88 if (!ST.enableFlatScratch()) { 89 // Non-entry functions have no special inputs for now, other registers 90 // required for scratch access. 91 ScratchRSrcReg = AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3; 92 93 ArgInfo.PrivateSegmentBuffer = 94 ArgDescriptor::createRegister(ScratchRSrcReg); 95 } 96 97 if (F.hasFnAttribute("amdgpu-implicitarg-ptr")) 98 ImplicitArgPtr = true; 99 } else { 100 if (F.hasFnAttribute("amdgpu-implicitarg-ptr")) { 101 KernargSegmentPtr = true; 102 MaxKernArgAlign = std::max(ST.getAlignmentForImplicitArgPtr(), 103 MaxKernArgAlign); 104 } 105 } 106 107 if (UseFixedABI) { 108 WorkGroupIDX = true; 109 WorkGroupIDY = true; 110 WorkGroupIDZ = true; 111 WorkItemIDX = true; 112 WorkItemIDY = true; 113 WorkItemIDZ = true; 114 ImplicitArgPtr = true; 115 } else { 116 if (F.hasFnAttribute("amdgpu-work-group-id-x")) 117 WorkGroupIDX = true; 118 119 if (F.hasFnAttribute("amdgpu-work-group-id-y")) 120 WorkGroupIDY = true; 121 122 if (F.hasFnAttribute("amdgpu-work-group-id-z")) 123 WorkGroupIDZ = true; 124 125 if (F.hasFnAttribute("amdgpu-work-item-id-x")) 126 WorkItemIDX = true; 127 128 if (F.hasFnAttribute("amdgpu-work-item-id-y")) 129 WorkItemIDY = true; 130 131 if (F.hasFnAttribute("amdgpu-work-item-id-z")) 132 WorkItemIDZ = true; 133 } 134 135 bool HasStackObjects = F.hasFnAttribute("amdgpu-stack-objects"); 136 if (isEntryFunction()) { 137 // X, XY, and XYZ are the only supported combinations, so make sure Y is 138 // enabled if Z is. 139 if (WorkItemIDZ) 140 WorkItemIDY = true; 141 142 if (!ST.flatScratchIsArchitected()) { 143 PrivateSegmentWaveByteOffset = true; 144 145 // HS and GS always have the scratch wave offset in SGPR5 on GFX9. 146 if (ST.getGeneration() >= AMDGPUSubtarget::GFX9 && 147 (CC == CallingConv::AMDGPU_HS || CC == CallingConv::AMDGPU_GS)) 148 ArgInfo.PrivateSegmentWaveByteOffset = 149 ArgDescriptor::createRegister(AMDGPU::SGPR5); 150 } 151 } 152 153 bool isAmdHsaOrMesa = ST.isAmdHsaOrMesa(F); 154 if (isAmdHsaOrMesa && !ST.enableFlatScratch()) 155 PrivateSegmentBuffer = true; 156 else if (ST.isMesaGfxShader(F)) 157 ImplicitBufferPtr = true; 158 159 if (!AMDGPU::isGraphics(CC)) { 160 if (UseFixedABI) { 161 DispatchPtr = true; 162 QueuePtr = true; 163 164 // FIXME: We don't need this? 165 DispatchID = true; 166 } else { 167 if (F.hasFnAttribute("amdgpu-dispatch-ptr")) 168 DispatchPtr = true; 169 170 if (F.hasFnAttribute("amdgpu-queue-ptr")) 171 QueuePtr = true; 172 173 if (F.hasFnAttribute("amdgpu-dispatch-id")) 174 DispatchID = true; 175 } 176 } 177 178 // TODO: This could be refined a lot. The attribute is a poor way of 179 // detecting calls or stack objects that may require it before argument 180 // lowering. 181 if (ST.hasFlatAddressSpace() && isEntryFunction() && 182 (isAmdHsaOrMesa || ST.enableFlatScratch()) && 183 (HasCalls || HasStackObjects || ST.enableFlatScratch()) && 184 !ST.flatScratchIsArchitected()) { 185 FlatScratchInit = true; 186 } 187 188 Attribute A = F.getFnAttribute("amdgpu-git-ptr-high"); 189 StringRef S = A.getValueAsString(); 190 if (!S.empty()) 191 S.consumeInteger(0, GITPtrHigh); 192 193 A = F.getFnAttribute("amdgpu-32bit-address-high-bits"); 194 S = A.getValueAsString(); 195 if (!S.empty()) 196 S.consumeInteger(0, HighBitsOf32BitAddress); 197 198 S = F.getFnAttribute("amdgpu-gds-size").getValueAsString(); 199 if (!S.empty()) 200 S.consumeInteger(0, GDSSize); 201 } 202 203 void SIMachineFunctionInfo::limitOccupancy(const MachineFunction &MF) { 204 limitOccupancy(getMaxWavesPerEU()); 205 const GCNSubtarget& ST = MF.getSubtarget<GCNSubtarget>(); 206 limitOccupancy(ST.getOccupancyWithLocalMemSize(getLDSSize(), 207 MF.getFunction())); 208 } 209 210 Register SIMachineFunctionInfo::addPrivateSegmentBuffer( 211 const SIRegisterInfo &TRI) { 212 ArgInfo.PrivateSegmentBuffer = 213 ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 214 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SGPR_128RegClass)); 215 NumUserSGPRs += 4; 216 return ArgInfo.PrivateSegmentBuffer.getRegister(); 217 } 218 219 Register SIMachineFunctionInfo::addDispatchPtr(const SIRegisterInfo &TRI) { 220 ArgInfo.DispatchPtr = ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 221 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass)); 222 NumUserSGPRs += 2; 223 return ArgInfo.DispatchPtr.getRegister(); 224 } 225 226 Register SIMachineFunctionInfo::addQueuePtr(const SIRegisterInfo &TRI) { 227 ArgInfo.QueuePtr = ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 228 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass)); 229 NumUserSGPRs += 2; 230 return ArgInfo.QueuePtr.getRegister(); 231 } 232 233 Register SIMachineFunctionInfo::addKernargSegmentPtr(const SIRegisterInfo &TRI) { 234 ArgInfo.KernargSegmentPtr 235 = ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 236 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass)); 237 NumUserSGPRs += 2; 238 return ArgInfo.KernargSegmentPtr.getRegister(); 239 } 240 241 Register SIMachineFunctionInfo::addDispatchID(const SIRegisterInfo &TRI) { 242 ArgInfo.DispatchID = ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 243 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass)); 244 NumUserSGPRs += 2; 245 return ArgInfo.DispatchID.getRegister(); 246 } 247 248 Register SIMachineFunctionInfo::addFlatScratchInit(const SIRegisterInfo &TRI) { 249 ArgInfo.FlatScratchInit = ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 250 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass)); 251 NumUserSGPRs += 2; 252 return ArgInfo.FlatScratchInit.getRegister(); 253 } 254 255 Register SIMachineFunctionInfo::addImplicitBufferPtr(const SIRegisterInfo &TRI) { 256 ArgInfo.ImplicitBufferPtr = ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 257 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass)); 258 NumUserSGPRs += 2; 259 return ArgInfo.ImplicitBufferPtr.getRegister(); 260 } 261 262 bool SIMachineFunctionInfo::isCalleeSavedReg(const MCPhysReg *CSRegs, 263 MCPhysReg Reg) { 264 for (unsigned I = 0; CSRegs[I]; ++I) { 265 if (CSRegs[I] == Reg) 266 return true; 267 } 268 269 return false; 270 } 271 272 /// \p returns true if \p NumLanes slots are available in VGPRs already used for 273 /// SGPR spilling. 274 // 275 // FIXME: This only works after processFunctionBeforeFrameFinalized 276 bool SIMachineFunctionInfo::haveFreeLanesForSGPRSpill(const MachineFunction &MF, 277 unsigned NumNeed) const { 278 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); 279 unsigned WaveSize = ST.getWavefrontSize(); 280 return NumVGPRSpillLanes + NumNeed <= WaveSize * SpillVGPRs.size(); 281 } 282 283 /// Reserve a slice of a VGPR to support spilling for FrameIndex \p FI. 284 bool SIMachineFunctionInfo::allocateSGPRSpillToVGPR(MachineFunction &MF, 285 int FI) { 286 std::vector<SpilledReg> &SpillLanes = SGPRToVGPRSpills[FI]; 287 288 // This has already been allocated. 289 if (!SpillLanes.empty()) 290 return true; 291 292 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); 293 const SIRegisterInfo *TRI = ST.getRegisterInfo(); 294 MachineFrameInfo &FrameInfo = MF.getFrameInfo(); 295 MachineRegisterInfo &MRI = MF.getRegInfo(); 296 unsigned WaveSize = ST.getWavefrontSize(); 297 SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>(); 298 299 unsigned Size = FrameInfo.getObjectSize(FI); 300 unsigned NumLanes = Size / 4; 301 302 if (NumLanes > WaveSize) 303 return false; 304 305 assert(Size >= 4 && "invalid sgpr spill size"); 306 assert(TRI->spillSGPRToVGPR() && "not spilling SGPRs to VGPRs"); 307 308 // Make sure to handle the case where a wide SGPR spill may span between two 309 // VGPRs. 310 for (unsigned I = 0; I < NumLanes; ++I, ++NumVGPRSpillLanes) { 311 Register LaneVGPR; 312 unsigned VGPRIndex = (NumVGPRSpillLanes % WaveSize); 313 314 // Reserve a VGPR (when NumVGPRSpillLanes = 0, WaveSize, 2*WaveSize, ..) and 315 // when one of the two conditions is true: 316 // 1. One reserved VGPR being tracked by VGPRReservedForSGPRSpill is not yet 317 // reserved. 318 // 2. All spill lanes of reserved VGPR(s) are full and another spill lane is 319 // required. 320 if (FuncInfo->VGPRReservedForSGPRSpill && NumVGPRSpillLanes < WaveSize) { 321 assert(FuncInfo->VGPRReservedForSGPRSpill == SpillVGPRs.back().VGPR); 322 LaneVGPR = FuncInfo->VGPRReservedForSGPRSpill; 323 } else if (VGPRIndex == 0) { 324 LaneVGPR = TRI->findUnusedRegister(MRI, &AMDGPU::VGPR_32RegClass, MF); 325 if (LaneVGPR == AMDGPU::NoRegister) { 326 // We have no VGPRs left for spilling SGPRs. Reset because we will not 327 // partially spill the SGPR to VGPRs. 328 SGPRToVGPRSpills.erase(FI); 329 NumVGPRSpillLanes -= I; 330 331 #if 0 332 DiagnosticInfoResourceLimit DiagOutOfRegs(MF.getFunction(), 333 "VGPRs for SGPR spilling", 334 0, DS_Error); 335 MF.getFunction().getContext().diagnose(DiagOutOfRegs); 336 #endif 337 return false; 338 } 339 340 Optional<int> SpillFI; 341 // We need to preserve inactive lanes, so always save, even caller-save 342 // registers. 343 if (!isEntryFunction()) { 344 SpillFI = FrameInfo.CreateSpillStackObject(4, Align(4)); 345 } 346 347 SpillVGPRs.push_back(SGPRSpillVGPR(LaneVGPR, SpillFI)); 348 349 // Add this register as live-in to all blocks to avoid machine verifer 350 // complaining about use of an undefined physical register. 351 for (MachineBasicBlock &BB : MF) 352 BB.addLiveIn(LaneVGPR); 353 } else { 354 LaneVGPR = SpillVGPRs.back().VGPR; 355 } 356 357 SpillLanes.push_back(SpilledReg(LaneVGPR, VGPRIndex)); 358 } 359 360 return true; 361 } 362 363 /// Reserve a VGPR for spilling of SGPRs 364 bool SIMachineFunctionInfo::reserveVGPRforSGPRSpills(MachineFunction &MF) { 365 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); 366 const SIRegisterInfo *TRI = ST.getRegisterInfo(); 367 SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>(); 368 369 Register LaneVGPR = TRI->findUnusedRegister( 370 MF.getRegInfo(), &AMDGPU::VGPR_32RegClass, MF, true); 371 if (LaneVGPR == Register()) 372 return false; 373 SpillVGPRs.push_back(SGPRSpillVGPR(LaneVGPR, None)); 374 FuncInfo->VGPRReservedForSGPRSpill = LaneVGPR; 375 return true; 376 } 377 378 /// Reserve AGPRs or VGPRs to support spilling for FrameIndex \p FI. 379 /// Either AGPR is spilled to VGPR to vice versa. 380 /// Returns true if a \p FI can be eliminated completely. 381 bool SIMachineFunctionInfo::allocateVGPRSpillToAGPR(MachineFunction &MF, 382 int FI, 383 bool isAGPRtoVGPR) { 384 MachineRegisterInfo &MRI = MF.getRegInfo(); 385 MachineFrameInfo &FrameInfo = MF.getFrameInfo(); 386 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); 387 388 assert(ST.hasMAIInsts() && FrameInfo.isSpillSlotObjectIndex(FI)); 389 390 auto &Spill = VGPRToAGPRSpills[FI]; 391 392 // This has already been allocated. 393 if (!Spill.Lanes.empty()) 394 return Spill.FullyAllocated; 395 396 unsigned Size = FrameInfo.getObjectSize(FI); 397 unsigned NumLanes = Size / 4; 398 Spill.Lanes.resize(NumLanes, AMDGPU::NoRegister); 399 400 const TargetRegisterClass &RC = 401 isAGPRtoVGPR ? AMDGPU::VGPR_32RegClass : AMDGPU::AGPR_32RegClass; 402 auto Regs = RC.getRegisters(); 403 404 auto &SpillRegs = isAGPRtoVGPR ? SpillAGPR : SpillVGPR; 405 const SIRegisterInfo *TRI = ST.getRegisterInfo(); 406 Spill.FullyAllocated = true; 407 408 // FIXME: Move allocation logic out of MachineFunctionInfo and initialize 409 // once. 410 BitVector OtherUsedRegs; 411 OtherUsedRegs.resize(TRI->getNumRegs()); 412 413 const uint32_t *CSRMask = 414 TRI->getCallPreservedMask(MF, MF.getFunction().getCallingConv()); 415 if (CSRMask) 416 OtherUsedRegs.setBitsInMask(CSRMask); 417 418 // TODO: Should include register tuples, but doesn't matter with current 419 // usage. 420 for (MCPhysReg Reg : SpillAGPR) 421 OtherUsedRegs.set(Reg); 422 for (MCPhysReg Reg : SpillVGPR) 423 OtherUsedRegs.set(Reg); 424 425 SmallVectorImpl<MCPhysReg>::const_iterator NextSpillReg = Regs.begin(); 426 for (int I = NumLanes - 1; I >= 0; --I) { 427 NextSpillReg = std::find_if( 428 NextSpillReg, Regs.end(), [&MRI, &OtherUsedRegs](MCPhysReg Reg) { 429 return MRI.isAllocatable(Reg) && !MRI.isPhysRegUsed(Reg) && 430 !OtherUsedRegs[Reg]; 431 }); 432 433 if (NextSpillReg == Regs.end()) { // Registers exhausted 434 Spill.FullyAllocated = false; 435 break; 436 } 437 438 OtherUsedRegs.set(*NextSpillReg); 439 SpillRegs.push_back(*NextSpillReg); 440 Spill.Lanes[I] = *NextSpillReg++; 441 } 442 443 return Spill.FullyAllocated; 444 } 445 446 void SIMachineFunctionInfo::removeDeadFrameIndices(MachineFrameInfo &MFI) { 447 // The FP & BP spills haven't been inserted yet, so keep them around. 448 for (auto &R : SGPRToVGPRSpills) { 449 if (R.first != FramePointerSaveIndex && R.first != BasePointerSaveIndex) 450 MFI.RemoveStackObject(R.first); 451 } 452 453 // All other SPGRs must be allocated on the default stack, so reset the stack 454 // ID. 455 for (int i = MFI.getObjectIndexBegin(), e = MFI.getObjectIndexEnd(); i != e; 456 ++i) 457 if (i != FramePointerSaveIndex && i != BasePointerSaveIndex) 458 MFI.setStackID(i, TargetStackID::Default); 459 460 for (auto &R : VGPRToAGPRSpills) { 461 if (R.second.FullyAllocated) 462 MFI.RemoveStackObject(R.first); 463 } 464 } 465 466 int SIMachineFunctionInfo::getScavengeFI(MachineFrameInfo &MFI, 467 const SIRegisterInfo &TRI) { 468 if (ScavengeFI) 469 return *ScavengeFI; 470 if (isEntryFunction()) { 471 ScavengeFI = MFI.CreateFixedObject( 472 TRI.getSpillSize(AMDGPU::SGPR_32RegClass), 0, false); 473 } else { 474 ScavengeFI = MFI.CreateStackObject( 475 TRI.getSpillSize(AMDGPU::SGPR_32RegClass), 476 TRI.getSpillAlign(AMDGPU::SGPR_32RegClass), false); 477 } 478 return *ScavengeFI; 479 } 480 481 MCPhysReg SIMachineFunctionInfo::getNextUserSGPR() const { 482 assert(NumSystemSGPRs == 0 && "System SGPRs must be added after user SGPRs"); 483 return AMDGPU::SGPR0 + NumUserSGPRs; 484 } 485 486 MCPhysReg SIMachineFunctionInfo::getNextSystemSGPR() const { 487 return AMDGPU::SGPR0 + NumUserSGPRs + NumSystemSGPRs; 488 } 489 490 Register 491 SIMachineFunctionInfo::getGITPtrLoReg(const MachineFunction &MF) const { 492 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); 493 if (!ST.isAmdPalOS()) 494 return Register(); 495 Register GitPtrLo = AMDGPU::SGPR0; // Low GIT address passed in 496 if (ST.hasMergedShaders()) { 497 switch (MF.getFunction().getCallingConv()) { 498 case CallingConv::AMDGPU_HS: 499 case CallingConv::AMDGPU_GS: 500 // Low GIT address is passed in s8 rather than s0 for an LS+HS or 501 // ES+GS merged shader on gfx9+. 502 GitPtrLo = AMDGPU::SGPR8; 503 return GitPtrLo; 504 default: 505 return GitPtrLo; 506 } 507 } 508 return GitPtrLo; 509 } 510 511 static yaml::StringValue regToString(Register Reg, 512 const TargetRegisterInfo &TRI) { 513 yaml::StringValue Dest; 514 { 515 raw_string_ostream OS(Dest.Value); 516 OS << printReg(Reg, &TRI); 517 } 518 return Dest; 519 } 520 521 static Optional<yaml::SIArgumentInfo> 522 convertArgumentInfo(const AMDGPUFunctionArgInfo &ArgInfo, 523 const TargetRegisterInfo &TRI) { 524 yaml::SIArgumentInfo AI; 525 526 auto convertArg = [&](Optional<yaml::SIArgument> &A, 527 const ArgDescriptor &Arg) { 528 if (!Arg) 529 return false; 530 531 // Create a register or stack argument. 532 yaml::SIArgument SA = yaml::SIArgument::createArgument(Arg.isRegister()); 533 if (Arg.isRegister()) { 534 raw_string_ostream OS(SA.RegisterName.Value); 535 OS << printReg(Arg.getRegister(), &TRI); 536 } else 537 SA.StackOffset = Arg.getStackOffset(); 538 // Check and update the optional mask. 539 if (Arg.isMasked()) 540 SA.Mask = Arg.getMask(); 541 542 A = SA; 543 return true; 544 }; 545 546 bool Any = false; 547 Any |= convertArg(AI.PrivateSegmentBuffer, ArgInfo.PrivateSegmentBuffer); 548 Any |= convertArg(AI.DispatchPtr, ArgInfo.DispatchPtr); 549 Any |= convertArg(AI.QueuePtr, ArgInfo.QueuePtr); 550 Any |= convertArg(AI.KernargSegmentPtr, ArgInfo.KernargSegmentPtr); 551 Any |= convertArg(AI.DispatchID, ArgInfo.DispatchID); 552 Any |= convertArg(AI.FlatScratchInit, ArgInfo.FlatScratchInit); 553 Any |= convertArg(AI.PrivateSegmentSize, ArgInfo.PrivateSegmentSize); 554 Any |= convertArg(AI.WorkGroupIDX, ArgInfo.WorkGroupIDX); 555 Any |= convertArg(AI.WorkGroupIDY, ArgInfo.WorkGroupIDY); 556 Any |= convertArg(AI.WorkGroupIDZ, ArgInfo.WorkGroupIDZ); 557 Any |= convertArg(AI.WorkGroupInfo, ArgInfo.WorkGroupInfo); 558 Any |= convertArg(AI.PrivateSegmentWaveByteOffset, 559 ArgInfo.PrivateSegmentWaveByteOffset); 560 Any |= convertArg(AI.ImplicitArgPtr, ArgInfo.ImplicitArgPtr); 561 Any |= convertArg(AI.ImplicitBufferPtr, ArgInfo.ImplicitBufferPtr); 562 Any |= convertArg(AI.WorkItemIDX, ArgInfo.WorkItemIDX); 563 Any |= convertArg(AI.WorkItemIDY, ArgInfo.WorkItemIDY); 564 Any |= convertArg(AI.WorkItemIDZ, ArgInfo.WorkItemIDZ); 565 566 if (Any) 567 return AI; 568 569 return None; 570 } 571 572 yaml::SIMachineFunctionInfo::SIMachineFunctionInfo( 573 const llvm::SIMachineFunctionInfo &MFI, const TargetRegisterInfo &TRI, 574 const llvm::MachineFunction &MF) 575 : ExplicitKernArgSize(MFI.getExplicitKernArgSize()), 576 MaxKernArgAlign(MFI.getMaxKernArgAlign()), LDSSize(MFI.getLDSSize()), 577 DynLDSAlign(MFI.getDynLDSAlign()), IsEntryFunction(MFI.isEntryFunction()), 578 NoSignedZerosFPMath(MFI.hasNoSignedZerosFPMath()), 579 MemoryBound(MFI.isMemoryBound()), WaveLimiter(MFI.needsWaveLimiter()), 580 HasSpilledSGPRs(MFI.hasSpilledSGPRs()), 581 HasSpilledVGPRs(MFI.hasSpilledVGPRs()), 582 HighBitsOf32BitAddress(MFI.get32BitAddressHighBits()), 583 Occupancy(MFI.getOccupancy()), 584 ScratchRSrcReg(regToString(MFI.getScratchRSrcReg(), TRI)), 585 FrameOffsetReg(regToString(MFI.getFrameOffsetReg(), TRI)), 586 StackPtrOffsetReg(regToString(MFI.getStackPtrOffsetReg(), TRI)), 587 ArgInfo(convertArgumentInfo(MFI.getArgInfo(), TRI)), Mode(MFI.getMode()) { 588 auto SFI = MFI.getOptionalScavengeFI(); 589 if (SFI) 590 ScavengeFI = yaml::FrameIndex(*SFI, MF.getFrameInfo()); 591 } 592 593 void yaml::SIMachineFunctionInfo::mappingImpl(yaml::IO &YamlIO) { 594 MappingTraits<SIMachineFunctionInfo>::mapping(YamlIO, *this); 595 } 596 597 bool SIMachineFunctionInfo::initializeBaseYamlFields( 598 const yaml::SIMachineFunctionInfo &YamlMFI, const MachineFunction &MF, 599 PerFunctionMIParsingState &PFS, SMDiagnostic &Error, SMRange &SourceRange) { 600 ExplicitKernArgSize = YamlMFI.ExplicitKernArgSize; 601 MaxKernArgAlign = assumeAligned(YamlMFI.MaxKernArgAlign); 602 LDSSize = YamlMFI.LDSSize; 603 DynLDSAlign = YamlMFI.DynLDSAlign; 604 HighBitsOf32BitAddress = YamlMFI.HighBitsOf32BitAddress; 605 Occupancy = YamlMFI.Occupancy; 606 IsEntryFunction = YamlMFI.IsEntryFunction; 607 NoSignedZerosFPMath = YamlMFI.NoSignedZerosFPMath; 608 MemoryBound = YamlMFI.MemoryBound; 609 WaveLimiter = YamlMFI.WaveLimiter; 610 HasSpilledSGPRs = YamlMFI.HasSpilledSGPRs; 611 HasSpilledVGPRs = YamlMFI.HasSpilledVGPRs; 612 613 if (YamlMFI.ScavengeFI) { 614 auto FIOrErr = YamlMFI.ScavengeFI->getFI(MF.getFrameInfo()); 615 if (!FIOrErr) { 616 // Create a diagnostic for a the frame index. 617 const MemoryBuffer &Buffer = 618 *PFS.SM->getMemoryBuffer(PFS.SM->getMainFileID()); 619 620 Error = SMDiagnostic(*PFS.SM, SMLoc(), Buffer.getBufferIdentifier(), 1, 1, 621 SourceMgr::DK_Error, toString(FIOrErr.takeError()), 622 "", None, None); 623 SourceRange = YamlMFI.ScavengeFI->SourceRange; 624 return true; 625 } 626 ScavengeFI = *FIOrErr; 627 } else { 628 ScavengeFI = None; 629 } 630 return false; 631 } 632 633 // Remove VGPR which was reserved for SGPR spills if there are no spilled SGPRs 634 bool SIMachineFunctionInfo::removeVGPRForSGPRSpill(Register ReservedVGPR, 635 MachineFunction &MF) { 636 for (auto *i = SpillVGPRs.begin(); i < SpillVGPRs.end(); i++) { 637 if (i->VGPR == ReservedVGPR) { 638 SpillVGPRs.erase(i); 639 640 for (MachineBasicBlock &MBB : MF) { 641 MBB.removeLiveIn(ReservedVGPR); 642 MBB.sortUniqueLiveIns(); 643 } 644 this->VGPRReservedForSGPRSpill = AMDGPU::NoRegister; 645 return true; 646 } 647 } 648 return false; 649 } 650