1 //===- MachineSink.cpp - Sinking for machine instructions -----------------===//
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 // This pass moves instructions into successor blocks when possible, so that
10 // they aren't executed on paths where their results aren't needed.
11 //
12 // This pass is not intended to be a replacement or a complete alternative
13 // for an LLVM-IR-level sinking pass. It is only designed to sink simple
14 // constructs that are not exposed before lowering and instruction selection.
15 //
16 //===----------------------------------------------------------------------===//
17 
18 #include "llvm/ADT/DenseSet.h"
19 #include "llvm/ADT/MapVector.h"
20 #include "llvm/ADT/PointerIntPair.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/SparseBitVector.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/Analysis/AliasAnalysis.h"
27 #include "llvm/CodeGen/MachineBasicBlock.h"
28 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
29 #include "llvm/CodeGen/MachineBranchProbabilityInfo.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/MachineLoopInfo.h"
35 #include "llvm/CodeGen/MachineOperand.h"
36 #include "llvm/CodeGen/MachinePostDominators.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/RegisterClassInfo.h"
39 #include "llvm/CodeGen/RegisterPressure.h"
40 #include "llvm/CodeGen/TargetInstrInfo.h"
41 #include "llvm/CodeGen/TargetRegisterInfo.h"
42 #include "llvm/CodeGen/TargetSubtargetInfo.h"
43 #include "llvm/IR/BasicBlock.h"
44 #include "llvm/IR/DebugInfoMetadata.h"
45 #include "llvm/IR/LLVMContext.h"
46 #include "llvm/InitializePasses.h"
47 #include "llvm/MC/MCRegisterInfo.h"
48 #include "llvm/Pass.h"
49 #include "llvm/Support/BranchProbability.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/raw_ostream.h"
53 #include <algorithm>
54 #include <cassert>
55 #include <cstdint>
56 #include <map>
57 #include <utility>
58 #include <vector>
59 
60 using namespace llvm;
61 
62 #define DEBUG_TYPE "machine-sink"
63 
64 static cl::opt<bool>
65 SplitEdges("machine-sink-split",
66            cl::desc("Split critical edges during machine sinking"),
67            cl::init(true), cl::Hidden);
68 
69 static cl::opt<bool>
70 UseBlockFreqInfo("machine-sink-bfi",
71            cl::desc("Use block frequency info to find successors to sink"),
72            cl::init(true), cl::Hidden);
73 
74 static cl::opt<unsigned> SplitEdgeProbabilityThreshold(
75     "machine-sink-split-probability-threshold",
76     cl::desc(
77         "Percentage threshold for splitting single-instruction critical edge. "
78         "If the branch threshold is higher than this threshold, we allow "
79         "speculative execution of up to 1 instruction to avoid branching to "
80         "splitted critical edge"),
81     cl::init(40), cl::Hidden);
82 
83 static cl::opt<unsigned> SinkLoadInstsPerBlockThreshold(
84     "machine-sink-load-instrs-threshold",
85     cl::desc("Do not try to find alias store for a load if there is a in-path "
86              "block whose instruction number is higher than this threshold."),
87     cl::init(2000), cl::Hidden);
88 
89 static cl::opt<unsigned> SinkLoadBlocksThreshold(
90     "machine-sink-load-blocks-threshold",
91     cl::desc("Do not try to find alias store for a load if the block number in "
92              "the straight line is higher than this threshold."),
93     cl::init(20), cl::Hidden);
94 
95 static cl::opt<bool>
96 SinkInstsIntoLoop("sink-insts-to-avoid-spills",
97                   cl::desc("Sink instructions into loops to avoid "
98                            "register spills"),
99                   cl::init(false), cl::Hidden);
100 
101 static cl::opt<unsigned> SinkIntoLoopLimit(
102     "machine-sink-loop-limit",
103     cl::desc("The maximum number of instructions considered for loop sinking."),
104     cl::init(50), cl::Hidden);
105 
106 STATISTIC(NumSunk,      "Number of machine instructions sunk");
107 STATISTIC(NumLoopSunk,  "Number of machine instructions sunk into a loop");
108 STATISTIC(NumSplit,     "Number of critical edges split");
109 STATISTIC(NumCoalesces, "Number of copies coalesced");
110 STATISTIC(NumPostRACopySink, "Number of copies sunk after RA");
111 
112 namespace {
113 
114   class MachineSinking : public MachineFunctionPass {
115     const TargetInstrInfo *TII;
116     const TargetRegisterInfo *TRI;
117     MachineRegisterInfo  *MRI;     // Machine register information
118     MachineDominatorTree *DT;      // Machine dominator tree
119     MachinePostDominatorTree *PDT; // Machine post dominator tree
120     MachineLoopInfo *LI;
121     MachineBlockFrequencyInfo *MBFI;
122     const MachineBranchProbabilityInfo *MBPI;
123     AliasAnalysis *AA;
124     RegisterClassInfo RegClassInfo;
125 
126     // Remember which edges have been considered for breaking.
127     SmallSet<std::pair<MachineBasicBlock*, MachineBasicBlock*>, 8>
128     CEBCandidates;
129     // Remember which edges we are about to split.
130     // This is different from CEBCandidates since those edges
131     // will be split.
132     SetVector<std::pair<MachineBasicBlock *, MachineBasicBlock *>> ToSplit;
133 
134     SparseBitVector<> RegsToClearKillFlags;
135 
136     using AllSuccsCache =
137         std::map<MachineBasicBlock *, SmallVector<MachineBasicBlock *, 4>>;
138 
139     /// DBG_VALUE pointer and flag. The flag is true if this DBG_VALUE is
140     /// post-dominated by another DBG_VALUE of the same variable location.
141     /// This is necessary to detect sequences such as:
142     ///     %0 = someinst
143     ///     DBG_VALUE %0, !123, !DIExpression()
144     ///     %1 = anotherinst
145     ///     DBG_VALUE %1, !123, !DIExpression()
146     /// Where if %0 were to sink, the DBG_VAUE should not sink with it, as that
147     /// would re-order assignments.
148     using SeenDbgUser = PointerIntPair<MachineInstr *, 1>;
149 
150     /// Record of DBG_VALUE uses of vregs in a block, so that we can identify
151     /// debug instructions to sink.
152     SmallDenseMap<unsigned, TinyPtrVector<SeenDbgUser>> SeenDbgUsers;
153 
154     /// Record of debug variables that have had their locations set in the
155     /// current block.
156     DenseSet<DebugVariable> SeenDbgVars;
157 
158     std::map<std::pair<MachineBasicBlock *, MachineBasicBlock *>, bool>
159         HasStoreCache;
160     std::map<std::pair<MachineBasicBlock *, MachineBasicBlock *>,
161              std::vector<MachineInstr *>>
162         StoreInstrCache;
163 
164     /// Cached BB's register pressure.
165     std::map<MachineBasicBlock *, std::vector<unsigned>> CachedRegisterPressure;
166 
167   public:
168     static char ID; // Pass identification
169 
170     MachineSinking() : MachineFunctionPass(ID) {
171       initializeMachineSinkingPass(*PassRegistry::getPassRegistry());
172     }
173 
174     bool runOnMachineFunction(MachineFunction &MF) override;
175 
176     void getAnalysisUsage(AnalysisUsage &AU) const override {
177       MachineFunctionPass::getAnalysisUsage(AU);
178       AU.addRequired<AAResultsWrapperPass>();
179       AU.addRequired<MachineDominatorTree>();
180       AU.addRequired<MachinePostDominatorTree>();
181       AU.addRequired<MachineLoopInfo>();
182       AU.addRequired<MachineBranchProbabilityInfo>();
183       AU.addPreserved<MachineLoopInfo>();
184       if (UseBlockFreqInfo)
185         AU.addRequired<MachineBlockFrequencyInfo>();
186     }
187 
188     void releaseMemory() override {
189       CEBCandidates.clear();
190     }
191 
192   private:
193     bool ProcessBlock(MachineBasicBlock &MBB);
194     void ProcessDbgInst(MachineInstr &MI);
195     bool isWorthBreakingCriticalEdge(MachineInstr &MI,
196                                      MachineBasicBlock *From,
197                                      MachineBasicBlock *To);
198 
199     bool hasStoreBetween(MachineBasicBlock *From, MachineBasicBlock *To,
200                          MachineInstr &MI);
201 
202     /// Postpone the splitting of the given critical
203     /// edge (\p From, \p To).
204     ///
205     /// We do not split the edges on the fly. Indeed, this invalidates
206     /// the dominance information and thus triggers a lot of updates
207     /// of that information underneath.
208     /// Instead, we postpone all the splits after each iteration of
209     /// the main loop. That way, the information is at least valid
210     /// for the lifetime of an iteration.
211     ///
212     /// \return True if the edge is marked as toSplit, false otherwise.
213     /// False can be returned if, for instance, this is not profitable.
214     bool PostponeSplitCriticalEdge(MachineInstr &MI,
215                                    MachineBasicBlock *From,
216                                    MachineBasicBlock *To,
217                                    bool BreakPHIEdge);
218     bool SinkInstruction(MachineInstr &MI, bool &SawStore,
219                          AllSuccsCache &AllSuccessors);
220 
221     /// If we sink a COPY inst, some debug users of it's destination may no
222     /// longer be dominated by the COPY, and will eventually be dropped.
223     /// This is easily rectified by forwarding the non-dominated debug uses
224     /// to the copy source.
225     void SalvageUnsunkDebugUsersOfCopy(MachineInstr &,
226                                        MachineBasicBlock *TargetBlock);
227     bool AllUsesDominatedByBlock(Register Reg, MachineBasicBlock *MBB,
228                                  MachineBasicBlock *DefMBB, bool &BreakPHIEdge,
229                                  bool &LocalUse) const;
230     MachineBasicBlock *FindSuccToSinkTo(MachineInstr &MI, MachineBasicBlock *MBB,
231                bool &BreakPHIEdge, AllSuccsCache &AllSuccessors);
232 
233     void FindLoopSinkCandidates(MachineLoop *L, MachineBasicBlock *BB,
234                                 SmallVectorImpl<MachineInstr *> &Candidates);
235     bool SinkIntoLoop(MachineLoop *L, MachineInstr &I);
236 
237     bool isProfitableToSinkTo(Register Reg, MachineInstr &MI,
238                               MachineBasicBlock *MBB,
239                               MachineBasicBlock *SuccToSinkTo,
240                               AllSuccsCache &AllSuccessors);
241 
242     bool PerformTrivialForwardCoalescing(MachineInstr &MI,
243                                          MachineBasicBlock *MBB);
244 
245     SmallVector<MachineBasicBlock *, 4> &
246     GetAllSortedSuccessors(MachineInstr &MI, MachineBasicBlock *MBB,
247                            AllSuccsCache &AllSuccessors) const;
248 
249     std::vector<unsigned> &getBBRegisterPressure(MachineBasicBlock &MBB);
250   };
251 
252 } // end anonymous namespace
253 
254 char MachineSinking::ID = 0;
255 
256 char &llvm::MachineSinkingID = MachineSinking::ID;
257 
258 INITIALIZE_PASS_BEGIN(MachineSinking, DEBUG_TYPE,
259                       "Machine code sinking", false, false)
260 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
261 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
262 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
263 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
264 INITIALIZE_PASS_END(MachineSinking, DEBUG_TYPE,
265                     "Machine code sinking", false, false)
266 
267 bool MachineSinking::PerformTrivialForwardCoalescing(MachineInstr &MI,
268                                                      MachineBasicBlock *MBB) {
269   if (!MI.isCopy())
270     return false;
271 
272   Register SrcReg = MI.getOperand(1).getReg();
273   Register DstReg = MI.getOperand(0).getReg();
274   if (!Register::isVirtualRegister(SrcReg) ||
275       !Register::isVirtualRegister(DstReg) || !MRI->hasOneNonDBGUse(SrcReg))
276     return false;
277 
278   const TargetRegisterClass *SRC = MRI->getRegClass(SrcReg);
279   const TargetRegisterClass *DRC = MRI->getRegClass(DstReg);
280   if (SRC != DRC)
281     return false;
282 
283   MachineInstr *DefMI = MRI->getVRegDef(SrcReg);
284   if (DefMI->isCopyLike())
285     return false;
286   LLVM_DEBUG(dbgs() << "Coalescing: " << *DefMI);
287   LLVM_DEBUG(dbgs() << "*** to: " << MI);
288   MRI->replaceRegWith(DstReg, SrcReg);
289   MI.eraseFromParent();
290 
291   // Conservatively, clear any kill flags, since it's possible that they are no
292   // longer correct.
293   MRI->clearKillFlags(SrcReg);
294 
295   ++NumCoalesces;
296   return true;
297 }
298 
299 /// AllUsesDominatedByBlock - Return true if all uses of the specified register
300 /// occur in blocks dominated by the specified block. If any use is in the
301 /// definition block, then return false since it is never legal to move def
302 /// after uses.
303 bool MachineSinking::AllUsesDominatedByBlock(Register Reg,
304                                              MachineBasicBlock *MBB,
305                                              MachineBasicBlock *DefMBB,
306                                              bool &BreakPHIEdge,
307                                              bool &LocalUse) const {
308   assert(Register::isVirtualRegister(Reg) && "Only makes sense for vregs");
309 
310   // Ignore debug uses because debug info doesn't affect the code.
311   if (MRI->use_nodbg_empty(Reg))
312     return true;
313 
314   // BreakPHIEdge is true if all the uses are in the successor MBB being sunken
315   // into and they are all PHI nodes. In this case, machine-sink must break
316   // the critical edge first. e.g.
317   //
318   // %bb.1:
319   //   Predecessors according to CFG: %bb.0
320   //     ...
321   //     %def = DEC64_32r %x, implicit-def dead %eflags
322   //     ...
323   //     JE_4 <%bb.37>, implicit %eflags
324   //   Successors according to CFG: %bb.37 %bb.2
325   //
326   // %bb.2:
327   //     %p = PHI %y, %bb.0, %def, %bb.1
328   if (all_of(MRI->use_nodbg_operands(Reg), [&](MachineOperand &MO) {
329         MachineInstr *UseInst = MO.getParent();
330         unsigned OpNo = UseInst->getOperandNo(&MO);
331         MachineBasicBlock *UseBlock = UseInst->getParent();
332         return UseBlock == MBB && UseInst->isPHI() &&
333                UseInst->getOperand(OpNo + 1).getMBB() == DefMBB;
334       })) {
335     BreakPHIEdge = true;
336     return true;
337   }
338 
339   for (MachineOperand &MO : MRI->use_nodbg_operands(Reg)) {
340     // Determine the block of the use.
341     MachineInstr *UseInst = MO.getParent();
342     unsigned OpNo = &MO - &UseInst->getOperand(0);
343     MachineBasicBlock *UseBlock = UseInst->getParent();
344     if (UseInst->isPHI()) {
345       // PHI nodes use the operand in the predecessor block, not the block with
346       // the PHI.
347       UseBlock = UseInst->getOperand(OpNo+1).getMBB();
348     } else if (UseBlock == DefMBB) {
349       LocalUse = true;
350       return false;
351     }
352 
353     // Check that it dominates.
354     if (!DT->dominates(MBB, UseBlock))
355       return false;
356   }
357 
358   return true;
359 }
360 
361 /// Return true if this machine instruction loads from global offset table or
362 /// constant pool.
363 static bool mayLoadFromGOTOrConstantPool(MachineInstr &MI) {
364   assert(MI.mayLoad() && "Expected MI that loads!");
365 
366   // If we lost memory operands, conservatively assume that the instruction
367   // reads from everything..
368   if (MI.memoperands_empty())
369     return true;
370 
371   for (MachineMemOperand *MemOp : MI.memoperands())
372     if (const PseudoSourceValue *PSV = MemOp->getPseudoValue())
373       if (PSV->isGOT() || PSV->isConstantPool())
374         return true;
375 
376   return false;
377 }
378 
379 void MachineSinking::FindLoopSinkCandidates(MachineLoop *L, MachineBasicBlock *BB,
380     SmallVectorImpl<MachineInstr *> &Candidates) {
381   for (auto &MI : *BB) {
382     LLVM_DEBUG(dbgs() << "LoopSink: Analysing candidate: " << MI);
383     if (!TII->shouldSink(MI)) {
384       LLVM_DEBUG(dbgs() << "LoopSink: Instruction not a candidate for this "
385                            "target\n");
386       continue;
387     }
388     if (!L->isLoopInvariant(MI)) {
389       LLVM_DEBUG(dbgs() << "LoopSink: Instruction is not loop invariant\n");
390       continue;
391     }
392     bool DontMoveAcrossStore = true;
393     if (!MI.isSafeToMove(AA, DontMoveAcrossStore)) {
394       LLVM_DEBUG(dbgs() << "LoopSink: Instruction not safe to move.\n");
395       continue;
396     }
397     if (MI.mayLoad() && !mayLoadFromGOTOrConstantPool(MI)) {
398       LLVM_DEBUG(dbgs() << "LoopSink: Dont sink GOT or constant pool loads\n");
399       continue;
400     }
401     if (MI.isConvergent())
402       continue;
403 
404     const MachineOperand &MO = MI.getOperand(0);
405     if (!MO.isReg() || !MO.getReg() || !MO.isDef())
406       continue;
407     if (!MRI->hasOneDef(MO.getReg()))
408       continue;
409 
410     LLVM_DEBUG(dbgs() << "LoopSink: Instruction added as candidate.\n");
411     Candidates.push_back(&MI);
412   }
413 }
414 
415 bool MachineSinking::runOnMachineFunction(MachineFunction &MF) {
416   if (skipFunction(MF.getFunction()))
417     return false;
418 
419   LLVM_DEBUG(dbgs() << "******** Machine Sinking ********\n");
420 
421   TII = MF.getSubtarget().getInstrInfo();
422   TRI = MF.getSubtarget().getRegisterInfo();
423   MRI = &MF.getRegInfo();
424   DT = &getAnalysis<MachineDominatorTree>();
425   PDT = &getAnalysis<MachinePostDominatorTree>();
426   LI = &getAnalysis<MachineLoopInfo>();
427   MBFI = UseBlockFreqInfo ? &getAnalysis<MachineBlockFrequencyInfo>() : nullptr;
428   MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
429   AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
430   RegClassInfo.runOnMachineFunction(MF);
431 
432   bool EverMadeChange = false;
433 
434   while (true) {
435     bool MadeChange = false;
436 
437     // Process all basic blocks.
438     CEBCandidates.clear();
439     ToSplit.clear();
440     for (auto &MBB: MF)
441       MadeChange |= ProcessBlock(MBB);
442 
443     // If we have anything we marked as toSplit, split it now.
444     for (auto &Pair : ToSplit) {
445       auto NewSucc = Pair.first->SplitCriticalEdge(Pair.second, *this);
446       if (NewSucc != nullptr) {
447         LLVM_DEBUG(dbgs() << " *** Splitting critical edge: "
448                           << printMBBReference(*Pair.first) << " -- "
449                           << printMBBReference(*NewSucc) << " -- "
450                           << printMBBReference(*Pair.second) << '\n');
451         if (MBFI)
452           MBFI->onEdgeSplit(*Pair.first, *NewSucc, *MBPI);
453 
454         MadeChange = true;
455         ++NumSplit;
456       } else
457         LLVM_DEBUG(dbgs() << " *** Not legal to break critical edge\n");
458     }
459     // If this iteration over the code changed anything, keep iterating.
460     if (!MadeChange) break;
461     EverMadeChange = true;
462   }
463 
464   if (SinkInstsIntoLoop) {
465     SmallVector<MachineLoop *, 8> Loops(LI->begin(), LI->end());
466     for (auto *L : Loops) {
467       MachineBasicBlock *Preheader = LI->findLoopPreheader(L);
468       if (!Preheader) {
469         LLVM_DEBUG(dbgs() << "LoopSink: Can't find preheader\n");
470         continue;
471       }
472       SmallVector<MachineInstr *, 8> Candidates;
473       FindLoopSinkCandidates(L, Preheader, Candidates);
474 
475       // Walk the candidates in reverse order so that we start with the use
476       // of a def-use chain, if there is any.
477       // TODO: Sort the candidates using a cost-model.
478       unsigned i = 0;
479       for (auto It = Candidates.rbegin(); It != Candidates.rend(); ++It) {
480         if (i++ == SinkIntoLoopLimit) {
481           LLVM_DEBUG(dbgs() << "LoopSink:   Limit reached of instructions to "
482                                "be analysed.");
483           break;
484         }
485 
486         MachineInstr *I = *It;
487         if (!SinkIntoLoop(L, *I))
488           break;
489         EverMadeChange = true;
490         ++NumLoopSunk;
491       }
492     }
493   }
494 
495   HasStoreCache.clear();
496   StoreInstrCache.clear();
497 
498   // Now clear any kill flags for recorded registers.
499   for (auto I : RegsToClearKillFlags)
500     MRI->clearKillFlags(I);
501   RegsToClearKillFlags.clear();
502 
503   return EverMadeChange;
504 }
505 
506 bool MachineSinking::ProcessBlock(MachineBasicBlock &MBB) {
507   // Can't sink anything out of a block that has less than two successors.
508   if (MBB.succ_size() <= 1 || MBB.empty()) return false;
509 
510   // Don't bother sinking code out of unreachable blocks. In addition to being
511   // unprofitable, it can also lead to infinite looping, because in an
512   // unreachable loop there may be nowhere to stop.
513   if (!DT->isReachableFromEntry(&MBB)) return false;
514 
515   bool MadeChange = false;
516 
517   // Cache all successors, sorted by frequency info and loop depth.
518   AllSuccsCache AllSuccessors;
519 
520   // Walk the basic block bottom-up.  Remember if we saw a store.
521   MachineBasicBlock::iterator I = MBB.end();
522   --I;
523   bool ProcessedBegin, SawStore = false;
524   do {
525     MachineInstr &MI = *I;  // The instruction to sink.
526 
527     // Predecrement I (if it's not begin) so that it isn't invalidated by
528     // sinking.
529     ProcessedBegin = I == MBB.begin();
530     if (!ProcessedBegin)
531       --I;
532 
533     if (MI.isDebugOrPseudoInstr()) {
534       if (MI.isDebugValue())
535         ProcessDbgInst(MI);
536       continue;
537     }
538 
539     bool Joined = PerformTrivialForwardCoalescing(MI, &MBB);
540     if (Joined) {
541       MadeChange = true;
542       continue;
543     }
544 
545     if (SinkInstruction(MI, SawStore, AllSuccessors)) {
546       ++NumSunk;
547       MadeChange = true;
548     }
549 
550     // If we just processed the first instruction in the block, we're done.
551   } while (!ProcessedBegin);
552 
553   SeenDbgUsers.clear();
554   SeenDbgVars.clear();
555   // recalculate the bb register pressure after sinking one BB.
556   CachedRegisterPressure.clear();
557 
558   return MadeChange;
559 }
560 
561 void MachineSinking::ProcessDbgInst(MachineInstr &MI) {
562   // When we see DBG_VALUEs for registers, record any vreg it reads, so that
563   // we know what to sink if the vreg def sinks.
564   assert(MI.isDebugValue() && "Expected DBG_VALUE for processing");
565 
566   DebugVariable Var(MI.getDebugVariable(), MI.getDebugExpression(),
567                     MI.getDebugLoc()->getInlinedAt());
568   bool SeenBefore = SeenDbgVars.contains(Var);
569 
570   for (MachineOperand &MO : MI.debug_operands()) {
571     if (MO.isReg() && MO.getReg().isVirtual())
572       SeenDbgUsers[MO.getReg()].push_back(SeenDbgUser(&MI, SeenBefore));
573   }
574 
575   // Record the variable for any DBG_VALUE, to avoid re-ordering any of them.
576   SeenDbgVars.insert(Var);
577 }
578 
579 bool MachineSinking::isWorthBreakingCriticalEdge(MachineInstr &MI,
580                                                  MachineBasicBlock *From,
581                                                  MachineBasicBlock *To) {
582   // FIXME: Need much better heuristics.
583 
584   // If the pass has already considered breaking this edge (during this pass
585   // through the function), then let's go ahead and break it. This means
586   // sinking multiple "cheap" instructions into the same block.
587   if (!CEBCandidates.insert(std::make_pair(From, To)).second)
588     return true;
589 
590   if (!MI.isCopy() && !TII->isAsCheapAsAMove(MI))
591     return true;
592 
593   if (From->isSuccessor(To) && MBPI->getEdgeProbability(From, To) <=
594       BranchProbability(SplitEdgeProbabilityThreshold, 100))
595     return true;
596 
597   // MI is cheap, we probably don't want to break the critical edge for it.
598   // However, if this would allow some definitions of its source operands
599   // to be sunk then it's probably worth it.
600   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
601     const MachineOperand &MO = MI.getOperand(i);
602     if (!MO.isReg() || !MO.isUse())
603       continue;
604     Register Reg = MO.getReg();
605     if (Reg == 0)
606       continue;
607 
608     // We don't move live definitions of physical registers,
609     // so sinking their uses won't enable any opportunities.
610     if (Register::isPhysicalRegister(Reg))
611       continue;
612 
613     // If this instruction is the only user of a virtual register,
614     // check if breaking the edge will enable sinking
615     // both this instruction and the defining instruction.
616     if (MRI->hasOneNonDBGUse(Reg)) {
617       // If the definition resides in same MBB,
618       // claim it's likely we can sink these together.
619       // If definition resides elsewhere, we aren't
620       // blocking it from being sunk so don't break the edge.
621       MachineInstr *DefMI = MRI->getVRegDef(Reg);
622       if (DefMI->getParent() == MI.getParent())
623         return true;
624     }
625   }
626 
627   return false;
628 }
629 
630 bool MachineSinking::PostponeSplitCriticalEdge(MachineInstr &MI,
631                                                MachineBasicBlock *FromBB,
632                                                MachineBasicBlock *ToBB,
633                                                bool BreakPHIEdge) {
634   if (!isWorthBreakingCriticalEdge(MI, FromBB, ToBB))
635     return false;
636 
637   // Avoid breaking back edge. From == To means backedge for single BB loop.
638   if (!SplitEdges || FromBB == ToBB)
639     return false;
640 
641   // Check for backedges of more "complex" loops.
642   if (LI->getLoopFor(FromBB) == LI->getLoopFor(ToBB) &&
643       LI->isLoopHeader(ToBB))
644     return false;
645 
646   // It's not always legal to break critical edges and sink the computation
647   // to the edge.
648   //
649   // %bb.1:
650   // v1024
651   // Beq %bb.3
652   // <fallthrough>
653   // %bb.2:
654   // ... no uses of v1024
655   // <fallthrough>
656   // %bb.3:
657   // ...
658   //       = v1024
659   //
660   // If %bb.1 -> %bb.3 edge is broken and computation of v1024 is inserted:
661   //
662   // %bb.1:
663   // ...
664   // Bne %bb.2
665   // %bb.4:
666   // v1024 =
667   // B %bb.3
668   // %bb.2:
669   // ... no uses of v1024
670   // <fallthrough>
671   // %bb.3:
672   // ...
673   //       = v1024
674   //
675   // This is incorrect since v1024 is not computed along the %bb.1->%bb.2->%bb.3
676   // flow. We need to ensure the new basic block where the computation is
677   // sunk to dominates all the uses.
678   // It's only legal to break critical edge and sink the computation to the
679   // new block if all the predecessors of "To", except for "From", are
680   // not dominated by "From". Given SSA property, this means these
681   // predecessors are dominated by "To".
682   //
683   // There is no need to do this check if all the uses are PHI nodes. PHI
684   // sources are only defined on the specific predecessor edges.
685   if (!BreakPHIEdge) {
686     for (MachineBasicBlock::pred_iterator PI = ToBB->pred_begin(),
687            E = ToBB->pred_end(); PI != E; ++PI) {
688       if (*PI == FromBB)
689         continue;
690       if (!DT->dominates(ToBB, *PI))
691         return false;
692     }
693   }
694 
695   ToSplit.insert(std::make_pair(FromBB, ToBB));
696 
697   return true;
698 }
699 
700 std::vector<unsigned> &
701 MachineSinking::getBBRegisterPressure(MachineBasicBlock &MBB) {
702   // Currently to save compiling time, MBB's register pressure will not change
703   // in one ProcessBlock iteration because of CachedRegisterPressure. but MBB's
704   // register pressure is changed after sinking any instructions into it.
705   // FIXME: need a accurate and cheap register pressure estiminate model here.
706   auto RP = CachedRegisterPressure.find(&MBB);
707   if (RP != CachedRegisterPressure.end())
708     return RP->second;
709 
710   RegionPressure Pressure;
711   RegPressureTracker RPTracker(Pressure);
712 
713   // Initialize the register pressure tracker.
714   RPTracker.init(MBB.getParent(), &RegClassInfo, nullptr, &MBB, MBB.end(),
715                  /*TrackLaneMasks*/ false, /*TrackUntiedDefs=*/true);
716 
717   for (MachineBasicBlock::iterator MII = MBB.instr_end(),
718                                    MIE = MBB.instr_begin();
719        MII != MIE; --MII) {
720     MachineInstr &MI = *std::prev(MII);
721     if (MI.isDebugInstr() || MI.isPseudoProbe())
722       continue;
723     RegisterOperands RegOpers;
724     RegOpers.collect(MI, *TRI, *MRI, false, false);
725     RPTracker.recedeSkipDebugValues();
726     assert(&*RPTracker.getPos() == &MI && "RPTracker sync error!");
727     RPTracker.recede(RegOpers);
728   }
729 
730   RPTracker.closeRegion();
731   auto It = CachedRegisterPressure.insert(
732       std::make_pair(&MBB, RPTracker.getPressure().MaxSetPressure));
733   return It.first->second;
734 }
735 
736 /// isProfitableToSinkTo - Return true if it is profitable to sink MI.
737 bool MachineSinking::isProfitableToSinkTo(Register Reg, MachineInstr &MI,
738                                           MachineBasicBlock *MBB,
739                                           MachineBasicBlock *SuccToSinkTo,
740                                           AllSuccsCache &AllSuccessors) {
741   assert (SuccToSinkTo && "Invalid SinkTo Candidate BB");
742 
743   if (MBB == SuccToSinkTo)
744     return false;
745 
746   // It is profitable if SuccToSinkTo does not post dominate current block.
747   if (!PDT->dominates(SuccToSinkTo, MBB))
748     return true;
749 
750   // It is profitable to sink an instruction from a deeper loop to a shallower
751   // loop, even if the latter post-dominates the former (PR21115).
752   if (LI->getLoopDepth(MBB) > LI->getLoopDepth(SuccToSinkTo))
753     return true;
754 
755   // Check if only use in post dominated block is PHI instruction.
756   bool NonPHIUse = false;
757   for (MachineInstr &UseInst : MRI->use_nodbg_instructions(Reg)) {
758     MachineBasicBlock *UseBlock = UseInst.getParent();
759     if (UseBlock == SuccToSinkTo && !UseInst.isPHI())
760       NonPHIUse = true;
761   }
762   if (!NonPHIUse)
763     return true;
764 
765   // If SuccToSinkTo post dominates then also it may be profitable if MI
766   // can further profitably sinked into another block in next round.
767   bool BreakPHIEdge = false;
768   // FIXME - If finding successor is compile time expensive then cache results.
769   if (MachineBasicBlock *MBB2 =
770           FindSuccToSinkTo(MI, SuccToSinkTo, BreakPHIEdge, AllSuccessors))
771     return isProfitableToSinkTo(Reg, MI, SuccToSinkTo, MBB2, AllSuccessors);
772 
773   MachineLoop *ML = LI->getLoopFor(MBB);
774 
775   // If the instruction is not inside a loop, it is not profitable to sink MI to
776   // a post dominate block SuccToSinkTo.
777   if (!ML)
778     return false;
779 
780   auto isRegisterPressureSetExceedLimit = [&](const TargetRegisterClass *RC) {
781     unsigned Weight = TRI->getRegClassWeight(RC).RegWeight;
782     const int *PS = TRI->getRegClassPressureSets(RC);
783     // Get register pressure for block SuccToSinkTo.
784     std::vector<unsigned> BBRegisterPressure =
785         getBBRegisterPressure(*SuccToSinkTo);
786     for (; *PS != -1; PS++)
787       // check if any register pressure set exceeds limit in block SuccToSinkTo
788       // after sinking.
789       if (Weight + BBRegisterPressure[*PS] >=
790           TRI->getRegPressureSetLimit(*MBB->getParent(), *PS))
791         return true;
792     return false;
793   };
794 
795   // If this instruction is inside a loop and sinking this instruction can make
796   // more registers live range shorten, it is still prifitable.
797   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
798     const MachineOperand &MO = MI.getOperand(i);
799     // Ignore non-register operands.
800     if (!MO.isReg())
801       continue;
802     Register Reg = MO.getReg();
803     if (Reg == 0)
804       continue;
805 
806     // Don't handle physical register.
807     if (Register::isPhysicalRegister(Reg))
808       return false;
809 
810     // Users for the defs are all dominated by SuccToSinkTo.
811     if (MO.isDef()) {
812       // This def register's live range is shortened after sinking.
813       bool LocalUse = false;
814       if (!AllUsesDominatedByBlock(Reg, SuccToSinkTo, MBB, BreakPHIEdge,
815                                    LocalUse))
816         return false;
817     } else {
818       MachineInstr *DefMI = MRI->getVRegDef(Reg);
819       // DefMI is defined outside of loop. There should be no live range
820       // impact for this operand. Defination outside of loop means:
821       // 1: defination is outside of loop.
822       // 2: defination is in this loop, but it is a PHI in the loop header.
823       if (LI->getLoopFor(DefMI->getParent()) != ML ||
824           (DefMI->isPHI() && LI->isLoopHeader(DefMI->getParent())))
825         continue;
826       // The DefMI is defined inside the loop.
827       // If sinking this operand makes some register pressure set exceed limit,
828       // it is not profitable.
829       if (isRegisterPressureSetExceedLimit(MRI->getRegClass(Reg))) {
830         LLVM_DEBUG(dbgs() << "register pressure exceed limit, not profitable.");
831         return false;
832       }
833     }
834   }
835 
836   // If MI is in loop and all its operands are alive across the whole loop or if
837   // no operand sinking make register pressure set exceed limit, it is
838   // profitable to sink MI.
839   return true;
840 }
841 
842 /// Get the sorted sequence of successors for this MachineBasicBlock, possibly
843 /// computing it if it was not already cached.
844 SmallVector<MachineBasicBlock *, 4> &
845 MachineSinking::GetAllSortedSuccessors(MachineInstr &MI, MachineBasicBlock *MBB,
846                                        AllSuccsCache &AllSuccessors) const {
847   // Do we have the sorted successors in cache ?
848   auto Succs = AllSuccessors.find(MBB);
849   if (Succs != AllSuccessors.end())
850     return Succs->second;
851 
852   SmallVector<MachineBasicBlock *, 4> AllSuccs(MBB->successors());
853 
854   // Handle cases where sinking can happen but where the sink point isn't a
855   // successor. For example:
856   //
857   //   x = computation
858   //   if () {} else {}
859   //   use x
860   //
861   for (MachineDomTreeNode *DTChild : DT->getNode(MBB)->children()) {
862     // DomTree children of MBB that have MBB as immediate dominator are added.
863     if (DTChild->getIDom()->getBlock() == MI.getParent() &&
864         // Skip MBBs already added to the AllSuccs vector above.
865         !MBB->isSuccessor(DTChild->getBlock()))
866       AllSuccs.push_back(DTChild->getBlock());
867   }
868 
869   // Sort Successors according to their loop depth or block frequency info.
870   llvm::stable_sort(
871       AllSuccs, [this](const MachineBasicBlock *L, const MachineBasicBlock *R) {
872         uint64_t LHSFreq = MBFI ? MBFI->getBlockFreq(L).getFrequency() : 0;
873         uint64_t RHSFreq = MBFI ? MBFI->getBlockFreq(R).getFrequency() : 0;
874         bool HasBlockFreq = LHSFreq != 0 && RHSFreq != 0;
875         return HasBlockFreq ? LHSFreq < RHSFreq
876                             : LI->getLoopDepth(L) < LI->getLoopDepth(R);
877       });
878 
879   auto it = AllSuccessors.insert(std::make_pair(MBB, AllSuccs));
880 
881   return it.first->second;
882 }
883 
884 /// FindSuccToSinkTo - Find a successor to sink this instruction to.
885 MachineBasicBlock *
886 MachineSinking::FindSuccToSinkTo(MachineInstr &MI, MachineBasicBlock *MBB,
887                                  bool &BreakPHIEdge,
888                                  AllSuccsCache &AllSuccessors) {
889   assert (MBB && "Invalid MachineBasicBlock!");
890 
891   // Loop over all the operands of the specified instruction.  If there is
892   // anything we can't handle, bail out.
893 
894   // SuccToSinkTo - This is the successor to sink this instruction to, once we
895   // decide.
896   MachineBasicBlock *SuccToSinkTo = nullptr;
897   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
898     const MachineOperand &MO = MI.getOperand(i);
899     if (!MO.isReg()) continue;  // Ignore non-register operands.
900 
901     Register Reg = MO.getReg();
902     if (Reg == 0) continue;
903 
904     if (Register::isPhysicalRegister(Reg)) {
905       if (MO.isUse()) {
906         // If the physreg has no defs anywhere, it's just an ambient register
907         // and we can freely move its uses. Alternatively, if it's allocatable,
908         // it could get allocated to something with a def during allocation.
909         if (!MRI->isConstantPhysReg(Reg))
910           return nullptr;
911       } else if (!MO.isDead()) {
912         // A def that isn't dead. We can't move it.
913         return nullptr;
914       }
915     } else {
916       // Virtual register uses are always safe to sink.
917       if (MO.isUse()) continue;
918 
919       // If it's not safe to move defs of the register class, then abort.
920       if (!TII->isSafeToMoveRegClassDefs(MRI->getRegClass(Reg)))
921         return nullptr;
922 
923       // Virtual register defs can only be sunk if all their uses are in blocks
924       // dominated by one of the successors.
925       if (SuccToSinkTo) {
926         // If a previous operand picked a block to sink to, then this operand
927         // must be sinkable to the same block.
928         bool LocalUse = false;
929         if (!AllUsesDominatedByBlock(Reg, SuccToSinkTo, MBB,
930                                      BreakPHIEdge, LocalUse))
931           return nullptr;
932 
933         continue;
934       }
935 
936       // Otherwise, we should look at all the successors and decide which one
937       // we should sink to. If we have reliable block frequency information
938       // (frequency != 0) available, give successors with smaller frequencies
939       // higher priority, otherwise prioritize smaller loop depths.
940       for (MachineBasicBlock *SuccBlock :
941            GetAllSortedSuccessors(MI, MBB, AllSuccessors)) {
942         bool LocalUse = false;
943         if (AllUsesDominatedByBlock(Reg, SuccBlock, MBB,
944                                     BreakPHIEdge, LocalUse)) {
945           SuccToSinkTo = SuccBlock;
946           break;
947         }
948         if (LocalUse)
949           // Def is used locally, it's never safe to move this def.
950           return nullptr;
951       }
952 
953       // If we couldn't find a block to sink to, ignore this instruction.
954       if (!SuccToSinkTo)
955         return nullptr;
956       if (!isProfitableToSinkTo(Reg, MI, MBB, SuccToSinkTo, AllSuccessors))
957         return nullptr;
958     }
959   }
960 
961   // It is not possible to sink an instruction into its own block.  This can
962   // happen with loops.
963   if (MBB == SuccToSinkTo)
964     return nullptr;
965 
966   // It's not safe to sink instructions to EH landing pad. Control flow into
967   // landing pad is implicitly defined.
968   if (SuccToSinkTo && SuccToSinkTo->isEHPad())
969     return nullptr;
970 
971   // It ought to be okay to sink instructions into an INLINEASM_BR target, but
972   // only if we make sure that MI occurs _before_ an INLINEASM_BR instruction in
973   // the source block (which this code does not yet do). So for now, forbid
974   // doing so.
975   if (SuccToSinkTo && SuccToSinkTo->isInlineAsmBrIndirectTarget())
976     return nullptr;
977 
978   return SuccToSinkTo;
979 }
980 
981 /// Return true if MI is likely to be usable as a memory operation by the
982 /// implicit null check optimization.
983 ///
984 /// This is a "best effort" heuristic, and should not be relied upon for
985 /// correctness.  This returning true does not guarantee that the implicit null
986 /// check optimization is legal over MI, and this returning false does not
987 /// guarantee MI cannot possibly be used to do a null check.
988 static bool SinkingPreventsImplicitNullCheck(MachineInstr &MI,
989                                              const TargetInstrInfo *TII,
990                                              const TargetRegisterInfo *TRI) {
991   using MachineBranchPredicate = TargetInstrInfo::MachineBranchPredicate;
992 
993   auto *MBB = MI.getParent();
994   if (MBB->pred_size() != 1)
995     return false;
996 
997   auto *PredMBB = *MBB->pred_begin();
998   auto *PredBB = PredMBB->getBasicBlock();
999 
1000   // Frontends that don't use implicit null checks have no reason to emit
1001   // branches with make.implicit metadata, and this function should always
1002   // return false for them.
1003   if (!PredBB ||
1004       !PredBB->getTerminator()->getMetadata(LLVMContext::MD_make_implicit))
1005     return false;
1006 
1007   const MachineOperand *BaseOp;
1008   int64_t Offset;
1009   bool OffsetIsScalable;
1010   if (!TII->getMemOperandWithOffset(MI, BaseOp, Offset, OffsetIsScalable, TRI))
1011     return false;
1012 
1013   if (!BaseOp->isReg())
1014     return false;
1015 
1016   if (!(MI.mayLoad() && !MI.isPredicable()))
1017     return false;
1018 
1019   MachineBranchPredicate MBP;
1020   if (TII->analyzeBranchPredicate(*PredMBB, MBP, false))
1021     return false;
1022 
1023   return MBP.LHS.isReg() && MBP.RHS.isImm() && MBP.RHS.getImm() == 0 &&
1024          (MBP.Predicate == MachineBranchPredicate::PRED_NE ||
1025           MBP.Predicate == MachineBranchPredicate::PRED_EQ) &&
1026          MBP.LHS.getReg() == BaseOp->getReg();
1027 }
1028 
1029 /// If the sunk instruction is a copy, try to forward the copy instead of
1030 /// leaving an 'undef' DBG_VALUE in the original location. Don't do this if
1031 /// there's any subregister weirdness involved. Returns true if copy
1032 /// propagation occurred.
1033 static bool attemptDebugCopyProp(MachineInstr &SinkInst, MachineInstr &DbgMI,
1034                                  Register Reg) {
1035   const MachineRegisterInfo &MRI = SinkInst.getMF()->getRegInfo();
1036   const TargetInstrInfo &TII = *SinkInst.getMF()->getSubtarget().getInstrInfo();
1037 
1038   // Copy DBG_VALUE operand and set the original to undef. We then check to
1039   // see whether this is something that can be copy-forwarded. If it isn't,
1040   // continue around the loop.
1041 
1042   const MachineOperand *SrcMO = nullptr, *DstMO = nullptr;
1043   auto CopyOperands = TII.isCopyInstr(SinkInst);
1044   if (!CopyOperands)
1045     return false;
1046   SrcMO = CopyOperands->Source;
1047   DstMO = CopyOperands->Destination;
1048 
1049   // Check validity of forwarding this copy.
1050   bool PostRA = MRI.getNumVirtRegs() == 0;
1051 
1052   // Trying to forward between physical and virtual registers is too hard.
1053   if (Reg.isVirtual() != SrcMO->getReg().isVirtual())
1054     return false;
1055 
1056   // Only try virtual register copy-forwarding before regalloc, and physical
1057   // register copy-forwarding after regalloc.
1058   bool arePhysRegs = !Reg.isVirtual();
1059   if (arePhysRegs != PostRA)
1060     return false;
1061 
1062   // Pre-regalloc, only forward if all subregisters agree (or there are no
1063   // subregs at all). More analysis might recover some forwardable copies.
1064   if (!PostRA)
1065     for (auto &DbgMO : DbgMI.getDebugOperandsForReg(Reg))
1066       if (DbgMO.getSubReg() != SrcMO->getSubReg() ||
1067           DbgMO.getSubReg() != DstMO->getSubReg())
1068         return false;
1069 
1070   // Post-regalloc, we may be sinking a DBG_VALUE of a sub or super-register
1071   // of this copy. Only forward the copy if the DBG_VALUE operand exactly
1072   // matches the copy destination.
1073   if (PostRA && Reg != DstMO->getReg())
1074     return false;
1075 
1076   for (auto &DbgMO : DbgMI.getDebugOperandsForReg(Reg)) {
1077     DbgMO.setReg(SrcMO->getReg());
1078     DbgMO.setSubReg(SrcMO->getSubReg());
1079   }
1080   return true;
1081 }
1082 
1083 using MIRegs = std::pair<MachineInstr *, SmallVector<unsigned, 2>>;
1084 /// Sink an instruction and its associated debug instructions.
1085 static void performSink(MachineInstr &MI, MachineBasicBlock &SuccToSinkTo,
1086                         MachineBasicBlock::iterator InsertPos,
1087                         SmallVectorImpl<MIRegs> &DbgValuesToSink) {
1088 
1089   // If we cannot find a location to use (merge with), then we erase the debug
1090   // location to prevent debug-info driven tools from potentially reporting
1091   // wrong location information.
1092   if (!SuccToSinkTo.empty() && InsertPos != SuccToSinkTo.end())
1093     MI.setDebugLoc(DILocation::getMergedLocation(MI.getDebugLoc(),
1094                                                  InsertPos->getDebugLoc()));
1095   else
1096     MI.setDebugLoc(DebugLoc());
1097 
1098   // Move the instruction.
1099   MachineBasicBlock *ParentBlock = MI.getParent();
1100   SuccToSinkTo.splice(InsertPos, ParentBlock, MI,
1101                       ++MachineBasicBlock::iterator(MI));
1102 
1103   // Sink a copy of debug users to the insert position. Mark the original
1104   // DBG_VALUE location as 'undef', indicating that any earlier variable
1105   // location should be terminated as we've optimised away the value at this
1106   // point.
1107   for (auto DbgValueToSink : DbgValuesToSink) {
1108     MachineInstr *DbgMI = DbgValueToSink.first;
1109     MachineInstr *NewDbgMI = DbgMI->getMF()->CloneMachineInstr(DbgMI);
1110     SuccToSinkTo.insert(InsertPos, NewDbgMI);
1111 
1112     bool PropagatedAllSunkOps = true;
1113     for (unsigned Reg : DbgValueToSink.second) {
1114       if (DbgMI->hasDebugOperandForReg(Reg)) {
1115         if (!attemptDebugCopyProp(MI, *DbgMI, Reg)) {
1116           PropagatedAllSunkOps = false;
1117           break;
1118         }
1119       }
1120     }
1121     if (!PropagatedAllSunkOps)
1122       DbgMI->setDebugValueUndef();
1123   }
1124 }
1125 
1126 /// hasStoreBetween - check if there is store betweeen straight line blocks From
1127 /// and To.
1128 bool MachineSinking::hasStoreBetween(MachineBasicBlock *From,
1129                                      MachineBasicBlock *To, MachineInstr &MI) {
1130   // Make sure From and To are in straight line which means From dominates To
1131   // and To post dominates From.
1132   if (!DT->dominates(From, To) || !PDT->dominates(To, From))
1133     return true;
1134 
1135   auto BlockPair = std::make_pair(From, To);
1136 
1137   // Does these two blocks pair be queried before and have a definite cached
1138   // result?
1139   if (HasStoreCache.find(BlockPair) != HasStoreCache.end())
1140     return HasStoreCache[BlockPair];
1141 
1142   if (StoreInstrCache.find(BlockPair) != StoreInstrCache.end())
1143     return llvm::any_of(StoreInstrCache[BlockPair], [&](MachineInstr *I) {
1144       return I->mayAlias(AA, MI, false);
1145     });
1146 
1147   bool SawStore = false;
1148   bool HasAliasedStore = false;
1149   DenseSet<MachineBasicBlock *> HandledBlocks;
1150   DenseSet<MachineBasicBlock *> HandledDomBlocks;
1151   // Go through all reachable blocks from From.
1152   for (MachineBasicBlock *BB : depth_first(From)) {
1153     // We insert the instruction at the start of block To, so no need to worry
1154     // about stores inside To.
1155     // Store in block From should be already considered when just enter function
1156     // SinkInstruction.
1157     if (BB == To || BB == From)
1158       continue;
1159 
1160     // We already handle this BB in previous iteration.
1161     if (HandledBlocks.count(BB))
1162       continue;
1163 
1164     HandledBlocks.insert(BB);
1165     // To post dominates BB, it must be a path from block From.
1166     if (PDT->dominates(To, BB)) {
1167       if (!HandledDomBlocks.count(BB))
1168         HandledDomBlocks.insert(BB);
1169 
1170       // If this BB is too big or the block number in straight line between From
1171       // and To is too big, stop searching to save compiling time.
1172       if (BB->size() > SinkLoadInstsPerBlockThreshold ||
1173           HandledDomBlocks.size() > SinkLoadBlocksThreshold) {
1174         for (auto *DomBB : HandledDomBlocks) {
1175           if (DomBB != BB && DT->dominates(DomBB, BB))
1176             HasStoreCache[std::make_pair(DomBB, To)] = true;
1177           else if(DomBB != BB && DT->dominates(BB, DomBB))
1178             HasStoreCache[std::make_pair(From, DomBB)] = true;
1179         }
1180         HasStoreCache[BlockPair] = true;
1181         return true;
1182       }
1183 
1184       for (MachineInstr &I : *BB) {
1185         // Treat as alias conservatively for a call or an ordered memory
1186         // operation.
1187         if (I.isCall() || I.hasOrderedMemoryRef()) {
1188           for (auto *DomBB : HandledDomBlocks) {
1189             if (DomBB != BB && DT->dominates(DomBB, BB))
1190               HasStoreCache[std::make_pair(DomBB, To)] = true;
1191             else if(DomBB != BB && DT->dominates(BB, DomBB))
1192               HasStoreCache[std::make_pair(From, DomBB)] = true;
1193           }
1194           HasStoreCache[BlockPair] = true;
1195           return true;
1196         }
1197 
1198         if (I.mayStore()) {
1199           SawStore = true;
1200           // We still have chance to sink MI if all stores between are not
1201           // aliased to MI.
1202           // Cache all store instructions, so that we don't need to go through
1203           // all From reachable blocks for next load instruction.
1204           if (I.mayAlias(AA, MI, false))
1205             HasAliasedStore = true;
1206           StoreInstrCache[BlockPair].push_back(&I);
1207         }
1208       }
1209     }
1210   }
1211   // If there is no store at all, cache the result.
1212   if (!SawStore)
1213     HasStoreCache[BlockPair] = false;
1214   return HasAliasedStore;
1215 }
1216 
1217 /// Sink instructions into loops if profitable. This especially tries to prevent
1218 /// register spills caused by register pressure if there is little to no
1219 /// overhead moving instructions into loops.
1220 bool MachineSinking::SinkIntoLoop(MachineLoop *L, MachineInstr &I) {
1221   LLVM_DEBUG(dbgs() << "LoopSink: Finding sink block for: " << I);
1222   MachineBasicBlock *Preheader = L->getLoopPreheader();
1223   assert(Preheader && "Loop sink needs a preheader block");
1224   MachineBasicBlock *SinkBlock = nullptr;
1225   bool CanSink = true;
1226   const MachineOperand &MO = I.getOperand(0);
1227 
1228   for (MachineInstr &MI : MRI->use_instructions(MO.getReg())) {
1229     LLVM_DEBUG(dbgs() << "LoopSink:   Analysing use: " << MI);
1230     if (!L->contains(&MI)) {
1231       LLVM_DEBUG(dbgs() << "LoopSink:   Use not in loop, can't sink.\n");
1232       CanSink = false;
1233       break;
1234     }
1235 
1236     // FIXME: Come up with a proper cost model that estimates whether sinking
1237     // the instruction (and thus possibly executing it on every loop
1238     // iteration) is more expensive than a register.
1239     // For now assumes that copies are cheap and thus almost always worth it.
1240     if (!MI.isCopy()) {
1241       LLVM_DEBUG(dbgs() << "LoopSink:   Use is not a copy\n");
1242       CanSink = false;
1243       break;
1244     }
1245     if (!SinkBlock) {
1246       SinkBlock = MI.getParent();
1247       LLVM_DEBUG(dbgs() << "LoopSink:   Setting sink block to: "
1248                         << printMBBReference(*SinkBlock) << "\n");
1249       continue;
1250     }
1251     SinkBlock = DT->findNearestCommonDominator(SinkBlock, MI.getParent());
1252     if (!SinkBlock) {
1253       LLVM_DEBUG(dbgs() << "LoopSink:   Can't find nearest dominator\n");
1254       CanSink = false;
1255       break;
1256     }
1257     LLVM_DEBUG(dbgs() << "LoopSink:   Setting nearest common dom block: " <<
1258                printMBBReference(*SinkBlock) << "\n");
1259   }
1260 
1261   if (!CanSink) {
1262     LLVM_DEBUG(dbgs() << "LoopSink: Can't sink instruction.\n");
1263     return false;
1264   }
1265   if (!SinkBlock) {
1266     LLVM_DEBUG(dbgs() << "LoopSink: Not sinking, can't find sink block.\n");
1267     return false;
1268   }
1269   if (SinkBlock == Preheader) {
1270     LLVM_DEBUG(dbgs() << "LoopSink: Not sinking, sink block is the preheader\n");
1271     return false;
1272   }
1273   if (SinkBlock->size() > SinkLoadInstsPerBlockThreshold) {
1274     LLVM_DEBUG(dbgs() << "LoopSink: Not Sinking, block too large to analyse.\n");
1275     return false;
1276   }
1277 
1278   LLVM_DEBUG(dbgs() << "LoopSink: Sinking instruction!\n");
1279   SinkBlock->splice(SinkBlock->getFirstNonPHI(), Preheader, I);
1280 
1281   // The instruction is moved from its basic block, so do not retain the
1282   // debug information.
1283   assert(!I.isDebugInstr() && "Should not sink debug inst");
1284   I.setDebugLoc(DebugLoc());
1285   return true;
1286 }
1287 
1288 /// SinkInstruction - Determine whether it is safe to sink the specified machine
1289 /// instruction out of its current block into a successor.
1290 bool MachineSinking::SinkInstruction(MachineInstr &MI, bool &SawStore,
1291                                      AllSuccsCache &AllSuccessors) {
1292   // Don't sink instructions that the target prefers not to sink.
1293   if (!TII->shouldSink(MI))
1294     return false;
1295 
1296   // Check if it's safe to move the instruction.
1297   if (!MI.isSafeToMove(AA, SawStore))
1298     return false;
1299 
1300   // Convergent operations may not be made control-dependent on additional
1301   // values.
1302   if (MI.isConvergent())
1303     return false;
1304 
1305   // Don't break implicit null checks.  This is a performance heuristic, and not
1306   // required for correctness.
1307   if (SinkingPreventsImplicitNullCheck(MI, TII, TRI))
1308     return false;
1309 
1310   // FIXME: This should include support for sinking instructions within the
1311   // block they are currently in to shorten the live ranges.  We often get
1312   // instructions sunk into the top of a large block, but it would be better to
1313   // also sink them down before their first use in the block.  This xform has to
1314   // be careful not to *increase* register pressure though, e.g. sinking
1315   // "x = y + z" down if it kills y and z would increase the live ranges of y
1316   // and z and only shrink the live range of x.
1317 
1318   bool BreakPHIEdge = false;
1319   MachineBasicBlock *ParentBlock = MI.getParent();
1320   MachineBasicBlock *SuccToSinkTo =
1321       FindSuccToSinkTo(MI, ParentBlock, BreakPHIEdge, AllSuccessors);
1322 
1323   // If there are no outputs, it must have side-effects.
1324   if (!SuccToSinkTo)
1325     return false;
1326 
1327   // If the instruction to move defines a dead physical register which is live
1328   // when leaving the basic block, don't move it because it could turn into a
1329   // "zombie" define of that preg. E.g., EFLAGS. (<rdar://problem/8030636>)
1330   for (unsigned I = 0, E = MI.getNumOperands(); I != E; ++I) {
1331     const MachineOperand &MO = MI.getOperand(I);
1332     if (!MO.isReg()) continue;
1333     Register Reg = MO.getReg();
1334     if (Reg == 0 || !Register::isPhysicalRegister(Reg))
1335       continue;
1336     if (SuccToSinkTo->isLiveIn(Reg))
1337       return false;
1338   }
1339 
1340   LLVM_DEBUG(dbgs() << "Sink instr " << MI << "\tinto block " << *SuccToSinkTo);
1341 
1342   // If the block has multiple predecessors, this is a critical edge.
1343   // Decide if we can sink along it or need to break the edge.
1344   if (SuccToSinkTo->pred_size() > 1) {
1345     // We cannot sink a load across a critical edge - there may be stores in
1346     // other code paths.
1347     bool TryBreak = false;
1348     bool Store =
1349         MI.mayLoad() ? hasStoreBetween(ParentBlock, SuccToSinkTo, MI) : true;
1350     if (!MI.isSafeToMove(AA, Store)) {
1351       LLVM_DEBUG(dbgs() << " *** NOTE: Won't sink load along critical edge.\n");
1352       TryBreak = true;
1353     }
1354 
1355     // We don't want to sink across a critical edge if we don't dominate the
1356     // successor. We could be introducing calculations to new code paths.
1357     if (!TryBreak && !DT->dominates(ParentBlock, SuccToSinkTo)) {
1358       LLVM_DEBUG(dbgs() << " *** NOTE: Critical edge found\n");
1359       TryBreak = true;
1360     }
1361 
1362     // Don't sink instructions into a loop.
1363     if (!TryBreak && LI->isLoopHeader(SuccToSinkTo)) {
1364       LLVM_DEBUG(dbgs() << " *** NOTE: Loop header found\n");
1365       TryBreak = true;
1366     }
1367 
1368     // Otherwise we are OK with sinking along a critical edge.
1369     if (!TryBreak)
1370       LLVM_DEBUG(dbgs() << "Sinking along critical edge.\n");
1371     else {
1372       // Mark this edge as to be split.
1373       // If the edge can actually be split, the next iteration of the main loop
1374       // will sink MI in the newly created block.
1375       bool Status =
1376         PostponeSplitCriticalEdge(MI, ParentBlock, SuccToSinkTo, BreakPHIEdge);
1377       if (!Status)
1378         LLVM_DEBUG(dbgs() << " *** PUNTING: Not legal or profitable to "
1379                              "break critical edge\n");
1380       // The instruction will not be sunk this time.
1381       return false;
1382     }
1383   }
1384 
1385   if (BreakPHIEdge) {
1386     // BreakPHIEdge is true if all the uses are in the successor MBB being
1387     // sunken into and they are all PHI nodes. In this case, machine-sink must
1388     // break the critical edge first.
1389     bool Status = PostponeSplitCriticalEdge(MI, ParentBlock,
1390                                             SuccToSinkTo, BreakPHIEdge);
1391     if (!Status)
1392       LLVM_DEBUG(dbgs() << " *** PUNTING: Not legal or profitable to "
1393                            "break critical edge\n");
1394     // The instruction will not be sunk this time.
1395     return false;
1396   }
1397 
1398   // Determine where to insert into. Skip phi nodes.
1399   MachineBasicBlock::iterator InsertPos = SuccToSinkTo->begin();
1400   while (InsertPos != SuccToSinkTo->end() && InsertPos->isPHI())
1401     ++InsertPos;
1402 
1403   // Collect debug users of any vreg that this inst defines.
1404   SmallVector<MIRegs, 4> DbgUsersToSink;
1405   for (auto &MO : MI.operands()) {
1406     if (!MO.isReg() || !MO.isDef() || !MO.getReg().isVirtual())
1407       continue;
1408     if (!SeenDbgUsers.count(MO.getReg()))
1409       continue;
1410 
1411     // Sink any users that don't pass any other DBG_VALUEs for this variable.
1412     auto &Users = SeenDbgUsers[MO.getReg()];
1413     for (auto &User : Users) {
1414       MachineInstr *DbgMI = User.getPointer();
1415       if (User.getInt()) {
1416         // This DBG_VALUE would re-order assignments. If we can't copy-propagate
1417         // it, it can't be recovered. Set it undef.
1418         if (!attemptDebugCopyProp(MI, *DbgMI, MO.getReg()))
1419           DbgMI->setDebugValueUndef();
1420       } else {
1421         DbgUsersToSink.push_back(
1422             {DbgMI, SmallVector<unsigned, 2>(1, MO.getReg())});
1423       }
1424     }
1425   }
1426 
1427   // After sinking, some debug users may not be dominated any more. If possible,
1428   // copy-propagate their operands. As it's expensive, don't do this if there's
1429   // no debuginfo in the program.
1430   if (MI.getMF()->getFunction().getSubprogram() && MI.isCopy())
1431     SalvageUnsunkDebugUsersOfCopy(MI, SuccToSinkTo);
1432 
1433   performSink(MI, *SuccToSinkTo, InsertPos, DbgUsersToSink);
1434 
1435   // Conservatively, clear any kill flags, since it's possible that they are no
1436   // longer correct.
1437   // Note that we have to clear the kill flags for any register this instruction
1438   // uses as we may sink over another instruction which currently kills the
1439   // used registers.
1440   for (MachineOperand &MO : MI.operands()) {
1441     if (MO.isReg() && MO.isUse())
1442       RegsToClearKillFlags.set(MO.getReg()); // Remember to clear kill flags.
1443   }
1444 
1445   return true;
1446 }
1447 
1448 void MachineSinking::SalvageUnsunkDebugUsersOfCopy(
1449     MachineInstr &MI, MachineBasicBlock *TargetBlock) {
1450   assert(MI.isCopy());
1451   assert(MI.getOperand(1).isReg());
1452 
1453   // Enumerate all users of vreg operands that are def'd. Skip those that will
1454   // be sunk. For the rest, if they are not dominated by the block we will sink
1455   // MI into, propagate the copy source to them.
1456   SmallVector<MachineInstr *, 4> DbgDefUsers;
1457   SmallVector<Register, 4> DbgUseRegs;
1458   const MachineRegisterInfo &MRI = MI.getMF()->getRegInfo();
1459   for (auto &MO : MI.operands()) {
1460     if (!MO.isReg() || !MO.isDef() || !MO.getReg().isVirtual())
1461       continue;
1462     DbgUseRegs.push_back(MO.getReg());
1463     for (auto &User : MRI.use_instructions(MO.getReg())) {
1464       if (!User.isDebugValue() || DT->dominates(TargetBlock, User.getParent()))
1465         continue;
1466 
1467       // If is in same block, will either sink or be use-before-def.
1468       if (User.getParent() == MI.getParent())
1469         continue;
1470 
1471       assert(User.hasDebugOperandForReg(MO.getReg()) &&
1472              "DBG_VALUE user of vreg, but has no operand for it?");
1473       DbgDefUsers.push_back(&User);
1474     }
1475   }
1476 
1477   // Point the users of this copy that are no longer dominated, at the source
1478   // of the copy.
1479   for (auto *User : DbgDefUsers) {
1480     for (auto &Reg : DbgUseRegs) {
1481       for (auto &DbgOp : User->getDebugOperandsForReg(Reg)) {
1482         DbgOp.setReg(MI.getOperand(1).getReg());
1483         DbgOp.setSubReg(MI.getOperand(1).getSubReg());
1484       }
1485     }
1486   }
1487 }
1488 
1489 //===----------------------------------------------------------------------===//
1490 // This pass is not intended to be a replacement or a complete alternative
1491 // for the pre-ra machine sink pass. It is only designed to sink COPY
1492 // instructions which should be handled after RA.
1493 //
1494 // This pass sinks COPY instructions into a successor block, if the COPY is not
1495 // used in the current block and the COPY is live-in to a single successor
1496 // (i.e., doesn't require the COPY to be duplicated).  This avoids executing the
1497 // copy on paths where their results aren't needed.  This also exposes
1498 // additional opportunites for dead copy elimination and shrink wrapping.
1499 //
1500 // These copies were either not handled by or are inserted after the MachineSink
1501 // pass. As an example of the former case, the MachineSink pass cannot sink
1502 // COPY instructions with allocatable source registers; for AArch64 these type
1503 // of copy instructions are frequently used to move function parameters (PhyReg)
1504 // into virtual registers in the entry block.
1505 //
1506 // For the machine IR below, this pass will sink %w19 in the entry into its
1507 // successor (%bb.1) because %w19 is only live-in in %bb.1.
1508 // %bb.0:
1509 //   %wzr = SUBSWri %w1, 1
1510 //   %w19 = COPY %w0
1511 //   Bcc 11, %bb.2
1512 // %bb.1:
1513 //   Live Ins: %w19
1514 //   BL @fun
1515 //   %w0 = ADDWrr %w0, %w19
1516 //   RET %w0
1517 // %bb.2:
1518 //   %w0 = COPY %wzr
1519 //   RET %w0
1520 // As we sink %w19 (CSR in AArch64) into %bb.1, the shrink-wrapping pass will be
1521 // able to see %bb.0 as a candidate.
1522 //===----------------------------------------------------------------------===//
1523 namespace {
1524 
1525 class PostRAMachineSinking : public MachineFunctionPass {
1526 public:
1527   bool runOnMachineFunction(MachineFunction &MF) override;
1528 
1529   static char ID;
1530   PostRAMachineSinking() : MachineFunctionPass(ID) {}
1531   StringRef getPassName() const override { return "PostRA Machine Sink"; }
1532 
1533   void getAnalysisUsage(AnalysisUsage &AU) const override {
1534     AU.setPreservesCFG();
1535     MachineFunctionPass::getAnalysisUsage(AU);
1536   }
1537 
1538   MachineFunctionProperties getRequiredProperties() const override {
1539     return MachineFunctionProperties().set(
1540         MachineFunctionProperties::Property::NoVRegs);
1541   }
1542 
1543 private:
1544   /// Track which register units have been modified and used.
1545   LiveRegUnits ModifiedRegUnits, UsedRegUnits;
1546 
1547   /// Track DBG_VALUEs of (unmodified) register units. Each DBG_VALUE has an
1548   /// entry in this map for each unit it touches. The DBG_VALUE's entry
1549   /// consists of a pointer to the instruction itself, and a vector of registers
1550   /// referred to by the instruction that overlap the key register unit.
1551   DenseMap<unsigned, SmallVector<MIRegs, 2>> SeenDbgInstrs;
1552 
1553   /// Sink Copy instructions unused in the same block close to their uses in
1554   /// successors.
1555   bool tryToSinkCopy(MachineBasicBlock &BB, MachineFunction &MF,
1556                      const TargetRegisterInfo *TRI, const TargetInstrInfo *TII);
1557 };
1558 } // namespace
1559 
1560 char PostRAMachineSinking::ID = 0;
1561 char &llvm::PostRAMachineSinkingID = PostRAMachineSinking::ID;
1562 
1563 INITIALIZE_PASS(PostRAMachineSinking, "postra-machine-sink",
1564                 "PostRA Machine Sink", false, false)
1565 
1566 static bool aliasWithRegsInLiveIn(MachineBasicBlock &MBB, unsigned Reg,
1567                                   const TargetRegisterInfo *TRI) {
1568   LiveRegUnits LiveInRegUnits(*TRI);
1569   LiveInRegUnits.addLiveIns(MBB);
1570   return !LiveInRegUnits.available(Reg);
1571 }
1572 
1573 static MachineBasicBlock *
1574 getSingleLiveInSuccBB(MachineBasicBlock &CurBB,
1575                       const SmallPtrSetImpl<MachineBasicBlock *> &SinkableBBs,
1576                       unsigned Reg, const TargetRegisterInfo *TRI) {
1577   // Try to find a single sinkable successor in which Reg is live-in.
1578   MachineBasicBlock *BB = nullptr;
1579   for (auto *SI : SinkableBBs) {
1580     if (aliasWithRegsInLiveIn(*SI, Reg, TRI)) {
1581       // If BB is set here, Reg is live-in to at least two sinkable successors,
1582       // so quit.
1583       if (BB)
1584         return nullptr;
1585       BB = SI;
1586     }
1587   }
1588   // Reg is not live-in to any sinkable successors.
1589   if (!BB)
1590     return nullptr;
1591 
1592   // Check if any register aliased with Reg is live-in in other successors.
1593   for (auto *SI : CurBB.successors()) {
1594     if (!SinkableBBs.count(SI) && aliasWithRegsInLiveIn(*SI, Reg, TRI))
1595       return nullptr;
1596   }
1597   return BB;
1598 }
1599 
1600 static MachineBasicBlock *
1601 getSingleLiveInSuccBB(MachineBasicBlock &CurBB,
1602                       const SmallPtrSetImpl<MachineBasicBlock *> &SinkableBBs,
1603                       ArrayRef<unsigned> DefedRegsInCopy,
1604                       const TargetRegisterInfo *TRI) {
1605   MachineBasicBlock *SingleBB = nullptr;
1606   for (auto DefReg : DefedRegsInCopy) {
1607     MachineBasicBlock *BB =
1608         getSingleLiveInSuccBB(CurBB, SinkableBBs, DefReg, TRI);
1609     if (!BB || (SingleBB && SingleBB != BB))
1610       return nullptr;
1611     SingleBB = BB;
1612   }
1613   return SingleBB;
1614 }
1615 
1616 static void clearKillFlags(MachineInstr *MI, MachineBasicBlock &CurBB,
1617                            SmallVectorImpl<unsigned> &UsedOpsInCopy,
1618                            LiveRegUnits &UsedRegUnits,
1619                            const TargetRegisterInfo *TRI) {
1620   for (auto U : UsedOpsInCopy) {
1621     MachineOperand &MO = MI->getOperand(U);
1622     Register SrcReg = MO.getReg();
1623     if (!UsedRegUnits.available(SrcReg)) {
1624       MachineBasicBlock::iterator NI = std::next(MI->getIterator());
1625       for (MachineInstr &UI : make_range(NI, CurBB.end())) {
1626         if (UI.killsRegister(SrcReg, TRI)) {
1627           UI.clearRegisterKills(SrcReg, TRI);
1628           MO.setIsKill(true);
1629           break;
1630         }
1631       }
1632     }
1633   }
1634 }
1635 
1636 static void updateLiveIn(MachineInstr *MI, MachineBasicBlock *SuccBB,
1637                          SmallVectorImpl<unsigned> &UsedOpsInCopy,
1638                          SmallVectorImpl<unsigned> &DefedRegsInCopy) {
1639   MachineFunction &MF = *SuccBB->getParent();
1640   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
1641   for (unsigned DefReg : DefedRegsInCopy)
1642     for (MCSubRegIterator S(DefReg, TRI, true); S.isValid(); ++S)
1643       SuccBB->removeLiveIn(*S);
1644   for (auto U : UsedOpsInCopy) {
1645     Register SrcReg = MI->getOperand(U).getReg();
1646     LaneBitmask Mask;
1647     for (MCRegUnitMaskIterator S(SrcReg, TRI); S.isValid(); ++S) {
1648       Mask |= (*S).second;
1649     }
1650     SuccBB->addLiveIn(SrcReg, Mask.any() ? Mask : LaneBitmask::getAll());
1651   }
1652   SuccBB->sortUniqueLiveIns();
1653 }
1654 
1655 static bool hasRegisterDependency(MachineInstr *MI,
1656                                   SmallVectorImpl<unsigned> &UsedOpsInCopy,
1657                                   SmallVectorImpl<unsigned> &DefedRegsInCopy,
1658                                   LiveRegUnits &ModifiedRegUnits,
1659                                   LiveRegUnits &UsedRegUnits) {
1660   bool HasRegDependency = false;
1661   for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1662     MachineOperand &MO = MI->getOperand(i);
1663     if (!MO.isReg())
1664       continue;
1665     Register Reg = MO.getReg();
1666     if (!Reg)
1667       continue;
1668     if (MO.isDef()) {
1669       if (!ModifiedRegUnits.available(Reg) || !UsedRegUnits.available(Reg)) {
1670         HasRegDependency = true;
1671         break;
1672       }
1673       DefedRegsInCopy.push_back(Reg);
1674 
1675       // FIXME: instead of isUse(), readsReg() would be a better fix here,
1676       // For example, we can ignore modifications in reg with undef. However,
1677       // it's not perfectly clear if skipping the internal read is safe in all
1678       // other targets.
1679     } else if (MO.isUse()) {
1680       if (!ModifiedRegUnits.available(Reg)) {
1681         HasRegDependency = true;
1682         break;
1683       }
1684       UsedOpsInCopy.push_back(i);
1685     }
1686   }
1687   return HasRegDependency;
1688 }
1689 
1690 static SmallSet<MCRegister, 4> getRegUnits(MCRegister Reg,
1691                                            const TargetRegisterInfo *TRI) {
1692   SmallSet<MCRegister, 4> RegUnits;
1693   for (auto RI = MCRegUnitIterator(Reg, TRI); RI.isValid(); ++RI)
1694     RegUnits.insert(*RI);
1695   return RegUnits;
1696 }
1697 
1698 bool PostRAMachineSinking::tryToSinkCopy(MachineBasicBlock &CurBB,
1699                                          MachineFunction &MF,
1700                                          const TargetRegisterInfo *TRI,
1701                                          const TargetInstrInfo *TII) {
1702   SmallPtrSet<MachineBasicBlock *, 2> SinkableBBs;
1703   // FIXME: For now, we sink only to a successor which has a single predecessor
1704   // so that we can directly sink COPY instructions to the successor without
1705   // adding any new block or branch instruction.
1706   for (MachineBasicBlock *SI : CurBB.successors())
1707     if (!SI->livein_empty() && SI->pred_size() == 1)
1708       SinkableBBs.insert(SI);
1709 
1710   if (SinkableBBs.empty())
1711     return false;
1712 
1713   bool Changed = false;
1714 
1715   // Track which registers have been modified and used between the end of the
1716   // block and the current instruction.
1717   ModifiedRegUnits.clear();
1718   UsedRegUnits.clear();
1719   SeenDbgInstrs.clear();
1720 
1721   for (auto I = CurBB.rbegin(), E = CurBB.rend(); I != E;) {
1722     MachineInstr *MI = &*I;
1723     ++I;
1724 
1725     // Track the operand index for use in Copy.
1726     SmallVector<unsigned, 2> UsedOpsInCopy;
1727     // Track the register number defed in Copy.
1728     SmallVector<unsigned, 2> DefedRegsInCopy;
1729 
1730     // We must sink this DBG_VALUE if its operand is sunk. To avoid searching
1731     // for DBG_VALUEs later, record them when they're encountered.
1732     if (MI->isDebugValue()) {
1733       SmallDenseMap<MCRegister, SmallVector<unsigned, 2>, 4> MIUnits;
1734       bool IsValid = true;
1735       for (MachineOperand &MO : MI->debug_operands()) {
1736         if (MO.isReg() && Register::isPhysicalRegister(MO.getReg())) {
1737           // Bail if we can already tell the sink would be rejected, rather
1738           // than needlessly accumulating lots of DBG_VALUEs.
1739           if (hasRegisterDependency(MI, UsedOpsInCopy, DefedRegsInCopy,
1740                                     ModifiedRegUnits, UsedRegUnits)) {
1741             IsValid = false;
1742             break;
1743           }
1744 
1745           // Record debug use of each reg unit.
1746           SmallSet<MCRegister, 4> RegUnits = getRegUnits(MO.getReg(), TRI);
1747           for (MCRegister Reg : RegUnits)
1748             MIUnits[Reg].push_back(MO.getReg());
1749         }
1750       }
1751       if (IsValid) {
1752         for (auto RegOps : MIUnits)
1753           SeenDbgInstrs[RegOps.first].push_back({MI, RegOps.second});
1754       }
1755       continue;
1756     }
1757 
1758     if (MI->isDebugOrPseudoInstr())
1759       continue;
1760 
1761     // Do not move any instruction across function call.
1762     if (MI->isCall())
1763       return false;
1764 
1765     if (!MI->isCopy() || !MI->getOperand(0).isRenamable()) {
1766       LiveRegUnits::accumulateUsedDefed(*MI, ModifiedRegUnits, UsedRegUnits,
1767                                         TRI);
1768       continue;
1769     }
1770 
1771     // Don't sink the COPY if it would violate a register dependency.
1772     if (hasRegisterDependency(MI, UsedOpsInCopy, DefedRegsInCopy,
1773                               ModifiedRegUnits, UsedRegUnits)) {
1774       LiveRegUnits::accumulateUsedDefed(*MI, ModifiedRegUnits, UsedRegUnits,
1775                                         TRI);
1776       continue;
1777     }
1778     assert((!UsedOpsInCopy.empty() && !DefedRegsInCopy.empty()) &&
1779            "Unexpect SrcReg or DefReg");
1780     MachineBasicBlock *SuccBB =
1781         getSingleLiveInSuccBB(CurBB, SinkableBBs, DefedRegsInCopy, TRI);
1782     // Don't sink if we cannot find a single sinkable successor in which Reg
1783     // is live-in.
1784     if (!SuccBB) {
1785       LiveRegUnits::accumulateUsedDefed(*MI, ModifiedRegUnits, UsedRegUnits,
1786                                         TRI);
1787       continue;
1788     }
1789     assert((SuccBB->pred_size() == 1 && *SuccBB->pred_begin() == &CurBB) &&
1790            "Unexpected predecessor");
1791 
1792     // Collect DBG_VALUEs that must sink with this copy. We've previously
1793     // recorded which reg units that DBG_VALUEs read, if this instruction
1794     // writes any of those units then the corresponding DBG_VALUEs must sink.
1795     MapVector<MachineInstr *, MIRegs::second_type> DbgValsToSinkMap;
1796     for (auto &MO : MI->operands()) {
1797       if (!MO.isReg() || !MO.isDef())
1798         continue;
1799 
1800       SmallSet<MCRegister, 4> Units = getRegUnits(MO.getReg(), TRI);
1801       for (MCRegister Reg : Units) {
1802         for (auto MIRegs : SeenDbgInstrs.lookup(Reg)) {
1803           auto &Regs = DbgValsToSinkMap[MIRegs.first];
1804           for (unsigned Reg : MIRegs.second)
1805             Regs.push_back(Reg);
1806         }
1807       }
1808     }
1809     SmallVector<MIRegs, 4> DbgValsToSink(DbgValsToSinkMap.begin(),
1810                                          DbgValsToSinkMap.end());
1811 
1812     // Clear the kill flag if SrcReg is killed between MI and the end of the
1813     // block.
1814     clearKillFlags(MI, CurBB, UsedOpsInCopy, UsedRegUnits, TRI);
1815     MachineBasicBlock::iterator InsertPos = SuccBB->getFirstNonPHI();
1816     performSink(*MI, *SuccBB, InsertPos, DbgValsToSink);
1817     updateLiveIn(MI, SuccBB, UsedOpsInCopy, DefedRegsInCopy);
1818 
1819     Changed = true;
1820     ++NumPostRACopySink;
1821   }
1822   return Changed;
1823 }
1824 
1825 bool PostRAMachineSinking::runOnMachineFunction(MachineFunction &MF) {
1826   if (skipFunction(MF.getFunction()))
1827     return false;
1828 
1829   bool Changed = false;
1830   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
1831   const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo();
1832 
1833   ModifiedRegUnits.init(*TRI);
1834   UsedRegUnits.init(*TRI);
1835   for (auto &BB : MF)
1836     Changed |= tryToSinkCopy(BB, MF, TRI, TII);
1837 
1838   return Changed;
1839 }
1840