1 //===- InlineSpiller.cpp - Insert spills and restores inline --------------===// 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 // The inline spiller modifies the machine function directly instead of 10 // inserting spills and restores in VirtRegMap. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "SplitKit.h" 15 #include "llvm/ADT/ArrayRef.h" 16 #include "llvm/ADT/DenseMap.h" 17 #include "llvm/ADT/MapVector.h" 18 #include "llvm/ADT/None.h" 19 #include "llvm/ADT/STLExtras.h" 20 #include "llvm/ADT/SetVector.h" 21 #include "llvm/ADT/SmallPtrSet.h" 22 #include "llvm/ADT/SmallVector.h" 23 #include "llvm/ADT/Statistic.h" 24 #include "llvm/Analysis/AliasAnalysis.h" 25 #include "llvm/CodeGen/LiveInterval.h" 26 #include "llvm/CodeGen/LiveIntervalCalc.h" 27 #include "llvm/CodeGen/LiveIntervals.h" 28 #include "llvm/CodeGen/LiveRangeEdit.h" 29 #include "llvm/CodeGen/LiveStacks.h" 30 #include "llvm/CodeGen/MachineBasicBlock.h" 31 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h" 32 #include "llvm/CodeGen/MachineDominators.h" 33 #include "llvm/CodeGen/MachineFunction.h" 34 #include "llvm/CodeGen/MachineFunctionPass.h" 35 #include "llvm/CodeGen/MachineInstr.h" 36 #include "llvm/CodeGen/MachineInstrBuilder.h" 37 #include "llvm/CodeGen/MachineInstrBundle.h" 38 #include "llvm/CodeGen/MachineLoopInfo.h" 39 #include "llvm/CodeGen/MachineOperand.h" 40 #include "llvm/CodeGen/MachineRegisterInfo.h" 41 #include "llvm/CodeGen/SlotIndexes.h" 42 #include "llvm/CodeGen/Spiller.h" 43 #include "llvm/CodeGen/StackMaps.h" 44 #include "llvm/CodeGen/TargetInstrInfo.h" 45 #include "llvm/CodeGen/TargetOpcodes.h" 46 #include "llvm/CodeGen/TargetRegisterInfo.h" 47 #include "llvm/CodeGen/TargetSubtargetInfo.h" 48 #include "llvm/CodeGen/VirtRegMap.h" 49 #include "llvm/Config/llvm-config.h" 50 #include "llvm/Support/BlockFrequency.h" 51 #include "llvm/Support/BranchProbability.h" 52 #include "llvm/Support/CommandLine.h" 53 #include "llvm/Support/Compiler.h" 54 #include "llvm/Support/Debug.h" 55 #include "llvm/Support/ErrorHandling.h" 56 #include "llvm/Support/raw_ostream.h" 57 #include <cassert> 58 #include <iterator> 59 #include <tuple> 60 #include <utility> 61 #include <vector> 62 63 using namespace llvm; 64 65 #define DEBUG_TYPE "regalloc" 66 67 STATISTIC(NumSpilledRanges, "Number of spilled live ranges"); 68 STATISTIC(NumSnippets, "Number of spilled snippets"); 69 STATISTIC(NumSpills, "Number of spills inserted"); 70 STATISTIC(NumSpillsRemoved, "Number of spills removed"); 71 STATISTIC(NumReloads, "Number of reloads inserted"); 72 STATISTIC(NumReloadsRemoved, "Number of reloads removed"); 73 STATISTIC(NumFolded, "Number of folded stack accesses"); 74 STATISTIC(NumFoldedLoads, "Number of folded loads"); 75 STATISTIC(NumRemats, "Number of rematerialized defs for spilling"); 76 77 static cl::opt<bool> DisableHoisting("disable-spill-hoist", cl::Hidden, 78 cl::desc("Disable inline spill hoisting")); 79 static cl::opt<bool> 80 RestrictStatepointRemat("restrict-statepoint-remat", 81 cl::init(false), cl::Hidden, 82 cl::desc("Restrict remat for statepoint operands")); 83 84 namespace { 85 86 class HoistSpillHelper : private LiveRangeEdit::Delegate { 87 MachineFunction &MF; 88 LiveIntervals &LIS; 89 LiveStacks &LSS; 90 AliasAnalysis *AA; 91 MachineDominatorTree &MDT; 92 MachineLoopInfo &Loops; 93 VirtRegMap &VRM; 94 MachineRegisterInfo &MRI; 95 const TargetInstrInfo &TII; 96 const TargetRegisterInfo &TRI; 97 const MachineBlockFrequencyInfo &MBFI; 98 99 InsertPointAnalysis IPA; 100 101 // Map from StackSlot to the LiveInterval of the original register. 102 // Note the LiveInterval of the original register may have been deleted 103 // after it is spilled. We keep a copy here to track the range where 104 // spills can be moved. 105 DenseMap<int, std::unique_ptr<LiveInterval>> StackSlotToOrigLI; 106 107 // Map from pair of (StackSlot and Original VNI) to a set of spills which 108 // have the same stackslot and have equal values defined by Original VNI. 109 // These spills are mergeable and are hoist candiates. 110 using MergeableSpillsMap = 111 MapVector<std::pair<int, VNInfo *>, SmallPtrSet<MachineInstr *, 16>>; 112 MergeableSpillsMap MergeableSpills; 113 114 /// This is the map from original register to a set containing all its 115 /// siblings. To hoist a spill to another BB, we need to find out a live 116 /// sibling there and use it as the source of the new spill. 117 DenseMap<Register, SmallSetVector<Register, 16>> Virt2SiblingsMap; 118 119 bool isSpillCandBB(LiveInterval &OrigLI, VNInfo &OrigVNI, 120 MachineBasicBlock &BB, Register &LiveReg); 121 122 void rmRedundantSpills( 123 SmallPtrSet<MachineInstr *, 16> &Spills, 124 SmallVectorImpl<MachineInstr *> &SpillsToRm, 125 DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill); 126 127 void getVisitOrders( 128 MachineBasicBlock *Root, SmallPtrSet<MachineInstr *, 16> &Spills, 129 SmallVectorImpl<MachineDomTreeNode *> &Orders, 130 SmallVectorImpl<MachineInstr *> &SpillsToRm, 131 DenseMap<MachineDomTreeNode *, unsigned> &SpillsToKeep, 132 DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill); 133 134 void runHoistSpills(LiveInterval &OrigLI, VNInfo &OrigVNI, 135 SmallPtrSet<MachineInstr *, 16> &Spills, 136 SmallVectorImpl<MachineInstr *> &SpillsToRm, 137 DenseMap<MachineBasicBlock *, unsigned> &SpillsToIns); 138 139 public: 140 HoistSpillHelper(MachineFunctionPass &pass, MachineFunction &mf, 141 VirtRegMap &vrm) 142 : MF(mf), LIS(pass.getAnalysis<LiveIntervals>()), 143 LSS(pass.getAnalysis<LiveStacks>()), 144 AA(&pass.getAnalysis<AAResultsWrapperPass>().getAAResults()), 145 MDT(pass.getAnalysis<MachineDominatorTree>()), 146 Loops(pass.getAnalysis<MachineLoopInfo>()), VRM(vrm), 147 MRI(mf.getRegInfo()), TII(*mf.getSubtarget().getInstrInfo()), 148 TRI(*mf.getSubtarget().getRegisterInfo()), 149 MBFI(pass.getAnalysis<MachineBlockFrequencyInfo>()), 150 IPA(LIS, mf.getNumBlockIDs()) {} 151 152 void addToMergeableSpills(MachineInstr &Spill, int StackSlot, 153 unsigned Original); 154 bool rmFromMergeableSpills(MachineInstr &Spill, int StackSlot); 155 void hoistAllSpills(); 156 void LRE_DidCloneVirtReg(unsigned, unsigned) override; 157 }; 158 159 class InlineSpiller : public Spiller { 160 MachineFunction &MF; 161 LiveIntervals &LIS; 162 LiveStacks &LSS; 163 AliasAnalysis *AA; 164 MachineDominatorTree &MDT; 165 MachineLoopInfo &Loops; 166 VirtRegMap &VRM; 167 MachineRegisterInfo &MRI; 168 const TargetInstrInfo &TII; 169 const TargetRegisterInfo &TRI; 170 const MachineBlockFrequencyInfo &MBFI; 171 172 // Variables that are valid during spill(), but used by multiple methods. 173 LiveRangeEdit *Edit; 174 LiveInterval *StackInt; 175 int StackSlot; 176 unsigned Original; 177 178 // All registers to spill to StackSlot, including the main register. 179 SmallVector<Register, 8> RegsToSpill; 180 181 // All COPY instructions to/from snippets. 182 // They are ignored since both operands refer to the same stack slot. 183 SmallPtrSet<MachineInstr*, 8> SnippetCopies; 184 185 // Values that failed to remat at some point. 186 SmallPtrSet<VNInfo*, 8> UsedValues; 187 188 // Dead defs generated during spilling. 189 SmallVector<MachineInstr*, 8> DeadDefs; 190 191 // Object records spills information and does the hoisting. 192 HoistSpillHelper HSpiller; 193 194 ~InlineSpiller() override = default; 195 196 public: 197 InlineSpiller(MachineFunctionPass &pass, MachineFunction &mf, VirtRegMap &vrm) 198 : MF(mf), LIS(pass.getAnalysis<LiveIntervals>()), 199 LSS(pass.getAnalysis<LiveStacks>()), 200 AA(&pass.getAnalysis<AAResultsWrapperPass>().getAAResults()), 201 MDT(pass.getAnalysis<MachineDominatorTree>()), 202 Loops(pass.getAnalysis<MachineLoopInfo>()), VRM(vrm), 203 MRI(mf.getRegInfo()), TII(*mf.getSubtarget().getInstrInfo()), 204 TRI(*mf.getSubtarget().getRegisterInfo()), 205 MBFI(pass.getAnalysis<MachineBlockFrequencyInfo>()), 206 HSpiller(pass, mf, vrm) {} 207 208 void spill(LiveRangeEdit &) override; 209 void postOptimization() override; 210 211 private: 212 bool isSnippet(const LiveInterval &SnipLI); 213 void collectRegsToSpill(); 214 215 bool isRegToSpill(Register Reg) { return is_contained(RegsToSpill, Reg); } 216 217 bool isSibling(Register Reg); 218 bool hoistSpillInsideBB(LiveInterval &SpillLI, MachineInstr &CopyMI); 219 void eliminateRedundantSpills(LiveInterval &LI, VNInfo *VNI); 220 221 void markValueUsed(LiveInterval*, VNInfo*); 222 bool canGuaranteeAssignmentAfterRemat(Register VReg, MachineInstr &MI); 223 bool reMaterializeFor(LiveInterval &, MachineInstr &MI); 224 void reMaterializeAll(); 225 226 bool coalesceStackAccess(MachineInstr *MI, Register Reg); 227 bool foldMemoryOperand(ArrayRef<std::pair<MachineInstr *, unsigned>>, 228 MachineInstr *LoadMI = nullptr); 229 void insertReload(Register VReg, SlotIndex, MachineBasicBlock::iterator MI); 230 void insertSpill(Register VReg, bool isKill, MachineBasicBlock::iterator MI); 231 232 void spillAroundUses(Register Reg); 233 void spillAll(); 234 }; 235 236 } // end anonymous namespace 237 238 Spiller::~Spiller() = default; 239 240 void Spiller::anchor() {} 241 242 Spiller *llvm::createInlineSpiller(MachineFunctionPass &pass, 243 MachineFunction &mf, 244 VirtRegMap &vrm) { 245 return new InlineSpiller(pass, mf, vrm); 246 } 247 248 //===----------------------------------------------------------------------===// 249 // Snippets 250 //===----------------------------------------------------------------------===// 251 252 // When spilling a virtual register, we also spill any snippets it is connected 253 // to. The snippets are small live ranges that only have a single real use, 254 // leftovers from live range splitting. Spilling them enables memory operand 255 // folding or tightens the live range around the single use. 256 // 257 // This minimizes register pressure and maximizes the store-to-load distance for 258 // spill slots which can be important in tight loops. 259 260 /// isFullCopyOf - If MI is a COPY to or from Reg, return the other register, 261 /// otherwise return 0. 262 static Register isFullCopyOf(const MachineInstr &MI, Register Reg) { 263 if (!MI.isFullCopy()) 264 return Register(); 265 if (MI.getOperand(0).getReg() == Reg) 266 return MI.getOperand(1).getReg(); 267 if (MI.getOperand(1).getReg() == Reg) 268 return MI.getOperand(0).getReg(); 269 return Register(); 270 } 271 272 /// isSnippet - Identify if a live interval is a snippet that should be spilled. 273 /// It is assumed that SnipLI is a virtual register with the same original as 274 /// Edit->getReg(). 275 bool InlineSpiller::isSnippet(const LiveInterval &SnipLI) { 276 Register Reg = Edit->getReg(); 277 278 // A snippet is a tiny live range with only a single instruction using it 279 // besides copies to/from Reg or spills/fills. We accept: 280 // 281 // %snip = COPY %Reg / FILL fi# 282 // %snip = USE %snip 283 // %Reg = COPY %snip / SPILL %snip, fi# 284 // 285 if (SnipLI.getNumValNums() > 2 || !LIS.intervalIsInOneMBB(SnipLI)) 286 return false; 287 288 MachineInstr *UseMI = nullptr; 289 290 // Check that all uses satisfy our criteria. 291 for (MachineRegisterInfo::reg_instr_nodbg_iterator 292 RI = MRI.reg_instr_nodbg_begin(SnipLI.reg), 293 E = MRI.reg_instr_nodbg_end(); RI != E; ) { 294 MachineInstr &MI = *RI++; 295 296 // Allow copies to/from Reg. 297 if (isFullCopyOf(MI, Reg)) 298 continue; 299 300 // Allow stack slot loads. 301 int FI; 302 if (SnipLI.reg == TII.isLoadFromStackSlot(MI, FI) && FI == StackSlot) 303 continue; 304 305 // Allow stack slot stores. 306 if (SnipLI.reg == TII.isStoreToStackSlot(MI, FI) && FI == StackSlot) 307 continue; 308 309 // Allow a single additional instruction. 310 if (UseMI && &MI != UseMI) 311 return false; 312 UseMI = &MI; 313 } 314 return true; 315 } 316 317 /// collectRegsToSpill - Collect live range snippets that only have a single 318 /// real use. 319 void InlineSpiller::collectRegsToSpill() { 320 Register Reg = Edit->getReg(); 321 322 // Main register always spills. 323 RegsToSpill.assign(1, Reg); 324 SnippetCopies.clear(); 325 326 // Snippets all have the same original, so there can't be any for an original 327 // register. 328 if (Original == Reg) 329 return; 330 331 for (MachineRegisterInfo::reg_instr_iterator 332 RI = MRI.reg_instr_begin(Reg), E = MRI.reg_instr_end(); RI != E; ) { 333 MachineInstr &MI = *RI++; 334 Register SnipReg = isFullCopyOf(MI, Reg); 335 if (!isSibling(SnipReg)) 336 continue; 337 LiveInterval &SnipLI = LIS.getInterval(SnipReg); 338 if (!isSnippet(SnipLI)) 339 continue; 340 SnippetCopies.insert(&MI); 341 if (isRegToSpill(SnipReg)) 342 continue; 343 RegsToSpill.push_back(SnipReg); 344 LLVM_DEBUG(dbgs() << "\talso spill snippet " << SnipLI << '\n'); 345 ++NumSnippets; 346 } 347 } 348 349 bool InlineSpiller::isSibling(Register Reg) { 350 return Reg.isVirtual() && VRM.getOriginal(Reg) == Original; 351 } 352 353 /// It is beneficial to spill to earlier place in the same BB in case 354 /// as follows: 355 /// There is an alternative def earlier in the same MBB. 356 /// Hoist the spill as far as possible in SpillMBB. This can ease 357 /// register pressure: 358 /// 359 /// x = def 360 /// y = use x 361 /// s = copy x 362 /// 363 /// Hoisting the spill of s to immediately after the def removes the 364 /// interference between x and y: 365 /// 366 /// x = def 367 /// spill x 368 /// y = use killed x 369 /// 370 /// This hoist only helps when the copy kills its source. 371 /// 372 bool InlineSpiller::hoistSpillInsideBB(LiveInterval &SpillLI, 373 MachineInstr &CopyMI) { 374 SlotIndex Idx = LIS.getInstructionIndex(CopyMI); 375 #ifndef NDEBUG 376 VNInfo *VNI = SpillLI.getVNInfoAt(Idx.getRegSlot()); 377 assert(VNI && VNI->def == Idx.getRegSlot() && "Not defined by copy"); 378 #endif 379 380 Register SrcReg = CopyMI.getOperand(1).getReg(); 381 LiveInterval &SrcLI = LIS.getInterval(SrcReg); 382 VNInfo *SrcVNI = SrcLI.getVNInfoAt(Idx); 383 LiveQueryResult SrcQ = SrcLI.Query(Idx); 384 MachineBasicBlock *DefMBB = LIS.getMBBFromIndex(SrcVNI->def); 385 if (DefMBB != CopyMI.getParent() || !SrcQ.isKill()) 386 return false; 387 388 // Conservatively extend the stack slot range to the range of the original 389 // value. We may be able to do better with stack slot coloring by being more 390 // careful here. 391 assert(StackInt && "No stack slot assigned yet."); 392 LiveInterval &OrigLI = LIS.getInterval(Original); 393 VNInfo *OrigVNI = OrigLI.getVNInfoAt(Idx); 394 StackInt->MergeValueInAsValue(OrigLI, OrigVNI, StackInt->getValNumInfo(0)); 395 LLVM_DEBUG(dbgs() << "\tmerged orig valno " << OrigVNI->id << ": " 396 << *StackInt << '\n'); 397 398 // We are going to spill SrcVNI immediately after its def, so clear out 399 // any later spills of the same value. 400 eliminateRedundantSpills(SrcLI, SrcVNI); 401 402 MachineBasicBlock *MBB = LIS.getMBBFromIndex(SrcVNI->def); 403 MachineBasicBlock::iterator MII; 404 if (SrcVNI->isPHIDef()) 405 MII = MBB->SkipPHIsLabelsAndDebug(MBB->begin()); 406 else { 407 MachineInstr *DefMI = LIS.getInstructionFromIndex(SrcVNI->def); 408 assert(DefMI && "Defining instruction disappeared"); 409 MII = DefMI; 410 ++MII; 411 } 412 // Insert spill without kill flag immediately after def. 413 TII.storeRegToStackSlot(*MBB, MII, SrcReg, false, StackSlot, 414 MRI.getRegClass(SrcReg), &TRI); 415 --MII; // Point to store instruction. 416 LIS.InsertMachineInstrInMaps(*MII); 417 LLVM_DEBUG(dbgs() << "\thoisted: " << SrcVNI->def << '\t' << *MII); 418 419 HSpiller.addToMergeableSpills(*MII, StackSlot, Original); 420 ++NumSpills; 421 return true; 422 } 423 424 /// eliminateRedundantSpills - SLI:VNI is known to be on the stack. Remove any 425 /// redundant spills of this value in SLI.reg and sibling copies. 426 void InlineSpiller::eliminateRedundantSpills(LiveInterval &SLI, VNInfo *VNI) { 427 assert(VNI && "Missing value"); 428 SmallVector<std::pair<LiveInterval*, VNInfo*>, 8> WorkList; 429 WorkList.push_back(std::make_pair(&SLI, VNI)); 430 assert(StackInt && "No stack slot assigned yet."); 431 432 do { 433 LiveInterval *LI; 434 std::tie(LI, VNI) = WorkList.pop_back_val(); 435 Register Reg = LI->reg; 436 LLVM_DEBUG(dbgs() << "Checking redundant spills for " << VNI->id << '@' 437 << VNI->def << " in " << *LI << '\n'); 438 439 // Regs to spill are taken care of. 440 if (isRegToSpill(Reg)) 441 continue; 442 443 // Add all of VNI's live range to StackInt. 444 StackInt->MergeValueInAsValue(*LI, VNI, StackInt->getValNumInfo(0)); 445 LLVM_DEBUG(dbgs() << "Merged to stack int: " << *StackInt << '\n'); 446 447 // Find all spills and copies of VNI. 448 for (MachineRegisterInfo::use_instr_nodbg_iterator 449 UI = MRI.use_instr_nodbg_begin(Reg), E = MRI.use_instr_nodbg_end(); 450 UI != E; ) { 451 MachineInstr &MI = *UI++; 452 if (!MI.isCopy() && !MI.mayStore()) 453 continue; 454 SlotIndex Idx = LIS.getInstructionIndex(MI); 455 if (LI->getVNInfoAt(Idx) != VNI) 456 continue; 457 458 // Follow sibling copies down the dominator tree. 459 if (Register DstReg = isFullCopyOf(MI, Reg)) { 460 if (isSibling(DstReg)) { 461 LiveInterval &DstLI = LIS.getInterval(DstReg); 462 VNInfo *DstVNI = DstLI.getVNInfoAt(Idx.getRegSlot()); 463 assert(DstVNI && "Missing defined value"); 464 assert(DstVNI->def == Idx.getRegSlot() && "Wrong copy def slot"); 465 WorkList.push_back(std::make_pair(&DstLI, DstVNI)); 466 } 467 continue; 468 } 469 470 // Erase spills. 471 int FI; 472 if (Reg == TII.isStoreToStackSlot(MI, FI) && FI == StackSlot) { 473 LLVM_DEBUG(dbgs() << "Redundant spill " << Idx << '\t' << MI); 474 // eliminateDeadDefs won't normally remove stores, so switch opcode. 475 MI.setDesc(TII.get(TargetOpcode::KILL)); 476 DeadDefs.push_back(&MI); 477 ++NumSpillsRemoved; 478 if (HSpiller.rmFromMergeableSpills(MI, StackSlot)) 479 --NumSpills; 480 } 481 } 482 } while (!WorkList.empty()); 483 } 484 485 //===----------------------------------------------------------------------===// 486 // Rematerialization 487 //===----------------------------------------------------------------------===// 488 489 /// markValueUsed - Remember that VNI failed to rematerialize, so its defining 490 /// instruction cannot be eliminated. See through snippet copies 491 void InlineSpiller::markValueUsed(LiveInterval *LI, VNInfo *VNI) { 492 SmallVector<std::pair<LiveInterval*, VNInfo*>, 8> WorkList; 493 WorkList.push_back(std::make_pair(LI, VNI)); 494 do { 495 std::tie(LI, VNI) = WorkList.pop_back_val(); 496 if (!UsedValues.insert(VNI).second) 497 continue; 498 499 if (VNI->isPHIDef()) { 500 MachineBasicBlock *MBB = LIS.getMBBFromIndex(VNI->def); 501 for (MachineBasicBlock *P : MBB->predecessors()) { 502 VNInfo *PVNI = LI->getVNInfoBefore(LIS.getMBBEndIdx(P)); 503 if (PVNI) 504 WorkList.push_back(std::make_pair(LI, PVNI)); 505 } 506 continue; 507 } 508 509 // Follow snippet copies. 510 MachineInstr *MI = LIS.getInstructionFromIndex(VNI->def); 511 if (!SnippetCopies.count(MI)) 512 continue; 513 LiveInterval &SnipLI = LIS.getInterval(MI->getOperand(1).getReg()); 514 assert(isRegToSpill(SnipLI.reg) && "Unexpected register in copy"); 515 VNInfo *SnipVNI = SnipLI.getVNInfoAt(VNI->def.getRegSlot(true)); 516 assert(SnipVNI && "Snippet undefined before copy"); 517 WorkList.push_back(std::make_pair(&SnipLI, SnipVNI)); 518 } while (!WorkList.empty()); 519 } 520 521 bool InlineSpiller::canGuaranteeAssignmentAfterRemat(Register VReg, 522 MachineInstr &MI) { 523 if (!RestrictStatepointRemat) 524 return true; 525 // Here's a quick explanation of the problem we're trying to handle here: 526 // * There are some pseudo instructions with more vreg uses than there are 527 // physical registers on the machine. 528 // * This is normally handled by spilling the vreg, and folding the reload 529 // into the user instruction. (Thus decreasing the number of used vregs 530 // until the remainder can be assigned to physregs.) 531 // * However, since we may try to spill vregs in any order, we can end up 532 // trying to spill each operand to the instruction, and then rematting it 533 // instead. When that happens, the new live intervals (for the remats) are 534 // expected to be trivially assignable (i.e. RS_Done). However, since we 535 // may have more remats than physregs, we're guaranteed to fail to assign 536 // one. 537 // At the moment, we only handle this for STATEPOINTs since they're the only 538 // pseudo op where we've seen this. If we start seeing other instructions 539 // with the same problem, we need to revisit this. 540 if (MI.getOpcode() != TargetOpcode::STATEPOINT) 541 return true; 542 // For STATEPOINTs we allow re-materialization for fixed arguments only hoping 543 // that number of physical registers is enough to cover all fixed arguments. 544 // If it is not true we need to revisit it. 545 for (unsigned Idx = StatepointOpers(&MI).getVarIdx(), 546 EndIdx = MI.getNumOperands(); 547 Idx < EndIdx; ++Idx) { 548 MachineOperand &MO = MI.getOperand(Idx); 549 if (MO.isReg() && MO.getReg() == VReg) 550 return false; 551 } 552 return true; 553 } 554 555 /// reMaterializeFor - Attempt to rematerialize before MI instead of reloading. 556 bool InlineSpiller::reMaterializeFor(LiveInterval &VirtReg, MachineInstr &MI) { 557 // Analyze instruction 558 SmallVector<std::pair<MachineInstr *, unsigned>, 8> Ops; 559 VirtRegInfo RI = AnalyzeVirtRegInBundle(MI, VirtReg.reg, &Ops); 560 561 if (!RI.Reads) 562 return false; 563 564 SlotIndex UseIdx = LIS.getInstructionIndex(MI).getRegSlot(true); 565 VNInfo *ParentVNI = VirtReg.getVNInfoAt(UseIdx.getBaseIndex()); 566 567 if (!ParentVNI) { 568 LLVM_DEBUG(dbgs() << "\tadding <undef> flags: "); 569 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { 570 MachineOperand &MO = MI.getOperand(i); 571 if (MO.isReg() && MO.isUse() && MO.getReg() == VirtReg.reg) 572 MO.setIsUndef(); 573 } 574 LLVM_DEBUG(dbgs() << UseIdx << '\t' << MI); 575 return true; 576 } 577 578 if (SnippetCopies.count(&MI)) 579 return false; 580 581 LiveInterval &OrigLI = LIS.getInterval(Original); 582 VNInfo *OrigVNI = OrigLI.getVNInfoAt(UseIdx); 583 LiveRangeEdit::Remat RM(ParentVNI); 584 RM.OrigMI = LIS.getInstructionFromIndex(OrigVNI->def); 585 586 if (!Edit->canRematerializeAt(RM, OrigVNI, UseIdx, false)) { 587 markValueUsed(&VirtReg, ParentVNI); 588 LLVM_DEBUG(dbgs() << "\tcannot remat for " << UseIdx << '\t' << MI); 589 return false; 590 } 591 592 // If the instruction also writes VirtReg.reg, it had better not require the 593 // same register for uses and defs. 594 if (RI.Tied) { 595 markValueUsed(&VirtReg, ParentVNI); 596 LLVM_DEBUG(dbgs() << "\tcannot remat tied reg: " << UseIdx << '\t' << MI); 597 return false; 598 } 599 600 // Before rematerializing into a register for a single instruction, try to 601 // fold a load into the instruction. That avoids allocating a new register. 602 if (RM.OrigMI->canFoldAsLoad() && 603 foldMemoryOperand(Ops, RM.OrigMI)) { 604 Edit->markRematerialized(RM.ParentVNI); 605 ++NumFoldedLoads; 606 return true; 607 } 608 609 // If we can't guarantee that we'll be able to actually assign the new vreg, 610 // we can't remat. 611 if (!canGuaranteeAssignmentAfterRemat(VirtReg.reg, MI)) { 612 markValueUsed(&VirtReg, ParentVNI); 613 LLVM_DEBUG(dbgs() << "\tcannot remat for " << UseIdx << '\t' << MI); 614 return false; 615 } 616 617 // Allocate a new register for the remat. 618 Register NewVReg = Edit->createFrom(Original); 619 620 // Finally we can rematerialize OrigMI before MI. 621 SlotIndex DefIdx = 622 Edit->rematerializeAt(*MI.getParent(), MI, NewVReg, RM, TRI); 623 624 // We take the DebugLoc from MI, since OrigMI may be attributed to a 625 // different source location. 626 auto *NewMI = LIS.getInstructionFromIndex(DefIdx); 627 NewMI->setDebugLoc(MI.getDebugLoc()); 628 629 (void)DefIdx; 630 LLVM_DEBUG(dbgs() << "\tremat: " << DefIdx << '\t' 631 << *LIS.getInstructionFromIndex(DefIdx)); 632 633 // Replace operands 634 for (const auto &OpPair : Ops) { 635 MachineOperand &MO = OpPair.first->getOperand(OpPair.second); 636 if (MO.isReg() && MO.isUse() && MO.getReg() == VirtReg.reg) { 637 MO.setReg(NewVReg); 638 MO.setIsKill(); 639 } 640 } 641 LLVM_DEBUG(dbgs() << "\t " << UseIdx << '\t' << MI << '\n'); 642 643 ++NumRemats; 644 return true; 645 } 646 647 /// reMaterializeAll - Try to rematerialize as many uses as possible, 648 /// and trim the live ranges after. 649 void InlineSpiller::reMaterializeAll() { 650 if (!Edit->anyRematerializable(AA)) 651 return; 652 653 UsedValues.clear(); 654 655 // Try to remat before all uses of snippets. 656 bool anyRemat = false; 657 for (Register Reg : RegsToSpill) { 658 LiveInterval &LI = LIS.getInterval(Reg); 659 for (MachineRegisterInfo::reg_bundle_iterator 660 RegI = MRI.reg_bundle_begin(Reg), E = MRI.reg_bundle_end(); 661 RegI != E; ) { 662 MachineInstr &MI = *RegI++; 663 664 // Debug values are not allowed to affect codegen. 665 if (MI.isDebugValue()) 666 continue; 667 668 assert(!MI.isDebugInstr() && "Did not expect to find a use in debug " 669 "instruction that isn't a DBG_VALUE"); 670 671 anyRemat |= reMaterializeFor(LI, MI); 672 } 673 } 674 if (!anyRemat) 675 return; 676 677 // Remove any values that were completely rematted. 678 for (Register Reg : RegsToSpill) { 679 LiveInterval &LI = LIS.getInterval(Reg); 680 for (LiveInterval::vni_iterator I = LI.vni_begin(), E = LI.vni_end(); 681 I != E; ++I) { 682 VNInfo *VNI = *I; 683 if (VNI->isUnused() || VNI->isPHIDef() || UsedValues.count(VNI)) 684 continue; 685 MachineInstr *MI = LIS.getInstructionFromIndex(VNI->def); 686 MI->addRegisterDead(Reg, &TRI); 687 if (!MI->allDefsAreDead()) 688 continue; 689 LLVM_DEBUG(dbgs() << "All defs dead: " << *MI); 690 DeadDefs.push_back(MI); 691 } 692 } 693 694 // Eliminate dead code after remat. Note that some snippet copies may be 695 // deleted here. 696 if (DeadDefs.empty()) 697 return; 698 LLVM_DEBUG(dbgs() << "Remat created " << DeadDefs.size() << " dead defs.\n"); 699 Edit->eliminateDeadDefs(DeadDefs, RegsToSpill, AA); 700 701 // LiveRangeEdit::eliminateDeadDef is used to remove dead define instructions 702 // after rematerialization. To remove a VNI for a vreg from its LiveInterval, 703 // LiveIntervals::removeVRegDefAt is used. However, after non-PHI VNIs are all 704 // removed, PHI VNI are still left in the LiveInterval. 705 // So to get rid of unused reg, we need to check whether it has non-dbg 706 // reference instead of whether it has non-empty interval. 707 unsigned ResultPos = 0; 708 for (Register Reg : RegsToSpill) { 709 if (MRI.reg_nodbg_empty(Reg)) { 710 Edit->eraseVirtReg(Reg); 711 continue; 712 } 713 714 assert(LIS.hasInterval(Reg) && 715 (!LIS.getInterval(Reg).empty() || !MRI.reg_nodbg_empty(Reg)) && 716 "Empty and not used live-range?!"); 717 718 RegsToSpill[ResultPos++] = Reg; 719 } 720 RegsToSpill.erase(RegsToSpill.begin() + ResultPos, RegsToSpill.end()); 721 LLVM_DEBUG(dbgs() << RegsToSpill.size() 722 << " registers to spill after remat.\n"); 723 } 724 725 //===----------------------------------------------------------------------===// 726 // Spilling 727 //===----------------------------------------------------------------------===// 728 729 /// If MI is a load or store of StackSlot, it can be removed. 730 bool InlineSpiller::coalesceStackAccess(MachineInstr *MI, Register Reg) { 731 int FI = 0; 732 Register InstrReg = TII.isLoadFromStackSlot(*MI, FI); 733 bool IsLoad = InstrReg; 734 if (!IsLoad) 735 InstrReg = TII.isStoreToStackSlot(*MI, FI); 736 737 // We have a stack access. Is it the right register and slot? 738 if (InstrReg != Reg || FI != StackSlot) 739 return false; 740 741 if (!IsLoad) 742 HSpiller.rmFromMergeableSpills(*MI, StackSlot); 743 744 LLVM_DEBUG(dbgs() << "Coalescing stack access: " << *MI); 745 LIS.RemoveMachineInstrFromMaps(*MI); 746 MI->eraseFromParent(); 747 748 if (IsLoad) { 749 ++NumReloadsRemoved; 750 --NumReloads; 751 } else { 752 ++NumSpillsRemoved; 753 --NumSpills; 754 } 755 756 return true; 757 } 758 759 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 760 LLVM_DUMP_METHOD 761 // Dump the range of instructions from B to E with their slot indexes. 762 static void dumpMachineInstrRangeWithSlotIndex(MachineBasicBlock::iterator B, 763 MachineBasicBlock::iterator E, 764 LiveIntervals const &LIS, 765 const char *const header, 766 Register VReg = Register()) { 767 char NextLine = '\n'; 768 char SlotIndent = '\t'; 769 770 if (std::next(B) == E) { 771 NextLine = ' '; 772 SlotIndent = ' '; 773 } 774 775 dbgs() << '\t' << header << ": " << NextLine; 776 777 for (MachineBasicBlock::iterator I = B; I != E; ++I) { 778 SlotIndex Idx = LIS.getInstructionIndex(*I).getRegSlot(); 779 780 // If a register was passed in and this instruction has it as a 781 // destination that is marked as an early clobber, print the 782 // early-clobber slot index. 783 if (VReg) { 784 MachineOperand *MO = I->findRegisterDefOperand(VReg); 785 if (MO && MO->isEarlyClobber()) 786 Idx = Idx.getRegSlot(true); 787 } 788 789 dbgs() << SlotIndent << Idx << '\t' << *I; 790 } 791 } 792 #endif 793 794 /// foldMemoryOperand - Try folding stack slot references in Ops into their 795 /// instructions. 796 /// 797 /// @param Ops Operand indices from AnalyzeVirtRegInBundle(). 798 /// @param LoadMI Load instruction to use instead of stack slot when non-null. 799 /// @return True on success. 800 bool InlineSpiller:: 801 foldMemoryOperand(ArrayRef<std::pair<MachineInstr *, unsigned>> Ops, 802 MachineInstr *LoadMI) { 803 if (Ops.empty()) 804 return false; 805 // Don't attempt folding in bundles. 806 MachineInstr *MI = Ops.front().first; 807 if (Ops.back().first != MI || MI->isBundled()) 808 return false; 809 810 bool WasCopy = MI->isCopy(); 811 Register ImpReg; 812 813 // Spill subregs if the target allows it. 814 // We always want to spill subregs for stackmap/patchpoint pseudos. 815 bool SpillSubRegs = TII.isSubregFoldable() || 816 MI->getOpcode() == TargetOpcode::STATEPOINT || 817 MI->getOpcode() == TargetOpcode::PATCHPOINT || 818 MI->getOpcode() == TargetOpcode::STACKMAP; 819 820 // TargetInstrInfo::foldMemoryOperand only expects explicit, non-tied 821 // operands. 822 SmallVector<unsigned, 8> FoldOps; 823 for (const auto &OpPair : Ops) { 824 unsigned Idx = OpPair.second; 825 assert(MI == OpPair.first && "Instruction conflict during operand folding"); 826 MachineOperand &MO = MI->getOperand(Idx); 827 if (MO.isImplicit()) { 828 ImpReg = MO.getReg(); 829 continue; 830 } 831 832 if (!SpillSubRegs && MO.getSubReg()) 833 return false; 834 // We cannot fold a load instruction into a def. 835 if (LoadMI && MO.isDef()) 836 return false; 837 // Tied use operands should not be passed to foldMemoryOperand. 838 if (!MI->isRegTiedToDefOperand(Idx)) 839 FoldOps.push_back(Idx); 840 } 841 842 // If we only have implicit uses, we won't be able to fold that. 843 // Moreover, TargetInstrInfo::foldMemoryOperand will assert if we try! 844 if (FoldOps.empty()) 845 return false; 846 847 MachineInstrSpan MIS(MI, MI->getParent()); 848 849 MachineInstr *FoldMI = 850 LoadMI ? TII.foldMemoryOperand(*MI, FoldOps, *LoadMI, &LIS) 851 : TII.foldMemoryOperand(*MI, FoldOps, StackSlot, &LIS, &VRM); 852 if (!FoldMI) 853 return false; 854 855 // Remove LIS for any dead defs in the original MI not in FoldMI. 856 for (MIBundleOperands MO(*MI); MO.isValid(); ++MO) { 857 if (!MO->isReg()) 858 continue; 859 Register Reg = MO->getReg(); 860 if (!Reg || Register::isVirtualRegister(Reg) || MRI.isReserved(Reg)) { 861 continue; 862 } 863 // Skip non-Defs, including undef uses and internal reads. 864 if (MO->isUse()) 865 continue; 866 PhysRegInfo RI = AnalyzePhysRegInBundle(*FoldMI, Reg, &TRI); 867 if (RI.FullyDefined) 868 continue; 869 // FoldMI does not define this physreg. Remove the LI segment. 870 assert(MO->isDead() && "Cannot fold physreg def"); 871 SlotIndex Idx = LIS.getInstructionIndex(*MI).getRegSlot(); 872 LIS.removePhysRegDefAt(Reg, Idx); 873 } 874 875 int FI; 876 if (TII.isStoreToStackSlot(*MI, FI) && 877 HSpiller.rmFromMergeableSpills(*MI, FI)) 878 --NumSpills; 879 LIS.ReplaceMachineInstrInMaps(*MI, *FoldMI); 880 // Update the call site info. 881 if (MI->isCandidateForCallSiteEntry()) 882 MI->getMF()->moveCallSiteInfo(MI, FoldMI); 883 MI->eraseFromParent(); 884 885 // Insert any new instructions other than FoldMI into the LIS maps. 886 assert(!MIS.empty() && "Unexpected empty span of instructions!"); 887 for (MachineInstr &MI : MIS) 888 if (&MI != FoldMI) 889 LIS.InsertMachineInstrInMaps(MI); 890 891 // TII.foldMemoryOperand may have left some implicit operands on the 892 // instruction. Strip them. 893 if (ImpReg) 894 for (unsigned i = FoldMI->getNumOperands(); i; --i) { 895 MachineOperand &MO = FoldMI->getOperand(i - 1); 896 if (!MO.isReg() || !MO.isImplicit()) 897 break; 898 if (MO.getReg() == ImpReg) 899 FoldMI->RemoveOperand(i - 1); 900 } 901 902 LLVM_DEBUG(dumpMachineInstrRangeWithSlotIndex(MIS.begin(), MIS.end(), LIS, 903 "folded")); 904 905 if (!WasCopy) 906 ++NumFolded; 907 else if (Ops.front().second == 0) { 908 ++NumSpills; 909 HSpiller.addToMergeableSpills(*FoldMI, StackSlot, Original); 910 } else 911 ++NumReloads; 912 return true; 913 } 914 915 void InlineSpiller::insertReload(Register NewVReg, 916 SlotIndex Idx, 917 MachineBasicBlock::iterator MI) { 918 MachineBasicBlock &MBB = *MI->getParent(); 919 920 MachineInstrSpan MIS(MI, &MBB); 921 TII.loadRegFromStackSlot(MBB, MI, NewVReg, StackSlot, 922 MRI.getRegClass(NewVReg), &TRI); 923 924 LIS.InsertMachineInstrRangeInMaps(MIS.begin(), MI); 925 926 LLVM_DEBUG(dumpMachineInstrRangeWithSlotIndex(MIS.begin(), MI, LIS, "reload", 927 NewVReg)); 928 ++NumReloads; 929 } 930 931 /// Check if \p Def fully defines a VReg with an undefined value. 932 /// If that's the case, that means the value of VReg is actually 933 /// not relevant. 934 static bool isRealSpill(const MachineInstr &Def) { 935 if (!Def.isImplicitDef()) 936 return true; 937 assert(Def.getNumOperands() == 1 && 938 "Implicit def with more than one definition"); 939 // We can say that the VReg defined by Def is undef, only if it is 940 // fully defined by Def. Otherwise, some of the lanes may not be 941 // undef and the value of the VReg matters. 942 return Def.getOperand(0).getSubReg(); 943 } 944 945 /// insertSpill - Insert a spill of NewVReg after MI. 946 void InlineSpiller::insertSpill(Register NewVReg, bool isKill, 947 MachineBasicBlock::iterator MI) { 948 // Spill are not terminators, so inserting spills after terminators will 949 // violate invariants in MachineVerifier. 950 assert(!MI->isTerminator() && "Inserting a spill after a terminator"); 951 MachineBasicBlock &MBB = *MI->getParent(); 952 953 MachineInstrSpan MIS(MI, &MBB); 954 MachineBasicBlock::iterator SpillBefore = std::next(MI); 955 bool IsRealSpill = isRealSpill(*MI); 956 if (IsRealSpill) 957 TII.storeRegToStackSlot(MBB, SpillBefore, NewVReg, isKill, StackSlot, 958 MRI.getRegClass(NewVReg), &TRI); 959 else 960 // Don't spill undef value. 961 // Anything works for undef, in particular keeping the memory 962 // uninitialized is a viable option and it saves code size and 963 // run time. 964 BuildMI(MBB, SpillBefore, MI->getDebugLoc(), TII.get(TargetOpcode::KILL)) 965 .addReg(NewVReg, getKillRegState(isKill)); 966 967 MachineBasicBlock::iterator Spill = std::next(MI); 968 LIS.InsertMachineInstrRangeInMaps(Spill, MIS.end()); 969 970 LLVM_DEBUG( 971 dumpMachineInstrRangeWithSlotIndex(Spill, MIS.end(), LIS, "spill")); 972 ++NumSpills; 973 if (IsRealSpill) 974 HSpiller.addToMergeableSpills(*Spill, StackSlot, Original); 975 } 976 977 /// spillAroundUses - insert spill code around each use of Reg. 978 void InlineSpiller::spillAroundUses(Register Reg) { 979 LLVM_DEBUG(dbgs() << "spillAroundUses " << printReg(Reg) << '\n'); 980 LiveInterval &OldLI = LIS.getInterval(Reg); 981 982 // Iterate over instructions using Reg. 983 for (MachineRegisterInfo::reg_bundle_iterator 984 RegI = MRI.reg_bundle_begin(Reg), E = MRI.reg_bundle_end(); 985 RegI != E; ) { 986 MachineInstr *MI = &*(RegI++); 987 988 // Debug values are not allowed to affect codegen. 989 if (MI->isDebugValue()) { 990 // Modify DBG_VALUE now that the value is in a spill slot. 991 MachineBasicBlock *MBB = MI->getParent(); 992 LLVM_DEBUG(dbgs() << "Modifying debug info due to spill:\t" << *MI); 993 buildDbgValueForSpill(*MBB, MI, *MI, StackSlot); 994 MBB->erase(MI); 995 continue; 996 } 997 998 assert(!MI->isDebugInstr() && "Did not expect to find a use in debug " 999 "instruction that isn't a DBG_VALUE"); 1000 1001 // Ignore copies to/from snippets. We'll delete them. 1002 if (SnippetCopies.count(MI)) 1003 continue; 1004 1005 // Stack slot accesses may coalesce away. 1006 if (coalesceStackAccess(MI, Reg)) 1007 continue; 1008 1009 // Analyze instruction. 1010 SmallVector<std::pair<MachineInstr*, unsigned>, 8> Ops; 1011 VirtRegInfo RI = AnalyzeVirtRegInBundle(*MI, Reg, &Ops); 1012 1013 // Find the slot index where this instruction reads and writes OldLI. 1014 // This is usually the def slot, except for tied early clobbers. 1015 SlotIndex Idx = LIS.getInstructionIndex(*MI).getRegSlot(); 1016 if (VNInfo *VNI = OldLI.getVNInfoAt(Idx.getRegSlot(true))) 1017 if (SlotIndex::isSameInstr(Idx, VNI->def)) 1018 Idx = VNI->def; 1019 1020 // Check for a sibling copy. 1021 Register SibReg = isFullCopyOf(*MI, Reg); 1022 if (SibReg && isSibling(SibReg)) { 1023 // This may actually be a copy between snippets. 1024 if (isRegToSpill(SibReg)) { 1025 LLVM_DEBUG(dbgs() << "Found new snippet copy: " << *MI); 1026 SnippetCopies.insert(MI); 1027 continue; 1028 } 1029 if (RI.Writes) { 1030 if (hoistSpillInsideBB(OldLI, *MI)) { 1031 // This COPY is now dead, the value is already in the stack slot. 1032 MI->getOperand(0).setIsDead(); 1033 DeadDefs.push_back(MI); 1034 continue; 1035 } 1036 } else { 1037 // This is a reload for a sib-reg copy. Drop spills downstream. 1038 LiveInterval &SibLI = LIS.getInterval(SibReg); 1039 eliminateRedundantSpills(SibLI, SibLI.getVNInfoAt(Idx)); 1040 // The COPY will fold to a reload below. 1041 } 1042 } 1043 1044 // Attempt to fold memory ops. 1045 if (foldMemoryOperand(Ops)) 1046 continue; 1047 1048 // Create a new virtual register for spill/fill. 1049 // FIXME: Infer regclass from instruction alone. 1050 Register NewVReg = Edit->createFrom(Reg); 1051 1052 if (RI.Reads) 1053 insertReload(NewVReg, Idx, MI); 1054 1055 // Rewrite instruction operands. 1056 bool hasLiveDef = false; 1057 for (const auto &OpPair : Ops) { 1058 MachineOperand &MO = OpPair.first->getOperand(OpPair.second); 1059 MO.setReg(NewVReg); 1060 if (MO.isUse()) { 1061 if (!OpPair.first->isRegTiedToDefOperand(OpPair.second)) 1062 MO.setIsKill(); 1063 } else { 1064 if (!MO.isDead()) 1065 hasLiveDef = true; 1066 } 1067 } 1068 LLVM_DEBUG(dbgs() << "\trewrite: " << Idx << '\t' << *MI << '\n'); 1069 1070 // FIXME: Use a second vreg if instruction has no tied ops. 1071 if (RI.Writes) 1072 if (hasLiveDef) 1073 insertSpill(NewVReg, true, MI); 1074 } 1075 } 1076 1077 /// spillAll - Spill all registers remaining after rematerialization. 1078 void InlineSpiller::spillAll() { 1079 // Update LiveStacks now that we are committed to spilling. 1080 if (StackSlot == VirtRegMap::NO_STACK_SLOT) { 1081 StackSlot = VRM.assignVirt2StackSlot(Original); 1082 StackInt = &LSS.getOrCreateInterval(StackSlot, MRI.getRegClass(Original)); 1083 StackInt->getNextValue(SlotIndex(), LSS.getVNInfoAllocator()); 1084 } else 1085 StackInt = &LSS.getInterval(StackSlot); 1086 1087 if (Original != Edit->getReg()) 1088 VRM.assignVirt2StackSlot(Edit->getReg(), StackSlot); 1089 1090 assert(StackInt->getNumValNums() == 1 && "Bad stack interval values"); 1091 for (Register Reg : RegsToSpill) 1092 StackInt->MergeSegmentsInAsValue(LIS.getInterval(Reg), 1093 StackInt->getValNumInfo(0)); 1094 LLVM_DEBUG(dbgs() << "Merged spilled regs: " << *StackInt << '\n'); 1095 1096 // Spill around uses of all RegsToSpill. 1097 for (Register Reg : RegsToSpill) 1098 spillAroundUses(Reg); 1099 1100 // Hoisted spills may cause dead code. 1101 if (!DeadDefs.empty()) { 1102 LLVM_DEBUG(dbgs() << "Eliminating " << DeadDefs.size() << " dead defs\n"); 1103 Edit->eliminateDeadDefs(DeadDefs, RegsToSpill, AA); 1104 } 1105 1106 // Finally delete the SnippetCopies. 1107 for (Register Reg : RegsToSpill) { 1108 for (MachineRegisterInfo::reg_instr_iterator 1109 RI = MRI.reg_instr_begin(Reg), E = MRI.reg_instr_end(); 1110 RI != E; ) { 1111 MachineInstr &MI = *(RI++); 1112 assert(SnippetCopies.count(&MI) && "Remaining use wasn't a snippet copy"); 1113 // FIXME: Do this with a LiveRangeEdit callback. 1114 LIS.RemoveMachineInstrFromMaps(MI); 1115 MI.eraseFromParent(); 1116 } 1117 } 1118 1119 // Delete all spilled registers. 1120 for (Register Reg : RegsToSpill) 1121 Edit->eraseVirtReg(Reg); 1122 } 1123 1124 void InlineSpiller::spill(LiveRangeEdit &edit) { 1125 ++NumSpilledRanges; 1126 Edit = &edit; 1127 assert(!Register::isStackSlot(edit.getReg()) && 1128 "Trying to spill a stack slot."); 1129 // Share a stack slot among all descendants of Original. 1130 Original = VRM.getOriginal(edit.getReg()); 1131 StackSlot = VRM.getStackSlot(Original); 1132 StackInt = nullptr; 1133 1134 LLVM_DEBUG(dbgs() << "Inline spilling " 1135 << TRI.getRegClassName(MRI.getRegClass(edit.getReg())) 1136 << ':' << edit.getParent() << "\nFrom original " 1137 << printReg(Original) << '\n'); 1138 assert(edit.getParent().isSpillable() && 1139 "Attempting to spill already spilled value."); 1140 assert(DeadDefs.empty() && "Previous spill didn't remove dead defs"); 1141 1142 collectRegsToSpill(); 1143 reMaterializeAll(); 1144 1145 // Remat may handle everything. 1146 if (!RegsToSpill.empty()) 1147 spillAll(); 1148 1149 Edit->calculateRegClassAndHint(MF, Loops, MBFI); 1150 } 1151 1152 /// Optimizations after all the reg selections and spills are done. 1153 void InlineSpiller::postOptimization() { HSpiller.hoistAllSpills(); } 1154 1155 /// When a spill is inserted, add the spill to MergeableSpills map. 1156 void HoistSpillHelper::addToMergeableSpills(MachineInstr &Spill, int StackSlot, 1157 unsigned Original) { 1158 BumpPtrAllocator &Allocator = LIS.getVNInfoAllocator(); 1159 LiveInterval &OrigLI = LIS.getInterval(Original); 1160 // save a copy of LiveInterval in StackSlotToOrigLI because the original 1161 // LiveInterval may be cleared after all its references are spilled. 1162 if (StackSlotToOrigLI.find(StackSlot) == StackSlotToOrigLI.end()) { 1163 auto LI = std::make_unique<LiveInterval>(OrigLI.reg, OrigLI.weight); 1164 LI->assign(OrigLI, Allocator); 1165 StackSlotToOrigLI[StackSlot] = std::move(LI); 1166 } 1167 SlotIndex Idx = LIS.getInstructionIndex(Spill); 1168 VNInfo *OrigVNI = StackSlotToOrigLI[StackSlot]->getVNInfoAt(Idx.getRegSlot()); 1169 std::pair<int, VNInfo *> MIdx = std::make_pair(StackSlot, OrigVNI); 1170 MergeableSpills[MIdx].insert(&Spill); 1171 } 1172 1173 /// When a spill is removed, remove the spill from MergeableSpills map. 1174 /// Return true if the spill is removed successfully. 1175 bool HoistSpillHelper::rmFromMergeableSpills(MachineInstr &Spill, 1176 int StackSlot) { 1177 auto It = StackSlotToOrigLI.find(StackSlot); 1178 if (It == StackSlotToOrigLI.end()) 1179 return false; 1180 SlotIndex Idx = LIS.getInstructionIndex(Spill); 1181 VNInfo *OrigVNI = It->second->getVNInfoAt(Idx.getRegSlot()); 1182 std::pair<int, VNInfo *> MIdx = std::make_pair(StackSlot, OrigVNI); 1183 return MergeableSpills[MIdx].erase(&Spill); 1184 } 1185 1186 /// Check BB to see if it is a possible target BB to place a hoisted spill, 1187 /// i.e., there should be a living sibling of OrigReg at the insert point. 1188 bool HoistSpillHelper::isSpillCandBB(LiveInterval &OrigLI, VNInfo &OrigVNI, 1189 MachineBasicBlock &BB, Register &LiveReg) { 1190 SlotIndex Idx; 1191 Register OrigReg = OrigLI.reg; 1192 MachineBasicBlock::iterator MI = IPA.getLastInsertPointIter(OrigLI, BB); 1193 if (MI != BB.end()) 1194 Idx = LIS.getInstructionIndex(*MI); 1195 else 1196 Idx = LIS.getMBBEndIdx(&BB).getPrevSlot(); 1197 SmallSetVector<Register, 16> &Siblings = Virt2SiblingsMap[OrigReg]; 1198 assert(OrigLI.getVNInfoAt(Idx) == &OrigVNI && "Unexpected VNI"); 1199 1200 for (const Register &SibReg : Siblings) { 1201 LiveInterval &LI = LIS.getInterval(SibReg); 1202 VNInfo *VNI = LI.getVNInfoAt(Idx); 1203 if (VNI) { 1204 LiveReg = SibReg; 1205 return true; 1206 } 1207 } 1208 return false; 1209 } 1210 1211 /// Remove redundant spills in the same BB. Save those redundant spills in 1212 /// SpillsToRm, and save the spill to keep and its BB in SpillBBToSpill map. 1213 void HoistSpillHelper::rmRedundantSpills( 1214 SmallPtrSet<MachineInstr *, 16> &Spills, 1215 SmallVectorImpl<MachineInstr *> &SpillsToRm, 1216 DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill) { 1217 // For each spill saw, check SpillBBToSpill[] and see if its BB already has 1218 // another spill inside. If a BB contains more than one spill, only keep the 1219 // earlier spill with smaller SlotIndex. 1220 for (const auto CurrentSpill : Spills) { 1221 MachineBasicBlock *Block = CurrentSpill->getParent(); 1222 MachineDomTreeNode *Node = MDT.getBase().getNode(Block); 1223 MachineInstr *PrevSpill = SpillBBToSpill[Node]; 1224 if (PrevSpill) { 1225 SlotIndex PIdx = LIS.getInstructionIndex(*PrevSpill); 1226 SlotIndex CIdx = LIS.getInstructionIndex(*CurrentSpill); 1227 MachineInstr *SpillToRm = (CIdx > PIdx) ? CurrentSpill : PrevSpill; 1228 MachineInstr *SpillToKeep = (CIdx > PIdx) ? PrevSpill : CurrentSpill; 1229 SpillsToRm.push_back(SpillToRm); 1230 SpillBBToSpill[MDT.getBase().getNode(Block)] = SpillToKeep; 1231 } else { 1232 SpillBBToSpill[MDT.getBase().getNode(Block)] = CurrentSpill; 1233 } 1234 } 1235 for (const auto SpillToRm : SpillsToRm) 1236 Spills.erase(SpillToRm); 1237 } 1238 1239 /// Starting from \p Root find a top-down traversal order of the dominator 1240 /// tree to visit all basic blocks containing the elements of \p Spills. 1241 /// Redundant spills will be found and put into \p SpillsToRm at the same 1242 /// time. \p SpillBBToSpill will be populated as part of the process and 1243 /// maps a basic block to the first store occurring in the basic block. 1244 /// \post SpillsToRm.union(Spills\@post) == Spills\@pre 1245 void HoistSpillHelper::getVisitOrders( 1246 MachineBasicBlock *Root, SmallPtrSet<MachineInstr *, 16> &Spills, 1247 SmallVectorImpl<MachineDomTreeNode *> &Orders, 1248 SmallVectorImpl<MachineInstr *> &SpillsToRm, 1249 DenseMap<MachineDomTreeNode *, unsigned> &SpillsToKeep, 1250 DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill) { 1251 // The set contains all the possible BB nodes to which we may hoist 1252 // original spills. 1253 SmallPtrSet<MachineDomTreeNode *, 8> WorkSet; 1254 // Save the BB nodes on the path from the first BB node containing 1255 // non-redundant spill to the Root node. 1256 SmallPtrSet<MachineDomTreeNode *, 8> NodesOnPath; 1257 // All the spills to be hoisted must originate from a single def instruction 1258 // to the OrigReg. It means the def instruction should dominate all the spills 1259 // to be hoisted. We choose the BB where the def instruction is located as 1260 // the Root. 1261 MachineDomTreeNode *RootIDomNode = MDT[Root]->getIDom(); 1262 // For every node on the dominator tree with spill, walk up on the dominator 1263 // tree towards the Root node until it is reached. If there is other node 1264 // containing spill in the middle of the path, the previous spill saw will 1265 // be redundant and the node containing it will be removed. All the nodes on 1266 // the path starting from the first node with non-redundant spill to the Root 1267 // node will be added to the WorkSet, which will contain all the possible 1268 // locations where spills may be hoisted to after the loop below is done. 1269 for (const auto Spill : Spills) { 1270 MachineBasicBlock *Block = Spill->getParent(); 1271 MachineDomTreeNode *Node = MDT[Block]; 1272 MachineInstr *SpillToRm = nullptr; 1273 while (Node != RootIDomNode) { 1274 // If Node dominates Block, and it already contains a spill, the spill in 1275 // Block will be redundant. 1276 if (Node != MDT[Block] && SpillBBToSpill[Node]) { 1277 SpillToRm = SpillBBToSpill[MDT[Block]]; 1278 break; 1279 /// If we see the Node already in WorkSet, the path from the Node to 1280 /// the Root node must already be traversed by another spill. 1281 /// Then no need to repeat. 1282 } else if (WorkSet.count(Node)) { 1283 break; 1284 } else { 1285 NodesOnPath.insert(Node); 1286 } 1287 Node = Node->getIDom(); 1288 } 1289 if (SpillToRm) { 1290 SpillsToRm.push_back(SpillToRm); 1291 } else { 1292 // Add a BB containing the original spills to SpillsToKeep -- i.e., 1293 // set the initial status before hoisting start. The value of BBs 1294 // containing original spills is set to 0, in order to descriminate 1295 // with BBs containing hoisted spills which will be inserted to 1296 // SpillsToKeep later during hoisting. 1297 SpillsToKeep[MDT[Block]] = 0; 1298 WorkSet.insert(NodesOnPath.begin(), NodesOnPath.end()); 1299 } 1300 NodesOnPath.clear(); 1301 } 1302 1303 // Sort the nodes in WorkSet in top-down order and save the nodes 1304 // in Orders. Orders will be used for hoisting in runHoistSpills. 1305 unsigned idx = 0; 1306 Orders.push_back(MDT.getBase().getNode(Root)); 1307 do { 1308 MachineDomTreeNode *Node = Orders[idx++]; 1309 for (MachineDomTreeNode *Child : Node->children()) { 1310 if (WorkSet.count(Child)) 1311 Orders.push_back(Child); 1312 } 1313 } while (idx != Orders.size()); 1314 assert(Orders.size() == WorkSet.size() && 1315 "Orders have different size with WorkSet"); 1316 1317 #ifndef NDEBUG 1318 LLVM_DEBUG(dbgs() << "Orders size is " << Orders.size() << "\n"); 1319 SmallVector<MachineDomTreeNode *, 32>::reverse_iterator RIt = Orders.rbegin(); 1320 for (; RIt != Orders.rend(); RIt++) 1321 LLVM_DEBUG(dbgs() << "BB" << (*RIt)->getBlock()->getNumber() << ","); 1322 LLVM_DEBUG(dbgs() << "\n"); 1323 #endif 1324 } 1325 1326 /// Try to hoist spills according to BB hotness. The spills to removed will 1327 /// be saved in \p SpillsToRm. The spills to be inserted will be saved in 1328 /// \p SpillsToIns. 1329 void HoistSpillHelper::runHoistSpills( 1330 LiveInterval &OrigLI, VNInfo &OrigVNI, 1331 SmallPtrSet<MachineInstr *, 16> &Spills, 1332 SmallVectorImpl<MachineInstr *> &SpillsToRm, 1333 DenseMap<MachineBasicBlock *, unsigned> &SpillsToIns) { 1334 // Visit order of dominator tree nodes. 1335 SmallVector<MachineDomTreeNode *, 32> Orders; 1336 // SpillsToKeep contains all the nodes where spills are to be inserted 1337 // during hoisting. If the spill to be inserted is an original spill 1338 // (not a hoisted one), the value of the map entry is 0. If the spill 1339 // is a hoisted spill, the value of the map entry is the VReg to be used 1340 // as the source of the spill. 1341 DenseMap<MachineDomTreeNode *, unsigned> SpillsToKeep; 1342 // Map from BB to the first spill inside of it. 1343 DenseMap<MachineDomTreeNode *, MachineInstr *> SpillBBToSpill; 1344 1345 rmRedundantSpills(Spills, SpillsToRm, SpillBBToSpill); 1346 1347 MachineBasicBlock *Root = LIS.getMBBFromIndex(OrigVNI.def); 1348 getVisitOrders(Root, Spills, Orders, SpillsToRm, SpillsToKeep, 1349 SpillBBToSpill); 1350 1351 // SpillsInSubTreeMap keeps the map from a dom tree node to a pair of 1352 // nodes set and the cost of all the spills inside those nodes. 1353 // The nodes set are the locations where spills are to be inserted 1354 // in the subtree of current node. 1355 using NodesCostPair = 1356 std::pair<SmallPtrSet<MachineDomTreeNode *, 16>, BlockFrequency>; 1357 DenseMap<MachineDomTreeNode *, NodesCostPair> SpillsInSubTreeMap; 1358 1359 // Iterate Orders set in reverse order, which will be a bottom-up order 1360 // in the dominator tree. Once we visit a dom tree node, we know its 1361 // children have already been visited and the spill locations in the 1362 // subtrees of all the children have been determined. 1363 SmallVector<MachineDomTreeNode *, 32>::reverse_iterator RIt = Orders.rbegin(); 1364 for (; RIt != Orders.rend(); RIt++) { 1365 MachineBasicBlock *Block = (*RIt)->getBlock(); 1366 1367 // If Block contains an original spill, simply continue. 1368 if (SpillsToKeep.find(*RIt) != SpillsToKeep.end() && !SpillsToKeep[*RIt]) { 1369 SpillsInSubTreeMap[*RIt].first.insert(*RIt); 1370 // SpillsInSubTreeMap[*RIt].second contains the cost of spill. 1371 SpillsInSubTreeMap[*RIt].second = MBFI.getBlockFreq(Block); 1372 continue; 1373 } 1374 1375 // Collect spills in subtree of current node (*RIt) to 1376 // SpillsInSubTreeMap[*RIt].first. 1377 for (MachineDomTreeNode *Child : (*RIt)->children()) { 1378 if (SpillsInSubTreeMap.find(Child) == SpillsInSubTreeMap.end()) 1379 continue; 1380 // The stmt "SpillsInSubTree = SpillsInSubTreeMap[*RIt].first" below 1381 // should be placed before getting the begin and end iterators of 1382 // SpillsInSubTreeMap[Child].first, or else the iterators may be 1383 // invalidated when SpillsInSubTreeMap[*RIt] is seen the first time 1384 // and the map grows and then the original buckets in the map are moved. 1385 SmallPtrSet<MachineDomTreeNode *, 16> &SpillsInSubTree = 1386 SpillsInSubTreeMap[*RIt].first; 1387 BlockFrequency &SubTreeCost = SpillsInSubTreeMap[*RIt].second; 1388 SubTreeCost += SpillsInSubTreeMap[Child].second; 1389 auto BI = SpillsInSubTreeMap[Child].first.begin(); 1390 auto EI = SpillsInSubTreeMap[Child].first.end(); 1391 SpillsInSubTree.insert(BI, EI); 1392 SpillsInSubTreeMap.erase(Child); 1393 } 1394 1395 SmallPtrSet<MachineDomTreeNode *, 16> &SpillsInSubTree = 1396 SpillsInSubTreeMap[*RIt].first; 1397 BlockFrequency &SubTreeCost = SpillsInSubTreeMap[*RIt].second; 1398 // No spills in subtree, simply continue. 1399 if (SpillsInSubTree.empty()) 1400 continue; 1401 1402 // Check whether Block is a possible candidate to insert spill. 1403 Register LiveReg; 1404 if (!isSpillCandBB(OrigLI, OrigVNI, *Block, LiveReg)) 1405 continue; 1406 1407 // If there are multiple spills that could be merged, bias a little 1408 // to hoist the spill. 1409 BranchProbability MarginProb = (SpillsInSubTree.size() > 1) 1410 ? BranchProbability(9, 10) 1411 : BranchProbability(1, 1); 1412 if (SubTreeCost > MBFI.getBlockFreq(Block) * MarginProb) { 1413 // Hoist: Move spills to current Block. 1414 for (const auto SpillBB : SpillsInSubTree) { 1415 // When SpillBB is a BB contains original spill, insert the spill 1416 // to SpillsToRm. 1417 if (SpillsToKeep.find(SpillBB) != SpillsToKeep.end() && 1418 !SpillsToKeep[SpillBB]) { 1419 MachineInstr *SpillToRm = SpillBBToSpill[SpillBB]; 1420 SpillsToRm.push_back(SpillToRm); 1421 } 1422 // SpillBB will not contain spill anymore, remove it from SpillsToKeep. 1423 SpillsToKeep.erase(SpillBB); 1424 } 1425 // Current Block is the BB containing the new hoisted spill. Add it to 1426 // SpillsToKeep. LiveReg is the source of the new spill. 1427 SpillsToKeep[*RIt] = LiveReg; 1428 LLVM_DEBUG({ 1429 dbgs() << "spills in BB: "; 1430 for (const auto Rspill : SpillsInSubTree) 1431 dbgs() << Rspill->getBlock()->getNumber() << " "; 1432 dbgs() << "were promoted to BB" << (*RIt)->getBlock()->getNumber() 1433 << "\n"; 1434 }); 1435 SpillsInSubTree.clear(); 1436 SpillsInSubTree.insert(*RIt); 1437 SubTreeCost = MBFI.getBlockFreq(Block); 1438 } 1439 } 1440 // For spills in SpillsToKeep with LiveReg set (i.e., not original spill), 1441 // save them to SpillsToIns. 1442 for (const auto &Ent : SpillsToKeep) { 1443 if (Ent.second) 1444 SpillsToIns[Ent.first->getBlock()] = Ent.second; 1445 } 1446 } 1447 1448 /// For spills with equal values, remove redundant spills and hoist those left 1449 /// to less hot spots. 1450 /// 1451 /// Spills with equal values will be collected into the same set in 1452 /// MergeableSpills when spill is inserted. These equal spills are originated 1453 /// from the same defining instruction and are dominated by the instruction. 1454 /// Before hoisting all the equal spills, redundant spills inside in the same 1455 /// BB are first marked to be deleted. Then starting from the spills left, walk 1456 /// up on the dominator tree towards the Root node where the define instruction 1457 /// is located, mark the dominated spills to be deleted along the way and 1458 /// collect the BB nodes on the path from non-dominated spills to the define 1459 /// instruction into a WorkSet. The nodes in WorkSet are the candidate places 1460 /// where we are considering to hoist the spills. We iterate the WorkSet in 1461 /// bottom-up order, and for each node, we will decide whether to hoist spills 1462 /// inside its subtree to that node. In this way, we can get benefit locally 1463 /// even if hoisting all the equal spills to one cold place is impossible. 1464 void HoistSpillHelper::hoistAllSpills() { 1465 SmallVector<Register, 4> NewVRegs; 1466 LiveRangeEdit Edit(nullptr, NewVRegs, MF, LIS, &VRM, this); 1467 1468 for (unsigned i = 0, e = MRI.getNumVirtRegs(); i != e; ++i) { 1469 Register Reg = Register::index2VirtReg(i); 1470 Register Original = VRM.getPreSplitReg(Reg); 1471 if (!MRI.def_empty(Reg)) 1472 Virt2SiblingsMap[Original].insert(Reg); 1473 } 1474 1475 // Each entry in MergeableSpills contains a spill set with equal values. 1476 for (auto &Ent : MergeableSpills) { 1477 int Slot = Ent.first.first; 1478 LiveInterval &OrigLI = *StackSlotToOrigLI[Slot]; 1479 VNInfo *OrigVNI = Ent.first.second; 1480 SmallPtrSet<MachineInstr *, 16> &EqValSpills = Ent.second; 1481 if (Ent.second.empty()) 1482 continue; 1483 1484 LLVM_DEBUG({ 1485 dbgs() << "\nFor Slot" << Slot << " and VN" << OrigVNI->id << ":\n" 1486 << "Equal spills in BB: "; 1487 for (const auto spill : EqValSpills) 1488 dbgs() << spill->getParent()->getNumber() << " "; 1489 dbgs() << "\n"; 1490 }); 1491 1492 // SpillsToRm is the spill set to be removed from EqValSpills. 1493 SmallVector<MachineInstr *, 16> SpillsToRm; 1494 // SpillsToIns is the spill set to be newly inserted after hoisting. 1495 DenseMap<MachineBasicBlock *, unsigned> SpillsToIns; 1496 1497 runHoistSpills(OrigLI, *OrigVNI, EqValSpills, SpillsToRm, SpillsToIns); 1498 1499 LLVM_DEBUG({ 1500 dbgs() << "Finally inserted spills in BB: "; 1501 for (const auto &Ispill : SpillsToIns) 1502 dbgs() << Ispill.first->getNumber() << " "; 1503 dbgs() << "\nFinally removed spills in BB: "; 1504 for (const auto Rspill : SpillsToRm) 1505 dbgs() << Rspill->getParent()->getNumber() << " "; 1506 dbgs() << "\n"; 1507 }); 1508 1509 // Stack live range update. 1510 LiveInterval &StackIntvl = LSS.getInterval(Slot); 1511 if (!SpillsToIns.empty() || !SpillsToRm.empty()) 1512 StackIntvl.MergeValueInAsValue(OrigLI, OrigVNI, 1513 StackIntvl.getValNumInfo(0)); 1514 1515 // Insert hoisted spills. 1516 for (auto const &Insert : SpillsToIns) { 1517 MachineBasicBlock *BB = Insert.first; 1518 Register LiveReg = Insert.second; 1519 MachineBasicBlock::iterator MI = IPA.getLastInsertPointIter(OrigLI, *BB); 1520 TII.storeRegToStackSlot(*BB, MI, LiveReg, false, Slot, 1521 MRI.getRegClass(LiveReg), &TRI); 1522 LIS.InsertMachineInstrRangeInMaps(std::prev(MI), MI); 1523 ++NumSpills; 1524 } 1525 1526 // Remove redundant spills or change them to dead instructions. 1527 NumSpills -= SpillsToRm.size(); 1528 for (auto const RMEnt : SpillsToRm) { 1529 RMEnt->setDesc(TII.get(TargetOpcode::KILL)); 1530 for (unsigned i = RMEnt->getNumOperands(); i; --i) { 1531 MachineOperand &MO = RMEnt->getOperand(i - 1); 1532 if (MO.isReg() && MO.isImplicit() && MO.isDef() && !MO.isDead()) 1533 RMEnt->RemoveOperand(i - 1); 1534 } 1535 } 1536 Edit.eliminateDeadDefs(SpillsToRm, None, AA); 1537 } 1538 } 1539 1540 /// For VirtReg clone, the \p New register should have the same physreg or 1541 /// stackslot as the \p old register. 1542 void HoistSpillHelper::LRE_DidCloneVirtReg(unsigned New, unsigned Old) { 1543 if (VRM.hasPhys(Old)) 1544 VRM.assignVirt2Phys(New, VRM.getPhys(Old)); 1545 else if (VRM.getStackSlot(Old) != VirtRegMap::NO_STACK_SLOT) 1546 VRM.assignVirt2StackSlot(New, VRM.getStackSlot(Old)); 1547 else 1548 llvm_unreachable("VReg should be assigned either physreg or stackslot"); 1549 } 1550