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#include "llvm/CodeGen/Passes.h"
25#include "llvm/ADT/BitVector.h"
26#include "llvm/ADT/DepthFirstIterator.h"
27#include "llvm/ADT/PostOrderIterator.h"
28#include "llvm/ADT/SetVector.h"
29#include "llvm/ADT/SmallPtrSet.h"
30#include "llvm/ADT/SparseSet.h"
31#include "llvm/ADT/Statistic.h"
32#include "llvm/Analysis/ValueTracking.h"
33#include "llvm/CodeGen/LiveInterval.h"
34#include "llvm/CodeGen/MachineBasicBlock.h"
35#include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
36#include "llvm/CodeGen/MachineDominators.h"
37#include "llvm/CodeGen/MachineFrameInfo.h"
38#include "llvm/CodeGen/MachineFunctionPass.h"
39#include "llvm/CodeGen/MachineLoopInfo.h"
40#include "llvm/CodeGen/MachineMemOperand.h"
41#include "llvm/CodeGen/MachineModuleInfo.h"
42#include "llvm/CodeGen/MachineRegisterInfo.h"
43#include "llvm/CodeGen/PseudoSourceValue.h"
44#include "llvm/CodeGen/SlotIndexes.h"
45#include "llvm/CodeGen/StackProtector.h"
46#include "llvm/IR/DebugInfo.h"
47#include "llvm/IR/Dominators.h"
48#include "llvm/IR/Function.h"
49#include "llvm/IR/Instructions.h"
50#include "llvm/IR/Module.h"
51#include "llvm/MC/MCInstrItineraries.h"
52#include "llvm/Support/CommandLine.h"
53#include "llvm/Support/Debug.h"
54#include "llvm/Support/raw_ostream.h"
55#include "llvm/Target/TargetInstrInfo.h"
56#include "llvm/Target/TargetRegisterInfo.h"
57
58using namespace llvm;
59
60#define DEBUG_TYPE "stackcoloring"
61
62static cl::opt<bool>
63DisableColoring("no-stack-coloring",
64        cl::init(false), cl::Hidden,
65        cl::desc("Disable stack coloring"));
66
67/// The user may write code that uses allocas outside of the declared lifetime
68/// zone. This can happen when the user returns a reference to a local
69/// data-structure. We can detect these cases and decide not to optimize the
70/// code. If this flag is enabled, we try to save the user.
71static cl::opt<bool>
72ProtectFromEscapedAllocas("protect-from-escaped-allocas",
73                          cl::init(false), cl::Hidden,
74                          cl::desc("Do not optimize lifetime zones that "
75                                   "are broken"));
76
77STATISTIC(NumMarkerSeen,  "Number of lifetime markers found.");
78STATISTIC(StackSpaceSaved, "Number of bytes saved due to merging slots.");
79STATISTIC(StackSlotMerged, "Number of stack slot merged.");
80STATISTIC(EscapedAllocas, "Number of allocas that escaped the lifetime region");
81
82//===----------------------------------------------------------------------===//
83//                           StackColoring Pass
84//===----------------------------------------------------------------------===//
85
86namespace {
87/// StackColoring - A machine pass for merging disjoint stack allocations,
88/// marked by the LIFETIME_START and LIFETIME_END pseudo instructions.
89class StackColoring : public MachineFunctionPass {
90  MachineFrameInfo *MFI;
91  MachineFunction *MF;
92
93  /// A class representing liveness information for a single basic block.
94  /// Each bit in the BitVector represents the liveness property
95  /// for a different stack slot.
96  struct BlockLifetimeInfo {
97    /// Which slots BEGINs in each basic block.
98    BitVector Begin;
99    /// Which slots ENDs in each basic block.
100    BitVector End;
101    /// Which slots are marked as LIVE_IN, coming into each basic block.
102    BitVector LiveIn;
103    /// Which slots are marked as LIVE_OUT, coming out of each basic block.
104    BitVector LiveOut;
105  };
106
107  /// Maps active slots (per bit) for each basic block.
108  typedef DenseMap<const MachineBasicBlock*, BlockLifetimeInfo> LivenessMap;
109  LivenessMap BlockLiveness;
110
111  /// Maps serial numbers to basic blocks.
112  DenseMap<const MachineBasicBlock*, int> BasicBlocks;
113  /// Maps basic blocks to a serial number.
114  SmallVector<const MachineBasicBlock*, 8> BasicBlockNumbering;
115
116  /// Maps liveness intervals for each slot.
117  SmallVector<std::unique_ptr<LiveInterval>, 16> Intervals;
118  /// VNInfo is used for the construction of LiveIntervals.
119  VNInfo::Allocator VNInfoAllocator;
120  /// SlotIndex analysis object.
121  SlotIndexes *Indexes;
122  /// The stack protector object.
123  StackProtector *SP;
124
125  /// The list of lifetime markers found. These markers are to be removed
126  /// once the coloring is done.
127  SmallVector<MachineInstr*, 8> Markers;
128
129public:
130  static char ID;
131  StackColoring() : MachineFunctionPass(ID) {
132    initializeStackColoringPass(*PassRegistry::getPassRegistry());
133  }
134  void getAnalysisUsage(AnalysisUsage &AU) const override;
135  bool runOnMachineFunction(MachineFunction &MF) override;
136
137private:
138  /// Debug.
139  void dump() const;
140
141  /// Removes all of the lifetime marker instructions from the function.
142  /// \returns true if any markers were removed.
143  bool removeAllMarkers();
144
145  /// Scan the machine function and find all of the lifetime markers.
146  /// Record the findings in the BEGIN and END vectors.
147  /// \returns the number of markers found.
148  unsigned collectMarkers(unsigned NumSlot);
149
150  /// Perform the dataflow calculation and calculate the lifetime for each of
151  /// the slots, based on the BEGIN/END vectors. Set the LifetimeLIVE_IN and
152  /// LifetimeLIVE_OUT maps that represent which stack slots are live coming
153  /// in and out blocks.
154  void calculateLocalLiveness();
155
156  /// Construct the LiveIntervals for the slots.
157  void calculateLiveIntervals(unsigned NumSlots);
158
159  /// Go over the machine function and change instructions which use stack
160  /// slots to use the joint slots.
161  void remapInstructions(DenseMap<int, int> &SlotRemap);
162
163  /// The input program may contain instructions which are not inside lifetime
164  /// markers. This can happen due to a bug in the compiler or due to a bug in
165  /// user code (for example, returning a reference to a local variable).
166  /// This procedure checks all of the instructions in the function and
167  /// invalidates lifetime ranges which do not contain all of the instructions
168  /// which access that frame slot.
169  void removeInvalidSlotRanges();
170
171  /// Map entries which point to other entries to their destination.
172  ///   A->B->C becomes A->C.
173   void expungeSlotMap(DenseMap<int, int> &SlotRemap, unsigned NumSlots);
174};
175} // end anonymous namespace
176
177char StackColoring::ID = 0;
178char &llvm::StackColoringID = StackColoring::ID;
179
180INITIALIZE_PASS_BEGIN(StackColoring,
181                   "stack-coloring", "Merge disjoint stack slots", false, false)
182INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
183INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
184INITIALIZE_PASS_DEPENDENCY(StackProtector)
185INITIALIZE_PASS_END(StackColoring,
186                   "stack-coloring", "Merge disjoint stack slots", false, false)
187
188void StackColoring::getAnalysisUsage(AnalysisUsage &AU) const {
189  AU.addRequired<MachineDominatorTree>();
190  AU.addPreserved<MachineDominatorTree>();
191  AU.addRequired<SlotIndexes>();
192  AU.addRequired<StackProtector>();
193  MachineFunctionPass::getAnalysisUsage(AU);
194}
195
196void StackColoring::dump() const {
197  for (MachineBasicBlock *MBB : depth_first(MF)) {
198    DEBUG(dbgs() << "Inspecting block #" << BasicBlocks.lookup(MBB) << " ["
199                 << MBB->getName() << "]\n");
200
201    LivenessMap::const_iterator BI = BlockLiveness.find(MBB);
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 (MachineBasicBlock *MBB : depth_first(MF)) {
235
236    // Assign a serial number to this basic block.
237    BasicBlocks[MBB] = BasicBlockNumbering.size();
238    BasicBlockNumbering.push_back(MBB);
239
240    // Keep a reference to avoid repeated lookups.
241    BlockLifetimeInfo &BlockInfo = BlockLiveness[MBB];
242
243    BlockInfo.Begin.resize(NumSlot);
244    BlockInfo.End.resize(NumSlot);
245
246    for (MachineInstr &MI : *MBB) {
247      if (MI.getOpcode() != TargetOpcode::LIFETIME_START &&
248          MI.getOpcode() != TargetOpcode::LIFETIME_END)
249        continue;
250
251      Markers.push_back(&MI);
252
253      bool IsStart = MI.getOpcode() == TargetOpcode::LIFETIME_START;
254      const MachineOperand &MO = MI.getOperand(0);
255      unsigned Slot = MO.getIndex();
256
257      MarkersFound++;
258
259      const AllocaInst *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        BlockInfo.Begin.set(Slot);
267      } else {
268        if (BlockInfo.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          BlockInfo.Begin.reset(Slot);
273        } else {
274          BlockInfo.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<const 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<const MachineBasicBlock*, 8> NextBBSet;
300
301    for (const MachineBasicBlock *BB : BasicBlockNumbering) {
302      if (!BBSet.count(BB)) continue;
303
304      // Use an iterator to avoid repeated lookups.
305      LivenessMap::iterator BI = BlockLiveness.find(BB);
306      assert(BI != BlockLiveness.end() && "Block not found");
307      BlockLifetimeInfo &BlockInfo = BI->second;
308
309      BitVector LocalLiveIn;
310      BitVector LocalLiveOut;
311
312      // Forward propagation from begins to ends.
313      for (MachineBasicBlock::const_pred_iterator PI = BB->pred_begin(),
314           PE = BB->pred_end(); PI != PE; ++PI) {
315        LivenessMap::const_iterator I = BlockLiveness.find(*PI);
316        assert(I != BlockLiveness.end() && "Predecessor not found");
317        LocalLiveIn |= I->second.LiveOut;
318      }
319      LocalLiveIn |= BlockInfo.End;
320      LocalLiveIn.reset(BlockInfo.Begin);
321
322      // Reverse propagation from ends to begins.
323      for (MachineBasicBlock::const_succ_iterator SI = BB->succ_begin(),
324           SE = BB->succ_end(); SI != SE; ++SI) {
325        LivenessMap::const_iterator I = BlockLiveness.find(*SI);
326        assert(I != BlockLiveness.end() && "Successor not found");
327        LocalLiveOut |= I->second.LiveIn;
328      }
329      LocalLiveOut |= BlockInfo.Begin;
330      LocalLiveOut.reset(BlockInfo.End);
331
332      LocalLiveIn |= LocalLiveOut;
333      LocalLiveOut |= LocalLiveIn;
334
335      // After adopting the live bits, we need to turn-off the bits which
336      // are de-activated in this block.
337      LocalLiveOut.reset(BlockInfo.End);
338      LocalLiveIn.reset(BlockInfo.Begin);
339
340      // If we have both BEGIN and END markers in the same basic block then
341      // we know that the BEGIN marker comes after the END, because we already
342      // handle the case where the BEGIN comes before the END when collecting
343      // the markers (and building the BEGIN/END vectore).
344      // Want to enable the LIVE_IN and LIVE_OUT of slots that have both
345      // BEGIN and END because it means that the value lives before and after
346      // this basic block.
347      BitVector LocalEndBegin = BlockInfo.End;
348      LocalEndBegin &= BlockInfo.Begin;
349      LocalLiveIn |= LocalEndBegin;
350      LocalLiveOut |= LocalEndBegin;
351
352      if (LocalLiveIn.test(BlockInfo.LiveIn)) {
353        changed = true;
354        BlockInfo.LiveIn |= LocalLiveIn;
355
356        NextBBSet.insert(BB->pred_begin(), BB->pred_end());
357      }
358
359      if (LocalLiveOut.test(BlockInfo.LiveOut)) {
360        changed = true;
361        BlockInfo.LiveOut |= LocalLiveOut;
362
363        NextBBSet.insert(BB->succ_begin(), BB->succ_end());
364      }
365    }
366
367    BBSet = NextBBSet;
368  }// while changed.
369}
370
371void StackColoring::calculateLiveIntervals(unsigned NumSlots) {
372  SmallVector<SlotIndex, 16> Starts;
373  SmallVector<SlotIndex, 16> Finishes;
374
375  // For each block, find which slots are active within this block
376  // and update the live intervals.
377  for (const MachineBasicBlock &MBB : *MF) {
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 (const MachineInstr *MI : Markers) {
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      const 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    BlockLifetimeInfo &MBBLiveness = BlockLiveness[&MBB];
410    for (int pos = MBBLiveness.LiveIn.find_first(); pos != -1;
411         pos = MBBLiveness.LiveIn.find_next(pos)) {
412      Starts[pos] = Indexes->getMBBStartIdx(&MBB);
413    }
414    for (int pos = MBBLiveness.LiveOut.find_first(); pos != -1;
415         pos = MBBLiveness.LiveOut.find_next(pos)) {
416      Finishes[pos] = Indexes->getMBBEndIdx(&MBB);
417    }
418
419    for (unsigned i = 0; i < NumSlots; ++i) {
420      assert(Starts[i].isValid() == Finishes[i].isValid() && "Unmatched range");
421      if (!Starts[i].isValid())
422        continue;
423
424      assert(Starts[i] && Finishes[i] && "Invalid interval");
425      VNInfo *ValNum = Intervals[i]->getValNumInfo(0);
426      SlotIndex S = Starts[i];
427      SlotIndex F = Finishes[i];
428      if (S < F) {
429        // We have a single consecutive region.
430        Intervals[i]->addSegment(LiveInterval::Segment(S, F, ValNum));
431      } else {
432        // We have two non-consecutive regions. This happens when
433        // LIFETIME_START appears after the LIFETIME_END marker.
434        SlotIndex NewStart = Indexes->getMBBStartIdx(&MBB);
435        SlotIndex NewFin = Indexes->getMBBEndIdx(&MBB);
436        Intervals[i]->addSegment(LiveInterval::Segment(NewStart, F, ValNum));
437        Intervals[i]->addSegment(LiveInterval::Segment(S, NewFin, ValNum));
438      }
439    }
440  }
441}
442
443bool StackColoring::removeAllMarkers() {
444  unsigned Count = 0;
445  for (MachineInstr *MI : Markers) {
446    MI->eraseFromParent();
447    Count++;
448  }
449  Markers.clear();
450
451  DEBUG(dbgs()<<"Removed "<<Count<<" markers.\n");
452  return Count;
453}
454
455void StackColoring::remapInstructions(DenseMap<int, int> &SlotRemap) {
456  unsigned FixedInstr = 0;
457  unsigned FixedMemOp = 0;
458  unsigned FixedDbg = 0;
459  MachineModuleInfo *MMI = &MF->getMMI();
460
461  // Remap debug information that refers to stack slots.
462  for (auto &VI : MMI->getVariableDbgInfo()) {
463    if (!VI.Var)
464      continue;
465    if (SlotRemap.count(VI.Slot)) {
466      DEBUG(dbgs()<<"Remapping debug info for ["<<VI.Var->getName()<<"].\n");
467      VI.Slot = SlotRemap[VI.Slot];
468      FixedDbg++;
469    }
470  }
471
472  // Keep a list of *allocas* which need to be remapped.
473  DenseMap<const AllocaInst*, const AllocaInst*> Allocas;
474  for (const std::pair<int, int> &SI : SlotRemap) {
475    const AllocaInst *From = MFI->getObjectAllocation(SI.first);
476    const AllocaInst *To = MFI->getObjectAllocation(SI.second);
477    assert(To && From && "Invalid allocation object");
478    Allocas[From] = To;
479
480    // AA might be used later for instruction scheduling, and we need it to be
481    // able to deduce the correct aliasing releationships between pointers
482    // derived from the alloca being remapped and the target of that remapping.
483    // The only safe way, without directly informing AA about the remapping
484    // somehow, is to directly update the IR to reflect the change being made
485    // here.
486    Instruction *Inst = const_cast<AllocaInst *>(To);
487    if (From->getType() != To->getType()) {
488      BitCastInst *Cast = new BitCastInst(Inst, From->getType());
489      Cast->insertAfter(Inst);
490      Inst = Cast;
491    }
492
493    // Allow the stack protector to adjust its value map to account for the
494    // upcoming replacement.
495    SP->adjustForColoring(From, To);
496
497    // Note that this will not replace uses in MMOs (which we'll update below),
498    // or anywhere else (which is why we won't delete the original
499    // instruction).
500    const_cast<AllocaInst *>(From)->replaceAllUsesWith(Inst);
501  }
502
503  // Remap all instructions to the new stack slots.
504  for (MachineBasicBlock &BB : *MF)
505    for (MachineInstr &I : BB) {
506      // Skip lifetime markers. We'll remove them soon.
507      if (I.getOpcode() == TargetOpcode::LIFETIME_START ||
508          I.getOpcode() == TargetOpcode::LIFETIME_END)
509        continue;
510
511      // Update the MachineMemOperand to use the new alloca.
512      for (MachineMemOperand *MMO : I.memoperands()) {
513        // FIXME: In order to enable the use of TBAA when using AA in CodeGen,
514        // we'll also need to update the TBAA nodes in MMOs with values
515        // derived from the merged allocas. When doing this, we'll need to use
516        // the same variant of GetUnderlyingObjects that is used by the
517        // instruction scheduler (that can look through ptrtoint/inttoptr
518        // pairs).
519
520        // We've replaced IR-level uses of the remapped allocas, so we only
521        // need to replace direct uses here.
522        const AllocaInst *AI = dyn_cast_or_null<AllocaInst>(MMO->getValue());
523        if (!AI)
524          continue;
525
526        if (!Allocas.count(AI))
527          continue;
528
529        MMO->setValue(Allocas[AI]);
530        FixedMemOp++;
531      }
532
533      // Update all of the machine instruction operands.
534      for (MachineOperand &MO : I.operands()) {
535        if (!MO.isFI())
536          continue;
537        int FromSlot = MO.getIndex();
538
539        // Don't touch arguments.
540        if (FromSlot<0)
541          continue;
542
543        // Only look at mapped slots.
544        if (!SlotRemap.count(FromSlot))
545          continue;
546
547        // In a debug build, check that the instruction that we are modifying is
548        // inside the expected live range. If the instruction is not inside
549        // the calculated range then it means that the alloca usage moved
550        // outside of the lifetime markers, or that the user has a bug.
551        // NOTE: Alloca address calculations which happen outside the lifetime
552        // zone are are okay, despite the fact that we don't have a good way
553        // for validating all of the usages of the calculation.
554#ifndef NDEBUG
555        bool TouchesMemory = I.mayLoad() || I.mayStore();
556        // If we *don't* protect the user from escaped allocas, don't bother
557        // validating the instructions.
558        if (!I.isDebugValue() && TouchesMemory && ProtectFromEscapedAllocas) {
559          SlotIndex Index = Indexes->getInstructionIndex(&I);
560          const LiveInterval *Interval = &*Intervals[FromSlot];
561          assert(Interval->find(Index) != Interval->end() &&
562                 "Found instruction usage outside of live range.");
563        }
564#endif
565
566        // Fix the machine instructions.
567        int ToSlot = SlotRemap[FromSlot];
568        MO.setIndex(ToSlot);
569        FixedInstr++;
570      }
571    }
572
573  DEBUG(dbgs()<<"Fixed "<<FixedMemOp<<" machine memory operands.\n");
574  DEBUG(dbgs()<<"Fixed "<<FixedDbg<<" debug locations.\n");
575  DEBUG(dbgs()<<"Fixed "<<FixedInstr<<" machine instructions.\n");
576}
577
578void StackColoring::removeInvalidSlotRanges() {
579  for (MachineBasicBlock &BB : *MF)
580    for (MachineInstr &I : BB) {
581      if (I.getOpcode() == TargetOpcode::LIFETIME_START ||
582          I.getOpcode() == TargetOpcode::LIFETIME_END || I.isDebugValue())
583        continue;
584
585      // Some intervals are suspicious! In some cases we find address
586      // calculations outside of the lifetime zone, but not actual memory
587      // read or write. Memory accesses outside of the lifetime zone are a clear
588      // violation, but address calculations are okay. This can happen when
589      // GEPs are hoisted outside of the lifetime zone.
590      // So, in here we only check instructions which can read or write memory.
591      if (!I.mayLoad() && !I.mayStore())
592        continue;
593
594      // Check all of the machine operands.
595      for (const MachineOperand &MO : I.operands()) {
596        if (!MO.isFI())
597          continue;
598
599        int Slot = MO.getIndex();
600
601        if (Slot<0)
602          continue;
603
604        if (Intervals[Slot]->empty())
605          continue;
606
607        // Check that the used slot is inside the calculated lifetime range.
608        // If it is not, warn about it and invalidate the range.
609        LiveInterval *Interval = &*Intervals[Slot];
610        SlotIndex Index = Indexes->getInstructionIndex(&I);
611        if (Interval->find(Index) == Interval->end()) {
612          Interval->clear();
613          DEBUG(dbgs()<<"Invalidating range #"<<Slot<<"\n");
614          EscapedAllocas++;
615        }
616      }
617    }
618}
619
620void StackColoring::expungeSlotMap(DenseMap<int, int> &SlotRemap,
621                                   unsigned NumSlots) {
622  // Expunge slot remap map.
623  for (unsigned i=0; i < NumSlots; ++i) {
624    // If we are remapping i
625    if (SlotRemap.count(i)) {
626      int Target = SlotRemap[i];
627      // As long as our target is mapped to something else, follow it.
628      while (SlotRemap.count(Target)) {
629        Target = SlotRemap[Target];
630        SlotRemap[i] = Target;
631      }
632    }
633  }
634}
635
636bool StackColoring::runOnMachineFunction(MachineFunction &Func) {
637  if (skipOptnoneFunction(*Func.getFunction()))
638    return false;
639
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  SP = &getAnalysis<StackProtector>();
647  BlockLiveness.clear();
648  BasicBlocks.clear();
649  BasicBlockNumbering.clear();
650  Markers.clear();
651  Intervals.clear();
652  VNInfoAllocator.Reset();
653
654  unsigned NumSlots = MFI->getObjectIndexEnd();
655
656  // If there are no stack slots then there are no markers to remove.
657  if (!NumSlots)
658    return false;
659
660  SmallVector<int, 8> SortedSlots;
661
662  SortedSlots.reserve(NumSlots);
663  Intervals.reserve(NumSlots);
664
665  unsigned NumMarkers = collectMarkers(NumSlots);
666
667  unsigned TotalSize = 0;
668  DEBUG(dbgs()<<"Found "<<NumMarkers<<" markers and "<<NumSlots<<" slots\n");
669  DEBUG(dbgs()<<"Slot structure:\n");
670
671  for (int i=0; i < MFI->getObjectIndexEnd(); ++i) {
672    DEBUG(dbgs()<<"Slot #"<<i<<" - "<<MFI->getObjectSize(i)<<" bytes.\n");
673    TotalSize += MFI->getObjectSize(i);
674  }
675
676  DEBUG(dbgs()<<"Total Stack size: "<<TotalSize<<" bytes\n\n");
677
678  // Don't continue because there are not enough lifetime markers, or the
679  // stack is too small, or we are told not to optimize the slots.
680  if (NumMarkers < 2 || TotalSize < 16 || DisableColoring) {
681    DEBUG(dbgs()<<"Will not try to merge slots.\n");
682    return removeAllMarkers();
683  }
684
685  for (unsigned i=0; i < NumSlots; ++i) {
686    std::unique_ptr<LiveInterval> LI(new LiveInterval(i, 0));
687    LI->getNextValue(Indexes->getZeroIndex(), VNInfoAllocator);
688    Intervals.push_back(std::move(LI));
689    SortedSlots.push_back(i);
690  }
691
692  // Calculate the liveness of each block.
693  calculateLocalLiveness();
694
695  // Propagate the liveness information.
696  calculateLiveIntervals(NumSlots);
697
698  // Search for allocas which are used outside of the declared lifetime
699  // markers.
700  if (ProtectFromEscapedAllocas)
701    removeInvalidSlotRanges();
702
703  // Maps old slots to new slots.
704  DenseMap<int, int> SlotRemap;
705  unsigned RemovedSlots = 0;
706  unsigned ReducedSize = 0;
707
708  // Do not bother looking at empty intervals.
709  for (unsigned I = 0; I < NumSlots; ++I) {
710    if (Intervals[SortedSlots[I]]->empty())
711      SortedSlots[I] = -1;
712  }
713
714  // This is a simple greedy algorithm for merging allocas. First, sort the
715  // slots, placing the largest slots first. Next, perform an n^2 scan and look
716  // for disjoint slots. When you find disjoint slots, merge the samller one
717  // into the bigger one and update the live interval. Remove the small alloca
718  // and continue.
719
720  // Sort the slots according to their size. Place unused slots at the end.
721  // Use stable sort to guarantee deterministic code generation.
722  std::stable_sort(SortedSlots.begin(), SortedSlots.end(),
723                   [this](int LHS, int RHS) {
724    // We use -1 to denote a uninteresting slot. Place these slots at the end.
725    if (LHS == -1) return false;
726    if (RHS == -1) return true;
727    // Sort according to size.
728    return MFI->getObjectSize(LHS) > MFI->getObjectSize(RHS);
729  });
730
731  bool Changed = true;
732  while (Changed) {
733    Changed = false;
734    for (unsigned I = 0; I < NumSlots; ++I) {
735      if (SortedSlots[I] == -1)
736        continue;
737
738      for (unsigned J=I+1; J < NumSlots; ++J) {
739        if (SortedSlots[J] == -1)
740          continue;
741
742        int FirstSlot = SortedSlots[I];
743        int SecondSlot = SortedSlots[J];
744        LiveInterval *First = &*Intervals[FirstSlot];
745        LiveInterval *Second = &*Intervals[SecondSlot];
746        assert (!First->empty() && !Second->empty() && "Found an empty range");
747
748        // Merge disjoint slots.
749        if (!First->overlaps(*Second)) {
750          Changed = true;
751          First->MergeSegmentsInAsValue(*Second, First->getValNumInfo(0));
752          SlotRemap[SecondSlot] = FirstSlot;
753          SortedSlots[J] = -1;
754          DEBUG(dbgs()<<"Merging #"<<FirstSlot<<" and slots #"<<
755                SecondSlot<<" together.\n");
756          unsigned MaxAlignment = std::max(MFI->getObjectAlignment(FirstSlot),
757                                           MFI->getObjectAlignment(SecondSlot));
758
759          assert(MFI->getObjectSize(FirstSlot) >=
760                 MFI->getObjectSize(SecondSlot) &&
761                 "Merging a small object into a larger one");
762
763          RemovedSlots+=1;
764          ReducedSize += MFI->getObjectSize(SecondSlot);
765          MFI->setObjectAlignment(FirstSlot, MaxAlignment);
766          MFI->RemoveStackObject(SecondSlot);
767        }
768      }
769    }
770  }// While changed.
771
772  // Record statistics.
773  StackSpaceSaved += ReducedSize;
774  StackSlotMerged += RemovedSlots;
775  DEBUG(dbgs()<<"Merge "<<RemovedSlots<<" slots. Saved "<<
776        ReducedSize<<" bytes\n");
777
778  // Scan the entire function and update all machine operands that use frame
779  // indices to use the remapped frame index.
780  expungeSlotMap(SlotRemap, NumSlots);
781  remapInstructions(SlotRemap);
782
783  return removeAllMarkers();
784}
785