StackColoring.cpp revision 36b56886974eae4f9c5ebc96befd3e7bfe5de338
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/ValueTracking.h" 34#include "llvm/CodeGen/LiveInterval.h" 35#include "llvm/CodeGen/MachineBasicBlock.h" 36#include "llvm/CodeGen/MachineBranchProbabilityInfo.h" 37#include "llvm/CodeGen/MachineDominators.h" 38#include "llvm/CodeGen/MachineFrameInfo.h" 39#include "llvm/CodeGen/MachineFunctionPass.h" 40#include "llvm/CodeGen/MachineLoopInfo.h" 41#include "llvm/CodeGen/MachineMemOperand.h" 42#include "llvm/CodeGen/MachineModuleInfo.h" 43#include "llvm/CodeGen/MachineRegisterInfo.h" 44#include "llvm/CodeGen/PseudoSourceValue.h" 45#include "llvm/CodeGen/SlotIndexes.h" 46#include "llvm/CodeGen/StackProtector.h" 47#include "llvm/IR/DebugInfo.h" 48#include "llvm/IR/Dominators.h" 49#include "llvm/IR/Function.h" 50#include "llvm/IR/Instructions.h" 51#include "llvm/IR/Module.h" 52#include "llvm/MC/MCInstrItineraries.h" 53#include "llvm/Support/CommandLine.h" 54#include "llvm/Support/Debug.h" 55#include "llvm/Support/raw_ostream.h" 56#include "llvm/Target/TargetInstrInfo.h" 57#include "llvm/Target/TargetRegisterInfo.h" 58 59using namespace llvm; 60 61static cl::opt<bool> 62DisableColoring("no-stack-coloring", 63 cl::init(false), cl::Hidden, 64 cl::desc("Disable stack coloring")); 65 66/// The user may write code that uses allocas outside of the declared lifetime 67/// zone. This can happen when the user returns a reference to a local 68/// data-structure. We can detect these cases and decide not to optimize the 69/// code. If this flag is enabled, we try to save the user. 70static cl::opt<bool> 71ProtectFromEscapedAllocas("protect-from-escaped-allocas", 72 cl::init(false), cl::Hidden, 73 cl::desc("Do not optimize lifetime zones that " 74 "are broken")); 75 76STATISTIC(NumMarkerSeen, "Number of lifetime markers found."); 77STATISTIC(StackSpaceSaved, "Number of bytes saved due to merging slots."); 78STATISTIC(StackSlotMerged, "Number of stack slot merged."); 79STATISTIC(EscapedAllocas, "Number of allocas that escaped the lifetime region"); 80 81//===----------------------------------------------------------------------===// 82// StackColoring Pass 83//===----------------------------------------------------------------------===// 84 85namespace { 86/// StackColoring - A machine pass for merging disjoint stack allocations, 87/// marked by the LIFETIME_START and LIFETIME_END pseudo instructions. 88class StackColoring : public MachineFunctionPass { 89 MachineFrameInfo *MFI; 90 MachineFunction *MF; 91 92 /// A class representing liveness information for a single basic block. 93 /// Each bit in the BitVector represents the liveness property 94 /// for a different stack slot. 95 struct BlockLifetimeInfo { 96 /// Which slots BEGINs in each basic block. 97 BitVector Begin; 98 /// Which slots ENDs in each basic block. 99 BitVector End; 100 /// Which slots are marked as LIVE_IN, coming into each basic block. 101 BitVector LiveIn; 102 /// Which slots are marked as LIVE_OUT, coming out of each basic block. 103 BitVector LiveOut; 104 }; 105 106 /// Maps active slots (per bit) for each basic block. 107 typedef DenseMap<const MachineBasicBlock*, BlockLifetimeInfo> LivenessMap; 108 LivenessMap BlockLiveness; 109 110 /// Maps serial numbers to basic blocks. 111 DenseMap<const MachineBasicBlock*, int> BasicBlocks; 112 /// Maps basic blocks to a serial number. 113 SmallVector<const MachineBasicBlock*, 8> BasicBlockNumbering; 114 115 /// Maps liveness intervals for each slot. 116 SmallVector<std::unique_ptr<LiveInterval>, 16> Intervals; 117 /// VNInfo is used for the construction of LiveIntervals. 118 VNInfo::Allocator VNInfoAllocator; 119 /// SlotIndex analysis object. 120 SlotIndexes *Indexes; 121 /// The stack protector object. 122 StackProtector *SP; 123 124 /// The list of lifetime markers found. These markers are to be removed 125 /// once the coloring is done. 126 SmallVector<MachineInstr*, 8> Markers; 127 128public: 129 static char ID; 130 StackColoring() : MachineFunctionPass(ID) { 131 initializeStackColoringPass(*PassRegistry::getPassRegistry()); 132 } 133 void getAnalysisUsage(AnalysisUsage &AU) const override; 134 bool runOnMachineFunction(MachineFunction &MF) override; 135 136private: 137 /// Debug. 138 void dump() const; 139 140 /// Removes all of the lifetime marker instructions from the function. 141 /// \returns true if any markers were removed. 142 bool removeAllMarkers(); 143 144 /// Scan the machine function and find all of the lifetime markers. 145 /// Record the findings in the BEGIN and END vectors. 146 /// \returns the number of markers found. 147 unsigned collectMarkers(unsigned NumSlot); 148 149 /// Perform the dataflow calculation and calculate the lifetime for each of 150 /// the slots, based on the BEGIN/END vectors. Set the LifetimeLIVE_IN and 151 /// LifetimeLIVE_OUT maps that represent which stack slots are live coming 152 /// in and out blocks. 153 void calculateLocalLiveness(); 154 155 /// Construct the LiveIntervals for the slots. 156 void calculateLiveIntervals(unsigned NumSlots); 157 158 /// Go over the machine function and change instructions which use stack 159 /// slots to use the joint slots. 160 void remapInstructions(DenseMap<int, int> &SlotRemap); 161 162 /// The input program may contain instructions which are not inside lifetime 163 /// markers. This can happen due to a bug in the compiler or due to a bug in 164 /// user code (for example, returning a reference to a local variable). 165 /// This procedure checks all of the instructions in the function and 166 /// invalidates lifetime ranges which do not contain all of the instructions 167 /// which access that frame slot. 168 void removeInvalidSlotRanges(); 169 170 /// Map entries which point to other entries to their destination. 171 /// A->B->C becomes A->C. 172 void expungeSlotMap(DenseMap<int, int> &SlotRemap, unsigned NumSlots); 173}; 174} // end anonymous namespace 175 176char StackColoring::ID = 0; 177char &llvm::StackColoringID = StackColoring::ID; 178 179INITIALIZE_PASS_BEGIN(StackColoring, 180 "stack-coloring", "Merge disjoint stack slots", false, false) 181INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) 182INITIALIZE_PASS_DEPENDENCY(SlotIndexes) 183INITIALIZE_PASS_DEPENDENCY(StackProtector) 184INITIALIZE_PASS_END(StackColoring, 185 "stack-coloring", "Merge disjoint stack slots", false, false) 186 187void StackColoring::getAnalysisUsage(AnalysisUsage &AU) const { 188 AU.addRequired<MachineDominatorTree>(); 189 AU.addPreserved<MachineDominatorTree>(); 190 AU.addRequired<SlotIndexes>(); 191 AU.addRequired<StackProtector>(); 192 MachineFunctionPass::getAnalysisUsage(AU); 193} 194 195void StackColoring::dump() const { 196 for (df_iterator<MachineFunction*> FI = df_begin(MF), FE = df_end(MF); 197 FI != FE; ++FI) { 198 DEBUG(dbgs()<<"Inspecting block #"<<BasicBlocks.lookup(*FI)<< 199 " ["<<FI->getName()<<"]\n"); 200 201 LivenessMap::const_iterator BI = BlockLiveness.find(*FI); 202 assert(BI != BlockLiveness.end() && "Block not found"); 203 const BlockLifetimeInfo &BlockInfo = BI->second; 204 205 DEBUG(dbgs()<<"BEGIN : {"); 206 for (unsigned i=0; i < BlockInfo.Begin.size(); ++i) 207 DEBUG(dbgs()<<BlockInfo.Begin.test(i)<<" "); 208 DEBUG(dbgs()<<"}\n"); 209 210 DEBUG(dbgs()<<"END : {"); 211 for (unsigned i=0; i < BlockInfo.End.size(); ++i) 212 DEBUG(dbgs()<<BlockInfo.End.test(i)<<" "); 213 214 DEBUG(dbgs()<<"}\n"); 215 216 DEBUG(dbgs()<<"LIVE_IN: {"); 217 for (unsigned i=0; i < BlockInfo.LiveIn.size(); ++i) 218 DEBUG(dbgs()<<BlockInfo.LiveIn.test(i)<<" "); 219 220 DEBUG(dbgs()<<"}\n"); 221 DEBUG(dbgs()<<"LIVEOUT: {"); 222 for (unsigned i=0; i < BlockInfo.LiveOut.size(); ++i) 223 DEBUG(dbgs()<<BlockInfo.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 // Keep a reference to avoid repeated lookups. 242 BlockLifetimeInfo &BlockInfo = BlockLiveness[*FI]; 243 244 BlockInfo.Begin.resize(NumSlot); 245 BlockInfo.End.resize(NumSlot); 246 247 for (MachineInstr &MI : **FI) { 248 if (MI.getOpcode() != TargetOpcode::LIFETIME_START && 249 MI.getOpcode() != TargetOpcode::LIFETIME_END) 250 continue; 251 252 Markers.push_back(&MI); 253 254 bool IsStart = MI.getOpcode() == TargetOpcode::LIFETIME_START; 255 const MachineOperand &MO = MI.getOperand(0); 256 unsigned Slot = MO.getIndex(); 257 258 MarkersFound++; 259 260 const AllocaInst *Allocation = MFI->getObjectAllocation(Slot); 261 if (Allocation) { 262 DEBUG(dbgs()<<"Found a lifetime marker for slot #"<<Slot<< 263 " with allocation: "<< Allocation->getName()<<"\n"); 264 } 265 266 if (IsStart) { 267 BlockInfo.Begin.set(Slot); 268 } else { 269 if (BlockInfo.Begin.test(Slot)) { 270 // Allocas that start and end within a single block are handled 271 // specially when computing the LiveIntervals to avoid pessimizing 272 // the liveness propagation. 273 BlockInfo.Begin.reset(Slot); 274 } else { 275 BlockInfo.End.set(Slot); 276 } 277 } 278 } 279 } 280 281 // Update statistics. 282 NumMarkerSeen += MarkersFound; 283 return MarkersFound; 284} 285 286void StackColoring::calculateLocalLiveness() { 287 // Perform a standard reverse dataflow computation to solve for 288 // global liveness. The BEGIN set here is equivalent to KILL in the standard 289 // formulation, and END is equivalent to GEN. The result of this computation 290 // is a map from blocks to bitvectors where the bitvectors represent which 291 // allocas are live in/out of that block. 292 SmallPtrSet<const MachineBasicBlock*, 8> BBSet(BasicBlockNumbering.begin(), 293 BasicBlockNumbering.end()); 294 unsigned NumSSMIters = 0; 295 bool changed = true; 296 while (changed) { 297 changed = false; 298 ++NumSSMIters; 299 300 SmallPtrSet<const MachineBasicBlock*, 8> NextBBSet; 301 302 for (const MachineBasicBlock *BB : BasicBlockNumbering) { 303 if (!BBSet.count(BB)) continue; 304 305 // Use an iterator to avoid repeated lookups. 306 LivenessMap::iterator BI = BlockLiveness.find(BB); 307 assert(BI != BlockLiveness.end() && "Block not found"); 308 BlockLifetimeInfo &BlockInfo = BI->second; 309 310 BitVector LocalLiveIn; 311 BitVector LocalLiveOut; 312 313 // Forward propagation from begins to ends. 314 for (MachineBasicBlock::const_pred_iterator PI = BB->pred_begin(), 315 PE = BB->pred_end(); PI != PE; ++PI) { 316 LivenessMap::const_iterator I = BlockLiveness.find(*PI); 317 assert(I != BlockLiveness.end() && "Predecessor not found"); 318 LocalLiveIn |= I->second.LiveOut; 319 } 320 LocalLiveIn |= BlockInfo.End; 321 LocalLiveIn.reset(BlockInfo.Begin); 322 323 // Reverse propagation from ends to begins. 324 for (MachineBasicBlock::const_succ_iterator SI = BB->succ_begin(), 325 SE = BB->succ_end(); SI != SE; ++SI) { 326 LivenessMap::const_iterator I = BlockLiveness.find(*SI); 327 assert(I != BlockLiveness.end() && "Successor not found"); 328 LocalLiveOut |= I->second.LiveIn; 329 } 330 LocalLiveOut |= BlockInfo.Begin; 331 LocalLiveOut.reset(BlockInfo.End); 332 333 LocalLiveIn |= LocalLiveOut; 334 LocalLiveOut |= LocalLiveIn; 335 336 // After adopting the live bits, we need to turn-off the bits which 337 // are de-activated in this block. 338 LocalLiveOut.reset(BlockInfo.End); 339 LocalLiveIn.reset(BlockInfo.Begin); 340 341 // If we have both BEGIN and END markers in the same basic block then 342 // we know that the BEGIN marker comes after the END, because we already 343 // handle the case where the BEGIN comes before the END when collecting 344 // the markers (and building the BEGIN/END vectore). 345 // Want to enable the LIVE_IN and LIVE_OUT of slots that have both 346 // BEGIN and END because it means that the value lives before and after 347 // this basic block. 348 BitVector LocalEndBegin = BlockInfo.End; 349 LocalEndBegin &= BlockInfo.Begin; 350 LocalLiveIn |= LocalEndBegin; 351 LocalLiveOut |= LocalEndBegin; 352 353 if (LocalLiveIn.test(BlockInfo.LiveIn)) { 354 changed = true; 355 BlockInfo.LiveIn |= LocalLiveIn; 356 357 NextBBSet.insert(BB->pred_begin(), BB->pred_end()); 358 } 359 360 if (LocalLiveOut.test(BlockInfo.LiveOut)) { 361 changed = true; 362 BlockInfo.LiveOut |= LocalLiveOut; 363 364 NextBBSet.insert(BB->succ_begin(), BB->succ_end()); 365 } 366 } 367 368 BBSet = NextBBSet; 369 }// while changed. 370} 371 372void StackColoring::calculateLiveIntervals(unsigned NumSlots) { 373 SmallVector<SlotIndex, 16> Starts; 374 SmallVector<SlotIndex, 16> Finishes; 375 376 // For each block, find which slots are active within this block 377 // and update the live intervals. 378 for (const MachineBasicBlock &MBB : *MF) { 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 (const MachineInstr *MI : Markers) { 386 if (MI->getParent() != &MBB) 387 continue; 388 389 assert((MI->getOpcode() == TargetOpcode::LIFETIME_START || 390 MI->getOpcode() == TargetOpcode::LIFETIME_END) && 391 "Invalid Lifetime marker"); 392 393 bool IsStart = MI->getOpcode() == TargetOpcode::LIFETIME_START; 394 const MachineOperand &Mo = MI->getOperand(0); 395 int Slot = Mo.getIndex(); 396 assert(Slot >= 0 && "Invalid slot"); 397 398 SlotIndex ThisIndex = Indexes->getInstructionIndex(MI); 399 400 if (IsStart) { 401 if (!Starts[Slot].isValid() || Starts[Slot] > ThisIndex) 402 Starts[Slot] = ThisIndex; 403 } else { 404 if (!Finishes[Slot].isValid() || Finishes[Slot] < ThisIndex) 405 Finishes[Slot] = ThisIndex; 406 } 407 } 408 409 // Create the interval of the blocks that we previously found to be 'alive'. 410 BlockLifetimeInfo &MBBLiveness = BlockLiveness[&MBB]; 411 for (int pos = MBBLiveness.LiveIn.find_first(); pos != -1; 412 pos = MBBLiveness.LiveIn.find_next(pos)) { 413 Starts[pos] = Indexes->getMBBStartIdx(&MBB); 414 } 415 for (int pos = MBBLiveness.LiveOut.find_first(); pos != -1; 416 pos = MBBLiveness.LiveOut.find_next(pos)) { 417 Finishes[pos] = Indexes->getMBBEndIdx(&MBB); 418 } 419 420 for (unsigned i = 0; i < NumSlots; ++i) { 421 assert(Starts[i].isValid() == Finishes[i].isValid() && "Unmatched range"); 422 if (!Starts[i].isValid()) 423 continue; 424 425 assert(Starts[i] && Finishes[i] && "Invalid interval"); 426 VNInfo *ValNum = Intervals[i]->getValNumInfo(0); 427 SlotIndex S = Starts[i]; 428 SlotIndex F = Finishes[i]; 429 if (S < F) { 430 // We have a single consecutive region. 431 Intervals[i]->addSegment(LiveInterval::Segment(S, F, ValNum)); 432 } else { 433 // We have two non-consecutive regions. This happens when 434 // LIFETIME_START appears after the LIFETIME_END marker. 435 SlotIndex NewStart = Indexes->getMBBStartIdx(&MBB); 436 SlotIndex NewFin = Indexes->getMBBEndIdx(&MBB); 437 Intervals[i]->addSegment(LiveInterval::Segment(NewStart, F, ValNum)); 438 Intervals[i]->addSegment(LiveInterval::Segment(S, NewFin, ValNum)); 439 } 440 } 441 } 442} 443 444bool StackColoring::removeAllMarkers() { 445 unsigned Count = 0; 446 for (MachineInstr *MI : Markers) { 447 MI->eraseFromParent(); 448 Count++; 449 } 450 Markers.clear(); 451 452 DEBUG(dbgs()<<"Removed "<<Count<<" markers.\n"); 453 return Count; 454} 455 456void StackColoring::remapInstructions(DenseMap<int, int> &SlotRemap) { 457 unsigned FixedInstr = 0; 458 unsigned FixedMemOp = 0; 459 unsigned FixedDbg = 0; 460 MachineModuleInfo *MMI = &MF->getMMI(); 461 462 // Remap debug information that refers to stack slots. 463 for (auto &VI : MMI->getVariableDbgInfo()) { 464 if (!VI.Var) 465 continue; 466 if (SlotRemap.count(VI.Slot)) { 467 DEBUG(dbgs()<<"Remapping debug info for ["<<VI.Var->getName()<<"].\n"); 468 VI.Slot = SlotRemap[VI.Slot]; 469 FixedDbg++; 470 } 471 } 472 473 // Keep a list of *allocas* which need to be remapped. 474 DenseMap<const AllocaInst*, const AllocaInst*> Allocas; 475 for (const std::pair<int, int> &SI : SlotRemap) { 476 const AllocaInst *From = MFI->getObjectAllocation(SI.first); 477 const AllocaInst *To = MFI->getObjectAllocation(SI.second); 478 assert(To && From && "Invalid allocation object"); 479 Allocas[From] = To; 480 481 // AA might be used later for instruction scheduling, and we need it to be 482 // able to deduce the correct aliasing releationships between pointers 483 // derived from the alloca being remapped and the target of that remapping. 484 // The only safe way, without directly informing AA about the remapping 485 // somehow, is to directly update the IR to reflect the change being made 486 // here. 487 Instruction *Inst = const_cast<AllocaInst *>(To); 488 if (From->getType() != To->getType()) { 489 BitCastInst *Cast = new BitCastInst(Inst, From->getType()); 490 Cast->insertAfter(Inst); 491 Inst = Cast; 492 } 493 494 // Allow the stack protector to adjust its value map to account for the 495 // upcoming replacement. 496 SP->adjustForColoring(From, To); 497 498 // Note that this will not replace uses in MMOs (which we'll update below), 499 // or anywhere else (which is why we won't delete the original 500 // instruction). 501 const_cast<AllocaInst *>(From)->replaceAllUsesWith(Inst); 502 } 503 504 // Remap all instructions to the new stack slots. 505 for (MachineBasicBlock &BB : *MF) 506 for (MachineInstr &I : BB) { 507 // Skip lifetime markers. We'll remove them soon. 508 if (I.getOpcode() == TargetOpcode::LIFETIME_START || 509 I.getOpcode() == TargetOpcode::LIFETIME_END) 510 continue; 511 512 // Update the MachineMemOperand to use the new alloca. 513 for (MachineMemOperand *MMO : I.memoperands()) { 514 const Value *V = MMO->getValue(); 515 516 if (!V) 517 continue; 518 519 // FIXME: In order to enable the use of TBAA when using AA in CodeGen, 520 // we'll also need to update the TBAA nodes in MMOs with values 521 // derived from the merged allocas. When doing this, we'll need to use 522 // the same variant of GetUnderlyingObjects that is used by the 523 // instruction scheduler (that can look through ptrtoint/inttoptr 524 // pairs). 525 526 // We've replaced IR-level uses of the remapped allocas, so we only 527 // need to replace direct uses here. 528 if (!isa<AllocaInst>(V)) 529 continue; 530 531 const AllocaInst *AI= cast<AllocaInst>(V); 532 if (!Allocas.count(AI)) 533 continue; 534 535 MMO->setValue(Allocas[AI]); 536 FixedMemOp++; 537 } 538 539 // Update all of the machine instruction operands. 540 for (MachineOperand &MO : I.operands()) { 541 if (!MO.isFI()) 542 continue; 543 int FromSlot = MO.getIndex(); 544 545 // Don't touch arguments. 546 if (FromSlot<0) 547 continue; 548 549 // Only look at mapped slots. 550 if (!SlotRemap.count(FromSlot)) 551 continue; 552 553 // In a debug build, check that the instruction that we are modifying is 554 // inside the expected live range. If the instruction is not inside 555 // the calculated range then it means that the alloca usage moved 556 // outside of the lifetime markers, or that the user has a bug. 557 // NOTE: Alloca address calculations which happen outside the lifetime 558 // zone are are okay, despite the fact that we don't have a good way 559 // for validating all of the usages of the calculation. 560#ifndef NDEBUG 561 bool TouchesMemory = I.mayLoad() || I.mayStore(); 562 // If we *don't* protect the user from escaped allocas, don't bother 563 // validating the instructions. 564 if (!I.isDebugValue() && TouchesMemory && ProtectFromEscapedAllocas) { 565 SlotIndex Index = Indexes->getInstructionIndex(&I); 566 const LiveInterval *Interval = &*Intervals[FromSlot]; 567 assert(Interval->find(Index) != Interval->end() && 568 "Found instruction usage outside of live range."); 569 } 570#endif 571 572 // Fix the machine instructions. 573 int ToSlot = SlotRemap[FromSlot]; 574 MO.setIndex(ToSlot); 575 FixedInstr++; 576 } 577 } 578 579 DEBUG(dbgs()<<"Fixed "<<FixedMemOp<<" machine memory operands.\n"); 580 DEBUG(dbgs()<<"Fixed "<<FixedDbg<<" debug locations.\n"); 581 DEBUG(dbgs()<<"Fixed "<<FixedInstr<<" machine instructions.\n"); 582} 583 584void StackColoring::removeInvalidSlotRanges() { 585 for (MachineBasicBlock &BB : *MF) 586 for (MachineInstr &I : BB) { 587 if (I.getOpcode() == TargetOpcode::LIFETIME_START || 588 I.getOpcode() == TargetOpcode::LIFETIME_END || I.isDebugValue()) 589 continue; 590 591 // Some intervals are suspicious! In some cases we find address 592 // calculations outside of the lifetime zone, but not actual memory 593 // read or write. Memory accesses outside of the lifetime zone are a clear 594 // violation, but address calculations are okay. This can happen when 595 // GEPs are hoisted outside of the lifetime zone. 596 // So, in here we only check instructions which can read or write memory. 597 if (!I.mayLoad() && !I.mayStore()) 598 continue; 599 600 // Check all of the machine operands. 601 for (const MachineOperand &MO : I.operands()) { 602 if (!MO.isFI()) 603 continue; 604 605 int Slot = MO.getIndex(); 606 607 if (Slot<0) 608 continue; 609 610 if (Intervals[Slot]->empty()) 611 continue; 612 613 // Check that the used slot is inside the calculated lifetime range. 614 // If it is not, warn about it and invalidate the range. 615 LiveInterval *Interval = &*Intervals[Slot]; 616 SlotIndex Index = Indexes->getInstructionIndex(&I); 617 if (Interval->find(Index) == Interval->end()) { 618 Interval->clear(); 619 DEBUG(dbgs()<<"Invalidating range #"<<Slot<<"\n"); 620 EscapedAllocas++; 621 } 622 } 623 } 624} 625 626void StackColoring::expungeSlotMap(DenseMap<int, int> &SlotRemap, 627 unsigned NumSlots) { 628 // Expunge slot remap map. 629 for (unsigned i=0; i < NumSlots; ++i) { 630 // If we are remapping i 631 if (SlotRemap.count(i)) { 632 int Target = SlotRemap[i]; 633 // As long as our target is mapped to something else, follow it. 634 while (SlotRemap.count(Target)) { 635 Target = SlotRemap[Target]; 636 SlotRemap[i] = Target; 637 } 638 } 639 } 640} 641 642bool StackColoring::runOnMachineFunction(MachineFunction &Func) { 643 if (skipOptnoneFunction(*Func.getFunction())) 644 return false; 645 646 DEBUG(dbgs() << "********** Stack Coloring **********\n" 647 << "********** Function: " 648 << ((const Value*)Func.getFunction())->getName() << '\n'); 649 MF = &Func; 650 MFI = MF->getFrameInfo(); 651 Indexes = &getAnalysis<SlotIndexes>(); 652 SP = &getAnalysis<StackProtector>(); 653 BlockLiveness.clear(); 654 BasicBlocks.clear(); 655 BasicBlockNumbering.clear(); 656 Markers.clear(); 657 Intervals.clear(); 658 VNInfoAllocator.Reset(); 659 660 unsigned NumSlots = MFI->getObjectIndexEnd(); 661 662 // If there are no stack slots then there are no markers to remove. 663 if (!NumSlots) 664 return false; 665 666 SmallVector<int, 8> SortedSlots; 667 668 SortedSlots.reserve(NumSlots); 669 Intervals.reserve(NumSlots); 670 671 unsigned NumMarkers = collectMarkers(NumSlots); 672 673 unsigned TotalSize = 0; 674 DEBUG(dbgs()<<"Found "<<NumMarkers<<" markers and "<<NumSlots<<" slots\n"); 675 DEBUG(dbgs()<<"Slot structure:\n"); 676 677 for (int i=0; i < MFI->getObjectIndexEnd(); ++i) { 678 DEBUG(dbgs()<<"Slot #"<<i<<" - "<<MFI->getObjectSize(i)<<" bytes.\n"); 679 TotalSize += MFI->getObjectSize(i); 680 } 681 682 DEBUG(dbgs()<<"Total Stack size: "<<TotalSize<<" bytes\n\n"); 683 684 // Don't continue because there are not enough lifetime markers, or the 685 // stack is too small, or we are told not to optimize the slots. 686 if (NumMarkers < 2 || TotalSize < 16 || DisableColoring) { 687 DEBUG(dbgs()<<"Will not try to merge slots.\n"); 688 return removeAllMarkers(); 689 } 690 691 for (unsigned i=0; i < NumSlots; ++i) { 692 std::unique_ptr<LiveInterval> LI(new LiveInterval(i, 0)); 693 LI->getNextValue(Indexes->getZeroIndex(), VNInfoAllocator); 694 Intervals.push_back(std::move(LI)); 695 SortedSlots.push_back(i); 696 } 697 698 // Calculate the liveness of each block. 699 calculateLocalLiveness(); 700 701 // Propagate the liveness information. 702 calculateLiveIntervals(NumSlots); 703 704 // Search for allocas which are used outside of the declared lifetime 705 // markers. 706 if (ProtectFromEscapedAllocas) 707 removeInvalidSlotRanges(); 708 709 // Maps old slots to new slots. 710 DenseMap<int, int> SlotRemap; 711 unsigned RemovedSlots = 0; 712 unsigned ReducedSize = 0; 713 714 // Do not bother looking at empty intervals. 715 for (unsigned I = 0; I < NumSlots; ++I) { 716 if (Intervals[SortedSlots[I]]->empty()) 717 SortedSlots[I] = -1; 718 } 719 720 // This is a simple greedy algorithm for merging allocas. First, sort the 721 // slots, placing the largest slots first. Next, perform an n^2 scan and look 722 // for disjoint slots. When you find disjoint slots, merge the samller one 723 // into the bigger one and update the live interval. Remove the small alloca 724 // and continue. 725 726 // Sort the slots according to their size. Place unused slots at the end. 727 // Use stable sort to guarantee deterministic code generation. 728 std::stable_sort(SortedSlots.begin(), SortedSlots.end(), 729 [this](int LHS, int RHS) { 730 // We use -1 to denote a uninteresting slot. Place these slots at the end. 731 if (LHS == -1) return false; 732 if (RHS == -1) return true; 733 // Sort according to size. 734 return MFI->getObjectSize(LHS) > MFI->getObjectSize(RHS); 735 }); 736 737 bool Changed = true; 738 while (Changed) { 739 Changed = false; 740 for (unsigned I = 0; I < NumSlots; ++I) { 741 if (SortedSlots[I] == -1) 742 continue; 743 744 for (unsigned J=I+1; J < NumSlots; ++J) { 745 if (SortedSlots[J] == -1) 746 continue; 747 748 int FirstSlot = SortedSlots[I]; 749 int SecondSlot = SortedSlots[J]; 750 LiveInterval *First = &*Intervals[FirstSlot]; 751 LiveInterval *Second = &*Intervals[SecondSlot]; 752 assert (!First->empty() && !Second->empty() && "Found an empty range"); 753 754 // Merge disjoint slots. 755 if (!First->overlaps(*Second)) { 756 Changed = true; 757 First->MergeSegmentsInAsValue(*Second, First->getValNumInfo(0)); 758 SlotRemap[SecondSlot] = FirstSlot; 759 SortedSlots[J] = -1; 760 DEBUG(dbgs()<<"Merging #"<<FirstSlot<<" and slots #"<< 761 SecondSlot<<" together.\n"); 762 unsigned MaxAlignment = std::max(MFI->getObjectAlignment(FirstSlot), 763 MFI->getObjectAlignment(SecondSlot)); 764 765 assert(MFI->getObjectSize(FirstSlot) >= 766 MFI->getObjectSize(SecondSlot) && 767 "Merging a small object into a larger one"); 768 769 RemovedSlots+=1; 770 ReducedSize += MFI->getObjectSize(SecondSlot); 771 MFI->setObjectAlignment(FirstSlot, MaxAlignment); 772 MFI->RemoveStackObject(SecondSlot); 773 } 774 } 775 } 776 }// While changed. 777 778 // Record statistics. 779 StackSpaceSaved += ReducedSize; 780 StackSlotMerged += RemovedSlots; 781 DEBUG(dbgs()<<"Merge "<<RemovedSlots<<" slots. Saved "<< 782 ReducedSize<<" bytes\n"); 783 784 // Scan the entire function and update all machine operands that use frame 785 // indices to use the remapped frame index. 786 expungeSlotMap(SlotRemap, NumSlots); 787 remapInstructions(SlotRemap); 788 789 return removeAllMarkers(); 790} 791