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/Support/CommandLine.h"
52 #include "llvm/Support/Debug.h"
53 #include "llvm/Support/raw_ostream.h"
54 #include "llvm/Target/TargetInstrInfo.h"
55 #include "llvm/Target/TargetRegisterInfo.h"
56
57 using namespace llvm;
58
59 #define DEBUG_TYPE "stackcoloring"
60
61 static cl::opt<bool>
62 DisableColoring("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.
70 static cl::opt<bool>
71 ProtectFromEscapedAllocas("protect-from-escaped-allocas",
72 cl::init(false), cl::Hidden,
73 cl::desc("Do not optimize lifetime zones that "
74 "are broken"));
75
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");
80
81 //===----------------------------------------------------------------------===//
82 // StackColoring Pass
83 //===----------------------------------------------------------------------===//
84
85 namespace {
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;
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
128 public:
129 static char ID;
StackColoring()130 StackColoring() : MachineFunctionPass(ID) {
131 initializeStackColoringPass(*PassRegistry::getPassRegistry());
132 }
133 void getAnalysisUsage(AnalysisUsage &AU) const override;
134 bool runOnMachineFunction(MachineFunction &MF) override;
135
136 private:
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
176 char StackColoring::ID = 0;
177 char &llvm::StackColoringID = StackColoring::ID;
178
179 INITIALIZE_PASS_BEGIN(StackColoring,
180 "stack-coloring", "Merge disjoint stack slots", false, false)
INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)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)
186
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);
193 }
194
dump() const195 void StackColoring::dump() const {
196 for (MachineBasicBlock *MBB : depth_first(MF)) {
197 DEBUG(dbgs() << "Inspecting block #" << BasicBlocks.lookup(MBB) << " ["
198 << MBB->getName() << "]\n");
199
200 LivenessMap::const_iterator BI = BlockLiveness.find(MBB);
201 assert(BI != BlockLiveness.end() && "Block not found");
202 const BlockLifetimeInfo &BlockInfo = BI->second;
203
204 DEBUG(dbgs()<<"BEGIN : {");
205 for (unsigned i=0; i < BlockInfo.Begin.size(); ++i)
206 DEBUG(dbgs()<<BlockInfo.Begin.test(i)<<" ");
207 DEBUG(dbgs()<<"}\n");
208
209 DEBUG(dbgs()<<"END : {");
210 for (unsigned i=0; i < BlockInfo.End.size(); ++i)
211 DEBUG(dbgs()<<BlockInfo.End.test(i)<<" ");
212
213 DEBUG(dbgs()<<"}\n");
214
215 DEBUG(dbgs()<<"LIVE_IN: {");
216 for (unsigned i=0; i < BlockInfo.LiveIn.size(); ++i)
217 DEBUG(dbgs()<<BlockInfo.LiveIn.test(i)<<" ");
218
219 DEBUG(dbgs()<<"}\n");
220 DEBUG(dbgs()<<"LIVEOUT: {");
221 for (unsigned i=0; i < BlockInfo.LiveOut.size(); ++i)
222 DEBUG(dbgs()<<BlockInfo.LiveOut.test(i)<<" ");
223 DEBUG(dbgs()<<"}\n");
224 }
225 }
226
collectMarkers(unsigned NumSlot)227 unsigned StackColoring::collectMarkers(unsigned NumSlot) {
228 unsigned MarkersFound = 0;
229 // Scan the function to find all lifetime markers.
230 // NOTE: We use a reverse-post-order iteration to ensure that we obtain a
231 // deterministic numbering, and because we'll need a post-order iteration
232 // later for solving the liveness dataflow problem.
233 for (MachineBasicBlock *MBB : depth_first(MF)) {
234
235 // Assign a serial number to this basic block.
236 BasicBlocks[MBB] = BasicBlockNumbering.size();
237 BasicBlockNumbering.push_back(MBB);
238
239 // Keep a reference to avoid repeated lookups.
240 BlockLifetimeInfo &BlockInfo = BlockLiveness[MBB];
241
242 BlockInfo.Begin.resize(NumSlot);
243 BlockInfo.End.resize(NumSlot);
244
245 for (MachineInstr &MI : *MBB) {
246 if (MI.getOpcode() != TargetOpcode::LIFETIME_START &&
247 MI.getOpcode() != TargetOpcode::LIFETIME_END)
248 continue;
249
250 Markers.push_back(&MI);
251
252 bool IsStart = MI.getOpcode() == TargetOpcode::LIFETIME_START;
253 const MachineOperand &MO = MI.getOperand(0);
254 unsigned Slot = MO.getIndex();
255
256 MarkersFound++;
257
258 const AllocaInst *Allocation = MFI->getObjectAllocation(Slot);
259 if (Allocation) {
260 DEBUG(dbgs()<<"Found a lifetime marker for slot #"<<Slot<<
261 " with allocation: "<< Allocation->getName()<<"\n");
262 }
263
264 if (IsStart) {
265 BlockInfo.Begin.set(Slot);
266 } else {
267 if (BlockInfo.Begin.test(Slot)) {
268 // Allocas that start and end within a single block are handled
269 // specially when computing the LiveIntervals to avoid pessimizing
270 // the liveness propagation.
271 BlockInfo.Begin.reset(Slot);
272 } else {
273 BlockInfo.End.set(Slot);
274 }
275 }
276 }
277 }
278
279 // Update statistics.
280 NumMarkerSeen += MarkersFound;
281 return MarkersFound;
282 }
283
calculateLocalLiveness()284 void StackColoring::calculateLocalLiveness() {
285 // Perform a standard reverse dataflow computation to solve for
286 // global liveness. The BEGIN set here is equivalent to KILL in the standard
287 // formulation, and END is equivalent to GEN. The result of this computation
288 // is a map from blocks to bitvectors where the bitvectors represent which
289 // allocas are live in/out of that block.
290 SmallPtrSet<const MachineBasicBlock*, 8> BBSet(BasicBlockNumbering.begin(),
291 BasicBlockNumbering.end());
292 unsigned NumSSMIters = 0;
293 bool changed = true;
294 while (changed) {
295 changed = false;
296 ++NumSSMIters;
297
298 SmallPtrSet<const MachineBasicBlock*, 8> NextBBSet;
299
300 for (const MachineBasicBlock *BB : BasicBlockNumbering) {
301 if (!BBSet.count(BB)) continue;
302
303 // Use an iterator to avoid repeated lookups.
304 LivenessMap::iterator BI = BlockLiveness.find(BB);
305 assert(BI != BlockLiveness.end() && "Block not found");
306 BlockLifetimeInfo &BlockInfo = BI->second;
307
308 BitVector LocalLiveIn;
309 BitVector LocalLiveOut;
310
311 // Forward propagation from begins to ends.
312 for (MachineBasicBlock::const_pred_iterator PI = BB->pred_begin(),
313 PE = BB->pred_end(); PI != PE; ++PI) {
314 LivenessMap::const_iterator I = BlockLiveness.find(*PI);
315 assert(I != BlockLiveness.end() && "Predecessor not found");
316 LocalLiveIn |= I->second.LiveOut;
317 }
318 LocalLiveIn |= BlockInfo.End;
319 LocalLiveIn.reset(BlockInfo.Begin);
320
321 // Reverse propagation from ends to begins.
322 for (MachineBasicBlock::const_succ_iterator SI = BB->succ_begin(),
323 SE = BB->succ_end(); SI != SE; ++SI) {
324 LivenessMap::const_iterator I = BlockLiveness.find(*SI);
325 assert(I != BlockLiveness.end() && "Successor not found");
326 LocalLiveOut |= I->second.LiveIn;
327 }
328 LocalLiveOut |= BlockInfo.Begin;
329 LocalLiveOut.reset(BlockInfo.End);
330
331 LocalLiveIn |= LocalLiveOut;
332 LocalLiveOut |= LocalLiveIn;
333
334 // After adopting the live bits, we need to turn-off the bits which
335 // are de-activated in this block.
336 LocalLiveOut.reset(BlockInfo.End);
337 LocalLiveIn.reset(BlockInfo.Begin);
338
339 // If we have both BEGIN and END markers in the same basic block then
340 // we know that the BEGIN marker comes after the END, because we already
341 // handle the case where the BEGIN comes before the END when collecting
342 // the markers (and building the BEGIN/END vectore).
343 // Want to enable the LIVE_IN and LIVE_OUT of slots that have both
344 // BEGIN and END because it means that the value lives before and after
345 // this basic block.
346 BitVector LocalEndBegin = BlockInfo.End;
347 LocalEndBegin &= BlockInfo.Begin;
348 LocalLiveIn |= LocalEndBegin;
349 LocalLiveOut |= LocalEndBegin;
350
351 if (LocalLiveIn.test(BlockInfo.LiveIn)) {
352 changed = true;
353 BlockInfo.LiveIn |= LocalLiveIn;
354
355 NextBBSet.insert(BB->pred_begin(), BB->pred_end());
356 }
357
358 if (LocalLiveOut.test(BlockInfo.LiveOut)) {
359 changed = true;
360 BlockInfo.LiveOut |= LocalLiveOut;
361
362 NextBBSet.insert(BB->succ_begin(), BB->succ_end());
363 }
364 }
365
366 BBSet = std::move(NextBBSet);
367 }// while changed.
368 }
369
calculateLiveIntervals(unsigned NumSlots)370 void StackColoring::calculateLiveIntervals(unsigned NumSlots) {
371 SmallVector<SlotIndex, 16> Starts;
372 SmallVector<SlotIndex, 16> Finishes;
373
374 // For each block, find which slots are active within this block
375 // and update the live intervals.
376 for (const MachineBasicBlock &MBB : *MF) {
377 Starts.clear();
378 Starts.resize(NumSlots);
379 Finishes.clear();
380 Finishes.resize(NumSlots);
381
382 // Create the interval for the basic blocks with lifetime markers in them.
383 for (const MachineInstr *MI : Markers) {
384 if (MI->getParent() != &MBB)
385 continue;
386
387 assert((MI->getOpcode() == TargetOpcode::LIFETIME_START ||
388 MI->getOpcode() == TargetOpcode::LIFETIME_END) &&
389 "Invalid Lifetime marker");
390
391 bool IsStart = MI->getOpcode() == TargetOpcode::LIFETIME_START;
392 const MachineOperand &Mo = MI->getOperand(0);
393 int Slot = Mo.getIndex();
394 assert(Slot >= 0 && "Invalid slot");
395
396 SlotIndex ThisIndex = Indexes->getInstructionIndex(MI);
397
398 if (IsStart) {
399 if (!Starts[Slot].isValid() || Starts[Slot] > ThisIndex)
400 Starts[Slot] = ThisIndex;
401 } else {
402 if (!Finishes[Slot].isValid() || Finishes[Slot] < ThisIndex)
403 Finishes[Slot] = ThisIndex;
404 }
405 }
406
407 // Create the interval of the blocks that we previously found to be 'alive'.
408 BlockLifetimeInfo &MBBLiveness = BlockLiveness[&MBB];
409 for (int pos = MBBLiveness.LiveIn.find_first(); pos != -1;
410 pos = MBBLiveness.LiveIn.find_next(pos)) {
411 Starts[pos] = Indexes->getMBBStartIdx(&MBB);
412 }
413 for (int pos = MBBLiveness.LiveOut.find_first(); pos != -1;
414 pos = MBBLiveness.LiveOut.find_next(pos)) {
415 Finishes[pos] = Indexes->getMBBEndIdx(&MBB);
416 }
417
418 for (unsigned i = 0; i < NumSlots; ++i) {
419 assert(Starts[i].isValid() == Finishes[i].isValid() && "Unmatched range");
420 if (!Starts[i].isValid())
421 continue;
422
423 assert(Starts[i] && Finishes[i] && "Invalid interval");
424 VNInfo *ValNum = Intervals[i]->getValNumInfo(0);
425 SlotIndex S = Starts[i];
426 SlotIndex F = Finishes[i];
427 if (S < F) {
428 // We have a single consecutive region.
429 Intervals[i]->addSegment(LiveInterval::Segment(S, F, ValNum));
430 } else {
431 // We have two non-consecutive regions. This happens when
432 // LIFETIME_START appears after the LIFETIME_END marker.
433 SlotIndex NewStart = Indexes->getMBBStartIdx(&MBB);
434 SlotIndex NewFin = Indexes->getMBBEndIdx(&MBB);
435 Intervals[i]->addSegment(LiveInterval::Segment(NewStart, F, ValNum));
436 Intervals[i]->addSegment(LiveInterval::Segment(S, NewFin, ValNum));
437 }
438 }
439 }
440 }
441
removeAllMarkers()442 bool StackColoring::removeAllMarkers() {
443 unsigned Count = 0;
444 for (MachineInstr *MI : Markers) {
445 MI->eraseFromParent();
446 Count++;
447 }
448 Markers.clear();
449
450 DEBUG(dbgs()<<"Removed "<<Count<<" markers.\n");
451 return Count;
452 }
453
remapInstructions(DenseMap<int,int> & SlotRemap)454 void StackColoring::remapInstructions(DenseMap<int, int> &SlotRemap) {
455 unsigned FixedInstr = 0;
456 unsigned FixedMemOp = 0;
457 unsigned FixedDbg = 0;
458 MachineModuleInfo *MMI = &MF->getMMI();
459
460 // Remap debug information that refers to stack slots.
461 for (auto &VI : MMI->getVariableDbgInfo()) {
462 if (!VI.Var)
463 continue;
464 if (SlotRemap.count(VI.Slot)) {
465 DEBUG(dbgs() << "Remapping debug info for ["
466 << cast<DILocalVariable>(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
removeInvalidSlotRanges()578 void 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
expungeSlotMap(DenseMap<int,int> & SlotRemap,unsigned NumSlots)620 void 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
runOnMachineFunction(MachineFunction & Func)636 bool 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