1 //===-- FixupStatepointCallerSaved.cpp - Fixup caller saved registers ----===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 /// 10 /// \file 11 /// Statepoint instruction in deopt parameters contains values which are 12 /// meaningful to the runtime and should be able to be read at the moment the 13 /// call returns. So we can say that we need to encode the fact that these 14 /// values are "late read" by runtime. If we could express this notion for 15 /// register allocator it would produce the right form for us. 16 /// The need to fixup (i.e this pass) is specifically handling the fact that 17 /// we cannot describe such a late read for the register allocator. 18 /// Register allocator may put the value on a register clobbered by the call. 19 /// This pass forces the spill of such registers and replaces corresponding 20 /// statepoint operands to added spill slots. 21 /// 22 //===----------------------------------------------------------------------===// 23 24 #include "llvm/ADT/SmallSet.h" 25 #include "llvm/ADT/Statistic.h" 26 #include "llvm/CodeGen/MachineFrameInfo.h" 27 #include "llvm/CodeGen/MachineFunctionPass.h" 28 #include "llvm/CodeGen/MachineRegisterInfo.h" 29 #include "llvm/CodeGen/Passes.h" 30 #include "llvm/CodeGen/StackMaps.h" 31 #include "llvm/CodeGen/TargetFrameLowering.h" 32 #include "llvm/CodeGen/TargetInstrInfo.h" 33 #include "llvm/IR/Statepoint.h" 34 #include "llvm/InitializePasses.h" 35 #include "llvm/Support/Debug.h" 36 37 using namespace llvm; 38 39 #define DEBUG_TYPE "fixup-statepoint-caller-saved" 40 STATISTIC(NumSpilledRegisters, "Number of spilled register"); 41 STATISTIC(NumSpillSlotsAllocated, "Number of spill slots allocated"); 42 STATISTIC(NumSpillSlotsExtended, "Number of spill slots extended"); 43 44 static cl::opt<bool> FixupSCSExtendSlotSize( 45 "fixup-scs-extend-slot-size", cl::Hidden, cl::init(false), 46 cl::desc("Allow spill in spill slot of greater size than register size"), 47 cl::Hidden); 48 49 static cl::opt<bool> PassGCPtrInCSR( 50 "fixup-allow-gcptr-in-csr", cl::Hidden, cl::init(false), 51 cl::desc("Allow passing GC Pointer arguments in callee saved registers")); 52 53 static cl::opt<bool> EnableCopyProp( 54 "fixup-scs-enable-copy-propagation", cl::Hidden, cl::init(true), 55 cl::desc("Enable simple copy propagation during register reloading")); 56 57 // This is purely debugging option. 58 // It may be handy for investigating statepoint spilling issues. 59 static cl::opt<unsigned> MaxStatepointsWithRegs( 60 "fixup-max-csr-statepoints", cl::Hidden, 61 cl::desc("Max number of statepoints allowed to pass GC Ptrs in registers")); 62 63 namespace { 64 65 class FixupStatepointCallerSaved : public MachineFunctionPass { 66 public: 67 static char ID; 68 69 FixupStatepointCallerSaved() : MachineFunctionPass(ID) { 70 initializeFixupStatepointCallerSavedPass(*PassRegistry::getPassRegistry()); 71 } 72 73 void getAnalysisUsage(AnalysisUsage &AU) const override { 74 AU.setPreservesCFG(); 75 MachineFunctionPass::getAnalysisUsage(AU); 76 } 77 78 StringRef getPassName() const override { 79 return "Fixup Statepoint Caller Saved"; 80 } 81 82 bool runOnMachineFunction(MachineFunction &MF) override; 83 }; 84 85 } // End anonymous namespace. 86 87 char FixupStatepointCallerSaved::ID = 0; 88 char &llvm::FixupStatepointCallerSavedID = FixupStatepointCallerSaved::ID; 89 90 INITIALIZE_PASS_BEGIN(FixupStatepointCallerSaved, DEBUG_TYPE, 91 "Fixup Statepoint Caller Saved", false, false) 92 INITIALIZE_PASS_END(FixupStatepointCallerSaved, DEBUG_TYPE, 93 "Fixup Statepoint Caller Saved", false, false) 94 95 // Utility function to get size of the register. 96 static unsigned getRegisterSize(const TargetRegisterInfo &TRI, Register Reg) { 97 const TargetRegisterClass *RC = TRI.getMinimalPhysRegClass(Reg); 98 return TRI.getSpillSize(*RC); 99 } 100 101 // Advance iterator to the next stack map entry 102 static MachineInstr::const_mop_iterator 103 advanceToNextStackMapElt(MachineInstr::const_mop_iterator MOI) { 104 if (MOI->isImm()) { 105 switch (MOI->getImm()) { 106 default: 107 llvm_unreachable("Unrecognized operand type."); 108 case StackMaps::DirectMemRefOp: 109 MOI += 2; // <Reg>, <Imm> 110 break; 111 case StackMaps::IndirectMemRefOp: 112 MOI += 3; // <Size>, <Reg>, <Imm> 113 break; 114 case StackMaps::ConstantOp: 115 MOI += 1; 116 break; 117 } 118 } 119 return ++MOI; 120 } 121 122 // Return statepoint GC args as a set 123 static SmallSet<Register, 8> collectGCRegs(MachineInstr &MI) { 124 StatepointOpers SO(&MI); 125 unsigned NumDeoptIdx = SO.getNumDeoptArgsIdx(); 126 unsigned NumDeoptArgs = MI.getOperand(NumDeoptIdx).getImm(); 127 MachineInstr::const_mop_iterator MOI(MI.operands_begin() + NumDeoptIdx + 1), 128 MOE(MI.operands_end()); 129 130 // Skip deopt args 131 while (NumDeoptArgs--) 132 MOI = advanceToNextStackMapElt(MOI); 133 134 SmallSet<Register, 8> Result; 135 while (MOI != MOE) { 136 if (MOI->isReg() && !MOI->isImplicit()) 137 Result.insert(MOI->getReg()); 138 MOI = advanceToNextStackMapElt(MOI); 139 } 140 return Result; 141 } 142 143 // Try to eliminate redundant copy to register which we're going to 144 // spill, i.e. try to change: 145 // X = COPY Y 146 // SPILL X 147 // to 148 // SPILL Y 149 // If there are no uses of X between copy and STATEPOINT, that COPY 150 // may be eliminated. 151 // Reg - register we're about to spill 152 // RI - On entry points to statepoint. 153 // On successful copy propagation set to new spill point. 154 // IsKill - set to true if COPY is Kill (there are no uses of Y) 155 // Returns either found source copy register or original one. 156 static Register performCopyPropagation(Register Reg, 157 MachineBasicBlock::iterator &RI, 158 bool &IsKill, const TargetInstrInfo &TII, 159 const TargetRegisterInfo &TRI) { 160 // First check if statepoint itself uses Reg in non-meta operands. 161 int Idx = RI->findRegisterUseOperandIdx(Reg, false, &TRI); 162 if (Idx >= 0 && (unsigned)Idx < StatepointOpers(&*RI).getNumDeoptArgsIdx()) { 163 IsKill = false; 164 return Reg; 165 } 166 167 if (!EnableCopyProp) 168 return Reg; 169 170 MachineBasicBlock *MBB = RI->getParent(); 171 MachineBasicBlock::reverse_iterator E = MBB->rend(); 172 MachineInstr *Def = nullptr, *Use = nullptr; 173 for (auto It = ++(RI.getReverse()); It != E; ++It) { 174 if (It->readsRegister(Reg, &TRI) && !Use) 175 Use = &*It; 176 if (It->modifiesRegister(Reg, &TRI)) { 177 Def = &*It; 178 break; 179 } 180 } 181 182 if (!Def) 183 return Reg; 184 185 auto DestSrc = TII.isCopyInstr(*Def); 186 if (!DestSrc || DestSrc->Destination->getReg() != Reg) 187 return Reg; 188 189 Register SrcReg = DestSrc->Source->getReg(); 190 191 if (getRegisterSize(TRI, Reg) != getRegisterSize(TRI, SrcReg)) 192 return Reg; 193 194 LLVM_DEBUG(dbgs() << "spillRegisters: perform copy propagation " 195 << printReg(Reg, &TRI) << " -> " << printReg(SrcReg, &TRI) 196 << "\n"); 197 198 // Insert spill immediately after Def 199 RI = ++MachineBasicBlock::iterator(Def); 200 IsKill = DestSrc->Source->isKill(); 201 202 // There are no uses of original register between COPY and STATEPOINT. 203 // There can't be any after STATEPOINT, so we can eliminate Def. 204 if (!Use) { 205 LLVM_DEBUG(dbgs() << "spillRegisters: removing dead copy " << *Def); 206 Def->eraseFromParent(); 207 } 208 return SrcReg; 209 } 210 211 namespace { 212 // Pair {Register, FrameIndex} 213 using RegSlotPair = std::pair<Register, int>; 214 215 // Keeps track of what reloads were inserted in MBB. 216 class RegReloadCache { 217 using ReloadSet = SmallSet<RegSlotPair, 8>; 218 DenseMap<const MachineBasicBlock *, ReloadSet> Reloads; 219 220 public: 221 RegReloadCache() = default; 222 223 // Record reload of Reg from FI in block MBB 224 void recordReload(Register Reg, int FI, const MachineBasicBlock *MBB) { 225 RegSlotPair RSP(Reg, FI); 226 auto Res = Reloads[MBB].insert(RSP); 227 assert(Res.second && "reload already exists"); 228 } 229 230 // Does basic block MBB contains reload of Reg from FI? 231 bool hasReload(Register Reg, int FI, const MachineBasicBlock *MBB) { 232 RegSlotPair RSP(Reg, FI); 233 return Reloads.count(MBB) && Reloads[MBB].count(RSP); 234 } 235 }; 236 237 // Cache used frame indexes during statepoint re-write to re-use them in 238 // processing next statepoint instruction. 239 // Two strategies. One is to preserve the size of spill slot while another one 240 // extends the size of spill slots to reduce the number of them, causing 241 // the less total frame size. But unspill will have "implicit" any extend. 242 class FrameIndexesCache { 243 private: 244 struct FrameIndexesPerSize { 245 // List of used frame indexes during processing previous statepoints. 246 SmallVector<int, 8> Slots; 247 // Current index of un-used yet frame index. 248 unsigned Index = 0; 249 }; 250 MachineFrameInfo &MFI; 251 const TargetRegisterInfo &TRI; 252 // Map size to list of frame indexes of this size. If the mode is 253 // FixupSCSExtendSlotSize then the key 0 is used to keep all frame indexes. 254 // If the size of required spill slot is greater than in a cache then the 255 // size will be increased. 256 DenseMap<unsigned, FrameIndexesPerSize> Cache; 257 258 // Keeps track of slots reserved for the shared landing pad processing. 259 // Initialized from GlobalIndices for the current EHPad. 260 SmallSet<int, 8> ReservedSlots; 261 262 // Landing pad can be destination of several statepoints. Every register 263 // defined by such statepoints must be spilled to the same stack slot. 264 // This map keeps that information. 265 DenseMap<const MachineBasicBlock *, SmallVector<RegSlotPair, 8>> 266 GlobalIndices; 267 268 FrameIndexesPerSize &getCacheBucket(unsigned Size) { 269 // In FixupSCSExtendSlotSize mode the bucket with 0 index is used 270 // for all sizes. 271 return Cache[FixupSCSExtendSlotSize ? 0 : Size]; 272 } 273 274 public: 275 FrameIndexesCache(MachineFrameInfo &MFI, const TargetRegisterInfo &TRI) 276 : MFI(MFI), TRI(TRI) {} 277 // Reset the current state of used frame indexes. After invocation of 278 // this function all frame indexes are available for allocation with 279 // the exception of slots reserved for landing pad processing (if any). 280 void reset(const MachineBasicBlock *EHPad) { 281 for (auto &It : Cache) 282 It.second.Index = 0; 283 284 ReservedSlots.clear(); 285 if (EHPad && GlobalIndices.count(EHPad)) 286 for (auto &RSP : GlobalIndices[EHPad]) 287 ReservedSlots.insert(RSP.second); 288 } 289 290 // Get frame index to spill the register. 291 int getFrameIndex(Register Reg, MachineBasicBlock *EHPad) { 292 // Check if slot for Reg is already reserved at EHPad. 293 auto It = GlobalIndices.find(EHPad); 294 if (It != GlobalIndices.end()) { 295 auto &Vec = It->second; 296 auto Idx = llvm::find_if( 297 Vec, [Reg](RegSlotPair &RSP) { return Reg == RSP.first; }); 298 if (Idx != Vec.end()) { 299 int FI = Idx->second; 300 LLVM_DEBUG(dbgs() << "Found global FI " << FI << " for register " 301 << printReg(Reg, &TRI) << " at " 302 << printMBBReference(*EHPad) << "\n"); 303 assert(ReservedSlots.count(FI) && "using unreserved slot"); 304 return FI; 305 } 306 } 307 308 unsigned Size = getRegisterSize(TRI, Reg); 309 FrameIndexesPerSize &Line = getCacheBucket(Size); 310 while (Line.Index < Line.Slots.size()) { 311 int FI = Line.Slots[Line.Index++]; 312 if (ReservedSlots.count(FI)) 313 continue; 314 // If all sizes are kept together we probably need to extend the 315 // spill slot size. 316 if (MFI.getObjectSize(FI) < Size) { 317 MFI.setObjectSize(FI, Size); 318 MFI.setObjectAlignment(FI, Align(Size)); 319 NumSpillSlotsExtended++; 320 } 321 return FI; 322 } 323 int FI = MFI.CreateSpillStackObject(Size, Align(Size)); 324 NumSpillSlotsAllocated++; 325 Line.Slots.push_back(FI); 326 ++Line.Index; 327 328 // Remember assignment {Reg, FI} for EHPad 329 if (EHPad) { 330 GlobalIndices[EHPad].push_back(std::make_pair(Reg, FI)); 331 LLVM_DEBUG(dbgs() << "Reserved FI " << FI << " for spilling reg " 332 << printReg(Reg, &TRI) << " at landing pad " 333 << printMBBReference(*EHPad) << "\n"); 334 } 335 336 return FI; 337 } 338 339 // Sort all registers to spill in descendent order. In the 340 // FixupSCSExtendSlotSize mode it will minimize the total frame size. 341 // In non FixupSCSExtendSlotSize mode we can skip this step. 342 void sortRegisters(SmallVectorImpl<Register> &Regs) { 343 if (!FixupSCSExtendSlotSize) 344 return; 345 llvm::sort(Regs.begin(), Regs.end(), [&](Register &A, Register &B) { 346 return getRegisterSize(TRI, A) > getRegisterSize(TRI, B); 347 }); 348 } 349 }; 350 351 // Describes the state of the current processing statepoint instruction. 352 class StatepointState { 353 private: 354 // statepoint instruction. 355 MachineInstr &MI; 356 MachineFunction &MF; 357 // If non-null then statepoint is invoke, and this points to the landing pad. 358 MachineBasicBlock *EHPad; 359 const TargetRegisterInfo &TRI; 360 const TargetInstrInfo &TII; 361 MachineFrameInfo &MFI; 362 // Mask with callee saved registers. 363 const uint32_t *Mask; 364 // Cache of frame indexes used on previous instruction processing. 365 FrameIndexesCache &CacheFI; 366 bool AllowGCPtrInCSR; 367 // Operands with physical registers requiring spilling. 368 SmallVector<unsigned, 8> OpsToSpill; 369 // Set of register to spill. 370 SmallVector<Register, 8> RegsToSpill; 371 // Set of registers to reload after statepoint. 372 SmallVector<Register, 8> RegsToReload; 373 // Map Register to Frame Slot index. 374 DenseMap<Register, int> RegToSlotIdx; 375 376 public: 377 StatepointState(MachineInstr &MI, const uint32_t *Mask, 378 FrameIndexesCache &CacheFI, bool AllowGCPtrInCSR) 379 : MI(MI), MF(*MI.getMF()), TRI(*MF.getSubtarget().getRegisterInfo()), 380 TII(*MF.getSubtarget().getInstrInfo()), MFI(MF.getFrameInfo()), 381 Mask(Mask), CacheFI(CacheFI), AllowGCPtrInCSR(AllowGCPtrInCSR) { 382 383 // Find statepoint's landing pad, if any. 384 EHPad = nullptr; 385 MachineBasicBlock *MBB = MI.getParent(); 386 // Invoke statepoint must be last one in block. 387 bool Last = std::none_of(++MI.getIterator(), MBB->end().getInstrIterator(), 388 [](MachineInstr &I) { 389 return I.getOpcode() == TargetOpcode::STATEPOINT; 390 }); 391 392 if (!Last) 393 return; 394 395 auto IsEHPad = [](MachineBasicBlock *B) { return B->isEHPad(); }; 396 397 assert(llvm::count_if(MBB->successors(), IsEHPad) < 2 && "multiple EHPads"); 398 399 auto It = llvm::find_if(MBB->successors(), IsEHPad); 400 if (It != MBB->succ_end()) 401 EHPad = *It; 402 } 403 404 MachineBasicBlock *getEHPad() const { return EHPad; } 405 406 // Return true if register is callee saved. 407 bool isCalleeSaved(Register Reg) { return (Mask[Reg / 32] >> Reg % 32) & 1; } 408 409 // Iterates over statepoint meta args to find caller saver registers. 410 // Also cache the size of found registers. 411 // Returns true if caller save registers found. 412 bool findRegistersToSpill() { 413 SmallSet<Register, 8> VisitedRegs; 414 SmallSet<Register, 8> GCRegs = collectGCRegs(MI); 415 for (unsigned Idx = StatepointOpers(&MI).getVarIdx(), 416 EndIdx = MI.getNumOperands(); 417 Idx < EndIdx; ++Idx) { 418 MachineOperand &MO = MI.getOperand(Idx); 419 if (!MO.isReg() || MO.isImplicit()) 420 continue; 421 Register Reg = MO.getReg(); 422 assert(Reg.isPhysical() && "Only physical regs are expected"); 423 424 if (isCalleeSaved(Reg) && (AllowGCPtrInCSR || !is_contained(GCRegs, Reg))) 425 continue; 426 427 LLVM_DEBUG(dbgs() << "Will spill " << printReg(Reg, &TRI) << " at index " 428 << Idx << "\n"); 429 430 if (VisitedRegs.insert(Reg).second) 431 RegsToSpill.push_back(Reg); 432 OpsToSpill.push_back(Idx); 433 } 434 CacheFI.sortRegisters(RegsToSpill); 435 return !RegsToSpill.empty(); 436 } 437 438 // Spill all caller saved registers right before statepoint instruction. 439 // Remember frame index where register is spilled. 440 void spillRegisters() { 441 for (Register Reg : RegsToSpill) { 442 int FI = CacheFI.getFrameIndex(Reg, EHPad); 443 const TargetRegisterClass *RC = TRI.getMinimalPhysRegClass(Reg); 444 445 NumSpilledRegisters++; 446 RegToSlotIdx[Reg] = FI; 447 448 LLVM_DEBUG(dbgs() << "Spilling " << printReg(Reg, &TRI) << " to FI " << FI 449 << "\n"); 450 451 // Perform trivial copy propagation 452 bool IsKill = true; 453 MachineBasicBlock::iterator InsertBefore(MI); 454 Reg = performCopyPropagation(Reg, InsertBefore, IsKill, TII, TRI); 455 456 LLVM_DEBUG(dbgs() << "Insert spill before " << *InsertBefore); 457 TII.storeRegToStackSlot(*MI.getParent(), InsertBefore, Reg, IsKill, FI, 458 RC, &TRI); 459 } 460 } 461 462 void insertReloadBefore(unsigned Reg, MachineBasicBlock::iterator It, 463 MachineBasicBlock *MBB) { 464 const TargetRegisterClass *RC = TRI.getMinimalPhysRegClass(Reg); 465 int FI = RegToSlotIdx[Reg]; 466 if (It != MBB->end()) { 467 TII.loadRegFromStackSlot(*MBB, It, Reg, FI, RC, &TRI); 468 return; 469 } 470 471 // To insert reload at the end of MBB, insert it before last instruction 472 // and then swap them. 473 assert(MBB->begin() != MBB->end() && "Empty block"); 474 --It; 475 TII.loadRegFromStackSlot(*MBB, It, Reg, FI, RC, &TRI); 476 MachineInstr *Reload = It->getPrevNode(); 477 int Dummy = 0; 478 assert(TII.isLoadFromStackSlot(*Reload, Dummy) == Reg); 479 assert(Dummy == FI); 480 MBB->remove(Reload); 481 MBB->insertAfter(It, Reload); 482 } 483 484 // Insert reloads of (relocated) registers spilled in statepoint. 485 void insertReloads(MachineInstr *NewStatepoint, RegReloadCache &RC) { 486 MachineBasicBlock *MBB = NewStatepoint->getParent(); 487 auto InsertPoint = std::next(NewStatepoint->getIterator()); 488 489 for (auto Reg : RegsToReload) { 490 insertReloadBefore(Reg, InsertPoint, MBB); 491 LLVM_DEBUG(dbgs() << "Reloading " << printReg(Reg, &TRI) << " from FI " 492 << RegToSlotIdx[Reg] << " after statepoint\n"); 493 494 if (EHPad && !RC.hasReload(Reg, RegToSlotIdx[Reg], EHPad)) { 495 RC.recordReload(Reg, RegToSlotIdx[Reg], EHPad); 496 auto EHPadInsertPoint = EHPad->SkipPHIsLabelsAndDebug(EHPad->begin()); 497 insertReloadBefore(Reg, EHPadInsertPoint, EHPad); 498 LLVM_DEBUG(dbgs() << "...also reload at EHPad " 499 << printMBBReference(*EHPad) << "\n"); 500 } 501 } 502 } 503 504 // Re-write statepoint machine instruction to replace caller saved operands 505 // with indirect memory location (frame index). 506 MachineInstr *rewriteStatepoint() { 507 MachineInstr *NewMI = 508 MF.CreateMachineInstr(TII.get(MI.getOpcode()), MI.getDebugLoc(), true); 509 MachineInstrBuilder MIB(MF, NewMI); 510 511 unsigned NumOps = MI.getNumOperands(); 512 513 // New indices for the remaining defs. 514 SmallVector<unsigned, 8> NewIndices; 515 unsigned NumDefs = MI.getNumDefs(); 516 for (unsigned I = 0; I < NumDefs; ++I) { 517 MachineOperand &DefMO = MI.getOperand(I); 518 assert(DefMO.isReg() && DefMO.isDef() && "Expected Reg Def operand"); 519 Register Reg = DefMO.getReg(); 520 if (!AllowGCPtrInCSR) { 521 assert(is_contained(RegsToSpill, Reg)); 522 RegsToReload.push_back(Reg); 523 } else { 524 if (isCalleeSaved(Reg)) { 525 NewIndices.push_back(NewMI->getNumOperands()); 526 MIB.addReg(Reg, RegState::Define); 527 } else { 528 NewIndices.push_back(NumOps); 529 RegsToReload.push_back(Reg); 530 } 531 } 532 } 533 534 // Add End marker. 535 OpsToSpill.push_back(MI.getNumOperands()); 536 unsigned CurOpIdx = 0; 537 538 for (unsigned I = NumDefs; I < MI.getNumOperands(); ++I) { 539 MachineOperand &MO = MI.getOperand(I); 540 if (I == OpsToSpill[CurOpIdx]) { 541 int FI = RegToSlotIdx[MO.getReg()]; 542 MIB.addImm(StackMaps::IndirectMemRefOp); 543 MIB.addImm(getRegisterSize(TRI, MO.getReg())); 544 assert(MO.isReg() && "Should be register"); 545 assert(MO.getReg().isPhysical() && "Should be physical register"); 546 MIB.addFrameIndex(FI); 547 MIB.addImm(0); 548 ++CurOpIdx; 549 } else { 550 MIB.add(MO); 551 unsigned OldDef; 552 if (AllowGCPtrInCSR && MI.isRegTiedToDefOperand(I, &OldDef)) { 553 assert(OldDef < NumDefs); 554 assert(NewIndices[OldDef] < NumOps); 555 MIB->tieOperands(NewIndices[OldDef], MIB->getNumOperands() - 1); 556 } 557 } 558 } 559 assert(CurOpIdx == (OpsToSpill.size() - 1) && "Not all operands processed"); 560 // Add mem operands. 561 NewMI->setMemRefs(MF, MI.memoperands()); 562 for (auto It : RegToSlotIdx) { 563 Register R = It.first; 564 int FrameIndex = It.second; 565 auto PtrInfo = MachinePointerInfo::getFixedStack(MF, FrameIndex); 566 MachineMemOperand::Flags Flags = MachineMemOperand::MOLoad; 567 if (is_contained(RegsToReload, R)) 568 Flags |= MachineMemOperand::MOStore; 569 auto *MMO = 570 MF.getMachineMemOperand(PtrInfo, Flags, getRegisterSize(TRI, R), 571 MFI.getObjectAlign(FrameIndex)); 572 NewMI->addMemOperand(MF, MMO); 573 } 574 575 // Insert new statepoint and erase old one. 576 MI.getParent()->insert(MI, NewMI); 577 578 LLVM_DEBUG(dbgs() << "rewritten statepoint to : " << *NewMI << "\n"); 579 MI.eraseFromParent(); 580 return NewMI; 581 } 582 }; 583 584 class StatepointProcessor { 585 private: 586 MachineFunction &MF; 587 const TargetRegisterInfo &TRI; 588 FrameIndexesCache CacheFI; 589 RegReloadCache ReloadCache; 590 591 public: 592 StatepointProcessor(MachineFunction &MF) 593 : MF(MF), TRI(*MF.getSubtarget().getRegisterInfo()), 594 CacheFI(MF.getFrameInfo(), TRI) {} 595 596 bool process(MachineInstr &MI, bool AllowGCPtrInCSR) { 597 StatepointOpers SO(&MI); 598 uint64_t Flags = SO.getFlags(); 599 // Do nothing for LiveIn, it supports all registers. 600 if (Flags & (uint64_t)StatepointFlags::DeoptLiveIn) 601 return false; 602 LLVM_DEBUG(dbgs() << "\nMBB " << MI.getParent()->getNumber() << " " 603 << MI.getParent()->getName() << " : process statepoint " 604 << MI); 605 CallingConv::ID CC = SO.getCallingConv(); 606 const uint32_t *Mask = TRI.getCallPreservedMask(MF, CC); 607 StatepointState SS(MI, Mask, CacheFI, AllowGCPtrInCSR); 608 CacheFI.reset(SS.getEHPad()); 609 610 if (!SS.findRegistersToSpill()) 611 return false; 612 613 SS.spillRegisters(); 614 auto *NewStatepoint = SS.rewriteStatepoint(); 615 SS.insertReloads(NewStatepoint, ReloadCache); 616 return true; 617 } 618 }; 619 } // namespace 620 621 bool FixupStatepointCallerSaved::runOnMachineFunction(MachineFunction &MF) { 622 if (skipFunction(MF.getFunction())) 623 return false; 624 625 const Function &F = MF.getFunction(); 626 if (!F.hasGC()) 627 return false; 628 629 SmallVector<MachineInstr *, 16> Statepoints; 630 for (MachineBasicBlock &BB : MF) 631 for (MachineInstr &I : BB) 632 if (I.getOpcode() == TargetOpcode::STATEPOINT) 633 Statepoints.push_back(&I); 634 635 if (Statepoints.empty()) 636 return false; 637 638 bool Changed = false; 639 StatepointProcessor SPP(MF); 640 unsigned NumStatepoints = 0; 641 bool AllowGCPtrInCSR = PassGCPtrInCSR; 642 for (MachineInstr *I : Statepoints) { 643 ++NumStatepoints; 644 if (MaxStatepointsWithRegs.getNumOccurrences() && 645 NumStatepoints >= MaxStatepointsWithRegs) 646 AllowGCPtrInCSR = false; 647 Changed |= SPP.process(*I, AllowGCPtrInCSR); 648 } 649 return Changed; 650 } 651