StackColoring.cpp revision f1af1feeee0f0ec797410762c006211f9c1e2a0f
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 DEBUG(dbgs()<<"Inspecting block #"<<BasicBlocks[*FI]<< 206 " ["<<FI->getName()<<"]\n"); 207 DEBUG(dbgs()<<"BEGIN : {"); 208 for (unsigned i=0; i < BlockLiveness[*FI].Begin.size(); ++i) 209 DEBUG(dbgs()<<BlockLiveness[*FI].Begin.test(i)<<" "); 210 DEBUG(dbgs()<<"}\n"); 211 212 DEBUG(dbgs()<<"END : {"); 213 for (unsigned i=0; i < BlockLiveness[*FI].End.size(); ++i) 214 DEBUG(dbgs()<<BlockLiveness[*FI].End.test(i)<<" "); 215 216 DEBUG(dbgs()<<"}\n"); 217 218 DEBUG(dbgs()<<"LIVE_IN: {"); 219 for (unsigned i=0; i < BlockLiveness[*FI].LiveIn.size(); ++i) 220 DEBUG(dbgs()<<BlockLiveness[*FI].LiveIn.test(i)<<" "); 221 222 DEBUG(dbgs()<<"}\n"); 223 DEBUG(dbgs()<<"LIVEOUT: {"); 224 for (unsigned i=0; i < BlockLiveness[*FI].LiveOut.size(); ++i) 225 DEBUG(dbgs()<<BlockLiveness[*FI].LiveOut.test(i)<<" "); 226 DEBUG(dbgs()<<"}\n"); 227 } 228} 229 230unsigned StackColoring::collectMarkers(unsigned NumSlot) { 231 unsigned MarkersFound = 0; 232 // Scan the function to find all lifetime markers. 233 // NOTE: We use the a reverse-post-order iteration to ensure that we obtain a 234 // deterministic numbering, and because we'll need a post-order iteration 235 // later for solving the liveness dataflow problem. 236 for (df_iterator<MachineFunction*> FI = df_begin(MF), FE = df_end(MF); 237 FI != FE; ++FI) { 238 239 // Assign a serial number to this basic block. 240 BasicBlocks[*FI] = BasicBlockNumbering.size(); 241 BasicBlockNumbering.push_back(*FI); 242 243 BlockLiveness[*FI].Begin.resize(NumSlot); 244 BlockLiveness[*FI].End.resize(NumSlot); 245 246 for (MachineBasicBlock::iterator BI = (*FI)->begin(), BE = (*FI)->end(); 247 BI != BE; ++BI) { 248 249 if (BI->getOpcode() != TargetOpcode::LIFETIME_START && 250 BI->getOpcode() != TargetOpcode::LIFETIME_END) 251 continue; 252 253 Markers.push_back(BI); 254 255 bool IsStart = BI->getOpcode() == TargetOpcode::LIFETIME_START; 256 MachineOperand &MI = BI->getOperand(0); 257 unsigned Slot = MI.getIndex(); 258 259 MarkersFound++; 260 261 const AllocaInst *Allocation = MFI->getObjectAllocation(Slot); 262 if (Allocation) { 263 DEBUG(dbgs()<<"Found a lifetime marker for slot #"<<Slot<< 264 " with allocation: "<< Allocation->getName()<<"\n"); 265 } 266 267 if (IsStart) { 268 BlockLiveness[*FI].Begin.set(Slot); 269 } else { 270 if (BlockLiveness[*FI].Begin.test(Slot)) { 271 // Allocas that start and end within a single block are handled 272 // specially when computing the LiveIntervals to avoid pessimizing 273 // the liveness propagation. 274 BlockLiveness[*FI].Begin.reset(Slot); 275 } else { 276 BlockLiveness[*FI].End.set(Slot); 277 } 278 } 279 } 280 } 281 282 // Update statistics. 283 NumMarkerSeen += MarkersFound; 284 return MarkersFound; 285} 286 287void StackColoring::calculateLocalLiveness() { 288 // Perform a standard reverse dataflow computation to solve for 289 // global liveness. The BEGIN set here is equivalent to KILL in the standard 290 // formulation, and END is equivalent to GEN. The result of this computation 291 // is a map from blocks to bitvectors where the bitvectors represent which 292 // allocas are live in/out of that block. 293 SmallPtrSet<MachineBasicBlock*, 8> BBSet(BasicBlockNumbering.begin(), 294 BasicBlockNumbering.end()); 295 unsigned NumSSMIters = 0; 296 bool changed = true; 297 while (changed) { 298 changed = false; 299 ++NumSSMIters; 300 301 SmallPtrSet<MachineBasicBlock*, 8> NextBBSet; 302 303 for (SmallVector<MachineBasicBlock*, 8>::iterator 304 PI = BasicBlockNumbering.begin(), PE = BasicBlockNumbering.end(); 305 PI != PE; ++PI) { 306 307 MachineBasicBlock *BB = *PI; 308 if (!BBSet.count(BB)) continue; 309 310 BitVector LocalLiveIn; 311 BitVector LocalLiveOut; 312 313 // Forward propagation from begins to ends. 314 for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(), 315 PE = BB->pred_end(); PI != PE; ++PI) 316 LocalLiveIn |= BlockLiveness[*PI].LiveOut; 317 LocalLiveIn |= BlockLiveness[BB].End; 318 LocalLiveIn.reset(BlockLiveness[BB].Begin); 319 320 // Reverse propagation from ends to begins. 321 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 322 SE = BB->succ_end(); SI != SE; ++SI) 323 LocalLiveOut |= BlockLiveness[*SI].LiveIn; 324 LocalLiveOut |= BlockLiveness[BB].Begin; 325 LocalLiveOut.reset(BlockLiveness[BB].End); 326 327 LocalLiveIn |= LocalLiveOut; 328 LocalLiveOut |= LocalLiveIn; 329 330 // After adopting the live bits, we need to turn-off the bits which 331 // are de-activated in this block. 332 LocalLiveOut.reset(BlockLiveness[BB].End); 333 LocalLiveIn.reset(BlockLiveness[BB].Begin); 334 335 // If we have both BEGIN and END markers in the same basic block then 336 // we know that the BEGIN marker comes after the END, because we already 337 // handle the case where the BEGIN comes before the END when collecting 338 // the markers (and building the BEGIN/END vectore). 339 // Want to enable the LIVE_IN and LIVE_OUT of slots that have both 340 // BEGIN and END because it means that the value lives before and after 341 // this basic block. 342 BitVector LocalEndBegin = BlockLiveness[BB].End; 343 LocalEndBegin &= BlockLiveness[BB].Begin; 344 LocalLiveIn |= LocalEndBegin; 345 LocalLiveOut |= LocalEndBegin; 346 347 if (LocalLiveIn.test(BlockLiveness[BB].LiveIn)) { 348 changed = true; 349 BlockLiveness[BB].LiveIn |= LocalLiveIn; 350 351 for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(), 352 PE = BB->pred_end(); PI != PE; ++PI) 353 NextBBSet.insert(*PI); 354 } 355 356 if (LocalLiveOut.test(BlockLiveness[BB].LiveOut)) { 357 changed = true; 358 BlockLiveness[BB].LiveOut |= LocalLiveOut; 359 360 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 361 SE = BB->succ_end(); SI != SE; ++SI) 362 NextBBSet.insert(*SI); 363 } 364 } 365 366 BBSet = NextBBSet; 367 }// while changed. 368} 369 370void StackColoring::calculateLiveIntervals(unsigned NumSlots) { 371 SmallVector<SlotIndex, 16> Starts; 372 SmallVector<SlotIndex, 16> Finishes; 373 374 // For each block, find which slots are active within this block 375 // and update the live intervals. 376 for (MachineFunction::iterator MBB = MF->begin(), MBBe = MF->end(); 377 MBB != MBBe; ++MBB) { 378 Starts.clear(); 379 Starts.resize(NumSlots); 380 Finishes.clear(); 381 Finishes.resize(NumSlots); 382 383 // Create the interval for the basic blocks with lifetime markers in them. 384 for (SmallVector<MachineInstr*, 8>::iterator it = Markers.begin(), 385 e = Markers.end(); it != e; ++it) { 386 MachineInstr *MI = *it; 387 if (MI->getParent() != MBB) 388 continue; 389 390 assert((MI->getOpcode() == TargetOpcode::LIFETIME_START || 391 MI->getOpcode() == TargetOpcode::LIFETIME_END) && 392 "Invalid Lifetime marker"); 393 394 bool IsStart = MI->getOpcode() == TargetOpcode::LIFETIME_START; 395 MachineOperand &Mo = MI->getOperand(0); 396 int Slot = Mo.getIndex(); 397 assert(Slot >= 0 && "Invalid slot"); 398 399 SlotIndex ThisIndex = Indexes->getInstructionIndex(MI); 400 401 if (IsStart) { 402 if (!Starts[Slot].isValid() || Starts[Slot] > ThisIndex) 403 Starts[Slot] = ThisIndex; 404 } else { 405 if (!Finishes[Slot].isValid() || Finishes[Slot] < ThisIndex) 406 Finishes[Slot] = ThisIndex; 407 } 408 } 409 410 // Create the interval of the blocks that we previously found to be 'alive'. 411 BitVector Alive = BlockLiveness[MBB].LiveIn; 412 Alive |= BlockLiveness[MBB].LiveOut; 413 414 if (Alive.any()) { 415 for (int pos = Alive.find_first(); pos != -1; 416 pos = Alive.find_next(pos)) { 417 if (!Starts[pos].isValid()) 418 Starts[pos] = Indexes->getMBBStartIdx(MBB); 419 if (!Finishes[pos].isValid()) 420 Finishes[pos] = Indexes->getMBBEndIdx(MBB); 421 } 422 } 423 424 for (unsigned i = 0; i < NumSlots; ++i) { 425 assert(Starts[i].isValid() == Finishes[i].isValid() && "Unmatched range"); 426 if (!Starts[i].isValid()) 427 continue; 428 429 assert(Starts[i] && Finishes[i] && "Invalid interval"); 430 VNInfo *ValNum = Intervals[i]->getValNumInfo(0); 431 SlotIndex S = Starts[i]; 432 SlotIndex F = Finishes[i]; 433 if (S < F) { 434 // We have a single consecutive region. 435 Intervals[i]->addRange(LiveRange(S, F, ValNum)); 436 } else { 437 // We have two non consecutive regions. This happens when 438 // LIFETIME_START appears after the LIFETIME_END marker. 439 SlotIndex NewStart = Indexes->getMBBStartIdx(MBB); 440 SlotIndex NewFin = Indexes->getMBBEndIdx(MBB); 441 Intervals[i]->addRange(LiveRange(NewStart, F, ValNum)); 442 Intervals[i]->addRange(LiveRange(S, NewFin, ValNum)); 443 } 444 } 445 } 446} 447 448bool StackColoring::removeAllMarkers() { 449 unsigned Count = 0; 450 for (unsigned i = 0; i < Markers.size(); ++i) { 451 Markers[i]->eraseFromParent(); 452 Count++; 453 } 454 Markers.clear(); 455 456 DEBUG(dbgs()<<"Removed "<<Count<<" markers.\n"); 457 return Count; 458} 459 460void StackColoring::remapInstructions(DenseMap<int, int> &SlotRemap) { 461 unsigned FixedInstr = 0; 462 unsigned FixedMemOp = 0; 463 unsigned FixedDbg = 0; 464 MachineModuleInfo *MMI = &MF->getMMI(); 465 466 // Remap debug information that refers to stack slots. 467 MachineModuleInfo::VariableDbgInfoMapTy &VMap = MMI->getVariableDbgInfo(); 468 for (MachineModuleInfo::VariableDbgInfoMapTy::iterator VI = VMap.begin(), 469 VE = VMap.end(); VI != VE; ++VI) { 470 const MDNode *Var = VI->first; 471 if (!Var) continue; 472 std::pair<unsigned, DebugLoc> &VP = VI->second; 473 if (SlotRemap.count(VP.first)) { 474 DEBUG(dbgs()<<"Remapping debug info for ["<<Var->getName()<<"].\n"); 475 VP.first = SlotRemap[VP.first]; 476 FixedDbg++; 477 } 478 } 479 480 // Keep a list of *allocas* which need to be remapped. 481 DenseMap<const AllocaInst*, const AllocaInst*> Allocas; 482 for (DenseMap<int, int>::iterator it = SlotRemap.begin(), 483 e = SlotRemap.end(); it != e; ++it) { 484 const AllocaInst *From = MFI->getObjectAllocation(it->first); 485 const AllocaInst *To = MFI->getObjectAllocation(it->second); 486 assert(To && From && "Invalid allocation object"); 487 Allocas[From] = To; 488 } 489 490 // Remap all instructions to the new stack slots. 491 MachineFunction::iterator BB, BBE; 492 MachineBasicBlock::iterator I, IE; 493 for (BB = MF->begin(), BBE = MF->end(); BB != BBE; ++BB) 494 for (I = BB->begin(), IE = BB->end(); I != IE; ++I) { 495 496 // Skip lifetime markers. We'll remove them soon. 497 if (I->getOpcode() == TargetOpcode::LIFETIME_START || 498 I->getOpcode() == TargetOpcode::LIFETIME_END) 499 continue; 500 501 // Update the MachineMemOperand to use the new alloca. 502 for (MachineInstr::mmo_iterator MM = I->memoperands_begin(), 503 E = I->memoperands_end(); MM != E; ++MM) { 504 MachineMemOperand *MMO = *MM; 505 506 const Value *V = MMO->getValue(); 507 508 if (!V) 509 continue; 510 511 // Climb up and find the original alloca. 512 V = GetUnderlyingObject(V); 513 // If we did not find one, or if the one that we found is not in our 514 // map, then move on. 515 if (!V || !isa<AllocaInst>(V)) { 516 // Clear mem operand since we don't know for sure that it doesn't 517 // alias a merged alloca. 518 MMO->setValue(0); 519 continue; 520 } 521 const AllocaInst *AI= cast<AllocaInst>(V); 522 if (!Allocas.count(AI)) 523 continue; 524 525 MMO->setValue(Allocas[AI]); 526 FixedMemOp++; 527 } 528 529 // Update all of the machine instruction operands. 530 for (unsigned i = 0 ; i < I->getNumOperands(); ++i) { 531 MachineOperand &MO = I->getOperand(i); 532 533 if (!MO.isFI()) 534 continue; 535 int FromSlot = MO.getIndex(); 536 537 // Don't touch arguments. 538 if (FromSlot<0) 539 continue; 540 541 // Only look at mapped slots. 542 if (!SlotRemap.count(FromSlot)) 543 continue; 544 545 // In a debug build, check that the instruction that we are modifying is 546 // inside the expected live range. If the instruction is not inside 547 // the calculated range then it means that the alloca usage moved 548 // outside of the lifetime markers, or that the user has a bug. 549 // NOTE: Alloca address calculations which happen outside the lifetime 550 // zone are are okay, despite the fact that we don't have a good way 551 // for validating all of the usages of the calculation. 552#ifndef NDEBUG 553 bool TouchesMemory = I->mayLoad() || I->mayStore(); 554 // If we *don't* protect the user from escaped allocas, don't bother 555 // validating the instructions. 556 if (!I->isDebugValue() && TouchesMemory && ProtectFromEscapedAllocas) { 557 SlotIndex Index = Indexes->getInstructionIndex(I); 558 LiveInterval *Interval = Intervals[FromSlot]; 559 assert(Interval->find(Index) != Interval->end() && 560 "Found instruction usage outside of live range."); 561 } 562#endif 563 564 // Fix the machine instructions. 565 int ToSlot = SlotRemap[FromSlot]; 566 MO.setIndex(ToSlot); 567 FixedInstr++; 568 } 569 } 570 571 DEBUG(dbgs()<<"Fixed "<<FixedMemOp<<" machine memory operands.\n"); 572 DEBUG(dbgs()<<"Fixed "<<FixedDbg<<" debug locations.\n"); 573 DEBUG(dbgs()<<"Fixed "<<FixedInstr<<" machine instructions.\n"); 574} 575 576void StackColoring::removeInvalidSlotRanges() { 577 MachineFunction::iterator BB, BBE; 578 MachineBasicBlock::iterator I, IE; 579 for (BB = MF->begin(), BBE = MF->end(); BB != BBE; ++BB) 580 for (I = BB->begin(), IE = BB->end(); I != IE; ++I) { 581 582 if (I->getOpcode() == TargetOpcode::LIFETIME_START || 583 I->getOpcode() == TargetOpcode::LIFETIME_END || I->isDebugValue()) 584 continue; 585 586 // Some intervals are suspicious! In some cases we find address 587 // calculations outside of the lifetime zone, but not actual memory 588 // read or write. Memory accesses outside of the lifetime zone are a clear 589 // violation, but address calculations are okay. This can happen when 590 // GEPs are hoisted outside of the lifetime zone. 591 // So, in here we only check instructions which can read or write memory. 592 if (!I->mayLoad() && !I->mayStore()) 593 continue; 594 595 // Check all of the machine operands. 596 for (unsigned i = 0 ; i < I->getNumOperands(); ++i) { 597 MachineOperand &MO = I->getOperand(i); 598 599 if (!MO.isFI()) 600 continue; 601 602 int Slot = MO.getIndex(); 603 604 if (Slot<0) 605 continue; 606 607 if (Intervals[Slot]->empty()) 608 continue; 609 610 // Check that the used slot is inside the calculated lifetime range. 611 // If it is not, warn about it and invalidate the range. 612 LiveInterval *Interval = Intervals[Slot]; 613 SlotIndex Index = Indexes->getInstructionIndex(I); 614 if (Interval->find(Index) == Interval->end()) { 615 Intervals[Slot]->clear(); 616 DEBUG(dbgs()<<"Invalidating range #"<<Slot<<"\n"); 617 EscapedAllocas++; 618 } 619 } 620 } 621} 622 623void StackColoring::expungeSlotMap(DenseMap<int, int> &SlotRemap, 624 unsigned NumSlots) { 625 // Expunge slot remap map. 626 for (unsigned i=0; i < NumSlots; ++i) { 627 // If we are remapping i 628 if (SlotRemap.count(i)) { 629 int Target = SlotRemap[i]; 630 // As long as our target is mapped to something else, follow it. 631 while (SlotRemap.count(Target)) { 632 Target = SlotRemap[Target]; 633 SlotRemap[i] = Target; 634 } 635 } 636 } 637} 638 639bool StackColoring::runOnMachineFunction(MachineFunction &Func) { 640 DEBUG(dbgs() << "********** Stack Coloring **********\n" 641 << "********** Function: " 642 << ((const Value*)Func.getFunction())->getName() << '\n'); 643 MF = &Func; 644 MFI = MF->getFrameInfo(); 645 Indexes = &getAnalysis<SlotIndexes>(); 646 BlockLiveness.clear(); 647 BasicBlocks.clear(); 648 BasicBlockNumbering.clear(); 649 Markers.clear(); 650 Intervals.clear(); 651 VNInfoAllocator.Reset(); 652 653 unsigned NumSlots = MFI->getObjectIndexEnd(); 654 655 // If there are no stack slots then there are no markers to remove. 656 if (!NumSlots) 657 return false; 658 659 SmallVector<int, 8> SortedSlots; 660 661 SortedSlots.reserve(NumSlots); 662 Intervals.reserve(NumSlots); 663 664 unsigned NumMarkers = collectMarkers(NumSlots); 665 666 unsigned TotalSize = 0; 667 DEBUG(dbgs()<<"Found "<<NumMarkers<<" markers and "<<NumSlots<<" slots\n"); 668 DEBUG(dbgs()<<"Slot structure:\n"); 669 670 for (int i=0; i < MFI->getObjectIndexEnd(); ++i) { 671 DEBUG(dbgs()<<"Slot #"<<i<<" - "<<MFI->getObjectSize(i)<<" bytes.\n"); 672 TotalSize += MFI->getObjectSize(i); 673 } 674 675 DEBUG(dbgs()<<"Total Stack size: "<<TotalSize<<" bytes\n\n"); 676 677 // Don't continue because there are not enough lifetime markers, or the 678 // stack is too small, or we are told not to optimize the slots. 679 if (NumMarkers < 2 || TotalSize < 16 || DisableColoring) { 680 DEBUG(dbgs()<<"Will not try to merge slots.\n"); 681 return removeAllMarkers(); 682 } 683 684 for (unsigned i=0; i < NumSlots; ++i) { 685 LiveInterval *LI = new LiveInterval(i, 0); 686 Intervals.push_back(LI); 687 LI->getNextValue(Indexes->getZeroIndex(), VNInfoAllocator); 688 SortedSlots.push_back(i); 689 } 690 691 // Calculate the liveness of each block. 692 calculateLocalLiveness(); 693 694 // Propagate the liveness information. 695 calculateLiveIntervals(NumSlots); 696 697 // Search for allocas which are used outside of the declared lifetime 698 // markers. 699 if (ProtectFromEscapedAllocas) 700 removeInvalidSlotRanges(); 701 702 // Maps old slots to new slots. 703 DenseMap<int, int> SlotRemap; 704 unsigned RemovedSlots = 0; 705 unsigned ReducedSize = 0; 706 707 // Do not bother looking at empty intervals. 708 for (unsigned I = 0; I < NumSlots; ++I) { 709 if (Intervals[SortedSlots[I]]->empty()) 710 SortedSlots[I] = -1; 711 } 712 713 // This is a simple greedy algorithm for merging allocas. First, sort the 714 // slots, placing the largest slots first. Next, perform an n^2 scan and look 715 // for disjoint slots. When you find disjoint slots, merge the samller one 716 // into the bigger one and update the live interval. Remove the small alloca 717 // and continue. 718 719 // Sort the slots according to their size. Place unused slots at the end. 720 // Use stable sort to guarantee deterministic code generation. 721 std::stable_sort(SortedSlots.begin(), SortedSlots.end(), 722 SlotSizeSorter(MFI)); 723 724 bool Chanded = true; 725 while (Chanded) { 726 Chanded = false; 727 for (unsigned I = 0; I < NumSlots; ++I) { 728 if (SortedSlots[I] == -1) 729 continue; 730 731 for (unsigned J=I+1; J < NumSlots; ++J) { 732 if (SortedSlots[J] == -1) 733 continue; 734 735 int FirstSlot = SortedSlots[I]; 736 int SecondSlot = SortedSlots[J]; 737 LiveInterval *First = Intervals[FirstSlot]; 738 LiveInterval *Second = Intervals[SecondSlot]; 739 assert (!First->empty() && !Second->empty() && "Found an empty range"); 740 741 // Merge disjoint slots. 742 if (!First->overlaps(*Second)) { 743 Chanded = true; 744 First->MergeRangesInAsValue(*Second, First->getValNumInfo(0)); 745 SlotRemap[SecondSlot] = FirstSlot; 746 SortedSlots[J] = -1; 747 DEBUG(dbgs()<<"Merging #"<<FirstSlot<<" and slots #"<< 748 SecondSlot<<" together.\n"); 749 unsigned MaxAlignment = std::max(MFI->getObjectAlignment(FirstSlot), 750 MFI->getObjectAlignment(SecondSlot)); 751 752 assert(MFI->getObjectSize(FirstSlot) >= 753 MFI->getObjectSize(SecondSlot) && 754 "Merging a small object into a larger one"); 755 756 RemovedSlots+=1; 757 ReducedSize += MFI->getObjectSize(SecondSlot); 758 MFI->setObjectAlignment(FirstSlot, MaxAlignment); 759 MFI->RemoveStackObject(SecondSlot); 760 } 761 } 762 } 763 }// While changed. 764 765 // Record statistics. 766 StackSpaceSaved += ReducedSize; 767 StackSlotMerged += RemovedSlots; 768 DEBUG(dbgs()<<"Merge "<<RemovedSlots<<" slots. Saved "<< 769 ReducedSize<<" bytes\n"); 770 771 // Scan the entire function and update all machine operands that use frame 772 // indices to use the remapped frame index. 773 expungeSlotMap(SlotRemap, NumSlots); 774 remapInstructions(SlotRemap); 775 776 // Release the intervals. 777 for (unsigned I = 0; I < NumSlots; ++I) { 778 delete Intervals[I]; 779 } 780 781 return removeAllMarkers(); 782} 783