InlineSpiller.cpp revision d96c9853192af2baea0d164228582f1dac8efbf2
1//===-------- InlineSpiller.cpp - Insert spills and restores inline -------===//
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// The inline spiller modifies the machine function directly instead of
11// inserting spills and restores in VirtRegMap.
12//
13//===----------------------------------------------------------------------===//
14
15#define DEBUG_TYPE "spiller"
16#include "Spiller.h"
17#include "LiveRangeEdit.h"
18#include "SplitKit.h"
19#include "VirtRegMap.h"
20#include "llvm/CodeGen/LiveIntervalAnalysis.h"
21#include "llvm/CodeGen/MachineFrameInfo.h"
22#include "llvm/CodeGen/MachineFunction.h"
23#include "llvm/CodeGen/MachineLoopInfo.h"
24#include "llvm/CodeGen/MachineRegisterInfo.h"
25#include "llvm/Target/TargetMachine.h"
26#include "llvm/Target/TargetInstrInfo.h"
27#include "llvm/Support/Debug.h"
28#include "llvm/Support/raw_ostream.h"
29
30using namespace llvm;
31
32namespace {
33class InlineSpiller : public Spiller {
34  MachineFunctionPass &pass_;
35  MachineFunction &mf_;
36  LiveIntervals &lis_;
37  MachineLoopInfo &loops_;
38  VirtRegMap &vrm_;
39  MachineFrameInfo &mfi_;
40  MachineRegisterInfo &mri_;
41  const TargetInstrInfo &tii_;
42  const TargetRegisterInfo &tri_;
43  const BitVector reserved_;
44
45  SplitAnalysis splitAnalysis_;
46
47  // Variables that are valid during spill(), but used by multiple methods.
48  LiveRangeEdit *edit_;
49  const TargetRegisterClass *rc_;
50  int stackSlot_;
51
52  // Values that failed to remat at some point.
53  SmallPtrSet<VNInfo*, 8> usedValues_;
54
55  ~InlineSpiller() {}
56
57public:
58  InlineSpiller(MachineFunctionPass &pass,
59                MachineFunction &mf,
60                VirtRegMap &vrm)
61    : pass_(pass),
62      mf_(mf),
63      lis_(pass.getAnalysis<LiveIntervals>()),
64      loops_(pass.getAnalysis<MachineLoopInfo>()),
65      vrm_(vrm),
66      mfi_(*mf.getFrameInfo()),
67      mri_(mf.getRegInfo()),
68      tii_(*mf.getTarget().getInstrInfo()),
69      tri_(*mf.getTarget().getRegisterInfo()),
70      reserved_(tri_.getReservedRegs(mf_)),
71      splitAnalysis_(mf, lis_, loops_) {}
72
73  void spill(LiveInterval *li,
74             SmallVectorImpl<LiveInterval*> &newIntervals,
75             SmallVectorImpl<LiveInterval*> &spillIs);
76
77  void spill(LiveRangeEdit &);
78
79private:
80  bool split();
81
82  bool reMaterializeFor(MachineBasicBlock::iterator MI);
83  void reMaterializeAll();
84
85  bool coalesceStackAccess(MachineInstr *MI);
86  bool foldMemoryOperand(MachineBasicBlock::iterator MI,
87                         const SmallVectorImpl<unsigned> &Ops);
88  void insertReload(LiveInterval &NewLI, MachineBasicBlock::iterator MI);
89  void insertSpill(LiveInterval &NewLI, MachineBasicBlock::iterator MI);
90};
91}
92
93namespace llvm {
94Spiller *createInlineSpiller(MachineFunctionPass &pass,
95                             MachineFunction &mf,
96                             VirtRegMap &vrm) {
97  return new InlineSpiller(pass, mf, vrm);
98}
99}
100
101/// split - try splitting the current interval into pieces that may allocate
102/// separately. Return true if successful.
103bool InlineSpiller::split() {
104  splitAnalysis_.analyze(&edit_->getParent());
105
106  // Try splitting around loops.
107  if (const MachineLoop *loop = splitAnalysis_.getBestSplitLoop()) {
108    SplitEditor(splitAnalysis_, lis_, vrm_, *edit_)
109      .splitAroundLoop(loop);
110    return true;
111  }
112
113  // Try splitting into single block intervals.
114  SplitAnalysis::BlockPtrSet blocks;
115  if (splitAnalysis_.getMultiUseBlocks(blocks)) {
116    SplitEditor(splitAnalysis_, lis_, vrm_, *edit_)
117      .splitSingleBlocks(blocks);
118    return true;
119  }
120
121  // Try splitting inside a basic block.
122  if (const MachineBasicBlock *MBB = splitAnalysis_.getBlockForInsideSplit()) {
123    SplitEditor(splitAnalysis_, lis_, vrm_, *edit_)
124      .splitInsideBlock(MBB);
125    return true;
126  }
127
128  return false;
129}
130
131/// reMaterializeFor - Attempt to rematerialize edit_->getReg() before MI instead of
132/// reloading it.
133bool InlineSpiller::reMaterializeFor(MachineBasicBlock::iterator MI) {
134  SlotIndex UseIdx = lis_.getInstructionIndex(MI).getUseIndex();
135  VNInfo *OrigVNI = edit_->getParent().getVNInfoAt(UseIdx);
136
137  if (!OrigVNI) {
138    DEBUG(dbgs() << "\tadding <undef> flags: ");
139    for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
140      MachineOperand &MO = MI->getOperand(i);
141      if (MO.isReg() && MO.isUse() && MO.getReg() == edit_->getReg())
142        MO.setIsUndef();
143    }
144    DEBUG(dbgs() << UseIdx << '\t' << *MI);
145    return true;
146  }
147
148  LiveRangeEdit::Remat RM = edit_->canRematerializeAt(OrigVNI, UseIdx, false,
149                                                      lis_);
150  if (!RM) {
151    usedValues_.insert(OrigVNI);
152    DEBUG(dbgs() << "\tcannot remat for " << UseIdx << '\t' << *MI);
153    return false;
154  }
155
156  // If the instruction also writes edit_->getReg(), it had better not require
157  // the same register for uses and defs.
158  bool Reads, Writes;
159  SmallVector<unsigned, 8> Ops;
160  tie(Reads, Writes) = MI->readsWritesVirtualRegister(edit_->getReg(), &Ops);
161  if (Writes) {
162    for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
163      MachineOperand &MO = MI->getOperand(Ops[i]);
164      if (MO.isUse() ? MI->isRegTiedToDefOperand(Ops[i]) : MO.getSubReg()) {
165        usedValues_.insert(OrigVNI);
166        DEBUG(dbgs() << "\tcannot remat tied reg: " << UseIdx << '\t' << *MI);
167        return false;
168      }
169    }
170  }
171
172  // Alocate a new register for the remat.
173  LiveInterval &NewLI = edit_->create(mri_, lis_, vrm_);
174  NewLI.markNotSpillable();
175
176  // Finally we can rematerialize OrigMI before MI.
177  SlotIndex DefIdx = edit_->rematerializeAt(*MI->getParent(), MI, NewLI.reg, RM,
178                                            lis_, tii_, tri_);
179  DEBUG(dbgs() << "\tremat:  " << DefIdx << '\n');
180
181  // Replace operands
182  for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
183    MachineOperand &MO = MI->getOperand(Ops[i]);
184    if (MO.isReg() && MO.isUse() && MO.getReg() == edit_->getReg()) {
185      MO.setReg(NewLI.reg);
186      MO.setIsKill();
187    }
188  }
189  DEBUG(dbgs() << "\t        " << UseIdx << '\t' << *MI);
190
191  VNInfo *DefVNI = NewLI.getNextValue(DefIdx, 0, lis_.getVNInfoAllocator());
192  NewLI.addRange(LiveRange(DefIdx, UseIdx.getDefIndex(), DefVNI));
193  DEBUG(dbgs() << "\tinterval: " << NewLI << '\n');
194  return true;
195}
196
197/// reMaterializeAll - Try to rematerialize as many uses as possible,
198/// and trim the live ranges after.
199void InlineSpiller::reMaterializeAll() {
200  // Do a quick scan of the interval values to find if any are remattable.
201  if (!edit_->anyRematerializable(lis_, tii_, 0))
202    return;
203
204  usedValues_.clear();
205
206  // Try to remat before all uses of edit_->getReg().
207  bool anyRemat = false;
208  for (MachineRegisterInfo::use_nodbg_iterator
209       RI = mri_.use_nodbg_begin(edit_->getReg());
210       MachineInstr *MI = RI.skipInstruction();)
211     anyRemat |= reMaterializeFor(MI);
212
213  if (!anyRemat)
214    return;
215
216  // Remove any values that were completely rematted.
217  bool anyRemoved = false;
218  for (LiveInterval::vni_iterator I = edit_->getParent().vni_begin(),
219       E = edit_->getParent().vni_end(); I != E; ++I) {
220    VNInfo *VNI = *I;
221    if (VNI->hasPHIKill() || !edit_->didRematerialize(VNI) ||
222        usedValues_.count(VNI))
223      continue;
224    MachineInstr *DefMI = lis_.getInstructionFromIndex(VNI->def);
225    DEBUG(dbgs() << "\tremoving dead def: " << VNI->def << '\t' << *DefMI);
226    lis_.RemoveMachineInstrFromMaps(DefMI);
227    vrm_.RemoveMachineInstrFromMaps(DefMI);
228    DefMI->eraseFromParent();
229    VNI->def = SlotIndex();
230    anyRemoved = true;
231  }
232
233  if (!anyRemoved)
234    return;
235
236  // Removing values may cause debug uses where parent is not live.
237  for (MachineRegisterInfo::use_iterator RI = mri_.use_begin(edit_->getReg());
238       MachineInstr *MI = RI.skipInstruction();) {
239    if (!MI->isDebugValue())
240      continue;
241    // Try to preserve the debug value if parent is live immediately after it.
242    MachineBasicBlock::iterator NextMI = MI;
243    ++NextMI;
244    if (NextMI != MI->getParent()->end() && !lis_.isNotInMIMap(NextMI)) {
245      SlotIndex Idx = lis_.getInstructionIndex(NextMI);
246      VNInfo *VNI = edit_->getParent().getVNInfoAt(Idx);
247      if (VNI && (VNI->hasPHIKill() || usedValues_.count(VNI)))
248        continue;
249    }
250    DEBUG(dbgs() << "Removing debug info due to remat:" << "\t" << *MI);
251    MI->eraseFromParent();
252  }
253}
254
255/// If MI is a load or store of stackSlot_, it can be removed.
256bool InlineSpiller::coalesceStackAccess(MachineInstr *MI) {
257  int FI = 0;
258  unsigned reg;
259  if (!(reg = tii_.isLoadFromStackSlot(MI, FI)) &&
260      !(reg = tii_.isStoreToStackSlot(MI, FI)))
261    return false;
262
263  // We have a stack access. Is it the right register and slot?
264  if (reg != edit_->getReg() || FI != stackSlot_)
265    return false;
266
267  DEBUG(dbgs() << "Coalescing stack access: " << *MI);
268  lis_.RemoveMachineInstrFromMaps(MI);
269  MI->eraseFromParent();
270  return true;
271}
272
273/// foldMemoryOperand - Try folding stack slot references in Ops into MI.
274/// Return true on success, and MI will be erased.
275bool InlineSpiller::foldMemoryOperand(MachineBasicBlock::iterator MI,
276                                      const SmallVectorImpl<unsigned> &Ops) {
277  // TargetInstrInfo::foldMemoryOperand only expects explicit, non-tied
278  // operands.
279  SmallVector<unsigned, 8> FoldOps;
280  for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
281    unsigned Idx = Ops[i];
282    MachineOperand &MO = MI->getOperand(Idx);
283    if (MO.isImplicit())
284      continue;
285    // FIXME: Teach targets to deal with subregs.
286    if (MO.getSubReg())
287      return false;
288    // Tied use operands should not be passed to foldMemoryOperand.
289    if (!MI->isRegTiedToDefOperand(Idx))
290      FoldOps.push_back(Idx);
291  }
292
293  MachineInstr *FoldMI = tii_.foldMemoryOperand(MI, FoldOps, stackSlot_);
294  if (!FoldMI)
295    return false;
296  lis_.ReplaceMachineInstrInMaps(MI, FoldMI);
297  vrm_.addSpillSlotUse(stackSlot_, FoldMI);
298  MI->eraseFromParent();
299  DEBUG(dbgs() << "\tfolded: " << *FoldMI);
300  return true;
301}
302
303/// insertReload - Insert a reload of NewLI.reg before MI.
304void InlineSpiller::insertReload(LiveInterval &NewLI,
305                                 MachineBasicBlock::iterator MI) {
306  MachineBasicBlock &MBB = *MI->getParent();
307  SlotIndex Idx = lis_.getInstructionIndex(MI).getDefIndex();
308  tii_.loadRegFromStackSlot(MBB, MI, NewLI.reg, stackSlot_, rc_, &tri_);
309  --MI; // Point to load instruction.
310  SlotIndex LoadIdx = lis_.InsertMachineInstrInMaps(MI).getDefIndex();
311  vrm_.addSpillSlotUse(stackSlot_, MI);
312  DEBUG(dbgs() << "\treload:  " << LoadIdx << '\t' << *MI);
313  VNInfo *LoadVNI = NewLI.getNextValue(LoadIdx, 0,
314                                       lis_.getVNInfoAllocator());
315  NewLI.addRange(LiveRange(LoadIdx, Idx, LoadVNI));
316}
317
318/// insertSpill - Insert a spill of NewLI.reg after MI.
319void InlineSpiller::insertSpill(LiveInterval &NewLI,
320                                MachineBasicBlock::iterator MI) {
321  MachineBasicBlock &MBB = *MI->getParent();
322  SlotIndex Idx = lis_.getInstructionIndex(MI).getDefIndex();
323  tii_.storeRegToStackSlot(MBB, ++MI, NewLI.reg, true, stackSlot_, rc_, &tri_);
324  --MI; // Point to store instruction.
325  SlotIndex StoreIdx = lis_.InsertMachineInstrInMaps(MI).getDefIndex();
326  vrm_.addSpillSlotUse(stackSlot_, MI);
327  DEBUG(dbgs() << "\tspilled: " << StoreIdx << '\t' << *MI);
328  VNInfo *StoreVNI = NewLI.getNextValue(Idx, 0, lis_.getVNInfoAllocator());
329  NewLI.addRange(LiveRange(Idx, StoreIdx, StoreVNI));
330}
331
332void InlineSpiller::spill(LiveInterval *li,
333                          SmallVectorImpl<LiveInterval*> &newIntervals,
334                          SmallVectorImpl<LiveInterval*> &spillIs) {
335  LiveRangeEdit edit(*li, newIntervals, spillIs);
336  spill(edit);
337}
338
339void InlineSpiller::spill(LiveRangeEdit &edit) {
340  edit_ = &edit;
341  DEBUG(dbgs() << "Inline spilling " << edit.getParent() << "\n");
342  assert(edit.getParent().isSpillable() &&
343         "Attempting to spill already spilled value.");
344  assert(!edit.getParent().isStackSlot() && "Trying to spill a stack slot.");
345
346  if (split())
347    return;
348
349  reMaterializeAll();
350
351  // Remat may handle everything.
352  if (edit_->getParent().empty())
353    return;
354
355  rc_ = mri_.getRegClass(edit.getReg());
356  stackSlot_ = edit.assignStackSlot(vrm_);
357
358  // Iterate over instructions using register.
359  for (MachineRegisterInfo::reg_iterator RI = mri_.reg_begin(edit.getReg());
360       MachineInstr *MI = RI.skipInstruction();) {
361
362    // Debug values are not allowed to affect codegen.
363    if (MI->isDebugValue()) {
364      // Modify DBG_VALUE now that the value is in a spill slot.
365      uint64_t Offset = MI->getOperand(1).getImm();
366      const MDNode *MDPtr = MI->getOperand(2).getMetadata();
367      DebugLoc DL = MI->getDebugLoc();
368      if (MachineInstr *NewDV = tii_.emitFrameIndexDebugValue(mf_, stackSlot_,
369                                                           Offset, MDPtr, DL)) {
370        DEBUG(dbgs() << "Modifying debug info due to spill:" << "\t" << *MI);
371        MachineBasicBlock *MBB = MI->getParent();
372        MBB->insert(MBB->erase(MI), NewDV);
373      } else {
374        DEBUG(dbgs() << "Removing debug info due to spill:" << "\t" << *MI);
375        MI->eraseFromParent();
376      }
377      continue;
378    }
379
380    // Stack slot accesses may coalesce away.
381    if (coalesceStackAccess(MI))
382      continue;
383
384    // Analyze instruction.
385    bool Reads, Writes;
386    SmallVector<unsigned, 8> Ops;
387    tie(Reads, Writes) = MI->readsWritesVirtualRegister(edit.getReg(), &Ops);
388
389    // Attempt to fold memory ops.
390    if (foldMemoryOperand(MI, Ops))
391      continue;
392
393    // Allocate interval around instruction.
394    // FIXME: Infer regclass from instruction alone.
395    LiveInterval &NewLI = edit.create(mri_, lis_, vrm_);
396    NewLI.markNotSpillable();
397
398    if (Reads)
399      insertReload(NewLI, MI);
400
401    // Rewrite instruction operands.
402    bool hasLiveDef = false;
403    for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
404      MachineOperand &MO = MI->getOperand(Ops[i]);
405      MO.setReg(NewLI.reg);
406      if (MO.isUse()) {
407        if (!MI->isRegTiedToDefOperand(Ops[i]))
408          MO.setIsKill();
409      } else {
410        if (!MO.isDead())
411          hasLiveDef = true;
412      }
413    }
414
415    // FIXME: Use a second vreg if instruction has no tied ops.
416    if (Writes && hasLiveDef)
417      insertSpill(NewLI, MI);
418
419    DEBUG(dbgs() << "\tinterval: " << NewLI << '\n');
420  }
421}
422