1 //===-- StackColoring.cpp -------------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
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
19 // 2. Merge this pass with StackSlotColoring and allow merging of allocas with
22 //===----------------------------------------------------------------------===//
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/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"
62 DisableColoring("no-stack-coloring",
63 cl::init(false), cl::Hidden,
64 cl::desc("Disable stack coloring"));
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.
71 ProtectFromEscapedAllocas("protect-from-escaped-allocas",
72 cl::init(false), cl::Hidden,
73 cl::desc("Do not optimize lifetime zones that "
76 STATISTIC(NumMarkerSeen, "Number of lifetime markers found.");
77 STATISTIC(StackSpaceSaved, "Number of bytes saved due to merging slots.");
78 STATISTIC(StackSlotMerged, "Number of stack slot merged.");
79 STATISTIC(EscapedAllocas, "Number of allocas that escaped the lifetime region");
81 //===----------------------------------------------------------------------===//
83 //===----------------------------------------------------------------------===//
86 /// StackColoring - A machine pass for merging disjoint stack allocations,
87 /// marked by the LIFETIME_START and LIFETIME_END pseudo instructions.
88 class StackColoring : public MachineFunctionPass {
89 MachineFrameInfo *MFI;
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.
98 /// Which slots ENDs in each basic block.
100 /// Which slots are marked as LIVE_IN, coming into each basic block.
102 /// Which slots are marked as LIVE_OUT, coming out of each basic block.
106 /// Maps active slots (per bit) for each basic block.
107 typedef DenseMap<const MachineBasicBlock*, BlockLifetimeInfo> LivenessMap;
108 LivenessMap BlockLiveness;
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;
115 /// Maps liveness intervals for each slot.
116 SmallVector<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.
124 /// The list of lifetime markers found. These markers are to be removed
125 /// once the coloring is done.
126 SmallVector<MachineInstr*, 8> Markers;
130 StackColoring() : MachineFunctionPass(ID) {
131 initializeStackColoringPass(*PassRegistry::getPassRegistry());
133 void getAnalysisUsage(AnalysisUsage &AU) const;
134 bool runOnMachineFunction(MachineFunction &MF);
140 /// Removes all of the lifetime marker instructions from the function.
141 /// \returns true if any markers were removed.
142 bool removeAllMarkers();
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);
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();
155 /// Construct the LiveIntervals for the slots.
156 void calculateLiveIntervals(unsigned NumSlots);
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);
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();
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);
174 } // end anonymous namespace
176 char StackColoring::ID = 0;
177 char &llvm::StackColoringID = StackColoring::ID;
179 INITIALIZE_PASS_BEGIN(StackColoring,
180 "stack-coloring", "Merge disjoint stack slots", false, false)
181 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
182 INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
183 INITIALIZE_PASS_DEPENDENCY(StackProtector)
184 INITIALIZE_PASS_END(StackColoring,
185 "stack-coloring", "Merge disjoint stack slots", false, false)
187 void 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);
195 void StackColoring::dump() const {
196 for (df_iterator<MachineFunction*> FI = df_begin(MF), FE = df_end(MF);
198 DEBUG(dbgs()<<"Inspecting block #"<<BasicBlocks.lookup(*FI)<<
199 " ["<<FI->getName()<<"]\n");
201 LivenessMap::const_iterator BI = BlockLiveness.find(*FI);
202 assert(BI != BlockLiveness.end() && "Block not found");
203 const BlockLifetimeInfo &BlockInfo = BI->second;
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");
210 DEBUG(dbgs()<<"END : {");
211 for (unsigned i=0; i < BlockInfo.End.size(); ++i)
212 DEBUG(dbgs()<<BlockInfo.End.test(i)<<" ");
214 DEBUG(dbgs()<<"}\n");
216 DEBUG(dbgs()<<"LIVE_IN: {");
217 for (unsigned i=0; i < BlockInfo.LiveIn.size(); ++i)
218 DEBUG(dbgs()<<BlockInfo.LiveIn.test(i)<<" ");
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");
228 unsigned 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);
237 // Assign a serial number to this basic block.
238 BasicBlocks[*FI] = BasicBlockNumbering.size();
239 BasicBlockNumbering.push_back(*FI);
241 // Keep a reference to avoid repeated lookups.
242 BlockLifetimeInfo &BlockInfo = BlockLiveness[*FI];
244 BlockInfo.Begin.resize(NumSlot);
245 BlockInfo.End.resize(NumSlot);
247 for (MachineBasicBlock::iterator BI = (*FI)->begin(), BE = (*FI)->end();
250 if (BI->getOpcode() != TargetOpcode::LIFETIME_START &&
251 BI->getOpcode() != TargetOpcode::LIFETIME_END)
254 Markers.push_back(BI);
256 bool IsStart = BI->getOpcode() == TargetOpcode::LIFETIME_START;
257 const MachineOperand &MI = BI->getOperand(0);
258 unsigned Slot = MI.getIndex();
262 const AllocaInst *Allocation = MFI->getObjectAllocation(Slot);
264 DEBUG(dbgs()<<"Found a lifetime marker for slot #"<<Slot<<
265 " with allocation: "<< Allocation->getName()<<"\n");
269 BlockInfo.Begin.set(Slot);
271 if (BlockInfo.Begin.test(Slot)) {
272 // Allocas that start and end within a single block are handled
273 // specially when computing the LiveIntervals to avoid pessimizing
274 // the liveness propagation.
275 BlockInfo.Begin.reset(Slot);
277 BlockInfo.End.set(Slot);
283 // Update statistics.
284 NumMarkerSeen += MarkersFound;
288 void StackColoring::calculateLocalLiveness() {
289 // Perform a standard reverse dataflow computation to solve for
290 // global liveness. The BEGIN set here is equivalent to KILL in the standard
291 // formulation, and END is equivalent to GEN. The result of this computation
292 // is a map from blocks to bitvectors where the bitvectors represent which
293 // allocas are live in/out of that block.
294 SmallPtrSet<const MachineBasicBlock*, 8> BBSet(BasicBlockNumbering.begin(),
295 BasicBlockNumbering.end());
296 unsigned NumSSMIters = 0;
302 SmallPtrSet<const MachineBasicBlock*, 8> NextBBSet;
304 for (SmallVectorImpl<const MachineBasicBlock *>::iterator
305 PI = BasicBlockNumbering.begin(), PE = BasicBlockNumbering.end();
308 const MachineBasicBlock *BB = *PI;
309 if (!BBSet.count(BB)) continue;
311 // Use an iterator to avoid repeated lookups.
312 LivenessMap::iterator BI = BlockLiveness.find(BB);
313 assert(BI != BlockLiveness.end() && "Block not found");
314 BlockLifetimeInfo &BlockInfo = BI->second;
316 BitVector LocalLiveIn;
317 BitVector LocalLiveOut;
319 // Forward propagation from begins to ends.
320 for (MachineBasicBlock::const_pred_iterator PI = BB->pred_begin(),
321 PE = BB->pred_end(); PI != PE; ++PI) {
322 LivenessMap::const_iterator I = BlockLiveness.find(*PI);
323 assert(I != BlockLiveness.end() && "Predecessor not found");
324 LocalLiveIn |= I->second.LiveOut;
326 LocalLiveIn |= BlockInfo.End;
327 LocalLiveIn.reset(BlockInfo.Begin);
329 // Reverse propagation from ends to begins.
330 for (MachineBasicBlock::const_succ_iterator SI = BB->succ_begin(),
331 SE = BB->succ_end(); SI != SE; ++SI) {
332 LivenessMap::const_iterator I = BlockLiveness.find(*SI);
333 assert(I != BlockLiveness.end() && "Successor not found");
334 LocalLiveOut |= I->second.LiveIn;
336 LocalLiveOut |= BlockInfo.Begin;
337 LocalLiveOut.reset(BlockInfo.End);
339 LocalLiveIn |= LocalLiveOut;
340 LocalLiveOut |= LocalLiveIn;
342 // After adopting the live bits, we need to turn-off the bits which
343 // are de-activated in this block.
344 LocalLiveOut.reset(BlockInfo.End);
345 LocalLiveIn.reset(BlockInfo.Begin);
347 // If we have both BEGIN and END markers in the same basic block then
348 // we know that the BEGIN marker comes after the END, because we already
349 // handle the case where the BEGIN comes before the END when collecting
350 // the markers (and building the BEGIN/END vectore).
351 // Want to enable the LIVE_IN and LIVE_OUT of slots that have both
352 // BEGIN and END because it means that the value lives before and after
354 BitVector LocalEndBegin = BlockInfo.End;
355 LocalEndBegin &= BlockInfo.Begin;
356 LocalLiveIn |= LocalEndBegin;
357 LocalLiveOut |= LocalEndBegin;
359 if (LocalLiveIn.test(BlockInfo.LiveIn)) {
361 BlockInfo.LiveIn |= LocalLiveIn;
363 for (MachineBasicBlock::const_pred_iterator PI = BB->pred_begin(),
364 PE = BB->pred_end(); PI != PE; ++PI)
365 NextBBSet.insert(*PI);
368 if (LocalLiveOut.test(BlockInfo.LiveOut)) {
370 BlockInfo.LiveOut |= LocalLiveOut;
372 for (MachineBasicBlock::const_succ_iterator SI = BB->succ_begin(),
373 SE = BB->succ_end(); SI != SE; ++SI)
374 NextBBSet.insert(*SI);
382 void StackColoring::calculateLiveIntervals(unsigned NumSlots) {
383 SmallVector<SlotIndex, 16> Starts;
384 SmallVector<SlotIndex, 16> Finishes;
386 // For each block, find which slots are active within this block
387 // and update the live intervals.
388 for (MachineFunction::iterator MBB = MF->begin(), MBBe = MF->end();
389 MBB != MBBe; ++MBB) {
391 Starts.resize(NumSlots);
393 Finishes.resize(NumSlots);
395 // Create the interval for the basic blocks with lifetime markers in them.
396 for (SmallVectorImpl<MachineInstr*>::const_iterator it = Markers.begin(),
397 e = Markers.end(); it != e; ++it) {
398 const MachineInstr *MI = *it;
399 if (MI->getParent() != MBB)
402 assert((MI->getOpcode() == TargetOpcode::LIFETIME_START ||
403 MI->getOpcode() == TargetOpcode::LIFETIME_END) &&
404 "Invalid Lifetime marker");
406 bool IsStart = MI->getOpcode() == TargetOpcode::LIFETIME_START;
407 const MachineOperand &Mo = MI->getOperand(0);
408 int Slot = Mo.getIndex();
409 assert(Slot >= 0 && "Invalid slot");
411 SlotIndex ThisIndex = Indexes->getInstructionIndex(MI);
414 if (!Starts[Slot].isValid() || Starts[Slot] > ThisIndex)
415 Starts[Slot] = ThisIndex;
417 if (!Finishes[Slot].isValid() || Finishes[Slot] < ThisIndex)
418 Finishes[Slot] = ThisIndex;
422 // Create the interval of the blocks that we previously found to be 'alive'.
423 BlockLifetimeInfo &MBBLiveness = BlockLiveness[MBB];
424 for (int pos = MBBLiveness.LiveIn.find_first(); pos != -1;
425 pos = MBBLiveness.LiveIn.find_next(pos)) {
426 Starts[pos] = Indexes->getMBBStartIdx(MBB);
428 for (int pos = MBBLiveness.LiveOut.find_first(); pos != -1;
429 pos = MBBLiveness.LiveOut.find_next(pos)) {
430 Finishes[pos] = Indexes->getMBBEndIdx(MBB);
433 for (unsigned i = 0; i < NumSlots; ++i) {
434 assert(Starts[i].isValid() == Finishes[i].isValid() && "Unmatched range");
435 if (!Starts[i].isValid())
438 assert(Starts[i] && Finishes[i] && "Invalid interval");
439 VNInfo *ValNum = Intervals[i]->getValNumInfo(0);
440 SlotIndex S = Starts[i];
441 SlotIndex F = Finishes[i];
443 // We have a single consecutive region.
444 Intervals[i]->addSegment(LiveInterval::Segment(S, F, ValNum));
446 // We have two non-consecutive regions. This happens when
447 // LIFETIME_START appears after the LIFETIME_END marker.
448 SlotIndex NewStart = Indexes->getMBBStartIdx(MBB);
449 SlotIndex NewFin = Indexes->getMBBEndIdx(MBB);
450 Intervals[i]->addSegment(LiveInterval::Segment(NewStart, F, ValNum));
451 Intervals[i]->addSegment(LiveInterval::Segment(S, NewFin, ValNum));
457 bool StackColoring::removeAllMarkers() {
459 for (unsigned i = 0; i < Markers.size(); ++i) {
460 Markers[i]->eraseFromParent();
465 DEBUG(dbgs()<<"Removed "<<Count<<" markers.\n");
469 void StackColoring::remapInstructions(DenseMap<int, int> &SlotRemap) {
470 unsigned FixedInstr = 0;
471 unsigned FixedMemOp = 0;
472 unsigned FixedDbg = 0;
473 MachineModuleInfo *MMI = &MF->getMMI();
475 // Remap debug information that refers to stack slots.
476 MachineModuleInfo::VariableDbgInfoMapTy &VMap = MMI->getVariableDbgInfo();
477 for (MachineModuleInfo::VariableDbgInfoMapTy::iterator VI = VMap.begin(),
478 VE = VMap.end(); VI != VE; ++VI) {
479 const MDNode *Var = VI->first;
481 std::pair<unsigned, DebugLoc> &VP = VI->second;
482 if (SlotRemap.count(VP.first)) {
483 DEBUG(dbgs()<<"Remapping debug info for ["<<Var->getName()<<"].\n");
484 VP.first = SlotRemap[VP.first];
489 // Keep a list of *allocas* which need to be remapped.
490 DenseMap<const AllocaInst*, const AllocaInst*> Allocas;
491 for (DenseMap<int, int>::const_iterator it = SlotRemap.begin(),
492 e = SlotRemap.end(); it != e; ++it) {
493 const AllocaInst *From = MFI->getObjectAllocation(it->first);
494 const AllocaInst *To = MFI->getObjectAllocation(it->second);
495 assert(To && From && "Invalid allocation object");
498 // AA might be used later for instruction scheduling, and we need it to be
499 // able to deduce the correct aliasing releationships between pointers
500 // derived from the alloca being remapped and the target of that remapping.
501 // The only safe way, without directly informing AA about the remapping
502 // somehow, is to directly update the IR to reflect the change being made
504 Instruction *Inst = const_cast<AllocaInst *>(To);
505 if (From->getType() != To->getType()) {
506 BitCastInst *Cast = new BitCastInst(Inst, From->getType());
507 Cast->insertAfter(Inst);
511 // Allow the stack protector to adjust its value map to account for the
512 // upcoming replacement.
513 SP->adjustForColoring(From, To);
515 // Note that this will not replace uses in MMOs (which we'll update below),
516 // or anywhere else (which is why we won't delete the original
518 const_cast<AllocaInst *>(From)->replaceAllUsesWith(Inst);
521 // Remap all instructions to the new stack slots.
522 MachineFunction::iterator BB, BBE;
523 MachineBasicBlock::iterator I, IE;
524 for (BB = MF->begin(), BBE = MF->end(); BB != BBE; ++BB)
525 for (I = BB->begin(), IE = BB->end(); I != IE; ++I) {
527 // Skip lifetime markers. We'll remove them soon.
528 if (I->getOpcode() == TargetOpcode::LIFETIME_START ||
529 I->getOpcode() == TargetOpcode::LIFETIME_END)
532 // Update the MachineMemOperand to use the new alloca.
533 for (MachineInstr::mmo_iterator MM = I->memoperands_begin(),
534 E = I->memoperands_end(); MM != E; ++MM) {
535 MachineMemOperand *MMO = *MM;
537 const Value *V = MMO->getValue();
542 // FIXME: In order to enable the use of TBAA when using AA in CodeGen,
543 // we'll also need to update the TBAA nodes in MMOs with values
544 // derived from the merged allocas. When doing this, we'll need to use
545 // the same variant of GetUnderlyingObjects that is used by the
546 // instruction scheduler (that can look through ptrtoint/inttoptr
549 // We've replaced IR-level uses of the remapped allocas, so we only
550 // need to replace direct uses here.
551 if (!isa<AllocaInst>(V))
554 const AllocaInst *AI= cast<AllocaInst>(V);
555 if (!Allocas.count(AI))
558 MMO->setValue(Allocas[AI]);
562 // Update all of the machine instruction operands.
563 for (unsigned i = 0 ; i < I->getNumOperands(); ++i) {
564 MachineOperand &MO = I->getOperand(i);
568 int FromSlot = MO.getIndex();
570 // Don't touch arguments.
574 // Only look at mapped slots.
575 if (!SlotRemap.count(FromSlot))
578 // In a debug build, check that the instruction that we are modifying is
579 // inside the expected live range. If the instruction is not inside
580 // the calculated range then it means that the alloca usage moved
581 // outside of the lifetime markers, or that the user has a bug.
582 // NOTE: Alloca address calculations which happen outside the lifetime
583 // zone are are okay, despite the fact that we don't have a good way
584 // for validating all of the usages of the calculation.
586 bool TouchesMemory = I->mayLoad() || I->mayStore();
587 // If we *don't* protect the user from escaped allocas, don't bother
588 // validating the instructions.
589 if (!I->isDebugValue() && TouchesMemory && ProtectFromEscapedAllocas) {
590 SlotIndex Index = Indexes->getInstructionIndex(I);
591 LiveInterval *Interval = Intervals[FromSlot];
592 assert(Interval->find(Index) != Interval->end() &&
593 "Found instruction usage outside of live range.");
597 // Fix the machine instructions.
598 int ToSlot = SlotRemap[FromSlot];
604 DEBUG(dbgs()<<"Fixed "<<FixedMemOp<<" machine memory operands.\n");
605 DEBUG(dbgs()<<"Fixed "<<FixedDbg<<" debug locations.\n");
606 DEBUG(dbgs()<<"Fixed "<<FixedInstr<<" machine instructions.\n");
609 void StackColoring::removeInvalidSlotRanges() {
610 MachineFunction::const_iterator BB, BBE;
611 MachineBasicBlock::const_iterator I, IE;
612 for (BB = MF->begin(), BBE = MF->end(); BB != BBE; ++BB)
613 for (I = BB->begin(), IE = BB->end(); I != IE; ++I) {
615 if (I->getOpcode() == TargetOpcode::LIFETIME_START ||
616 I->getOpcode() == TargetOpcode::LIFETIME_END || I->isDebugValue())
619 // Some intervals are suspicious! In some cases we find address
620 // calculations outside of the lifetime zone, but not actual memory
621 // read or write. Memory accesses outside of the lifetime zone are a clear
622 // violation, but address calculations are okay. This can happen when
623 // GEPs are hoisted outside of the lifetime zone.
624 // So, in here we only check instructions which can read or write memory.
625 if (!I->mayLoad() && !I->mayStore())
628 // Check all of the machine operands.
629 for (unsigned i = 0 ; i < I->getNumOperands(); ++i) {
630 const MachineOperand &MO = I->getOperand(i);
635 int Slot = MO.getIndex();
640 if (Intervals[Slot]->empty())
643 // Check that the used slot is inside the calculated lifetime range.
644 // If it is not, warn about it and invalidate the range.
645 LiveInterval *Interval = Intervals[Slot];
646 SlotIndex Index = Indexes->getInstructionIndex(I);
647 if (Interval->find(Index) == Interval->end()) {
648 Intervals[Slot]->clear();
649 DEBUG(dbgs()<<"Invalidating range #"<<Slot<<"\n");
656 void StackColoring::expungeSlotMap(DenseMap<int, int> &SlotRemap,
658 // Expunge slot remap map.
659 for (unsigned i=0; i < NumSlots; ++i) {
660 // If we are remapping i
661 if (SlotRemap.count(i)) {
662 int Target = SlotRemap[i];
663 // As long as our target is mapped to something else, follow it.
664 while (SlotRemap.count(Target)) {
665 Target = SlotRemap[Target];
666 SlotRemap[i] = Target;
672 bool StackColoring::runOnMachineFunction(MachineFunction &Func) {
673 DEBUG(dbgs() << "********** Stack Coloring **********\n"
674 << "********** Function: "
675 << ((const Value*)Func.getFunction())->getName() << '\n');
677 MFI = MF->getFrameInfo();
678 Indexes = &getAnalysis<SlotIndexes>();
679 SP = &getAnalysis<StackProtector>();
680 BlockLiveness.clear();
682 BasicBlockNumbering.clear();
685 VNInfoAllocator.Reset();
687 unsigned NumSlots = MFI->getObjectIndexEnd();
689 // If there are no stack slots then there are no markers to remove.
693 SmallVector<int, 8> SortedSlots;
695 SortedSlots.reserve(NumSlots);
696 Intervals.reserve(NumSlots);
698 unsigned NumMarkers = collectMarkers(NumSlots);
700 unsigned TotalSize = 0;
701 DEBUG(dbgs()<<"Found "<<NumMarkers<<" markers and "<<NumSlots<<" slots\n");
702 DEBUG(dbgs()<<"Slot structure:\n");
704 for (int i=0; i < MFI->getObjectIndexEnd(); ++i) {
705 DEBUG(dbgs()<<"Slot #"<<i<<" - "<<MFI->getObjectSize(i)<<" bytes.\n");
706 TotalSize += MFI->getObjectSize(i);
709 DEBUG(dbgs()<<"Total Stack size: "<<TotalSize<<" bytes\n\n");
711 // Don't continue because there are not enough lifetime markers, or the
712 // stack is too small, or we are told not to optimize the slots.
713 if (NumMarkers < 2 || TotalSize < 16 || DisableColoring) {
714 DEBUG(dbgs()<<"Will not try to merge slots.\n");
715 return removeAllMarkers();
718 for (unsigned i=0; i < NumSlots; ++i) {
719 LiveInterval *LI = new LiveInterval(i, 0);
720 Intervals.push_back(LI);
721 LI->getNextValue(Indexes->getZeroIndex(), VNInfoAllocator);
722 SortedSlots.push_back(i);
725 // Calculate the liveness of each block.
726 calculateLocalLiveness();
728 // Propagate the liveness information.
729 calculateLiveIntervals(NumSlots);
731 // Search for allocas which are used outside of the declared lifetime
733 if (ProtectFromEscapedAllocas)
734 removeInvalidSlotRanges();
736 // Maps old slots to new slots.
737 DenseMap<int, int> SlotRemap;
738 unsigned RemovedSlots = 0;
739 unsigned ReducedSize = 0;
741 // Do not bother looking at empty intervals.
742 for (unsigned I = 0; I < NumSlots; ++I) {
743 if (Intervals[SortedSlots[I]]->empty())
747 // This is a simple greedy algorithm for merging allocas. First, sort the
748 // slots, placing the largest slots first. Next, perform an n^2 scan and look
749 // for disjoint slots. When you find disjoint slots, merge the samller one
750 // into the bigger one and update the live interval. Remove the small alloca
753 // Sort the slots according to their size. Place unused slots at the end.
754 // Use stable sort to guarantee deterministic code generation.
755 std::stable_sort(SortedSlots.begin(), SortedSlots.end(),
756 [this](int LHS, int RHS) {
757 // We use -1 to denote a uninteresting slot. Place these slots at the end.
758 if (LHS == -1) return false;
759 if (RHS == -1) return true;
760 // Sort according to size.
761 return MFI->getObjectSize(LHS) > MFI->getObjectSize(RHS);
767 for (unsigned I = 0; I < NumSlots; ++I) {
768 if (SortedSlots[I] == -1)
771 for (unsigned J=I+1; J < NumSlots; ++J) {
772 if (SortedSlots[J] == -1)
775 int FirstSlot = SortedSlots[I];
776 int SecondSlot = SortedSlots[J];
777 LiveInterval *First = Intervals[FirstSlot];
778 LiveInterval *Second = Intervals[SecondSlot];
779 assert (!First->empty() && !Second->empty() && "Found an empty range");
781 // Merge disjoint slots.
782 if (!First->overlaps(*Second)) {
784 First->MergeSegmentsInAsValue(*Second, First->getValNumInfo(0));
785 SlotRemap[SecondSlot] = FirstSlot;
787 DEBUG(dbgs()<<"Merging #"<<FirstSlot<<" and slots #"<<
788 SecondSlot<<" together.\n");
789 unsigned MaxAlignment = std::max(MFI->getObjectAlignment(FirstSlot),
790 MFI->getObjectAlignment(SecondSlot));
792 assert(MFI->getObjectSize(FirstSlot) >=
793 MFI->getObjectSize(SecondSlot) &&
794 "Merging a small object into a larger one");
797 ReducedSize += MFI->getObjectSize(SecondSlot);
798 MFI->setObjectAlignment(FirstSlot, MaxAlignment);
799 MFI->RemoveStackObject(SecondSlot);
805 // Record statistics.
806 StackSpaceSaved += ReducedSize;
807 StackSlotMerged += RemovedSlots;
808 DEBUG(dbgs()<<"Merge "<<RemovedSlots<<" slots. Saved "<<
809 ReducedSize<<" bytes\n");
811 // Scan the entire function and update all machine operands that use frame
812 // indices to use the remapped frame index.
813 expungeSlotMap(SlotRemap, NumSlots);
814 remapInstructions(SlotRemap);
816 // Release the intervals.
817 for (unsigned I = 0; I < NumSlots; ++I) {
821 return removeAllMarkers();