InlineSpiller.cpp revision 360660
1//===- InlineSpiller.cpp - Insert spills and restores inline --------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// The inline spiller modifies the machine function directly instead of
10// inserting spills and restores in VirtRegMap.
11//
12//===----------------------------------------------------------------------===//
13
14#include "LiveRangeCalc.h"
15#include "Spiller.h"
16#include "SplitKit.h"
17#include "llvm/ADT/ArrayRef.h"
18#include "llvm/ADT/DenseMap.h"
19#include "llvm/ADT/MapVector.h"
20#include "llvm/ADT/None.h"
21#include "llvm/ADT/STLExtras.h"
22#include "llvm/ADT/SetVector.h"
23#include "llvm/ADT/SmallPtrSet.h"
24#include "llvm/ADT/SmallVector.h"
25#include "llvm/ADT/Statistic.h"
26#include "llvm/Analysis/AliasAnalysis.h"
27#include "llvm/CodeGen/LiveInterval.h"
28#include "llvm/CodeGen/LiveIntervals.h"
29#include "llvm/CodeGen/LiveRangeEdit.h"
30#include "llvm/CodeGen/LiveStacks.h"
31#include "llvm/CodeGen/MachineBasicBlock.h"
32#include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
33#include "llvm/CodeGen/MachineDominators.h"
34#include "llvm/CodeGen/MachineFunction.h"
35#include "llvm/CodeGen/MachineFunctionPass.h"
36#include "llvm/CodeGen/MachineInstr.h"
37#include "llvm/CodeGen/MachineInstrBuilder.h"
38#include "llvm/CodeGen/MachineInstrBundle.h"
39#include "llvm/CodeGen/MachineLoopInfo.h"
40#include "llvm/CodeGen/MachineOperand.h"
41#include "llvm/CodeGen/MachineRegisterInfo.h"
42#include "llvm/CodeGen/SlotIndexes.h"
43#include "llvm/CodeGen/TargetInstrInfo.h"
44#include "llvm/CodeGen/TargetOpcodes.h"
45#include "llvm/CodeGen/TargetRegisterInfo.h"
46#include "llvm/CodeGen/TargetSubtargetInfo.h"
47#include "llvm/CodeGen/VirtRegMap.h"
48#include "llvm/Config/llvm-config.h"
49#include "llvm/Support/BlockFrequency.h"
50#include "llvm/Support/BranchProbability.h"
51#include "llvm/Support/CommandLine.h"
52#include "llvm/Support/Compiler.h"
53#include "llvm/Support/Debug.h"
54#include "llvm/Support/ErrorHandling.h"
55#include "llvm/Support/raw_ostream.h"
56#include <cassert>
57#include <iterator>
58#include <tuple>
59#include <utility>
60#include <vector>
61
62using namespace llvm;
63
64#define DEBUG_TYPE "regalloc"
65
66STATISTIC(NumSpilledRanges,   "Number of spilled live ranges");
67STATISTIC(NumSnippets,        "Number of spilled snippets");
68STATISTIC(NumSpills,          "Number of spills inserted");
69STATISTIC(NumSpillsRemoved,   "Number of spills removed");
70STATISTIC(NumReloads,         "Number of reloads inserted");
71STATISTIC(NumReloadsRemoved,  "Number of reloads removed");
72STATISTIC(NumFolded,          "Number of folded stack accesses");
73STATISTIC(NumFoldedLoads,     "Number of folded loads");
74STATISTIC(NumRemats,          "Number of rematerialized defs for spilling");
75
76static cl::opt<bool> DisableHoisting("disable-spill-hoist", cl::Hidden,
77                                     cl::desc("Disable inline spill hoisting"));
78static cl::opt<bool>
79RestrictStatepointRemat("restrict-statepoint-remat",
80                       cl::init(false), cl::Hidden,
81                       cl::desc("Restrict remat for statepoint operands"));
82
83namespace {
84
85class HoistSpillHelper : private LiveRangeEdit::Delegate {
86  MachineFunction &MF;
87  LiveIntervals &LIS;
88  LiveStacks &LSS;
89  AliasAnalysis *AA;
90  MachineDominatorTree &MDT;
91  MachineLoopInfo &Loops;
92  VirtRegMap &VRM;
93  MachineRegisterInfo &MRI;
94  const TargetInstrInfo &TII;
95  const TargetRegisterInfo &TRI;
96  const MachineBlockFrequencyInfo &MBFI;
97
98  InsertPointAnalysis IPA;
99
100  // Map from StackSlot to the LiveInterval of the original register.
101  // Note the LiveInterval of the original register may have been deleted
102  // after it is spilled. We keep a copy here to track the range where
103  // spills can be moved.
104  DenseMap<int, std::unique_ptr<LiveInterval>> StackSlotToOrigLI;
105
106  // Map from pair of (StackSlot and Original VNI) to a set of spills which
107  // have the same stackslot and have equal values defined by Original VNI.
108  // These spills are mergeable and are hoist candiates.
109  using MergeableSpillsMap =
110      MapVector<std::pair<int, VNInfo *>, SmallPtrSet<MachineInstr *, 16>>;
111  MergeableSpillsMap MergeableSpills;
112
113  /// This is the map from original register to a set containing all its
114  /// siblings. To hoist a spill to another BB, we need to find out a live
115  /// sibling there and use it as the source of the new spill.
116  DenseMap<unsigned, SmallSetVector<unsigned, 16>> Virt2SiblingsMap;
117
118  bool isSpillCandBB(LiveInterval &OrigLI, VNInfo &OrigVNI,
119                     MachineBasicBlock &BB, unsigned &LiveReg);
120
121  void rmRedundantSpills(
122      SmallPtrSet<MachineInstr *, 16> &Spills,
123      SmallVectorImpl<MachineInstr *> &SpillsToRm,
124      DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill);
125
126  void getVisitOrders(
127      MachineBasicBlock *Root, SmallPtrSet<MachineInstr *, 16> &Spills,
128      SmallVectorImpl<MachineDomTreeNode *> &Orders,
129      SmallVectorImpl<MachineInstr *> &SpillsToRm,
130      DenseMap<MachineDomTreeNode *, unsigned> &SpillsToKeep,
131      DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill);
132
133  void runHoistSpills(LiveInterval &OrigLI, VNInfo &OrigVNI,
134                      SmallPtrSet<MachineInstr *, 16> &Spills,
135                      SmallVectorImpl<MachineInstr *> &SpillsToRm,
136                      DenseMap<MachineBasicBlock *, unsigned> &SpillsToIns);
137
138public:
139  HoistSpillHelper(MachineFunctionPass &pass, MachineFunction &mf,
140                   VirtRegMap &vrm)
141      : MF(mf), LIS(pass.getAnalysis<LiveIntervals>()),
142        LSS(pass.getAnalysis<LiveStacks>()),
143        AA(&pass.getAnalysis<AAResultsWrapperPass>().getAAResults()),
144        MDT(pass.getAnalysis<MachineDominatorTree>()),
145        Loops(pass.getAnalysis<MachineLoopInfo>()), VRM(vrm),
146        MRI(mf.getRegInfo()), TII(*mf.getSubtarget().getInstrInfo()),
147        TRI(*mf.getSubtarget().getRegisterInfo()),
148        MBFI(pass.getAnalysis<MachineBlockFrequencyInfo>()),
149        IPA(LIS, mf.getNumBlockIDs()) {}
150
151  void addToMergeableSpills(MachineInstr &Spill, int StackSlot,
152                            unsigned Original);
153  bool rmFromMergeableSpills(MachineInstr &Spill, int StackSlot);
154  void hoistAllSpills();
155  void LRE_DidCloneVirtReg(unsigned, unsigned) override;
156};
157
158class InlineSpiller : public Spiller {
159  MachineFunction &MF;
160  LiveIntervals &LIS;
161  LiveStacks &LSS;
162  AliasAnalysis *AA;
163  MachineDominatorTree &MDT;
164  MachineLoopInfo &Loops;
165  VirtRegMap &VRM;
166  MachineRegisterInfo &MRI;
167  const TargetInstrInfo &TII;
168  const TargetRegisterInfo &TRI;
169  const MachineBlockFrequencyInfo &MBFI;
170
171  // Variables that are valid during spill(), but used by multiple methods.
172  LiveRangeEdit *Edit;
173  LiveInterval *StackInt;
174  int StackSlot;
175  unsigned Original;
176
177  // All registers to spill to StackSlot, including the main register.
178  SmallVector<unsigned, 8> RegsToSpill;
179
180  // All COPY instructions to/from snippets.
181  // They are ignored since both operands refer to the same stack slot.
182  SmallPtrSet<MachineInstr*, 8> SnippetCopies;
183
184  // Values that failed to remat at some point.
185  SmallPtrSet<VNInfo*, 8> UsedValues;
186
187  // Dead defs generated during spilling.
188  SmallVector<MachineInstr*, 8> DeadDefs;
189
190  // Object records spills information and does the hoisting.
191  HoistSpillHelper HSpiller;
192
193  ~InlineSpiller() override = default;
194
195public:
196  InlineSpiller(MachineFunctionPass &pass, MachineFunction &mf, VirtRegMap &vrm)
197      : MF(mf), LIS(pass.getAnalysis<LiveIntervals>()),
198        LSS(pass.getAnalysis<LiveStacks>()),
199        AA(&pass.getAnalysis<AAResultsWrapperPass>().getAAResults()),
200        MDT(pass.getAnalysis<MachineDominatorTree>()),
201        Loops(pass.getAnalysis<MachineLoopInfo>()), VRM(vrm),
202        MRI(mf.getRegInfo()), TII(*mf.getSubtarget().getInstrInfo()),
203        TRI(*mf.getSubtarget().getRegisterInfo()),
204        MBFI(pass.getAnalysis<MachineBlockFrequencyInfo>()),
205        HSpiller(pass, mf, vrm) {}
206
207  void spill(LiveRangeEdit &) override;
208  void postOptimization() override;
209
210private:
211  bool isSnippet(const LiveInterval &SnipLI);
212  void collectRegsToSpill();
213
214  bool isRegToSpill(unsigned Reg) { return is_contained(RegsToSpill, Reg); }
215
216  bool isSibling(unsigned Reg);
217  bool hoistSpillInsideBB(LiveInterval &SpillLI, MachineInstr &CopyMI);
218  void eliminateRedundantSpills(LiveInterval &LI, VNInfo *VNI);
219
220  void markValueUsed(LiveInterval*, VNInfo*);
221  bool canGuaranteeAssignmentAfterRemat(unsigned VReg, MachineInstr &MI);
222  bool reMaterializeFor(LiveInterval &, MachineInstr &MI);
223  void reMaterializeAll();
224
225  bool coalesceStackAccess(MachineInstr *MI, unsigned Reg);
226  bool foldMemoryOperand(ArrayRef<std::pair<MachineInstr *, unsigned>>,
227                         MachineInstr *LoadMI = nullptr);
228  void insertReload(unsigned VReg, SlotIndex, MachineBasicBlock::iterator MI);
229  void insertSpill(unsigned VReg, bool isKill, MachineBasicBlock::iterator MI);
230
231  void spillAroundUses(unsigned Reg);
232  void spillAll();
233};
234
235} // end anonymous namespace
236
237Spiller::~Spiller() = default;
238
239void Spiller::anchor() {}
240
241Spiller *llvm::createInlineSpiller(MachineFunctionPass &pass,
242                                   MachineFunction &mf,
243                                   VirtRegMap &vrm) {
244  return new InlineSpiller(pass, mf, vrm);
245}
246
247//===----------------------------------------------------------------------===//
248//                                Snippets
249//===----------------------------------------------------------------------===//
250
251// When spilling a virtual register, we also spill any snippets it is connected
252// to. The snippets are small live ranges that only have a single real use,
253// leftovers from live range splitting. Spilling them enables memory operand
254// folding or tightens the live range around the single use.
255//
256// This minimizes register pressure and maximizes the store-to-load distance for
257// spill slots which can be important in tight loops.
258
259/// isFullCopyOf - If MI is a COPY to or from Reg, return the other register,
260/// otherwise return 0.
261static unsigned isFullCopyOf(const MachineInstr &MI, unsigned Reg) {
262  if (!MI.isFullCopy())
263    return 0;
264  if (MI.getOperand(0).getReg() == Reg)
265    return MI.getOperand(1).getReg();
266  if (MI.getOperand(1).getReg() == Reg)
267    return MI.getOperand(0).getReg();
268  return 0;
269}
270
271/// isSnippet - Identify if a live interval is a snippet that should be spilled.
272/// It is assumed that SnipLI is a virtual register with the same original as
273/// Edit->getReg().
274bool InlineSpiller::isSnippet(const LiveInterval &SnipLI) {
275  unsigned Reg = Edit->getReg();
276
277  // A snippet is a tiny live range with only a single instruction using it
278  // besides copies to/from Reg or spills/fills. We accept:
279  //
280  //   %snip = COPY %Reg / FILL fi#
281  //   %snip = USE %snip
282  //   %Reg = COPY %snip / SPILL %snip, fi#
283  //
284  if (SnipLI.getNumValNums() > 2 || !LIS.intervalIsInOneMBB(SnipLI))
285    return false;
286
287  MachineInstr *UseMI = nullptr;
288
289  // Check that all uses satisfy our criteria.
290  for (MachineRegisterInfo::reg_instr_nodbg_iterator
291       RI = MRI.reg_instr_nodbg_begin(SnipLI.reg),
292       E = MRI.reg_instr_nodbg_end(); RI != E; ) {
293    MachineInstr &MI = *RI++;
294
295    // Allow copies to/from Reg.
296    if (isFullCopyOf(MI, Reg))
297      continue;
298
299    // Allow stack slot loads.
300    int FI;
301    if (SnipLI.reg == TII.isLoadFromStackSlot(MI, FI) && FI == StackSlot)
302      continue;
303
304    // Allow stack slot stores.
305    if (SnipLI.reg == TII.isStoreToStackSlot(MI, FI) && FI == StackSlot)
306      continue;
307
308    // Allow a single additional instruction.
309    if (UseMI && &MI != UseMI)
310      return false;
311    UseMI = &MI;
312  }
313  return true;
314}
315
316/// collectRegsToSpill - Collect live range snippets that only have a single
317/// real use.
318void InlineSpiller::collectRegsToSpill() {
319  unsigned Reg = Edit->getReg();
320
321  // Main register always spills.
322  RegsToSpill.assign(1, Reg);
323  SnippetCopies.clear();
324
325  // Snippets all have the same original, so there can't be any for an original
326  // register.
327  if (Original == Reg)
328    return;
329
330  for (MachineRegisterInfo::reg_instr_iterator
331       RI = MRI.reg_instr_begin(Reg), E = MRI.reg_instr_end(); RI != E; ) {
332    MachineInstr &MI = *RI++;
333    unsigned SnipReg = isFullCopyOf(MI, Reg);
334    if (!isSibling(SnipReg))
335      continue;
336    LiveInterval &SnipLI = LIS.getInterval(SnipReg);
337    if (!isSnippet(SnipLI))
338      continue;
339    SnippetCopies.insert(&MI);
340    if (isRegToSpill(SnipReg))
341      continue;
342    RegsToSpill.push_back(SnipReg);
343    LLVM_DEBUG(dbgs() << "\talso spill snippet " << SnipLI << '\n');
344    ++NumSnippets;
345  }
346}
347
348bool InlineSpiller::isSibling(unsigned Reg) {
349  return TargetRegisterInfo::isVirtualRegister(Reg) &&
350           VRM.getOriginal(Reg) == Original;
351}
352
353/// It is beneficial to spill to earlier place in the same BB in case
354/// as follows:
355/// There is an alternative def earlier in the same MBB.
356/// Hoist the spill as far as possible in SpillMBB. This can ease
357/// register pressure:
358///
359///   x = def
360///   y = use x
361///   s = copy x
362///
363/// Hoisting the spill of s to immediately after the def removes the
364/// interference between x and y:
365///
366///   x = def
367///   spill x
368///   y = use killed x
369///
370/// This hoist only helps when the copy kills its source.
371///
372bool InlineSpiller::hoistSpillInsideBB(LiveInterval &SpillLI,
373                                       MachineInstr &CopyMI) {
374  SlotIndex Idx = LIS.getInstructionIndex(CopyMI);
375#ifndef NDEBUG
376  VNInfo *VNI = SpillLI.getVNInfoAt(Idx.getRegSlot());
377  assert(VNI && VNI->def == Idx.getRegSlot() && "Not defined by copy");
378#endif
379
380  unsigned SrcReg = CopyMI.getOperand(1).getReg();
381  LiveInterval &SrcLI = LIS.getInterval(SrcReg);
382  VNInfo *SrcVNI = SrcLI.getVNInfoAt(Idx);
383  LiveQueryResult SrcQ = SrcLI.Query(Idx);
384  MachineBasicBlock *DefMBB = LIS.getMBBFromIndex(SrcVNI->def);
385  if (DefMBB != CopyMI.getParent() || !SrcQ.isKill())
386    return false;
387
388  // Conservatively extend the stack slot range to the range of the original
389  // value. We may be able to do better with stack slot coloring by being more
390  // careful here.
391  assert(StackInt && "No stack slot assigned yet.");
392  LiveInterval &OrigLI = LIS.getInterval(Original);
393  VNInfo *OrigVNI = OrigLI.getVNInfoAt(Idx);
394  StackInt->MergeValueInAsValue(OrigLI, OrigVNI, StackInt->getValNumInfo(0));
395  LLVM_DEBUG(dbgs() << "\tmerged orig valno " << OrigVNI->id << ": "
396                    << *StackInt << '\n');
397
398  // We are going to spill SrcVNI immediately after its def, so clear out
399  // any later spills of the same value.
400  eliminateRedundantSpills(SrcLI, SrcVNI);
401
402  MachineBasicBlock *MBB = LIS.getMBBFromIndex(SrcVNI->def);
403  MachineBasicBlock::iterator MII;
404  if (SrcVNI->isPHIDef())
405    MII = MBB->SkipPHIsLabelsAndDebug(MBB->begin());
406  else {
407    MachineInstr *DefMI = LIS.getInstructionFromIndex(SrcVNI->def);
408    assert(DefMI && "Defining instruction disappeared");
409    MII = DefMI;
410    ++MII;
411  }
412  // Insert spill without kill flag immediately after def.
413  TII.storeRegToStackSlot(*MBB, MII, SrcReg, false, StackSlot,
414                          MRI.getRegClass(SrcReg), &TRI);
415  --MII; // Point to store instruction.
416  LIS.InsertMachineInstrInMaps(*MII);
417  LLVM_DEBUG(dbgs() << "\thoisted: " << SrcVNI->def << '\t' << *MII);
418
419  HSpiller.addToMergeableSpills(*MII, StackSlot, Original);
420  ++NumSpills;
421  return true;
422}
423
424/// eliminateRedundantSpills - SLI:VNI is known to be on the stack. Remove any
425/// redundant spills of this value in SLI.reg and sibling copies.
426void InlineSpiller::eliminateRedundantSpills(LiveInterval &SLI, VNInfo *VNI) {
427  assert(VNI && "Missing value");
428  SmallVector<std::pair<LiveInterval*, VNInfo*>, 8> WorkList;
429  WorkList.push_back(std::make_pair(&SLI, VNI));
430  assert(StackInt && "No stack slot assigned yet.");
431
432  do {
433    LiveInterval *LI;
434    std::tie(LI, VNI) = WorkList.pop_back_val();
435    unsigned Reg = LI->reg;
436    LLVM_DEBUG(dbgs() << "Checking redundant spills for " << VNI->id << '@'
437                      << VNI->def << " in " << *LI << '\n');
438
439    // Regs to spill are taken care of.
440    if (isRegToSpill(Reg))
441      continue;
442
443    // Add all of VNI's live range to StackInt.
444    StackInt->MergeValueInAsValue(*LI, VNI, StackInt->getValNumInfo(0));
445    LLVM_DEBUG(dbgs() << "Merged to stack int: " << *StackInt << '\n');
446
447    // Find all spills and copies of VNI.
448    for (MachineRegisterInfo::use_instr_nodbg_iterator
449         UI = MRI.use_instr_nodbg_begin(Reg), E = MRI.use_instr_nodbg_end();
450         UI != E; ) {
451      MachineInstr &MI = *UI++;
452      if (!MI.isCopy() && !MI.mayStore())
453        continue;
454      SlotIndex Idx = LIS.getInstructionIndex(MI);
455      if (LI->getVNInfoAt(Idx) != VNI)
456        continue;
457
458      // Follow sibling copies down the dominator tree.
459      if (unsigned DstReg = isFullCopyOf(MI, Reg)) {
460        if (isSibling(DstReg)) {
461           LiveInterval &DstLI = LIS.getInterval(DstReg);
462           VNInfo *DstVNI = DstLI.getVNInfoAt(Idx.getRegSlot());
463           assert(DstVNI && "Missing defined value");
464           assert(DstVNI->def == Idx.getRegSlot() && "Wrong copy def slot");
465           WorkList.push_back(std::make_pair(&DstLI, DstVNI));
466        }
467        continue;
468      }
469
470      // Erase spills.
471      int FI;
472      if (Reg == TII.isStoreToStackSlot(MI, FI) && FI == StackSlot) {
473        LLVM_DEBUG(dbgs() << "Redundant spill " << Idx << '\t' << MI);
474        // eliminateDeadDefs won't normally remove stores, so switch opcode.
475        MI.setDesc(TII.get(TargetOpcode::KILL));
476        DeadDefs.push_back(&MI);
477        ++NumSpillsRemoved;
478        if (HSpiller.rmFromMergeableSpills(MI, StackSlot))
479          --NumSpills;
480      }
481    }
482  } while (!WorkList.empty());
483}
484
485//===----------------------------------------------------------------------===//
486//                            Rematerialization
487//===----------------------------------------------------------------------===//
488
489/// markValueUsed - Remember that VNI failed to rematerialize, so its defining
490/// instruction cannot be eliminated. See through snippet copies
491void InlineSpiller::markValueUsed(LiveInterval *LI, VNInfo *VNI) {
492  SmallVector<std::pair<LiveInterval*, VNInfo*>, 8> WorkList;
493  WorkList.push_back(std::make_pair(LI, VNI));
494  do {
495    std::tie(LI, VNI) = WorkList.pop_back_val();
496    if (!UsedValues.insert(VNI).second)
497      continue;
498
499    if (VNI->isPHIDef()) {
500      MachineBasicBlock *MBB = LIS.getMBBFromIndex(VNI->def);
501      for (MachineBasicBlock *P : MBB->predecessors()) {
502        VNInfo *PVNI = LI->getVNInfoBefore(LIS.getMBBEndIdx(P));
503        if (PVNI)
504          WorkList.push_back(std::make_pair(LI, PVNI));
505      }
506      continue;
507    }
508
509    // Follow snippet copies.
510    MachineInstr *MI = LIS.getInstructionFromIndex(VNI->def);
511    if (!SnippetCopies.count(MI))
512      continue;
513    LiveInterval &SnipLI = LIS.getInterval(MI->getOperand(1).getReg());
514    assert(isRegToSpill(SnipLI.reg) && "Unexpected register in copy");
515    VNInfo *SnipVNI = SnipLI.getVNInfoAt(VNI->def.getRegSlot(true));
516    assert(SnipVNI && "Snippet undefined before copy");
517    WorkList.push_back(std::make_pair(&SnipLI, SnipVNI));
518  } while (!WorkList.empty());
519}
520
521bool InlineSpiller::canGuaranteeAssignmentAfterRemat(unsigned VReg,
522                                                     MachineInstr &MI) {
523  if (!RestrictStatepointRemat)
524    return true;
525  // Here's a quick explanation of the problem we're trying to handle here:
526  // * There are some pseudo instructions with more vreg uses than there are
527  //   physical registers on the machine.
528  // * This is normally handled by spilling the vreg, and folding the reload
529  //   into the user instruction.  (Thus decreasing the number of used vregs
530  //   until the remainder can be assigned to physregs.)
531  // * However, since we may try to spill vregs in any order, we can end up
532  //   trying to spill each operand to the instruction, and then rematting it
533  //   instead.  When that happens, the new live intervals (for the remats) are
534  //   expected to be trivially assignable (i.e. RS_Done).  However, since we
535  //   may have more remats than physregs, we're guaranteed to fail to assign
536  //   one.
537  // At the moment, we only handle this for STATEPOINTs since they're the only
538  // psuedo op where we've seen this.  If we start seeing other instructions
539  // with the same problem, we need to revisit this.
540  return (MI.getOpcode() != TargetOpcode::STATEPOINT);
541}
542
543/// reMaterializeFor - Attempt to rematerialize before MI instead of reloading.
544bool InlineSpiller::reMaterializeFor(LiveInterval &VirtReg, MachineInstr &MI) {
545  // Analyze instruction
546  SmallVector<std::pair<MachineInstr *, unsigned>, 8> Ops;
547  MIBundleOperands::VirtRegInfo RI =
548      MIBundleOperands(MI).analyzeVirtReg(VirtReg.reg, &Ops);
549
550  if (!RI.Reads)
551    return false;
552
553  SlotIndex UseIdx = LIS.getInstructionIndex(MI).getRegSlot(true);
554  VNInfo *ParentVNI = VirtReg.getVNInfoAt(UseIdx.getBaseIndex());
555
556  if (!ParentVNI) {
557    LLVM_DEBUG(dbgs() << "\tadding <undef> flags: ");
558    for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
559      MachineOperand &MO = MI.getOperand(i);
560      if (MO.isReg() && MO.isUse() && MO.getReg() == VirtReg.reg)
561        MO.setIsUndef();
562    }
563    LLVM_DEBUG(dbgs() << UseIdx << '\t' << MI);
564    return true;
565  }
566
567  if (SnippetCopies.count(&MI))
568    return false;
569
570  LiveInterval &OrigLI = LIS.getInterval(Original);
571  VNInfo *OrigVNI = OrigLI.getVNInfoAt(UseIdx);
572  LiveRangeEdit::Remat RM(ParentVNI);
573  RM.OrigMI = LIS.getInstructionFromIndex(OrigVNI->def);
574
575  if (!Edit->canRematerializeAt(RM, OrigVNI, UseIdx, false)) {
576    markValueUsed(&VirtReg, ParentVNI);
577    LLVM_DEBUG(dbgs() << "\tcannot remat for " << UseIdx << '\t' << MI);
578    return false;
579  }
580
581  // If the instruction also writes VirtReg.reg, it had better not require the
582  // same register for uses and defs.
583  if (RI.Tied) {
584    markValueUsed(&VirtReg, ParentVNI);
585    LLVM_DEBUG(dbgs() << "\tcannot remat tied reg: " << UseIdx << '\t' << MI);
586    return false;
587  }
588
589  // Before rematerializing into a register for a single instruction, try to
590  // fold a load into the instruction. That avoids allocating a new register.
591  if (RM.OrigMI->canFoldAsLoad() &&
592      foldMemoryOperand(Ops, RM.OrigMI)) {
593    Edit->markRematerialized(RM.ParentVNI);
594    ++NumFoldedLoads;
595    return true;
596  }
597
598  // If we can't guarantee that we'll be able to actually assign the new vreg,
599  // we can't remat.
600  if (!canGuaranteeAssignmentAfterRemat(VirtReg.reg, MI)) {
601    markValueUsed(&VirtReg, ParentVNI);
602    LLVM_DEBUG(dbgs() << "\tcannot remat for " << UseIdx << '\t' << MI);
603    return false;
604  }
605
606  // Allocate a new register for the remat.
607  unsigned NewVReg = Edit->createFrom(Original);
608
609  // Finally we can rematerialize OrigMI before MI.
610  SlotIndex DefIdx =
611      Edit->rematerializeAt(*MI.getParent(), MI, NewVReg, RM, TRI);
612
613  // We take the DebugLoc from MI, since OrigMI may be attributed to a
614  // different source location.
615  auto *NewMI = LIS.getInstructionFromIndex(DefIdx);
616  NewMI->setDebugLoc(MI.getDebugLoc());
617
618  (void)DefIdx;
619  LLVM_DEBUG(dbgs() << "\tremat:  " << DefIdx << '\t'
620                    << *LIS.getInstructionFromIndex(DefIdx));
621
622  // Replace operands
623  for (const auto &OpPair : Ops) {
624    MachineOperand &MO = OpPair.first->getOperand(OpPair.second);
625    if (MO.isReg() && MO.isUse() && MO.getReg() == VirtReg.reg) {
626      MO.setReg(NewVReg);
627      MO.setIsKill();
628    }
629  }
630  LLVM_DEBUG(dbgs() << "\t        " << UseIdx << '\t' << MI << '\n');
631
632  ++NumRemats;
633  return true;
634}
635
636/// reMaterializeAll - Try to rematerialize as many uses as possible,
637/// and trim the live ranges after.
638void InlineSpiller::reMaterializeAll() {
639  if (!Edit->anyRematerializable(AA))
640    return;
641
642  UsedValues.clear();
643
644  // Try to remat before all uses of snippets.
645  bool anyRemat = false;
646  for (unsigned Reg : RegsToSpill) {
647    LiveInterval &LI = LIS.getInterval(Reg);
648    for (MachineRegisterInfo::reg_bundle_iterator
649           RegI = MRI.reg_bundle_begin(Reg), E = MRI.reg_bundle_end();
650         RegI != E; ) {
651      MachineInstr &MI = *RegI++;
652
653      // Debug values are not allowed to affect codegen.
654      if (MI.isDebugValue())
655        continue;
656
657      assert(!MI.isDebugInstr() && "Did not expect to find a use in debug "
658             "instruction that isn't a DBG_VALUE");
659
660      anyRemat |= reMaterializeFor(LI, MI);
661    }
662  }
663  if (!anyRemat)
664    return;
665
666  // Remove any values that were completely rematted.
667  for (unsigned Reg : RegsToSpill) {
668    LiveInterval &LI = LIS.getInterval(Reg);
669    for (LiveInterval::vni_iterator I = LI.vni_begin(), E = LI.vni_end();
670         I != E; ++I) {
671      VNInfo *VNI = *I;
672      if (VNI->isUnused() || VNI->isPHIDef() || UsedValues.count(VNI))
673        continue;
674      MachineInstr *MI = LIS.getInstructionFromIndex(VNI->def);
675      MI->addRegisterDead(Reg, &TRI);
676      if (!MI->allDefsAreDead())
677        continue;
678      LLVM_DEBUG(dbgs() << "All defs dead: " << *MI);
679      DeadDefs.push_back(MI);
680    }
681  }
682
683  // Eliminate dead code after remat. Note that some snippet copies may be
684  // deleted here.
685  if (DeadDefs.empty())
686    return;
687  LLVM_DEBUG(dbgs() << "Remat created " << DeadDefs.size() << " dead defs.\n");
688  Edit->eliminateDeadDefs(DeadDefs, RegsToSpill, AA);
689
690  // LiveRangeEdit::eliminateDeadDef is used to remove dead define instructions
691  // after rematerialization.  To remove a VNI for a vreg from its LiveInterval,
692  // LiveIntervals::removeVRegDefAt is used. However, after non-PHI VNIs are all
693  // removed, PHI VNI are still left in the LiveInterval.
694  // So to get rid of unused reg, we need to check whether it has non-dbg
695  // reference instead of whether it has non-empty interval.
696  unsigned ResultPos = 0;
697  for (unsigned Reg : RegsToSpill) {
698    if (MRI.reg_nodbg_empty(Reg)) {
699      Edit->eraseVirtReg(Reg);
700      continue;
701    }
702
703    assert(LIS.hasInterval(Reg) &&
704           (!LIS.getInterval(Reg).empty() || !MRI.reg_nodbg_empty(Reg)) &&
705           "Empty and not used live-range?!");
706
707    RegsToSpill[ResultPos++] = Reg;
708  }
709  RegsToSpill.erase(RegsToSpill.begin() + ResultPos, RegsToSpill.end());
710  LLVM_DEBUG(dbgs() << RegsToSpill.size()
711                    << " registers to spill after remat.\n");
712}
713
714//===----------------------------------------------------------------------===//
715//                                 Spilling
716//===----------------------------------------------------------------------===//
717
718/// If MI is a load or store of StackSlot, it can be removed.
719bool InlineSpiller::coalesceStackAccess(MachineInstr *MI, unsigned Reg) {
720  int FI = 0;
721  unsigned InstrReg = TII.isLoadFromStackSlot(*MI, FI);
722  bool IsLoad = InstrReg;
723  if (!IsLoad)
724    InstrReg = TII.isStoreToStackSlot(*MI, FI);
725
726  // We have a stack access. Is it the right register and slot?
727  if (InstrReg != Reg || FI != StackSlot)
728    return false;
729
730  if (!IsLoad)
731    HSpiller.rmFromMergeableSpills(*MI, StackSlot);
732
733  LLVM_DEBUG(dbgs() << "Coalescing stack access: " << *MI);
734  LIS.RemoveMachineInstrFromMaps(*MI);
735  MI->eraseFromParent();
736
737  if (IsLoad) {
738    ++NumReloadsRemoved;
739    --NumReloads;
740  } else {
741    ++NumSpillsRemoved;
742    --NumSpills;
743  }
744
745  return true;
746}
747
748#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
749LLVM_DUMP_METHOD
750// Dump the range of instructions from B to E with their slot indexes.
751static void dumpMachineInstrRangeWithSlotIndex(MachineBasicBlock::iterator B,
752                                               MachineBasicBlock::iterator E,
753                                               LiveIntervals const &LIS,
754                                               const char *const header,
755                                               unsigned VReg =0) {
756  char NextLine = '\n';
757  char SlotIndent = '\t';
758
759  if (std::next(B) == E) {
760    NextLine = ' ';
761    SlotIndent = ' ';
762  }
763
764  dbgs() << '\t' << header << ": " << NextLine;
765
766  for (MachineBasicBlock::iterator I = B; I != E; ++I) {
767    SlotIndex Idx = LIS.getInstructionIndex(*I).getRegSlot();
768
769    // If a register was passed in and this instruction has it as a
770    // destination that is marked as an early clobber, print the
771    // early-clobber slot index.
772    if (VReg) {
773      MachineOperand *MO = I->findRegisterDefOperand(VReg);
774      if (MO && MO->isEarlyClobber())
775        Idx = Idx.getRegSlot(true);
776    }
777
778    dbgs() << SlotIndent << Idx << '\t' << *I;
779  }
780}
781#endif
782
783/// foldMemoryOperand - Try folding stack slot references in Ops into their
784/// instructions.
785///
786/// @param Ops    Operand indices from analyzeVirtReg().
787/// @param LoadMI Load instruction to use instead of stack slot when non-null.
788/// @return       True on success.
789bool InlineSpiller::
790foldMemoryOperand(ArrayRef<std::pair<MachineInstr *, unsigned>> Ops,
791                  MachineInstr *LoadMI) {
792  if (Ops.empty())
793    return false;
794  // Don't attempt folding in bundles.
795  MachineInstr *MI = Ops.front().first;
796  if (Ops.back().first != MI || MI->isBundled())
797    return false;
798
799  bool WasCopy = MI->isCopy();
800  unsigned ImpReg = 0;
801
802  // Spill subregs if the target allows it.
803  // We always want to spill subregs for stackmap/patchpoint pseudos.
804  bool SpillSubRegs = TII.isSubregFoldable() ||
805                      MI->getOpcode() == TargetOpcode::STATEPOINT ||
806                      MI->getOpcode() == TargetOpcode::PATCHPOINT ||
807                      MI->getOpcode() == TargetOpcode::STACKMAP;
808
809  // TargetInstrInfo::foldMemoryOperand only expects explicit, non-tied
810  // operands.
811  SmallVector<unsigned, 8> FoldOps;
812  for (const auto &OpPair : Ops) {
813    unsigned Idx = OpPair.second;
814    assert(MI == OpPair.first && "Instruction conflict during operand folding");
815    MachineOperand &MO = MI->getOperand(Idx);
816    if (MO.isImplicit()) {
817      ImpReg = MO.getReg();
818      continue;
819    }
820
821    if (!SpillSubRegs && MO.getSubReg())
822      return false;
823    // We cannot fold a load instruction into a def.
824    if (LoadMI && MO.isDef())
825      return false;
826    // Tied use operands should not be passed to foldMemoryOperand.
827    if (!MI->isRegTiedToDefOperand(Idx))
828      FoldOps.push_back(Idx);
829  }
830
831  // If we only have implicit uses, we won't be able to fold that.
832  // Moreover, TargetInstrInfo::foldMemoryOperand will assert if we try!
833  if (FoldOps.empty())
834    return false;
835
836  MachineInstrSpan MIS(MI, MI->getParent());
837
838  MachineInstr *FoldMI =
839      LoadMI ? TII.foldMemoryOperand(*MI, FoldOps, *LoadMI, &LIS)
840             : TII.foldMemoryOperand(*MI, FoldOps, StackSlot, &LIS, &VRM);
841  if (!FoldMI)
842    return false;
843
844  // Remove LIS for any dead defs in the original MI not in FoldMI.
845  for (MIBundleOperands MO(*MI); MO.isValid(); ++MO) {
846    if (!MO->isReg())
847      continue;
848    unsigned Reg = MO->getReg();
849    if (!Reg || TargetRegisterInfo::isVirtualRegister(Reg) ||
850        MRI.isReserved(Reg)) {
851      continue;
852    }
853    // Skip non-Defs, including undef uses and internal reads.
854    if (MO->isUse())
855      continue;
856    MIBundleOperands::PhysRegInfo RI =
857        MIBundleOperands(*FoldMI).analyzePhysReg(Reg, &TRI);
858    if (RI.FullyDefined)
859      continue;
860    // FoldMI does not define this physreg. Remove the LI segment.
861    assert(MO->isDead() && "Cannot fold physreg def");
862    SlotIndex Idx = LIS.getInstructionIndex(*MI).getRegSlot();
863    LIS.removePhysRegDefAt(Reg, Idx);
864  }
865
866  int FI;
867  if (TII.isStoreToStackSlot(*MI, FI) &&
868      HSpiller.rmFromMergeableSpills(*MI, FI))
869    --NumSpills;
870  LIS.ReplaceMachineInstrInMaps(*MI, *FoldMI);
871  if (MI->isCall())
872    MI->getMF()->updateCallSiteInfo(MI, FoldMI);
873  MI->eraseFromParent();
874
875  // Insert any new instructions other than FoldMI into the LIS maps.
876  assert(!MIS.empty() && "Unexpected empty span of instructions!");
877  for (MachineInstr &MI : MIS)
878    if (&MI != FoldMI)
879      LIS.InsertMachineInstrInMaps(MI);
880
881  // TII.foldMemoryOperand may have left some implicit operands on the
882  // instruction.  Strip them.
883  if (ImpReg)
884    for (unsigned i = FoldMI->getNumOperands(); i; --i) {
885      MachineOperand &MO = FoldMI->getOperand(i - 1);
886      if (!MO.isReg() || !MO.isImplicit())
887        break;
888      if (MO.getReg() == ImpReg)
889        FoldMI->RemoveOperand(i - 1);
890    }
891
892  LLVM_DEBUG(dumpMachineInstrRangeWithSlotIndex(MIS.begin(), MIS.end(), LIS,
893                                                "folded"));
894
895  if (!WasCopy)
896    ++NumFolded;
897  else if (Ops.front().second == 0) {
898    ++NumSpills;
899    HSpiller.addToMergeableSpills(*FoldMI, StackSlot, Original);
900  } else
901    ++NumReloads;
902  return true;
903}
904
905void InlineSpiller::insertReload(unsigned NewVReg,
906                                 SlotIndex Idx,
907                                 MachineBasicBlock::iterator MI) {
908  MachineBasicBlock &MBB = *MI->getParent();
909
910  MachineInstrSpan MIS(MI, &MBB);
911  TII.loadRegFromStackSlot(MBB, MI, NewVReg, StackSlot,
912                           MRI.getRegClass(NewVReg), &TRI);
913
914  LIS.InsertMachineInstrRangeInMaps(MIS.begin(), MI);
915
916  LLVM_DEBUG(dumpMachineInstrRangeWithSlotIndex(MIS.begin(), MI, LIS, "reload",
917                                                NewVReg));
918  ++NumReloads;
919}
920
921/// Check if \p Def fully defines a VReg with an undefined value.
922/// If that's the case, that means the value of VReg is actually
923/// not relevant.
924static bool isFullUndefDef(const MachineInstr &Def) {
925  if (!Def.isImplicitDef())
926    return false;
927  assert(Def.getNumOperands() == 1 &&
928         "Implicit def with more than one definition");
929  // We can say that the VReg defined by Def is undef, only if it is
930  // fully defined by Def. Otherwise, some of the lanes may not be
931  // undef and the value of the VReg matters.
932  return !Def.getOperand(0).getSubReg();
933}
934
935/// insertSpill - Insert a spill of NewVReg after MI.
936void InlineSpiller::insertSpill(unsigned NewVReg, bool isKill,
937                                 MachineBasicBlock::iterator MI) {
938  MachineBasicBlock &MBB = *MI->getParent();
939
940  MachineInstrSpan MIS(MI, &MBB);
941  bool IsRealSpill = true;
942  if (isFullUndefDef(*MI)) {
943    // Don't spill undef value.
944    // Anything works for undef, in particular keeping the memory
945    // uninitialized is a viable option and it saves code size and
946    // run time.
947    BuildMI(MBB, std::next(MI), MI->getDebugLoc(), TII.get(TargetOpcode::KILL))
948        .addReg(NewVReg, getKillRegState(isKill));
949    IsRealSpill = false;
950  } else
951    TII.storeRegToStackSlot(MBB, std::next(MI), NewVReg, isKill, StackSlot,
952                            MRI.getRegClass(NewVReg), &TRI);
953
954  LIS.InsertMachineInstrRangeInMaps(std::next(MI), MIS.end());
955
956  LLVM_DEBUG(dumpMachineInstrRangeWithSlotIndex(std::next(MI), MIS.end(), LIS,
957                                                "spill"));
958  ++NumSpills;
959  if (IsRealSpill)
960    HSpiller.addToMergeableSpills(*std::next(MI), StackSlot, Original);
961}
962
963/// spillAroundUses - insert spill code around each use of Reg.
964void InlineSpiller::spillAroundUses(unsigned Reg) {
965  LLVM_DEBUG(dbgs() << "spillAroundUses " << printReg(Reg) << '\n');
966  LiveInterval &OldLI = LIS.getInterval(Reg);
967
968  // Iterate over instructions using Reg.
969  for (MachineRegisterInfo::reg_bundle_iterator
970       RegI = MRI.reg_bundle_begin(Reg), E = MRI.reg_bundle_end();
971       RegI != E; ) {
972    MachineInstr *MI = &*(RegI++);
973
974    // Debug values are not allowed to affect codegen.
975    if (MI->isDebugValue()) {
976      // Modify DBG_VALUE now that the value is in a spill slot.
977      MachineBasicBlock *MBB = MI->getParent();
978      LLVM_DEBUG(dbgs() << "Modifying debug info due to spill:\t" << *MI);
979      buildDbgValueForSpill(*MBB, MI, *MI, StackSlot);
980      MBB->erase(MI);
981      continue;
982    }
983
984    assert(!MI->isDebugInstr() && "Did not expect to find a use in debug "
985           "instruction that isn't a DBG_VALUE");
986
987    // Ignore copies to/from snippets. We'll delete them.
988    if (SnippetCopies.count(MI))
989      continue;
990
991    // Stack slot accesses may coalesce away.
992    if (coalesceStackAccess(MI, Reg))
993      continue;
994
995    // Analyze instruction.
996    SmallVector<std::pair<MachineInstr*, unsigned>, 8> Ops;
997    MIBundleOperands::VirtRegInfo RI =
998        MIBundleOperands(*MI).analyzeVirtReg(Reg, &Ops);
999
1000    // Find the slot index where this instruction reads and writes OldLI.
1001    // This is usually the def slot, except for tied early clobbers.
1002    SlotIndex Idx = LIS.getInstructionIndex(*MI).getRegSlot();
1003    if (VNInfo *VNI = OldLI.getVNInfoAt(Idx.getRegSlot(true)))
1004      if (SlotIndex::isSameInstr(Idx, VNI->def))
1005        Idx = VNI->def;
1006
1007    // Check for a sibling copy.
1008    unsigned SibReg = isFullCopyOf(*MI, Reg);
1009    if (SibReg && isSibling(SibReg)) {
1010      // This may actually be a copy between snippets.
1011      if (isRegToSpill(SibReg)) {
1012        LLVM_DEBUG(dbgs() << "Found new snippet copy: " << *MI);
1013        SnippetCopies.insert(MI);
1014        continue;
1015      }
1016      if (RI.Writes) {
1017        if (hoistSpillInsideBB(OldLI, *MI)) {
1018          // This COPY is now dead, the value is already in the stack slot.
1019          MI->getOperand(0).setIsDead();
1020          DeadDefs.push_back(MI);
1021          continue;
1022        }
1023      } else {
1024        // This is a reload for a sib-reg copy. Drop spills downstream.
1025        LiveInterval &SibLI = LIS.getInterval(SibReg);
1026        eliminateRedundantSpills(SibLI, SibLI.getVNInfoAt(Idx));
1027        // The COPY will fold to a reload below.
1028      }
1029    }
1030
1031    // Attempt to fold memory ops.
1032    if (foldMemoryOperand(Ops))
1033      continue;
1034
1035    // Create a new virtual register for spill/fill.
1036    // FIXME: Infer regclass from instruction alone.
1037    unsigned NewVReg = Edit->createFrom(Reg);
1038
1039    if (RI.Reads)
1040      insertReload(NewVReg, Idx, MI);
1041
1042    // Rewrite instruction operands.
1043    bool hasLiveDef = false;
1044    for (const auto &OpPair : Ops) {
1045      MachineOperand &MO = OpPair.first->getOperand(OpPair.second);
1046      MO.setReg(NewVReg);
1047      if (MO.isUse()) {
1048        if (!OpPair.first->isRegTiedToDefOperand(OpPair.second))
1049          MO.setIsKill();
1050      } else {
1051        if (!MO.isDead())
1052          hasLiveDef = true;
1053      }
1054    }
1055    LLVM_DEBUG(dbgs() << "\trewrite: " << Idx << '\t' << *MI << '\n');
1056
1057    // FIXME: Use a second vreg if instruction has no tied ops.
1058    if (RI.Writes)
1059      if (hasLiveDef)
1060        insertSpill(NewVReg, true, MI);
1061  }
1062}
1063
1064/// spillAll - Spill all registers remaining after rematerialization.
1065void InlineSpiller::spillAll() {
1066  // Update LiveStacks now that we are committed to spilling.
1067  if (StackSlot == VirtRegMap::NO_STACK_SLOT) {
1068    StackSlot = VRM.assignVirt2StackSlot(Original);
1069    StackInt = &LSS.getOrCreateInterval(StackSlot, MRI.getRegClass(Original));
1070    StackInt->getNextValue(SlotIndex(), LSS.getVNInfoAllocator());
1071  } else
1072    StackInt = &LSS.getInterval(StackSlot);
1073
1074  if (Original != Edit->getReg())
1075    VRM.assignVirt2StackSlot(Edit->getReg(), StackSlot);
1076
1077  assert(StackInt->getNumValNums() == 1 && "Bad stack interval values");
1078  for (unsigned Reg : RegsToSpill)
1079    StackInt->MergeSegmentsInAsValue(LIS.getInterval(Reg),
1080                                     StackInt->getValNumInfo(0));
1081  LLVM_DEBUG(dbgs() << "Merged spilled regs: " << *StackInt << '\n');
1082
1083  // Spill around uses of all RegsToSpill.
1084  for (unsigned Reg : RegsToSpill)
1085    spillAroundUses(Reg);
1086
1087  // Hoisted spills may cause dead code.
1088  if (!DeadDefs.empty()) {
1089    LLVM_DEBUG(dbgs() << "Eliminating " << DeadDefs.size() << " dead defs\n");
1090    Edit->eliminateDeadDefs(DeadDefs, RegsToSpill, AA);
1091  }
1092
1093  // Finally delete the SnippetCopies.
1094  for (unsigned Reg : RegsToSpill) {
1095    for (MachineRegisterInfo::reg_instr_iterator
1096         RI = MRI.reg_instr_begin(Reg), E = MRI.reg_instr_end();
1097         RI != E; ) {
1098      MachineInstr &MI = *(RI++);
1099      assert(SnippetCopies.count(&MI) && "Remaining use wasn't a snippet copy");
1100      // FIXME: Do this with a LiveRangeEdit callback.
1101      LIS.RemoveMachineInstrFromMaps(MI);
1102      MI.eraseFromParent();
1103    }
1104  }
1105
1106  // Delete all spilled registers.
1107  for (unsigned Reg : RegsToSpill)
1108    Edit->eraseVirtReg(Reg);
1109}
1110
1111void InlineSpiller::spill(LiveRangeEdit &edit) {
1112  ++NumSpilledRanges;
1113  Edit = &edit;
1114  assert(!TargetRegisterInfo::isStackSlot(edit.getReg())
1115         && "Trying to spill a stack slot.");
1116  // Share a stack slot among all descendants of Original.
1117  Original = VRM.getOriginal(edit.getReg());
1118  StackSlot = VRM.getStackSlot(Original);
1119  StackInt = nullptr;
1120
1121  LLVM_DEBUG(dbgs() << "Inline spilling "
1122                    << TRI.getRegClassName(MRI.getRegClass(edit.getReg()))
1123                    << ':' << edit.getParent() << "\nFrom original "
1124                    << printReg(Original) << '\n');
1125  assert(edit.getParent().isSpillable() &&
1126         "Attempting to spill already spilled value.");
1127  assert(DeadDefs.empty() && "Previous spill didn't remove dead defs");
1128
1129  collectRegsToSpill();
1130  reMaterializeAll();
1131
1132  // Remat may handle everything.
1133  if (!RegsToSpill.empty())
1134    spillAll();
1135
1136  Edit->calculateRegClassAndHint(MF, Loops, MBFI);
1137}
1138
1139/// Optimizations after all the reg selections and spills are done.
1140void InlineSpiller::postOptimization() { HSpiller.hoistAllSpills(); }
1141
1142/// When a spill is inserted, add the spill to MergeableSpills map.
1143void HoistSpillHelper::addToMergeableSpills(MachineInstr &Spill, int StackSlot,
1144                                            unsigned Original) {
1145  BumpPtrAllocator &Allocator = LIS.getVNInfoAllocator();
1146  LiveInterval &OrigLI = LIS.getInterval(Original);
1147  // save a copy of LiveInterval in StackSlotToOrigLI because the original
1148  // LiveInterval may be cleared after all its references are spilled.
1149  if (StackSlotToOrigLI.find(StackSlot) == StackSlotToOrigLI.end()) {
1150    auto LI = llvm::make_unique<LiveInterval>(OrigLI.reg, OrigLI.weight);
1151    LI->assign(OrigLI, Allocator);
1152    StackSlotToOrigLI[StackSlot] = std::move(LI);
1153  }
1154  SlotIndex Idx = LIS.getInstructionIndex(Spill);
1155  VNInfo *OrigVNI = StackSlotToOrigLI[StackSlot]->getVNInfoAt(Idx.getRegSlot());
1156  std::pair<int, VNInfo *> MIdx = std::make_pair(StackSlot, OrigVNI);
1157  MergeableSpills[MIdx].insert(&Spill);
1158}
1159
1160/// When a spill is removed, remove the spill from MergeableSpills map.
1161/// Return true if the spill is removed successfully.
1162bool HoistSpillHelper::rmFromMergeableSpills(MachineInstr &Spill,
1163                                             int StackSlot) {
1164  auto It = StackSlotToOrigLI.find(StackSlot);
1165  if (It == StackSlotToOrigLI.end())
1166    return false;
1167  SlotIndex Idx = LIS.getInstructionIndex(Spill);
1168  VNInfo *OrigVNI = It->second->getVNInfoAt(Idx.getRegSlot());
1169  std::pair<int, VNInfo *> MIdx = std::make_pair(StackSlot, OrigVNI);
1170  return MergeableSpills[MIdx].erase(&Spill);
1171}
1172
1173/// Check BB to see if it is a possible target BB to place a hoisted spill,
1174/// i.e., there should be a living sibling of OrigReg at the insert point.
1175bool HoistSpillHelper::isSpillCandBB(LiveInterval &OrigLI, VNInfo &OrigVNI,
1176                                     MachineBasicBlock &BB, unsigned &LiveReg) {
1177  SlotIndex Idx;
1178  unsigned OrigReg = OrigLI.reg;
1179  MachineBasicBlock::iterator MI = IPA.getLastInsertPointIter(OrigLI, BB);
1180  if (MI != BB.end())
1181    Idx = LIS.getInstructionIndex(*MI);
1182  else
1183    Idx = LIS.getMBBEndIdx(&BB).getPrevSlot();
1184  SmallSetVector<unsigned, 16> &Siblings = Virt2SiblingsMap[OrigReg];
1185  assert(OrigLI.getVNInfoAt(Idx) == &OrigVNI && "Unexpected VNI");
1186
1187  for (auto const SibReg : Siblings) {
1188    LiveInterval &LI = LIS.getInterval(SibReg);
1189    VNInfo *VNI = LI.getVNInfoAt(Idx);
1190    if (VNI) {
1191      LiveReg = SibReg;
1192      return true;
1193    }
1194  }
1195  return false;
1196}
1197
1198/// Remove redundant spills in the same BB. Save those redundant spills in
1199/// SpillsToRm, and save the spill to keep and its BB in SpillBBToSpill map.
1200void HoistSpillHelper::rmRedundantSpills(
1201    SmallPtrSet<MachineInstr *, 16> &Spills,
1202    SmallVectorImpl<MachineInstr *> &SpillsToRm,
1203    DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill) {
1204  // For each spill saw, check SpillBBToSpill[] and see if its BB already has
1205  // another spill inside. If a BB contains more than one spill, only keep the
1206  // earlier spill with smaller SlotIndex.
1207  for (const auto CurrentSpill : Spills) {
1208    MachineBasicBlock *Block = CurrentSpill->getParent();
1209    MachineDomTreeNode *Node = MDT.getBase().getNode(Block);
1210    MachineInstr *PrevSpill = SpillBBToSpill[Node];
1211    if (PrevSpill) {
1212      SlotIndex PIdx = LIS.getInstructionIndex(*PrevSpill);
1213      SlotIndex CIdx = LIS.getInstructionIndex(*CurrentSpill);
1214      MachineInstr *SpillToRm = (CIdx > PIdx) ? CurrentSpill : PrevSpill;
1215      MachineInstr *SpillToKeep = (CIdx > PIdx) ? PrevSpill : CurrentSpill;
1216      SpillsToRm.push_back(SpillToRm);
1217      SpillBBToSpill[MDT.getBase().getNode(Block)] = SpillToKeep;
1218    } else {
1219      SpillBBToSpill[MDT.getBase().getNode(Block)] = CurrentSpill;
1220    }
1221  }
1222  for (const auto SpillToRm : SpillsToRm)
1223    Spills.erase(SpillToRm);
1224}
1225
1226/// Starting from \p Root find a top-down traversal order of the dominator
1227/// tree to visit all basic blocks containing the elements of \p Spills.
1228/// Redundant spills will be found and put into \p SpillsToRm at the same
1229/// time. \p SpillBBToSpill will be populated as part of the process and
1230/// maps a basic block to the first store occurring in the basic block.
1231/// \post SpillsToRm.union(Spills\@post) == Spills\@pre
1232void HoistSpillHelper::getVisitOrders(
1233    MachineBasicBlock *Root, SmallPtrSet<MachineInstr *, 16> &Spills,
1234    SmallVectorImpl<MachineDomTreeNode *> &Orders,
1235    SmallVectorImpl<MachineInstr *> &SpillsToRm,
1236    DenseMap<MachineDomTreeNode *, unsigned> &SpillsToKeep,
1237    DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill) {
1238  // The set contains all the possible BB nodes to which we may hoist
1239  // original spills.
1240  SmallPtrSet<MachineDomTreeNode *, 8> WorkSet;
1241  // Save the BB nodes on the path from the first BB node containing
1242  // non-redundant spill to the Root node.
1243  SmallPtrSet<MachineDomTreeNode *, 8> NodesOnPath;
1244  // All the spills to be hoisted must originate from a single def instruction
1245  // to the OrigReg. It means the def instruction should dominate all the spills
1246  // to be hoisted. We choose the BB where the def instruction is located as
1247  // the Root.
1248  MachineDomTreeNode *RootIDomNode = MDT[Root]->getIDom();
1249  // For every node on the dominator tree with spill, walk up on the dominator
1250  // tree towards the Root node until it is reached. If there is other node
1251  // containing spill in the middle of the path, the previous spill saw will
1252  // be redundant and the node containing it will be removed. All the nodes on
1253  // the path starting from the first node with non-redundant spill to the Root
1254  // node will be added to the WorkSet, which will contain all the possible
1255  // locations where spills may be hoisted to after the loop below is done.
1256  for (const auto Spill : Spills) {
1257    MachineBasicBlock *Block = Spill->getParent();
1258    MachineDomTreeNode *Node = MDT[Block];
1259    MachineInstr *SpillToRm = nullptr;
1260    while (Node != RootIDomNode) {
1261      // If Node dominates Block, and it already contains a spill, the spill in
1262      // Block will be redundant.
1263      if (Node != MDT[Block] && SpillBBToSpill[Node]) {
1264        SpillToRm = SpillBBToSpill[MDT[Block]];
1265        break;
1266        /// If we see the Node already in WorkSet, the path from the Node to
1267        /// the Root node must already be traversed by another spill.
1268        /// Then no need to repeat.
1269      } else if (WorkSet.count(Node)) {
1270        break;
1271      } else {
1272        NodesOnPath.insert(Node);
1273      }
1274      Node = Node->getIDom();
1275    }
1276    if (SpillToRm) {
1277      SpillsToRm.push_back(SpillToRm);
1278    } else {
1279      // Add a BB containing the original spills to SpillsToKeep -- i.e.,
1280      // set the initial status before hoisting start. The value of BBs
1281      // containing original spills is set to 0, in order to descriminate
1282      // with BBs containing hoisted spills which will be inserted to
1283      // SpillsToKeep later during hoisting.
1284      SpillsToKeep[MDT[Block]] = 0;
1285      WorkSet.insert(NodesOnPath.begin(), NodesOnPath.end());
1286    }
1287    NodesOnPath.clear();
1288  }
1289
1290  // Sort the nodes in WorkSet in top-down order and save the nodes
1291  // in Orders. Orders will be used for hoisting in runHoistSpills.
1292  unsigned idx = 0;
1293  Orders.push_back(MDT.getBase().getNode(Root));
1294  do {
1295    MachineDomTreeNode *Node = Orders[idx++];
1296    const std::vector<MachineDomTreeNode *> &Children = Node->getChildren();
1297    unsigned NumChildren = Children.size();
1298    for (unsigned i = 0; i != NumChildren; ++i) {
1299      MachineDomTreeNode *Child = Children[i];
1300      if (WorkSet.count(Child))
1301        Orders.push_back(Child);
1302    }
1303  } while (idx != Orders.size());
1304  assert(Orders.size() == WorkSet.size() &&
1305         "Orders have different size with WorkSet");
1306
1307#ifndef NDEBUG
1308  LLVM_DEBUG(dbgs() << "Orders size is " << Orders.size() << "\n");
1309  SmallVector<MachineDomTreeNode *, 32>::reverse_iterator RIt = Orders.rbegin();
1310  for (; RIt != Orders.rend(); RIt++)
1311    LLVM_DEBUG(dbgs() << "BB" << (*RIt)->getBlock()->getNumber() << ",");
1312  LLVM_DEBUG(dbgs() << "\n");
1313#endif
1314}
1315
1316/// Try to hoist spills according to BB hotness. The spills to removed will
1317/// be saved in \p SpillsToRm. The spills to be inserted will be saved in
1318/// \p SpillsToIns.
1319void HoistSpillHelper::runHoistSpills(
1320    LiveInterval &OrigLI, VNInfo &OrigVNI,
1321    SmallPtrSet<MachineInstr *, 16> &Spills,
1322    SmallVectorImpl<MachineInstr *> &SpillsToRm,
1323    DenseMap<MachineBasicBlock *, unsigned> &SpillsToIns) {
1324  // Visit order of dominator tree nodes.
1325  SmallVector<MachineDomTreeNode *, 32> Orders;
1326  // SpillsToKeep contains all the nodes where spills are to be inserted
1327  // during hoisting. If the spill to be inserted is an original spill
1328  // (not a hoisted one), the value of the map entry is 0. If the spill
1329  // is a hoisted spill, the value of the map entry is the VReg to be used
1330  // as the source of the spill.
1331  DenseMap<MachineDomTreeNode *, unsigned> SpillsToKeep;
1332  // Map from BB to the first spill inside of it.
1333  DenseMap<MachineDomTreeNode *, MachineInstr *> SpillBBToSpill;
1334
1335  rmRedundantSpills(Spills, SpillsToRm, SpillBBToSpill);
1336
1337  MachineBasicBlock *Root = LIS.getMBBFromIndex(OrigVNI.def);
1338  getVisitOrders(Root, Spills, Orders, SpillsToRm, SpillsToKeep,
1339                 SpillBBToSpill);
1340
1341  // SpillsInSubTreeMap keeps the map from a dom tree node to a pair of
1342  // nodes set and the cost of all the spills inside those nodes.
1343  // The nodes set are the locations where spills are to be inserted
1344  // in the subtree of current node.
1345  using NodesCostPair =
1346      std::pair<SmallPtrSet<MachineDomTreeNode *, 16>, BlockFrequency>;
1347  DenseMap<MachineDomTreeNode *, NodesCostPair> SpillsInSubTreeMap;
1348
1349  // Iterate Orders set in reverse order, which will be a bottom-up order
1350  // in the dominator tree. Once we visit a dom tree node, we know its
1351  // children have already been visited and the spill locations in the
1352  // subtrees of all the children have been determined.
1353  SmallVector<MachineDomTreeNode *, 32>::reverse_iterator RIt = Orders.rbegin();
1354  for (; RIt != Orders.rend(); RIt++) {
1355    MachineBasicBlock *Block = (*RIt)->getBlock();
1356
1357    // If Block contains an original spill, simply continue.
1358    if (SpillsToKeep.find(*RIt) != SpillsToKeep.end() && !SpillsToKeep[*RIt]) {
1359      SpillsInSubTreeMap[*RIt].first.insert(*RIt);
1360      // SpillsInSubTreeMap[*RIt].second contains the cost of spill.
1361      SpillsInSubTreeMap[*RIt].second = MBFI.getBlockFreq(Block);
1362      continue;
1363    }
1364
1365    // Collect spills in subtree of current node (*RIt) to
1366    // SpillsInSubTreeMap[*RIt].first.
1367    const std::vector<MachineDomTreeNode *> &Children = (*RIt)->getChildren();
1368    unsigned NumChildren = Children.size();
1369    for (unsigned i = 0; i != NumChildren; ++i) {
1370      MachineDomTreeNode *Child = Children[i];
1371      if (SpillsInSubTreeMap.find(Child) == SpillsInSubTreeMap.end())
1372        continue;
1373      // The stmt "SpillsInSubTree = SpillsInSubTreeMap[*RIt].first" below
1374      // should be placed before getting the begin and end iterators of
1375      // SpillsInSubTreeMap[Child].first, or else the iterators may be
1376      // invalidated when SpillsInSubTreeMap[*RIt] is seen the first time
1377      // and the map grows and then the original buckets in the map are moved.
1378      SmallPtrSet<MachineDomTreeNode *, 16> &SpillsInSubTree =
1379          SpillsInSubTreeMap[*RIt].first;
1380      BlockFrequency &SubTreeCost = SpillsInSubTreeMap[*RIt].second;
1381      SubTreeCost += SpillsInSubTreeMap[Child].second;
1382      auto BI = SpillsInSubTreeMap[Child].first.begin();
1383      auto EI = SpillsInSubTreeMap[Child].first.end();
1384      SpillsInSubTree.insert(BI, EI);
1385      SpillsInSubTreeMap.erase(Child);
1386    }
1387
1388    SmallPtrSet<MachineDomTreeNode *, 16> &SpillsInSubTree =
1389          SpillsInSubTreeMap[*RIt].first;
1390    BlockFrequency &SubTreeCost = SpillsInSubTreeMap[*RIt].second;
1391    // No spills in subtree, simply continue.
1392    if (SpillsInSubTree.empty())
1393      continue;
1394
1395    // Check whether Block is a possible candidate to insert spill.
1396    unsigned LiveReg = 0;
1397    if (!isSpillCandBB(OrigLI, OrigVNI, *Block, LiveReg))
1398      continue;
1399
1400    // If there are multiple spills that could be merged, bias a little
1401    // to hoist the spill.
1402    BranchProbability MarginProb = (SpillsInSubTree.size() > 1)
1403                                       ? BranchProbability(9, 10)
1404                                       : BranchProbability(1, 1);
1405    if (SubTreeCost > MBFI.getBlockFreq(Block) * MarginProb) {
1406      // Hoist: Move spills to current Block.
1407      for (const auto SpillBB : SpillsInSubTree) {
1408        // When SpillBB is a BB contains original spill, insert the spill
1409        // to SpillsToRm.
1410        if (SpillsToKeep.find(SpillBB) != SpillsToKeep.end() &&
1411            !SpillsToKeep[SpillBB]) {
1412          MachineInstr *SpillToRm = SpillBBToSpill[SpillBB];
1413          SpillsToRm.push_back(SpillToRm);
1414        }
1415        // SpillBB will not contain spill anymore, remove it from SpillsToKeep.
1416        SpillsToKeep.erase(SpillBB);
1417      }
1418      // Current Block is the BB containing the new hoisted spill. Add it to
1419      // SpillsToKeep. LiveReg is the source of the new spill.
1420      SpillsToKeep[*RIt] = LiveReg;
1421      LLVM_DEBUG({
1422        dbgs() << "spills in BB: ";
1423        for (const auto Rspill : SpillsInSubTree)
1424          dbgs() << Rspill->getBlock()->getNumber() << " ";
1425        dbgs() << "were promoted to BB" << (*RIt)->getBlock()->getNumber()
1426               << "\n";
1427      });
1428      SpillsInSubTree.clear();
1429      SpillsInSubTree.insert(*RIt);
1430      SubTreeCost = MBFI.getBlockFreq(Block);
1431    }
1432  }
1433  // For spills in SpillsToKeep with LiveReg set (i.e., not original spill),
1434  // save them to SpillsToIns.
1435  for (const auto Ent : SpillsToKeep) {
1436    if (Ent.second)
1437      SpillsToIns[Ent.first->getBlock()] = Ent.second;
1438  }
1439}
1440
1441/// For spills with equal values, remove redundant spills and hoist those left
1442/// to less hot spots.
1443///
1444/// Spills with equal values will be collected into the same set in
1445/// MergeableSpills when spill is inserted. These equal spills are originated
1446/// from the same defining instruction and are dominated by the instruction.
1447/// Before hoisting all the equal spills, redundant spills inside in the same
1448/// BB are first marked to be deleted. Then starting from the spills left, walk
1449/// up on the dominator tree towards the Root node where the define instruction
1450/// is located, mark the dominated spills to be deleted along the way and
1451/// collect the BB nodes on the path from non-dominated spills to the define
1452/// instruction into a WorkSet. The nodes in WorkSet are the candidate places
1453/// where we are considering to hoist the spills. We iterate the WorkSet in
1454/// bottom-up order, and for each node, we will decide whether to hoist spills
1455/// inside its subtree to that node. In this way, we can get benefit locally
1456/// even if hoisting all the equal spills to one cold place is impossible.
1457void HoistSpillHelper::hoistAllSpills() {
1458  SmallVector<unsigned, 4> NewVRegs;
1459  LiveRangeEdit Edit(nullptr, NewVRegs, MF, LIS, &VRM, this);
1460
1461  for (unsigned i = 0, e = MRI.getNumVirtRegs(); i != e; ++i) {
1462    unsigned Reg = TargetRegisterInfo::index2VirtReg(i);
1463    unsigned Original = VRM.getPreSplitReg(Reg);
1464    if (!MRI.def_empty(Reg))
1465      Virt2SiblingsMap[Original].insert(Reg);
1466  }
1467
1468  // Each entry in MergeableSpills contains a spill set with equal values.
1469  for (auto &Ent : MergeableSpills) {
1470    int Slot = Ent.first.first;
1471    LiveInterval &OrigLI = *StackSlotToOrigLI[Slot];
1472    VNInfo *OrigVNI = Ent.first.second;
1473    SmallPtrSet<MachineInstr *, 16> &EqValSpills = Ent.second;
1474    if (Ent.second.empty())
1475      continue;
1476
1477    LLVM_DEBUG({
1478      dbgs() << "\nFor Slot" << Slot << " and VN" << OrigVNI->id << ":\n"
1479             << "Equal spills in BB: ";
1480      for (const auto spill : EqValSpills)
1481        dbgs() << spill->getParent()->getNumber() << " ";
1482      dbgs() << "\n";
1483    });
1484
1485    // SpillsToRm is the spill set to be removed from EqValSpills.
1486    SmallVector<MachineInstr *, 16> SpillsToRm;
1487    // SpillsToIns is the spill set to be newly inserted after hoisting.
1488    DenseMap<MachineBasicBlock *, unsigned> SpillsToIns;
1489
1490    runHoistSpills(OrigLI, *OrigVNI, EqValSpills, SpillsToRm, SpillsToIns);
1491
1492    LLVM_DEBUG({
1493      dbgs() << "Finally inserted spills in BB: ";
1494      for (const auto Ispill : SpillsToIns)
1495        dbgs() << Ispill.first->getNumber() << " ";
1496      dbgs() << "\nFinally removed spills in BB: ";
1497      for (const auto Rspill : SpillsToRm)
1498        dbgs() << Rspill->getParent()->getNumber() << " ";
1499      dbgs() << "\n";
1500    });
1501
1502    // Stack live range update.
1503    LiveInterval &StackIntvl = LSS.getInterval(Slot);
1504    if (!SpillsToIns.empty() || !SpillsToRm.empty())
1505      StackIntvl.MergeValueInAsValue(OrigLI, OrigVNI,
1506                                     StackIntvl.getValNumInfo(0));
1507
1508    // Insert hoisted spills.
1509    for (auto const Insert : SpillsToIns) {
1510      MachineBasicBlock *BB = Insert.first;
1511      unsigned LiveReg = Insert.second;
1512      MachineBasicBlock::iterator MI = IPA.getLastInsertPointIter(OrigLI, *BB);
1513      TII.storeRegToStackSlot(*BB, MI, LiveReg, false, Slot,
1514                              MRI.getRegClass(LiveReg), &TRI);
1515      LIS.InsertMachineInstrRangeInMaps(std::prev(MI), MI);
1516      ++NumSpills;
1517    }
1518
1519    // Remove redundant spills or change them to dead instructions.
1520    NumSpills -= SpillsToRm.size();
1521    for (auto const RMEnt : SpillsToRm) {
1522      RMEnt->setDesc(TII.get(TargetOpcode::KILL));
1523      for (unsigned i = RMEnt->getNumOperands(); i; --i) {
1524        MachineOperand &MO = RMEnt->getOperand(i - 1);
1525        if (MO.isReg() && MO.isImplicit() && MO.isDef() && !MO.isDead())
1526          RMEnt->RemoveOperand(i - 1);
1527      }
1528    }
1529    Edit.eliminateDeadDefs(SpillsToRm, None, AA);
1530  }
1531}
1532
1533/// For VirtReg clone, the \p New register should have the same physreg or
1534/// stackslot as the \p old register.
1535void HoistSpillHelper::LRE_DidCloneVirtReg(unsigned New, unsigned Old) {
1536  if (VRM.hasPhys(Old))
1537    VRM.assignVirt2Phys(New, VRM.getPhys(Old));
1538  else if (VRM.getStackSlot(Old) != VirtRegMap::NO_STACK_SLOT)
1539    VRM.assignVirt2StackSlot(New, VRM.getStackSlot(Old));
1540  else
1541    llvm_unreachable("VReg should be assigned either physreg or stackslot");
1542}
1543