1 //===-- RegAllocGreedy.cpp - greedy register allocator --------------------===// 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 // This file defines the RAGreedy function pass for register allocation in 11 // optimized builds. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "llvm/CodeGen/Passes.h" 16 #include "AllocationOrder.h" 17 #include "InterferenceCache.h" 18 #include "LiveDebugVariables.h" 19 #include "RegAllocBase.h" 20 #include "SpillPlacement.h" 21 #include "Spiller.h" 22 #include "SplitKit.h" 23 #include "llvm/ADT/Statistic.h" 24 #include "llvm/Analysis/AliasAnalysis.h" 25 #include "llvm/CodeGen/CalcSpillWeights.h" 26 #include "llvm/CodeGen/EdgeBundles.h" 27 #include "llvm/CodeGen/LiveIntervalAnalysis.h" 28 #include "llvm/CodeGen/LiveRangeEdit.h" 29 #include "llvm/CodeGen/LiveRegMatrix.h" 30 #include "llvm/CodeGen/LiveStackAnalysis.h" 31 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h" 32 #include "llvm/CodeGen/MachineDominators.h" 33 #include "llvm/CodeGen/MachineFunctionPass.h" 34 #include "llvm/CodeGen/MachineLoopInfo.h" 35 #include "llvm/CodeGen/MachineRegisterInfo.h" 36 #include "llvm/CodeGen/RegAllocRegistry.h" 37 #include "llvm/CodeGen/RegisterClassInfo.h" 38 #include "llvm/CodeGen/VirtRegMap.h" 39 #include "llvm/IR/LLVMContext.h" 40 #include "llvm/PassAnalysisSupport.h" 41 #include "llvm/Support/BranchProbability.h" 42 #include "llvm/Support/CommandLine.h" 43 #include "llvm/Support/Debug.h" 44 #include "llvm/Support/ErrorHandling.h" 45 #include "llvm/Support/Timer.h" 46 #include "llvm/Support/raw_ostream.h" 47 #include "llvm/Target/TargetSubtargetInfo.h" 48 #include <queue> 49 50 using namespace llvm; 51 52 #define DEBUG_TYPE "regalloc" 53 54 STATISTIC(NumGlobalSplits, "Number of split global live ranges"); 55 STATISTIC(NumLocalSplits, "Number of split local live ranges"); 56 STATISTIC(NumEvicted, "Number of interferences evicted"); 57 58 static cl::opt<SplitEditor::ComplementSpillMode> 59 SplitSpillMode("split-spill-mode", cl::Hidden, 60 cl::desc("Spill mode for splitting live ranges"), 61 cl::values(clEnumValN(SplitEditor::SM_Partition, "default", "Default"), 62 clEnumValN(SplitEditor::SM_Size, "size", "Optimize for size"), 63 clEnumValN(SplitEditor::SM_Speed, "speed", "Optimize for speed"), 64 clEnumValEnd), 65 cl::init(SplitEditor::SM_Partition)); 66 67 static cl::opt<unsigned> 68 LastChanceRecoloringMaxDepth("lcr-max-depth", cl::Hidden, 69 cl::desc("Last chance recoloring max depth"), 70 cl::init(5)); 71 72 static cl::opt<unsigned> LastChanceRecoloringMaxInterference( 73 "lcr-max-interf", cl::Hidden, 74 cl::desc("Last chance recoloring maximum number of considered" 75 " interference at a time"), 76 cl::init(8)); 77 78 static cl::opt<bool> 79 ExhaustiveSearch("exhaustive-register-search", cl::NotHidden, 80 cl::desc("Exhaustive Search for registers bypassing the depth " 81 "and interference cutoffs of last chance recoloring")); 82 83 static cl::opt<bool> EnableLocalReassignment( 84 "enable-local-reassign", cl::Hidden, 85 cl::desc("Local reassignment can yield better allocation decisions, but " 86 "may be compile time intensive"), 87 cl::init(false)); 88 89 static cl::opt<bool> EnableDeferredSpilling( 90 "enable-deferred-spilling", cl::Hidden, 91 cl::desc("Instead of spilling a variable right away, defer the actual " 92 "code insertion to the end of the allocation. That way the " 93 "allocator might still find a suitable coloring for this " 94 "variable because of other evicted variables."), 95 cl::init(false)); 96 97 // FIXME: Find a good default for this flag and remove the flag. 98 static cl::opt<unsigned> 99 CSRFirstTimeCost("regalloc-csr-first-time-cost", 100 cl::desc("Cost for first time use of callee-saved register."), 101 cl::init(0), cl::Hidden); 102 103 static RegisterRegAlloc greedyRegAlloc("greedy", "greedy register allocator", 104 createGreedyRegisterAllocator); 105 106 namespace { 107 class RAGreedy : public MachineFunctionPass, 108 public RegAllocBase, 109 private LiveRangeEdit::Delegate { 110 // Convenient shortcuts. 111 typedef std::priority_queue<std::pair<unsigned, unsigned> > PQueue; 112 typedef SmallPtrSet<LiveInterval *, 4> SmallLISet; 113 typedef SmallSet<unsigned, 16> SmallVirtRegSet; 114 115 // context 116 MachineFunction *MF; 117 118 // Shortcuts to some useful interface. 119 const TargetInstrInfo *TII; 120 const TargetRegisterInfo *TRI; 121 RegisterClassInfo RCI; 122 123 // analyses 124 SlotIndexes *Indexes; 125 MachineBlockFrequencyInfo *MBFI; 126 MachineDominatorTree *DomTree; 127 MachineLoopInfo *Loops; 128 EdgeBundles *Bundles; 129 SpillPlacement *SpillPlacer; 130 LiveDebugVariables *DebugVars; 131 132 // state 133 std::unique_ptr<Spiller> SpillerInstance; 134 PQueue Queue; 135 unsigned NextCascade; 136 137 // Live ranges pass through a number of stages as we try to allocate them. 138 // Some of the stages may also create new live ranges: 139 // 140 // - Region splitting. 141 // - Per-block splitting. 142 // - Local splitting. 143 // - Spilling. 144 // 145 // Ranges produced by one of the stages skip the previous stages when they are 146 // dequeued. This improves performance because we can skip interference checks 147 // that are unlikely to give any results. It also guarantees that the live 148 // range splitting algorithm terminates, something that is otherwise hard to 149 // ensure. 150 enum LiveRangeStage { 151 /// Newly created live range that has never been queued. 152 RS_New, 153 154 /// Only attempt assignment and eviction. Then requeue as RS_Split. 155 RS_Assign, 156 157 /// Attempt live range splitting if assignment is impossible. 158 RS_Split, 159 160 /// Attempt more aggressive live range splitting that is guaranteed to make 161 /// progress. This is used for split products that may not be making 162 /// progress. 163 RS_Split2, 164 165 /// Live range will be spilled. No more splitting will be attempted. 166 RS_Spill, 167 168 169 /// Live range is in memory. Because of other evictions, it might get moved 170 /// in a register in the end. 171 RS_Memory, 172 173 /// There is nothing more we can do to this live range. Abort compilation 174 /// if it can't be assigned. 175 RS_Done 176 }; 177 178 // Enum CutOffStage to keep a track whether the register allocation failed 179 // because of the cutoffs encountered in last chance recoloring. 180 // Note: This is used as bitmask. New value should be next power of 2. 181 enum CutOffStage { 182 // No cutoffs encountered 183 CO_None = 0, 184 185 // lcr-max-depth cutoff encountered 186 CO_Depth = 1, 187 188 // lcr-max-interf cutoff encountered 189 CO_Interf = 2 190 }; 191 192 uint8_t CutOffInfo; 193 194 #ifndef NDEBUG 195 static const char *const StageName[]; 196 #endif 197 198 // RegInfo - Keep additional information about each live range. 199 struct RegInfo { 200 LiveRangeStage Stage; 201 202 // Cascade - Eviction loop prevention. See canEvictInterference(). 203 unsigned Cascade; 204 205 RegInfo() : Stage(RS_New), Cascade(0) {} 206 }; 207 208 IndexedMap<RegInfo, VirtReg2IndexFunctor> ExtraRegInfo; 209 210 LiveRangeStage getStage(const LiveInterval &VirtReg) const { 211 return ExtraRegInfo[VirtReg.reg].Stage; 212 } 213 214 void setStage(const LiveInterval &VirtReg, LiveRangeStage Stage) { 215 ExtraRegInfo.resize(MRI->getNumVirtRegs()); 216 ExtraRegInfo[VirtReg.reg].Stage = Stage; 217 } 218 219 template<typename Iterator> 220 void setStage(Iterator Begin, Iterator End, LiveRangeStage NewStage) { 221 ExtraRegInfo.resize(MRI->getNumVirtRegs()); 222 for (;Begin != End; ++Begin) { 223 unsigned Reg = *Begin; 224 if (ExtraRegInfo[Reg].Stage == RS_New) 225 ExtraRegInfo[Reg].Stage = NewStage; 226 } 227 } 228 229 /// Cost of evicting interference. 230 struct EvictionCost { 231 unsigned BrokenHints; ///< Total number of broken hints. 232 float MaxWeight; ///< Maximum spill weight evicted. 233 234 EvictionCost(): BrokenHints(0), MaxWeight(0) {} 235 236 bool isMax() const { return BrokenHints == ~0u; } 237 238 void setMax() { BrokenHints = ~0u; } 239 240 void setBrokenHints(unsigned NHints) { BrokenHints = NHints; } 241 242 bool operator<(const EvictionCost &O) const { 243 return std::tie(BrokenHints, MaxWeight) < 244 std::tie(O.BrokenHints, O.MaxWeight); 245 } 246 }; 247 248 // splitting state. 249 std::unique_ptr<SplitAnalysis> SA; 250 std::unique_ptr<SplitEditor> SE; 251 252 /// Cached per-block interference maps 253 InterferenceCache IntfCache; 254 255 /// All basic blocks where the current register has uses. 256 SmallVector<SpillPlacement::BlockConstraint, 8> SplitConstraints; 257 258 /// Global live range splitting candidate info. 259 struct GlobalSplitCandidate { 260 // Register intended for assignment, or 0. 261 unsigned PhysReg; 262 263 // SplitKit interval index for this candidate. 264 unsigned IntvIdx; 265 266 // Interference for PhysReg. 267 InterferenceCache::Cursor Intf; 268 269 // Bundles where this candidate should be live. 270 BitVector LiveBundles; 271 SmallVector<unsigned, 8> ActiveBlocks; 272 273 void reset(InterferenceCache &Cache, unsigned Reg) { 274 PhysReg = Reg; 275 IntvIdx = 0; 276 Intf.setPhysReg(Cache, Reg); 277 LiveBundles.clear(); 278 ActiveBlocks.clear(); 279 } 280 281 // Set B[i] = C for every live bundle where B[i] was NoCand. 282 unsigned getBundles(SmallVectorImpl<unsigned> &B, unsigned C) { 283 unsigned Count = 0; 284 for (int i = LiveBundles.find_first(); i >= 0; 285 i = LiveBundles.find_next(i)) 286 if (B[i] == NoCand) { 287 B[i] = C; 288 Count++; 289 } 290 return Count; 291 } 292 }; 293 294 /// Candidate info for each PhysReg in AllocationOrder. 295 /// This vector never shrinks, but grows to the size of the largest register 296 /// class. 297 SmallVector<GlobalSplitCandidate, 32> GlobalCand; 298 299 enum : unsigned { NoCand = ~0u }; 300 301 /// Candidate map. Each edge bundle is assigned to a GlobalCand entry, or to 302 /// NoCand which indicates the stack interval. 303 SmallVector<unsigned, 32> BundleCand; 304 305 /// Callee-save register cost, calculated once per machine function. 306 BlockFrequency CSRCost; 307 308 /// Run or not the local reassignment heuristic. This information is 309 /// obtained from the TargetSubtargetInfo. 310 bool EnableLocalReassign; 311 312 /// Set of broken hints that may be reconciled later because of eviction. 313 SmallSetVector<LiveInterval *, 8> SetOfBrokenHints; 314 315 public: 316 RAGreedy(); 317 318 /// Return the pass name. 319 const char* getPassName() const override { 320 return "Greedy Register Allocator"; 321 } 322 323 /// RAGreedy analysis usage. 324 void getAnalysisUsage(AnalysisUsage &AU) const override; 325 void releaseMemory() override; 326 Spiller &spiller() override { return *SpillerInstance; } 327 void enqueue(LiveInterval *LI) override; 328 LiveInterval *dequeue() override; 329 unsigned selectOrSplit(LiveInterval&, SmallVectorImpl<unsigned>&) override; 330 void aboutToRemoveInterval(LiveInterval &) override; 331 332 /// Perform register allocation. 333 bool runOnMachineFunction(MachineFunction &mf) override; 334 335 static char ID; 336 337 private: 338 unsigned selectOrSplitImpl(LiveInterval &, SmallVectorImpl<unsigned> &, 339 SmallVirtRegSet &, unsigned = 0); 340 341 bool LRE_CanEraseVirtReg(unsigned) override; 342 void LRE_WillShrinkVirtReg(unsigned) override; 343 void LRE_DidCloneVirtReg(unsigned, unsigned) override; 344 void enqueue(PQueue &CurQueue, LiveInterval *LI); 345 LiveInterval *dequeue(PQueue &CurQueue); 346 347 BlockFrequency calcSpillCost(); 348 bool addSplitConstraints(InterferenceCache::Cursor, BlockFrequency&); 349 void addThroughConstraints(InterferenceCache::Cursor, ArrayRef<unsigned>); 350 void growRegion(GlobalSplitCandidate &Cand); 351 BlockFrequency calcGlobalSplitCost(GlobalSplitCandidate&); 352 bool calcCompactRegion(GlobalSplitCandidate&); 353 void splitAroundRegion(LiveRangeEdit&, ArrayRef<unsigned>); 354 void calcGapWeights(unsigned, SmallVectorImpl<float>&); 355 unsigned canReassign(LiveInterval &VirtReg, unsigned PhysReg); 356 bool shouldEvict(LiveInterval &A, bool, LiveInterval &B, bool); 357 bool canEvictInterference(LiveInterval&, unsigned, bool, EvictionCost&); 358 void evictInterference(LiveInterval&, unsigned, 359 SmallVectorImpl<unsigned>&); 360 bool mayRecolorAllInterferences(unsigned PhysReg, LiveInterval &VirtReg, 361 SmallLISet &RecoloringCandidates, 362 const SmallVirtRegSet &FixedRegisters); 363 364 unsigned tryAssign(LiveInterval&, AllocationOrder&, 365 SmallVectorImpl<unsigned>&); 366 unsigned tryEvict(LiveInterval&, AllocationOrder&, 367 SmallVectorImpl<unsigned>&, unsigned = ~0u); 368 unsigned tryRegionSplit(LiveInterval&, AllocationOrder&, 369 SmallVectorImpl<unsigned>&); 370 /// Calculate cost of region splitting. 371 unsigned calculateRegionSplitCost(LiveInterval &VirtReg, 372 AllocationOrder &Order, 373 BlockFrequency &BestCost, 374 unsigned &NumCands, bool IgnoreCSR); 375 /// Perform region splitting. 376 unsigned doRegionSplit(LiveInterval &VirtReg, unsigned BestCand, 377 bool HasCompact, 378 SmallVectorImpl<unsigned> &NewVRegs); 379 /// Check other options before using a callee-saved register for the first 380 /// time. 381 unsigned tryAssignCSRFirstTime(LiveInterval &VirtReg, AllocationOrder &Order, 382 unsigned PhysReg, unsigned &CostPerUseLimit, 383 SmallVectorImpl<unsigned> &NewVRegs); 384 void initializeCSRCost(); 385 unsigned tryBlockSplit(LiveInterval&, AllocationOrder&, 386 SmallVectorImpl<unsigned>&); 387 unsigned tryInstructionSplit(LiveInterval&, AllocationOrder&, 388 SmallVectorImpl<unsigned>&); 389 unsigned tryLocalSplit(LiveInterval&, AllocationOrder&, 390 SmallVectorImpl<unsigned>&); 391 unsigned trySplit(LiveInterval&, AllocationOrder&, 392 SmallVectorImpl<unsigned>&); 393 unsigned tryLastChanceRecoloring(LiveInterval &, AllocationOrder &, 394 SmallVectorImpl<unsigned> &, 395 SmallVirtRegSet &, unsigned); 396 bool tryRecoloringCandidates(PQueue &, SmallVectorImpl<unsigned> &, 397 SmallVirtRegSet &, unsigned); 398 void tryHintRecoloring(LiveInterval &); 399 void tryHintsRecoloring(); 400 401 /// Model the information carried by one end of a copy. 402 struct HintInfo { 403 /// The frequency of the copy. 404 BlockFrequency Freq; 405 /// The virtual register or physical register. 406 unsigned Reg; 407 /// Its currently assigned register. 408 /// In case of a physical register Reg == PhysReg. 409 unsigned PhysReg; 410 HintInfo(BlockFrequency Freq, unsigned Reg, unsigned PhysReg) 411 : Freq(Freq), Reg(Reg), PhysReg(PhysReg) {} 412 }; 413 typedef SmallVector<HintInfo, 4> HintsInfo; 414 BlockFrequency getBrokenHintFreq(const HintsInfo &, unsigned); 415 void collectHintInfo(unsigned, HintsInfo &); 416 417 bool isUnusedCalleeSavedReg(unsigned PhysReg) const; 418 }; 419 } // end anonymous namespace 420 421 char RAGreedy::ID = 0; 422 423 #ifndef NDEBUG 424 const char *const RAGreedy::StageName[] = { 425 "RS_New", 426 "RS_Assign", 427 "RS_Split", 428 "RS_Split2", 429 "RS_Spill", 430 "RS_Memory", 431 "RS_Done" 432 }; 433 #endif 434 435 // Hysteresis to use when comparing floats. 436 // This helps stabilize decisions based on float comparisons. 437 const float Hysteresis = (2007 / 2048.0f); // 0.97998046875 438 439 440 FunctionPass* llvm::createGreedyRegisterAllocator() { 441 return new RAGreedy(); 442 } 443 444 RAGreedy::RAGreedy(): MachineFunctionPass(ID) { 445 initializeLiveDebugVariablesPass(*PassRegistry::getPassRegistry()); 446 initializeSlotIndexesPass(*PassRegistry::getPassRegistry()); 447 initializeLiveIntervalsPass(*PassRegistry::getPassRegistry()); 448 initializeSlotIndexesPass(*PassRegistry::getPassRegistry()); 449 initializeRegisterCoalescerPass(*PassRegistry::getPassRegistry()); 450 initializeMachineSchedulerPass(*PassRegistry::getPassRegistry()); 451 initializeLiveStacksPass(*PassRegistry::getPassRegistry()); 452 initializeMachineDominatorTreePass(*PassRegistry::getPassRegistry()); 453 initializeMachineLoopInfoPass(*PassRegistry::getPassRegistry()); 454 initializeVirtRegMapPass(*PassRegistry::getPassRegistry()); 455 initializeLiveRegMatrixPass(*PassRegistry::getPassRegistry()); 456 initializeEdgeBundlesPass(*PassRegistry::getPassRegistry()); 457 initializeSpillPlacementPass(*PassRegistry::getPassRegistry()); 458 } 459 460 void RAGreedy::getAnalysisUsage(AnalysisUsage &AU) const { 461 AU.setPreservesCFG(); 462 AU.addRequired<MachineBlockFrequencyInfo>(); 463 AU.addPreserved<MachineBlockFrequencyInfo>(); 464 AU.addRequired<AAResultsWrapperPass>(); 465 AU.addPreserved<AAResultsWrapperPass>(); 466 AU.addRequired<LiveIntervals>(); 467 AU.addPreserved<LiveIntervals>(); 468 AU.addRequired<SlotIndexes>(); 469 AU.addPreserved<SlotIndexes>(); 470 AU.addRequired<LiveDebugVariables>(); 471 AU.addPreserved<LiveDebugVariables>(); 472 AU.addRequired<LiveStacks>(); 473 AU.addPreserved<LiveStacks>(); 474 AU.addRequired<MachineDominatorTree>(); 475 AU.addPreserved<MachineDominatorTree>(); 476 AU.addRequired<MachineLoopInfo>(); 477 AU.addPreserved<MachineLoopInfo>(); 478 AU.addRequired<VirtRegMap>(); 479 AU.addPreserved<VirtRegMap>(); 480 AU.addRequired<LiveRegMatrix>(); 481 AU.addPreserved<LiveRegMatrix>(); 482 AU.addRequired<EdgeBundles>(); 483 AU.addRequired<SpillPlacement>(); 484 MachineFunctionPass::getAnalysisUsage(AU); 485 } 486 487 488 //===----------------------------------------------------------------------===// 489 // LiveRangeEdit delegate methods 490 //===----------------------------------------------------------------------===// 491 492 bool RAGreedy::LRE_CanEraseVirtReg(unsigned VirtReg) { 493 if (VRM->hasPhys(VirtReg)) { 494 LiveInterval &LI = LIS->getInterval(VirtReg); 495 Matrix->unassign(LI); 496 aboutToRemoveInterval(LI); 497 return true; 498 } 499 // Unassigned virtreg is probably in the priority queue. 500 // RegAllocBase will erase it after dequeueing. 501 return false; 502 } 503 504 void RAGreedy::LRE_WillShrinkVirtReg(unsigned VirtReg) { 505 if (!VRM->hasPhys(VirtReg)) 506 return; 507 508 // Register is assigned, put it back on the queue for reassignment. 509 LiveInterval &LI = LIS->getInterval(VirtReg); 510 Matrix->unassign(LI); 511 enqueue(&LI); 512 } 513 514 void RAGreedy::LRE_DidCloneVirtReg(unsigned New, unsigned Old) { 515 // Cloning a register we haven't even heard about yet? Just ignore it. 516 if (!ExtraRegInfo.inBounds(Old)) 517 return; 518 519 // LRE may clone a virtual register because dead code elimination causes it to 520 // be split into connected components. The new components are much smaller 521 // than the original, so they should get a new chance at being assigned. 522 // same stage as the parent. 523 ExtraRegInfo[Old].Stage = RS_Assign; 524 ExtraRegInfo.grow(New); 525 ExtraRegInfo[New] = ExtraRegInfo[Old]; 526 } 527 528 void RAGreedy::releaseMemory() { 529 SpillerInstance.reset(); 530 ExtraRegInfo.clear(); 531 GlobalCand.clear(); 532 } 533 534 void RAGreedy::enqueue(LiveInterval *LI) { enqueue(Queue, LI); } 535 536 void RAGreedy::enqueue(PQueue &CurQueue, LiveInterval *LI) { 537 // Prioritize live ranges by size, assigning larger ranges first. 538 // The queue holds (size, reg) pairs. 539 const unsigned Size = LI->getSize(); 540 const unsigned Reg = LI->reg; 541 assert(TargetRegisterInfo::isVirtualRegister(Reg) && 542 "Can only enqueue virtual registers"); 543 unsigned Prio; 544 545 ExtraRegInfo.grow(Reg); 546 if (ExtraRegInfo[Reg].Stage == RS_New) 547 ExtraRegInfo[Reg].Stage = RS_Assign; 548 549 if (ExtraRegInfo[Reg].Stage == RS_Split) { 550 // Unsplit ranges that couldn't be allocated immediately are deferred until 551 // everything else has been allocated. 552 Prio = Size; 553 } else if (ExtraRegInfo[Reg].Stage == RS_Memory) { 554 // Memory operand should be considered last. 555 // Change the priority such that Memory operand are assigned in 556 // the reverse order that they came in. 557 // TODO: Make this a member variable and probably do something about hints. 558 static unsigned MemOp = 0; 559 Prio = MemOp++; 560 } else { 561 // Giant live ranges fall back to the global assignment heuristic, which 562 // prevents excessive spilling in pathological cases. 563 bool ReverseLocal = TRI->reverseLocalAssignment(); 564 const TargetRegisterClass &RC = *MRI->getRegClass(Reg); 565 bool ForceGlobal = !ReverseLocal && 566 (Size / SlotIndex::InstrDist) > (2 * RC.getNumRegs()); 567 568 if (ExtraRegInfo[Reg].Stage == RS_Assign && !ForceGlobal && !LI->empty() && 569 LIS->intervalIsInOneMBB(*LI)) { 570 // Allocate original local ranges in linear instruction order. Since they 571 // are singly defined, this produces optimal coloring in the absence of 572 // global interference and other constraints. 573 if (!ReverseLocal) 574 Prio = LI->beginIndex().getInstrDistance(Indexes->getLastIndex()); 575 else { 576 // Allocating bottom up may allow many short LRGs to be assigned first 577 // to one of the cheap registers. This could be much faster for very 578 // large blocks on targets with many physical registers. 579 Prio = Indexes->getZeroIndex().getInstrDistance(LI->endIndex()); 580 } 581 Prio |= RC.AllocationPriority << 24; 582 } else { 583 // Allocate global and split ranges in long->short order. Long ranges that 584 // don't fit should be spilled (or split) ASAP so they don't create 585 // interference. Mark a bit to prioritize global above local ranges. 586 Prio = (1u << 29) + Size; 587 } 588 // Mark a higher bit to prioritize global and local above RS_Split. 589 Prio |= (1u << 31); 590 591 // Boost ranges that have a physical register hint. 592 if (VRM->hasKnownPreference(Reg)) 593 Prio |= (1u << 30); 594 } 595 // The virtual register number is a tie breaker for same-sized ranges. 596 // Give lower vreg numbers higher priority to assign them first. 597 CurQueue.push(std::make_pair(Prio, ~Reg)); 598 } 599 600 LiveInterval *RAGreedy::dequeue() { return dequeue(Queue); } 601 602 LiveInterval *RAGreedy::dequeue(PQueue &CurQueue) { 603 if (CurQueue.empty()) 604 return nullptr; 605 LiveInterval *LI = &LIS->getInterval(~CurQueue.top().second); 606 CurQueue.pop(); 607 return LI; 608 } 609 610 611 //===----------------------------------------------------------------------===// 612 // Direct Assignment 613 //===----------------------------------------------------------------------===// 614 615 /// tryAssign - Try to assign VirtReg to an available register. 616 unsigned RAGreedy::tryAssign(LiveInterval &VirtReg, 617 AllocationOrder &Order, 618 SmallVectorImpl<unsigned> &NewVRegs) { 619 Order.rewind(); 620 unsigned PhysReg; 621 while ((PhysReg = Order.next())) 622 if (!Matrix->checkInterference(VirtReg, PhysReg)) 623 break; 624 if (!PhysReg || Order.isHint()) 625 return PhysReg; 626 627 // PhysReg is available, but there may be a better choice. 628 629 // If we missed a simple hint, try to cheaply evict interference from the 630 // preferred register. 631 if (unsigned Hint = MRI->getSimpleHint(VirtReg.reg)) 632 if (Order.isHint(Hint)) { 633 DEBUG(dbgs() << "missed hint " << PrintReg(Hint, TRI) << '\n'); 634 EvictionCost MaxCost; 635 MaxCost.setBrokenHints(1); 636 if (canEvictInterference(VirtReg, Hint, true, MaxCost)) { 637 evictInterference(VirtReg, Hint, NewVRegs); 638 return Hint; 639 } 640 } 641 642 // Try to evict interference from a cheaper alternative. 643 unsigned Cost = TRI->getCostPerUse(PhysReg); 644 645 // Most registers have 0 additional cost. 646 if (!Cost) 647 return PhysReg; 648 649 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << " is available at cost " << Cost 650 << '\n'); 651 unsigned CheapReg = tryEvict(VirtReg, Order, NewVRegs, Cost); 652 return CheapReg ? CheapReg : PhysReg; 653 } 654 655 656 //===----------------------------------------------------------------------===// 657 // Interference eviction 658 //===----------------------------------------------------------------------===// 659 660 unsigned RAGreedy::canReassign(LiveInterval &VirtReg, unsigned PrevReg) { 661 AllocationOrder Order(VirtReg.reg, *VRM, RegClassInfo, Matrix); 662 unsigned PhysReg; 663 while ((PhysReg = Order.next())) { 664 if (PhysReg == PrevReg) 665 continue; 666 667 MCRegUnitIterator Units(PhysReg, TRI); 668 for (; Units.isValid(); ++Units) { 669 // Instantiate a "subquery", not to be confused with the Queries array. 670 LiveIntervalUnion::Query subQ(&VirtReg, &Matrix->getLiveUnions()[*Units]); 671 if (subQ.checkInterference()) 672 break; 673 } 674 // If no units have interference, break out with the current PhysReg. 675 if (!Units.isValid()) 676 break; 677 } 678 if (PhysReg) 679 DEBUG(dbgs() << "can reassign: " << VirtReg << " from " 680 << PrintReg(PrevReg, TRI) << " to " << PrintReg(PhysReg, TRI) 681 << '\n'); 682 return PhysReg; 683 } 684 685 /// shouldEvict - determine if A should evict the assigned live range B. The 686 /// eviction policy defined by this function together with the allocation order 687 /// defined by enqueue() decides which registers ultimately end up being split 688 /// and spilled. 689 /// 690 /// Cascade numbers are used to prevent infinite loops if this function is a 691 /// cyclic relation. 692 /// 693 /// @param A The live range to be assigned. 694 /// @param IsHint True when A is about to be assigned to its preferred 695 /// register. 696 /// @param B The live range to be evicted. 697 /// @param BreaksHint True when B is already assigned to its preferred register. 698 bool RAGreedy::shouldEvict(LiveInterval &A, bool IsHint, 699 LiveInterval &B, bool BreaksHint) { 700 bool CanSplit = getStage(B) < RS_Spill; 701 702 // Be fairly aggressive about following hints as long as the evictee can be 703 // split. 704 if (CanSplit && IsHint && !BreaksHint) 705 return true; 706 707 if (A.weight > B.weight) { 708 DEBUG(dbgs() << "should evict: " << B << " w= " << B.weight << '\n'); 709 return true; 710 } 711 return false; 712 } 713 714 /// canEvictInterference - Return true if all interferences between VirtReg and 715 /// PhysReg can be evicted. 716 /// 717 /// @param VirtReg Live range that is about to be assigned. 718 /// @param PhysReg Desired register for assignment. 719 /// @param IsHint True when PhysReg is VirtReg's preferred register. 720 /// @param MaxCost Only look for cheaper candidates and update with new cost 721 /// when returning true. 722 /// @returns True when interference can be evicted cheaper than MaxCost. 723 bool RAGreedy::canEvictInterference(LiveInterval &VirtReg, unsigned PhysReg, 724 bool IsHint, EvictionCost &MaxCost) { 725 // It is only possible to evict virtual register interference. 726 if (Matrix->checkInterference(VirtReg, PhysReg) > LiveRegMatrix::IK_VirtReg) 727 return false; 728 729 bool IsLocal = LIS->intervalIsInOneMBB(VirtReg); 730 731 // Find VirtReg's cascade number. This will be unassigned if VirtReg was never 732 // involved in an eviction before. If a cascade number was assigned, deny 733 // evicting anything with the same or a newer cascade number. This prevents 734 // infinite eviction loops. 735 // 736 // This works out so a register without a cascade number is allowed to evict 737 // anything, and it can be evicted by anything. 738 unsigned Cascade = ExtraRegInfo[VirtReg.reg].Cascade; 739 if (!Cascade) 740 Cascade = NextCascade; 741 742 EvictionCost Cost; 743 for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) { 744 LiveIntervalUnion::Query &Q = Matrix->query(VirtReg, *Units); 745 // If there is 10 or more interferences, chances are one is heavier. 746 if (Q.collectInterferingVRegs(10) >= 10) 747 return false; 748 749 // Check if any interfering live range is heavier than MaxWeight. 750 for (unsigned i = Q.interferingVRegs().size(); i; --i) { 751 LiveInterval *Intf = Q.interferingVRegs()[i - 1]; 752 assert(TargetRegisterInfo::isVirtualRegister(Intf->reg) && 753 "Only expecting virtual register interference from query"); 754 // Never evict spill products. They cannot split or spill. 755 if (getStage(*Intf) == RS_Done) 756 return false; 757 // Once a live range becomes small enough, it is urgent that we find a 758 // register for it. This is indicated by an infinite spill weight. These 759 // urgent live ranges get to evict almost anything. 760 // 761 // Also allow urgent evictions of unspillable ranges from a strictly 762 // larger allocation order. 763 bool Urgent = !VirtReg.isSpillable() && 764 (Intf->isSpillable() || 765 RegClassInfo.getNumAllocatableRegs(MRI->getRegClass(VirtReg.reg)) < 766 RegClassInfo.getNumAllocatableRegs(MRI->getRegClass(Intf->reg))); 767 // Only evict older cascades or live ranges without a cascade. 768 unsigned IntfCascade = ExtraRegInfo[Intf->reg].Cascade; 769 if (Cascade <= IntfCascade) { 770 if (!Urgent) 771 return false; 772 // We permit breaking cascades for urgent evictions. It should be the 773 // last resort, though, so make it really expensive. 774 Cost.BrokenHints += 10; 775 } 776 // Would this break a satisfied hint? 777 bool BreaksHint = VRM->hasPreferredPhys(Intf->reg); 778 // Update eviction cost. 779 Cost.BrokenHints += BreaksHint; 780 Cost.MaxWeight = std::max(Cost.MaxWeight, Intf->weight); 781 // Abort if this would be too expensive. 782 if (!(Cost < MaxCost)) 783 return false; 784 if (Urgent) 785 continue; 786 // Apply the eviction policy for non-urgent evictions. 787 if (!shouldEvict(VirtReg, IsHint, *Intf, BreaksHint)) 788 return false; 789 // If !MaxCost.isMax(), then we're just looking for a cheap register. 790 // Evicting another local live range in this case could lead to suboptimal 791 // coloring. 792 if (!MaxCost.isMax() && IsLocal && LIS->intervalIsInOneMBB(*Intf) && 793 (!EnableLocalReassign || !canReassign(*Intf, PhysReg))) { 794 return false; 795 } 796 } 797 } 798 MaxCost = Cost; 799 return true; 800 } 801 802 /// evictInterference - Evict any interferring registers that prevent VirtReg 803 /// from being assigned to Physreg. This assumes that canEvictInterference 804 /// returned true. 805 void RAGreedy::evictInterference(LiveInterval &VirtReg, unsigned PhysReg, 806 SmallVectorImpl<unsigned> &NewVRegs) { 807 // Make sure that VirtReg has a cascade number, and assign that cascade 808 // number to every evicted register. These live ranges than then only be 809 // evicted by a newer cascade, preventing infinite loops. 810 unsigned Cascade = ExtraRegInfo[VirtReg.reg].Cascade; 811 if (!Cascade) 812 Cascade = ExtraRegInfo[VirtReg.reg].Cascade = NextCascade++; 813 814 DEBUG(dbgs() << "evicting " << PrintReg(PhysReg, TRI) 815 << " interference: Cascade " << Cascade << '\n'); 816 817 // Collect all interfering virtregs first. 818 SmallVector<LiveInterval*, 8> Intfs; 819 for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) { 820 LiveIntervalUnion::Query &Q = Matrix->query(VirtReg, *Units); 821 assert(Q.seenAllInterferences() && "Didn't check all interfererences."); 822 ArrayRef<LiveInterval*> IVR = Q.interferingVRegs(); 823 Intfs.append(IVR.begin(), IVR.end()); 824 } 825 826 // Evict them second. This will invalidate the queries. 827 for (unsigned i = 0, e = Intfs.size(); i != e; ++i) { 828 LiveInterval *Intf = Intfs[i]; 829 // The same VirtReg may be present in multiple RegUnits. Skip duplicates. 830 if (!VRM->hasPhys(Intf->reg)) 831 continue; 832 Matrix->unassign(*Intf); 833 assert((ExtraRegInfo[Intf->reg].Cascade < Cascade || 834 VirtReg.isSpillable() < Intf->isSpillable()) && 835 "Cannot decrease cascade number, illegal eviction"); 836 ExtraRegInfo[Intf->reg].Cascade = Cascade; 837 ++NumEvicted; 838 NewVRegs.push_back(Intf->reg); 839 } 840 } 841 842 /// Returns true if the given \p PhysReg is a callee saved register and has not 843 /// been used for allocation yet. 844 bool RAGreedy::isUnusedCalleeSavedReg(unsigned PhysReg) const { 845 unsigned CSR = RegClassInfo.getLastCalleeSavedAlias(PhysReg); 846 if (CSR == 0) 847 return false; 848 849 return !Matrix->isPhysRegUsed(PhysReg); 850 } 851 852 /// tryEvict - Try to evict all interferences for a physreg. 853 /// @param VirtReg Currently unassigned virtual register. 854 /// @param Order Physregs to try. 855 /// @return Physreg to assign VirtReg, or 0. 856 unsigned RAGreedy::tryEvict(LiveInterval &VirtReg, 857 AllocationOrder &Order, 858 SmallVectorImpl<unsigned> &NewVRegs, 859 unsigned CostPerUseLimit) { 860 NamedRegionTimer T("Evict", TimerGroupName, TimePassesIsEnabled); 861 862 // Keep track of the cheapest interference seen so far. 863 EvictionCost BestCost; 864 BestCost.setMax(); 865 unsigned BestPhys = 0; 866 unsigned OrderLimit = Order.getOrder().size(); 867 868 // When we are just looking for a reduced cost per use, don't break any 869 // hints, and only evict smaller spill weights. 870 if (CostPerUseLimit < ~0u) { 871 BestCost.BrokenHints = 0; 872 BestCost.MaxWeight = VirtReg.weight; 873 874 // Check of any registers in RC are below CostPerUseLimit. 875 const TargetRegisterClass *RC = MRI->getRegClass(VirtReg.reg); 876 unsigned MinCost = RegClassInfo.getMinCost(RC); 877 if (MinCost >= CostPerUseLimit) { 878 DEBUG(dbgs() << TRI->getRegClassName(RC) << " minimum cost = " << MinCost 879 << ", no cheaper registers to be found.\n"); 880 return 0; 881 } 882 883 // It is normal for register classes to have a long tail of registers with 884 // the same cost. We don't need to look at them if they're too expensive. 885 if (TRI->getCostPerUse(Order.getOrder().back()) >= CostPerUseLimit) { 886 OrderLimit = RegClassInfo.getLastCostChange(RC); 887 DEBUG(dbgs() << "Only trying the first " << OrderLimit << " regs.\n"); 888 } 889 } 890 891 Order.rewind(); 892 while (unsigned PhysReg = Order.next(OrderLimit)) { 893 if (TRI->getCostPerUse(PhysReg) >= CostPerUseLimit) 894 continue; 895 // The first use of a callee-saved register in a function has cost 1. 896 // Don't start using a CSR when the CostPerUseLimit is low. 897 if (CostPerUseLimit == 1 && isUnusedCalleeSavedReg(PhysReg)) { 898 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << " would clobber CSR " 899 << PrintReg(RegClassInfo.getLastCalleeSavedAlias(PhysReg), TRI) 900 << '\n'); 901 continue; 902 } 903 904 if (!canEvictInterference(VirtReg, PhysReg, false, BestCost)) 905 continue; 906 907 // Best so far. 908 BestPhys = PhysReg; 909 910 // Stop if the hint can be used. 911 if (Order.isHint()) 912 break; 913 } 914 915 if (!BestPhys) 916 return 0; 917 918 evictInterference(VirtReg, BestPhys, NewVRegs); 919 return BestPhys; 920 } 921 922 923 //===----------------------------------------------------------------------===// 924 // Region Splitting 925 //===----------------------------------------------------------------------===// 926 927 /// addSplitConstraints - Fill out the SplitConstraints vector based on the 928 /// interference pattern in Physreg and its aliases. Add the constraints to 929 /// SpillPlacement and return the static cost of this split in Cost, assuming 930 /// that all preferences in SplitConstraints are met. 931 /// Return false if there are no bundles with positive bias. 932 bool RAGreedy::addSplitConstraints(InterferenceCache::Cursor Intf, 933 BlockFrequency &Cost) { 934 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks(); 935 936 // Reset interference dependent info. 937 SplitConstraints.resize(UseBlocks.size()); 938 BlockFrequency StaticCost = 0; 939 for (unsigned i = 0; i != UseBlocks.size(); ++i) { 940 const SplitAnalysis::BlockInfo &BI = UseBlocks[i]; 941 SpillPlacement::BlockConstraint &BC = SplitConstraints[i]; 942 943 BC.Number = BI.MBB->getNumber(); 944 Intf.moveToBlock(BC.Number); 945 BC.Entry = BI.LiveIn ? SpillPlacement::PrefReg : SpillPlacement::DontCare; 946 BC.Exit = BI.LiveOut ? SpillPlacement::PrefReg : SpillPlacement::DontCare; 947 BC.ChangesValue = BI.FirstDef.isValid(); 948 949 if (!Intf.hasInterference()) 950 continue; 951 952 // Number of spill code instructions to insert. 953 unsigned Ins = 0; 954 955 // Interference for the live-in value. 956 if (BI.LiveIn) { 957 if (Intf.first() <= Indexes->getMBBStartIdx(BC.Number)) 958 BC.Entry = SpillPlacement::MustSpill, ++Ins; 959 else if (Intf.first() < BI.FirstInstr) 960 BC.Entry = SpillPlacement::PrefSpill, ++Ins; 961 else if (Intf.first() < BI.LastInstr) 962 ++Ins; 963 } 964 965 // Interference for the live-out value. 966 if (BI.LiveOut) { 967 if (Intf.last() >= SA->getLastSplitPoint(BC.Number)) 968 BC.Exit = SpillPlacement::MustSpill, ++Ins; 969 else if (Intf.last() > BI.LastInstr) 970 BC.Exit = SpillPlacement::PrefSpill, ++Ins; 971 else if (Intf.last() > BI.FirstInstr) 972 ++Ins; 973 } 974 975 // Accumulate the total frequency of inserted spill code. 976 while (Ins--) 977 StaticCost += SpillPlacer->getBlockFrequency(BC.Number); 978 } 979 Cost = StaticCost; 980 981 // Add constraints for use-blocks. Note that these are the only constraints 982 // that may add a positive bias, it is downhill from here. 983 SpillPlacer->addConstraints(SplitConstraints); 984 return SpillPlacer->scanActiveBundles(); 985 } 986 987 988 /// addThroughConstraints - Add constraints and links to SpillPlacer from the 989 /// live-through blocks in Blocks. 990 void RAGreedy::addThroughConstraints(InterferenceCache::Cursor Intf, 991 ArrayRef<unsigned> Blocks) { 992 const unsigned GroupSize = 8; 993 SpillPlacement::BlockConstraint BCS[GroupSize]; 994 unsigned TBS[GroupSize]; 995 unsigned B = 0, T = 0; 996 997 for (unsigned i = 0; i != Blocks.size(); ++i) { 998 unsigned Number = Blocks[i]; 999 Intf.moveToBlock(Number); 1000 1001 if (!Intf.hasInterference()) { 1002 assert(T < GroupSize && "Array overflow"); 1003 TBS[T] = Number; 1004 if (++T == GroupSize) { 1005 SpillPlacer->addLinks(makeArrayRef(TBS, T)); 1006 T = 0; 1007 } 1008 continue; 1009 } 1010 1011 assert(B < GroupSize && "Array overflow"); 1012 BCS[B].Number = Number; 1013 1014 // Interference for the live-in value. 1015 if (Intf.first() <= Indexes->getMBBStartIdx(Number)) 1016 BCS[B].Entry = SpillPlacement::MustSpill; 1017 else 1018 BCS[B].Entry = SpillPlacement::PrefSpill; 1019 1020 // Interference for the live-out value. 1021 if (Intf.last() >= SA->getLastSplitPoint(Number)) 1022 BCS[B].Exit = SpillPlacement::MustSpill; 1023 else 1024 BCS[B].Exit = SpillPlacement::PrefSpill; 1025 1026 if (++B == GroupSize) { 1027 SpillPlacer->addConstraints(makeArrayRef(BCS, B)); 1028 B = 0; 1029 } 1030 } 1031 1032 SpillPlacer->addConstraints(makeArrayRef(BCS, B)); 1033 SpillPlacer->addLinks(makeArrayRef(TBS, T)); 1034 } 1035 1036 void RAGreedy::growRegion(GlobalSplitCandidate &Cand) { 1037 // Keep track of through blocks that have not been added to SpillPlacer. 1038 BitVector Todo = SA->getThroughBlocks(); 1039 SmallVectorImpl<unsigned> &ActiveBlocks = Cand.ActiveBlocks; 1040 unsigned AddedTo = 0; 1041 #ifndef NDEBUG 1042 unsigned Visited = 0; 1043 #endif 1044 1045 for (;;) { 1046 ArrayRef<unsigned> NewBundles = SpillPlacer->getRecentPositive(); 1047 // Find new through blocks in the periphery of PrefRegBundles. 1048 for (int i = 0, e = NewBundles.size(); i != e; ++i) { 1049 unsigned Bundle = NewBundles[i]; 1050 // Look at all blocks connected to Bundle in the full graph. 1051 ArrayRef<unsigned> Blocks = Bundles->getBlocks(Bundle); 1052 for (ArrayRef<unsigned>::iterator I = Blocks.begin(), E = Blocks.end(); 1053 I != E; ++I) { 1054 unsigned Block = *I; 1055 if (!Todo.test(Block)) 1056 continue; 1057 Todo.reset(Block); 1058 // This is a new through block. Add it to SpillPlacer later. 1059 ActiveBlocks.push_back(Block); 1060 #ifndef NDEBUG 1061 ++Visited; 1062 #endif 1063 } 1064 } 1065 // Any new blocks to add? 1066 if (ActiveBlocks.size() == AddedTo) 1067 break; 1068 1069 // Compute through constraints from the interference, or assume that all 1070 // through blocks prefer spilling when forming compact regions. 1071 auto NewBlocks = makeArrayRef(ActiveBlocks).slice(AddedTo); 1072 if (Cand.PhysReg) 1073 addThroughConstraints(Cand.Intf, NewBlocks); 1074 else 1075 // Provide a strong negative bias on through blocks to prevent unwanted 1076 // liveness on loop backedges. 1077 SpillPlacer->addPrefSpill(NewBlocks, /* Strong= */ true); 1078 AddedTo = ActiveBlocks.size(); 1079 1080 // Perhaps iterating can enable more bundles? 1081 SpillPlacer->iterate(); 1082 } 1083 DEBUG(dbgs() << ", v=" << Visited); 1084 } 1085 1086 /// calcCompactRegion - Compute the set of edge bundles that should be live 1087 /// when splitting the current live range into compact regions. Compact 1088 /// regions can be computed without looking at interference. They are the 1089 /// regions formed by removing all the live-through blocks from the live range. 1090 /// 1091 /// Returns false if the current live range is already compact, or if the 1092 /// compact regions would form single block regions anyway. 1093 bool RAGreedy::calcCompactRegion(GlobalSplitCandidate &Cand) { 1094 // Without any through blocks, the live range is already compact. 1095 if (!SA->getNumThroughBlocks()) 1096 return false; 1097 1098 // Compact regions don't correspond to any physreg. 1099 Cand.reset(IntfCache, 0); 1100 1101 DEBUG(dbgs() << "Compact region bundles"); 1102 1103 // Use the spill placer to determine the live bundles. GrowRegion pretends 1104 // that all the through blocks have interference when PhysReg is unset. 1105 SpillPlacer->prepare(Cand.LiveBundles); 1106 1107 // The static split cost will be zero since Cand.Intf reports no interference. 1108 BlockFrequency Cost; 1109 if (!addSplitConstraints(Cand.Intf, Cost)) { 1110 DEBUG(dbgs() << ", none.\n"); 1111 return false; 1112 } 1113 1114 growRegion(Cand); 1115 SpillPlacer->finish(); 1116 1117 if (!Cand.LiveBundles.any()) { 1118 DEBUG(dbgs() << ", none.\n"); 1119 return false; 1120 } 1121 1122 DEBUG({ 1123 for (int i = Cand.LiveBundles.find_first(); i>=0; 1124 i = Cand.LiveBundles.find_next(i)) 1125 dbgs() << " EB#" << i; 1126 dbgs() << ".\n"; 1127 }); 1128 return true; 1129 } 1130 1131 /// calcSpillCost - Compute how expensive it would be to split the live range in 1132 /// SA around all use blocks instead of forming bundle regions. 1133 BlockFrequency RAGreedy::calcSpillCost() { 1134 BlockFrequency Cost = 0; 1135 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks(); 1136 for (unsigned i = 0; i != UseBlocks.size(); ++i) { 1137 const SplitAnalysis::BlockInfo &BI = UseBlocks[i]; 1138 unsigned Number = BI.MBB->getNumber(); 1139 // We normally only need one spill instruction - a load or a store. 1140 Cost += SpillPlacer->getBlockFrequency(Number); 1141 1142 // Unless the value is redefined in the block. 1143 if (BI.LiveIn && BI.LiveOut && BI.FirstDef) 1144 Cost += SpillPlacer->getBlockFrequency(Number); 1145 } 1146 return Cost; 1147 } 1148 1149 /// calcGlobalSplitCost - Return the global split cost of following the split 1150 /// pattern in LiveBundles. This cost should be added to the local cost of the 1151 /// interference pattern in SplitConstraints. 1152 /// 1153 BlockFrequency RAGreedy::calcGlobalSplitCost(GlobalSplitCandidate &Cand) { 1154 BlockFrequency GlobalCost = 0; 1155 const BitVector &LiveBundles = Cand.LiveBundles; 1156 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks(); 1157 for (unsigned i = 0; i != UseBlocks.size(); ++i) { 1158 const SplitAnalysis::BlockInfo &BI = UseBlocks[i]; 1159 SpillPlacement::BlockConstraint &BC = SplitConstraints[i]; 1160 bool RegIn = LiveBundles[Bundles->getBundle(BC.Number, 0)]; 1161 bool RegOut = LiveBundles[Bundles->getBundle(BC.Number, 1)]; 1162 unsigned Ins = 0; 1163 1164 if (BI.LiveIn) 1165 Ins += RegIn != (BC.Entry == SpillPlacement::PrefReg); 1166 if (BI.LiveOut) 1167 Ins += RegOut != (BC.Exit == SpillPlacement::PrefReg); 1168 while (Ins--) 1169 GlobalCost += SpillPlacer->getBlockFrequency(BC.Number); 1170 } 1171 1172 for (unsigned i = 0, e = Cand.ActiveBlocks.size(); i != e; ++i) { 1173 unsigned Number = Cand.ActiveBlocks[i]; 1174 bool RegIn = LiveBundles[Bundles->getBundle(Number, 0)]; 1175 bool RegOut = LiveBundles[Bundles->getBundle(Number, 1)]; 1176 if (!RegIn && !RegOut) 1177 continue; 1178 if (RegIn && RegOut) { 1179 // We need double spill code if this block has interference. 1180 Cand.Intf.moveToBlock(Number); 1181 if (Cand.Intf.hasInterference()) { 1182 GlobalCost += SpillPlacer->getBlockFrequency(Number); 1183 GlobalCost += SpillPlacer->getBlockFrequency(Number); 1184 } 1185 continue; 1186 } 1187 // live-in / stack-out or stack-in live-out. 1188 GlobalCost += SpillPlacer->getBlockFrequency(Number); 1189 } 1190 return GlobalCost; 1191 } 1192 1193 /// splitAroundRegion - Split the current live range around the regions 1194 /// determined by BundleCand and GlobalCand. 1195 /// 1196 /// Before calling this function, GlobalCand and BundleCand must be initialized 1197 /// so each bundle is assigned to a valid candidate, or NoCand for the 1198 /// stack-bound bundles. The shared SA/SE SplitAnalysis and SplitEditor 1199 /// objects must be initialized for the current live range, and intervals 1200 /// created for the used candidates. 1201 /// 1202 /// @param LREdit The LiveRangeEdit object handling the current split. 1203 /// @param UsedCands List of used GlobalCand entries. Every BundleCand value 1204 /// must appear in this list. 1205 void RAGreedy::splitAroundRegion(LiveRangeEdit &LREdit, 1206 ArrayRef<unsigned> UsedCands) { 1207 // These are the intervals created for new global ranges. We may create more 1208 // intervals for local ranges. 1209 const unsigned NumGlobalIntvs = LREdit.size(); 1210 DEBUG(dbgs() << "splitAroundRegion with " << NumGlobalIntvs << " globals.\n"); 1211 assert(NumGlobalIntvs && "No global intervals configured"); 1212 1213 // Isolate even single instructions when dealing with a proper sub-class. 1214 // That guarantees register class inflation for the stack interval because it 1215 // is all copies. 1216 unsigned Reg = SA->getParent().reg; 1217 bool SingleInstrs = RegClassInfo.isProperSubClass(MRI->getRegClass(Reg)); 1218 1219 // First handle all the blocks with uses. 1220 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks(); 1221 for (unsigned i = 0; i != UseBlocks.size(); ++i) { 1222 const SplitAnalysis::BlockInfo &BI = UseBlocks[i]; 1223 unsigned Number = BI.MBB->getNumber(); 1224 unsigned IntvIn = 0, IntvOut = 0; 1225 SlotIndex IntfIn, IntfOut; 1226 if (BI.LiveIn) { 1227 unsigned CandIn = BundleCand[Bundles->getBundle(Number, 0)]; 1228 if (CandIn != NoCand) { 1229 GlobalSplitCandidate &Cand = GlobalCand[CandIn]; 1230 IntvIn = Cand.IntvIdx; 1231 Cand.Intf.moveToBlock(Number); 1232 IntfIn = Cand.Intf.first(); 1233 } 1234 } 1235 if (BI.LiveOut) { 1236 unsigned CandOut = BundleCand[Bundles->getBundle(Number, 1)]; 1237 if (CandOut != NoCand) { 1238 GlobalSplitCandidate &Cand = GlobalCand[CandOut]; 1239 IntvOut = Cand.IntvIdx; 1240 Cand.Intf.moveToBlock(Number); 1241 IntfOut = Cand.Intf.last(); 1242 } 1243 } 1244 1245 // Create separate intervals for isolated blocks with multiple uses. 1246 if (!IntvIn && !IntvOut) { 1247 DEBUG(dbgs() << "BB#" << BI.MBB->getNumber() << " isolated.\n"); 1248 if (SA->shouldSplitSingleBlock(BI, SingleInstrs)) 1249 SE->splitSingleBlock(BI); 1250 continue; 1251 } 1252 1253 if (IntvIn && IntvOut) 1254 SE->splitLiveThroughBlock(Number, IntvIn, IntfIn, IntvOut, IntfOut); 1255 else if (IntvIn) 1256 SE->splitRegInBlock(BI, IntvIn, IntfIn); 1257 else 1258 SE->splitRegOutBlock(BI, IntvOut, IntfOut); 1259 } 1260 1261 // Handle live-through blocks. The relevant live-through blocks are stored in 1262 // the ActiveBlocks list with each candidate. We need to filter out 1263 // duplicates. 1264 BitVector Todo = SA->getThroughBlocks(); 1265 for (unsigned c = 0; c != UsedCands.size(); ++c) { 1266 ArrayRef<unsigned> Blocks = GlobalCand[UsedCands[c]].ActiveBlocks; 1267 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) { 1268 unsigned Number = Blocks[i]; 1269 if (!Todo.test(Number)) 1270 continue; 1271 Todo.reset(Number); 1272 1273 unsigned IntvIn = 0, IntvOut = 0; 1274 SlotIndex IntfIn, IntfOut; 1275 1276 unsigned CandIn = BundleCand[Bundles->getBundle(Number, 0)]; 1277 if (CandIn != NoCand) { 1278 GlobalSplitCandidate &Cand = GlobalCand[CandIn]; 1279 IntvIn = Cand.IntvIdx; 1280 Cand.Intf.moveToBlock(Number); 1281 IntfIn = Cand.Intf.first(); 1282 } 1283 1284 unsigned CandOut = BundleCand[Bundles->getBundle(Number, 1)]; 1285 if (CandOut != NoCand) { 1286 GlobalSplitCandidate &Cand = GlobalCand[CandOut]; 1287 IntvOut = Cand.IntvIdx; 1288 Cand.Intf.moveToBlock(Number); 1289 IntfOut = Cand.Intf.last(); 1290 } 1291 if (!IntvIn && !IntvOut) 1292 continue; 1293 SE->splitLiveThroughBlock(Number, IntvIn, IntfIn, IntvOut, IntfOut); 1294 } 1295 } 1296 1297 ++NumGlobalSplits; 1298 1299 SmallVector<unsigned, 8> IntvMap; 1300 SE->finish(&IntvMap); 1301 DebugVars->splitRegister(Reg, LREdit.regs(), *LIS); 1302 1303 ExtraRegInfo.resize(MRI->getNumVirtRegs()); 1304 unsigned OrigBlocks = SA->getNumLiveBlocks(); 1305 1306 // Sort out the new intervals created by splitting. We get four kinds: 1307 // - Remainder intervals should not be split again. 1308 // - Candidate intervals can be assigned to Cand.PhysReg. 1309 // - Block-local splits are candidates for local splitting. 1310 // - DCE leftovers should go back on the queue. 1311 for (unsigned i = 0, e = LREdit.size(); i != e; ++i) { 1312 LiveInterval &Reg = LIS->getInterval(LREdit.get(i)); 1313 1314 // Ignore old intervals from DCE. 1315 if (getStage(Reg) != RS_New) 1316 continue; 1317 1318 // Remainder interval. Don't try splitting again, spill if it doesn't 1319 // allocate. 1320 if (IntvMap[i] == 0) { 1321 setStage(Reg, RS_Spill); 1322 continue; 1323 } 1324 1325 // Global intervals. Allow repeated splitting as long as the number of live 1326 // blocks is strictly decreasing. 1327 if (IntvMap[i] < NumGlobalIntvs) { 1328 if (SA->countLiveBlocks(&Reg) >= OrigBlocks) { 1329 DEBUG(dbgs() << "Main interval covers the same " << OrigBlocks 1330 << " blocks as original.\n"); 1331 // Don't allow repeated splitting as a safe guard against looping. 1332 setStage(Reg, RS_Split2); 1333 } 1334 continue; 1335 } 1336 1337 // Other intervals are treated as new. This includes local intervals created 1338 // for blocks with multiple uses, and anything created by DCE. 1339 } 1340 1341 if (VerifyEnabled) 1342 MF->verify(this, "After splitting live range around region"); 1343 } 1344 1345 unsigned RAGreedy::tryRegionSplit(LiveInterval &VirtReg, AllocationOrder &Order, 1346 SmallVectorImpl<unsigned> &NewVRegs) { 1347 unsigned NumCands = 0; 1348 BlockFrequency BestCost; 1349 1350 // Check if we can split this live range around a compact region. 1351 bool HasCompact = calcCompactRegion(GlobalCand.front()); 1352 if (HasCompact) { 1353 // Yes, keep GlobalCand[0] as the compact region candidate. 1354 NumCands = 1; 1355 BestCost = BlockFrequency::getMaxFrequency(); 1356 } else { 1357 // No benefit from the compact region, our fallback will be per-block 1358 // splitting. Make sure we find a solution that is cheaper than spilling. 1359 BestCost = calcSpillCost(); 1360 DEBUG(dbgs() << "Cost of isolating all blocks = "; 1361 MBFI->printBlockFreq(dbgs(), BestCost) << '\n'); 1362 } 1363 1364 unsigned BestCand = 1365 calculateRegionSplitCost(VirtReg, Order, BestCost, NumCands, 1366 false/*IgnoreCSR*/); 1367 1368 // No solutions found, fall back to single block splitting. 1369 if (!HasCompact && BestCand == NoCand) 1370 return 0; 1371 1372 return doRegionSplit(VirtReg, BestCand, HasCompact, NewVRegs); 1373 } 1374 1375 unsigned RAGreedy::calculateRegionSplitCost(LiveInterval &VirtReg, 1376 AllocationOrder &Order, 1377 BlockFrequency &BestCost, 1378 unsigned &NumCands, 1379 bool IgnoreCSR) { 1380 unsigned BestCand = NoCand; 1381 Order.rewind(); 1382 while (unsigned PhysReg = Order.next()) { 1383 if (IgnoreCSR && isUnusedCalleeSavedReg(PhysReg)) 1384 continue; 1385 1386 // Discard bad candidates before we run out of interference cache cursors. 1387 // This will only affect register classes with a lot of registers (>32). 1388 if (NumCands == IntfCache.getMaxCursors()) { 1389 unsigned WorstCount = ~0u; 1390 unsigned Worst = 0; 1391 for (unsigned i = 0; i != NumCands; ++i) { 1392 if (i == BestCand || !GlobalCand[i].PhysReg) 1393 continue; 1394 unsigned Count = GlobalCand[i].LiveBundles.count(); 1395 if (Count < WorstCount) 1396 Worst = i, WorstCount = Count; 1397 } 1398 --NumCands; 1399 GlobalCand[Worst] = GlobalCand[NumCands]; 1400 if (BestCand == NumCands) 1401 BestCand = Worst; 1402 } 1403 1404 if (GlobalCand.size() <= NumCands) 1405 GlobalCand.resize(NumCands+1); 1406 GlobalSplitCandidate &Cand = GlobalCand[NumCands]; 1407 Cand.reset(IntfCache, PhysReg); 1408 1409 SpillPlacer->prepare(Cand.LiveBundles); 1410 BlockFrequency Cost; 1411 if (!addSplitConstraints(Cand.Intf, Cost)) { 1412 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << "\tno positive bundles\n"); 1413 continue; 1414 } 1415 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << "\tstatic = "; 1416 MBFI->printBlockFreq(dbgs(), Cost)); 1417 if (Cost >= BestCost) { 1418 DEBUG({ 1419 if (BestCand == NoCand) 1420 dbgs() << " worse than no bundles\n"; 1421 else 1422 dbgs() << " worse than " 1423 << PrintReg(GlobalCand[BestCand].PhysReg, TRI) << '\n'; 1424 }); 1425 continue; 1426 } 1427 growRegion(Cand); 1428 1429 SpillPlacer->finish(); 1430 1431 // No live bundles, defer to splitSingleBlocks(). 1432 if (!Cand.LiveBundles.any()) { 1433 DEBUG(dbgs() << " no bundles.\n"); 1434 continue; 1435 } 1436 1437 Cost += calcGlobalSplitCost(Cand); 1438 DEBUG({ 1439 dbgs() << ", total = "; MBFI->printBlockFreq(dbgs(), Cost) 1440 << " with bundles"; 1441 for (int i = Cand.LiveBundles.find_first(); i>=0; 1442 i = Cand.LiveBundles.find_next(i)) 1443 dbgs() << " EB#" << i; 1444 dbgs() << ".\n"; 1445 }); 1446 if (Cost < BestCost) { 1447 BestCand = NumCands; 1448 BestCost = Cost; 1449 } 1450 ++NumCands; 1451 } 1452 return BestCand; 1453 } 1454 1455 unsigned RAGreedy::doRegionSplit(LiveInterval &VirtReg, unsigned BestCand, 1456 bool HasCompact, 1457 SmallVectorImpl<unsigned> &NewVRegs) { 1458 SmallVector<unsigned, 8> UsedCands; 1459 // Prepare split editor. 1460 LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this); 1461 SE->reset(LREdit, SplitSpillMode); 1462 1463 // Assign all edge bundles to the preferred candidate, or NoCand. 1464 BundleCand.assign(Bundles->getNumBundles(), NoCand); 1465 1466 // Assign bundles for the best candidate region. 1467 if (BestCand != NoCand) { 1468 GlobalSplitCandidate &Cand = GlobalCand[BestCand]; 1469 if (unsigned B = Cand.getBundles(BundleCand, BestCand)) { 1470 UsedCands.push_back(BestCand); 1471 Cand.IntvIdx = SE->openIntv(); 1472 DEBUG(dbgs() << "Split for " << PrintReg(Cand.PhysReg, TRI) << " in " 1473 << B << " bundles, intv " << Cand.IntvIdx << ".\n"); 1474 (void)B; 1475 } 1476 } 1477 1478 // Assign bundles for the compact region. 1479 if (HasCompact) { 1480 GlobalSplitCandidate &Cand = GlobalCand.front(); 1481 assert(!Cand.PhysReg && "Compact region has no physreg"); 1482 if (unsigned B = Cand.getBundles(BundleCand, 0)) { 1483 UsedCands.push_back(0); 1484 Cand.IntvIdx = SE->openIntv(); 1485 DEBUG(dbgs() << "Split for compact region in " << B << " bundles, intv " 1486 << Cand.IntvIdx << ".\n"); 1487 (void)B; 1488 } 1489 } 1490 1491 splitAroundRegion(LREdit, UsedCands); 1492 return 0; 1493 } 1494 1495 1496 //===----------------------------------------------------------------------===// 1497 // Per-Block Splitting 1498 //===----------------------------------------------------------------------===// 1499 1500 /// tryBlockSplit - Split a global live range around every block with uses. This 1501 /// creates a lot of local live ranges, that will be split by tryLocalSplit if 1502 /// they don't allocate. 1503 unsigned RAGreedy::tryBlockSplit(LiveInterval &VirtReg, AllocationOrder &Order, 1504 SmallVectorImpl<unsigned> &NewVRegs) { 1505 assert(&SA->getParent() == &VirtReg && "Live range wasn't analyzed"); 1506 unsigned Reg = VirtReg.reg; 1507 bool SingleInstrs = RegClassInfo.isProperSubClass(MRI->getRegClass(Reg)); 1508 LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this); 1509 SE->reset(LREdit, SplitSpillMode); 1510 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks(); 1511 for (unsigned i = 0; i != UseBlocks.size(); ++i) { 1512 const SplitAnalysis::BlockInfo &BI = UseBlocks[i]; 1513 if (SA->shouldSplitSingleBlock(BI, SingleInstrs)) 1514 SE->splitSingleBlock(BI); 1515 } 1516 // No blocks were split. 1517 if (LREdit.empty()) 1518 return 0; 1519 1520 // We did split for some blocks. 1521 SmallVector<unsigned, 8> IntvMap; 1522 SE->finish(&IntvMap); 1523 1524 // Tell LiveDebugVariables about the new ranges. 1525 DebugVars->splitRegister(Reg, LREdit.regs(), *LIS); 1526 1527 ExtraRegInfo.resize(MRI->getNumVirtRegs()); 1528 1529 // Sort out the new intervals created by splitting. The remainder interval 1530 // goes straight to spilling, the new local ranges get to stay RS_New. 1531 for (unsigned i = 0, e = LREdit.size(); i != e; ++i) { 1532 LiveInterval &LI = LIS->getInterval(LREdit.get(i)); 1533 if (getStage(LI) == RS_New && IntvMap[i] == 0) 1534 setStage(LI, RS_Spill); 1535 } 1536 1537 if (VerifyEnabled) 1538 MF->verify(this, "After splitting live range around basic blocks"); 1539 return 0; 1540 } 1541 1542 1543 //===----------------------------------------------------------------------===// 1544 // Per-Instruction Splitting 1545 //===----------------------------------------------------------------------===// 1546 1547 /// Get the number of allocatable registers that match the constraints of \p Reg 1548 /// on \p MI and that are also in \p SuperRC. 1549 static unsigned getNumAllocatableRegsForConstraints( 1550 const MachineInstr *MI, unsigned Reg, const TargetRegisterClass *SuperRC, 1551 const TargetInstrInfo *TII, const TargetRegisterInfo *TRI, 1552 const RegisterClassInfo &RCI) { 1553 assert(SuperRC && "Invalid register class"); 1554 1555 const TargetRegisterClass *ConstrainedRC = 1556 MI->getRegClassConstraintEffectForVReg(Reg, SuperRC, TII, TRI, 1557 /* ExploreBundle */ true); 1558 if (!ConstrainedRC) 1559 return 0; 1560 return RCI.getNumAllocatableRegs(ConstrainedRC); 1561 } 1562 1563 /// tryInstructionSplit - Split a live range around individual instructions. 1564 /// This is normally not worthwhile since the spiller is doing essentially the 1565 /// same thing. However, when the live range is in a constrained register 1566 /// class, it may help to insert copies such that parts of the live range can 1567 /// be moved to a larger register class. 1568 /// 1569 /// This is similar to spilling to a larger register class. 1570 unsigned 1571 RAGreedy::tryInstructionSplit(LiveInterval &VirtReg, AllocationOrder &Order, 1572 SmallVectorImpl<unsigned> &NewVRegs) { 1573 const TargetRegisterClass *CurRC = MRI->getRegClass(VirtReg.reg); 1574 // There is no point to this if there are no larger sub-classes. 1575 if (!RegClassInfo.isProperSubClass(CurRC)) 1576 return 0; 1577 1578 // Always enable split spill mode, since we're effectively spilling to a 1579 // register. 1580 LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this); 1581 SE->reset(LREdit, SplitEditor::SM_Size); 1582 1583 ArrayRef<SlotIndex> Uses = SA->getUseSlots(); 1584 if (Uses.size() <= 1) 1585 return 0; 1586 1587 DEBUG(dbgs() << "Split around " << Uses.size() << " individual instrs.\n"); 1588 1589 const TargetRegisterClass *SuperRC = 1590 TRI->getLargestLegalSuperClass(CurRC, *MF); 1591 unsigned SuperRCNumAllocatableRegs = RCI.getNumAllocatableRegs(SuperRC); 1592 // Split around every non-copy instruction if this split will relax 1593 // the constraints on the virtual register. 1594 // Otherwise, splitting just inserts uncoalescable copies that do not help 1595 // the allocation. 1596 for (unsigned i = 0; i != Uses.size(); ++i) { 1597 if (const MachineInstr *MI = Indexes->getInstructionFromIndex(Uses[i])) 1598 if (MI->isFullCopy() || 1599 SuperRCNumAllocatableRegs == 1600 getNumAllocatableRegsForConstraints(MI, VirtReg.reg, SuperRC, TII, 1601 TRI, RCI)) { 1602 DEBUG(dbgs() << " skip:\t" << Uses[i] << '\t' << *MI); 1603 continue; 1604 } 1605 SE->openIntv(); 1606 SlotIndex SegStart = SE->enterIntvBefore(Uses[i]); 1607 SlotIndex SegStop = SE->leaveIntvAfter(Uses[i]); 1608 SE->useIntv(SegStart, SegStop); 1609 } 1610 1611 if (LREdit.empty()) { 1612 DEBUG(dbgs() << "All uses were copies.\n"); 1613 return 0; 1614 } 1615 1616 SmallVector<unsigned, 8> IntvMap; 1617 SE->finish(&IntvMap); 1618 DebugVars->splitRegister(VirtReg.reg, LREdit.regs(), *LIS); 1619 ExtraRegInfo.resize(MRI->getNumVirtRegs()); 1620 1621 // Assign all new registers to RS_Spill. This was the last chance. 1622 setStage(LREdit.begin(), LREdit.end(), RS_Spill); 1623 return 0; 1624 } 1625 1626 1627 //===----------------------------------------------------------------------===// 1628 // Local Splitting 1629 //===----------------------------------------------------------------------===// 1630 1631 1632 /// calcGapWeights - Compute the maximum spill weight that needs to be evicted 1633 /// in order to use PhysReg between two entries in SA->UseSlots. 1634 /// 1635 /// GapWeight[i] represents the gap between UseSlots[i] and UseSlots[i+1]. 1636 /// 1637 void RAGreedy::calcGapWeights(unsigned PhysReg, 1638 SmallVectorImpl<float> &GapWeight) { 1639 assert(SA->getUseBlocks().size() == 1 && "Not a local interval"); 1640 const SplitAnalysis::BlockInfo &BI = SA->getUseBlocks().front(); 1641 ArrayRef<SlotIndex> Uses = SA->getUseSlots(); 1642 const unsigned NumGaps = Uses.size()-1; 1643 1644 // Start and end points for the interference check. 1645 SlotIndex StartIdx = 1646 BI.LiveIn ? BI.FirstInstr.getBaseIndex() : BI.FirstInstr; 1647 SlotIndex StopIdx = 1648 BI.LiveOut ? BI.LastInstr.getBoundaryIndex() : BI.LastInstr; 1649 1650 GapWeight.assign(NumGaps, 0.0f); 1651 1652 // Add interference from each overlapping register. 1653 for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) { 1654 if (!Matrix->query(const_cast<LiveInterval&>(SA->getParent()), *Units) 1655 .checkInterference()) 1656 continue; 1657 1658 // We know that VirtReg is a continuous interval from FirstInstr to 1659 // LastInstr, so we don't need InterferenceQuery. 1660 // 1661 // Interference that overlaps an instruction is counted in both gaps 1662 // surrounding the instruction. The exception is interference before 1663 // StartIdx and after StopIdx. 1664 // 1665 LiveIntervalUnion::SegmentIter IntI = 1666 Matrix->getLiveUnions()[*Units] .find(StartIdx); 1667 for (unsigned Gap = 0; IntI.valid() && IntI.start() < StopIdx; ++IntI) { 1668 // Skip the gaps before IntI. 1669 while (Uses[Gap+1].getBoundaryIndex() < IntI.start()) 1670 if (++Gap == NumGaps) 1671 break; 1672 if (Gap == NumGaps) 1673 break; 1674 1675 // Update the gaps covered by IntI. 1676 const float weight = IntI.value()->weight; 1677 for (; Gap != NumGaps; ++Gap) { 1678 GapWeight[Gap] = std::max(GapWeight[Gap], weight); 1679 if (Uses[Gap+1].getBaseIndex() >= IntI.stop()) 1680 break; 1681 } 1682 if (Gap == NumGaps) 1683 break; 1684 } 1685 } 1686 1687 // Add fixed interference. 1688 for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) { 1689 const LiveRange &LR = LIS->getRegUnit(*Units); 1690 LiveRange::const_iterator I = LR.find(StartIdx); 1691 LiveRange::const_iterator E = LR.end(); 1692 1693 // Same loop as above. Mark any overlapped gaps as HUGE_VALF. 1694 for (unsigned Gap = 0; I != E && I->start < StopIdx; ++I) { 1695 while (Uses[Gap+1].getBoundaryIndex() < I->start) 1696 if (++Gap == NumGaps) 1697 break; 1698 if (Gap == NumGaps) 1699 break; 1700 1701 for (; Gap != NumGaps; ++Gap) { 1702 GapWeight[Gap] = llvm::huge_valf; 1703 if (Uses[Gap+1].getBaseIndex() >= I->end) 1704 break; 1705 } 1706 if (Gap == NumGaps) 1707 break; 1708 } 1709 } 1710 } 1711 1712 /// tryLocalSplit - Try to split VirtReg into smaller intervals inside its only 1713 /// basic block. 1714 /// 1715 unsigned RAGreedy::tryLocalSplit(LiveInterval &VirtReg, AllocationOrder &Order, 1716 SmallVectorImpl<unsigned> &NewVRegs) { 1717 assert(SA->getUseBlocks().size() == 1 && "Not a local interval"); 1718 const SplitAnalysis::BlockInfo &BI = SA->getUseBlocks().front(); 1719 1720 // Note that it is possible to have an interval that is live-in or live-out 1721 // while only covering a single block - A phi-def can use undef values from 1722 // predecessors, and the block could be a single-block loop. 1723 // We don't bother doing anything clever about such a case, we simply assume 1724 // that the interval is continuous from FirstInstr to LastInstr. We should 1725 // make sure that we don't do anything illegal to such an interval, though. 1726 1727 ArrayRef<SlotIndex> Uses = SA->getUseSlots(); 1728 if (Uses.size() <= 2) 1729 return 0; 1730 const unsigned NumGaps = Uses.size()-1; 1731 1732 DEBUG({ 1733 dbgs() << "tryLocalSplit: "; 1734 for (unsigned i = 0, e = Uses.size(); i != e; ++i) 1735 dbgs() << ' ' << Uses[i]; 1736 dbgs() << '\n'; 1737 }); 1738 1739 // If VirtReg is live across any register mask operands, compute a list of 1740 // gaps with register masks. 1741 SmallVector<unsigned, 8> RegMaskGaps; 1742 if (Matrix->checkRegMaskInterference(VirtReg)) { 1743 // Get regmask slots for the whole block. 1744 ArrayRef<SlotIndex> RMS = LIS->getRegMaskSlotsInBlock(BI.MBB->getNumber()); 1745 DEBUG(dbgs() << RMS.size() << " regmasks in block:"); 1746 // Constrain to VirtReg's live range. 1747 unsigned ri = std::lower_bound(RMS.begin(), RMS.end(), 1748 Uses.front().getRegSlot()) - RMS.begin(); 1749 unsigned re = RMS.size(); 1750 for (unsigned i = 0; i != NumGaps && ri != re; ++i) { 1751 // Look for Uses[i] <= RMS <= Uses[i+1]. 1752 assert(!SlotIndex::isEarlierInstr(RMS[ri], Uses[i])); 1753 if (SlotIndex::isEarlierInstr(Uses[i+1], RMS[ri])) 1754 continue; 1755 // Skip a regmask on the same instruction as the last use. It doesn't 1756 // overlap the live range. 1757 if (SlotIndex::isSameInstr(Uses[i+1], RMS[ri]) && i+1 == NumGaps) 1758 break; 1759 DEBUG(dbgs() << ' ' << RMS[ri] << ':' << Uses[i] << '-' << Uses[i+1]); 1760 RegMaskGaps.push_back(i); 1761 // Advance ri to the next gap. A regmask on one of the uses counts in 1762 // both gaps. 1763 while (ri != re && SlotIndex::isEarlierInstr(RMS[ri], Uses[i+1])) 1764 ++ri; 1765 } 1766 DEBUG(dbgs() << '\n'); 1767 } 1768 1769 // Since we allow local split results to be split again, there is a risk of 1770 // creating infinite loops. It is tempting to require that the new live 1771 // ranges have less instructions than the original. That would guarantee 1772 // convergence, but it is too strict. A live range with 3 instructions can be 1773 // split 2+3 (including the COPY), and we want to allow that. 1774 // 1775 // Instead we use these rules: 1776 // 1777 // 1. Allow any split for ranges with getStage() < RS_Split2. (Except for the 1778 // noop split, of course). 1779 // 2. Require progress be made for ranges with getStage() == RS_Split2. All 1780 // the new ranges must have fewer instructions than before the split. 1781 // 3. New ranges with the same number of instructions are marked RS_Split2, 1782 // smaller ranges are marked RS_New. 1783 // 1784 // These rules allow a 3 -> 2+3 split once, which we need. They also prevent 1785 // excessive splitting and infinite loops. 1786 // 1787 bool ProgressRequired = getStage(VirtReg) >= RS_Split2; 1788 1789 // Best split candidate. 1790 unsigned BestBefore = NumGaps; 1791 unsigned BestAfter = 0; 1792 float BestDiff = 0; 1793 1794 const float blockFreq = 1795 SpillPlacer->getBlockFrequency(BI.MBB->getNumber()).getFrequency() * 1796 (1.0f / MBFI->getEntryFreq()); 1797 SmallVector<float, 8> GapWeight; 1798 1799 Order.rewind(); 1800 while (unsigned PhysReg = Order.next()) { 1801 // Keep track of the largest spill weight that would need to be evicted in 1802 // order to make use of PhysReg between UseSlots[i] and UseSlots[i+1]. 1803 calcGapWeights(PhysReg, GapWeight); 1804 1805 // Remove any gaps with regmask clobbers. 1806 if (Matrix->checkRegMaskInterference(VirtReg, PhysReg)) 1807 for (unsigned i = 0, e = RegMaskGaps.size(); i != e; ++i) 1808 GapWeight[RegMaskGaps[i]] = llvm::huge_valf; 1809 1810 // Try to find the best sequence of gaps to close. 1811 // The new spill weight must be larger than any gap interference. 1812 1813 // We will split before Uses[SplitBefore] and after Uses[SplitAfter]. 1814 unsigned SplitBefore = 0, SplitAfter = 1; 1815 1816 // MaxGap should always be max(GapWeight[SplitBefore..SplitAfter-1]). 1817 // It is the spill weight that needs to be evicted. 1818 float MaxGap = GapWeight[0]; 1819 1820 for (;;) { 1821 // Live before/after split? 1822 const bool LiveBefore = SplitBefore != 0 || BI.LiveIn; 1823 const bool LiveAfter = SplitAfter != NumGaps || BI.LiveOut; 1824 1825 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << ' ' 1826 << Uses[SplitBefore] << '-' << Uses[SplitAfter] 1827 << " i=" << MaxGap); 1828 1829 // Stop before the interval gets so big we wouldn't be making progress. 1830 if (!LiveBefore && !LiveAfter) { 1831 DEBUG(dbgs() << " all\n"); 1832 break; 1833 } 1834 // Should the interval be extended or shrunk? 1835 bool Shrink = true; 1836 1837 // How many gaps would the new range have? 1838 unsigned NewGaps = LiveBefore + SplitAfter - SplitBefore + LiveAfter; 1839 1840 // Legally, without causing looping? 1841 bool Legal = !ProgressRequired || NewGaps < NumGaps; 1842 1843 if (Legal && MaxGap < llvm::huge_valf) { 1844 // Estimate the new spill weight. Each instruction reads or writes the 1845 // register. Conservatively assume there are no read-modify-write 1846 // instructions. 1847 // 1848 // Try to guess the size of the new interval. 1849 const float EstWeight = normalizeSpillWeight( 1850 blockFreq * (NewGaps + 1), 1851 Uses[SplitBefore].distance(Uses[SplitAfter]) + 1852 (LiveBefore + LiveAfter) * SlotIndex::InstrDist, 1853 1); 1854 // Would this split be possible to allocate? 1855 // Never allocate all gaps, we wouldn't be making progress. 1856 DEBUG(dbgs() << " w=" << EstWeight); 1857 if (EstWeight * Hysteresis >= MaxGap) { 1858 Shrink = false; 1859 float Diff = EstWeight - MaxGap; 1860 if (Diff > BestDiff) { 1861 DEBUG(dbgs() << " (best)"); 1862 BestDiff = Hysteresis * Diff; 1863 BestBefore = SplitBefore; 1864 BestAfter = SplitAfter; 1865 } 1866 } 1867 } 1868 1869 // Try to shrink. 1870 if (Shrink) { 1871 if (++SplitBefore < SplitAfter) { 1872 DEBUG(dbgs() << " shrink\n"); 1873 // Recompute the max when necessary. 1874 if (GapWeight[SplitBefore - 1] >= MaxGap) { 1875 MaxGap = GapWeight[SplitBefore]; 1876 for (unsigned i = SplitBefore + 1; i != SplitAfter; ++i) 1877 MaxGap = std::max(MaxGap, GapWeight[i]); 1878 } 1879 continue; 1880 } 1881 MaxGap = 0; 1882 } 1883 1884 // Try to extend the interval. 1885 if (SplitAfter >= NumGaps) { 1886 DEBUG(dbgs() << " end\n"); 1887 break; 1888 } 1889 1890 DEBUG(dbgs() << " extend\n"); 1891 MaxGap = std::max(MaxGap, GapWeight[SplitAfter++]); 1892 } 1893 } 1894 1895 // Didn't find any candidates? 1896 if (BestBefore == NumGaps) 1897 return 0; 1898 1899 DEBUG(dbgs() << "Best local split range: " << Uses[BestBefore] 1900 << '-' << Uses[BestAfter] << ", " << BestDiff 1901 << ", " << (BestAfter - BestBefore + 1) << " instrs\n"); 1902 1903 LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this); 1904 SE->reset(LREdit); 1905 1906 SE->openIntv(); 1907 SlotIndex SegStart = SE->enterIntvBefore(Uses[BestBefore]); 1908 SlotIndex SegStop = SE->leaveIntvAfter(Uses[BestAfter]); 1909 SE->useIntv(SegStart, SegStop); 1910 SmallVector<unsigned, 8> IntvMap; 1911 SE->finish(&IntvMap); 1912 DebugVars->splitRegister(VirtReg.reg, LREdit.regs(), *LIS); 1913 1914 // If the new range has the same number of instructions as before, mark it as 1915 // RS_Split2 so the next split will be forced to make progress. Otherwise, 1916 // leave the new intervals as RS_New so they can compete. 1917 bool LiveBefore = BestBefore != 0 || BI.LiveIn; 1918 bool LiveAfter = BestAfter != NumGaps || BI.LiveOut; 1919 unsigned NewGaps = LiveBefore + BestAfter - BestBefore + LiveAfter; 1920 if (NewGaps >= NumGaps) { 1921 DEBUG(dbgs() << "Tagging non-progress ranges: "); 1922 assert(!ProgressRequired && "Didn't make progress when it was required."); 1923 for (unsigned i = 0, e = IntvMap.size(); i != e; ++i) 1924 if (IntvMap[i] == 1) { 1925 setStage(LIS->getInterval(LREdit.get(i)), RS_Split2); 1926 DEBUG(dbgs() << PrintReg(LREdit.get(i))); 1927 } 1928 DEBUG(dbgs() << '\n'); 1929 } 1930 ++NumLocalSplits; 1931 1932 return 0; 1933 } 1934 1935 //===----------------------------------------------------------------------===// 1936 // Live Range Splitting 1937 //===----------------------------------------------------------------------===// 1938 1939 /// trySplit - Try to split VirtReg or one of its interferences, making it 1940 /// assignable. 1941 /// @return Physreg when VirtReg may be assigned and/or new NewVRegs. 1942 unsigned RAGreedy::trySplit(LiveInterval &VirtReg, AllocationOrder &Order, 1943 SmallVectorImpl<unsigned>&NewVRegs) { 1944 // Ranges must be Split2 or less. 1945 if (getStage(VirtReg) >= RS_Spill) 1946 return 0; 1947 1948 // Local intervals are handled separately. 1949 if (LIS->intervalIsInOneMBB(VirtReg)) { 1950 NamedRegionTimer T("Local Splitting", TimerGroupName, TimePassesIsEnabled); 1951 SA->analyze(&VirtReg); 1952 unsigned PhysReg = tryLocalSplit(VirtReg, Order, NewVRegs); 1953 if (PhysReg || !NewVRegs.empty()) 1954 return PhysReg; 1955 return tryInstructionSplit(VirtReg, Order, NewVRegs); 1956 } 1957 1958 NamedRegionTimer T("Global Splitting", TimerGroupName, TimePassesIsEnabled); 1959 1960 SA->analyze(&VirtReg); 1961 1962 // FIXME: SplitAnalysis may repair broken live ranges coming from the 1963 // coalescer. That may cause the range to become allocatable which means that 1964 // tryRegionSplit won't be making progress. This check should be replaced with 1965 // an assertion when the coalescer is fixed. 1966 if (SA->didRepairRange()) { 1967 // VirtReg has changed, so all cached queries are invalid. 1968 Matrix->invalidateVirtRegs(); 1969 if (unsigned PhysReg = tryAssign(VirtReg, Order, NewVRegs)) 1970 return PhysReg; 1971 } 1972 1973 // First try to split around a region spanning multiple blocks. RS_Split2 1974 // ranges already made dubious progress with region splitting, so they go 1975 // straight to single block splitting. 1976 if (getStage(VirtReg) < RS_Split2) { 1977 unsigned PhysReg = tryRegionSplit(VirtReg, Order, NewVRegs); 1978 if (PhysReg || !NewVRegs.empty()) 1979 return PhysReg; 1980 } 1981 1982 // Then isolate blocks. 1983 return tryBlockSplit(VirtReg, Order, NewVRegs); 1984 } 1985 1986 //===----------------------------------------------------------------------===// 1987 // Last Chance Recoloring 1988 //===----------------------------------------------------------------------===// 1989 1990 /// mayRecolorAllInterferences - Check if the virtual registers that 1991 /// interfere with \p VirtReg on \p PhysReg (or one of its aliases) may be 1992 /// recolored to free \p PhysReg. 1993 /// When true is returned, \p RecoloringCandidates has been augmented with all 1994 /// the live intervals that need to be recolored in order to free \p PhysReg 1995 /// for \p VirtReg. 1996 /// \p FixedRegisters contains all the virtual registers that cannot be 1997 /// recolored. 1998 bool 1999 RAGreedy::mayRecolorAllInterferences(unsigned PhysReg, LiveInterval &VirtReg, 2000 SmallLISet &RecoloringCandidates, 2001 const SmallVirtRegSet &FixedRegisters) { 2002 const TargetRegisterClass *CurRC = MRI->getRegClass(VirtReg.reg); 2003 2004 for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) { 2005 LiveIntervalUnion::Query &Q = Matrix->query(VirtReg, *Units); 2006 // If there is LastChanceRecoloringMaxInterference or more interferences, 2007 // chances are one would not be recolorable. 2008 if (Q.collectInterferingVRegs(LastChanceRecoloringMaxInterference) >= 2009 LastChanceRecoloringMaxInterference && !ExhaustiveSearch) { 2010 DEBUG(dbgs() << "Early abort: too many interferences.\n"); 2011 CutOffInfo |= CO_Interf; 2012 return false; 2013 } 2014 for (unsigned i = Q.interferingVRegs().size(); i; --i) { 2015 LiveInterval *Intf = Q.interferingVRegs()[i - 1]; 2016 // If Intf is done and sit on the same register class as VirtReg, 2017 // it would not be recolorable as it is in the same state as VirtReg. 2018 if ((getStage(*Intf) == RS_Done && 2019 MRI->getRegClass(Intf->reg) == CurRC) || 2020 FixedRegisters.count(Intf->reg)) { 2021 DEBUG(dbgs() << "Early abort: the inteference is not recolorable.\n"); 2022 return false; 2023 } 2024 RecoloringCandidates.insert(Intf); 2025 } 2026 } 2027 return true; 2028 } 2029 2030 /// tryLastChanceRecoloring - Try to assign a color to \p VirtReg by recoloring 2031 /// its interferences. 2032 /// Last chance recoloring chooses a color for \p VirtReg and recolors every 2033 /// virtual register that was using it. The recoloring process may recursively 2034 /// use the last chance recoloring. Therefore, when a virtual register has been 2035 /// assigned a color by this mechanism, it is marked as Fixed, i.e., it cannot 2036 /// be last-chance-recolored again during this recoloring "session". 2037 /// E.g., 2038 /// Let 2039 /// vA can use {R1, R2 } 2040 /// vB can use { R2, R3} 2041 /// vC can use {R1 } 2042 /// Where vA, vB, and vC cannot be split anymore (they are reloads for 2043 /// instance) and they all interfere. 2044 /// 2045 /// vA is assigned R1 2046 /// vB is assigned R2 2047 /// vC tries to evict vA but vA is already done. 2048 /// Regular register allocation fails. 2049 /// 2050 /// Last chance recoloring kicks in: 2051 /// vC does as if vA was evicted => vC uses R1. 2052 /// vC is marked as fixed. 2053 /// vA needs to find a color. 2054 /// None are available. 2055 /// vA cannot evict vC: vC is a fixed virtual register now. 2056 /// vA does as if vB was evicted => vA uses R2. 2057 /// vB needs to find a color. 2058 /// R3 is available. 2059 /// Recoloring => vC = R1, vA = R2, vB = R3 2060 /// 2061 /// \p Order defines the preferred allocation order for \p VirtReg. 2062 /// \p NewRegs will contain any new virtual register that have been created 2063 /// (split, spill) during the process and that must be assigned. 2064 /// \p FixedRegisters contains all the virtual registers that cannot be 2065 /// recolored. 2066 /// \p Depth gives the current depth of the last chance recoloring. 2067 /// \return a physical register that can be used for VirtReg or ~0u if none 2068 /// exists. 2069 unsigned RAGreedy::tryLastChanceRecoloring(LiveInterval &VirtReg, 2070 AllocationOrder &Order, 2071 SmallVectorImpl<unsigned> &NewVRegs, 2072 SmallVirtRegSet &FixedRegisters, 2073 unsigned Depth) { 2074 DEBUG(dbgs() << "Try last chance recoloring for " << VirtReg << '\n'); 2075 // Ranges must be Done. 2076 assert((getStage(VirtReg) >= RS_Done || !VirtReg.isSpillable()) && 2077 "Last chance recoloring should really be last chance"); 2078 // Set the max depth to LastChanceRecoloringMaxDepth. 2079 // We may want to reconsider that if we end up with a too large search space 2080 // for target with hundreds of registers. 2081 // Indeed, in that case we may want to cut the search space earlier. 2082 if (Depth >= LastChanceRecoloringMaxDepth && !ExhaustiveSearch) { 2083 DEBUG(dbgs() << "Abort because max depth has been reached.\n"); 2084 CutOffInfo |= CO_Depth; 2085 return ~0u; 2086 } 2087 2088 // Set of Live intervals that will need to be recolored. 2089 SmallLISet RecoloringCandidates; 2090 // Record the original mapping virtual register to physical register in case 2091 // the recoloring fails. 2092 DenseMap<unsigned, unsigned> VirtRegToPhysReg; 2093 // Mark VirtReg as fixed, i.e., it will not be recolored pass this point in 2094 // this recoloring "session". 2095 FixedRegisters.insert(VirtReg.reg); 2096 2097 Order.rewind(); 2098 while (unsigned PhysReg = Order.next()) { 2099 DEBUG(dbgs() << "Try to assign: " << VirtReg << " to " 2100 << PrintReg(PhysReg, TRI) << '\n'); 2101 RecoloringCandidates.clear(); 2102 VirtRegToPhysReg.clear(); 2103 2104 // It is only possible to recolor virtual register interference. 2105 if (Matrix->checkInterference(VirtReg, PhysReg) > 2106 LiveRegMatrix::IK_VirtReg) { 2107 DEBUG(dbgs() << "Some inteferences are not with virtual registers.\n"); 2108 2109 continue; 2110 } 2111 2112 // Early give up on this PhysReg if it is obvious we cannot recolor all 2113 // the interferences. 2114 if (!mayRecolorAllInterferences(PhysReg, VirtReg, RecoloringCandidates, 2115 FixedRegisters)) { 2116 DEBUG(dbgs() << "Some inteferences cannot be recolored.\n"); 2117 continue; 2118 } 2119 2120 // RecoloringCandidates contains all the virtual registers that interfer 2121 // with VirtReg on PhysReg (or one of its aliases). 2122 // Enqueue them for recoloring and perform the actual recoloring. 2123 PQueue RecoloringQueue; 2124 for (SmallLISet::iterator It = RecoloringCandidates.begin(), 2125 EndIt = RecoloringCandidates.end(); 2126 It != EndIt; ++It) { 2127 unsigned ItVirtReg = (*It)->reg; 2128 enqueue(RecoloringQueue, *It); 2129 assert(VRM->hasPhys(ItVirtReg) && 2130 "Interferences are supposed to be with allocated vairables"); 2131 2132 // Record the current allocation. 2133 VirtRegToPhysReg[ItVirtReg] = VRM->getPhys(ItVirtReg); 2134 // unset the related struct. 2135 Matrix->unassign(**It); 2136 } 2137 2138 // Do as if VirtReg was assigned to PhysReg so that the underlying 2139 // recoloring has the right information about the interferes and 2140 // available colors. 2141 Matrix->assign(VirtReg, PhysReg); 2142 2143 // Save the current recoloring state. 2144 // If we cannot recolor all the interferences, we will have to start again 2145 // at this point for the next physical register. 2146 SmallVirtRegSet SaveFixedRegisters(FixedRegisters); 2147 if (tryRecoloringCandidates(RecoloringQueue, NewVRegs, FixedRegisters, 2148 Depth)) { 2149 // Do not mess up with the global assignment process. 2150 // I.e., VirtReg must be unassigned. 2151 Matrix->unassign(VirtReg); 2152 return PhysReg; 2153 } 2154 2155 DEBUG(dbgs() << "Fail to assign: " << VirtReg << " to " 2156 << PrintReg(PhysReg, TRI) << '\n'); 2157 2158 // The recoloring attempt failed, undo the changes. 2159 FixedRegisters = SaveFixedRegisters; 2160 Matrix->unassign(VirtReg); 2161 2162 for (SmallLISet::iterator It = RecoloringCandidates.begin(), 2163 EndIt = RecoloringCandidates.end(); 2164 It != EndIt; ++It) { 2165 unsigned ItVirtReg = (*It)->reg; 2166 if (VRM->hasPhys(ItVirtReg)) 2167 Matrix->unassign(**It); 2168 unsigned ItPhysReg = VirtRegToPhysReg[ItVirtReg]; 2169 Matrix->assign(**It, ItPhysReg); 2170 } 2171 } 2172 2173 // Last chance recoloring did not worked either, give up. 2174 return ~0u; 2175 } 2176 2177 /// tryRecoloringCandidates - Try to assign a new color to every register 2178 /// in \RecoloringQueue. 2179 /// \p NewRegs will contain any new virtual register created during the 2180 /// recoloring process. 2181 /// \p FixedRegisters[in/out] contains all the registers that have been 2182 /// recolored. 2183 /// \return true if all virtual registers in RecoloringQueue were successfully 2184 /// recolored, false otherwise. 2185 bool RAGreedy::tryRecoloringCandidates(PQueue &RecoloringQueue, 2186 SmallVectorImpl<unsigned> &NewVRegs, 2187 SmallVirtRegSet &FixedRegisters, 2188 unsigned Depth) { 2189 while (!RecoloringQueue.empty()) { 2190 LiveInterval *LI = dequeue(RecoloringQueue); 2191 DEBUG(dbgs() << "Try to recolor: " << *LI << '\n'); 2192 unsigned PhysReg; 2193 PhysReg = selectOrSplitImpl(*LI, NewVRegs, FixedRegisters, Depth + 1); 2194 if (PhysReg == ~0u || !PhysReg) 2195 return false; 2196 DEBUG(dbgs() << "Recoloring of " << *LI 2197 << " succeeded with: " << PrintReg(PhysReg, TRI) << '\n'); 2198 Matrix->assign(*LI, PhysReg); 2199 FixedRegisters.insert(LI->reg); 2200 } 2201 return true; 2202 } 2203 2204 //===----------------------------------------------------------------------===// 2205 // Main Entry Point 2206 //===----------------------------------------------------------------------===// 2207 2208 unsigned RAGreedy::selectOrSplit(LiveInterval &VirtReg, 2209 SmallVectorImpl<unsigned> &NewVRegs) { 2210 CutOffInfo = CO_None; 2211 LLVMContext &Ctx = MF->getFunction()->getContext(); 2212 SmallVirtRegSet FixedRegisters; 2213 unsigned Reg = selectOrSplitImpl(VirtReg, NewVRegs, FixedRegisters); 2214 if (Reg == ~0U && (CutOffInfo != CO_None)) { 2215 uint8_t CutOffEncountered = CutOffInfo & (CO_Depth | CO_Interf); 2216 if (CutOffEncountered == CO_Depth) 2217 Ctx.emitError("register allocation failed: maximum depth for recoloring " 2218 "reached. Use -fexhaustive-register-search to skip " 2219 "cutoffs"); 2220 else if (CutOffEncountered == CO_Interf) 2221 Ctx.emitError("register allocation failed: maximum interference for " 2222 "recoloring reached. Use -fexhaustive-register-search " 2223 "to skip cutoffs"); 2224 else if (CutOffEncountered == (CO_Depth | CO_Interf)) 2225 Ctx.emitError("register allocation failed: maximum interference and " 2226 "depth for recoloring reached. Use " 2227 "-fexhaustive-register-search to skip cutoffs"); 2228 } 2229 return Reg; 2230 } 2231 2232 /// Using a CSR for the first time has a cost because it causes push|pop 2233 /// to be added to prologue|epilogue. Splitting a cold section of the live 2234 /// range can have lower cost than using the CSR for the first time; 2235 /// Spilling a live range in the cold path can have lower cost than using 2236 /// the CSR for the first time. Returns the physical register if we decide 2237 /// to use the CSR; otherwise return 0. 2238 unsigned RAGreedy::tryAssignCSRFirstTime(LiveInterval &VirtReg, 2239 AllocationOrder &Order, 2240 unsigned PhysReg, 2241 unsigned &CostPerUseLimit, 2242 SmallVectorImpl<unsigned> &NewVRegs) { 2243 if (getStage(VirtReg) == RS_Spill && VirtReg.isSpillable()) { 2244 // We choose spill over using the CSR for the first time if the spill cost 2245 // is lower than CSRCost. 2246 SA->analyze(&VirtReg); 2247 if (calcSpillCost() >= CSRCost) 2248 return PhysReg; 2249 2250 // We are going to spill, set CostPerUseLimit to 1 to make sure that 2251 // we will not use a callee-saved register in tryEvict. 2252 CostPerUseLimit = 1; 2253 return 0; 2254 } 2255 if (getStage(VirtReg) < RS_Split) { 2256 // We choose pre-splitting over using the CSR for the first time if 2257 // the cost of splitting is lower than CSRCost. 2258 SA->analyze(&VirtReg); 2259 unsigned NumCands = 0; 2260 BlockFrequency BestCost = CSRCost; // Don't modify CSRCost. 2261 unsigned BestCand = calculateRegionSplitCost(VirtReg, Order, BestCost, 2262 NumCands, true /*IgnoreCSR*/); 2263 if (BestCand == NoCand) 2264 // Use the CSR if we can't find a region split below CSRCost. 2265 return PhysReg; 2266 2267 // Perform the actual pre-splitting. 2268 doRegionSplit(VirtReg, BestCand, false/*HasCompact*/, NewVRegs); 2269 return 0; 2270 } 2271 return PhysReg; 2272 } 2273 2274 void RAGreedy::aboutToRemoveInterval(LiveInterval &LI) { 2275 // Do not keep invalid information around. 2276 SetOfBrokenHints.remove(&LI); 2277 } 2278 2279 void RAGreedy::initializeCSRCost() { 2280 // We use the larger one out of the command-line option and the value report 2281 // by TRI. 2282 CSRCost = BlockFrequency( 2283 std::max((unsigned)CSRFirstTimeCost, TRI->getCSRFirstUseCost())); 2284 if (!CSRCost.getFrequency()) 2285 return; 2286 2287 // Raw cost is relative to Entry == 2^14; scale it appropriately. 2288 uint64_t ActualEntry = MBFI->getEntryFreq(); 2289 if (!ActualEntry) { 2290 CSRCost = 0; 2291 return; 2292 } 2293 uint64_t FixedEntry = 1 << 14; 2294 if (ActualEntry < FixedEntry) 2295 CSRCost *= BranchProbability(ActualEntry, FixedEntry); 2296 else if (ActualEntry <= UINT32_MAX) 2297 // Invert the fraction and divide. 2298 CSRCost /= BranchProbability(FixedEntry, ActualEntry); 2299 else 2300 // Can't use BranchProbability in general, since it takes 32-bit numbers. 2301 CSRCost = CSRCost.getFrequency() * (ActualEntry / FixedEntry); 2302 } 2303 2304 /// \brief Collect the hint info for \p Reg. 2305 /// The results are stored into \p Out. 2306 /// \p Out is not cleared before being populated. 2307 void RAGreedy::collectHintInfo(unsigned Reg, HintsInfo &Out) { 2308 for (const MachineInstr &Instr : MRI->reg_nodbg_instructions(Reg)) { 2309 if (!Instr.isFullCopy()) 2310 continue; 2311 // Look for the other end of the copy. 2312 unsigned OtherReg = Instr.getOperand(0).getReg(); 2313 if (OtherReg == Reg) { 2314 OtherReg = Instr.getOperand(1).getReg(); 2315 if (OtherReg == Reg) 2316 continue; 2317 } 2318 // Get the current assignment. 2319 unsigned OtherPhysReg = TargetRegisterInfo::isPhysicalRegister(OtherReg) 2320 ? OtherReg 2321 : VRM->getPhys(OtherReg); 2322 // Push the collected information. 2323 Out.push_back(HintInfo(MBFI->getBlockFreq(Instr.getParent()), OtherReg, 2324 OtherPhysReg)); 2325 } 2326 } 2327 2328 /// \brief Using the given \p List, compute the cost of the broken hints if 2329 /// \p PhysReg was used. 2330 /// \return The cost of \p List for \p PhysReg. 2331 BlockFrequency RAGreedy::getBrokenHintFreq(const HintsInfo &List, 2332 unsigned PhysReg) { 2333 BlockFrequency Cost = 0; 2334 for (const HintInfo &Info : List) { 2335 if (Info.PhysReg != PhysReg) 2336 Cost += Info.Freq; 2337 } 2338 return Cost; 2339 } 2340 2341 /// \brief Using the register assigned to \p VirtReg, try to recolor 2342 /// all the live ranges that are copy-related with \p VirtReg. 2343 /// The recoloring is then propagated to all the live-ranges that have 2344 /// been recolored and so on, until no more copies can be coalesced or 2345 /// it is not profitable. 2346 /// For a given live range, profitability is determined by the sum of the 2347 /// frequencies of the non-identity copies it would introduce with the old 2348 /// and new register. 2349 void RAGreedy::tryHintRecoloring(LiveInterval &VirtReg) { 2350 // We have a broken hint, check if it is possible to fix it by 2351 // reusing PhysReg for the copy-related live-ranges. Indeed, we evicted 2352 // some register and PhysReg may be available for the other live-ranges. 2353 SmallSet<unsigned, 4> Visited; 2354 SmallVector<unsigned, 2> RecoloringCandidates; 2355 HintsInfo Info; 2356 unsigned Reg = VirtReg.reg; 2357 unsigned PhysReg = VRM->getPhys(Reg); 2358 // Start the recoloring algorithm from the input live-interval, then 2359 // it will propagate to the ones that are copy-related with it. 2360 Visited.insert(Reg); 2361 RecoloringCandidates.push_back(Reg); 2362 2363 DEBUG(dbgs() << "Trying to reconcile hints for: " << PrintReg(Reg, TRI) << '(' 2364 << PrintReg(PhysReg, TRI) << ")\n"); 2365 2366 do { 2367 Reg = RecoloringCandidates.pop_back_val(); 2368 2369 // We cannot recolor physcal register. 2370 if (TargetRegisterInfo::isPhysicalRegister(Reg)) 2371 continue; 2372 2373 assert(VRM->hasPhys(Reg) && "We have unallocated variable!!"); 2374 2375 // Get the live interval mapped with this virtual register to be able 2376 // to check for the interference with the new color. 2377 LiveInterval &LI = LIS->getInterval(Reg); 2378 unsigned CurrPhys = VRM->getPhys(Reg); 2379 // Check that the new color matches the register class constraints and 2380 // that it is free for this live range. 2381 if (CurrPhys != PhysReg && (!MRI->getRegClass(Reg)->contains(PhysReg) || 2382 Matrix->checkInterference(LI, PhysReg))) 2383 continue; 2384 2385 DEBUG(dbgs() << PrintReg(Reg, TRI) << '(' << PrintReg(CurrPhys, TRI) 2386 << ") is recolorable.\n"); 2387 2388 // Gather the hint info. 2389 Info.clear(); 2390 collectHintInfo(Reg, Info); 2391 // Check if recoloring the live-range will increase the cost of the 2392 // non-identity copies. 2393 if (CurrPhys != PhysReg) { 2394 DEBUG(dbgs() << "Checking profitability:\n"); 2395 BlockFrequency OldCopiesCost = getBrokenHintFreq(Info, CurrPhys); 2396 BlockFrequency NewCopiesCost = getBrokenHintFreq(Info, PhysReg); 2397 DEBUG(dbgs() << "Old Cost: " << OldCopiesCost.getFrequency() 2398 << "\nNew Cost: " << NewCopiesCost.getFrequency() << '\n'); 2399 if (OldCopiesCost < NewCopiesCost) { 2400 DEBUG(dbgs() << "=> Not profitable.\n"); 2401 continue; 2402 } 2403 // At this point, the cost is either cheaper or equal. If it is 2404 // equal, we consider this is profitable because it may expose 2405 // more recoloring opportunities. 2406 DEBUG(dbgs() << "=> Profitable.\n"); 2407 // Recolor the live-range. 2408 Matrix->unassign(LI); 2409 Matrix->assign(LI, PhysReg); 2410 } 2411 // Push all copy-related live-ranges to keep reconciling the broken 2412 // hints. 2413 for (const HintInfo &HI : Info) { 2414 if (Visited.insert(HI.Reg).second) 2415 RecoloringCandidates.push_back(HI.Reg); 2416 } 2417 } while (!RecoloringCandidates.empty()); 2418 } 2419 2420 /// \brief Try to recolor broken hints. 2421 /// Broken hints may be repaired by recoloring when an evicted variable 2422 /// freed up a register for a larger live-range. 2423 /// Consider the following example: 2424 /// BB1: 2425 /// a = 2426 /// b = 2427 /// BB2: 2428 /// ... 2429 /// = b 2430 /// = a 2431 /// Let us assume b gets split: 2432 /// BB1: 2433 /// a = 2434 /// b = 2435 /// BB2: 2436 /// c = b 2437 /// ... 2438 /// d = c 2439 /// = d 2440 /// = a 2441 /// Because of how the allocation work, b, c, and d may be assigned different 2442 /// colors. Now, if a gets evicted later: 2443 /// BB1: 2444 /// a = 2445 /// st a, SpillSlot 2446 /// b = 2447 /// BB2: 2448 /// c = b 2449 /// ... 2450 /// d = c 2451 /// = d 2452 /// e = ld SpillSlot 2453 /// = e 2454 /// This is likely that we can assign the same register for b, c, and d, 2455 /// getting rid of 2 copies. 2456 void RAGreedy::tryHintsRecoloring() { 2457 for (LiveInterval *LI : SetOfBrokenHints) { 2458 assert(TargetRegisterInfo::isVirtualRegister(LI->reg) && 2459 "Recoloring is possible only for virtual registers"); 2460 // Some dead defs may be around (e.g., because of debug uses). 2461 // Ignore those. 2462 if (!VRM->hasPhys(LI->reg)) 2463 continue; 2464 tryHintRecoloring(*LI); 2465 } 2466 } 2467 2468 unsigned RAGreedy::selectOrSplitImpl(LiveInterval &VirtReg, 2469 SmallVectorImpl<unsigned> &NewVRegs, 2470 SmallVirtRegSet &FixedRegisters, 2471 unsigned Depth) { 2472 unsigned CostPerUseLimit = ~0u; 2473 // First try assigning a free register. 2474 AllocationOrder Order(VirtReg.reg, *VRM, RegClassInfo, Matrix); 2475 if (unsigned PhysReg = tryAssign(VirtReg, Order, NewVRegs)) { 2476 // When NewVRegs is not empty, we may have made decisions such as evicting 2477 // a virtual register, go with the earlier decisions and use the physical 2478 // register. 2479 if (CSRCost.getFrequency() && isUnusedCalleeSavedReg(PhysReg) && 2480 NewVRegs.empty()) { 2481 unsigned CSRReg = tryAssignCSRFirstTime(VirtReg, Order, PhysReg, 2482 CostPerUseLimit, NewVRegs); 2483 if (CSRReg || !NewVRegs.empty()) 2484 // Return now if we decide to use a CSR or create new vregs due to 2485 // pre-splitting. 2486 return CSRReg; 2487 } else 2488 return PhysReg; 2489 } 2490 2491 LiveRangeStage Stage = getStage(VirtReg); 2492 DEBUG(dbgs() << StageName[Stage] 2493 << " Cascade " << ExtraRegInfo[VirtReg.reg].Cascade << '\n'); 2494 2495 // Try to evict a less worthy live range, but only for ranges from the primary 2496 // queue. The RS_Split ranges already failed to do this, and they should not 2497 // get a second chance until they have been split. 2498 if (Stage != RS_Split) 2499 if (unsigned PhysReg = 2500 tryEvict(VirtReg, Order, NewVRegs, CostPerUseLimit)) { 2501 unsigned Hint = MRI->getSimpleHint(VirtReg.reg); 2502 // If VirtReg has a hint and that hint is broken record this 2503 // virtual register as a recoloring candidate for broken hint. 2504 // Indeed, since we evicted a variable in its neighborhood it is 2505 // likely we can at least partially recolor some of the 2506 // copy-related live-ranges. 2507 if (Hint && Hint != PhysReg) 2508 SetOfBrokenHints.insert(&VirtReg); 2509 return PhysReg; 2510 } 2511 2512 assert(NewVRegs.empty() && "Cannot append to existing NewVRegs"); 2513 2514 // The first time we see a live range, don't try to split or spill. 2515 // Wait until the second time, when all smaller ranges have been allocated. 2516 // This gives a better picture of the interference to split around. 2517 if (Stage < RS_Split) { 2518 setStage(VirtReg, RS_Split); 2519 DEBUG(dbgs() << "wait for second round\n"); 2520 NewVRegs.push_back(VirtReg.reg); 2521 return 0; 2522 } 2523 2524 // If we couldn't allocate a register from spilling, there is probably some 2525 // invalid inline assembly. The base class wil report it. 2526 if (Stage >= RS_Done || !VirtReg.isSpillable()) 2527 return tryLastChanceRecoloring(VirtReg, Order, NewVRegs, FixedRegisters, 2528 Depth); 2529 2530 // Try splitting VirtReg or interferences. 2531 unsigned PhysReg = trySplit(VirtReg, Order, NewVRegs); 2532 if (PhysReg || !NewVRegs.empty()) 2533 return PhysReg; 2534 2535 // Finally spill VirtReg itself. 2536 if (EnableDeferredSpilling && getStage(VirtReg) < RS_Memory) { 2537 // TODO: This is experimental and in particular, we do not model 2538 // the live range splitting done by spilling correctly. 2539 // We would need a deep integration with the spiller to do the 2540 // right thing here. Anyway, that is still good for early testing. 2541 setStage(VirtReg, RS_Memory); 2542 DEBUG(dbgs() << "Do as if this register is in memory\n"); 2543 NewVRegs.push_back(VirtReg.reg); 2544 } else { 2545 NamedRegionTimer T("Spiller", TimerGroupName, TimePassesIsEnabled); 2546 LiveRangeEdit LRE(&VirtReg, NewVRegs, *MF, *LIS, VRM, this); 2547 spiller().spill(LRE); 2548 setStage(NewVRegs.begin(), NewVRegs.end(), RS_Done); 2549 2550 if (VerifyEnabled) 2551 MF->verify(this, "After spilling"); 2552 } 2553 2554 // The live virtual register requesting allocation was spilled, so tell 2555 // the caller not to allocate anything during this round. 2556 return 0; 2557 } 2558 2559 bool RAGreedy::runOnMachineFunction(MachineFunction &mf) { 2560 DEBUG(dbgs() << "********** GREEDY REGISTER ALLOCATION **********\n" 2561 << "********** Function: " << mf.getName() << '\n'); 2562 2563 MF = &mf; 2564 TRI = MF->getSubtarget().getRegisterInfo(); 2565 TII = MF->getSubtarget().getInstrInfo(); 2566 RCI.runOnMachineFunction(mf); 2567 2568 EnableLocalReassign = EnableLocalReassignment || 2569 MF->getSubtarget().enableRALocalReassignment( 2570 MF->getTarget().getOptLevel()); 2571 2572 if (VerifyEnabled) 2573 MF->verify(this, "Before greedy register allocator"); 2574 2575 RegAllocBase::init(getAnalysis<VirtRegMap>(), 2576 getAnalysis<LiveIntervals>(), 2577 getAnalysis<LiveRegMatrix>()); 2578 Indexes = &getAnalysis<SlotIndexes>(); 2579 MBFI = &getAnalysis<MachineBlockFrequencyInfo>(); 2580 DomTree = &getAnalysis<MachineDominatorTree>(); 2581 SpillerInstance.reset(createInlineSpiller(*this, *MF, *VRM)); 2582 Loops = &getAnalysis<MachineLoopInfo>(); 2583 Bundles = &getAnalysis<EdgeBundles>(); 2584 SpillPlacer = &getAnalysis<SpillPlacement>(); 2585 DebugVars = &getAnalysis<LiveDebugVariables>(); 2586 2587 initializeCSRCost(); 2588 2589 calculateSpillWeightsAndHints(*LIS, mf, VRM, *Loops, *MBFI); 2590 2591 DEBUG(LIS->dump()); 2592 2593 SA.reset(new SplitAnalysis(*VRM, *LIS, *Loops)); 2594 SE.reset(new SplitEditor(*SA, *LIS, *VRM, *DomTree, *MBFI)); 2595 ExtraRegInfo.clear(); 2596 ExtraRegInfo.resize(MRI->getNumVirtRegs()); 2597 NextCascade = 1; 2598 IntfCache.init(MF, Matrix->getLiveUnions(), Indexes, LIS, TRI); 2599 GlobalCand.resize(32); // This will grow as needed. 2600 SetOfBrokenHints.clear(); 2601 2602 allocatePhysRegs(); 2603 tryHintsRecoloring(); 2604 releaseMemory(); 2605 return true; 2606 } 2607