StackColoring.cpp revision 261abf5f4011e5b1e8949d7404190a4f4eaff8d8
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 typedef DenseMap<const MachineBasicBlock*, BlockLifetimeInfo> LivenessMap; 106 LivenessMap BlockLiveness; 107 108 /// Maps serial numbers to basic blocks. 109 DenseMap<const MachineBasicBlock*, int> BasicBlocks; 110 /// Maps basic blocks to a serial number. 111 SmallVector<const MachineBasicBlock*, 8> BasicBlockNumbering; 112 113 /// Maps liveness intervals for each slot. 114 SmallVector<LiveInterval*, 16> Intervals; 115 /// VNInfo is used for the construction of LiveIntervals. 116 VNInfo::Allocator VNInfoAllocator; 117 /// SlotIndex analysis object. 118 SlotIndexes *Indexes; 119 120 /// The list of lifetime markers found. These markers are to be removed 121 /// once the coloring is done. 122 SmallVector<MachineInstr*, 8> Markers; 123 124 /// SlotSizeSorter - A Sort utility for arranging stack slots according 125 /// to their size. 126 struct SlotSizeSorter { 127 MachineFrameInfo *MFI; 128 SlotSizeSorter(MachineFrameInfo *mfi) : MFI(mfi) { } 129 bool operator()(int LHS, int RHS) { 130 // We use -1 to denote a uninteresting slot. Place these slots at the end. 131 if (LHS == -1) return false; 132 if (RHS == -1) return true; 133 // Sort according to size. 134 return MFI->getObjectSize(LHS) > MFI->getObjectSize(RHS); 135 } 136}; 137 138public: 139 static char ID; 140 StackColoring() : MachineFunctionPass(ID) { 141 initializeStackColoringPass(*PassRegistry::getPassRegistry()); 142 } 143 void getAnalysisUsage(AnalysisUsage &AU) const; 144 bool runOnMachineFunction(MachineFunction &MF); 145 146private: 147 /// Debug. 148 void dump() const; 149 150 /// Removes all of the lifetime marker instructions from the function. 151 /// \returns true if any markers were removed. 152 bool removeAllMarkers(); 153 154 /// Scan the machine function and find all of the lifetime markers. 155 /// Record the findings in the BEGIN and END vectors. 156 /// \returns the number of markers found. 157 unsigned collectMarkers(unsigned NumSlot); 158 159 /// Perform the dataflow calculation and calculate the lifetime for each of 160 /// the slots, based on the BEGIN/END vectors. Set the LifetimeLIVE_IN and 161 /// LifetimeLIVE_OUT maps that represent which stack slots are live coming 162 /// in and out blocks. 163 void calculateLocalLiveness(); 164 165 /// Construct the LiveIntervals for the slots. 166 void calculateLiveIntervals(unsigned NumSlots); 167 168 /// Go over the machine function and change instructions which use stack 169 /// slots to use the joint slots. 170 void remapInstructions(DenseMap<int, int> &SlotRemap); 171 172 /// The input program may contain intructions which are not inside lifetime 173 /// markers. This can happen due to a bug in the compiler or due to a bug in 174 /// user code (for example, returning a reference to a local variable). 175 /// This procedure checks all of the instructions in the function and 176 /// invalidates lifetime ranges which do not contain all of the instructions 177 /// which access that frame slot. 178 void removeInvalidSlotRanges(); 179 180 /// Map entries which point to other entries to their destination. 181 /// A->B->C becomes A->C. 182 void expungeSlotMap(DenseMap<int, int> &SlotRemap, unsigned NumSlots); 183}; 184} // end anonymous namespace 185 186char StackColoring::ID = 0; 187char &llvm::StackColoringID = StackColoring::ID; 188 189INITIALIZE_PASS_BEGIN(StackColoring, 190 "stack-coloring", "Merge disjoint stack slots", false, false) 191INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) 192INITIALIZE_PASS_DEPENDENCY(SlotIndexes) 193INITIALIZE_PASS_END(StackColoring, 194 "stack-coloring", "Merge disjoint stack slots", false, false) 195 196void StackColoring::getAnalysisUsage(AnalysisUsage &AU) const { 197 AU.addRequired<MachineDominatorTree>(); 198 AU.addPreserved<MachineDominatorTree>(); 199 AU.addRequired<SlotIndexes>(); 200 MachineFunctionPass::getAnalysisUsage(AU); 201} 202 203void StackColoring::dump() const { 204 for (df_iterator<MachineFunction*> FI = df_begin(MF), FE = df_end(MF); 205 FI != FE; ++FI) { 206 DEBUG(dbgs()<<"Inspecting block #"<<BasicBlocks.lookup(*FI)<< 207 " ["<<FI->getName()<<"]\n"); 208 209 LivenessMap::const_iterator BI = BlockLiveness.find(*FI); 210 assert(BI != BlockLiveness.end() && "Block not found"); 211 const BlockLifetimeInfo &BlockInfo = BI->second; 212 213 DEBUG(dbgs()<<"BEGIN : {"); 214 for (unsigned i=0; i < BlockInfo.Begin.size(); ++i) 215 DEBUG(dbgs()<<BlockInfo.Begin.test(i)<<" "); 216 DEBUG(dbgs()<<"}\n"); 217 218 DEBUG(dbgs()<<"END : {"); 219 for (unsigned i=0; i < BlockInfo.End.size(); ++i) 220 DEBUG(dbgs()<<BlockInfo.End.test(i)<<" "); 221 222 DEBUG(dbgs()<<"}\n"); 223 224 DEBUG(dbgs()<<"LIVE_IN: {"); 225 for (unsigned i=0; i < BlockInfo.LiveIn.size(); ++i) 226 DEBUG(dbgs()<<BlockInfo.LiveIn.test(i)<<" "); 227 228 DEBUG(dbgs()<<"}\n"); 229 DEBUG(dbgs()<<"LIVEOUT: {"); 230 for (unsigned i=0; i < BlockInfo.LiveOut.size(); ++i) 231 DEBUG(dbgs()<<BlockInfo.LiveOut.test(i)<<" "); 232 DEBUG(dbgs()<<"}\n"); 233 } 234} 235 236unsigned StackColoring::collectMarkers(unsigned NumSlot) { 237 unsigned MarkersFound = 0; 238 // Scan the function to find all lifetime markers. 239 // NOTE: We use the a reverse-post-order iteration to ensure that we obtain a 240 // deterministic numbering, and because we'll need a post-order iteration 241 // later for solving the liveness dataflow problem. 242 for (df_iterator<MachineFunction*> FI = df_begin(MF), FE = df_end(MF); 243 FI != FE; ++FI) { 244 245 // Assign a serial number to this basic block. 246 BasicBlocks[*FI] = BasicBlockNumbering.size(); 247 BasicBlockNumbering.push_back(*FI); 248 249 // Keep a reference to avoid repeated lookups. 250 BlockLifetimeInfo &BlockInfo = BlockLiveness[*FI]; 251 252 BlockInfo.Begin.resize(NumSlot); 253 BlockInfo.End.resize(NumSlot); 254 255 for (MachineBasicBlock::iterator BI = (*FI)->begin(), BE = (*FI)->end(); 256 BI != BE; ++BI) { 257 258 if (BI->getOpcode() != TargetOpcode::LIFETIME_START && 259 BI->getOpcode() != TargetOpcode::LIFETIME_END) 260 continue; 261 262 Markers.push_back(BI); 263 264 bool IsStart = BI->getOpcode() == TargetOpcode::LIFETIME_START; 265 const MachineOperand &MI = BI->getOperand(0); 266 unsigned Slot = MI.getIndex(); 267 268 MarkersFound++; 269 270 const AllocaInst *Allocation = MFI->getObjectAllocation(Slot); 271 if (Allocation) { 272 DEBUG(dbgs()<<"Found a lifetime marker for slot #"<<Slot<< 273 " with allocation: "<< Allocation->getName()<<"\n"); 274 } 275 276 if (IsStart) { 277 BlockInfo.Begin.set(Slot); 278 } else { 279 if (BlockInfo.Begin.test(Slot)) { 280 // Allocas that start and end within a single block are handled 281 // specially when computing the LiveIntervals to avoid pessimizing 282 // the liveness propagation. 283 BlockInfo.Begin.reset(Slot); 284 } else { 285 BlockInfo.End.set(Slot); 286 } 287 } 288 } 289 } 290 291 // Update statistics. 292 NumMarkerSeen += MarkersFound; 293 return MarkersFound; 294} 295 296void StackColoring::calculateLocalLiveness() { 297 // Perform a standard reverse dataflow computation to solve for 298 // global liveness. The BEGIN set here is equivalent to KILL in the standard 299 // formulation, and END is equivalent to GEN. The result of this computation 300 // is a map from blocks to bitvectors where the bitvectors represent which 301 // allocas are live in/out of that block. 302 SmallPtrSet<const MachineBasicBlock*, 8> BBSet(BasicBlockNumbering.begin(), 303 BasicBlockNumbering.end()); 304 unsigned NumSSMIters = 0; 305 bool changed = true; 306 while (changed) { 307 changed = false; 308 ++NumSSMIters; 309 310 SmallPtrSet<const MachineBasicBlock*, 8> NextBBSet; 311 312 for (SmallVector<const MachineBasicBlock*, 8>::iterator 313 PI = BasicBlockNumbering.begin(), PE = BasicBlockNumbering.end(); 314 PI != PE; ++PI) { 315 316 const MachineBasicBlock *BB = *PI; 317 if (!BBSet.count(BB)) continue; 318 319 // Use an iterator to avoid repeated lookups. 320 LivenessMap::iterator BI = BlockLiveness.find(BB); 321 assert(BI != BlockLiveness.end() && "Block not found"); 322 BlockLifetimeInfo &BlockInfo = BI->second; 323 324 BitVector LocalLiveIn; 325 BitVector LocalLiveOut; 326 327 // Forward propagation from begins to ends. 328 for (MachineBasicBlock::const_pred_iterator PI = BB->pred_begin(), 329 PE = BB->pred_end(); PI != PE; ++PI) { 330 LivenessMap::const_iterator I = BlockLiveness.find(*PI); 331 assert(I != BlockLiveness.end() && "Predecessor not found"); 332 LocalLiveIn |= I->second.LiveOut; 333 } 334 LocalLiveIn |= BlockInfo.End; 335 LocalLiveIn.reset(BlockInfo.Begin); 336 337 // Reverse propagation from ends to begins. 338 for (MachineBasicBlock::const_succ_iterator SI = BB->succ_begin(), 339 SE = BB->succ_end(); SI != SE; ++SI) { 340 LivenessMap::const_iterator I = BlockLiveness.find(*SI); 341 assert(I != BlockLiveness.end() && "Successor not found"); 342 LocalLiveOut |= I->second.LiveIn; 343 } 344 LocalLiveOut |= BlockInfo.Begin; 345 LocalLiveOut.reset(BlockInfo.End); 346 347 LocalLiveIn |= LocalLiveOut; 348 LocalLiveOut |= LocalLiveIn; 349 350 // After adopting the live bits, we need to turn-off the bits which 351 // are de-activated in this block. 352 LocalLiveOut.reset(BlockInfo.End); 353 LocalLiveIn.reset(BlockInfo.Begin); 354 355 // If we have both BEGIN and END markers in the same basic block then 356 // we know that the BEGIN marker comes after the END, because we already 357 // handle the case where the BEGIN comes before the END when collecting 358 // the markers (and building the BEGIN/END vectore). 359 // Want to enable the LIVE_IN and LIVE_OUT of slots that have both 360 // BEGIN and END because it means that the value lives before and after 361 // this basic block. 362 BitVector LocalEndBegin = BlockInfo.End; 363 LocalEndBegin &= BlockInfo.Begin; 364 LocalLiveIn |= LocalEndBegin; 365 LocalLiveOut |= LocalEndBegin; 366 367 if (LocalLiveIn.test(BlockInfo.LiveIn)) { 368 changed = true; 369 BlockInfo.LiveIn |= LocalLiveIn; 370 371 for (MachineBasicBlock::const_pred_iterator PI = BB->pred_begin(), 372 PE = BB->pred_end(); PI != PE; ++PI) 373 NextBBSet.insert(*PI); 374 } 375 376 if (LocalLiveOut.test(BlockInfo.LiveOut)) { 377 changed = true; 378 BlockInfo.LiveOut |= LocalLiveOut; 379 380 for (MachineBasicBlock::const_succ_iterator SI = BB->succ_begin(), 381 SE = BB->succ_end(); SI != SE; ++SI) 382 NextBBSet.insert(*SI); 383 } 384 } 385 386 BBSet = NextBBSet; 387 }// while changed. 388} 389 390void StackColoring::calculateLiveIntervals(unsigned NumSlots) { 391 SmallVector<SlotIndex, 16> Starts; 392 SmallVector<SlotIndex, 16> Finishes; 393 394 // For each block, find which slots are active within this block 395 // and update the live intervals. 396 for (MachineFunction::iterator MBB = MF->begin(), MBBe = MF->end(); 397 MBB != MBBe; ++MBB) { 398 Starts.clear(); 399 Starts.resize(NumSlots); 400 Finishes.clear(); 401 Finishes.resize(NumSlots); 402 403 // Create the interval for the basic blocks with lifetime markers in them. 404 for (SmallVectorImpl<MachineInstr*>::const_iterator it = Markers.begin(), 405 e = Markers.end(); it != e; ++it) { 406 const MachineInstr *MI = *it; 407 if (MI->getParent() != MBB) 408 continue; 409 410 assert((MI->getOpcode() == TargetOpcode::LIFETIME_START || 411 MI->getOpcode() == TargetOpcode::LIFETIME_END) && 412 "Invalid Lifetime marker"); 413 414 bool IsStart = MI->getOpcode() == TargetOpcode::LIFETIME_START; 415 const MachineOperand &Mo = MI->getOperand(0); 416 int Slot = Mo.getIndex(); 417 assert(Slot >= 0 && "Invalid slot"); 418 419 SlotIndex ThisIndex = Indexes->getInstructionIndex(MI); 420 421 if (IsStart) { 422 if (!Starts[Slot].isValid() || Starts[Slot] > ThisIndex) 423 Starts[Slot] = ThisIndex; 424 } else { 425 if (!Finishes[Slot].isValid() || Finishes[Slot] < ThisIndex) 426 Finishes[Slot] = ThisIndex; 427 } 428 } 429 430 // Create the interval of the blocks that we previously found to be 'alive'. 431 BitVector Alive = BlockLiveness[MBB].LiveIn; 432 Alive |= BlockLiveness[MBB].LiveOut; 433 434 if (Alive.any()) { 435 for (int pos = Alive.find_first(); pos != -1; 436 pos = Alive.find_next(pos)) { 437 if (!Starts[pos].isValid()) 438 Starts[pos] = Indexes->getMBBStartIdx(MBB); 439 if (!Finishes[pos].isValid()) 440 Finishes[pos] = Indexes->getMBBEndIdx(MBB); 441 } 442 } 443 444 for (unsigned i = 0; i < NumSlots; ++i) { 445 assert(Starts[i].isValid() == Finishes[i].isValid() && "Unmatched range"); 446 if (!Starts[i].isValid()) 447 continue; 448 449 assert(Starts[i] && Finishes[i] && "Invalid interval"); 450 VNInfo *ValNum = Intervals[i]->getValNumInfo(0); 451 SlotIndex S = Starts[i]; 452 SlotIndex F = Finishes[i]; 453 if (S < F) { 454 // We have a single consecutive region. 455 Intervals[i]->addRange(LiveRange(S, F, ValNum)); 456 } else { 457 // We have two non consecutive regions. This happens when 458 // LIFETIME_START appears after the LIFETIME_END marker. 459 SlotIndex NewStart = Indexes->getMBBStartIdx(MBB); 460 SlotIndex NewFin = Indexes->getMBBEndIdx(MBB); 461 Intervals[i]->addRange(LiveRange(NewStart, F, ValNum)); 462 Intervals[i]->addRange(LiveRange(S, NewFin, ValNum)); 463 } 464 } 465 } 466} 467 468bool StackColoring::removeAllMarkers() { 469 unsigned Count = 0; 470 for (unsigned i = 0; i < Markers.size(); ++i) { 471 Markers[i]->eraseFromParent(); 472 Count++; 473 } 474 Markers.clear(); 475 476 DEBUG(dbgs()<<"Removed "<<Count<<" markers.\n"); 477 return Count; 478} 479 480void StackColoring::remapInstructions(DenseMap<int, int> &SlotRemap) { 481 unsigned FixedInstr = 0; 482 unsigned FixedMemOp = 0; 483 unsigned FixedDbg = 0; 484 MachineModuleInfo *MMI = &MF->getMMI(); 485 486 // Remap debug information that refers to stack slots. 487 MachineModuleInfo::VariableDbgInfoMapTy &VMap = MMI->getVariableDbgInfo(); 488 for (MachineModuleInfo::VariableDbgInfoMapTy::iterator VI = VMap.begin(), 489 VE = VMap.end(); VI != VE; ++VI) { 490 const MDNode *Var = VI->first; 491 if (!Var) continue; 492 std::pair<unsigned, DebugLoc> &VP = VI->second; 493 if (SlotRemap.count(VP.first)) { 494 DEBUG(dbgs()<<"Remapping debug info for ["<<Var->getName()<<"].\n"); 495 VP.first = SlotRemap[VP.first]; 496 FixedDbg++; 497 } 498 } 499 500 // Keep a list of *allocas* which need to be remapped. 501 DenseMap<const AllocaInst*, const AllocaInst*> Allocas; 502 for (DenseMap<int, int>::const_iterator it = SlotRemap.begin(), 503 e = SlotRemap.end(); it != e; ++it) { 504 const AllocaInst *From = MFI->getObjectAllocation(it->first); 505 const AllocaInst *To = MFI->getObjectAllocation(it->second); 506 assert(To && From && "Invalid allocation object"); 507 Allocas[From] = To; 508 } 509 510 // Remap all instructions to the new stack slots. 511 MachineFunction::iterator BB, BBE; 512 MachineBasicBlock::iterator I, IE; 513 for (BB = MF->begin(), BBE = MF->end(); BB != BBE; ++BB) 514 for (I = BB->begin(), IE = BB->end(); I != IE; ++I) { 515 516 // Skip lifetime markers. We'll remove them soon. 517 if (I->getOpcode() == TargetOpcode::LIFETIME_START || 518 I->getOpcode() == TargetOpcode::LIFETIME_END) 519 continue; 520 521 // Update the MachineMemOperand to use the new alloca. 522 for (MachineInstr::mmo_iterator MM = I->memoperands_begin(), 523 E = I->memoperands_end(); MM != E; ++MM) { 524 MachineMemOperand *MMO = *MM; 525 526 const Value *V = MMO->getValue(); 527 528 if (!V) 529 continue; 530 531 // Climb up and find the original alloca. 532 V = GetUnderlyingObject(V); 533 // If we did not find one, or if the one that we found is not in our 534 // map, then move on. 535 if (!V || !isa<AllocaInst>(V)) { 536 // Clear mem operand since we don't know for sure that it doesn't 537 // alias a merged alloca. 538 MMO->setValue(0); 539 continue; 540 } 541 const AllocaInst *AI= cast<AllocaInst>(V); 542 if (!Allocas.count(AI)) 543 continue; 544 545 MMO->setValue(Allocas[AI]); 546 FixedMemOp++; 547 } 548 549 // Update all of the machine instruction operands. 550 for (unsigned i = 0 ; i < I->getNumOperands(); ++i) { 551 MachineOperand &MO = I->getOperand(i); 552 553 if (!MO.isFI()) 554 continue; 555 int FromSlot = MO.getIndex(); 556 557 // Don't touch arguments. 558 if (FromSlot<0) 559 continue; 560 561 // Only look at mapped slots. 562 if (!SlotRemap.count(FromSlot)) 563 continue; 564 565 // In a debug build, check that the instruction that we are modifying is 566 // inside the expected live range. If the instruction is not inside 567 // the calculated range then it means that the alloca usage moved 568 // outside of the lifetime markers, or that the user has a bug. 569 // NOTE: Alloca address calculations which happen outside the lifetime 570 // zone are are okay, despite the fact that we don't have a good way 571 // for validating all of the usages of the calculation. 572#ifndef NDEBUG 573 bool TouchesMemory = I->mayLoad() || I->mayStore(); 574 // If we *don't* protect the user from escaped allocas, don't bother 575 // validating the instructions. 576 if (!I->isDebugValue() && TouchesMemory && ProtectFromEscapedAllocas) { 577 SlotIndex Index = Indexes->getInstructionIndex(I); 578 LiveInterval *Interval = Intervals[FromSlot]; 579 assert(Interval->find(Index) != Interval->end() && 580 "Found instruction usage outside of live range."); 581 } 582#endif 583 584 // Fix the machine instructions. 585 int ToSlot = SlotRemap[FromSlot]; 586 MO.setIndex(ToSlot); 587 FixedInstr++; 588 } 589 } 590 591 DEBUG(dbgs()<<"Fixed "<<FixedMemOp<<" machine memory operands.\n"); 592 DEBUG(dbgs()<<"Fixed "<<FixedDbg<<" debug locations.\n"); 593 DEBUG(dbgs()<<"Fixed "<<FixedInstr<<" machine instructions.\n"); 594} 595 596void StackColoring::removeInvalidSlotRanges() { 597 MachineFunction::const_iterator BB, BBE; 598 MachineBasicBlock::const_iterator I, IE; 599 for (BB = MF->begin(), BBE = MF->end(); BB != BBE; ++BB) 600 for (I = BB->begin(), IE = BB->end(); I != IE; ++I) { 601 602 if (I->getOpcode() == TargetOpcode::LIFETIME_START || 603 I->getOpcode() == TargetOpcode::LIFETIME_END || I->isDebugValue()) 604 continue; 605 606 // Some intervals are suspicious! In some cases we find address 607 // calculations outside of the lifetime zone, but not actual memory 608 // read or write. Memory accesses outside of the lifetime zone are a clear 609 // violation, but address calculations are okay. This can happen when 610 // GEPs are hoisted outside of the lifetime zone. 611 // So, in here we only check instructions which can read or write memory. 612 if (!I->mayLoad() && !I->mayStore()) 613 continue; 614 615 // Check all of the machine operands. 616 for (unsigned i = 0 ; i < I->getNumOperands(); ++i) { 617 const MachineOperand &MO = I->getOperand(i); 618 619 if (!MO.isFI()) 620 continue; 621 622 int Slot = MO.getIndex(); 623 624 if (Slot<0) 625 continue; 626 627 if (Intervals[Slot]->empty()) 628 continue; 629 630 // Check that the used slot is inside the calculated lifetime range. 631 // If it is not, warn about it and invalidate the range. 632 LiveInterval *Interval = Intervals[Slot]; 633 SlotIndex Index = Indexes->getInstructionIndex(I); 634 if (Interval->find(Index) == Interval->end()) { 635 Intervals[Slot]->clear(); 636 DEBUG(dbgs()<<"Invalidating range #"<<Slot<<"\n"); 637 EscapedAllocas++; 638 } 639 } 640 } 641} 642 643void StackColoring::expungeSlotMap(DenseMap<int, int> &SlotRemap, 644 unsigned NumSlots) { 645 // Expunge slot remap map. 646 for (unsigned i=0; i < NumSlots; ++i) { 647 // If we are remapping i 648 if (SlotRemap.count(i)) { 649 int Target = SlotRemap[i]; 650 // As long as our target is mapped to something else, follow it. 651 while (SlotRemap.count(Target)) { 652 Target = SlotRemap[Target]; 653 SlotRemap[i] = Target; 654 } 655 } 656 } 657} 658 659bool StackColoring::runOnMachineFunction(MachineFunction &Func) { 660 DEBUG(dbgs() << "********** Stack Coloring **********\n" 661 << "********** Function: " 662 << ((const Value*)Func.getFunction())->getName() << '\n'); 663 MF = &Func; 664 MFI = MF->getFrameInfo(); 665 Indexes = &getAnalysis<SlotIndexes>(); 666 BlockLiveness.clear(); 667 BasicBlocks.clear(); 668 BasicBlockNumbering.clear(); 669 Markers.clear(); 670 Intervals.clear(); 671 VNInfoAllocator.Reset(); 672 673 unsigned NumSlots = MFI->getObjectIndexEnd(); 674 675 // If there are no stack slots then there are no markers to remove. 676 if (!NumSlots) 677 return false; 678 679 SmallVector<int, 8> SortedSlots; 680 681 SortedSlots.reserve(NumSlots); 682 Intervals.reserve(NumSlots); 683 684 unsigned NumMarkers = collectMarkers(NumSlots); 685 686 unsigned TotalSize = 0; 687 DEBUG(dbgs()<<"Found "<<NumMarkers<<" markers and "<<NumSlots<<" slots\n"); 688 DEBUG(dbgs()<<"Slot structure:\n"); 689 690 for (int i=0; i < MFI->getObjectIndexEnd(); ++i) { 691 DEBUG(dbgs()<<"Slot #"<<i<<" - "<<MFI->getObjectSize(i)<<" bytes.\n"); 692 TotalSize += MFI->getObjectSize(i); 693 } 694 695 DEBUG(dbgs()<<"Total Stack size: "<<TotalSize<<" bytes\n\n"); 696 697 // Don't continue because there are not enough lifetime markers, or the 698 // stack is too small, or we are told not to optimize the slots. 699 if (NumMarkers < 2 || TotalSize < 16 || DisableColoring) { 700 DEBUG(dbgs()<<"Will not try to merge slots.\n"); 701 return removeAllMarkers(); 702 } 703 704 for (unsigned i=0; i < NumSlots; ++i) { 705 LiveInterval *LI = new LiveInterval(i, 0); 706 Intervals.push_back(LI); 707 LI->getNextValue(Indexes->getZeroIndex(), VNInfoAllocator); 708 SortedSlots.push_back(i); 709 } 710 711 // Calculate the liveness of each block. 712 calculateLocalLiveness(); 713 714 // Propagate the liveness information. 715 calculateLiveIntervals(NumSlots); 716 717 // Search for allocas which are used outside of the declared lifetime 718 // markers. 719 if (ProtectFromEscapedAllocas) 720 removeInvalidSlotRanges(); 721 722 // Maps old slots to new slots. 723 DenseMap<int, int> SlotRemap; 724 unsigned RemovedSlots = 0; 725 unsigned ReducedSize = 0; 726 727 // Do not bother looking at empty intervals. 728 for (unsigned I = 0; I < NumSlots; ++I) { 729 if (Intervals[SortedSlots[I]]->empty()) 730 SortedSlots[I] = -1; 731 } 732 733 // This is a simple greedy algorithm for merging allocas. First, sort the 734 // slots, placing the largest slots first. Next, perform an n^2 scan and look 735 // for disjoint slots. When you find disjoint slots, merge the samller one 736 // into the bigger one and update the live interval. Remove the small alloca 737 // and continue. 738 739 // Sort the slots according to their size. Place unused slots at the end. 740 // Use stable sort to guarantee deterministic code generation. 741 std::stable_sort(SortedSlots.begin(), SortedSlots.end(), 742 SlotSizeSorter(MFI)); 743 744 bool Chanded = true; 745 while (Chanded) { 746 Chanded = false; 747 for (unsigned I = 0; I < NumSlots; ++I) { 748 if (SortedSlots[I] == -1) 749 continue; 750 751 for (unsigned J=I+1; J < NumSlots; ++J) { 752 if (SortedSlots[J] == -1) 753 continue; 754 755 int FirstSlot = SortedSlots[I]; 756 int SecondSlot = SortedSlots[J]; 757 LiveInterval *First = Intervals[FirstSlot]; 758 LiveInterval *Second = Intervals[SecondSlot]; 759 assert (!First->empty() && !Second->empty() && "Found an empty range"); 760 761 // Merge disjoint slots. 762 if (!First->overlaps(*Second)) { 763 Chanded = true; 764 First->MergeRangesInAsValue(*Second, First->getValNumInfo(0)); 765 SlotRemap[SecondSlot] = FirstSlot; 766 SortedSlots[J] = -1; 767 DEBUG(dbgs()<<"Merging #"<<FirstSlot<<" and slots #"<< 768 SecondSlot<<" together.\n"); 769 unsigned MaxAlignment = std::max(MFI->getObjectAlignment(FirstSlot), 770 MFI->getObjectAlignment(SecondSlot)); 771 772 assert(MFI->getObjectSize(FirstSlot) >= 773 MFI->getObjectSize(SecondSlot) && 774 "Merging a small object into a larger one"); 775 776 RemovedSlots+=1; 777 ReducedSize += MFI->getObjectSize(SecondSlot); 778 MFI->setObjectAlignment(FirstSlot, MaxAlignment); 779 MFI->RemoveStackObject(SecondSlot); 780 } 781 } 782 } 783 }// While changed. 784 785 // Record statistics. 786 StackSpaceSaved += ReducedSize; 787 StackSlotMerged += RemovedSlots; 788 DEBUG(dbgs()<<"Merge "<<RemovedSlots<<" slots. Saved "<< 789 ReducedSize<<" bytes\n"); 790 791 // Scan the entire function and update all machine operands that use frame 792 // indices to use the remapped frame index. 793 expungeSlotMap(SlotRemap, NumSlots); 794 remapInstructions(SlotRemap); 795 796 // Release the intervals. 797 for (unsigned I = 0; I < NumSlots; ++I) { 798 delete Intervals[I]; 799 } 800 801 return removeAllMarkers(); 802} 803