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