StackColoring.cpp revision 0cd19b93017fcaba737b15ea4da39c460feb5670
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 "MachineTraceMetrics.h" 26#include "llvm/Function.h" 27#include "llvm/Module.h" 28#include "llvm/ADT/BitVector.h" 29#include "llvm/Analysis/Dominators.h" 30#include "llvm/Analysis/ValueTracking.h" 31#include "llvm/ADT/DepthFirstIterator.h" 32#include "llvm/ADT/PostOrderIterator.h" 33#include "llvm/ADT/SetVector.h" 34#include "llvm/ADT/SmallPtrSet.h" 35#include "llvm/ADT/SparseSet.h" 36#include "llvm/ADT/Statistic.h" 37#include "llvm/CodeGen/LiveInterval.h" 38#include "llvm/CodeGen/MachineLoopInfo.h" 39#include "llvm/CodeGen/MachineBranchProbabilityInfo.h" 40#include "llvm/CodeGen/MachineDominators.h" 41#include "llvm/CodeGen/MachineBasicBlock.h" 42#include "llvm/CodeGen/MachineFunctionPass.h" 43#include "llvm/CodeGen/MachineLoopInfo.h" 44#include "llvm/CodeGen/MachineModuleInfo.h" 45#include "llvm/CodeGen/MachineRegisterInfo.h" 46#include "llvm/CodeGen/MachineFrameInfo.h" 47#include "llvm/CodeGen/MachineMemOperand.h" 48#include "llvm/CodeGen/Passes.h" 49#include "llvm/CodeGen/SlotIndexes.h" 50#include "llvm/DebugInfo.h" 51#include "llvm/MC/MCInstrItineraries.h" 52#include "llvm/Target/TargetInstrInfo.h" 53#include "llvm/Target/TargetRegisterInfo.h" 54#include "llvm/Support/CommandLine.h" 55#include "llvm/Support/Debug.h" 56#include "llvm/Support/raw_ostream.h" 57 58using namespace llvm; 59 60static cl::opt<bool> 61DisableColoring("no-stack-coloring", 62 cl::init(false), cl::Hidden, 63 cl::desc("Disable stack coloring")); 64 65static cl::opt<bool> 66CheckEscapedAllocas("stack-coloring-check-escaped", 67 cl::init(true), cl::Hidden, 68 cl::desc("Look for allocas which escaped the lifetime region")); 69 70STATISTIC(NumMarkerSeen, "Number of lifetime markers found."); 71STATISTIC(StackSpaceSaved, "Number of bytes saved due to merging slots."); 72STATISTIC(StackSlotMerged, "Number of stack slot merged."); 73STATISTIC(EscapedAllocas, 74 "Number of allocas that escaped the lifetime region"); 75 76//===----------------------------------------------------------------------===// 77// StackColoring Pass 78//===----------------------------------------------------------------------===// 79 80namespace { 81/// StackColoring - A machine pass for merging disjoint stack allocations, 82/// marked by the LIFETIME_START and LIFETIME_END pseudo instructions. 83class StackColoring : public MachineFunctionPass { 84 MachineFrameInfo *MFI; 85 MachineFunction *MF; 86 87 /// A class representing liveness information for a single basic block. 88 /// Each bit in the BitVector represents the liveness property 89 /// for a different stack slot. 90 struct BlockLifetimeInfo { 91 /// Which slots BEGINs in each basic block. 92 BitVector Begin; 93 /// Which slots ENDs in each basic block. 94 BitVector End; 95 /// Which slots are marked as LIVE_IN, coming into each basic block. 96 BitVector LiveIn; 97 /// Which slots are marked as LIVE_OUT, coming out of each basic block. 98 BitVector LiveOut; 99 }; 100 101 /// Maps active slots (per bit) for each basic block. 102 DenseMap<MachineBasicBlock*, BlockLifetimeInfo> BlockLiveness; 103 104 /// Maps serial numbers to basic blocks. 105 DenseMap<MachineBasicBlock*, int> BasicBlocks; 106 /// Maps basic blocks to a serial number. 107 SmallVector<MachineBasicBlock*, 8> BasicBlockNumbering; 108 109 /// Maps liveness intervals for each slot. 110 SmallVector<LiveInterval*, 16> Intervals; 111 /// VNInfo is used for the construction of LiveIntervals. 112 VNInfo::Allocator VNInfoAllocator; 113 /// SlotIndex analysis object. 114 SlotIndexes *Indexes; 115 116 /// The list of lifetime markers found. These markers are to be removed 117 /// once the coloring is done. 118 SmallVector<MachineInstr*, 8> Markers; 119 120 /// SlotSizeSorter - A Sort utility for arranging stack slots according 121 /// to their size. 122 struct SlotSizeSorter { 123 MachineFrameInfo *MFI; 124 SlotSizeSorter(MachineFrameInfo *mfi) : MFI(mfi) { } 125 bool operator()(int LHS, int RHS) { 126 // We use -1 to denote a uninteresting slot. Place these slots at the end. 127 if (LHS == -1) return false; 128 if (RHS == -1) return true; 129 // Sort according to size. 130 return MFI->getObjectSize(LHS) > MFI->getObjectSize(RHS); 131 } 132}; 133 134public: 135 static char ID; 136 StackColoring() : MachineFunctionPass(ID) { 137 initializeStackColoringPass(*PassRegistry::getPassRegistry()); 138 } 139 void getAnalysisUsage(AnalysisUsage &AU) const; 140 bool runOnMachineFunction(MachineFunction &MF); 141 142private: 143 /// Debug. 144 void dump(); 145 146 /// Removes all of the lifetime marker instructions from the function. 147 /// \returns true if any markers were removed. 148 bool removeAllMarkers(); 149 150 /// Scan the machine function and find all of the lifetime markers. 151 /// Record the findings in the BEGIN and END vectors. 152 /// \returns the number of markers found. 153 unsigned collectMarkers(unsigned NumSlot); 154 155 /// Perform the dataflow calculation and calculate the lifetime for each of 156 /// the slots, based on the BEGIN/END vectors. Set the LifetimeLIVE_IN and 157 /// LifetimeLIVE_OUT maps that represent which stack slots are live coming 158 /// in and out blocks. 159 void calculateLocalLiveness(); 160 161 /// Construct the LiveIntervals for the slots. 162 void calculateLiveIntervals(unsigned NumSlots); 163 164 /// Go over the machine function and change instructions which use stack 165 /// slots to use the joint slots. 166 void remapInstructions(DenseMap<int, int> &SlotRemap); 167 168 /// The input program may contain intructions which are not inside lifetime 169 /// markers. This can happen due to a bug in the compiler or due to a bug in 170 /// user code (for example, returning a reference to a local variable). 171 /// This procedure checks all of the instructions in the function and 172 /// invalidates lifetime ranges which do not contain all of the instructions 173 /// which access that frame slot. 174 void removeInvalidSlotRanges(); 175 176 /// Map entries which point to other entries to their destination. 177 /// A->B->C becomes A->C. 178 void expungeSlotMap(DenseMap<int, int> &SlotRemap, unsigned NumSlots); 179}; 180} // end anonymous namespace 181 182char StackColoring::ID = 0; 183char &llvm::StackColoringID = StackColoring::ID; 184 185INITIALIZE_PASS_BEGIN(StackColoring, 186 "stack-coloring", "Merge disjoint stack slots", false, false) 187INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) 188INITIALIZE_PASS_DEPENDENCY(SlotIndexes) 189INITIALIZE_PASS_END(StackColoring, 190 "stack-coloring", "Merge disjoint stack slots", false, false) 191 192void StackColoring::getAnalysisUsage(AnalysisUsage &AU) const { 193 AU.addRequired<MachineDominatorTree>(); 194 AU.addPreserved<MachineDominatorTree>(); 195 AU.addRequired<SlotIndexes>(); 196 MachineFunctionPass::getAnalysisUsage(AU); 197} 198 199void StackColoring::dump() { 200 for (df_iterator<MachineFunction*> FI = df_begin(MF), FE = df_end(MF); 201 FI != FE; ++FI) { 202 unsigned Num = BasicBlocks[*FI]; 203 DEBUG(dbgs()<<"Inspecting block #"<<Num<<" ["<<FI->getName()<<"]\n"); 204 Num = 0; 205 DEBUG(dbgs()<<"BEGIN : {"); 206 for (unsigned i=0; i < BlockLiveness[*FI].Begin.size(); ++i) 207 DEBUG(dbgs()<<BlockLiveness[*FI].Begin.test(i)<<" "); 208 DEBUG(dbgs()<<"}\n"); 209 210 DEBUG(dbgs()<<"END : {"); 211 for (unsigned i=0; i < BlockLiveness[*FI].End.size(); ++i) 212 DEBUG(dbgs()<<BlockLiveness[*FI].End.test(i)<<" "); 213 214 DEBUG(dbgs()<<"}\n"); 215 216 DEBUG(dbgs()<<"LIVE_IN: {"); 217 for (unsigned i=0; i < BlockLiveness[*FI].LiveIn.size(); ++i) 218 DEBUG(dbgs()<<BlockLiveness[*FI].LiveIn.test(i)<<" "); 219 220 DEBUG(dbgs()<<"}\n"); 221 DEBUG(dbgs()<<"LIVEOUT: {"); 222 for (unsigned i=0; i < BlockLiveness[*FI].LiveOut.size(); ++i) 223 DEBUG(dbgs()<<BlockLiveness[*FI].LiveOut.test(i)<<" "); 224 DEBUG(dbgs()<<"}\n"); 225 } 226} 227 228unsigned StackColoring::collectMarkers(unsigned NumSlot) { 229 unsigned MarkersFound = 0; 230 // Scan the function to find all lifetime markers. 231 // NOTE: We use the a reverse-post-order iteration to ensure that we obtain a 232 // deterministic numbering, and because we'll need a post-order iteration 233 // later for solving the liveness dataflow problem. 234 for (df_iterator<MachineFunction*> FI = df_begin(MF), FE = df_end(MF); 235 FI != FE; ++FI) { 236 237 // Assign a serial number to this basic block. 238 BasicBlocks[*FI] = BasicBlockNumbering.size(); 239 BasicBlockNumbering.push_back(*FI); 240 241 BlockLiveness[*FI].Begin.resize(NumSlot); 242 BlockLiveness[*FI].End.resize(NumSlot); 243 244 for (MachineBasicBlock::iterator BI = (*FI)->begin(), BE = (*FI)->end(); 245 BI != BE; ++BI) { 246 247 if (BI->getOpcode() != TargetOpcode::LIFETIME_START && 248 BI->getOpcode() != TargetOpcode::LIFETIME_END) 249 continue; 250 251 Markers.push_back(BI); 252 253 bool IsStart = BI->getOpcode() == TargetOpcode::LIFETIME_START; 254 MachineOperand &MI = BI->getOperand(0); 255 unsigned Slot = MI.getIndex(); 256 257 MarkersFound++; 258 259 const Value *Allocation = MFI->getObjectAllocation(Slot); 260 if (Allocation) { 261 DEBUG(dbgs()<<"Found a lifetime marker for slot #"<<Slot<< 262 " with allocation: "<< Allocation->getName()<<"\n"); 263 } 264 265 if (IsStart) { 266 BlockLiveness[*FI].Begin.set(Slot); 267 } else { 268 if (BlockLiveness[*FI].Begin.test(Slot)) { 269 // Allocas that start and end within a single block are handled 270 // specially when computing the LiveIntervals to avoid pessimizing 271 // the liveness propagation. 272 BlockLiveness[*FI].Begin.reset(Slot); 273 } else { 274 BlockLiveness[*FI].End.set(Slot); 275 } 276 } 277 } 278 } 279 280 // Update statistics. 281 NumMarkerSeen += MarkersFound; 282 return MarkersFound; 283} 284 285void StackColoring::calculateLocalLiveness() { 286 // Perform a standard reverse dataflow computation to solve for 287 // global liveness. The BEGIN set here is equivalent to KILL in the standard 288 // formulation, and END is equivalent to GEN. The result of this computation 289 // is a map from blocks to bitvectors where the bitvectors represent which 290 // allocas are live in/out of that block. 291 SmallPtrSet<MachineBasicBlock*, 8> BBSet(BasicBlockNumbering.begin(), 292 BasicBlockNumbering.end()); 293 unsigned NumSSMIters = 0; 294 bool changed = true; 295 while (changed) { 296 changed = false; 297 ++NumSSMIters; 298 299 SmallPtrSet<MachineBasicBlock*, 8> NextBBSet; 300 301 for (SmallVector<MachineBasicBlock*, 8>::iterator 302 PI = BasicBlockNumbering.begin(), PE = BasicBlockNumbering.end(); 303 PI != PE; ++PI) { 304 305 MachineBasicBlock *BB = *PI; 306 if (!BBSet.count(BB)) continue; 307 308 BitVector LocalLiveIn; 309 BitVector LocalLiveOut; 310 311 // Forward propagation from begins to ends. 312 for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(), 313 PE = BB->pred_end(); PI != PE; ++PI) 314 LocalLiveIn |= BlockLiveness[*PI].LiveOut; 315 LocalLiveIn |= BlockLiveness[BB].End; 316 LocalLiveIn.reset(BlockLiveness[BB].Begin); 317 318 // Reverse propagation from ends to begins. 319 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 320 SE = BB->succ_end(); SI != SE; ++SI) 321 LocalLiveOut |= BlockLiveness[*SI].LiveIn; 322 LocalLiveOut |= BlockLiveness[BB].Begin; 323 LocalLiveOut.reset(BlockLiveness[BB].End); 324 325 LocalLiveIn |= LocalLiveOut; 326 LocalLiveOut |= LocalLiveIn; 327 328 // After adopting the live bits, we need to turn-off the bits which 329 // are de-activated in this block. 330 LocalLiveOut.reset(BlockLiveness[BB].End); 331 LocalLiveIn.reset(BlockLiveness[BB].Begin); 332 333 // If we have both BEGIN and END markers in the same basic block then 334 // we know that the BEGIN marker comes after the END, because we already 335 // handle the case where the BEGIN comes before the END when collecting 336 // the markers (and building the BEGIN/END vectore). 337 // Want to enable the LIVE_IN and LIVE_OUT of slots that have both 338 // BEGIN and END because it means that the value lives before and after 339 // this basic block. 340 BitVector LocalEndBegin = BlockLiveness[BB].End; 341 LocalEndBegin &= BlockLiveness[BB].Begin; 342 LocalLiveIn |= LocalEndBegin; 343 LocalLiveOut |= LocalEndBegin; 344 345 if (LocalLiveIn.test(BlockLiveness[BB].LiveIn)) { 346 changed = true; 347 BlockLiveness[BB].LiveIn |= LocalLiveIn; 348 349 for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(), 350 PE = BB->pred_end(); PI != PE; ++PI) 351 NextBBSet.insert(*PI); 352 } 353 354 if (LocalLiveOut.test(BlockLiveness[BB].LiveOut)) { 355 changed = true; 356 BlockLiveness[BB].LiveOut |= LocalLiveOut; 357 358 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 359 SE = BB->succ_end(); SI != SE; ++SI) 360 NextBBSet.insert(*SI); 361 } 362 } 363 364 BBSet = NextBBSet; 365 }// while changed. 366} 367 368void StackColoring::calculateLiveIntervals(unsigned NumSlots) { 369 SmallVector<SlotIndex, 16> Starts; 370 SmallVector<SlotIndex, 16> Finishes; 371 372 // For each block, find which slots are active within this block 373 // and update the live intervals. 374 for (MachineFunction::iterator MBB = MF->begin(), MBBe = MF->end(); 375 MBB != MBBe; ++MBB) { 376 Starts.clear(); 377 Starts.resize(NumSlots); 378 Finishes.clear(); 379 Finishes.resize(NumSlots); 380 381 // Create the interval for the basic blocks with lifetime markers in them. 382 for (SmallVector<MachineInstr*, 8>::iterator it = Markers.begin(), 383 e = Markers.end(); it != e; ++it) { 384 MachineInstr *MI = *it; 385 if (MI->getParent() != MBB) 386 continue; 387 388 assert((MI->getOpcode() == TargetOpcode::LIFETIME_START || 389 MI->getOpcode() == TargetOpcode::LIFETIME_END) && 390 "Invalid Lifetime marker"); 391 392 bool IsStart = MI->getOpcode() == TargetOpcode::LIFETIME_START; 393 MachineOperand &Mo = MI->getOperand(0); 394 int Slot = Mo.getIndex(); 395 assert(Slot >= 0 && "Invalid slot"); 396 397 SlotIndex ThisIndex = Indexes->getInstructionIndex(MI); 398 399 if (IsStart) { 400 if (!Starts[Slot].isValid() || Starts[Slot] > ThisIndex) 401 Starts[Slot] = ThisIndex; 402 } else { 403 if (!Finishes[Slot].isValid() || Finishes[Slot] < ThisIndex) 404 Finishes[Slot] = ThisIndex; 405 } 406 } 407 408 // Create the interval of the blocks that we previously found to be 'alive'. 409 BitVector Alive = BlockLiveness[MBB].LiveIn; 410 Alive |= BlockLiveness[MBB].LiveOut; 411 412 if (Alive.any()) { 413 for (int pos = Alive.find_first(); pos != -1; 414 pos = Alive.find_next(pos)) { 415 if (!Starts[pos].isValid()) 416 Starts[pos] = Indexes->getMBBStartIdx(MBB); 417 if (!Finishes[pos].isValid()) 418 Finishes[pos] = Indexes->getMBBEndIdx(MBB); 419 } 420 } 421 422 for (unsigned i = 0; i < NumSlots; ++i) { 423 assert(Starts[i].isValid() == Finishes[i].isValid() && "Unmatched range"); 424 if (!Starts[i].isValid()) 425 continue; 426 427 assert(Starts[i] && Finishes[i] && "Invalid interval"); 428 VNInfo *ValNum = Intervals[i]->getValNumInfo(0); 429 SlotIndex S = Starts[i]; 430 SlotIndex F = Finishes[i]; 431 if (S < F) { 432 // We have a single consecutive region. 433 Intervals[i]->addRange(LiveRange(S, F, ValNum)); 434 } else { 435 // We have two non consecutive regions. This happens when 436 // LIFETIME_START appears after the LIFETIME_END marker. 437 SlotIndex NewStart = Indexes->getMBBStartIdx(MBB); 438 SlotIndex NewFin = Indexes->getMBBEndIdx(MBB); 439 Intervals[i]->addRange(LiveRange(NewStart, F, ValNum)); 440 Intervals[i]->addRange(LiveRange(S, NewFin, ValNum)); 441 } 442 } 443 } 444} 445 446bool StackColoring::removeAllMarkers() { 447 unsigned Count = 0; 448 for (unsigned i = 0; i < Markers.size(); ++i) { 449 Markers[i]->eraseFromParent(); 450 Count++; 451 } 452 Markers.clear(); 453 454 DEBUG(dbgs()<<"Removed "<<Count<<" markers.\n"); 455 return Count; 456} 457 458void StackColoring::remapInstructions(DenseMap<int, int> &SlotRemap) { 459 unsigned FixedInstr = 0; 460 unsigned FixedMemOp = 0; 461 unsigned FixedDbg = 0; 462 MachineModuleInfo *MMI = &MF->getMMI(); 463 464 // Remap debug information that refers to stack slots. 465 MachineModuleInfo::VariableDbgInfoMapTy &VMap = MMI->getVariableDbgInfo(); 466 for (MachineModuleInfo::VariableDbgInfoMapTy::iterator VI = VMap.begin(), 467 VE = VMap.end(); VI != VE; ++VI) { 468 const MDNode *Var = VI->first; 469 if (!Var) continue; 470 std::pair<unsigned, DebugLoc> &VP = VI->second; 471 if (SlotRemap.count(VP.first)) { 472 DEBUG(dbgs()<<"Remapping debug info for ["<<Var->getName()<<"].\n"); 473 VP.first = SlotRemap[VP.first]; 474 FixedDbg++; 475 } 476 } 477 478 // Keep a list of *allocas* which need to be remapped. 479 DenseMap<const Value*, const Value*> Allocas; 480 for (DenseMap<int, int>::iterator it = SlotRemap.begin(), 481 e = SlotRemap.end(); it != e; ++it) { 482 const Value *From = MFI->getObjectAllocation(it->first); 483 const Value *To = MFI->getObjectAllocation(it->second); 484 assert(To && From && "Invalid allocation object"); 485 Allocas[From] = To; 486 } 487 488 // Remap all instructions to the new stack slots. 489 MachineFunction::iterator BB, BBE; 490 MachineBasicBlock::iterator I, IE; 491 for (BB = MF->begin(), BBE = MF->end(); BB != BBE; ++BB) 492 for (I = BB->begin(), IE = BB->end(); I != IE; ++I) { 493 494 // Skip lifetime markers. We'll remove them soon. 495 if (I->getOpcode() == TargetOpcode::LIFETIME_START || 496 I->getOpcode() == TargetOpcode::LIFETIME_END) 497 continue; 498 499 // Update the MachineMemOperand to use the new alloca. 500 for (MachineInstr::mmo_iterator MM = I->memoperands_begin(), 501 E = I->memoperands_end(); MM != E; ++MM) { 502 MachineMemOperand *MMO = *MM; 503 504 const Value *V = MMO->getValue(); 505 506 if (!V) 507 continue; 508 509 // Climb up and find the original alloca. 510 V = GetUnderlyingObject(V); 511 // If we did not find one, or if the one that we found is not in our 512 // map, then move on. 513 if (!V || !Allocas.count(V)) 514 continue; 515 516 MMO->setValue(Allocas[V]); 517 FixedMemOp++; 518 } 519 520 // Update all of the machine instruction operands. 521 for (unsigned i = 0 ; i < I->getNumOperands(); ++i) { 522 MachineOperand &MO = I->getOperand(i); 523 524 if (!MO.isFI()) 525 continue; 526 int FromSlot = MO.getIndex(); 527 528 // Don't touch arguments. 529 if (FromSlot<0) 530 continue; 531 532 // Only look at mapped slots. 533 if (!SlotRemap.count(FromSlot)) 534 continue; 535 536 // In a debug build, check that the instruction that we are modifying is 537 // inside the expected live range. If the instruction is not inside 538 // the calculated range then it means that the alloca usage moved 539 // outside of the lifetime markers. 540 // NOTE: Alloca address calculations which happen outside the lifetime 541 // zone are are okay, despite the fact that we don't have a good way 542 // for validating all of the usages of the calculation. 543#ifndef NDEBUG 544 bool TouchesMemory = I->mayLoad() || I->mayStore(); 545 if (!I->isDebugValue() && TouchesMemory) { 546 SlotIndex Index = Indexes->getInstructionIndex(I); 547 LiveInterval *Interval = Intervals[FromSlot]; 548 assert(Interval->find(Index) != Interval->end() && 549 "Found instruction usage outside of live range."); 550 } 551#endif 552 553 // Fix the machine instructions. 554 int ToSlot = SlotRemap[FromSlot]; 555 MO.setIndex(ToSlot); 556 FixedInstr++; 557 } 558 } 559 560 DEBUG(dbgs()<<"Fixed "<<FixedMemOp<<" machine memory operands.\n"); 561 DEBUG(dbgs()<<"Fixed "<<FixedDbg<<" debug locations.\n"); 562 DEBUG(dbgs()<<"Fixed "<<FixedInstr<<" machine instructions.\n"); 563} 564 565void StackColoring::removeInvalidSlotRanges() { 566 MachineFunction::iterator BB, BBE; 567 MachineBasicBlock::iterator I, IE; 568 for (BB = MF->begin(), BBE = MF->end(); BB != BBE; ++BB) 569 for (I = BB->begin(), IE = BB->end(); I != IE; ++I) { 570 571 if (I->getOpcode() == TargetOpcode::LIFETIME_START || 572 I->getOpcode() == TargetOpcode::LIFETIME_END || I->isDebugValue()) 573 continue; 574 575 // Some intervals are suspicious! In some cases we find address 576 // calculations outside of the lifetime zone, but not actual memory 577 // read or write. Memory accesses outside of the lifetime zone are a clear 578 // violation, but address calculations are okay. This can happen when 579 // GEPs are hoisted outside of the lifetime zone. 580 // So, in here we only check instrucitons which can read or write memory. 581 if (!I->mayLoad() && !I->mayStore()) 582 continue; 583 584 // Check all of the machine operands. 585 for (unsigned i = 0 ; i < I->getNumOperands(); ++i) { 586 MachineOperand &MO = I->getOperand(i); 587 588 if (!MO.isFI()) 589 continue; 590 591 int Slot = MO.getIndex(); 592 593 if (Slot<0) 594 continue; 595 596 if (Intervals[Slot]->empty()) 597 continue; 598 599 // Check that the used slot is inside the calculated lifetime range. 600 // If it is not, warn about it and invalidate the range. 601 LiveInterval *Interval = Intervals[Slot]; 602 SlotIndex Index = Indexes->getInstructionIndex(I); 603 if (Interval->find(Index) == Interval->end()) { 604 Intervals[Slot]->clear(); 605 DEBUG(dbgs()<<"Invalidating range #"<<Slot<<"\n"); 606 EscapedAllocas++; 607 } 608 } 609 } 610} 611 612void StackColoring::expungeSlotMap(DenseMap<int, int> &SlotRemap, 613 unsigned NumSlots) { 614 // Expunge slot remap map. 615 for (unsigned i=0; i < NumSlots; ++i) { 616 // If we are remapping i 617 if (SlotRemap.count(i)) { 618 int Target = SlotRemap[i]; 619 // As long as our target is mapped to something else, follow it. 620 while (SlotRemap.count(Target)) { 621 Target = SlotRemap[Target]; 622 SlotRemap[i] = Target; 623 } 624 } 625 } 626} 627 628bool StackColoring::runOnMachineFunction(MachineFunction &Func) { 629 DEBUG(dbgs() << "********** Stack Coloring **********\n" 630 << "********** Function: " 631 << ((const Value*)Func.getFunction())->getName() << '\n'); 632 MF = &Func; 633 MFI = MF->getFrameInfo(); 634 Indexes = &getAnalysis<SlotIndexes>(); 635 BlockLiveness.clear(); 636 BasicBlocks.clear(); 637 BasicBlockNumbering.clear(); 638 Markers.clear(); 639 Intervals.clear(); 640 VNInfoAllocator.Reset(); 641 642 unsigned NumSlots = MFI->getObjectIndexEnd(); 643 644 // If there are no stack slots then there are no markers to remove. 645 if (!NumSlots) 646 return false; 647 648 SmallVector<int, 8> SortedSlots; 649 650 SortedSlots.reserve(NumSlots); 651 Intervals.reserve(NumSlots); 652 653 unsigned NumMarkers = collectMarkers(NumSlots); 654 655 unsigned TotalSize = 0; 656 DEBUG(dbgs()<<"Found "<<NumMarkers<<" markers and "<<NumSlots<<" slots\n"); 657 DEBUG(dbgs()<<"Slot structure:\n"); 658 659 for (int i=0; i < MFI->getObjectIndexEnd(); ++i) { 660 DEBUG(dbgs()<<"Slot #"<<i<<" - "<<MFI->getObjectSize(i)<<" bytes.\n"); 661 TotalSize += MFI->getObjectSize(i); 662 } 663 664 DEBUG(dbgs()<<"Total Stack size: "<<TotalSize<<" bytes\n\n"); 665 666 // Don't continue because there are not enough lifetime markers, or the 667 // stack is too small, or we are told not to optimize the slots. 668 if (NumMarkers < 2 || TotalSize < 16 || DisableColoring) { 669 DEBUG(dbgs()<<"Will not try to merge slots.\n"); 670 return removeAllMarkers(); 671 } 672 673 for (unsigned i=0; i < NumSlots; ++i) { 674 LiveInterval *LI = new LiveInterval(i, 0); 675 Intervals.push_back(LI); 676 LI->getNextValue(Indexes->getZeroIndex(), VNInfoAllocator); 677 SortedSlots.push_back(i); 678 } 679 680 // Calculate the liveness of each block. 681 calculateLocalLiveness(); 682 683 // Propagate the liveness information. 684 calculateLiveIntervals(NumSlots); 685 686 // Search for allocas which are used outside of the declared lifetime 687 // markers. 688 if (CheckEscapedAllocas) 689 removeInvalidSlotRanges(); 690 691 // Maps old slots to new slots. 692 DenseMap<int, int> SlotRemap; 693 unsigned RemovedSlots = 0; 694 unsigned ReducedSize = 0; 695 696 // Do not bother looking at empty intervals. 697 for (unsigned I = 0; I < NumSlots; ++I) { 698 if (Intervals[SortedSlots[I]]->empty()) 699 SortedSlots[I] = -1; 700 } 701 702 // This is a simple greedy algorithm for merging allocas. First, sort the 703 // slots, placing the largest slots first. Next, perform an n^2 scan and look 704 // for disjoint slots. When you find disjoint slots, merge the samller one 705 // into the bigger one and update the live interval. Remove the small alloca 706 // and continue. 707 708 // Sort the slots according to their size. Place unused slots at the end. 709 std::sort(SortedSlots.begin(), SortedSlots.end(), SlotSizeSorter(MFI)); 710 711 bool Chanded = true; 712 while (Chanded) { 713 Chanded = false; 714 for (unsigned I = 0; I < NumSlots; ++I) { 715 if (SortedSlots[I] == -1) 716 continue; 717 718 for (unsigned J=I+1; J < NumSlots; ++J) { 719 if (SortedSlots[J] == -1) 720 continue; 721 722 int FirstSlot = SortedSlots[I]; 723 int SecondSlot = SortedSlots[J]; 724 LiveInterval *First = Intervals[FirstSlot]; 725 LiveInterval *Second = Intervals[SecondSlot]; 726 assert (!First->empty() && !Second->empty() && "Found an empty range"); 727 728 // Merge disjoint slots. 729 if (!First->overlaps(*Second)) { 730 Chanded = true; 731 First->MergeRangesInAsValue(*Second, First->getValNumInfo(0)); 732 SlotRemap[SecondSlot] = FirstSlot; 733 SortedSlots[J] = -1; 734 DEBUG(dbgs()<<"Merging #"<<FirstSlot<<" and slots #"<< 735 SecondSlot<<" together.\n"); 736 unsigned MaxAlignment = std::max(MFI->getObjectAlignment(FirstSlot), 737 MFI->getObjectAlignment(SecondSlot)); 738 739 assert(MFI->getObjectSize(FirstSlot) >= 740 MFI->getObjectSize(SecondSlot) && 741 "Merging a small object into a larger one"); 742 743 RemovedSlots+=1; 744 ReducedSize += MFI->getObjectSize(SecondSlot); 745 MFI->setObjectAlignment(FirstSlot, MaxAlignment); 746 MFI->RemoveStackObject(SecondSlot); 747 } 748 } 749 } 750 }// While changed. 751 752 // Record statistics. 753 StackSpaceSaved += ReducedSize; 754 StackSlotMerged += RemovedSlots; 755 DEBUG(dbgs()<<"Merge "<<RemovedSlots<<" slots. Saved "<< 756 ReducedSize<<" bytes\n"); 757 758 // Scan the entire function and update all machine operands that use frame 759 // indices to use the remapped frame index. 760 expungeSlotMap(SlotRemap, NumSlots); 761 remapInstructions(SlotRemap); 762 763 // Release the intervals. 764 for (unsigned I = 0; I < NumSlots; ++I) { 765 delete Intervals[I]; 766 } 767 768 return removeAllMarkers(); 769} 770