SplitKit.cpp revision 521a453721aeefbb6783b6acc8ea36b3c18b4931
1//===---------- SplitKit.cpp - Toolkit for splitting live ranges ----------===// 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 contains the SplitAnalysis class as well as mutator functions for 11// live range splitting. 12// 13//===----------------------------------------------------------------------===// 14 15#define DEBUG_TYPE "regalloc" 16#include "SplitKit.h" 17#include "LiveRangeEdit.h" 18#include "VirtRegMap.h" 19#include "llvm/CodeGen/CalcSpillWeights.h" 20#include "llvm/CodeGen/LiveIntervalAnalysis.h" 21#include "llvm/CodeGen/MachineDominators.h" 22#include "llvm/CodeGen/MachineInstrBuilder.h" 23#include "llvm/CodeGen/MachineLoopInfo.h" 24#include "llvm/CodeGen/MachineRegisterInfo.h" 25#include "llvm/Support/CommandLine.h" 26#include "llvm/Support/Debug.h" 27#include "llvm/Support/raw_ostream.h" 28#include "llvm/Target/TargetInstrInfo.h" 29#include "llvm/Target/TargetMachine.h" 30 31using namespace llvm; 32 33static cl::opt<bool> 34AllowSplit("spiller-splits-edges", 35 cl::desc("Allow critical edge splitting during spilling")); 36 37//===----------------------------------------------------------------------===// 38// Split Analysis 39//===----------------------------------------------------------------------===// 40 41SplitAnalysis::SplitAnalysis(const MachineFunction &mf, 42 const LiveIntervals &lis, 43 const MachineLoopInfo &mli) 44 : mf_(mf), 45 lis_(lis), 46 loops_(mli), 47 tii_(*mf.getTarget().getInstrInfo()), 48 curli_(0) {} 49 50void SplitAnalysis::clear() { 51 usingInstrs_.clear(); 52 usingBlocks_.clear(); 53 usingLoops_.clear(); 54 curli_ = 0; 55} 56 57bool SplitAnalysis::canAnalyzeBranch(const MachineBasicBlock *MBB) { 58 MachineBasicBlock *T, *F; 59 SmallVector<MachineOperand, 4> Cond; 60 return !tii_.AnalyzeBranch(const_cast<MachineBasicBlock&>(*MBB), T, F, Cond); 61} 62 63/// analyzeUses - Count instructions, basic blocks, and loops using curli. 64void SplitAnalysis::analyzeUses() { 65 const MachineRegisterInfo &MRI = mf_.getRegInfo(); 66 for (MachineRegisterInfo::reg_iterator I = MRI.reg_begin(curli_->reg); 67 MachineInstr *MI = I.skipInstruction();) { 68 if (MI->isDebugValue() || !usingInstrs_.insert(MI)) 69 continue; 70 MachineBasicBlock *MBB = MI->getParent(); 71 if (usingBlocks_[MBB]++) 72 continue; 73 for (MachineLoop *Loop = loops_.getLoopFor(MBB); Loop; 74 Loop = Loop->getParentLoop()) 75 usingLoops_[Loop]++; 76 } 77 DEBUG(dbgs() << " counted " 78 << usingInstrs_.size() << " instrs, " 79 << usingBlocks_.size() << " blocks, " 80 << usingLoops_.size() << " loops.\n"); 81} 82 83void SplitAnalysis::print(const BlockPtrSet &B, raw_ostream &OS) const { 84 for (BlockPtrSet::const_iterator I = B.begin(), E = B.end(); I != E; ++I) { 85 unsigned count = usingBlocks_.lookup(*I); 86 OS << " BB#" << (*I)->getNumber(); 87 if (count) 88 OS << '(' << count << ')'; 89 } 90} 91 92// Get three sets of basic blocks surrounding a loop: Blocks inside the loop, 93// predecessor blocks, and exit blocks. 94void SplitAnalysis::getLoopBlocks(const MachineLoop *Loop, LoopBlocks &Blocks) { 95 Blocks.clear(); 96 97 // Blocks in the loop. 98 Blocks.Loop.insert(Loop->block_begin(), Loop->block_end()); 99 100 // Predecessor blocks. 101 const MachineBasicBlock *Header = Loop->getHeader(); 102 for (MachineBasicBlock::const_pred_iterator I = Header->pred_begin(), 103 E = Header->pred_end(); I != E; ++I) 104 if (!Blocks.Loop.count(*I)) 105 Blocks.Preds.insert(*I); 106 107 // Exit blocks. 108 for (MachineLoop::block_iterator I = Loop->block_begin(), 109 E = Loop->block_end(); I != E; ++I) { 110 const MachineBasicBlock *MBB = *I; 111 for (MachineBasicBlock::const_succ_iterator SI = MBB->succ_begin(), 112 SE = MBB->succ_end(); SI != SE; ++SI) 113 if (!Blocks.Loop.count(*SI)) 114 Blocks.Exits.insert(*SI); 115 } 116} 117 118void SplitAnalysis::print(const LoopBlocks &B, raw_ostream &OS) const { 119 OS << "Loop:"; 120 print(B.Loop, OS); 121 OS << ", preds:"; 122 print(B.Preds, OS); 123 OS << ", exits:"; 124 print(B.Exits, OS); 125} 126 127/// analyzeLoopPeripheralUse - Return an enum describing how curli_ is used in 128/// and around the Loop. 129SplitAnalysis::LoopPeripheralUse SplitAnalysis:: 130analyzeLoopPeripheralUse(const SplitAnalysis::LoopBlocks &Blocks) { 131 LoopPeripheralUse use = ContainedInLoop; 132 for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end(); 133 I != E; ++I) { 134 const MachineBasicBlock *MBB = I->first; 135 // Is this a peripheral block? 136 if (use < MultiPeripheral && 137 (Blocks.Preds.count(MBB) || Blocks.Exits.count(MBB))) { 138 if (I->second > 1) use = MultiPeripheral; 139 else use = SinglePeripheral; 140 continue; 141 } 142 // Is it a loop block? 143 if (Blocks.Loop.count(MBB)) 144 continue; 145 // It must be an unrelated block. 146 DEBUG(dbgs() << ", outside: BB#" << MBB->getNumber()); 147 return OutsideLoop; 148 } 149 return use; 150} 151 152/// getCriticalExits - It may be necessary to partially break critical edges 153/// leaving the loop if an exit block has predecessors from outside the loop 154/// periphery. 155void SplitAnalysis::getCriticalExits(const SplitAnalysis::LoopBlocks &Blocks, 156 BlockPtrSet &CriticalExits) { 157 CriticalExits.clear(); 158 159 // A critical exit block has curli live-in, and has a predecessor that is not 160 // in the loop nor a loop predecessor. For such an exit block, the edges 161 // carrying the new variable must be moved to a new pre-exit block. 162 for (BlockPtrSet::iterator I = Blocks.Exits.begin(), E = Blocks.Exits.end(); 163 I != E; ++I) { 164 const MachineBasicBlock *Exit = *I; 165 // A single-predecessor exit block is definitely not a critical edge. 166 if (Exit->pred_size() == 1) 167 continue; 168 // This exit may not have curli live in at all. No need to split. 169 if (!lis_.isLiveInToMBB(*curli_, Exit)) 170 continue; 171 // Does this exit block have a predecessor that is not a loop block or loop 172 // predecessor? 173 for (MachineBasicBlock::const_pred_iterator PI = Exit->pred_begin(), 174 PE = Exit->pred_end(); PI != PE; ++PI) { 175 const MachineBasicBlock *Pred = *PI; 176 if (Blocks.Loop.count(Pred) || Blocks.Preds.count(Pred)) 177 continue; 178 // This is a critical exit block, and we need to split the exit edge. 179 CriticalExits.insert(Exit); 180 break; 181 } 182 } 183} 184 185void SplitAnalysis::getCriticalPreds(const SplitAnalysis::LoopBlocks &Blocks, 186 BlockPtrSet &CriticalPreds) { 187 CriticalPreds.clear(); 188 189 // A critical predecessor block has curli live-out, and has a successor that 190 // has curli live-in and is not in the loop nor a loop exit block. For such a 191 // predecessor block, we must carry the value in both the 'inside' and 192 // 'outside' registers. 193 for (BlockPtrSet::iterator I = Blocks.Preds.begin(), E = Blocks.Preds.end(); 194 I != E; ++I) { 195 const MachineBasicBlock *Pred = *I; 196 // Definitely not a critical edge. 197 if (Pred->succ_size() == 1) 198 continue; 199 // This block may not have curli live out at all if there is a PHI. 200 if (!lis_.isLiveOutOfMBB(*curli_, Pred)) 201 continue; 202 // Does this block have a successor outside the loop? 203 for (MachineBasicBlock::const_pred_iterator SI = Pred->succ_begin(), 204 SE = Pred->succ_end(); SI != SE; ++SI) { 205 const MachineBasicBlock *Succ = *SI; 206 if (Blocks.Loop.count(Succ) || Blocks.Exits.count(Succ)) 207 continue; 208 if (!lis_.isLiveInToMBB(*curli_, Succ)) 209 continue; 210 // This is a critical predecessor block. 211 CriticalPreds.insert(Pred); 212 break; 213 } 214 } 215} 216 217/// canSplitCriticalExits - Return true if it is possible to insert new exit 218/// blocks before the blocks in CriticalExits. 219bool 220SplitAnalysis::canSplitCriticalExits(const SplitAnalysis::LoopBlocks &Blocks, 221 BlockPtrSet &CriticalExits) { 222 // If we don't allow critical edge splitting, require no critical exits. 223 if (!AllowSplit) 224 return CriticalExits.empty(); 225 226 for (BlockPtrSet::iterator I = CriticalExits.begin(), E = CriticalExits.end(); 227 I != E; ++I) { 228 const MachineBasicBlock *Succ = *I; 229 // We want to insert a new pre-exit MBB before Succ, and change all the 230 // in-loop blocks to branch to the pre-exit instead of Succ. 231 // Check that all the in-loop predecessors can be changed. 232 for (MachineBasicBlock::const_pred_iterator PI = Succ->pred_begin(), 233 PE = Succ->pred_end(); PI != PE; ++PI) { 234 const MachineBasicBlock *Pred = *PI; 235 // The external predecessors won't be altered. 236 if (!Blocks.Loop.count(Pred) && !Blocks.Preds.count(Pred)) 237 continue; 238 if (!canAnalyzeBranch(Pred)) 239 return false; 240 } 241 242 // If Succ's layout predecessor falls through, that too must be analyzable. 243 // We need to insert the pre-exit block in the gap. 244 MachineFunction::const_iterator MFI = Succ; 245 if (MFI == mf_.begin()) 246 continue; 247 if (!canAnalyzeBranch(--MFI)) 248 return false; 249 } 250 // No problems found. 251 return true; 252} 253 254void SplitAnalysis::analyze(const LiveInterval *li) { 255 clear(); 256 curli_ = li; 257 analyzeUses(); 258} 259 260void SplitAnalysis::getSplitLoops(LoopPtrSet &Loops) { 261 assert(curli_ && "Call analyze() before getSplitLoops"); 262 if (usingLoops_.empty()) 263 return; 264 265 LoopBlocks Blocks; 266 BlockPtrSet CriticalExits; 267 268 // We split around loops where curli is used outside the periphery. 269 for (LoopCountMap::const_iterator I = usingLoops_.begin(), 270 E = usingLoops_.end(); I != E; ++I) { 271 const MachineLoop *Loop = I->first; 272 getLoopBlocks(Loop, Blocks); 273 DEBUG({ dbgs() << " "; print(Blocks, dbgs()); }); 274 275 switch(analyzeLoopPeripheralUse(Blocks)) { 276 case OutsideLoop: 277 break; 278 case MultiPeripheral: 279 // FIXME: We could split a live range with multiple uses in a peripheral 280 // block and still make progress. However, it is possible that splitting 281 // another live range will insert copies into a peripheral block, and 282 // there is a small chance we can enter an infinite loop, inserting copies 283 // forever. 284 // For safety, stick to splitting live ranges with uses outside the 285 // periphery. 286 DEBUG(dbgs() << ": multiple peripheral uses"); 287 break; 288 case ContainedInLoop: 289 DEBUG(dbgs() << ": fully contained\n"); 290 continue; 291 case SinglePeripheral: 292 DEBUG(dbgs() << ": single peripheral use\n"); 293 continue; 294 } 295 // Will it be possible to split around this loop? 296 getCriticalExits(Blocks, CriticalExits); 297 DEBUG(dbgs() << ": " << CriticalExits.size() << " critical exits\n"); 298 if (!canSplitCriticalExits(Blocks, CriticalExits)) 299 continue; 300 // This is a possible split. 301 Loops.insert(Loop); 302 } 303 304 DEBUG(dbgs() << " getSplitLoops found " << Loops.size() 305 << " candidate loops.\n"); 306} 307 308const MachineLoop *SplitAnalysis::getBestSplitLoop() { 309 LoopPtrSet Loops; 310 getSplitLoops(Loops); 311 if (Loops.empty()) 312 return 0; 313 314 // Pick the earliest loop. 315 // FIXME: Are there other heuristics to consider? 316 const MachineLoop *Best = 0; 317 SlotIndex BestIdx; 318 for (LoopPtrSet::const_iterator I = Loops.begin(), E = Loops.end(); I != E; 319 ++I) { 320 SlotIndex Idx = lis_.getMBBStartIdx((*I)->getHeader()); 321 if (!Best || Idx < BestIdx) 322 Best = *I, BestIdx = Idx; 323 } 324 DEBUG(dbgs() << " getBestSplitLoop found " << *Best); 325 return Best; 326} 327 328//===----------------------------------------------------------------------===// 329// LiveIntervalMap 330//===----------------------------------------------------------------------===// 331 332// Work around the fact that the std::pair constructors are broken for pointer 333// pairs in some implementations. makeVV(x, 0) works. 334static inline std::pair<const VNInfo*, VNInfo*> 335makeVV(const VNInfo *a, VNInfo *b) { 336 return std::make_pair(a, b); 337} 338 339void LiveIntervalMap::reset(LiveInterval *li) { 340 li_ = li; 341 valueMap_.clear(); 342 liveOutCache_.clear(); 343} 344 345bool LiveIntervalMap::isComplexMapped(const VNInfo *ParentVNI) const { 346 ValueMap::const_iterator i = valueMap_.find(ParentVNI); 347 return i != valueMap_.end() && i->second == 0; 348} 349 350// defValue - Introduce a li_ def for ParentVNI that could be later than 351// ParentVNI->def. 352VNInfo *LiveIntervalMap::defValue(const VNInfo *ParentVNI, SlotIndex Idx) { 353 assert(li_ && "call reset first"); 354 assert(ParentVNI && "Mapping NULL value"); 355 assert(Idx.isValid() && "Invalid SlotIndex"); 356 assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI"); 357 358 // Create a new value. 359 VNInfo *VNI = li_->getNextValue(Idx, 0, lis_.getVNInfoAllocator()); 360 361 // Preserve the PHIDef bit. 362 if (ParentVNI->isPHIDef() && Idx == ParentVNI->def) 363 VNI->setIsPHIDef(true); 364 365 // Use insert for lookup, so we can add missing values with a second lookup. 366 std::pair<ValueMap::iterator,bool> InsP = 367 valueMap_.insert(makeVV(ParentVNI, Idx == ParentVNI->def ? VNI : 0)); 368 369 // This is now a complex def. Mark with a NULL in valueMap. 370 if (!InsP.second) 371 InsP.first->second = 0; 372 373 return VNI; 374} 375 376 377// mapValue - Find the mapped value for ParentVNI at Idx. 378// Potentially create phi-def values. 379VNInfo *LiveIntervalMap::mapValue(const VNInfo *ParentVNI, SlotIndex Idx, 380 bool *simple) { 381 assert(li_ && "call reset first"); 382 assert(ParentVNI && "Mapping NULL value"); 383 assert(Idx.isValid() && "Invalid SlotIndex"); 384 assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI"); 385 386 // Use insert for lookup, so we can add missing values with a second lookup. 387 std::pair<ValueMap::iterator,bool> InsP = 388 valueMap_.insert(makeVV(ParentVNI, 0)); 389 390 // This was an unknown value. Create a simple mapping. 391 if (InsP.second) { 392 if (simple) *simple = true; 393 return InsP.first->second = li_->createValueCopy(ParentVNI, 394 lis_.getVNInfoAllocator()); 395 } 396 397 // This was a simple mapped value. 398 if (InsP.first->second) { 399 if (simple) *simple = true; 400 return InsP.first->second; 401 } 402 403 // This is a complex mapped value. There may be multiple defs, and we may need 404 // to create phi-defs. 405 if (simple) *simple = false; 406 MachineBasicBlock *IdxMBB = lis_.getMBBFromIndex(Idx); 407 assert(IdxMBB && "No MBB at Idx"); 408 409 // Is there a def in the same MBB we can extend? 410 if (VNInfo *VNI = extendTo(IdxMBB, Idx)) 411 return VNI; 412 413 // Now for the fun part. We know that ParentVNI potentially has multiple defs, 414 // and we may need to create even more phi-defs to preserve VNInfo SSA form. 415 // Perform a search for all predecessor blocks where we know the dominating 416 // VNInfo. Insert phi-def VNInfos along the path back to IdxMBB. 417 DEBUG(dbgs() << "\n Reaching defs for BB#" << IdxMBB->getNumber() 418 << " at " << Idx << " in " << *li_ << '\n'); 419 420 // Blocks where li_ should be live-in. 421 SmallVector<MachineDomTreeNode*, 16> LiveIn; 422 LiveIn.push_back(mdt_[IdxMBB]); 423 424 // Using liveOutCache_ as a visited set, perform a BFS for all reaching defs. 425 for (unsigned i = 0; i != LiveIn.size(); ++i) { 426 MachineBasicBlock *MBB = LiveIn[i]->getBlock(); 427 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 428 PE = MBB->pred_end(); PI != PE; ++PI) { 429 MachineBasicBlock *Pred = *PI; 430 // Is this a known live-out block? 431 std::pair<LiveOutMap::iterator,bool> LOIP = 432 liveOutCache_.insert(std::make_pair(Pred, LiveOutPair())); 433 // Yes, we have been here before. 434 if (!LOIP.second) { 435 DEBUG(if (VNInfo *VNI = LOIP.first->second.first) 436 dbgs() << " known valno #" << VNI->id 437 << " at BB#" << Pred->getNumber() << '\n'); 438 continue; 439 } 440 441 // Does Pred provide a live-out value? 442 SlotIndex Last = lis_.getMBBEndIdx(Pred).getPrevSlot(); 443 if (VNInfo *VNI = extendTo(Pred, Last)) { 444 MachineBasicBlock *DefMBB = lis_.getMBBFromIndex(VNI->def); 445 DEBUG(dbgs() << " found valno #" << VNI->id 446 << " from BB#" << DefMBB->getNumber() 447 << " at BB#" << Pred->getNumber() << '\n'); 448 LiveOutPair &LOP = LOIP.first->second; 449 LOP.first = VNI; 450 LOP.second = mdt_[DefMBB]; 451 continue; 452 } 453 // No, we need a live-in value for Pred as well 454 if (Pred != IdxMBB) 455 LiveIn.push_back(mdt_[Pred]); 456 } 457 } 458 459 // We may need to add phi-def values to preserve the SSA form. 460 // This is essentially the same iterative algorithm that SSAUpdater uses, 461 // except we already have a dominator tree, so we don't have to recompute it. 462 VNInfo *IdxVNI = 0; 463 unsigned Changes; 464 do { 465 Changes = 0; 466 DEBUG(dbgs() << " Iterating over " << LiveIn.size() << " blocks.\n"); 467 // Propagate live-out values down the dominator tree, inserting phi-defs when 468 // necessary. Since LiveIn was created by a BFS, going backwards makes it more 469 // likely for us to visit immediate dominators before their children. 470 for (unsigned i = LiveIn.size(); i; --i) { 471 MachineDomTreeNode *Node = LiveIn[i-1]; 472 MachineBasicBlock *MBB = Node->getBlock(); 473 MachineDomTreeNode *IDom = Node->getIDom(); 474 LiveOutPair IDomValue; 475 // We need a live-in value to a block with no immediate dominator? 476 // This is probably an unreachable block that has survived somehow. 477 bool needPHI = !IDom; 478 479 // Get the IDom live-out value. 480 if (!needPHI) { 481 LiveOutMap::iterator I = liveOutCache_.find(IDom->getBlock()); 482 if (I != liveOutCache_.end()) 483 IDomValue = I->second; 484 else 485 // If IDom is outside our set of live-out blocks, there must be new 486 // defs, and we need a phi-def here. 487 needPHI = true; 488 } 489 490 // IDom dominates all of our predecessors, but it may not be the immediate 491 // dominator. Check if any of them have live-out values that are properly 492 // dominated by IDom. If so, we need a phi-def here. 493 if (!needPHI) { 494 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 495 PE = MBB->pred_end(); PI != PE; ++PI) { 496 LiveOutPair Value = liveOutCache_[*PI]; 497 if (!Value.first || Value.first == IDomValue.first) 498 continue; 499 // This predecessor is carrying something other than IDomValue. 500 // It could be because IDomValue hasn't propagated yet, or it could be 501 // because MBB is in the dominance frontier of that value. 502 if (mdt_.dominates(IDom, Value.second)) { 503 needPHI = true; 504 break; 505 } 506 } 507 } 508 509 // Create a phi-def if required. 510 if (needPHI) { 511 ++Changes; 512 SlotIndex Start = lis_.getMBBStartIdx(MBB); 513 VNInfo *VNI = li_->getNextValue(Start, 0, lis_.getVNInfoAllocator()); 514 VNI->setIsPHIDef(true); 515 DEBUG(dbgs() << " - BB#" << MBB->getNumber() 516 << " phi-def #" << VNI->id << " at " << Start << '\n'); 517 // We no longer need li_ to be live-in. 518 LiveIn.erase(LiveIn.begin()+(i-1)); 519 // Blocks in LiveIn are either IdxMBB, or have a value live-through. 520 if (MBB == IdxMBB) 521 IdxVNI = VNI; 522 // Check if we need to update live-out info. 523 LiveOutMap::iterator I = liveOutCache_.find(MBB); 524 if (I == liveOutCache_.end() || I->second.second == Node) { 525 // We already have a live-out defined in MBB, so this must be IdxMBB. 526 assert(MBB == IdxMBB && "Adding phi-def to known live-out"); 527 li_->addRange(LiveRange(Start, Idx.getNextSlot(), VNI)); 528 } else { 529 // This phi-def is also live-out, so color the whole block. 530 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI)); 531 I->second = LiveOutPair(VNI, Node); 532 } 533 } else if (IDomValue.first) { 534 // No phi-def here. Remember incoming value for IdxMBB. 535 if (MBB == IdxMBB) 536 IdxVNI = IDomValue.first; 537 // Propagate IDomValue if needed: 538 // MBB is live-out and doesn't define its own value. 539 LiveOutMap::iterator I = liveOutCache_.find(MBB); 540 if (I != liveOutCache_.end() && I->second.second != Node && 541 I->second.first != IDomValue.first) { 542 ++Changes; 543 I->second = IDomValue; 544 DEBUG(dbgs() << " - BB#" << MBB->getNumber() 545 << " idom valno #" << IDomValue.first->id 546 << " from BB#" << IDom->getBlock()->getNumber() << '\n'); 547 } 548 } 549 } 550 DEBUG(dbgs() << " - made " << Changes << " changes.\n"); 551 } while (Changes); 552 553 assert(IdxVNI && "Didn't find value for Idx"); 554 555#ifndef NDEBUG 556 // Check the liveOutCache_ invariants. 557 for (LiveOutMap::iterator I = liveOutCache_.begin(), E = liveOutCache_.end(); 558 I != E; ++I) { 559 assert(I->first && "Null MBB entry in cache"); 560 assert(I->second.first && "Null VNInfo in cache"); 561 assert(I->second.second && "Null DomTreeNode in cache"); 562 if (I->second.second->getBlock() == I->first) 563 continue; 564 for (MachineBasicBlock::pred_iterator PI = I->first->pred_begin(), 565 PE = I->first->pred_end(); PI != PE; ++PI) 566 assert(liveOutCache_.lookup(*PI) == I->second && "Bad invariant"); 567 } 568#endif 569 570 // Since we went through the trouble of a full BFS visiting all reaching defs, 571 // the values in LiveIn are now accurate. No more phi-defs are needed 572 // for these blocks, so we can color the live ranges. 573 // This makes the next mapValue call much faster. 574 for (unsigned i = 0, e = LiveIn.size(); i != e; ++i) { 575 MachineBasicBlock *MBB = LiveIn[i]->getBlock(); 576 SlotIndex Start = lis_.getMBBStartIdx(MBB); 577 if (MBB == IdxMBB) { 578 li_->addRange(LiveRange(Start, Idx.getNextSlot(), IdxVNI)); 579 continue; 580 } 581 // Anything in LiveIn other than IdxMBB is live-through. 582 VNInfo *VNI = liveOutCache_.lookup(MBB).first; 583 assert(VNI && "Missing block value"); 584 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI)); 585 } 586 587 return IdxVNI; 588} 589 590// extendTo - Find the last li_ value defined in MBB at or before Idx. The 591// parentli_ is assumed to be live at Idx. Extend the live range to Idx. 592// Return the found VNInfo, or NULL. 593VNInfo *LiveIntervalMap::extendTo(const MachineBasicBlock *MBB, SlotIndex Idx) { 594 assert(li_ && "call reset first"); 595 LiveInterval::iterator I = std::upper_bound(li_->begin(), li_->end(), Idx); 596 if (I == li_->begin()) 597 return 0; 598 --I; 599 if (I->end <= lis_.getMBBStartIdx(MBB)) 600 return 0; 601 if (I->end <= Idx) 602 I->end = Idx.getNextSlot(); 603 return I->valno; 604} 605 606// addSimpleRange - Add a simple range from parentli_ to li_. 607// ParentVNI must be live in the [Start;End) interval. 608void LiveIntervalMap::addSimpleRange(SlotIndex Start, SlotIndex End, 609 const VNInfo *ParentVNI) { 610 assert(li_ && "call reset first"); 611 bool simple; 612 VNInfo *VNI = mapValue(ParentVNI, Start, &simple); 613 // A simple mapping is easy. 614 if (simple) { 615 li_->addRange(LiveRange(Start, End, VNI)); 616 return; 617 } 618 619 // ParentVNI is a complex value. We must map per MBB. 620 MachineFunction::iterator MBB = lis_.getMBBFromIndex(Start); 621 MachineFunction::iterator MBBE = lis_.getMBBFromIndex(End.getPrevSlot()); 622 623 if (MBB == MBBE) { 624 li_->addRange(LiveRange(Start, End, VNI)); 625 return; 626 } 627 628 // First block. 629 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI)); 630 631 // Run sequence of full blocks. 632 for (++MBB; MBB != MBBE; ++MBB) { 633 Start = lis_.getMBBStartIdx(MBB); 634 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), 635 mapValue(ParentVNI, Start))); 636 } 637 638 // Final block. 639 Start = lis_.getMBBStartIdx(MBB); 640 if (Start != End) 641 li_->addRange(LiveRange(Start, End, mapValue(ParentVNI, Start))); 642} 643 644/// addRange - Add live ranges to li_ where [Start;End) intersects parentli_. 645/// All needed values whose def is not inside [Start;End) must be defined 646/// beforehand so mapValue will work. 647void LiveIntervalMap::addRange(SlotIndex Start, SlotIndex End) { 648 assert(li_ && "call reset first"); 649 LiveInterval::const_iterator B = parentli_.begin(), E = parentli_.end(); 650 LiveInterval::const_iterator I = std::lower_bound(B, E, Start); 651 652 // Check if --I begins before Start and overlaps. 653 if (I != B) { 654 --I; 655 if (I->end > Start) 656 addSimpleRange(Start, std::min(End, I->end), I->valno); 657 ++I; 658 } 659 660 // The remaining ranges begin after Start. 661 for (;I != E && I->start < End; ++I) 662 addSimpleRange(I->start, std::min(End, I->end), I->valno); 663} 664 665 666//===----------------------------------------------------------------------===// 667// Split Editor 668//===----------------------------------------------------------------------===// 669 670/// Create a new SplitEditor for editing the LiveInterval analyzed by SA. 671SplitEditor::SplitEditor(SplitAnalysis &sa, 672 LiveIntervals &lis, 673 VirtRegMap &vrm, 674 MachineDominatorTree &mdt, 675 LiveRangeEdit &edit) 676 : sa_(sa), lis_(lis), vrm_(vrm), 677 mri_(vrm.getMachineFunction().getRegInfo()), 678 tii_(*vrm.getMachineFunction().getTarget().getInstrInfo()), 679 tri_(*vrm.getMachineFunction().getTarget().getRegisterInfo()), 680 edit_(edit), 681 dupli_(lis_, mdt, edit.getParent()), 682 openli_(lis_, mdt, edit.getParent()) 683{ 684 // We don't need an AliasAnalysis since we will only be performing 685 // cheap-as-a-copy remats anyway. 686 edit_.anyRematerializable(lis_, tii_, 0); 687} 688 689bool SplitEditor::intervalsLiveAt(SlotIndex Idx) const { 690 for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I) 691 if (*I != dupli_.getLI() && (*I)->liveAt(Idx)) 692 return true; 693 return false; 694} 695 696VNInfo *SplitEditor::defFromParent(LiveIntervalMap &Reg, 697 VNInfo *ParentVNI, 698 SlotIndex UseIdx, 699 MachineBasicBlock &MBB, 700 MachineBasicBlock::iterator I) { 701 VNInfo *VNI = 0; 702 MachineInstr *CopyMI = 0; 703 SlotIndex Def; 704 705 // Attempt cheap-as-a-copy rematerialization. 706 LiveRangeEdit::Remat RM(ParentVNI); 707 if (edit_.canRematerializeAt(RM, UseIdx, true, lis_)) { 708 Def = edit_.rematerializeAt(MBB, I, Reg.getLI()->reg, RM, 709 lis_, tii_, tri_); 710 } else { 711 // Can't remat, just insert a copy from parent. 712 CopyMI = BuildMI(MBB, I, DebugLoc(), tii_.get(TargetOpcode::COPY), 713 Reg.getLI()->reg).addReg(edit_.getReg()); 714 Def = lis_.InsertMachineInstrInMaps(CopyMI).getDefIndex(); 715 } 716 717 // Define the value in Reg. 718 VNI = Reg.defValue(ParentVNI, Def); 719 VNI->setCopy(CopyMI); 720 721 // Add minimal liveness for the new value. 722 if (UseIdx < Def) 723 UseIdx = Def; 724 Reg.getLI()->addRange(LiveRange(Def, UseIdx.getNextSlot(), VNI)); 725 return VNI; 726} 727 728/// Create a new virtual register and live interval. 729void SplitEditor::openIntv() { 730 assert(!openli_.getLI() && "Previous LI not closed before openIntv"); 731 if (!dupli_.getLI()) 732 dupli_.reset(&edit_.create(mri_, lis_, vrm_)); 733 734 openli_.reset(&edit_.create(mri_, lis_, vrm_)); 735} 736 737/// enterIntvBefore - Enter openli before the instruction at Idx. If curli is 738/// not live before Idx, a COPY is not inserted. 739void SplitEditor::enterIntvBefore(SlotIndex Idx) { 740 assert(openli_.getLI() && "openIntv not called before enterIntvBefore"); 741 Idx = Idx.getUseIndex(); 742 DEBUG(dbgs() << " enterIntvBefore " << Idx); 743 VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(Idx); 744 if (!ParentVNI) { 745 DEBUG(dbgs() << ": not live\n"); 746 return; 747 } 748 DEBUG(dbgs() << ": valno " << ParentVNI->id); 749 truncatedValues.insert(ParentVNI); 750 MachineInstr *MI = lis_.getInstructionFromIndex(Idx); 751 assert(MI && "enterIntvBefore called with invalid index"); 752 753 defFromParent(openli_, ParentVNI, Idx, *MI->getParent(), MI); 754 755 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n'); 756} 757 758/// enterIntvAtEnd - Enter openli at the end of MBB. 759void SplitEditor::enterIntvAtEnd(MachineBasicBlock &MBB) { 760 assert(openli_.getLI() && "openIntv not called before enterIntvAtEnd"); 761 SlotIndex End = lis_.getMBBEndIdx(&MBB).getPrevSlot(); 762 DEBUG(dbgs() << " enterIntvAtEnd BB#" << MBB.getNumber() << ", " << End); 763 VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(End); 764 if (!ParentVNI) { 765 DEBUG(dbgs() << ": not live\n"); 766 return; 767 } 768 DEBUG(dbgs() << ": valno " << ParentVNI->id); 769 truncatedValues.insert(ParentVNI); 770 defFromParent(openli_, ParentVNI, End, MBB, MBB.getFirstTerminator()); 771 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n'); 772} 773 774/// useIntv - indicate that all instructions in MBB should use openli. 775void SplitEditor::useIntv(const MachineBasicBlock &MBB) { 776 useIntv(lis_.getMBBStartIdx(&MBB), lis_.getMBBEndIdx(&MBB)); 777} 778 779void SplitEditor::useIntv(SlotIndex Start, SlotIndex End) { 780 assert(openli_.getLI() && "openIntv not called before useIntv"); 781 openli_.addRange(Start, End); 782 DEBUG(dbgs() << " use [" << Start << ';' << End << "): " 783 << *openli_.getLI() << '\n'); 784} 785 786/// leaveIntvAfter - Leave openli after the instruction at Idx. 787void SplitEditor::leaveIntvAfter(SlotIndex Idx) { 788 assert(openli_.getLI() && "openIntv not called before leaveIntvAfter"); 789 DEBUG(dbgs() << " leaveIntvAfter " << Idx); 790 791 // The interval must be live beyond the instruction at Idx. 792 Idx = Idx.getBoundaryIndex(); 793 VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(Idx); 794 if (!ParentVNI) { 795 DEBUG(dbgs() << ": not live\n"); 796 return; 797 } 798 DEBUG(dbgs() << ": valno " << ParentVNI->id); 799 800 MachineBasicBlock::iterator MII = lis_.getInstructionFromIndex(Idx); 801 VNInfo *VNI = defFromParent(dupli_, ParentVNI, Idx, 802 *MII->getParent(), llvm::next(MII)); 803 804 // Make sure that openli is properly extended from Idx to the new copy. 805 // FIXME: This shouldn't be necessary for remats. 806 openli_.addSimpleRange(Idx, VNI->def, ParentVNI); 807 808 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n'); 809} 810 811/// leaveIntvAtTop - Leave the interval at the top of MBB. 812/// Currently, only one value can leave the interval. 813void SplitEditor::leaveIntvAtTop(MachineBasicBlock &MBB) { 814 assert(openli_.getLI() && "openIntv not called before leaveIntvAtTop"); 815 SlotIndex Start = lis_.getMBBStartIdx(&MBB); 816 DEBUG(dbgs() << " leaveIntvAtTop BB#" << MBB.getNumber() << ", " << Start); 817 818 VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(Start); 819 if (!ParentVNI) { 820 DEBUG(dbgs() << ": not live\n"); 821 return; 822 } 823 824 VNInfo *VNI = defFromParent(dupli_, ParentVNI, Start, MBB, 825 MBB.SkipPHIsAndLabels(MBB.begin())); 826 827 // Finally we must make sure that openli is properly extended from Start to 828 // the new copy. 829 openli_.addSimpleRange(Start, VNI->def, ParentVNI); 830 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n'); 831} 832 833/// closeIntv - Indicate that we are done editing the currently open 834/// LiveInterval, and ranges can be trimmed. 835void SplitEditor::closeIntv() { 836 assert(openli_.getLI() && "openIntv not called before closeIntv"); 837 838 DEBUG(dbgs() << " closeIntv cleaning up\n"); 839 DEBUG(dbgs() << " open " << *openli_.getLI() << '\n'); 840 openli_.reset(0); 841} 842 843/// rewrite - Rewrite all uses of reg to use the new registers. 844void SplitEditor::rewrite(unsigned reg) { 845 for (MachineRegisterInfo::reg_iterator RI = mri_.reg_begin(reg), 846 RE = mri_.reg_end(); RI != RE;) { 847 MachineOperand &MO = RI.getOperand(); 848 unsigned OpNum = RI.getOperandNo(); 849 MachineInstr *MI = MO.getParent(); 850 ++RI; 851 if (MI->isDebugValue()) { 852 DEBUG(dbgs() << "Zapping " << *MI); 853 // FIXME: We can do much better with debug values. 854 MO.setReg(0); 855 continue; 856 } 857 SlotIndex Idx = lis_.getInstructionIndex(MI); 858 Idx = MO.isUse() ? Idx.getUseIndex() : Idx.getDefIndex(); 859 LiveInterval *LI = 0; 860 for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; 861 ++I) { 862 LiveInterval *testli = *I; 863 if (testli->liveAt(Idx)) { 864 LI = testli; 865 break; 866 } 867 } 868 DEBUG(dbgs() << " rewr BB#" << MI->getParent()->getNumber() << '\t'<< Idx); 869 assert(LI && "No register was live at use"); 870 MO.setReg(LI->reg); 871 if (MO.isUse() && !MI->isRegTiedToDefOperand(OpNum)) 872 MO.setIsKill(LI->killedAt(Idx.getDefIndex())); 873 DEBUG(dbgs() << '\t' << *MI); 874 } 875} 876 877void 878SplitEditor::addTruncSimpleRange(SlotIndex Start, SlotIndex End, VNInfo *VNI) { 879 // Build vector of iterator pairs from the intervals. 880 typedef std::pair<LiveInterval::const_iterator, 881 LiveInterval::const_iterator> IIPair; 882 SmallVector<IIPair, 8> Iters; 883 for (LiveRangeEdit::iterator LI = edit_.begin(), LE = edit_.end(); LI != LE; 884 ++LI) { 885 if (*LI == dupli_.getLI()) 886 continue; 887 LiveInterval::const_iterator I = (*LI)->find(Start); 888 LiveInterval::const_iterator E = (*LI)->end(); 889 if (I != E) 890 Iters.push_back(std::make_pair(I, E)); 891 } 892 893 SlotIndex sidx = Start; 894 // Break [Start;End) into segments that don't overlap any intervals. 895 for (;;) { 896 SlotIndex next = sidx, eidx = End; 897 // Find overlapping intervals. 898 for (unsigned i = 0; i != Iters.size() && sidx < eidx; ++i) { 899 LiveInterval::const_iterator I = Iters[i].first; 900 // Interval I is overlapping [sidx;eidx). Trim sidx. 901 if (I->start <= sidx) { 902 sidx = I->end; 903 // Move to the next run, remove iters when all are consumed. 904 I = ++Iters[i].first; 905 if (I == Iters[i].second) { 906 Iters.erase(Iters.begin() + i); 907 --i; 908 continue; 909 } 910 } 911 // Trim eidx too if needed. 912 if (I->start >= eidx) 913 continue; 914 eidx = I->start; 915 next = I->end; 916 } 917 // Now, [sidx;eidx) doesn't overlap anything in intervals_. 918 if (sidx < eidx) 919 dupli_.addSimpleRange(sidx, eidx, VNI); 920 // If the interval end was truncated, we can try again from next. 921 if (next <= sidx) 922 break; 923 sidx = next; 924 } 925} 926 927void SplitEditor::computeRemainder() { 928 // First we need to fill in the live ranges in dupli. 929 // If values were redefined, we need a full recoloring with SSA update. 930 // If values were truncated, we only need to truncate the ranges. 931 // If values were partially rematted, we should shrink to uses. 932 // If values were fully rematted, they should be omitted. 933 // FIXME: If a single value is redefined, just move the def and truncate. 934 LiveInterval &parent = edit_.getParent(); 935 936 // Values that are fully contained in the split intervals. 937 SmallPtrSet<const VNInfo*, 8> deadValues; 938 // Map all curli values that should have live defs in dupli. 939 for (LiveInterval::const_vni_iterator I = parent.vni_begin(), 940 E = parent.vni_end(); I != E; ++I) { 941 const VNInfo *VNI = *I; 942 // Don't transfer unused values to the new intervals. 943 if (VNI->isUnused()) 944 continue; 945 // Original def is contained in the split intervals. 946 if (intervalsLiveAt(VNI->def)) { 947 // Did this value escape? 948 if (dupli_.isMapped(VNI)) 949 truncatedValues.insert(VNI); 950 else 951 deadValues.insert(VNI); 952 continue; 953 } 954 // Add minimal live range at the definition. 955 VNInfo *DVNI = dupli_.defValue(VNI, VNI->def); 956 dupli_.getLI()->addRange(LiveRange(VNI->def, VNI->def.getNextSlot(), DVNI)); 957 } 958 959 // Add all ranges to dupli. 960 for (LiveInterval::const_iterator I = parent.begin(), E = parent.end(); 961 I != E; ++I) { 962 const LiveRange &LR = *I; 963 if (truncatedValues.count(LR.valno)) { 964 // recolor after removing intervals_. 965 addTruncSimpleRange(LR.start, LR.end, LR.valno); 966 } else if (!deadValues.count(LR.valno)) { 967 // recolor without truncation. 968 dupli_.addSimpleRange(LR.start, LR.end, LR.valno); 969 } 970 } 971 972 // Extend dupli_ to be live out of any critical loop predecessors. 973 // This means we have multiple registers live out of those blocks. 974 // The alternative would be to split the critical edges. 975 if (criticalPreds_.empty()) 976 return; 977 for (SplitAnalysis::BlockPtrSet::iterator I = criticalPreds_.begin(), 978 E = criticalPreds_.end(); I != E; ++I) 979 dupli_.extendTo(*I, lis_.getMBBEndIdx(*I).getPrevSlot()); 980 criticalPreds_.clear(); 981} 982 983void SplitEditor::finish() { 984 assert(!openli_.getLI() && "Previous LI not closed before rewrite"); 985 assert(dupli_.getLI() && "No dupli for rewrite. Noop spilt?"); 986 987 // Complete dupli liveness. 988 computeRemainder(); 989 990 // Get rid of unused values and set phi-kill flags. 991 for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I) 992 (*I)->RenumberValues(lis_); 993 994 // Rewrite instructions. 995 rewrite(edit_.getReg()); 996 997 // Now check if any registers were separated into multiple components. 998 ConnectedVNInfoEqClasses ConEQ(lis_); 999 for (unsigned i = 0, e = edit_.size(); i != e; ++i) { 1000 // Don't use iterators, they are invalidated by create() below. 1001 LiveInterval *li = edit_.get(i); 1002 unsigned NumComp = ConEQ.Classify(li); 1003 if (NumComp <= 1) 1004 continue; 1005 DEBUG(dbgs() << " " << NumComp << " components: " << *li << '\n'); 1006 SmallVector<LiveInterval*, 8> dups; 1007 dups.push_back(li); 1008 for (unsigned i = 1; i != NumComp; ++i) 1009 dups.push_back(&edit_.create(mri_, lis_, vrm_)); 1010 ConEQ.Distribute(&dups[0]); 1011 // Rewrite uses to the new regs. 1012 rewrite(li->reg); 1013 } 1014 1015 // Calculate spill weight and allocation hints for new intervals. 1016 VirtRegAuxInfo vrai(vrm_.getMachineFunction(), lis_, sa_.loops_); 1017 for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I){ 1018 LiveInterval &li = **I; 1019 vrai.CalculateRegClass(li.reg); 1020 vrai.CalculateWeightAndHint(li); 1021 DEBUG(dbgs() << " new interval " << mri_.getRegClass(li.reg)->getName() 1022 << ":" << li << '\n'); 1023 } 1024} 1025 1026 1027//===----------------------------------------------------------------------===// 1028// Loop Splitting 1029//===----------------------------------------------------------------------===// 1030 1031void SplitEditor::splitAroundLoop(const MachineLoop *Loop) { 1032 SplitAnalysis::LoopBlocks Blocks; 1033 sa_.getLoopBlocks(Loop, Blocks); 1034 1035 DEBUG({ 1036 dbgs() << " splitAround"; sa_.print(Blocks, dbgs()); dbgs() << '\n'; 1037 }); 1038 1039 // Break critical edges as needed. 1040 SplitAnalysis::BlockPtrSet CriticalExits; 1041 sa_.getCriticalExits(Blocks, CriticalExits); 1042 assert(CriticalExits.empty() && "Cannot break critical exits yet"); 1043 1044 // Get critical predecessors so computeRemainder can deal with them. 1045 sa_.getCriticalPreds(Blocks, criticalPreds_); 1046 1047 // Create new live interval for the loop. 1048 openIntv(); 1049 1050 // Insert copies in the predecessors. 1051 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Preds.begin(), 1052 E = Blocks.Preds.end(); I != E; ++I) { 1053 MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I); 1054 enterIntvAtEnd(MBB); 1055 } 1056 1057 // Switch all loop blocks. 1058 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Loop.begin(), 1059 E = Blocks.Loop.end(); I != E; ++I) 1060 useIntv(**I); 1061 1062 // Insert back copies in the exit blocks. 1063 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Exits.begin(), 1064 E = Blocks.Exits.end(); I != E; ++I) { 1065 MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I); 1066 leaveIntvAtTop(MBB); 1067 } 1068 1069 // Done. 1070 closeIntv(); 1071 finish(); 1072} 1073 1074 1075//===----------------------------------------------------------------------===// 1076// Single Block Splitting 1077//===----------------------------------------------------------------------===// 1078 1079/// getMultiUseBlocks - if curli has more than one use in a basic block, it 1080/// may be an advantage to split curli for the duration of the block. 1081bool SplitAnalysis::getMultiUseBlocks(BlockPtrSet &Blocks) { 1082 // If curli is local to one block, there is no point to splitting it. 1083 if (usingBlocks_.size() <= 1) 1084 return false; 1085 // Add blocks with multiple uses. 1086 for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end(); 1087 I != E; ++I) 1088 switch (I->second) { 1089 case 0: 1090 case 1: 1091 continue; 1092 case 2: { 1093 // When there are only two uses and curli is both live in and live out, 1094 // we don't really win anything by isolating the block since we would be 1095 // inserting two copies. 1096 // The remaing register would still have two uses in the block. (Unless it 1097 // separates into disconnected components). 1098 if (lis_.isLiveInToMBB(*curli_, I->first) && 1099 lis_.isLiveOutOfMBB(*curli_, I->first)) 1100 continue; 1101 } // Fall through. 1102 default: 1103 Blocks.insert(I->first); 1104 } 1105 return !Blocks.empty(); 1106} 1107 1108/// splitSingleBlocks - Split curli into a separate live interval inside each 1109/// basic block in Blocks. 1110void SplitEditor::splitSingleBlocks(const SplitAnalysis::BlockPtrSet &Blocks) { 1111 DEBUG(dbgs() << " splitSingleBlocks for " << Blocks.size() << " blocks.\n"); 1112 // Determine the first and last instruction using curli in each block. 1113 typedef std::pair<SlotIndex,SlotIndex> IndexPair; 1114 typedef DenseMap<const MachineBasicBlock*,IndexPair> IndexPairMap; 1115 IndexPairMap MBBRange; 1116 for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(), 1117 E = sa_.usingInstrs_.end(); I != E; ++I) { 1118 const MachineBasicBlock *MBB = (*I)->getParent(); 1119 if (!Blocks.count(MBB)) 1120 continue; 1121 SlotIndex Idx = lis_.getInstructionIndex(*I); 1122 DEBUG(dbgs() << " BB#" << MBB->getNumber() << '\t' << Idx << '\t' << **I); 1123 IndexPair &IP = MBBRange[MBB]; 1124 if (!IP.first.isValid() || Idx < IP.first) 1125 IP.first = Idx; 1126 if (!IP.second.isValid() || Idx > IP.second) 1127 IP.second = Idx; 1128 } 1129 1130 // Create a new interval for each block. 1131 for (SplitAnalysis::BlockPtrSet::const_iterator I = Blocks.begin(), 1132 E = Blocks.end(); I != E; ++I) { 1133 IndexPair &IP = MBBRange[*I]; 1134 DEBUG(dbgs() << " splitting for BB#" << (*I)->getNumber() << ": [" 1135 << IP.first << ';' << IP.second << ")\n"); 1136 assert(IP.first.isValid() && IP.second.isValid()); 1137 1138 openIntv(); 1139 enterIntvBefore(IP.first); 1140 useIntv(IP.first.getBaseIndex(), IP.second.getBoundaryIndex()); 1141 leaveIntvAfter(IP.second); 1142 closeIntv(); 1143 } 1144 finish(); 1145} 1146 1147 1148//===----------------------------------------------------------------------===// 1149// Sub Block Splitting 1150//===----------------------------------------------------------------------===// 1151 1152/// getBlockForInsideSplit - If curli is contained inside a single basic block, 1153/// and it wou pay to subdivide the interval inside that block, return it. 1154/// Otherwise return NULL. The returned block can be passed to 1155/// SplitEditor::splitInsideBlock. 1156const MachineBasicBlock *SplitAnalysis::getBlockForInsideSplit() { 1157 // The interval must be exclusive to one block. 1158 if (usingBlocks_.size() != 1) 1159 return 0; 1160 // Don't to this for less than 4 instructions. We want to be sure that 1161 // splitting actually reduces the instruction count per interval. 1162 if (usingInstrs_.size() < 4) 1163 return 0; 1164 return usingBlocks_.begin()->first; 1165} 1166 1167/// splitInsideBlock - Split curli into multiple intervals inside MBB. 1168void SplitEditor::splitInsideBlock(const MachineBasicBlock *MBB) { 1169 SmallVector<SlotIndex, 32> Uses; 1170 Uses.reserve(sa_.usingInstrs_.size()); 1171 for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(), 1172 E = sa_.usingInstrs_.end(); I != E; ++I) 1173 if ((*I)->getParent() == MBB) 1174 Uses.push_back(lis_.getInstructionIndex(*I)); 1175 DEBUG(dbgs() << " splitInsideBlock BB#" << MBB->getNumber() << " for " 1176 << Uses.size() << " instructions.\n"); 1177 assert(Uses.size() >= 3 && "Need at least 3 instructions"); 1178 array_pod_sort(Uses.begin(), Uses.end()); 1179 1180 // Simple algorithm: Find the largest gap between uses as determined by slot 1181 // indices. Create new intervals for instructions before the gap and after the 1182 // gap. 1183 unsigned bestPos = 0; 1184 int bestGap = 0; 1185 DEBUG(dbgs() << " dist (" << Uses[0]); 1186 for (unsigned i = 1, e = Uses.size(); i != e; ++i) { 1187 int g = Uses[i-1].distance(Uses[i]); 1188 DEBUG(dbgs() << ") -" << g << "- (" << Uses[i]); 1189 if (g > bestGap) 1190 bestPos = i, bestGap = g; 1191 } 1192 DEBUG(dbgs() << "), best: -" << bestGap << "-\n"); 1193 1194 // bestPos points to the first use after the best gap. 1195 assert(bestPos > 0 && "Invalid gap"); 1196 1197 // FIXME: Don't create intervals for low densities. 1198 1199 // First interval before the gap. Don't create single-instr intervals. 1200 if (bestPos > 1) { 1201 openIntv(); 1202 enterIntvBefore(Uses.front()); 1203 useIntv(Uses.front().getBaseIndex(), Uses[bestPos-1].getBoundaryIndex()); 1204 leaveIntvAfter(Uses[bestPos-1]); 1205 closeIntv(); 1206 } 1207 1208 // Second interval after the gap. 1209 if (bestPos < Uses.size()-1) { 1210 openIntv(); 1211 enterIntvBefore(Uses[bestPos]); 1212 useIntv(Uses[bestPos].getBaseIndex(), Uses.back().getBoundaryIndex()); 1213 leaveIntvAfter(Uses.back()); 1214 closeIntv(); 1215 } 1216 1217 finish(); 1218} 1219