StackColoring.cpp revision 5ed625c3cff2511469e9b3c5131c29fd89ddd482
1//===-- StackColoring.cpp -------------------------------------------------===// 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 pass implements the stack-coloring optimization that looks for 11// lifetime markers machine instructions (LIFESTART_BEGIN and LIFESTART_END), 12// which represent the possible lifetime of stack slots. It attempts to 13// merge disjoint stack slots and reduce the used stack space. 14// NOTE: This pass is not StackSlotColoring, which optimizes spill slots. 15// 16// TODO: In the future we plan to improve stack coloring in the following ways: 17// 1. Allow merging multiple small slots into a single larger slot at different 18// offsets. 19// 2. Merge this pass with StackSlotColoring and allow merging of allocas with 20// spill slots. 21// 22//===----------------------------------------------------------------------===// 23 24#define DEBUG_TYPE "stackcoloring" 25#include "llvm/CodeGen/Passes.h" 26#include "llvm/ADT/BitVector.h" 27#include "llvm/ADT/DepthFirstIterator.h" 28#include "llvm/ADT/PostOrderIterator.h" 29#include "llvm/ADT/SetVector.h" 30#include "llvm/ADT/SmallPtrSet.h" 31#include "llvm/ADT/SparseSet.h" 32#include "llvm/ADT/Statistic.h" 33#include "llvm/Analysis/Dominators.h" 34#include "llvm/Analysis/ValueTracking.h" 35#include "llvm/CodeGen/LiveInterval.h" 36#include "llvm/CodeGen/MachineBasicBlock.h" 37#include "llvm/CodeGen/MachineBranchProbabilityInfo.h" 38#include "llvm/CodeGen/MachineDominators.h" 39#include "llvm/CodeGen/MachineFrameInfo.h" 40#include "llvm/CodeGen/MachineFunctionPass.h" 41#include "llvm/CodeGen/MachineLoopInfo.h" 42#include "llvm/CodeGen/MachineMemOperand.h" 43#include "llvm/CodeGen/MachineModuleInfo.h" 44#include "llvm/CodeGen/MachineRegisterInfo.h" 45#include "llvm/CodeGen/SlotIndexes.h" 46#include "llvm/DebugInfo.h" 47#include "llvm/IR/Function.h" 48#include "llvm/IR/Instructions.h" 49#include "llvm/IR/Module.h" 50#include "llvm/MC/MCInstrItineraries.h" 51#include "llvm/Support/CommandLine.h" 52#include "llvm/Support/Debug.h" 53#include "llvm/Support/raw_ostream.h" 54#include "llvm/Target/TargetInstrInfo.h" 55#include "llvm/Target/TargetRegisterInfo.h" 56 57using namespace llvm; 58 59static cl::opt<bool> 60DisableColoring("no-stack-coloring", 61 cl::init(false), cl::Hidden, 62 cl::desc("Disable stack coloring")); 63 64/// The user may write code that uses allocas outside of the declared lifetime 65/// zone. This can happen when the user returns a reference to a local 66/// data-structure. We can detect these cases and decide not to optimize the 67/// code. If this flag is enabled, we try to save the user. 68static cl::opt<bool> 69ProtectFromEscapedAllocas("protect-from-escaped-allocas", 70 cl::init(false), cl::Hidden, 71 cl::desc("Do not optimize lifetime zones that are broken")); 72 73STATISTIC(NumMarkerSeen, "Number of lifetime markers found."); 74STATISTIC(StackSpaceSaved, "Number of bytes saved due to merging slots."); 75STATISTIC(StackSlotMerged, "Number of stack slot merged."); 76STATISTIC(EscapedAllocas, 77 "Number of allocas that escaped the lifetime region"); 78 79//===----------------------------------------------------------------------===// 80// StackColoring Pass 81//===----------------------------------------------------------------------===// 82 83namespace { 84/// StackColoring - A machine pass for merging disjoint stack allocations, 85/// marked by the LIFETIME_START and LIFETIME_END pseudo instructions. 86class StackColoring : public MachineFunctionPass { 87 MachineFrameInfo *MFI; 88 MachineFunction *MF; 89 90 /// A class representing liveness information for a single basic block. 91 /// Each bit in the BitVector represents the liveness property 92 /// for a different stack slot. 93 struct BlockLifetimeInfo { 94 /// Which slots BEGINs in each basic block. 95 BitVector Begin; 96 /// Which slots ENDs in each basic block. 97 BitVector End; 98 /// Which slots are marked as LIVE_IN, coming into each basic block. 99 BitVector LiveIn; 100 /// Which slots are marked as LIVE_OUT, coming out of each basic block. 101 BitVector LiveOut; 102 }; 103 104 /// Maps active slots (per bit) for each basic block. 105 DenseMap<MachineBasicBlock*, BlockLifetimeInfo> BlockLiveness; 106 107 /// Maps serial numbers to basic blocks. 108 DenseMap<MachineBasicBlock*, int> BasicBlocks; 109 /// Maps basic blocks to a serial number. 110 SmallVector<MachineBasicBlock*, 8> BasicBlockNumbering; 111 112 /// Maps liveness intervals for each slot. 113 SmallVector<LiveInterval*, 16> Intervals; 114 /// VNInfo is used for the construction of LiveIntervals. 115 VNInfo::Allocator VNInfoAllocator; 116 /// SlotIndex analysis object. 117 SlotIndexes *Indexes; 118 119 /// The list of lifetime markers found. These markers are to be removed 120 /// once the coloring is done. 121 SmallVector<MachineInstr*, 8> Markers; 122 123 /// SlotSizeSorter - A Sort utility for arranging stack slots according 124 /// to their size. 125 struct SlotSizeSorter { 126 MachineFrameInfo *MFI; 127 SlotSizeSorter(MachineFrameInfo *mfi) : MFI(mfi) { } 128 bool operator()(int LHS, int RHS) { 129 // We use -1 to denote a uninteresting slot. Place these slots at the end. 130 if (LHS == -1) return false; 131 if (RHS == -1) return true; 132 // Sort according to size. 133 return MFI->getObjectSize(LHS) > MFI->getObjectSize(RHS); 134 } 135}; 136 137public: 138 static char ID; 139 StackColoring() : MachineFunctionPass(ID) { 140 initializeStackColoringPass(*PassRegistry::getPassRegistry()); 141 } 142 void getAnalysisUsage(AnalysisUsage &AU) const; 143 bool runOnMachineFunction(MachineFunction &MF); 144 145private: 146 /// Debug. 147 void dump(); 148 149 /// Removes all of the lifetime marker instructions from the function. 150 /// \returns true if any markers were removed. 151 bool removeAllMarkers(); 152 153 /// Scan the machine function and find all of the lifetime markers. 154 /// Record the findings in the BEGIN and END vectors. 155 /// \returns the number of markers found. 156 unsigned collectMarkers(unsigned NumSlot); 157 158 /// Perform the dataflow calculation and calculate the lifetime for each of 159 /// the slots, based on the BEGIN/END vectors. Set the LifetimeLIVE_IN and 160 /// LifetimeLIVE_OUT maps that represent which stack slots are live coming 161 /// in and out blocks. 162 void calculateLocalLiveness(); 163 164 /// Construct the LiveIntervals for the slots. 165 void calculateLiveIntervals(unsigned NumSlots); 166 167 /// Go over the machine function and change instructions which use stack 168 /// slots to use the joint slots. 169 void remapInstructions(DenseMap<int, int> &SlotRemap); 170 171 /// The input program may contain intructions which are not inside lifetime 172 /// markers. This can happen due to a bug in the compiler or due to a bug in 173 /// user code (for example, returning a reference to a local variable). 174 /// This procedure checks all of the instructions in the function and 175 /// invalidates lifetime ranges which do not contain all of the instructions 176 /// which access that frame slot. 177 void removeInvalidSlotRanges(); 178 179 /// Map entries which point to other entries to their destination. 180 /// A->B->C becomes A->C. 181 void expungeSlotMap(DenseMap<int, int> &SlotRemap, unsigned NumSlots); 182}; 183} // end anonymous namespace 184 185char StackColoring::ID = 0; 186char &llvm::StackColoringID = StackColoring::ID; 187 188INITIALIZE_PASS_BEGIN(StackColoring, 189 "stack-coloring", "Merge disjoint stack slots", false, false) 190INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) 191INITIALIZE_PASS_DEPENDENCY(SlotIndexes) 192INITIALIZE_PASS_END(StackColoring, 193 "stack-coloring", "Merge disjoint stack slots", false, false) 194 195void StackColoring::getAnalysisUsage(AnalysisUsage &AU) const { 196 AU.addRequired<MachineDominatorTree>(); 197 AU.addPreserved<MachineDominatorTree>(); 198 AU.addRequired<SlotIndexes>(); 199 MachineFunctionPass::getAnalysisUsage(AU); 200} 201 202void StackColoring::dump() { 203 for (df_iterator<MachineFunction*> FI = df_begin(MF), FE = df_end(MF); 204 FI != FE; ++FI) { 205 unsigned Num = BasicBlocks[*FI]; 206 DEBUG(dbgs()<<"Inspecting block #"<<Num<<" ["<<FI->getName()<<"]\n"); 207 Num = 0; 208 DEBUG(dbgs()<<"BEGIN : {"); 209 for (unsigned i=0; i < BlockLiveness[*FI].Begin.size(); ++i) 210 DEBUG(dbgs()<<BlockLiveness[*FI].Begin.test(i)<<" "); 211 DEBUG(dbgs()<<"}\n"); 212 213 DEBUG(dbgs()<<"END : {"); 214 for (unsigned i=0; i < BlockLiveness[*FI].End.size(); ++i) 215 DEBUG(dbgs()<<BlockLiveness[*FI].End.test(i)<<" "); 216 217 DEBUG(dbgs()<<"}\n"); 218 219 DEBUG(dbgs()<<"LIVE_IN: {"); 220 for (unsigned i=0; i < BlockLiveness[*FI].LiveIn.size(); ++i) 221 DEBUG(dbgs()<<BlockLiveness[*FI].LiveIn.test(i)<<" "); 222 223 DEBUG(dbgs()<<"}\n"); 224 DEBUG(dbgs()<<"LIVEOUT: {"); 225 for (unsigned i=0; i < BlockLiveness[*FI].LiveOut.size(); ++i) 226 DEBUG(dbgs()<<BlockLiveness[*FI].LiveOut.test(i)<<" "); 227 DEBUG(dbgs()<<"}\n"); 228 } 229} 230 231unsigned StackColoring::collectMarkers(unsigned NumSlot) { 232 unsigned MarkersFound = 0; 233 // Scan the function to find all lifetime markers. 234 // NOTE: We use the a reverse-post-order iteration to ensure that we obtain a 235 // deterministic numbering, and because we'll need a post-order iteration 236 // later for solving the liveness dataflow problem. 237 for (df_iterator<MachineFunction*> FI = df_begin(MF), FE = df_end(MF); 238 FI != FE; ++FI) { 239 240 // Assign a serial number to this basic block. 241 BasicBlocks[*FI] = BasicBlockNumbering.size(); 242 BasicBlockNumbering.push_back(*FI); 243 244 BlockLiveness[*FI].Begin.resize(NumSlot); 245 BlockLiveness[*FI].End.resize(NumSlot); 246 247 for (MachineBasicBlock::iterator BI = (*FI)->begin(), BE = (*FI)->end(); 248 BI != BE; ++BI) { 249 250 if (BI->getOpcode() != TargetOpcode::LIFETIME_START && 251 BI->getOpcode() != TargetOpcode::LIFETIME_END) 252 continue; 253 254 Markers.push_back(BI); 255 256 bool IsStart = BI->getOpcode() == TargetOpcode::LIFETIME_START; 257 MachineOperand &MI = BI->getOperand(0); 258 unsigned Slot = MI.getIndex(); 259 260 MarkersFound++; 261 262 const AllocaInst *Allocation = MFI->getObjectAllocation(Slot); 263 if (Allocation) { 264 DEBUG(dbgs()<<"Found a lifetime marker for slot #"<<Slot<< 265 " with allocation: "<< Allocation->getName()<<"\n"); 266 } 267 268 if (IsStart) { 269 BlockLiveness[*FI].Begin.set(Slot); 270 } else { 271 if (BlockLiveness[*FI].Begin.test(Slot)) { 272 // Allocas that start and end within a single block are handled 273 // specially when computing the LiveIntervals to avoid pessimizing 274 // the liveness propagation. 275 BlockLiveness[*FI].Begin.reset(Slot); 276 } else { 277 BlockLiveness[*FI].End.set(Slot); 278 } 279 } 280 } 281 } 282 283 // Update statistics. 284 NumMarkerSeen += MarkersFound; 285 return MarkersFound; 286} 287 288void StackColoring::calculateLocalLiveness() { 289 // Perform a standard reverse dataflow computation to solve for 290 // global liveness. The BEGIN set here is equivalent to KILL in the standard 291 // formulation, and END is equivalent to GEN. The result of this computation 292 // is a map from blocks to bitvectors where the bitvectors represent which 293 // allocas are live in/out of that block. 294 SmallPtrSet<MachineBasicBlock*, 8> BBSet(BasicBlockNumbering.begin(), 295 BasicBlockNumbering.end()); 296 unsigned NumSSMIters = 0; 297 bool changed = true; 298 while (changed) { 299 changed = false; 300 ++NumSSMIters; 301 302 SmallPtrSet<MachineBasicBlock*, 8> NextBBSet; 303 304 for (SmallVector<MachineBasicBlock*, 8>::iterator 305 PI = BasicBlockNumbering.begin(), PE = BasicBlockNumbering.end(); 306 PI != PE; ++PI) { 307 308 MachineBasicBlock *BB = *PI; 309 if (!BBSet.count(BB)) continue; 310 311 BitVector LocalLiveIn; 312 BitVector LocalLiveOut; 313 314 // Forward propagation from begins to ends. 315 for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(), 316 PE = BB->pred_end(); PI != PE; ++PI) 317 LocalLiveIn |= BlockLiveness[*PI].LiveOut; 318 LocalLiveIn |= BlockLiveness[BB].End; 319 LocalLiveIn.reset(BlockLiveness[BB].Begin); 320 321 // Reverse propagation from ends to begins. 322 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 323 SE = BB->succ_end(); SI != SE; ++SI) 324 LocalLiveOut |= BlockLiveness[*SI].LiveIn; 325 LocalLiveOut |= BlockLiveness[BB].Begin; 326 LocalLiveOut.reset(BlockLiveness[BB].End); 327 328 LocalLiveIn |= LocalLiveOut; 329 LocalLiveOut |= LocalLiveIn; 330 331 // After adopting the live bits, we need to turn-off the bits which 332 // are de-activated in this block. 333 LocalLiveOut.reset(BlockLiveness[BB].End); 334 LocalLiveIn.reset(BlockLiveness[BB].Begin); 335 336 // If we have both BEGIN and END markers in the same basic block then 337 // we know that the BEGIN marker comes after the END, because we already 338 // handle the case where the BEGIN comes before the END when collecting 339 // the markers (and building the BEGIN/END vectore). 340 // Want to enable the LIVE_IN and LIVE_OUT of slots that have both 341 // BEGIN and END because it means that the value lives before and after 342 // this basic block. 343 BitVector LocalEndBegin = BlockLiveness[BB].End; 344 LocalEndBegin &= BlockLiveness[BB].Begin; 345 LocalLiveIn |= LocalEndBegin; 346 LocalLiveOut |= LocalEndBegin; 347 348 if (LocalLiveIn.test(BlockLiveness[BB].LiveIn)) { 349 changed = true; 350 BlockLiveness[BB].LiveIn |= LocalLiveIn; 351 352 for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(), 353 PE = BB->pred_end(); PI != PE; ++PI) 354 NextBBSet.insert(*PI); 355 } 356 357 if (LocalLiveOut.test(BlockLiveness[BB].LiveOut)) { 358 changed = true; 359 BlockLiveness[BB].LiveOut |= LocalLiveOut; 360 361 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 362 SE = BB->succ_end(); SI != SE; ++SI) 363 NextBBSet.insert(*SI); 364 } 365 } 366 367 BBSet = NextBBSet; 368 }// while changed. 369} 370 371void StackColoring::calculateLiveIntervals(unsigned NumSlots) { 372 SmallVector<SlotIndex, 16> Starts; 373 SmallVector<SlotIndex, 16> Finishes; 374 375 // For each block, find which slots are active within this block 376 // and update the live intervals. 377 for (MachineFunction::iterator MBB = MF->begin(), MBBe = MF->end(); 378 MBB != MBBe; ++MBB) { 379 Starts.clear(); 380 Starts.resize(NumSlots); 381 Finishes.clear(); 382 Finishes.resize(NumSlots); 383 384 // Create the interval for the basic blocks with lifetime markers in them. 385 for (SmallVector<MachineInstr*, 8>::iterator it = Markers.begin(), 386 e = Markers.end(); it != e; ++it) { 387 MachineInstr *MI = *it; 388 if (MI->getParent() != MBB) 389 continue; 390 391 assert((MI->getOpcode() == TargetOpcode::LIFETIME_START || 392 MI->getOpcode() == TargetOpcode::LIFETIME_END) && 393 "Invalid Lifetime marker"); 394 395 bool IsStart = MI->getOpcode() == TargetOpcode::LIFETIME_START; 396 MachineOperand &Mo = MI->getOperand(0); 397 int Slot = Mo.getIndex(); 398 assert(Slot >= 0 && "Invalid slot"); 399 400 SlotIndex ThisIndex = Indexes->getInstructionIndex(MI); 401 402 if (IsStart) { 403 if (!Starts[Slot].isValid() || Starts[Slot] > ThisIndex) 404 Starts[Slot] = ThisIndex; 405 } else { 406 if (!Finishes[Slot].isValid() || Finishes[Slot] < ThisIndex) 407 Finishes[Slot] = ThisIndex; 408 } 409 } 410 411 // Create the interval of the blocks that we previously found to be 'alive'. 412 BitVector Alive = BlockLiveness[MBB].LiveIn; 413 Alive |= BlockLiveness[MBB].LiveOut; 414 415 if (Alive.any()) { 416 for (int pos = Alive.find_first(); pos != -1; 417 pos = Alive.find_next(pos)) { 418 if (!Starts[pos].isValid()) 419 Starts[pos] = Indexes->getMBBStartIdx(MBB); 420 if (!Finishes[pos].isValid()) 421 Finishes[pos] = Indexes->getMBBEndIdx(MBB); 422 } 423 } 424 425 for (unsigned i = 0; i < NumSlots; ++i) { 426 assert(Starts[i].isValid() == Finishes[i].isValid() && "Unmatched range"); 427 if (!Starts[i].isValid()) 428 continue; 429 430 assert(Starts[i] && Finishes[i] && "Invalid interval"); 431 VNInfo *ValNum = Intervals[i]->getValNumInfo(0); 432 SlotIndex S = Starts[i]; 433 SlotIndex F = Finishes[i]; 434 if (S < F) { 435 // We have a single consecutive region. 436 Intervals[i]->addRange(LiveRange(S, F, ValNum)); 437 } else { 438 // We have two non consecutive regions. This happens when 439 // LIFETIME_START appears after the LIFETIME_END marker. 440 SlotIndex NewStart = Indexes->getMBBStartIdx(MBB); 441 SlotIndex NewFin = Indexes->getMBBEndIdx(MBB); 442 Intervals[i]->addRange(LiveRange(NewStart, F, ValNum)); 443 Intervals[i]->addRange(LiveRange(S, NewFin, ValNum)); 444 } 445 } 446 } 447} 448 449bool StackColoring::removeAllMarkers() { 450 unsigned Count = 0; 451 for (unsigned i = 0; i < Markers.size(); ++i) { 452 Markers[i]->eraseFromParent(); 453 Count++; 454 } 455 Markers.clear(); 456 457 DEBUG(dbgs()<<"Removed "<<Count<<" markers.\n"); 458 return Count; 459} 460 461void StackColoring::remapInstructions(DenseMap<int, int> &SlotRemap) { 462 unsigned FixedInstr = 0; 463 unsigned FixedMemOp = 0; 464 unsigned FixedDbg = 0; 465 MachineModuleInfo *MMI = &MF->getMMI(); 466 467 // Remap debug information that refers to stack slots. 468 MachineModuleInfo::VariableDbgInfoMapTy &VMap = MMI->getVariableDbgInfo(); 469 for (MachineModuleInfo::VariableDbgInfoMapTy::iterator VI = VMap.begin(), 470 VE = VMap.end(); VI != VE; ++VI) { 471 const MDNode *Var = VI->first; 472 if (!Var) continue; 473 std::pair<unsigned, DebugLoc> &VP = VI->second; 474 if (SlotRemap.count(VP.first)) { 475 DEBUG(dbgs()<<"Remapping debug info for ["<<Var->getName()<<"].\n"); 476 VP.first = SlotRemap[VP.first]; 477 FixedDbg++; 478 } 479 } 480 481 // Keep a list of *allocas* which need to be remapped. 482 DenseMap<const AllocaInst*, const AllocaInst*> Allocas; 483 for (DenseMap<int, int>::iterator it = SlotRemap.begin(), 484 e = SlotRemap.end(); it != e; ++it) { 485 const AllocaInst *From = MFI->getObjectAllocation(it->first); 486 const AllocaInst *To = MFI->getObjectAllocation(it->second); 487 assert(To && From && "Invalid allocation object"); 488 Allocas[From] = To; 489 } 490 491 // Remap all instructions to the new stack slots. 492 MachineFunction::iterator BB, BBE; 493 MachineBasicBlock::iterator I, IE; 494 for (BB = MF->begin(), BBE = MF->end(); BB != BBE; ++BB) 495 for (I = BB->begin(), IE = BB->end(); I != IE; ++I) { 496 497 // Skip lifetime markers. We'll remove them soon. 498 if (I->getOpcode() == TargetOpcode::LIFETIME_START || 499 I->getOpcode() == TargetOpcode::LIFETIME_END) 500 continue; 501 502 // Update the MachineMemOperand to use the new alloca. 503 for (MachineInstr::mmo_iterator MM = I->memoperands_begin(), 504 E = I->memoperands_end(); MM != E; ++MM) { 505 MachineMemOperand *MMO = *MM; 506 507 const Value *V = MMO->getValue(); 508 509 if (!V) 510 continue; 511 512 // Climb up and find the original alloca. 513 V = GetUnderlyingObject(V); 514 // If we did not find one, or if the one that we found is not in our 515 // map, then move on. 516 if (!V || !isa<AllocaInst>(V)) { 517 // Clear mem operand since we don't know for sure that it doesn't 518 // alias a merged alloca. 519 MMO->setValue(0); 520 continue; 521 } 522 const AllocaInst *AI= cast<AllocaInst>(V); 523 if (!Allocas.count(AI)) 524 continue; 525 526 MMO->setValue(Allocas[AI]); 527 FixedMemOp++; 528 } 529 530 // Update all of the machine instruction operands. 531 for (unsigned i = 0 ; i < I->getNumOperands(); ++i) { 532 MachineOperand &MO = I->getOperand(i); 533 534 if (!MO.isFI()) 535 continue; 536 int FromSlot = MO.getIndex(); 537 538 // Don't touch arguments. 539 if (FromSlot<0) 540 continue; 541 542 // Only look at mapped slots. 543 if (!SlotRemap.count(FromSlot)) 544 continue; 545 546 // In a debug build, check that the instruction that we are modifying is 547 // inside the expected live range. If the instruction is not inside 548 // the calculated range then it means that the alloca usage moved 549 // outside of the lifetime markers, or that the user has a bug. 550 // NOTE: Alloca address calculations which happen outside the lifetime 551 // zone are are okay, despite the fact that we don't have a good way 552 // for validating all of the usages of the calculation. 553#ifndef NDEBUG 554 bool TouchesMemory = I->mayLoad() || I->mayStore(); 555 // If we *don't* protect the user from escaped allocas, don't bother 556 // validating the instructions. 557 if (!I->isDebugValue() && TouchesMemory && ProtectFromEscapedAllocas) { 558 SlotIndex Index = Indexes->getInstructionIndex(I); 559 LiveInterval *Interval = Intervals[FromSlot]; 560 assert(Interval->find(Index) != Interval->end() && 561 "Found instruction usage outside of live range."); 562 } 563#endif 564 565 // Fix the machine instructions. 566 int ToSlot = SlotRemap[FromSlot]; 567 MO.setIndex(ToSlot); 568 FixedInstr++; 569 } 570 } 571 572 DEBUG(dbgs()<<"Fixed "<<FixedMemOp<<" machine memory operands.\n"); 573 DEBUG(dbgs()<<"Fixed "<<FixedDbg<<" debug locations.\n"); 574 DEBUG(dbgs()<<"Fixed "<<FixedInstr<<" machine instructions.\n"); 575} 576 577void StackColoring::removeInvalidSlotRanges() { 578 MachineFunction::iterator BB, BBE; 579 MachineBasicBlock::iterator I, IE; 580 for (BB = MF->begin(), BBE = MF->end(); BB != BBE; ++BB) 581 for (I = BB->begin(), IE = BB->end(); I != IE; ++I) { 582 583 if (I->getOpcode() == TargetOpcode::LIFETIME_START || 584 I->getOpcode() == TargetOpcode::LIFETIME_END || I->isDebugValue()) 585 continue; 586 587 // Some intervals are suspicious! In some cases we find address 588 // calculations outside of the lifetime zone, but not actual memory 589 // read or write. Memory accesses outside of the lifetime zone are a clear 590 // violation, but address calculations are okay. This can happen when 591 // GEPs are hoisted outside of the lifetime zone. 592 // So, in here we only check instructions which can read or write memory. 593 if (!I->mayLoad() && !I->mayStore()) 594 continue; 595 596 // Check all of the machine operands. 597 for (unsigned i = 0 ; i < I->getNumOperands(); ++i) { 598 MachineOperand &MO = I->getOperand(i); 599 600 if (!MO.isFI()) 601 continue; 602 603 int Slot = MO.getIndex(); 604 605 if (Slot<0) 606 continue; 607 608 if (Intervals[Slot]->empty()) 609 continue; 610 611 // Check that the used slot is inside the calculated lifetime range. 612 // If it is not, warn about it and invalidate the range. 613 LiveInterval *Interval = Intervals[Slot]; 614 SlotIndex Index = Indexes->getInstructionIndex(I); 615 if (Interval->find(Index) == Interval->end()) { 616 Intervals[Slot]->clear(); 617 DEBUG(dbgs()<<"Invalidating range #"<<Slot<<"\n"); 618 EscapedAllocas++; 619 } 620 } 621 } 622} 623 624void StackColoring::expungeSlotMap(DenseMap<int, int> &SlotRemap, 625 unsigned NumSlots) { 626 // Expunge slot remap map. 627 for (unsigned i=0; i < NumSlots; ++i) { 628 // If we are remapping i 629 if (SlotRemap.count(i)) { 630 int Target = SlotRemap[i]; 631 // As long as our target is mapped to something else, follow it. 632 while (SlotRemap.count(Target)) { 633 Target = SlotRemap[Target]; 634 SlotRemap[i] = Target; 635 } 636 } 637 } 638} 639 640bool StackColoring::runOnMachineFunction(MachineFunction &Func) { 641 DEBUG(dbgs() << "********** Stack Coloring **********\n" 642 << "********** Function: " 643 << ((const Value*)Func.getFunction())->getName() << '\n'); 644 MF = &Func; 645 MFI = MF->getFrameInfo(); 646 Indexes = &getAnalysis<SlotIndexes>(); 647 BlockLiveness.clear(); 648 BasicBlocks.clear(); 649 BasicBlockNumbering.clear(); 650 Markers.clear(); 651 Intervals.clear(); 652 VNInfoAllocator.Reset(); 653 654 unsigned NumSlots = MFI->getObjectIndexEnd(); 655 656 // If there are no stack slots then there are no markers to remove. 657 if (!NumSlots) 658 return false; 659 660 SmallVector<int, 8> SortedSlots; 661 662 SortedSlots.reserve(NumSlots); 663 Intervals.reserve(NumSlots); 664 665 unsigned NumMarkers = collectMarkers(NumSlots); 666 667 unsigned TotalSize = 0; 668 DEBUG(dbgs()<<"Found "<<NumMarkers<<" markers and "<<NumSlots<<" slots\n"); 669 DEBUG(dbgs()<<"Slot structure:\n"); 670 671 for (int i=0; i < MFI->getObjectIndexEnd(); ++i) { 672 DEBUG(dbgs()<<"Slot #"<<i<<" - "<<MFI->getObjectSize(i)<<" bytes.\n"); 673 TotalSize += MFI->getObjectSize(i); 674 } 675 676 DEBUG(dbgs()<<"Total Stack size: "<<TotalSize<<" bytes\n\n"); 677 678 // Don't continue because there are not enough lifetime markers, or the 679 // stack is too small, or we are told not to optimize the slots. 680 if (NumMarkers < 2 || TotalSize < 16 || DisableColoring) { 681 DEBUG(dbgs()<<"Will not try to merge slots.\n"); 682 return removeAllMarkers(); 683 } 684 685 for (unsigned i=0; i < NumSlots; ++i) { 686 LiveInterval *LI = new LiveInterval(i, 0); 687 Intervals.push_back(LI); 688 LI->getNextValue(Indexes->getZeroIndex(), VNInfoAllocator); 689 SortedSlots.push_back(i); 690 } 691 692 // Calculate the liveness of each block. 693 calculateLocalLiveness(); 694 695 // Propagate the liveness information. 696 calculateLiveIntervals(NumSlots); 697 698 // Search for allocas which are used outside of the declared lifetime 699 // markers. 700 if (ProtectFromEscapedAllocas) 701 removeInvalidSlotRanges(); 702 703 // Maps old slots to new slots. 704 DenseMap<int, int> SlotRemap; 705 unsigned RemovedSlots = 0; 706 unsigned ReducedSize = 0; 707 708 // Do not bother looking at empty intervals. 709 for (unsigned I = 0; I < NumSlots; ++I) { 710 if (Intervals[SortedSlots[I]]->empty()) 711 SortedSlots[I] = -1; 712 } 713 714 // This is a simple greedy algorithm for merging allocas. First, sort the 715 // slots, placing the largest slots first. Next, perform an n^2 scan and look 716 // for disjoint slots. When you find disjoint slots, merge the samller one 717 // into the bigger one and update the live interval. Remove the small alloca 718 // and continue. 719 720 // Sort the slots according to their size. Place unused slots at the end. 721 // Use stable sort to guarantee deterministic code generation. 722 std::stable_sort(SortedSlots.begin(), SortedSlots.end(), 723 SlotSizeSorter(MFI)); 724 725 bool Chanded = true; 726 while (Chanded) { 727 Chanded = false; 728 for (unsigned I = 0; I < NumSlots; ++I) { 729 if (SortedSlots[I] == -1) 730 continue; 731 732 for (unsigned J=I+1; J < NumSlots; ++J) { 733 if (SortedSlots[J] == -1) 734 continue; 735 736 int FirstSlot = SortedSlots[I]; 737 int SecondSlot = SortedSlots[J]; 738 LiveInterval *First = Intervals[FirstSlot]; 739 LiveInterval *Second = Intervals[SecondSlot]; 740 assert (!First->empty() && !Second->empty() && "Found an empty range"); 741 742 // Merge disjoint slots. 743 if (!First->overlaps(*Second)) { 744 Chanded = true; 745 First->MergeRangesInAsValue(*Second, First->getValNumInfo(0)); 746 SlotRemap[SecondSlot] = FirstSlot; 747 SortedSlots[J] = -1; 748 DEBUG(dbgs()<<"Merging #"<<FirstSlot<<" and slots #"<< 749 SecondSlot<<" together.\n"); 750 unsigned MaxAlignment = std::max(MFI->getObjectAlignment(FirstSlot), 751 MFI->getObjectAlignment(SecondSlot)); 752 753 assert(MFI->getObjectSize(FirstSlot) >= 754 MFI->getObjectSize(SecondSlot) && 755 "Merging a small object into a larger one"); 756 757 RemovedSlots+=1; 758 ReducedSize += MFI->getObjectSize(SecondSlot); 759 MFI->setObjectAlignment(FirstSlot, MaxAlignment); 760 MFI->RemoveStackObject(SecondSlot); 761 } 762 } 763 } 764 }// While changed. 765 766 // Record statistics. 767 StackSpaceSaved += ReducedSize; 768 StackSlotMerged += RemovedSlots; 769 DEBUG(dbgs()<<"Merge "<<RemovedSlots<<" slots. Saved "<< 770 ReducedSize<<" bytes\n"); 771 772 // Scan the entire function and update all machine operands that use frame 773 // indices to use the remapped frame index. 774 expungeSlotMap(SlotRemap, NumSlots); 775 remapInstructions(SlotRemap); 776 777 // Release the intervals. 778 for (unsigned I = 0; I < NumSlots; ++I) { 779 delete Intervals[I]; 780 } 781 782 return removeAllMarkers(); 783} 784