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 "SplitKit.h"
15 #include "llvm/ADT/ArrayRef.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/MapVector.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/SetVector.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/CodeGen/LiveInterval.h"
25 #include "llvm/CodeGen/LiveIntervals.h"
26 #include "llvm/CodeGen/LiveRangeEdit.h"
27 #include "llvm/CodeGen/LiveStacks.h"
28 #include "llvm/CodeGen/MachineBasicBlock.h"
29 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
30 #include "llvm/CodeGen/MachineDominators.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineFunctionPass.h"
33 #include "llvm/CodeGen/MachineInstr.h"
34 #include "llvm/CodeGen/MachineInstrBuilder.h"
35 #include "llvm/CodeGen/MachineInstrBundle.h"
36 #include "llvm/CodeGen/MachineLoopInfo.h"
37 #include "llvm/CodeGen/MachineOperand.h"
38 #include "llvm/CodeGen/MachineRegisterInfo.h"
39 #include "llvm/CodeGen/SlotIndexes.h"
40 #include "llvm/CodeGen/Spiller.h"
41 #include "llvm/CodeGen/StackMaps.h"
42 #include "llvm/CodeGen/TargetInstrInfo.h"
43 #include "llvm/CodeGen/TargetOpcodes.h"
44 #include "llvm/CodeGen/TargetRegisterInfo.h"
45 #include "llvm/CodeGen/TargetSubtargetInfo.h"
46 #include "llvm/CodeGen/VirtRegMap.h"
47 #include "llvm/Config/llvm-config.h"
48 #include "llvm/Support/BlockFrequency.h"
49 #include "llvm/Support/BranchProbability.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/Support/Compiler.h"
52 #include "llvm/Support/Debug.h"
53 #include "llvm/Support/ErrorHandling.h"
54 #include "llvm/Support/raw_ostream.h"
55 #include <cassert>
56 #include <iterator>
57 #include <tuple>
58 #include <utility>
59 #include <vector>
60
61 using namespace llvm;
62
63 #define DEBUG_TYPE "regalloc"
64
65 STATISTIC(NumSpilledRanges, "Number of spilled live ranges");
66 STATISTIC(NumSnippets, "Number of spilled snippets");
67 STATISTIC(NumSpills, "Number of spills inserted");
68 STATISTIC(NumSpillsRemoved, "Number of spills removed");
69 STATISTIC(NumReloads, "Number of reloads inserted");
70 STATISTIC(NumReloadsRemoved, "Number of reloads removed");
71 STATISTIC(NumFolded, "Number of folded stack accesses");
72 STATISTIC(NumFoldedLoads, "Number of folded loads");
73 STATISTIC(NumRemats, "Number of rematerialized defs for spilling");
74
75 static cl::opt<bool> DisableHoisting("disable-spill-hoist", cl::Hidden,
76 cl::desc("Disable inline spill hoisting"));
77 static cl::opt<bool>
78 RestrictStatepointRemat("restrict-statepoint-remat",
79 cl::init(false), cl::Hidden,
80 cl::desc("Restrict remat for statepoint operands"));
81
82 namespace {
83
84 class HoistSpillHelper : private LiveRangeEdit::Delegate {
85 MachineFunction &MF;
86 LiveIntervals &LIS;
87 LiveStacks &LSS;
88 MachineDominatorTree &MDT;
89 MachineLoopInfo &Loops;
90 VirtRegMap &VRM;
91 MachineRegisterInfo &MRI;
92 const TargetInstrInfo &TII;
93 const TargetRegisterInfo &TRI;
94 const MachineBlockFrequencyInfo &MBFI;
95
96 InsertPointAnalysis IPA;
97
98 // Map from StackSlot to the LiveInterval of the original register.
99 // Note the LiveInterval of the original register may have been deleted
100 // after it is spilled. We keep a copy here to track the range where
101 // spills can be moved.
102 DenseMap<int, std::unique_ptr<LiveInterval>> StackSlotToOrigLI;
103
104 // Map from pair of (StackSlot and Original VNI) to a set of spills which
105 // have the same stackslot and have equal values defined by Original VNI.
106 // These spills are mergeable and are hoist candidates.
107 using MergeableSpillsMap =
108 MapVector<std::pair<int, VNInfo *>, SmallPtrSet<MachineInstr *, 16>>;
109 MergeableSpillsMap MergeableSpills;
110
111 /// This is the map from original register to a set containing all its
112 /// siblings. To hoist a spill to another BB, we need to find out a live
113 /// sibling there and use it as the source of the new spill.
114 DenseMap<Register, SmallSetVector<Register, 16>> Virt2SiblingsMap;
115
116 bool isSpillCandBB(LiveInterval &OrigLI, VNInfo &OrigVNI,
117 MachineBasicBlock &BB, Register &LiveReg);
118
119 void rmRedundantSpills(
120 SmallPtrSet<MachineInstr *, 16> &Spills,
121 SmallVectorImpl<MachineInstr *> &SpillsToRm,
122 DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill);
123
124 void getVisitOrders(
125 MachineBasicBlock *Root, SmallPtrSet<MachineInstr *, 16> &Spills,
126 SmallVectorImpl<MachineDomTreeNode *> &Orders,
127 SmallVectorImpl<MachineInstr *> &SpillsToRm,
128 DenseMap<MachineDomTreeNode *, unsigned> &SpillsToKeep,
129 DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill);
130
131 void runHoistSpills(LiveInterval &OrigLI, VNInfo &OrigVNI,
132 SmallPtrSet<MachineInstr *, 16> &Spills,
133 SmallVectorImpl<MachineInstr *> &SpillsToRm,
134 DenseMap<MachineBasicBlock *, unsigned> &SpillsToIns);
135
136 public:
HoistSpillHelper(MachineFunctionPass & pass,MachineFunction & mf,VirtRegMap & vrm)137 HoistSpillHelper(MachineFunctionPass &pass, MachineFunction &mf,
138 VirtRegMap &vrm)
139 : MF(mf), LIS(pass.getAnalysis<LiveIntervals>()),
140 LSS(pass.getAnalysis<LiveStacks>()),
141 MDT(pass.getAnalysis<MachineDominatorTree>()),
142 Loops(pass.getAnalysis<MachineLoopInfo>()), VRM(vrm),
143 MRI(mf.getRegInfo()), TII(*mf.getSubtarget().getInstrInfo()),
144 TRI(*mf.getSubtarget().getRegisterInfo()),
145 MBFI(pass.getAnalysis<MachineBlockFrequencyInfo>()),
146 IPA(LIS, mf.getNumBlockIDs()) {}
147
148 void addToMergeableSpills(MachineInstr &Spill, int StackSlot,
149 unsigned Original);
150 bool rmFromMergeableSpills(MachineInstr &Spill, int StackSlot);
151 void hoistAllSpills();
152 void LRE_DidCloneVirtReg(Register, Register) override;
153 };
154
155 class InlineSpiller : public Spiller {
156 MachineFunction &MF;
157 LiveIntervals &LIS;
158 LiveStacks &LSS;
159 MachineDominatorTree &MDT;
160 MachineLoopInfo &Loops;
161 VirtRegMap &VRM;
162 MachineRegisterInfo &MRI;
163 const TargetInstrInfo &TII;
164 const TargetRegisterInfo &TRI;
165 const MachineBlockFrequencyInfo &MBFI;
166
167 // Variables that are valid during spill(), but used by multiple methods.
168 LiveRangeEdit *Edit;
169 LiveInterval *StackInt;
170 int StackSlot;
171 Register Original;
172
173 // All registers to spill to StackSlot, including the main register.
174 SmallVector<Register, 8> RegsToSpill;
175
176 // All COPY instructions to/from snippets.
177 // They are ignored since both operands refer to the same stack slot.
178 SmallPtrSet<MachineInstr*, 8> SnippetCopies;
179
180 // Values that failed to remat at some point.
181 SmallPtrSet<VNInfo*, 8> UsedValues;
182
183 // Dead defs generated during spilling.
184 SmallVector<MachineInstr*, 8> DeadDefs;
185
186 // Object records spills information and does the hoisting.
187 HoistSpillHelper HSpiller;
188
189 // Live range weight calculator.
190 VirtRegAuxInfo &VRAI;
191
192 ~InlineSpiller() override = default;
193
194 public:
InlineSpiller(MachineFunctionPass & Pass,MachineFunction & MF,VirtRegMap & VRM,VirtRegAuxInfo & VRAI)195 InlineSpiller(MachineFunctionPass &Pass, MachineFunction &MF, VirtRegMap &VRM,
196 VirtRegAuxInfo &VRAI)
197 : MF(MF), LIS(Pass.getAnalysis<LiveIntervals>()),
198 LSS(Pass.getAnalysis<LiveStacks>()),
199 MDT(Pass.getAnalysis<MachineDominatorTree>()),
200 Loops(Pass.getAnalysis<MachineLoopInfo>()), VRM(VRM),
201 MRI(MF.getRegInfo()), TII(*MF.getSubtarget().getInstrInfo()),
202 TRI(*MF.getSubtarget().getRegisterInfo()),
203 MBFI(Pass.getAnalysis<MachineBlockFrequencyInfo>()),
204 HSpiller(Pass, MF, VRM), VRAI(VRAI) {}
205
206 void spill(LiveRangeEdit &) override;
207 void postOptimization() override;
208
209 private:
210 bool isSnippet(const LiveInterval &SnipLI);
211 void collectRegsToSpill();
212
isRegToSpill(Register Reg)213 bool isRegToSpill(Register Reg) { return is_contained(RegsToSpill, Reg); }
214
215 bool isSibling(Register Reg);
216 bool hoistSpillInsideBB(LiveInterval &SpillLI, MachineInstr &CopyMI);
217 void eliminateRedundantSpills(LiveInterval &LI, VNInfo *VNI);
218
219 void markValueUsed(LiveInterval*, VNInfo*);
220 bool canGuaranteeAssignmentAfterRemat(Register VReg, MachineInstr &MI);
221 bool reMaterializeFor(LiveInterval &, MachineInstr &MI);
222 void reMaterializeAll();
223
224 bool coalesceStackAccess(MachineInstr *MI, Register Reg);
225 bool foldMemoryOperand(ArrayRef<std::pair<MachineInstr *, unsigned>>,
226 MachineInstr *LoadMI = nullptr);
227 void insertReload(Register VReg, SlotIndex, MachineBasicBlock::iterator MI);
228 void insertSpill(Register VReg, bool isKill, MachineBasicBlock::iterator MI);
229
230 void spillAroundUses(Register Reg);
231 void spillAll();
232 };
233
234 } // end anonymous namespace
235
236 Spiller::~Spiller() = default;
237
anchor()238 void Spiller::anchor() {}
239
createInlineSpiller(MachineFunctionPass & Pass,MachineFunction & MF,VirtRegMap & VRM,VirtRegAuxInfo & VRAI)240 Spiller *llvm::createInlineSpiller(MachineFunctionPass &Pass,
241 MachineFunction &MF, VirtRegMap &VRM,
242 VirtRegAuxInfo &VRAI) {
243 return new InlineSpiller(Pass, MF, VRM, VRAI);
244 }
245
246 //===----------------------------------------------------------------------===//
247 // Snippets
248 //===----------------------------------------------------------------------===//
249
250 // When spilling a virtual register, we also spill any snippets it is connected
251 // to. The snippets are small live ranges that only have a single real use,
252 // leftovers from live range splitting. Spilling them enables memory operand
253 // folding or tightens the live range around the single use.
254 //
255 // This minimizes register pressure and maximizes the store-to-load distance for
256 // spill slots which can be important in tight loops.
257
258 /// isFullCopyOf - If MI is a COPY to or from Reg, return the other register,
259 /// otherwise return 0.
isFullCopyOf(const MachineInstr & MI,Register Reg)260 static Register isFullCopyOf(const MachineInstr &MI, Register Reg) {
261 if (!MI.isFullCopy())
262 return Register();
263 if (MI.getOperand(0).getReg() == Reg)
264 return MI.getOperand(1).getReg();
265 if (MI.getOperand(1).getReg() == Reg)
266 return MI.getOperand(0).getReg();
267 return Register();
268 }
269
getVDefInterval(const MachineInstr & MI,LiveIntervals & LIS)270 static void getVDefInterval(const MachineInstr &MI, LiveIntervals &LIS) {
271 for (const MachineOperand &MO : MI.operands())
272 if (MO.isReg() && MO.isDef() && MO.getReg().isVirtual())
273 LIS.getInterval(MO.getReg());
274 }
275
276 /// isSnippet - Identify if a live interval is a snippet that should be spilled.
277 /// It is assumed that SnipLI is a virtual register with the same original as
278 /// Edit->getReg().
isSnippet(const LiveInterval & SnipLI)279 bool InlineSpiller::isSnippet(const LiveInterval &SnipLI) {
280 Register Reg = Edit->getReg();
281
282 // A snippet is a tiny live range with only a single instruction using it
283 // besides copies to/from Reg or spills/fills.
284 // Exception is done for statepoint instructions which will fold fills
285 // into their operands.
286 // We accept:
287 //
288 // %snip = COPY %Reg / FILL fi#
289 // %snip = USE %snip
290 // %snip = STATEPOINT %snip in var arg area
291 // %Reg = COPY %snip / SPILL %snip, fi#
292 //
293 if (!LIS.intervalIsInOneMBB(SnipLI))
294 return false;
295
296 // Number of defs should not exceed 2 not accounting defs coming from
297 // statepoint instructions.
298 unsigned NumValNums = SnipLI.getNumValNums();
299 for (auto *VNI : SnipLI.vnis()) {
300 MachineInstr *MI = LIS.getInstructionFromIndex(VNI->def);
301 if (MI->getOpcode() == TargetOpcode::STATEPOINT)
302 --NumValNums;
303 }
304 if (NumValNums > 2)
305 return false;
306
307 MachineInstr *UseMI = nullptr;
308
309 // Check that all uses satisfy our criteria.
310 for (MachineRegisterInfo::reg_instr_nodbg_iterator
311 RI = MRI.reg_instr_nodbg_begin(SnipLI.reg()),
312 E = MRI.reg_instr_nodbg_end();
313 RI != E;) {
314 MachineInstr &MI = *RI++;
315
316 // Allow copies to/from Reg.
317 if (isFullCopyOf(MI, Reg))
318 continue;
319
320 // Allow stack slot loads.
321 int FI;
322 if (SnipLI.reg() == TII.isLoadFromStackSlot(MI, FI) && FI == StackSlot)
323 continue;
324
325 // Allow stack slot stores.
326 if (SnipLI.reg() == TII.isStoreToStackSlot(MI, FI) && FI == StackSlot)
327 continue;
328
329 if (StatepointOpers::isFoldableReg(&MI, SnipLI.reg()))
330 continue;
331
332 // Allow a single additional instruction.
333 if (UseMI && &MI != UseMI)
334 return false;
335 UseMI = &MI;
336 }
337 return true;
338 }
339
340 /// collectRegsToSpill - Collect live range snippets that only have a single
341 /// real use.
collectRegsToSpill()342 void InlineSpiller::collectRegsToSpill() {
343 Register Reg = Edit->getReg();
344
345 // Main register always spills.
346 RegsToSpill.assign(1, Reg);
347 SnippetCopies.clear();
348
349 // Snippets all have the same original, so there can't be any for an original
350 // register.
351 if (Original == Reg)
352 return;
353
354 for (MachineInstr &MI :
355 llvm::make_early_inc_range(MRI.reg_instructions(Reg))) {
356 Register SnipReg = isFullCopyOf(MI, Reg);
357 if (!isSibling(SnipReg))
358 continue;
359 LiveInterval &SnipLI = LIS.getInterval(SnipReg);
360 if (!isSnippet(SnipLI))
361 continue;
362 SnippetCopies.insert(&MI);
363 if (isRegToSpill(SnipReg))
364 continue;
365 RegsToSpill.push_back(SnipReg);
366 LLVM_DEBUG(dbgs() << "\talso spill snippet " << SnipLI << '\n');
367 ++NumSnippets;
368 }
369 }
370
isSibling(Register Reg)371 bool InlineSpiller::isSibling(Register Reg) {
372 return Reg.isVirtual() && VRM.getOriginal(Reg) == Original;
373 }
374
375 /// It is beneficial to spill to earlier place in the same BB in case
376 /// as follows:
377 /// There is an alternative def earlier in the same MBB.
378 /// Hoist the spill as far as possible in SpillMBB. This can ease
379 /// register pressure:
380 ///
381 /// x = def
382 /// y = use x
383 /// s = copy x
384 ///
385 /// Hoisting the spill of s to immediately after the def removes the
386 /// interference between x and y:
387 ///
388 /// x = def
389 /// spill x
390 /// y = use killed x
391 ///
392 /// This hoist only helps when the copy kills its source.
393 ///
hoistSpillInsideBB(LiveInterval & SpillLI,MachineInstr & CopyMI)394 bool InlineSpiller::hoistSpillInsideBB(LiveInterval &SpillLI,
395 MachineInstr &CopyMI) {
396 SlotIndex Idx = LIS.getInstructionIndex(CopyMI);
397 #ifndef NDEBUG
398 VNInfo *VNI = SpillLI.getVNInfoAt(Idx.getRegSlot());
399 assert(VNI && VNI->def == Idx.getRegSlot() && "Not defined by copy");
400 #endif
401
402 Register SrcReg = CopyMI.getOperand(1).getReg();
403 LiveInterval &SrcLI = LIS.getInterval(SrcReg);
404 VNInfo *SrcVNI = SrcLI.getVNInfoAt(Idx);
405 LiveQueryResult SrcQ = SrcLI.Query(Idx);
406 MachineBasicBlock *DefMBB = LIS.getMBBFromIndex(SrcVNI->def);
407 if (DefMBB != CopyMI.getParent() || !SrcQ.isKill())
408 return false;
409
410 // Conservatively extend the stack slot range to the range of the original
411 // value. We may be able to do better with stack slot coloring by being more
412 // careful here.
413 assert(StackInt && "No stack slot assigned yet.");
414 LiveInterval &OrigLI = LIS.getInterval(Original);
415 VNInfo *OrigVNI = OrigLI.getVNInfoAt(Idx);
416 StackInt->MergeValueInAsValue(OrigLI, OrigVNI, StackInt->getValNumInfo(0));
417 LLVM_DEBUG(dbgs() << "\tmerged orig valno " << OrigVNI->id << ": "
418 << *StackInt << '\n');
419
420 // We are going to spill SrcVNI immediately after its def, so clear out
421 // any later spills of the same value.
422 eliminateRedundantSpills(SrcLI, SrcVNI);
423
424 MachineBasicBlock *MBB = LIS.getMBBFromIndex(SrcVNI->def);
425 MachineBasicBlock::iterator MII;
426 if (SrcVNI->isPHIDef())
427 MII = MBB->SkipPHIsLabelsAndDebug(MBB->begin());
428 else {
429 MachineInstr *DefMI = LIS.getInstructionFromIndex(SrcVNI->def);
430 assert(DefMI && "Defining instruction disappeared");
431 MII = DefMI;
432 ++MII;
433 }
434 MachineInstrSpan MIS(MII, MBB);
435 // Insert spill without kill flag immediately after def.
436 TII.storeRegToStackSlot(*MBB, MII, SrcReg, false, StackSlot,
437 MRI.getRegClass(SrcReg), &TRI, Register());
438 LIS.InsertMachineInstrRangeInMaps(MIS.begin(), MII);
439 for (const MachineInstr &MI : make_range(MIS.begin(), MII))
440 getVDefInterval(MI, LIS);
441 --MII; // Point to store instruction.
442 LLVM_DEBUG(dbgs() << "\thoisted: " << SrcVNI->def << '\t' << *MII);
443
444 // If there is only 1 store instruction is required for spill, add it
445 // to mergeable list. In X86 AMX, 2 intructions are required to store.
446 // We disable the merge for this case.
447 if (MIS.begin() == MII)
448 HSpiller.addToMergeableSpills(*MII, StackSlot, Original);
449 ++NumSpills;
450 return true;
451 }
452
453 /// eliminateRedundantSpills - SLI:VNI is known to be on the stack. Remove any
454 /// redundant spills of this value in SLI.reg and sibling copies.
eliminateRedundantSpills(LiveInterval & SLI,VNInfo * VNI)455 void InlineSpiller::eliminateRedundantSpills(LiveInterval &SLI, VNInfo *VNI) {
456 assert(VNI && "Missing value");
457 SmallVector<std::pair<LiveInterval*, VNInfo*>, 8> WorkList;
458 WorkList.push_back(std::make_pair(&SLI, VNI));
459 assert(StackInt && "No stack slot assigned yet.");
460
461 do {
462 LiveInterval *LI;
463 std::tie(LI, VNI) = WorkList.pop_back_val();
464 Register Reg = LI->reg();
465 LLVM_DEBUG(dbgs() << "Checking redundant spills for " << VNI->id << '@'
466 << VNI->def << " in " << *LI << '\n');
467
468 // Regs to spill are taken care of.
469 if (isRegToSpill(Reg))
470 continue;
471
472 // Add all of VNI's live range to StackInt.
473 StackInt->MergeValueInAsValue(*LI, VNI, StackInt->getValNumInfo(0));
474 LLVM_DEBUG(dbgs() << "Merged to stack int: " << *StackInt << '\n');
475
476 // Find all spills and copies of VNI.
477 for (MachineInstr &MI :
478 llvm::make_early_inc_range(MRI.use_nodbg_instructions(Reg))) {
479 if (!MI.isCopy() && !MI.mayStore())
480 continue;
481 SlotIndex Idx = LIS.getInstructionIndex(MI);
482 if (LI->getVNInfoAt(Idx) != VNI)
483 continue;
484
485 // Follow sibling copies down the dominator tree.
486 if (Register DstReg = isFullCopyOf(MI, Reg)) {
487 if (isSibling(DstReg)) {
488 LiveInterval &DstLI = LIS.getInterval(DstReg);
489 VNInfo *DstVNI = DstLI.getVNInfoAt(Idx.getRegSlot());
490 assert(DstVNI && "Missing defined value");
491 assert(DstVNI->def == Idx.getRegSlot() && "Wrong copy def slot");
492 WorkList.push_back(std::make_pair(&DstLI, DstVNI));
493 }
494 continue;
495 }
496
497 // Erase spills.
498 int FI;
499 if (Reg == TII.isStoreToStackSlot(MI, FI) && FI == StackSlot) {
500 LLVM_DEBUG(dbgs() << "Redundant spill " << Idx << '\t' << MI);
501 // eliminateDeadDefs won't normally remove stores, so switch opcode.
502 MI.setDesc(TII.get(TargetOpcode::KILL));
503 DeadDefs.push_back(&MI);
504 ++NumSpillsRemoved;
505 if (HSpiller.rmFromMergeableSpills(MI, StackSlot))
506 --NumSpills;
507 }
508 }
509 } while (!WorkList.empty());
510 }
511
512 //===----------------------------------------------------------------------===//
513 // Rematerialization
514 //===----------------------------------------------------------------------===//
515
516 /// markValueUsed - Remember that VNI failed to rematerialize, so its defining
517 /// instruction cannot be eliminated. See through snippet copies
markValueUsed(LiveInterval * LI,VNInfo * VNI)518 void InlineSpiller::markValueUsed(LiveInterval *LI, VNInfo *VNI) {
519 SmallVector<std::pair<LiveInterval*, VNInfo*>, 8> WorkList;
520 WorkList.push_back(std::make_pair(LI, VNI));
521 do {
522 std::tie(LI, VNI) = WorkList.pop_back_val();
523 if (!UsedValues.insert(VNI).second)
524 continue;
525
526 if (VNI->isPHIDef()) {
527 MachineBasicBlock *MBB = LIS.getMBBFromIndex(VNI->def);
528 for (MachineBasicBlock *P : MBB->predecessors()) {
529 VNInfo *PVNI = LI->getVNInfoBefore(LIS.getMBBEndIdx(P));
530 if (PVNI)
531 WorkList.push_back(std::make_pair(LI, PVNI));
532 }
533 continue;
534 }
535
536 // Follow snippet copies.
537 MachineInstr *MI = LIS.getInstructionFromIndex(VNI->def);
538 if (!SnippetCopies.count(MI))
539 continue;
540 LiveInterval &SnipLI = LIS.getInterval(MI->getOperand(1).getReg());
541 assert(isRegToSpill(SnipLI.reg()) && "Unexpected register in copy");
542 VNInfo *SnipVNI = SnipLI.getVNInfoAt(VNI->def.getRegSlot(true));
543 assert(SnipVNI && "Snippet undefined before copy");
544 WorkList.push_back(std::make_pair(&SnipLI, SnipVNI));
545 } while (!WorkList.empty());
546 }
547
canGuaranteeAssignmentAfterRemat(Register VReg,MachineInstr & MI)548 bool InlineSpiller::canGuaranteeAssignmentAfterRemat(Register VReg,
549 MachineInstr &MI) {
550 if (!RestrictStatepointRemat)
551 return true;
552 // Here's a quick explanation of the problem we're trying to handle here:
553 // * There are some pseudo instructions with more vreg uses than there are
554 // physical registers on the machine.
555 // * This is normally handled by spilling the vreg, and folding the reload
556 // into the user instruction. (Thus decreasing the number of used vregs
557 // until the remainder can be assigned to physregs.)
558 // * However, since we may try to spill vregs in any order, we can end up
559 // trying to spill each operand to the instruction, and then rematting it
560 // instead. When that happens, the new live intervals (for the remats) are
561 // expected to be trivially assignable (i.e. RS_Done). However, since we
562 // may have more remats than physregs, we're guaranteed to fail to assign
563 // one.
564 // At the moment, we only handle this for STATEPOINTs since they're the only
565 // pseudo op where we've seen this. If we start seeing other instructions
566 // with the same problem, we need to revisit this.
567 if (MI.getOpcode() != TargetOpcode::STATEPOINT)
568 return true;
569 // For STATEPOINTs we allow re-materialization for fixed arguments only hoping
570 // that number of physical registers is enough to cover all fixed arguments.
571 // If it is not true we need to revisit it.
572 for (unsigned Idx = StatepointOpers(&MI).getVarIdx(),
573 EndIdx = MI.getNumOperands();
574 Idx < EndIdx; ++Idx) {
575 MachineOperand &MO = MI.getOperand(Idx);
576 if (MO.isReg() && MO.getReg() == VReg)
577 return false;
578 }
579 return true;
580 }
581
582 /// reMaterializeFor - Attempt to rematerialize before MI instead of reloading.
reMaterializeFor(LiveInterval & VirtReg,MachineInstr & MI)583 bool InlineSpiller::reMaterializeFor(LiveInterval &VirtReg, MachineInstr &MI) {
584 // Analyze instruction
585 SmallVector<std::pair<MachineInstr *, unsigned>, 8> Ops;
586 VirtRegInfo RI = AnalyzeVirtRegInBundle(MI, VirtReg.reg(), &Ops);
587
588 if (!RI.Reads)
589 return false;
590
591 SlotIndex UseIdx = LIS.getInstructionIndex(MI).getRegSlot(true);
592 VNInfo *ParentVNI = VirtReg.getVNInfoAt(UseIdx.getBaseIndex());
593
594 if (!ParentVNI) {
595 LLVM_DEBUG(dbgs() << "\tadding <undef> flags: ");
596 for (MachineOperand &MO : MI.operands())
597 if (MO.isReg() && MO.isUse() && MO.getReg() == VirtReg.reg())
598 MO.setIsUndef();
599 LLVM_DEBUG(dbgs() << UseIdx << '\t' << MI);
600 return true;
601 }
602
603 if (SnippetCopies.count(&MI))
604 return false;
605
606 LiveInterval &OrigLI = LIS.getInterval(Original);
607 VNInfo *OrigVNI = OrigLI.getVNInfoAt(UseIdx);
608 LiveRangeEdit::Remat RM(ParentVNI);
609 RM.OrigMI = LIS.getInstructionFromIndex(OrigVNI->def);
610
611 if (!Edit->canRematerializeAt(RM, OrigVNI, UseIdx, false)) {
612 markValueUsed(&VirtReg, ParentVNI);
613 LLVM_DEBUG(dbgs() << "\tcannot remat for " << UseIdx << '\t' << MI);
614 return false;
615 }
616
617 // If the instruction also writes VirtReg.reg, it had better not require the
618 // same register for uses and defs.
619 if (RI.Tied) {
620 markValueUsed(&VirtReg, ParentVNI);
621 LLVM_DEBUG(dbgs() << "\tcannot remat tied reg: " << UseIdx << '\t' << MI);
622 return false;
623 }
624
625 // Before rematerializing into a register for a single instruction, try to
626 // fold a load into the instruction. That avoids allocating a new register.
627 if (RM.OrigMI->canFoldAsLoad() &&
628 foldMemoryOperand(Ops, RM.OrigMI)) {
629 Edit->markRematerialized(RM.ParentVNI);
630 ++NumFoldedLoads;
631 return true;
632 }
633
634 // If we can't guarantee that we'll be able to actually assign the new vreg,
635 // we can't remat.
636 if (!canGuaranteeAssignmentAfterRemat(VirtReg.reg(), MI)) {
637 markValueUsed(&VirtReg, ParentVNI);
638 LLVM_DEBUG(dbgs() << "\tcannot remat for " << UseIdx << '\t' << MI);
639 return false;
640 }
641
642 // Allocate a new register for the remat.
643 Register NewVReg = Edit->createFrom(Original);
644
645 // Finally we can rematerialize OrigMI before MI.
646 SlotIndex DefIdx =
647 Edit->rematerializeAt(*MI.getParent(), MI, NewVReg, RM, TRI);
648
649 // We take the DebugLoc from MI, since OrigMI may be attributed to a
650 // different source location.
651 auto *NewMI = LIS.getInstructionFromIndex(DefIdx);
652 NewMI->setDebugLoc(MI.getDebugLoc());
653
654 (void)DefIdx;
655 LLVM_DEBUG(dbgs() << "\tremat: " << DefIdx << '\t'
656 << *LIS.getInstructionFromIndex(DefIdx));
657
658 // Replace operands
659 for (const auto &OpPair : Ops) {
660 MachineOperand &MO = OpPair.first->getOperand(OpPair.second);
661 if (MO.isReg() && MO.isUse() && MO.getReg() == VirtReg.reg()) {
662 MO.setReg(NewVReg);
663 MO.setIsKill();
664 }
665 }
666 LLVM_DEBUG(dbgs() << "\t " << UseIdx << '\t' << MI << '\n');
667
668 ++NumRemats;
669 return true;
670 }
671
672 /// reMaterializeAll - Try to rematerialize as many uses as possible,
673 /// and trim the live ranges after.
reMaterializeAll()674 void InlineSpiller::reMaterializeAll() {
675 if (!Edit->anyRematerializable())
676 return;
677
678 UsedValues.clear();
679
680 // Try to remat before all uses of snippets.
681 bool anyRemat = false;
682 for (Register Reg : RegsToSpill) {
683 LiveInterval &LI = LIS.getInterval(Reg);
684 for (MachineInstr &MI : llvm::make_early_inc_range(MRI.reg_bundles(Reg))) {
685 // Debug values are not allowed to affect codegen.
686 if (MI.isDebugValue())
687 continue;
688
689 assert(!MI.isDebugInstr() && "Did not expect to find a use in debug "
690 "instruction that isn't a DBG_VALUE");
691
692 anyRemat |= reMaterializeFor(LI, MI);
693 }
694 }
695 if (!anyRemat)
696 return;
697
698 // Remove any values that were completely rematted.
699 for (Register Reg : RegsToSpill) {
700 LiveInterval &LI = LIS.getInterval(Reg);
701 for (VNInfo *VNI : LI.vnis()) {
702 if (VNI->isUnused() || VNI->isPHIDef() || UsedValues.count(VNI))
703 continue;
704 MachineInstr *MI = LIS.getInstructionFromIndex(VNI->def);
705 MI->addRegisterDead(Reg, &TRI);
706 if (!MI->allDefsAreDead())
707 continue;
708 LLVM_DEBUG(dbgs() << "All defs dead: " << *MI);
709 DeadDefs.push_back(MI);
710 }
711 }
712
713 // Eliminate dead code after remat. Note that some snippet copies may be
714 // deleted here.
715 if (DeadDefs.empty())
716 return;
717 LLVM_DEBUG(dbgs() << "Remat created " << DeadDefs.size() << " dead defs.\n");
718 Edit->eliminateDeadDefs(DeadDefs, RegsToSpill);
719
720 // LiveRangeEdit::eliminateDeadDef is used to remove dead define instructions
721 // after rematerialization. To remove a VNI for a vreg from its LiveInterval,
722 // LiveIntervals::removeVRegDefAt is used. However, after non-PHI VNIs are all
723 // removed, PHI VNI are still left in the LiveInterval.
724 // So to get rid of unused reg, we need to check whether it has non-dbg
725 // reference instead of whether it has non-empty interval.
726 unsigned ResultPos = 0;
727 for (Register Reg : RegsToSpill) {
728 if (MRI.reg_nodbg_empty(Reg)) {
729 Edit->eraseVirtReg(Reg);
730 continue;
731 }
732
733 assert(LIS.hasInterval(Reg) &&
734 (!LIS.getInterval(Reg).empty() || !MRI.reg_nodbg_empty(Reg)) &&
735 "Empty and not used live-range?!");
736
737 RegsToSpill[ResultPos++] = Reg;
738 }
739 RegsToSpill.erase(RegsToSpill.begin() + ResultPos, RegsToSpill.end());
740 LLVM_DEBUG(dbgs() << RegsToSpill.size()
741 << " registers to spill after remat.\n");
742 }
743
744 //===----------------------------------------------------------------------===//
745 // Spilling
746 //===----------------------------------------------------------------------===//
747
748 /// If MI is a load or store of StackSlot, it can be removed.
coalesceStackAccess(MachineInstr * MI,Register Reg)749 bool InlineSpiller::coalesceStackAccess(MachineInstr *MI, Register Reg) {
750 int FI = 0;
751 Register InstrReg = TII.isLoadFromStackSlot(*MI, FI);
752 bool IsLoad = InstrReg;
753 if (!IsLoad)
754 InstrReg = TII.isStoreToStackSlot(*MI, FI);
755
756 // We have a stack access. Is it the right register and slot?
757 if (InstrReg != Reg || FI != StackSlot)
758 return false;
759
760 if (!IsLoad)
761 HSpiller.rmFromMergeableSpills(*MI, StackSlot);
762
763 LLVM_DEBUG(dbgs() << "Coalescing stack access: " << *MI);
764 LIS.RemoveMachineInstrFromMaps(*MI);
765 MI->eraseFromParent();
766
767 if (IsLoad) {
768 ++NumReloadsRemoved;
769 --NumReloads;
770 } else {
771 ++NumSpillsRemoved;
772 --NumSpills;
773 }
774
775 return true;
776 }
777
778 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
779 LLVM_DUMP_METHOD
780 // Dump the range of instructions from B to E with their slot indexes.
dumpMachineInstrRangeWithSlotIndex(MachineBasicBlock::iterator B,MachineBasicBlock::iterator E,LiveIntervals const & LIS,const char * const header,Register VReg=Register ())781 static void dumpMachineInstrRangeWithSlotIndex(MachineBasicBlock::iterator B,
782 MachineBasicBlock::iterator E,
783 LiveIntervals const &LIS,
784 const char *const header,
785 Register VReg = Register()) {
786 char NextLine = '\n';
787 char SlotIndent = '\t';
788
789 if (std::next(B) == E) {
790 NextLine = ' ';
791 SlotIndent = ' ';
792 }
793
794 dbgs() << '\t' << header << ": " << NextLine;
795
796 for (MachineBasicBlock::iterator I = B; I != E; ++I) {
797 SlotIndex Idx = LIS.getInstructionIndex(*I).getRegSlot();
798
799 // If a register was passed in and this instruction has it as a
800 // destination that is marked as an early clobber, print the
801 // early-clobber slot index.
802 if (VReg) {
803 MachineOperand *MO = I->findRegisterDefOperand(VReg);
804 if (MO && MO->isEarlyClobber())
805 Idx = Idx.getRegSlot(true);
806 }
807
808 dbgs() << SlotIndent << Idx << '\t' << *I;
809 }
810 }
811 #endif
812
813 /// foldMemoryOperand - Try folding stack slot references in Ops into their
814 /// instructions.
815 ///
816 /// @param Ops Operand indices from AnalyzeVirtRegInBundle().
817 /// @param LoadMI Load instruction to use instead of stack slot when non-null.
818 /// @return True on success.
819 bool InlineSpiller::
foldMemoryOperand(ArrayRef<std::pair<MachineInstr *,unsigned>> Ops,MachineInstr * LoadMI)820 foldMemoryOperand(ArrayRef<std::pair<MachineInstr *, unsigned>> Ops,
821 MachineInstr *LoadMI) {
822 if (Ops.empty())
823 return false;
824 // Don't attempt folding in bundles.
825 MachineInstr *MI = Ops.front().first;
826 if (Ops.back().first != MI || MI->isBundled())
827 return false;
828
829 bool WasCopy = MI->isCopy();
830 Register ImpReg;
831
832 // TII::foldMemoryOperand will do what we need here for statepoint
833 // (fold load into use and remove corresponding def). We will replace
834 // uses of removed def with loads (spillAroundUses).
835 // For that to work we need to untie def and use to pass it through
836 // foldMemoryOperand and signal foldPatchpoint that it is allowed to
837 // fold them.
838 bool UntieRegs = MI->getOpcode() == TargetOpcode::STATEPOINT;
839
840 // Spill subregs if the target allows it.
841 // We always want to spill subregs for stackmap/patchpoint pseudos.
842 bool SpillSubRegs = TII.isSubregFoldable() ||
843 MI->getOpcode() == TargetOpcode::STATEPOINT ||
844 MI->getOpcode() == TargetOpcode::PATCHPOINT ||
845 MI->getOpcode() == TargetOpcode::STACKMAP;
846
847 // TargetInstrInfo::foldMemoryOperand only expects explicit, non-tied
848 // operands.
849 SmallVector<unsigned, 8> FoldOps;
850 for (const auto &OpPair : Ops) {
851 unsigned Idx = OpPair.second;
852 assert(MI == OpPair.first && "Instruction conflict during operand folding");
853 MachineOperand &MO = MI->getOperand(Idx);
854
855 // No point restoring an undef read, and we'll produce an invalid live
856 // interval.
857 // TODO: Is this really the correct way to handle undef tied uses?
858 if (MO.isUse() && !MO.readsReg() && !MO.isTied())
859 continue;
860
861 if (MO.isImplicit()) {
862 ImpReg = MO.getReg();
863 continue;
864 }
865
866 if (!SpillSubRegs && MO.getSubReg())
867 return false;
868 // We cannot fold a load instruction into a def.
869 if (LoadMI && MO.isDef())
870 return false;
871 // Tied use operands should not be passed to foldMemoryOperand.
872 if (UntieRegs || !MI->isRegTiedToDefOperand(Idx))
873 FoldOps.push_back(Idx);
874 }
875
876 // If we only have implicit uses, we won't be able to fold that.
877 // Moreover, TargetInstrInfo::foldMemoryOperand will assert if we try!
878 if (FoldOps.empty())
879 return false;
880
881 MachineInstrSpan MIS(MI, MI->getParent());
882
883 SmallVector<std::pair<unsigned, unsigned> > TiedOps;
884 if (UntieRegs)
885 for (unsigned Idx : FoldOps) {
886 MachineOperand &MO = MI->getOperand(Idx);
887 if (!MO.isTied())
888 continue;
889 unsigned Tied = MI->findTiedOperandIdx(Idx);
890 if (MO.isUse())
891 TiedOps.emplace_back(Tied, Idx);
892 else {
893 assert(MO.isDef() && "Tied to not use and def?");
894 TiedOps.emplace_back(Idx, Tied);
895 }
896 MI->untieRegOperand(Idx);
897 }
898
899 MachineInstr *FoldMI =
900 LoadMI ? TII.foldMemoryOperand(*MI, FoldOps, *LoadMI, &LIS)
901 : TII.foldMemoryOperand(*MI, FoldOps, StackSlot, &LIS, &VRM);
902 if (!FoldMI) {
903 // Re-tie operands.
904 for (auto Tied : TiedOps)
905 MI->tieOperands(Tied.first, Tied.second);
906 return false;
907 }
908
909 // Remove LIS for any dead defs in the original MI not in FoldMI.
910 for (MIBundleOperands MO(*MI); MO.isValid(); ++MO) {
911 if (!MO->isReg())
912 continue;
913 Register Reg = MO->getReg();
914 if (!Reg || Reg.isVirtual() || MRI.isReserved(Reg)) {
915 continue;
916 }
917 // Skip non-Defs, including undef uses and internal reads.
918 if (MO->isUse())
919 continue;
920 PhysRegInfo RI = AnalyzePhysRegInBundle(*FoldMI, Reg, &TRI);
921 if (RI.FullyDefined)
922 continue;
923 // FoldMI does not define this physreg. Remove the LI segment.
924 assert(MO->isDead() && "Cannot fold physreg def");
925 SlotIndex Idx = LIS.getInstructionIndex(*MI).getRegSlot();
926 LIS.removePhysRegDefAt(Reg.asMCReg(), Idx);
927 }
928
929 int FI;
930 if (TII.isStoreToStackSlot(*MI, FI) &&
931 HSpiller.rmFromMergeableSpills(*MI, FI))
932 --NumSpills;
933 LIS.ReplaceMachineInstrInMaps(*MI, *FoldMI);
934 // Update the call site info.
935 if (MI->isCandidateForCallSiteEntry())
936 MI->getMF()->moveCallSiteInfo(MI, FoldMI);
937
938 // If we've folded a store into an instruction labelled with debug-info,
939 // record a substitution from the old operand to the memory operand. Handle
940 // the simple common case where operand 0 is the one being folded, plus when
941 // the destination operand is also a tied def. More values could be
942 // substituted / preserved with more analysis.
943 if (MI->peekDebugInstrNum() && Ops[0].second == 0) {
944 // Helper lambda.
945 auto MakeSubstitution = [this,FoldMI,MI,&Ops]() {
946 // Substitute old operand zero to the new instructions memory operand.
947 unsigned OldOperandNum = Ops[0].second;
948 unsigned NewNum = FoldMI->getDebugInstrNum();
949 unsigned OldNum = MI->getDebugInstrNum();
950 MF.makeDebugValueSubstitution({OldNum, OldOperandNum},
951 {NewNum, MachineFunction::DebugOperandMemNumber});
952 };
953
954 const MachineOperand &Op0 = MI->getOperand(Ops[0].second);
955 if (Ops.size() == 1 && Op0.isDef()) {
956 MakeSubstitution();
957 } else if (Ops.size() == 2 && Op0.isDef() && MI->getOperand(1).isTied() &&
958 Op0.getReg() == MI->getOperand(1).getReg()) {
959 MakeSubstitution();
960 }
961 } else if (MI->peekDebugInstrNum()) {
962 // This is a debug-labelled instruction, but the operand being folded isn't
963 // at operand zero. Most likely this means it's a load being folded in.
964 // Substitute any register defs from operand zero up to the one being
965 // folded -- past that point, we don't know what the new operand indexes
966 // will be.
967 MF.substituteDebugValuesForInst(*MI, *FoldMI, Ops[0].second);
968 }
969
970 MI->eraseFromParent();
971
972 // Insert any new instructions other than FoldMI into the LIS maps.
973 assert(!MIS.empty() && "Unexpected empty span of instructions!");
974 for (MachineInstr &MI : MIS)
975 if (&MI != FoldMI)
976 LIS.InsertMachineInstrInMaps(MI);
977
978 // TII.foldMemoryOperand may have left some implicit operands on the
979 // instruction. Strip them.
980 if (ImpReg)
981 for (unsigned i = FoldMI->getNumOperands(); i; --i) {
982 MachineOperand &MO = FoldMI->getOperand(i - 1);
983 if (!MO.isReg() || !MO.isImplicit())
984 break;
985 if (MO.getReg() == ImpReg)
986 FoldMI->removeOperand(i - 1);
987 }
988
989 LLVM_DEBUG(dumpMachineInstrRangeWithSlotIndex(MIS.begin(), MIS.end(), LIS,
990 "folded"));
991
992 if (!WasCopy)
993 ++NumFolded;
994 else if (Ops.front().second == 0) {
995 ++NumSpills;
996 // If there is only 1 store instruction is required for spill, add it
997 // to mergeable list. In X86 AMX, 2 intructions are required to store.
998 // We disable the merge for this case.
999 if (std::distance(MIS.begin(), MIS.end()) <= 1)
1000 HSpiller.addToMergeableSpills(*FoldMI, StackSlot, Original);
1001 } else
1002 ++NumReloads;
1003 return true;
1004 }
1005
insertReload(Register NewVReg,SlotIndex Idx,MachineBasicBlock::iterator MI)1006 void InlineSpiller::insertReload(Register NewVReg,
1007 SlotIndex Idx,
1008 MachineBasicBlock::iterator MI) {
1009 MachineBasicBlock &MBB = *MI->getParent();
1010
1011 MachineInstrSpan MIS(MI, &MBB);
1012 TII.loadRegFromStackSlot(MBB, MI, NewVReg, StackSlot,
1013 MRI.getRegClass(NewVReg), &TRI, Register());
1014
1015 LIS.InsertMachineInstrRangeInMaps(MIS.begin(), MI);
1016
1017 LLVM_DEBUG(dumpMachineInstrRangeWithSlotIndex(MIS.begin(), MI, LIS, "reload",
1018 NewVReg));
1019 ++NumReloads;
1020 }
1021
1022 /// Check if \p Def fully defines a VReg with an undefined value.
1023 /// If that's the case, that means the value of VReg is actually
1024 /// not relevant.
isRealSpill(const MachineInstr & Def)1025 static bool isRealSpill(const MachineInstr &Def) {
1026 if (!Def.isImplicitDef())
1027 return true;
1028 assert(Def.getNumOperands() == 1 &&
1029 "Implicit def with more than one definition");
1030 // We can say that the VReg defined by Def is undef, only if it is
1031 // fully defined by Def. Otherwise, some of the lanes may not be
1032 // undef and the value of the VReg matters.
1033 return Def.getOperand(0).getSubReg();
1034 }
1035
1036 /// insertSpill - Insert a spill of NewVReg after MI.
insertSpill(Register NewVReg,bool isKill,MachineBasicBlock::iterator MI)1037 void InlineSpiller::insertSpill(Register NewVReg, bool isKill,
1038 MachineBasicBlock::iterator MI) {
1039 // Spill are not terminators, so inserting spills after terminators will
1040 // violate invariants in MachineVerifier.
1041 assert(!MI->isTerminator() && "Inserting a spill after a terminator");
1042 MachineBasicBlock &MBB = *MI->getParent();
1043
1044 MachineInstrSpan MIS(MI, &MBB);
1045 MachineBasicBlock::iterator SpillBefore = std::next(MI);
1046 bool IsRealSpill = isRealSpill(*MI);
1047
1048 if (IsRealSpill)
1049 TII.storeRegToStackSlot(MBB, SpillBefore, NewVReg, isKill, StackSlot,
1050 MRI.getRegClass(NewVReg), &TRI, Register());
1051 else
1052 // Don't spill undef value.
1053 // Anything works for undef, in particular keeping the memory
1054 // uninitialized is a viable option and it saves code size and
1055 // run time.
1056 BuildMI(MBB, SpillBefore, MI->getDebugLoc(), TII.get(TargetOpcode::KILL))
1057 .addReg(NewVReg, getKillRegState(isKill));
1058
1059 MachineBasicBlock::iterator Spill = std::next(MI);
1060 LIS.InsertMachineInstrRangeInMaps(Spill, MIS.end());
1061 for (const MachineInstr &MI : make_range(Spill, MIS.end()))
1062 getVDefInterval(MI, LIS);
1063
1064 LLVM_DEBUG(
1065 dumpMachineInstrRangeWithSlotIndex(Spill, MIS.end(), LIS, "spill"));
1066 ++NumSpills;
1067 // If there is only 1 store instruction is required for spill, add it
1068 // to mergeable list. In X86 AMX, 2 intructions are required to store.
1069 // We disable the merge for this case.
1070 if (IsRealSpill && std::distance(Spill, MIS.end()) <= 1)
1071 HSpiller.addToMergeableSpills(*Spill, StackSlot, Original);
1072 }
1073
1074 /// spillAroundUses - insert spill code around each use of Reg.
spillAroundUses(Register Reg)1075 void InlineSpiller::spillAroundUses(Register Reg) {
1076 LLVM_DEBUG(dbgs() << "spillAroundUses " << printReg(Reg) << '\n');
1077 LiveInterval &OldLI = LIS.getInterval(Reg);
1078
1079 // Iterate over instructions using Reg.
1080 for (MachineInstr &MI : llvm::make_early_inc_range(MRI.reg_bundles(Reg))) {
1081 // Debug values are not allowed to affect codegen.
1082 if (MI.isDebugValue()) {
1083 // Modify DBG_VALUE now that the value is in a spill slot.
1084 MachineBasicBlock *MBB = MI.getParent();
1085 LLVM_DEBUG(dbgs() << "Modifying debug info due to spill:\t" << MI);
1086 buildDbgValueForSpill(*MBB, &MI, MI, StackSlot, Reg);
1087 MBB->erase(MI);
1088 continue;
1089 }
1090
1091 assert(!MI.isDebugInstr() && "Did not expect to find a use in debug "
1092 "instruction that isn't a DBG_VALUE");
1093
1094 // Ignore copies to/from snippets. We'll delete them.
1095 if (SnippetCopies.count(&MI))
1096 continue;
1097
1098 // Stack slot accesses may coalesce away.
1099 if (coalesceStackAccess(&MI, Reg))
1100 continue;
1101
1102 // Analyze instruction.
1103 SmallVector<std::pair<MachineInstr*, unsigned>, 8> Ops;
1104 VirtRegInfo RI = AnalyzeVirtRegInBundle(MI, Reg, &Ops);
1105
1106 // Find the slot index where this instruction reads and writes OldLI.
1107 // This is usually the def slot, except for tied early clobbers.
1108 SlotIndex Idx = LIS.getInstructionIndex(MI).getRegSlot();
1109 if (VNInfo *VNI = OldLI.getVNInfoAt(Idx.getRegSlot(true)))
1110 if (SlotIndex::isSameInstr(Idx, VNI->def))
1111 Idx = VNI->def;
1112
1113 // Check for a sibling copy.
1114 Register SibReg = isFullCopyOf(MI, Reg);
1115 if (SibReg && isSibling(SibReg)) {
1116 // This may actually be a copy between snippets.
1117 if (isRegToSpill(SibReg)) {
1118 LLVM_DEBUG(dbgs() << "Found new snippet copy: " << MI);
1119 SnippetCopies.insert(&MI);
1120 continue;
1121 }
1122 if (RI.Writes) {
1123 if (hoistSpillInsideBB(OldLI, MI)) {
1124 // This COPY is now dead, the value is already in the stack slot.
1125 MI.getOperand(0).setIsDead();
1126 DeadDefs.push_back(&MI);
1127 continue;
1128 }
1129 } else {
1130 // This is a reload for a sib-reg copy. Drop spills downstream.
1131 LiveInterval &SibLI = LIS.getInterval(SibReg);
1132 eliminateRedundantSpills(SibLI, SibLI.getVNInfoAt(Idx));
1133 // The COPY will fold to a reload below.
1134 }
1135 }
1136
1137 // Attempt to fold memory ops.
1138 if (foldMemoryOperand(Ops))
1139 continue;
1140
1141 // Create a new virtual register for spill/fill.
1142 // FIXME: Infer regclass from instruction alone.
1143 Register NewVReg = Edit->createFrom(Reg);
1144
1145 if (RI.Reads)
1146 insertReload(NewVReg, Idx, &MI);
1147
1148 // Rewrite instruction operands.
1149 bool hasLiveDef = false;
1150 for (const auto &OpPair : Ops) {
1151 MachineOperand &MO = OpPair.first->getOperand(OpPair.second);
1152 MO.setReg(NewVReg);
1153 if (MO.isUse()) {
1154 if (!OpPair.first->isRegTiedToDefOperand(OpPair.second))
1155 MO.setIsKill();
1156 } else {
1157 if (!MO.isDead())
1158 hasLiveDef = true;
1159 }
1160 }
1161 LLVM_DEBUG(dbgs() << "\trewrite: " << Idx << '\t' << MI << '\n');
1162
1163 // FIXME: Use a second vreg if instruction has no tied ops.
1164 if (RI.Writes)
1165 if (hasLiveDef)
1166 insertSpill(NewVReg, true, &MI);
1167 }
1168 }
1169
1170 /// spillAll - Spill all registers remaining after rematerialization.
spillAll()1171 void InlineSpiller::spillAll() {
1172 // Update LiveStacks now that we are committed to spilling.
1173 if (StackSlot == VirtRegMap::NO_STACK_SLOT) {
1174 StackSlot = VRM.assignVirt2StackSlot(Original);
1175 StackInt = &LSS.getOrCreateInterval(StackSlot, MRI.getRegClass(Original));
1176 StackInt->getNextValue(SlotIndex(), LSS.getVNInfoAllocator());
1177 } else
1178 StackInt = &LSS.getInterval(StackSlot);
1179
1180 if (Original != Edit->getReg())
1181 VRM.assignVirt2StackSlot(Edit->getReg(), StackSlot);
1182
1183 assert(StackInt->getNumValNums() == 1 && "Bad stack interval values");
1184 for (Register Reg : RegsToSpill)
1185 StackInt->MergeSegmentsInAsValue(LIS.getInterval(Reg),
1186 StackInt->getValNumInfo(0));
1187 LLVM_DEBUG(dbgs() << "Merged spilled regs: " << *StackInt << '\n');
1188
1189 // Spill around uses of all RegsToSpill.
1190 for (Register Reg : RegsToSpill)
1191 spillAroundUses(Reg);
1192
1193 // Hoisted spills may cause dead code.
1194 if (!DeadDefs.empty()) {
1195 LLVM_DEBUG(dbgs() << "Eliminating " << DeadDefs.size() << " dead defs\n");
1196 Edit->eliminateDeadDefs(DeadDefs, RegsToSpill);
1197 }
1198
1199 // Finally delete the SnippetCopies.
1200 for (Register Reg : RegsToSpill) {
1201 for (MachineInstr &MI :
1202 llvm::make_early_inc_range(MRI.reg_instructions(Reg))) {
1203 assert(SnippetCopies.count(&MI) && "Remaining use wasn't a snippet copy");
1204 // FIXME: Do this with a LiveRangeEdit callback.
1205 LIS.RemoveMachineInstrFromMaps(MI);
1206 MI.eraseFromParent();
1207 }
1208 }
1209
1210 // Delete all spilled registers.
1211 for (Register Reg : RegsToSpill)
1212 Edit->eraseVirtReg(Reg);
1213 }
1214
spill(LiveRangeEdit & edit)1215 void InlineSpiller::spill(LiveRangeEdit &edit) {
1216 ++NumSpilledRanges;
1217 Edit = &edit;
1218 assert(!Register::isStackSlot(edit.getReg()) &&
1219 "Trying to spill a stack slot.");
1220 // Share a stack slot among all descendants of Original.
1221 Original = VRM.getOriginal(edit.getReg());
1222 StackSlot = VRM.getStackSlot(Original);
1223 StackInt = nullptr;
1224
1225 LLVM_DEBUG(dbgs() << "Inline spilling "
1226 << TRI.getRegClassName(MRI.getRegClass(edit.getReg()))
1227 << ':' << edit.getParent() << "\nFrom original "
1228 << printReg(Original) << '\n');
1229 assert(edit.getParent().isSpillable() &&
1230 "Attempting to spill already spilled value.");
1231 assert(DeadDefs.empty() && "Previous spill didn't remove dead defs");
1232
1233 collectRegsToSpill();
1234 reMaterializeAll();
1235
1236 // Remat may handle everything.
1237 if (!RegsToSpill.empty())
1238 spillAll();
1239
1240 Edit->calculateRegClassAndHint(MF, VRAI);
1241 }
1242
1243 /// Optimizations after all the reg selections and spills are done.
postOptimization()1244 void InlineSpiller::postOptimization() { HSpiller.hoistAllSpills(); }
1245
1246 /// When a spill is inserted, add the spill to MergeableSpills map.
addToMergeableSpills(MachineInstr & Spill,int StackSlot,unsigned Original)1247 void HoistSpillHelper::addToMergeableSpills(MachineInstr &Spill, int StackSlot,
1248 unsigned Original) {
1249 BumpPtrAllocator &Allocator = LIS.getVNInfoAllocator();
1250 LiveInterval &OrigLI = LIS.getInterval(Original);
1251 // save a copy of LiveInterval in StackSlotToOrigLI because the original
1252 // LiveInterval may be cleared after all its references are spilled.
1253 if (StackSlotToOrigLI.find(StackSlot) == StackSlotToOrigLI.end()) {
1254 auto LI = std::make_unique<LiveInterval>(OrigLI.reg(), OrigLI.weight());
1255 LI->assign(OrigLI, Allocator);
1256 StackSlotToOrigLI[StackSlot] = std::move(LI);
1257 }
1258 SlotIndex Idx = LIS.getInstructionIndex(Spill);
1259 VNInfo *OrigVNI = StackSlotToOrigLI[StackSlot]->getVNInfoAt(Idx.getRegSlot());
1260 std::pair<int, VNInfo *> MIdx = std::make_pair(StackSlot, OrigVNI);
1261 MergeableSpills[MIdx].insert(&Spill);
1262 }
1263
1264 /// When a spill is removed, remove the spill from MergeableSpills map.
1265 /// Return true if the spill is removed successfully.
rmFromMergeableSpills(MachineInstr & Spill,int StackSlot)1266 bool HoistSpillHelper::rmFromMergeableSpills(MachineInstr &Spill,
1267 int StackSlot) {
1268 auto It = StackSlotToOrigLI.find(StackSlot);
1269 if (It == StackSlotToOrigLI.end())
1270 return false;
1271 SlotIndex Idx = LIS.getInstructionIndex(Spill);
1272 VNInfo *OrigVNI = It->second->getVNInfoAt(Idx.getRegSlot());
1273 std::pair<int, VNInfo *> MIdx = std::make_pair(StackSlot, OrigVNI);
1274 return MergeableSpills[MIdx].erase(&Spill);
1275 }
1276
1277 /// Check BB to see if it is a possible target BB to place a hoisted spill,
1278 /// i.e., there should be a living sibling of OrigReg at the insert point.
isSpillCandBB(LiveInterval & OrigLI,VNInfo & OrigVNI,MachineBasicBlock & BB,Register & LiveReg)1279 bool HoistSpillHelper::isSpillCandBB(LiveInterval &OrigLI, VNInfo &OrigVNI,
1280 MachineBasicBlock &BB, Register &LiveReg) {
1281 SlotIndex Idx = IPA.getLastInsertPoint(OrigLI, BB);
1282 // The original def could be after the last insert point in the root block,
1283 // we can't hoist to here.
1284 if (Idx < OrigVNI.def) {
1285 // TODO: We could be better here. If LI is not alive in landing pad
1286 // we could hoist spill after LIP.
1287 LLVM_DEBUG(dbgs() << "can't spill in root block - def after LIP\n");
1288 return false;
1289 }
1290 Register OrigReg = OrigLI.reg();
1291 SmallSetVector<Register, 16> &Siblings = Virt2SiblingsMap[OrigReg];
1292 assert(OrigLI.getVNInfoAt(Idx) == &OrigVNI && "Unexpected VNI");
1293
1294 for (const Register &SibReg : Siblings) {
1295 LiveInterval &LI = LIS.getInterval(SibReg);
1296 VNInfo *VNI = LI.getVNInfoAt(Idx);
1297 if (VNI) {
1298 LiveReg = SibReg;
1299 return true;
1300 }
1301 }
1302 return false;
1303 }
1304
1305 /// Remove redundant spills in the same BB. Save those redundant spills in
1306 /// SpillsToRm, and save the spill to keep and its BB in SpillBBToSpill map.
rmRedundantSpills(SmallPtrSet<MachineInstr *,16> & Spills,SmallVectorImpl<MachineInstr * > & SpillsToRm,DenseMap<MachineDomTreeNode *,MachineInstr * > & SpillBBToSpill)1307 void HoistSpillHelper::rmRedundantSpills(
1308 SmallPtrSet<MachineInstr *, 16> &Spills,
1309 SmallVectorImpl<MachineInstr *> &SpillsToRm,
1310 DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill) {
1311 // For each spill saw, check SpillBBToSpill[] and see if its BB already has
1312 // another spill inside. If a BB contains more than one spill, only keep the
1313 // earlier spill with smaller SlotIndex.
1314 for (auto *const CurrentSpill : Spills) {
1315 MachineBasicBlock *Block = CurrentSpill->getParent();
1316 MachineDomTreeNode *Node = MDT.getBase().getNode(Block);
1317 MachineInstr *PrevSpill = SpillBBToSpill[Node];
1318 if (PrevSpill) {
1319 SlotIndex PIdx = LIS.getInstructionIndex(*PrevSpill);
1320 SlotIndex CIdx = LIS.getInstructionIndex(*CurrentSpill);
1321 MachineInstr *SpillToRm = (CIdx > PIdx) ? CurrentSpill : PrevSpill;
1322 MachineInstr *SpillToKeep = (CIdx > PIdx) ? PrevSpill : CurrentSpill;
1323 SpillsToRm.push_back(SpillToRm);
1324 SpillBBToSpill[MDT.getBase().getNode(Block)] = SpillToKeep;
1325 } else {
1326 SpillBBToSpill[MDT.getBase().getNode(Block)] = CurrentSpill;
1327 }
1328 }
1329 for (auto *const SpillToRm : SpillsToRm)
1330 Spills.erase(SpillToRm);
1331 }
1332
1333 /// Starting from \p Root find a top-down traversal order of the dominator
1334 /// tree to visit all basic blocks containing the elements of \p Spills.
1335 /// Redundant spills will be found and put into \p SpillsToRm at the same
1336 /// time. \p SpillBBToSpill will be populated as part of the process and
1337 /// maps a basic block to the first store occurring in the basic block.
1338 /// \post SpillsToRm.union(Spills\@post) == Spills\@pre
getVisitOrders(MachineBasicBlock * Root,SmallPtrSet<MachineInstr *,16> & Spills,SmallVectorImpl<MachineDomTreeNode * > & Orders,SmallVectorImpl<MachineInstr * > & SpillsToRm,DenseMap<MachineDomTreeNode *,unsigned> & SpillsToKeep,DenseMap<MachineDomTreeNode *,MachineInstr * > & SpillBBToSpill)1339 void HoistSpillHelper::getVisitOrders(
1340 MachineBasicBlock *Root, SmallPtrSet<MachineInstr *, 16> &Spills,
1341 SmallVectorImpl<MachineDomTreeNode *> &Orders,
1342 SmallVectorImpl<MachineInstr *> &SpillsToRm,
1343 DenseMap<MachineDomTreeNode *, unsigned> &SpillsToKeep,
1344 DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill) {
1345 // The set contains all the possible BB nodes to which we may hoist
1346 // original spills.
1347 SmallPtrSet<MachineDomTreeNode *, 8> WorkSet;
1348 // Save the BB nodes on the path from the first BB node containing
1349 // non-redundant spill to the Root node.
1350 SmallPtrSet<MachineDomTreeNode *, 8> NodesOnPath;
1351 // All the spills to be hoisted must originate from a single def instruction
1352 // to the OrigReg. It means the def instruction should dominate all the spills
1353 // to be hoisted. We choose the BB where the def instruction is located as
1354 // the Root.
1355 MachineDomTreeNode *RootIDomNode = MDT[Root]->getIDom();
1356 // For every node on the dominator tree with spill, walk up on the dominator
1357 // tree towards the Root node until it is reached. If there is other node
1358 // containing spill in the middle of the path, the previous spill saw will
1359 // be redundant and the node containing it will be removed. All the nodes on
1360 // the path starting from the first node with non-redundant spill to the Root
1361 // node will be added to the WorkSet, which will contain all the possible
1362 // locations where spills may be hoisted to after the loop below is done.
1363 for (auto *const Spill : Spills) {
1364 MachineBasicBlock *Block = Spill->getParent();
1365 MachineDomTreeNode *Node = MDT[Block];
1366 MachineInstr *SpillToRm = nullptr;
1367 while (Node != RootIDomNode) {
1368 // If Node dominates Block, and it already contains a spill, the spill in
1369 // Block will be redundant.
1370 if (Node != MDT[Block] && SpillBBToSpill[Node]) {
1371 SpillToRm = SpillBBToSpill[MDT[Block]];
1372 break;
1373 /// If we see the Node already in WorkSet, the path from the Node to
1374 /// the Root node must already be traversed by another spill.
1375 /// Then no need to repeat.
1376 } else if (WorkSet.count(Node)) {
1377 break;
1378 } else {
1379 NodesOnPath.insert(Node);
1380 }
1381 Node = Node->getIDom();
1382 }
1383 if (SpillToRm) {
1384 SpillsToRm.push_back(SpillToRm);
1385 } else {
1386 // Add a BB containing the original spills to SpillsToKeep -- i.e.,
1387 // set the initial status before hoisting start. The value of BBs
1388 // containing original spills is set to 0, in order to descriminate
1389 // with BBs containing hoisted spills which will be inserted to
1390 // SpillsToKeep later during hoisting.
1391 SpillsToKeep[MDT[Block]] = 0;
1392 WorkSet.insert(NodesOnPath.begin(), NodesOnPath.end());
1393 }
1394 NodesOnPath.clear();
1395 }
1396
1397 // Sort the nodes in WorkSet in top-down order and save the nodes
1398 // in Orders. Orders will be used for hoisting in runHoistSpills.
1399 unsigned idx = 0;
1400 Orders.push_back(MDT.getBase().getNode(Root));
1401 do {
1402 MachineDomTreeNode *Node = Orders[idx++];
1403 for (MachineDomTreeNode *Child : Node->children()) {
1404 if (WorkSet.count(Child))
1405 Orders.push_back(Child);
1406 }
1407 } while (idx != Orders.size());
1408 assert(Orders.size() == WorkSet.size() &&
1409 "Orders have different size with WorkSet");
1410
1411 #ifndef NDEBUG
1412 LLVM_DEBUG(dbgs() << "Orders size is " << Orders.size() << "\n");
1413 SmallVector<MachineDomTreeNode *, 32>::reverse_iterator RIt = Orders.rbegin();
1414 for (; RIt != Orders.rend(); RIt++)
1415 LLVM_DEBUG(dbgs() << "BB" << (*RIt)->getBlock()->getNumber() << ",");
1416 LLVM_DEBUG(dbgs() << "\n");
1417 #endif
1418 }
1419
1420 /// Try to hoist spills according to BB hotness. The spills to removed will
1421 /// be saved in \p SpillsToRm. The spills to be inserted will be saved in
1422 /// \p SpillsToIns.
runHoistSpills(LiveInterval & OrigLI,VNInfo & OrigVNI,SmallPtrSet<MachineInstr *,16> & Spills,SmallVectorImpl<MachineInstr * > & SpillsToRm,DenseMap<MachineBasicBlock *,unsigned> & SpillsToIns)1423 void HoistSpillHelper::runHoistSpills(
1424 LiveInterval &OrigLI, VNInfo &OrigVNI,
1425 SmallPtrSet<MachineInstr *, 16> &Spills,
1426 SmallVectorImpl<MachineInstr *> &SpillsToRm,
1427 DenseMap<MachineBasicBlock *, unsigned> &SpillsToIns) {
1428 // Visit order of dominator tree nodes.
1429 SmallVector<MachineDomTreeNode *, 32> Orders;
1430 // SpillsToKeep contains all the nodes where spills are to be inserted
1431 // during hoisting. If the spill to be inserted is an original spill
1432 // (not a hoisted one), the value of the map entry is 0. If the spill
1433 // is a hoisted spill, the value of the map entry is the VReg to be used
1434 // as the source of the spill.
1435 DenseMap<MachineDomTreeNode *, unsigned> SpillsToKeep;
1436 // Map from BB to the first spill inside of it.
1437 DenseMap<MachineDomTreeNode *, MachineInstr *> SpillBBToSpill;
1438
1439 rmRedundantSpills(Spills, SpillsToRm, SpillBBToSpill);
1440
1441 MachineBasicBlock *Root = LIS.getMBBFromIndex(OrigVNI.def);
1442 getVisitOrders(Root, Spills, Orders, SpillsToRm, SpillsToKeep,
1443 SpillBBToSpill);
1444
1445 // SpillsInSubTreeMap keeps the map from a dom tree node to a pair of
1446 // nodes set and the cost of all the spills inside those nodes.
1447 // The nodes set are the locations where spills are to be inserted
1448 // in the subtree of current node.
1449 using NodesCostPair =
1450 std::pair<SmallPtrSet<MachineDomTreeNode *, 16>, BlockFrequency>;
1451 DenseMap<MachineDomTreeNode *, NodesCostPair> SpillsInSubTreeMap;
1452
1453 // Iterate Orders set in reverse order, which will be a bottom-up order
1454 // in the dominator tree. Once we visit a dom tree node, we know its
1455 // children have already been visited and the spill locations in the
1456 // subtrees of all the children have been determined.
1457 SmallVector<MachineDomTreeNode *, 32>::reverse_iterator RIt = Orders.rbegin();
1458 for (; RIt != Orders.rend(); RIt++) {
1459 MachineBasicBlock *Block = (*RIt)->getBlock();
1460
1461 // If Block contains an original spill, simply continue.
1462 if (SpillsToKeep.find(*RIt) != SpillsToKeep.end() && !SpillsToKeep[*RIt]) {
1463 SpillsInSubTreeMap[*RIt].first.insert(*RIt);
1464 // SpillsInSubTreeMap[*RIt].second contains the cost of spill.
1465 SpillsInSubTreeMap[*RIt].second = MBFI.getBlockFreq(Block);
1466 continue;
1467 }
1468
1469 // Collect spills in subtree of current node (*RIt) to
1470 // SpillsInSubTreeMap[*RIt].first.
1471 for (MachineDomTreeNode *Child : (*RIt)->children()) {
1472 if (SpillsInSubTreeMap.find(Child) == SpillsInSubTreeMap.end())
1473 continue;
1474 // The stmt "SpillsInSubTree = SpillsInSubTreeMap[*RIt].first" below
1475 // should be placed before getting the begin and end iterators of
1476 // SpillsInSubTreeMap[Child].first, or else the iterators may be
1477 // invalidated when SpillsInSubTreeMap[*RIt] is seen the first time
1478 // and the map grows and then the original buckets in the map are moved.
1479 SmallPtrSet<MachineDomTreeNode *, 16> &SpillsInSubTree =
1480 SpillsInSubTreeMap[*RIt].first;
1481 BlockFrequency &SubTreeCost = SpillsInSubTreeMap[*RIt].second;
1482 SubTreeCost += SpillsInSubTreeMap[Child].second;
1483 auto BI = SpillsInSubTreeMap[Child].first.begin();
1484 auto EI = SpillsInSubTreeMap[Child].first.end();
1485 SpillsInSubTree.insert(BI, EI);
1486 SpillsInSubTreeMap.erase(Child);
1487 }
1488
1489 SmallPtrSet<MachineDomTreeNode *, 16> &SpillsInSubTree =
1490 SpillsInSubTreeMap[*RIt].first;
1491 BlockFrequency &SubTreeCost = SpillsInSubTreeMap[*RIt].second;
1492 // No spills in subtree, simply continue.
1493 if (SpillsInSubTree.empty())
1494 continue;
1495
1496 // Check whether Block is a possible candidate to insert spill.
1497 Register LiveReg;
1498 if (!isSpillCandBB(OrigLI, OrigVNI, *Block, LiveReg))
1499 continue;
1500
1501 // If there are multiple spills that could be merged, bias a little
1502 // to hoist the spill.
1503 BranchProbability MarginProb = (SpillsInSubTree.size() > 1)
1504 ? BranchProbability(9, 10)
1505 : BranchProbability(1, 1);
1506 if (SubTreeCost > MBFI.getBlockFreq(Block) * MarginProb) {
1507 // Hoist: Move spills to current Block.
1508 for (auto *const SpillBB : SpillsInSubTree) {
1509 // When SpillBB is a BB contains original spill, insert the spill
1510 // to SpillsToRm.
1511 if (SpillsToKeep.find(SpillBB) != SpillsToKeep.end() &&
1512 !SpillsToKeep[SpillBB]) {
1513 MachineInstr *SpillToRm = SpillBBToSpill[SpillBB];
1514 SpillsToRm.push_back(SpillToRm);
1515 }
1516 // SpillBB will not contain spill anymore, remove it from SpillsToKeep.
1517 SpillsToKeep.erase(SpillBB);
1518 }
1519 // Current Block is the BB containing the new hoisted spill. Add it to
1520 // SpillsToKeep. LiveReg is the source of the new spill.
1521 SpillsToKeep[*RIt] = LiveReg;
1522 LLVM_DEBUG({
1523 dbgs() << "spills in BB: ";
1524 for (const auto Rspill : SpillsInSubTree)
1525 dbgs() << Rspill->getBlock()->getNumber() << " ";
1526 dbgs() << "were promoted to BB" << (*RIt)->getBlock()->getNumber()
1527 << "\n";
1528 });
1529 SpillsInSubTree.clear();
1530 SpillsInSubTree.insert(*RIt);
1531 SubTreeCost = MBFI.getBlockFreq(Block);
1532 }
1533 }
1534 // For spills in SpillsToKeep with LiveReg set (i.e., not original spill),
1535 // save them to SpillsToIns.
1536 for (const auto &Ent : SpillsToKeep) {
1537 if (Ent.second)
1538 SpillsToIns[Ent.first->getBlock()] = Ent.second;
1539 }
1540 }
1541
1542 /// For spills with equal values, remove redundant spills and hoist those left
1543 /// to less hot spots.
1544 ///
1545 /// Spills with equal values will be collected into the same set in
1546 /// MergeableSpills when spill is inserted. These equal spills are originated
1547 /// from the same defining instruction and are dominated by the instruction.
1548 /// Before hoisting all the equal spills, redundant spills inside in the same
1549 /// BB are first marked to be deleted. Then starting from the spills left, walk
1550 /// up on the dominator tree towards the Root node where the define instruction
1551 /// is located, mark the dominated spills to be deleted along the way and
1552 /// collect the BB nodes on the path from non-dominated spills to the define
1553 /// instruction into a WorkSet. The nodes in WorkSet are the candidate places
1554 /// where we are considering to hoist the spills. We iterate the WorkSet in
1555 /// bottom-up order, and for each node, we will decide whether to hoist spills
1556 /// inside its subtree to that node. In this way, we can get benefit locally
1557 /// even if hoisting all the equal spills to one cold place is impossible.
hoistAllSpills()1558 void HoistSpillHelper::hoistAllSpills() {
1559 SmallVector<Register, 4> NewVRegs;
1560 LiveRangeEdit Edit(nullptr, NewVRegs, MF, LIS, &VRM, this);
1561
1562 for (unsigned i = 0, e = MRI.getNumVirtRegs(); i != e; ++i) {
1563 Register Reg = Register::index2VirtReg(i);
1564 Register Original = VRM.getPreSplitReg(Reg);
1565 if (!MRI.def_empty(Reg))
1566 Virt2SiblingsMap[Original].insert(Reg);
1567 }
1568
1569 // Each entry in MergeableSpills contains a spill set with equal values.
1570 for (auto &Ent : MergeableSpills) {
1571 int Slot = Ent.first.first;
1572 LiveInterval &OrigLI = *StackSlotToOrigLI[Slot];
1573 VNInfo *OrigVNI = Ent.first.second;
1574 SmallPtrSet<MachineInstr *, 16> &EqValSpills = Ent.second;
1575 if (Ent.second.empty())
1576 continue;
1577
1578 LLVM_DEBUG({
1579 dbgs() << "\nFor Slot" << Slot << " and VN" << OrigVNI->id << ":\n"
1580 << "Equal spills in BB: ";
1581 for (const auto spill : EqValSpills)
1582 dbgs() << spill->getParent()->getNumber() << " ";
1583 dbgs() << "\n";
1584 });
1585
1586 // SpillsToRm is the spill set to be removed from EqValSpills.
1587 SmallVector<MachineInstr *, 16> SpillsToRm;
1588 // SpillsToIns is the spill set to be newly inserted after hoisting.
1589 DenseMap<MachineBasicBlock *, unsigned> SpillsToIns;
1590
1591 runHoistSpills(OrigLI, *OrigVNI, EqValSpills, SpillsToRm, SpillsToIns);
1592
1593 LLVM_DEBUG({
1594 dbgs() << "Finally inserted spills in BB: ";
1595 for (const auto &Ispill : SpillsToIns)
1596 dbgs() << Ispill.first->getNumber() << " ";
1597 dbgs() << "\nFinally removed spills in BB: ";
1598 for (const auto Rspill : SpillsToRm)
1599 dbgs() << Rspill->getParent()->getNumber() << " ";
1600 dbgs() << "\n";
1601 });
1602
1603 // Stack live range update.
1604 LiveInterval &StackIntvl = LSS.getInterval(Slot);
1605 if (!SpillsToIns.empty() || !SpillsToRm.empty())
1606 StackIntvl.MergeValueInAsValue(OrigLI, OrigVNI,
1607 StackIntvl.getValNumInfo(0));
1608
1609 // Insert hoisted spills.
1610 for (auto const &Insert : SpillsToIns) {
1611 MachineBasicBlock *BB = Insert.first;
1612 Register LiveReg = Insert.second;
1613 MachineBasicBlock::iterator MII = IPA.getLastInsertPointIter(OrigLI, *BB);
1614 MachineInstrSpan MIS(MII, BB);
1615 TII.storeRegToStackSlot(*BB, MII, LiveReg, false, Slot,
1616 MRI.getRegClass(LiveReg), &TRI, Register());
1617 LIS.InsertMachineInstrRangeInMaps(MIS.begin(), MII);
1618 for (const MachineInstr &MI : make_range(MIS.begin(), MII))
1619 getVDefInterval(MI, LIS);
1620 ++NumSpills;
1621 }
1622
1623 // Remove redundant spills or change them to dead instructions.
1624 NumSpills -= SpillsToRm.size();
1625 for (auto *const RMEnt : SpillsToRm) {
1626 RMEnt->setDesc(TII.get(TargetOpcode::KILL));
1627 for (unsigned i = RMEnt->getNumOperands(); i; --i) {
1628 MachineOperand &MO = RMEnt->getOperand(i - 1);
1629 if (MO.isReg() && MO.isImplicit() && MO.isDef() && !MO.isDead())
1630 RMEnt->removeOperand(i - 1);
1631 }
1632 }
1633 Edit.eliminateDeadDefs(SpillsToRm, std::nullopt);
1634 }
1635 }
1636
1637 /// For VirtReg clone, the \p New register should have the same physreg or
1638 /// stackslot as the \p old register.
LRE_DidCloneVirtReg(Register New,Register Old)1639 void HoistSpillHelper::LRE_DidCloneVirtReg(Register New, Register Old) {
1640 if (VRM.hasPhys(Old))
1641 VRM.assignVirt2Phys(New, VRM.getPhys(Old));
1642 else if (VRM.getStackSlot(Old) != VirtRegMap::NO_STACK_SLOT)
1643 VRM.assignVirt2StackSlot(New, VRM.getStackSlot(Old));
1644 else
1645 llvm_unreachable("VReg should be assigned either physreg or stackslot");
1646 if (VRM.hasShape(Old))
1647 VRM.assignVirt2Shape(New, VRM.getShape(Old));
1648 }
1649