xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/MachineSink.cpp (revision 5036d9652a5701d00e9e40ea942c278e9f77d33d)
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/DepthFirstIterator.h"
20 #include "llvm/ADT/MapVector.h"
21 #include "llvm/ADT/PointerIntPair.h"
22 #include "llvm/ADT/PostOrderIterator.h"
23 #include "llvm/ADT/SetVector.h"
24 #include "llvm/ADT/SmallSet.h"
25 #include "llvm/ADT/SmallVector.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/CFG.h"
29 #include "llvm/CodeGen/MachineBasicBlock.h"
30 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
31 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
32 #include "llvm/CodeGen/MachineCycleAnalysis.h"
33 #include "llvm/CodeGen/MachineDominators.h"
34 #include "llvm/CodeGen/MachineFunction.h"
35 #include "llvm/CodeGen/MachineFunctionPass.h"
36 #include "llvm/CodeGen/MachineInstr.h"
37 #include "llvm/CodeGen/MachineLoopInfo.h"
38 #include "llvm/CodeGen/MachineOperand.h"
39 #include "llvm/CodeGen/MachinePostDominators.h"
40 #include "llvm/CodeGen/MachineRegisterInfo.h"
41 #include "llvm/CodeGen/RegisterClassInfo.h"
42 #include "llvm/CodeGen/RegisterPressure.h"
43 #include "llvm/CodeGen/TargetInstrInfo.h"
44 #include "llvm/CodeGen/TargetPassConfig.h"
45 #include "llvm/CodeGen/TargetRegisterInfo.h"
46 #include "llvm/CodeGen/TargetSubtargetInfo.h"
47 #include "llvm/IR/BasicBlock.h"
48 #include "llvm/IR/DebugInfoMetadata.h"
49 #include "llvm/IR/LLVMContext.h"
50 #include "llvm/InitializePasses.h"
51 #include "llvm/MC/MCRegisterInfo.h"
52 #include "llvm/Pass.h"
53 #include "llvm/Support/BranchProbability.h"
54 #include "llvm/Support/CommandLine.h"
55 #include "llvm/Support/Debug.h"
56 #include "llvm/Support/raw_ostream.h"
57 #include <algorithm>
58 #include <cassert>
59 #include <cstdint>
60 #include <utility>
61 #include <vector>
62 
63 using namespace llvm;
64 
65 #define DEBUG_TYPE "machine-sink"
66 
67 static cl::opt<bool>
68 SplitEdges("machine-sink-split",
69            cl::desc("Split critical edges during machine sinking"),
70            cl::init(true), cl::Hidden);
71 
72 static cl::opt<bool>
73 UseBlockFreqInfo("machine-sink-bfi",
74            cl::desc("Use block frequency info to find successors to sink"),
75            cl::init(true), cl::Hidden);
76 
77 static cl::opt<unsigned> SplitEdgeProbabilityThreshold(
78     "machine-sink-split-probability-threshold",
79     cl::desc(
80         "Percentage threshold for splitting single-instruction critical edge. "
81         "If the branch threshold is higher than this threshold, we allow "
82         "speculative execution of up to 1 instruction to avoid branching to "
83         "splitted critical edge"),
84     cl::init(40), cl::Hidden);
85 
86 static cl::opt<unsigned> SinkLoadInstsPerBlockThreshold(
87     "machine-sink-load-instrs-threshold",
88     cl::desc("Do not try to find alias store for a load if there is a in-path "
89              "block whose instruction number is higher than this threshold."),
90     cl::init(2000), cl::Hidden);
91 
92 static cl::opt<unsigned> SinkLoadBlocksThreshold(
93     "machine-sink-load-blocks-threshold",
94     cl::desc("Do not try to find alias store for a load if the block number in "
95              "the straight line is higher than this threshold."),
96     cl::init(20), cl::Hidden);
97 
98 static cl::opt<bool>
99     SinkInstsIntoCycle("sink-insts-to-avoid-spills",
100                        cl::desc("Sink instructions into cycles to avoid "
101                                 "register spills"),
102                        cl::init(false), cl::Hidden);
103 
104 static cl::opt<unsigned> SinkIntoCycleLimit(
105     "machine-sink-cycle-limit",
106     cl::desc("The maximum number of instructions considered for cycle sinking."),
107     cl::init(50), cl::Hidden);
108 
109 STATISTIC(NumSunk,      "Number of machine instructions sunk");
110 STATISTIC(NumCycleSunk,  "Number of machine instructions sunk into a cycle");
111 STATISTIC(NumSplit,     "Number of critical edges split");
112 STATISTIC(NumCoalesces, "Number of copies coalesced");
113 STATISTIC(NumPostRACopySink, "Number of copies sunk after RA");
114 
115 namespace {
116 
117   class MachineSinking : public MachineFunctionPass {
118     const TargetSubtargetInfo *STI = nullptr;
119     const TargetInstrInfo *TII = nullptr;
120     const TargetRegisterInfo *TRI = nullptr;
121     MachineRegisterInfo *MRI = nullptr;      // Machine register information
122     MachineDominatorTree *DT = nullptr;      // Machine dominator tree
123     MachinePostDominatorTree *PDT = nullptr; // Machine post dominator tree
124     MachineCycleInfo *CI = nullptr;
125     MachineBlockFrequencyInfo *MBFI = nullptr;
126     const MachineBranchProbabilityInfo *MBPI = nullptr;
127     AliasAnalysis *AA = nullptr;
128     RegisterClassInfo RegClassInfo;
129 
130     // Remember which edges have been considered for breaking.
131     SmallSet<std::pair<MachineBasicBlock*, MachineBasicBlock*>, 8>
132     CEBCandidates;
133     // Memorize the register that also wanted to sink into the same block along
134     // a different critical edge.
135     // {register to sink, sink-to block} -> the first sink-from block.
136     // We're recording the first sink-from block because that (critical) edge
137     // was deferred until we see another register that's going to sink into the
138     // same block.
139     DenseMap<std::pair<Register, MachineBasicBlock *>, MachineBasicBlock *>
140         CEMergeCandidates;
141     // Remember which edges we are about to split.
142     // This is different from CEBCandidates since those edges
143     // will be split.
144     SetVector<std::pair<MachineBasicBlock *, MachineBasicBlock *>> ToSplit;
145 
146     DenseSet<Register> RegsToClearKillFlags;
147 
148     using AllSuccsCache =
149         SmallDenseMap<MachineBasicBlock *, SmallVector<MachineBasicBlock *, 4>>;
150 
151     /// DBG_VALUE pointer and flag. The flag is true if this DBG_VALUE is
152     /// post-dominated by another DBG_VALUE of the same variable location.
153     /// This is necessary to detect sequences such as:
154     ///     %0 = someinst
155     ///     DBG_VALUE %0, !123, !DIExpression()
156     ///     %1 = anotherinst
157     ///     DBG_VALUE %1, !123, !DIExpression()
158     /// Where if %0 were to sink, the DBG_VAUE should not sink with it, as that
159     /// would re-order assignments.
160     using SeenDbgUser = PointerIntPair<MachineInstr *, 1>;
161 
162     /// Record of DBG_VALUE uses of vregs in a block, so that we can identify
163     /// debug instructions to sink.
164     SmallDenseMap<unsigned, TinyPtrVector<SeenDbgUser>> SeenDbgUsers;
165 
166     /// Record of debug variables that have had their locations set in the
167     /// current block.
168     DenseSet<DebugVariable> SeenDbgVars;
169 
170     DenseMap<std::pair<MachineBasicBlock *, MachineBasicBlock *>, bool>
171         HasStoreCache;
172 
173     DenseMap<std::pair<MachineBasicBlock *, MachineBasicBlock *>,
174              SmallVector<MachineInstr *>>
175         StoreInstrCache;
176 
177     /// Cached BB's register pressure.
178     DenseMap<const MachineBasicBlock *, std::vector<unsigned>>
179         CachedRegisterPressure;
180 
181     bool EnableSinkAndFold;
182 
183   public:
184     static char ID; // Pass identification
185 
186     MachineSinking() : MachineFunctionPass(ID) {
187       initializeMachineSinkingPass(*PassRegistry::getPassRegistry());
188     }
189 
190     bool runOnMachineFunction(MachineFunction &MF) override;
191 
192     void getAnalysisUsage(AnalysisUsage &AU) const override {
193       MachineFunctionPass::getAnalysisUsage(AU);
194       AU.addRequired<AAResultsWrapperPass>();
195       AU.addRequired<MachineDominatorTreeWrapperPass>();
196       AU.addRequired<MachinePostDominatorTreeWrapperPass>();
197       AU.addRequired<MachineCycleInfoWrapperPass>();
198       AU.addRequired<MachineBranchProbabilityInfoWrapperPass>();
199       AU.addPreserved<MachineCycleInfoWrapperPass>();
200       AU.addPreserved<MachineLoopInfoWrapperPass>();
201       if (UseBlockFreqInfo)
202         AU.addRequired<MachineBlockFrequencyInfoWrapperPass>();
203       AU.addRequired<TargetPassConfig>();
204     }
205 
206     void releaseMemory() override {
207       CEBCandidates.clear();
208       CEMergeCandidates.clear();
209     }
210 
211   private:
212     bool ProcessBlock(MachineBasicBlock &MBB);
213     void ProcessDbgInst(MachineInstr &MI);
214     bool isLegalToBreakCriticalEdge(MachineInstr &MI, MachineBasicBlock *From,
215                                     MachineBasicBlock *To, bool BreakPHIEdge);
216     bool isWorthBreakingCriticalEdge(MachineInstr &MI, MachineBasicBlock *From,
217                                      MachineBasicBlock *To,
218                                      MachineBasicBlock *&DeferredFromBlock);
219 
220     bool hasStoreBetween(MachineBasicBlock *From, MachineBasicBlock *To,
221                          MachineInstr &MI);
222 
223     /// Postpone the splitting of the given critical
224     /// edge (\p From, \p To).
225     ///
226     /// We do not split the edges on the fly. Indeed, this invalidates
227     /// the dominance information and thus triggers a lot of updates
228     /// of that information underneath.
229     /// Instead, we postpone all the splits after each iteration of
230     /// the main loop. That way, the information is at least valid
231     /// for the lifetime of an iteration.
232     ///
233     /// \return True if the edge is marked as toSplit, false otherwise.
234     /// False can be returned if, for instance, this is not profitable.
235     bool PostponeSplitCriticalEdge(MachineInstr &MI,
236                                    MachineBasicBlock *From,
237                                    MachineBasicBlock *To,
238                                    bool BreakPHIEdge);
239     bool SinkInstruction(MachineInstr &MI, bool &SawStore,
240                          AllSuccsCache &AllSuccessors);
241 
242     /// If we sink a COPY inst, some debug users of it's destination may no
243     /// longer be dominated by the COPY, and will eventually be dropped.
244     /// This is easily rectified by forwarding the non-dominated debug uses
245     /// to the copy source.
246     void SalvageUnsunkDebugUsersOfCopy(MachineInstr &,
247                                        MachineBasicBlock *TargetBlock);
248     bool AllUsesDominatedByBlock(Register Reg, MachineBasicBlock *MBB,
249                                  MachineBasicBlock *DefMBB, bool &BreakPHIEdge,
250                                  bool &LocalUse) const;
251     MachineBasicBlock *FindSuccToSinkTo(MachineInstr &MI, MachineBasicBlock *MBB,
252                bool &BreakPHIEdge, AllSuccsCache &AllSuccessors);
253 
254     void FindCycleSinkCandidates(MachineCycle *Cycle, MachineBasicBlock *BB,
255                                  SmallVectorImpl<MachineInstr *> &Candidates);
256     bool SinkIntoCycle(MachineCycle *Cycle, MachineInstr &I);
257 
258     bool isProfitableToSinkTo(Register Reg, MachineInstr &MI,
259                               MachineBasicBlock *MBB,
260                               MachineBasicBlock *SuccToSinkTo,
261                               AllSuccsCache &AllSuccessors);
262 
263     bool PerformTrivialForwardCoalescing(MachineInstr &MI,
264                                          MachineBasicBlock *MBB);
265 
266     bool PerformSinkAndFold(MachineInstr &MI, MachineBasicBlock *MBB);
267 
268     SmallVector<MachineBasicBlock *, 4> &
269     GetAllSortedSuccessors(MachineInstr &MI, MachineBasicBlock *MBB,
270                            AllSuccsCache &AllSuccessors) const;
271 
272     std::vector<unsigned> &getBBRegisterPressure(const MachineBasicBlock &MBB);
273 
274     bool registerPressureSetExceedsLimit(unsigned NRegs,
275                                          const TargetRegisterClass *RC,
276                                          const MachineBasicBlock &MBB);
277   };
278 
279 } // end anonymous namespace
280 
281 char MachineSinking::ID = 0;
282 
283 char &llvm::MachineSinkingID = MachineSinking::ID;
284 
285 INITIALIZE_PASS_BEGIN(MachineSinking, DEBUG_TYPE,
286                       "Machine code sinking", false, false)
287 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfoWrapperPass)
288 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass)
289 INITIALIZE_PASS_DEPENDENCY(MachineCycleInfoWrapperPass)
290 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
291 INITIALIZE_PASS_END(MachineSinking, DEBUG_TYPE,
292                     "Machine code sinking", false, false)
293 
294 /// Return true if a target defined block prologue instruction interferes
295 /// with a sink candidate.
296 static bool blockPrologueInterferes(const MachineBasicBlock *BB,
297                                     MachineBasicBlock::const_iterator End,
298                                     const MachineInstr &MI,
299                                     const TargetRegisterInfo *TRI,
300                                     const TargetInstrInfo *TII,
301                                     const MachineRegisterInfo *MRI) {
302   for (MachineBasicBlock::const_iterator PI = BB->getFirstNonPHI(); PI != End;
303        ++PI) {
304     // Only check target defined prologue instructions
305     if (!TII->isBasicBlockPrologue(*PI))
306       continue;
307     for (auto &MO : MI.operands()) {
308       if (!MO.isReg())
309         continue;
310       Register Reg = MO.getReg();
311       if (!Reg)
312         continue;
313       if (MO.isUse()) {
314         if (Reg.isPhysical() &&
315             (TII->isIgnorableUse(MO) || (MRI && MRI->isConstantPhysReg(Reg))))
316           continue;
317         if (PI->modifiesRegister(Reg, TRI))
318           return true;
319       } else {
320         if (PI->readsRegister(Reg, TRI))
321           return true;
322         // Check for interference with non-dead defs
323         auto *DefOp = PI->findRegisterDefOperand(Reg, TRI, false, true);
324         if (DefOp && !DefOp->isDead())
325           return true;
326       }
327     }
328   }
329 
330   return false;
331 }
332 
333 bool MachineSinking::PerformTrivialForwardCoalescing(MachineInstr &MI,
334                                                      MachineBasicBlock *MBB) {
335   if (!MI.isCopy())
336     return false;
337 
338   Register SrcReg = MI.getOperand(1).getReg();
339   Register DstReg = MI.getOperand(0).getReg();
340   if (!SrcReg.isVirtual() || !DstReg.isVirtual() ||
341       !MRI->hasOneNonDBGUse(SrcReg))
342     return false;
343 
344   const TargetRegisterClass *SRC = MRI->getRegClass(SrcReg);
345   const TargetRegisterClass *DRC = MRI->getRegClass(DstReg);
346   if (SRC != DRC)
347     return false;
348 
349   MachineInstr *DefMI = MRI->getVRegDef(SrcReg);
350   if (DefMI->isCopyLike())
351     return false;
352   LLVM_DEBUG(dbgs() << "Coalescing: " << *DefMI);
353   LLVM_DEBUG(dbgs() << "*** to: " << MI);
354   MRI->replaceRegWith(DstReg, SrcReg);
355   MI.eraseFromParent();
356 
357   // Conservatively, clear any kill flags, since it's possible that they are no
358   // longer correct.
359   MRI->clearKillFlags(SrcReg);
360 
361   ++NumCoalesces;
362   return true;
363 }
364 
365 bool MachineSinking::PerformSinkAndFold(MachineInstr &MI,
366                                         MachineBasicBlock *MBB) {
367   if (MI.isCopy() || MI.mayLoadOrStore() ||
368       MI.getOpcode() == TargetOpcode::REG_SEQUENCE)
369     return false;
370 
371   // Don't sink instructions that the target prefers not to sink.
372   if (!TII->shouldSink(MI))
373     return false;
374 
375   // Check if it's safe to move the instruction.
376   bool SawStore = true;
377   if (!MI.isSafeToMove(AA, SawStore))
378     return false;
379 
380   // Convergent operations may not be made control-dependent on additional
381   // values.
382   if (MI.isConvergent())
383     return false;
384 
385   // Don't sink defs/uses of hard registers or if the instruction defines more
386   // than one register.
387   // Don't sink more than two register uses - it'll cover most of the cases and
388   // greatly simplifies the register pressure checks.
389   Register DefReg;
390   Register UsedRegA, UsedRegB;
391   for (const MachineOperand &MO : MI.operands()) {
392     if (MO.isImm() || MO.isRegMask() || MO.isRegLiveOut() || MO.isMetadata() ||
393         MO.isMCSymbol() || MO.isDbgInstrRef() || MO.isCFIIndex() ||
394         MO.isIntrinsicID() || MO.isPredicate() || MO.isShuffleMask())
395       continue;
396     if (!MO.isReg())
397       return false;
398 
399     Register Reg = MO.getReg();
400     if (Reg == 0)
401       continue;
402 
403     if (Reg.isVirtual()) {
404       if (MO.isDef()) {
405         if (DefReg)
406           return false;
407         DefReg = Reg;
408         continue;
409       }
410 
411       if (UsedRegA == 0)
412         UsedRegA = Reg;
413       else if (UsedRegB == 0)
414         UsedRegB = Reg;
415       else
416         return false;
417       continue;
418     }
419 
420     if (Reg.isPhysical() && MO.isUse() &&
421         (MRI->isConstantPhysReg(Reg) || TII->isIgnorableUse(MO)))
422       continue;
423 
424     return false;
425   }
426 
427   // Scan uses of the destination register. Every use, except the last, must be
428   // a copy, with a chain of copies terminating with either a copy into a hard
429   // register, or a load/store instruction where the use is part of the
430   // address (*not* the stored value).
431   using SinkInfo = std::pair<MachineInstr *, ExtAddrMode>;
432   SmallVector<SinkInfo> SinkInto;
433   SmallVector<Register> Worklist;
434 
435   const TargetRegisterClass *RC = MRI->getRegClass(DefReg);
436   const TargetRegisterClass *RCA =
437       UsedRegA == 0 ? nullptr : MRI->getRegClass(UsedRegA);
438   const TargetRegisterClass *RCB =
439       UsedRegB == 0 ? nullptr : MRI->getRegClass(UsedRegB);
440 
441   Worklist.push_back(DefReg);
442   while (!Worklist.empty()) {
443     Register Reg = Worklist.pop_back_val();
444 
445     for (MachineOperand &MO : MRI->use_nodbg_operands(Reg)) {
446       ExtAddrMode MaybeAM;
447       MachineInstr &UseInst = *MO.getParent();
448       if (UseInst.isCopy()) {
449         Register DstReg;
450         if (const MachineOperand &O = UseInst.getOperand(0); O.isReg())
451           DstReg = O.getReg();
452         if (DstReg == 0)
453           return false;
454         if (DstReg.isVirtual()) {
455           Worklist.push_back(DstReg);
456           continue;
457         }
458         // If we are going to replace a copy, the original instruction must be
459         // as cheap as a copy.
460         if (!TII->isAsCheapAsAMove(MI))
461           return false;
462         // The hard register must be in the register class of the original
463         // instruction's destination register.
464         if (!RC->contains(DstReg))
465           return false;
466       } else if (UseInst.mayLoadOrStore()) {
467         ExtAddrMode AM;
468         if (!TII->canFoldIntoAddrMode(UseInst, Reg, MI, AM))
469           return false;
470         MaybeAM = AM;
471       } else {
472         return false;
473       }
474 
475       if (UseInst.getParent() != MI.getParent()) {
476         // If the register class of the register we are replacing is a superset
477         // of any of the register classes of the operands of the materialized
478         // instruction don't consider that live range extended.
479         const TargetRegisterClass *RCS = MRI->getRegClass(Reg);
480         if (RCA && RCA->hasSuperClassEq(RCS))
481           RCA = nullptr;
482         else if (RCB && RCB->hasSuperClassEq(RCS))
483           RCB = nullptr;
484         if (RCA || RCB) {
485           if (RCA == nullptr) {
486             RCA = RCB;
487             RCB = nullptr;
488           }
489 
490           unsigned NRegs = !!RCA + !!RCB;
491           if (RCA == RCB)
492             RCB = nullptr;
493 
494           // Check we don't exceed register pressure at the destination.
495           const MachineBasicBlock &MBB = *UseInst.getParent();
496           if (RCB == nullptr) {
497             if (registerPressureSetExceedsLimit(NRegs, RCA, MBB))
498               return false;
499           } else if (registerPressureSetExceedsLimit(1, RCA, MBB) ||
500                      registerPressureSetExceedsLimit(1, RCB, MBB)) {
501             return false;
502           }
503         }
504       }
505 
506       SinkInto.emplace_back(&UseInst, MaybeAM);
507     }
508   }
509 
510   if (SinkInto.empty())
511     return false;
512 
513   // Now we know we can fold the instruction in all its users.
514   for (auto &[SinkDst, MaybeAM] : SinkInto) {
515     MachineInstr *New = nullptr;
516     LLVM_DEBUG(dbgs() << "Sinking copy of"; MI.dump(); dbgs() << "into";
517                SinkDst->dump());
518     if (SinkDst->isCopy()) {
519       // TODO: After performing the sink-and-fold, the original instruction is
520       // deleted. Its value is still available (in a hard register), so if there
521       // are debug instructions which refer to the (now deleted) virtual
522       // register they could be updated to refer to the hard register, in
523       // principle. However, it's not clear how to do that, moreover in some
524       // cases the debug instructions may need to be replicated proportionally
525       // to the number of the COPY instructions replaced and in some extreme
526       // cases we can end up with quadratic increase in the number of debug
527       // instructions.
528 
529       // Sink a copy of the instruction, replacing a COPY instruction.
530       MachineBasicBlock::iterator InsertPt = SinkDst->getIterator();
531       Register DstReg = SinkDst->getOperand(0).getReg();
532       TII->reMaterialize(*SinkDst->getParent(), InsertPt, DstReg, 0, MI, *TRI);
533       New = &*std::prev(InsertPt);
534       if (!New->getDebugLoc())
535         New->setDebugLoc(SinkDst->getDebugLoc());
536 
537       // The operand registers of the "sunk" instruction have their live range
538       // extended and their kill flags may no longer be correct. Conservatively
539       // clear the kill flags.
540       if (UsedRegA)
541         MRI->clearKillFlags(UsedRegA);
542       if (UsedRegB)
543         MRI->clearKillFlags(UsedRegB);
544     } else {
545       // Fold instruction into the addressing mode of a memory instruction.
546       New = TII->emitLdStWithAddr(*SinkDst, MaybeAM);
547 
548       // The registers of the addressing mode may have their live range extended
549       // and their kill flags may no longer be correct. Conservatively clear the
550       // kill flags.
551       if (Register R = MaybeAM.BaseReg; R.isValid() && R.isVirtual())
552         MRI->clearKillFlags(R);
553       if (Register R = MaybeAM.ScaledReg; R.isValid() && R.isVirtual())
554         MRI->clearKillFlags(R);
555     }
556     LLVM_DEBUG(dbgs() << "yielding"; New->dump());
557     // Clear the StoreInstrCache, since we may invalidate it by erasing.
558     if (SinkDst->mayStore() && !SinkDst->hasOrderedMemoryRef())
559       StoreInstrCache.clear();
560     SinkDst->eraseFromParent();
561   }
562 
563   // Collect operands that need to be cleaned up because the registers no longer
564   // exist (in COPYs and debug instructions). We cannot delete instructions or
565   // clear operands while traversing register uses.
566   SmallVector<MachineOperand *> Cleanup;
567   Worklist.push_back(DefReg);
568   while (!Worklist.empty()) {
569     Register Reg = Worklist.pop_back_val();
570     for (MachineOperand &MO : MRI->use_operands(Reg)) {
571       MachineInstr *U = MO.getParent();
572       assert((U->isCopy() || U->isDebugInstr()) &&
573              "Only debug uses and copies must remain");
574       if (U->isCopy())
575         Worklist.push_back(U->getOperand(0).getReg());
576       Cleanup.push_back(&MO);
577     }
578   }
579 
580   // Delete the dead COPYs and clear operands in debug instructions
581   for (MachineOperand *MO : Cleanup) {
582     MachineInstr *I = MO->getParent();
583     if (I->isCopy()) {
584       I->eraseFromParent();
585     } else {
586       MO->setReg(0);
587       MO->setSubReg(0);
588     }
589   }
590 
591   MI.eraseFromParent();
592   return true;
593 }
594 
595 /// AllUsesDominatedByBlock - Return true if all uses of the specified register
596 /// occur in blocks dominated by the specified block. If any use is in the
597 /// definition block, then return false since it is never legal to move def
598 /// after uses.
599 bool MachineSinking::AllUsesDominatedByBlock(Register Reg,
600                                              MachineBasicBlock *MBB,
601                                              MachineBasicBlock *DefMBB,
602                                              bool &BreakPHIEdge,
603                                              bool &LocalUse) const {
604   assert(Reg.isVirtual() && "Only makes sense for vregs");
605 
606   // Ignore debug uses because debug info doesn't affect the code.
607   if (MRI->use_nodbg_empty(Reg))
608     return true;
609 
610   // BreakPHIEdge is true if all the uses are in the successor MBB being sunken
611   // into and they are all PHI nodes. In this case, machine-sink must break
612   // the critical edge first. e.g.
613   //
614   // %bb.1:
615   //   Predecessors according to CFG: %bb.0
616   //     ...
617   //     %def = DEC64_32r %x, implicit-def dead %eflags
618   //     ...
619   //     JE_4 <%bb.37>, implicit %eflags
620   //   Successors according to CFG: %bb.37 %bb.2
621   //
622   // %bb.2:
623   //     %p = PHI %y, %bb.0, %def, %bb.1
624   if (all_of(MRI->use_nodbg_operands(Reg), [&](MachineOperand &MO) {
625         MachineInstr *UseInst = MO.getParent();
626         unsigned OpNo = MO.getOperandNo();
627         MachineBasicBlock *UseBlock = UseInst->getParent();
628         return UseBlock == MBB && UseInst->isPHI() &&
629                UseInst->getOperand(OpNo + 1).getMBB() == DefMBB;
630       })) {
631     BreakPHIEdge = true;
632     return true;
633   }
634 
635   for (MachineOperand &MO : MRI->use_nodbg_operands(Reg)) {
636     // Determine the block of the use.
637     MachineInstr *UseInst = MO.getParent();
638     unsigned OpNo = &MO - &UseInst->getOperand(0);
639     MachineBasicBlock *UseBlock = UseInst->getParent();
640     if (UseInst->isPHI()) {
641       // PHI nodes use the operand in the predecessor block, not the block with
642       // the PHI.
643       UseBlock = UseInst->getOperand(OpNo+1).getMBB();
644     } else if (UseBlock == DefMBB) {
645       LocalUse = true;
646       return false;
647     }
648 
649     // Check that it dominates.
650     if (!DT->dominates(MBB, UseBlock))
651       return false;
652   }
653 
654   return true;
655 }
656 
657 /// Return true if this machine instruction loads from global offset table or
658 /// constant pool.
659 static bool mayLoadFromGOTOrConstantPool(MachineInstr &MI) {
660   assert(MI.mayLoad() && "Expected MI that loads!");
661 
662   // If we lost memory operands, conservatively assume that the instruction
663   // reads from everything..
664   if (MI.memoperands_empty())
665     return true;
666 
667   for (MachineMemOperand *MemOp : MI.memoperands())
668     if (const PseudoSourceValue *PSV = MemOp->getPseudoValue())
669       if (PSV->isGOT() || PSV->isConstantPool())
670         return true;
671 
672   return false;
673 }
674 
675 void MachineSinking::FindCycleSinkCandidates(
676     MachineCycle *Cycle, MachineBasicBlock *BB,
677     SmallVectorImpl<MachineInstr *> &Candidates) {
678   for (auto &MI : *BB) {
679     LLVM_DEBUG(dbgs() << "CycleSink: Analysing candidate: " << MI);
680     if (!TII->shouldSink(MI)) {
681       LLVM_DEBUG(dbgs() << "CycleSink: Instruction not a candidate for this "
682                            "target\n");
683       continue;
684     }
685     if (!isCycleInvariant(Cycle, MI)) {
686       LLVM_DEBUG(dbgs() << "CycleSink: Instruction is not cycle invariant\n");
687       continue;
688     }
689     bool DontMoveAcrossStore = true;
690     if (!MI.isSafeToMove(AA, DontMoveAcrossStore)) {
691       LLVM_DEBUG(dbgs() << "CycleSink: Instruction not safe to move.\n");
692       continue;
693     }
694     if (MI.mayLoad() && !mayLoadFromGOTOrConstantPool(MI)) {
695       LLVM_DEBUG(dbgs() << "CycleSink: Dont sink GOT or constant pool loads\n");
696       continue;
697     }
698     if (MI.isConvergent())
699       continue;
700 
701     const MachineOperand &MO = MI.getOperand(0);
702     if (!MO.isReg() || !MO.getReg() || !MO.isDef())
703       continue;
704     if (!MRI->hasOneDef(MO.getReg()))
705       continue;
706 
707     LLVM_DEBUG(dbgs() << "CycleSink: Instruction added as candidate.\n");
708     Candidates.push_back(&MI);
709   }
710 }
711 
712 bool MachineSinking::runOnMachineFunction(MachineFunction &MF) {
713   if (skipFunction(MF.getFunction()))
714     return false;
715 
716   LLVM_DEBUG(dbgs() << "******** Machine Sinking ********\n");
717 
718   STI = &MF.getSubtarget();
719   TII = STI->getInstrInfo();
720   TRI = STI->getRegisterInfo();
721   MRI = &MF.getRegInfo();
722   DT = &getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree();
723   PDT = &getAnalysis<MachinePostDominatorTreeWrapperPass>().getPostDomTree();
724   CI = &getAnalysis<MachineCycleInfoWrapperPass>().getCycleInfo();
725   MBFI = UseBlockFreqInfo
726              ? &getAnalysis<MachineBlockFrequencyInfoWrapperPass>().getMBFI()
727              : nullptr;
728   MBPI = &getAnalysis<MachineBranchProbabilityInfoWrapperPass>().getMBPI();
729   AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
730   RegClassInfo.runOnMachineFunction(MF);
731   TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
732   EnableSinkAndFold = PassConfig->getEnableSinkAndFold();
733 
734   bool EverMadeChange = false;
735 
736   while (true) {
737     bool MadeChange = false;
738 
739     // Process all basic blocks.
740     CEBCandidates.clear();
741     CEMergeCandidates.clear();
742     ToSplit.clear();
743     for (auto &MBB: MF)
744       MadeChange |= ProcessBlock(MBB);
745 
746     // If we have anything we marked as toSplit, split it now.
747     for (const auto &Pair : ToSplit) {
748       auto NewSucc = Pair.first->SplitCriticalEdge(Pair.second, *this);
749       if (NewSucc != nullptr) {
750         LLVM_DEBUG(dbgs() << " *** Splitting critical edge: "
751                           << printMBBReference(*Pair.first) << " -- "
752                           << printMBBReference(*NewSucc) << " -- "
753                           << printMBBReference(*Pair.second) << '\n');
754         if (MBFI)
755           MBFI->onEdgeSplit(*Pair.first, *NewSucc, *MBPI);
756 
757         MadeChange = true;
758         ++NumSplit;
759         CI->splitCriticalEdge(Pair.first, Pair.second, NewSucc);
760       } else
761         LLVM_DEBUG(dbgs() << " *** Not legal to break critical edge\n");
762     }
763     // If this iteration over the code changed anything, keep iterating.
764     if (!MadeChange) break;
765     EverMadeChange = true;
766   }
767 
768   if (SinkInstsIntoCycle) {
769     SmallVector<MachineCycle *, 8> Cycles(CI->toplevel_begin(),
770                                           CI->toplevel_end());
771     for (auto *Cycle : Cycles) {
772       MachineBasicBlock *Preheader = Cycle->getCyclePreheader();
773       if (!Preheader) {
774         LLVM_DEBUG(dbgs() << "CycleSink: Can't find preheader\n");
775         continue;
776       }
777       SmallVector<MachineInstr *, 8> Candidates;
778       FindCycleSinkCandidates(Cycle, Preheader, Candidates);
779 
780       // Walk the candidates in reverse order so that we start with the use
781       // of a def-use chain, if there is any.
782       // TODO: Sort the candidates using a cost-model.
783       unsigned i = 0;
784       for (MachineInstr *I : llvm::reverse(Candidates)) {
785         if (i++ == SinkIntoCycleLimit) {
786           LLVM_DEBUG(dbgs() << "CycleSink:   Limit reached of instructions to "
787                                "be analysed.");
788           break;
789         }
790 
791         if (!SinkIntoCycle(Cycle, *I))
792           break;
793         EverMadeChange = true;
794         ++NumCycleSunk;
795       }
796     }
797   }
798 
799   HasStoreCache.clear();
800   StoreInstrCache.clear();
801 
802   // Now clear any kill flags for recorded registers.
803   for (auto I : RegsToClearKillFlags)
804     MRI->clearKillFlags(I);
805   RegsToClearKillFlags.clear();
806 
807   return EverMadeChange;
808 }
809 
810 bool MachineSinking::ProcessBlock(MachineBasicBlock &MBB) {
811   if ((!EnableSinkAndFold && MBB.succ_size() <= 1) || MBB.empty())
812     return false;
813 
814   // Don't bother sinking code out of unreachable blocks. In addition to being
815   // unprofitable, it can also lead to infinite looping, because in an
816   // unreachable cycle there may be nowhere to stop.
817   if (!DT->isReachableFromEntry(&MBB)) return false;
818 
819   bool MadeChange = false;
820 
821   // Cache all successors, sorted by frequency info and cycle depth.
822   AllSuccsCache AllSuccessors;
823 
824   // Walk the basic block bottom-up.  Remember if we saw a store.
825   MachineBasicBlock::iterator I = MBB.end();
826   --I;
827   bool ProcessedBegin, SawStore = false;
828   do {
829     MachineInstr &MI = *I;  // The instruction to sink.
830 
831     // Predecrement I (if it's not begin) so that it isn't invalidated by
832     // sinking.
833     ProcessedBegin = I == MBB.begin();
834     if (!ProcessedBegin)
835       --I;
836 
837     if (MI.isDebugOrPseudoInstr()) {
838       if (MI.isDebugValue())
839         ProcessDbgInst(MI);
840       continue;
841     }
842 
843     if (EnableSinkAndFold && PerformSinkAndFold(MI, &MBB)) {
844       MadeChange = true;
845       continue;
846     }
847 
848     // Can't sink anything out of a block that has less than two successors.
849     if (MBB.succ_size() <= 1)
850       continue;
851 
852     if (PerformTrivialForwardCoalescing(MI, &MBB)) {
853       MadeChange = true;
854       continue;
855     }
856 
857     if (SinkInstruction(MI, SawStore, AllSuccessors)) {
858       ++NumSunk;
859       MadeChange = true;
860     }
861 
862     // If we just processed the first instruction in the block, we're done.
863   } while (!ProcessedBegin);
864 
865   SeenDbgUsers.clear();
866   SeenDbgVars.clear();
867   // recalculate the bb register pressure after sinking one BB.
868   CachedRegisterPressure.clear();
869   return MadeChange;
870 }
871 
872 void MachineSinking::ProcessDbgInst(MachineInstr &MI) {
873   // When we see DBG_VALUEs for registers, record any vreg it reads, so that
874   // we know what to sink if the vreg def sinks.
875   assert(MI.isDebugValue() && "Expected DBG_VALUE for processing");
876 
877   DebugVariable Var(MI.getDebugVariable(), MI.getDebugExpression(),
878                     MI.getDebugLoc()->getInlinedAt());
879   bool SeenBefore = SeenDbgVars.contains(Var);
880 
881   for (MachineOperand &MO : MI.debug_operands()) {
882     if (MO.isReg() && MO.getReg().isVirtual())
883       SeenDbgUsers[MO.getReg()].push_back(SeenDbgUser(&MI, SeenBefore));
884   }
885 
886   // Record the variable for any DBG_VALUE, to avoid re-ordering any of them.
887   SeenDbgVars.insert(Var);
888 }
889 
890 bool MachineSinking::isWorthBreakingCriticalEdge(
891     MachineInstr &MI, MachineBasicBlock *From, MachineBasicBlock *To,
892     MachineBasicBlock *&DeferredFromBlock) {
893   // FIXME: Need much better heuristics.
894 
895   // If the pass has already considered breaking this edge (during this pass
896   // through the function), then let's go ahead and break it. This means
897   // sinking multiple "cheap" instructions into the same block.
898   if (!CEBCandidates.insert(std::make_pair(From, To)).second)
899     return true;
900 
901   if (!MI.isCopy() && !TII->isAsCheapAsAMove(MI))
902     return true;
903 
904   // Check and record the register and the destination block we want to sink
905   // into. Note that we want to do the following before the next check on branch
906   // probability. Because we want to record the initial candidate even if it's
907   // on hot edge, so that other candidates that might not on hot edges can be
908   // sinked as well.
909   for (const auto &MO : MI.all_defs()) {
910     Register Reg = MO.getReg();
911     if (!Reg)
912       continue;
913     Register SrcReg = Reg.isVirtual() ? TRI->lookThruCopyLike(Reg, MRI) : Reg;
914     auto Key = std::make_pair(SrcReg, To);
915     auto Res = CEMergeCandidates.try_emplace(Key, From);
916     // We wanted to sink the same register into the same block, consider it to
917     // be profitable.
918     if (!Res.second) {
919       // Return the source block that was previously held off.
920       DeferredFromBlock = Res.first->second;
921       return true;
922     }
923   }
924 
925   if (From->isSuccessor(To) && MBPI->getEdgeProbability(From, To) <=
926       BranchProbability(SplitEdgeProbabilityThreshold, 100))
927     return true;
928 
929   // MI is cheap, we probably don't want to break the critical edge for it.
930   // However, if this would allow some definitions of its source operands
931   // to be sunk then it's probably worth it.
932   for (const MachineOperand &MO : MI.all_uses()) {
933     Register Reg = MO.getReg();
934     if (Reg == 0)
935       continue;
936 
937     // We don't move live definitions of physical registers,
938     // so sinking their uses won't enable any opportunities.
939     if (Reg.isPhysical())
940       continue;
941 
942     // If this instruction is the only user of a virtual register,
943     // check if breaking the edge will enable sinking
944     // both this instruction and the defining instruction.
945     if (MRI->hasOneNonDBGUse(Reg)) {
946       // If the definition resides in same MBB,
947       // claim it's likely we can sink these together.
948       // If definition resides elsewhere, we aren't
949       // blocking it from being sunk so don't break the edge.
950       MachineInstr *DefMI = MRI->getVRegDef(Reg);
951       if (DefMI->getParent() == MI.getParent())
952         return true;
953     }
954   }
955 
956   return false;
957 }
958 
959 bool MachineSinking::isLegalToBreakCriticalEdge(MachineInstr &MI,
960                                                 MachineBasicBlock *FromBB,
961                                                 MachineBasicBlock *ToBB,
962                                                 bool BreakPHIEdge) {
963   // Avoid breaking back edge. From == To means backedge for single BB cycle.
964   if (!SplitEdges || FromBB == ToBB || !FromBB->isSuccessor(ToBB))
965     return false;
966 
967   MachineCycle *FromCycle = CI->getCycle(FromBB);
968   MachineCycle *ToCycle = CI->getCycle(ToBB);
969 
970   // Check for backedges of more "complex" cycles.
971   if (FromCycle == ToCycle && FromCycle &&
972       (!FromCycle->isReducible() || FromCycle->getHeader() == ToBB))
973     return false;
974 
975   // It's not always legal to break critical edges and sink the computation
976   // to the edge.
977   //
978   // %bb.1:
979   // v1024
980   // Beq %bb.3
981   // <fallthrough>
982   // %bb.2:
983   // ... no uses of v1024
984   // <fallthrough>
985   // %bb.3:
986   // ...
987   //       = v1024
988   //
989   // If %bb.1 -> %bb.3 edge is broken and computation of v1024 is inserted:
990   //
991   // %bb.1:
992   // ...
993   // Bne %bb.2
994   // %bb.4:
995   // v1024 =
996   // B %bb.3
997   // %bb.2:
998   // ... no uses of v1024
999   // <fallthrough>
1000   // %bb.3:
1001   // ...
1002   //       = v1024
1003   //
1004   // This is incorrect since v1024 is not computed along the %bb.1->%bb.2->%bb.3
1005   // flow. We need to ensure the new basic block where the computation is
1006   // sunk to dominates all the uses.
1007   // It's only legal to break critical edge and sink the computation to the
1008   // new block if all the predecessors of "To", except for "From", are
1009   // not dominated by "From". Given SSA property, this means these
1010   // predecessors are dominated by "To".
1011   //
1012   // There is no need to do this check if all the uses are PHI nodes. PHI
1013   // sources are only defined on the specific predecessor edges.
1014   if (!BreakPHIEdge) {
1015     for (MachineBasicBlock *Pred : ToBB->predecessors())
1016       if (Pred != FromBB && !DT->dominates(ToBB, Pred))
1017         return false;
1018   }
1019 
1020   return true;
1021 }
1022 
1023 bool MachineSinking::PostponeSplitCriticalEdge(MachineInstr &MI,
1024                                                MachineBasicBlock *FromBB,
1025                                                MachineBasicBlock *ToBB,
1026                                                bool BreakPHIEdge) {
1027   bool Status = false;
1028   MachineBasicBlock *DeferredFromBB = nullptr;
1029   if (isWorthBreakingCriticalEdge(MI, FromBB, ToBB, DeferredFromBB)) {
1030     // If there is a DeferredFromBB, we consider FromBB only if _both_
1031     // of them are legal to split.
1032     if ((!DeferredFromBB ||
1033          ToSplit.count(std::make_pair(DeferredFromBB, ToBB)) ||
1034          isLegalToBreakCriticalEdge(MI, DeferredFromBB, ToBB, BreakPHIEdge)) &&
1035         isLegalToBreakCriticalEdge(MI, FromBB, ToBB, BreakPHIEdge)) {
1036       ToSplit.insert(std::make_pair(FromBB, ToBB));
1037       if (DeferredFromBB)
1038         ToSplit.insert(std::make_pair(DeferredFromBB, ToBB));
1039       Status = true;
1040     }
1041   }
1042 
1043   return Status;
1044 }
1045 
1046 std::vector<unsigned> &
1047 MachineSinking::getBBRegisterPressure(const MachineBasicBlock &MBB) {
1048   // Currently to save compiling time, MBB's register pressure will not change
1049   // in one ProcessBlock iteration because of CachedRegisterPressure. but MBB's
1050   // register pressure is changed after sinking any instructions into it.
1051   // FIXME: need a accurate and cheap register pressure estiminate model here.
1052   auto RP = CachedRegisterPressure.find(&MBB);
1053   if (RP != CachedRegisterPressure.end())
1054     return RP->second;
1055 
1056   RegionPressure Pressure;
1057   RegPressureTracker RPTracker(Pressure);
1058 
1059   // Initialize the register pressure tracker.
1060   RPTracker.init(MBB.getParent(), &RegClassInfo, nullptr, &MBB, MBB.end(),
1061                  /*TrackLaneMasks*/ false, /*TrackUntiedDefs=*/true);
1062 
1063   for (MachineBasicBlock::const_iterator MII = MBB.instr_end(),
1064                                          MIE = MBB.instr_begin();
1065        MII != MIE; --MII) {
1066     const MachineInstr &MI = *std::prev(MII);
1067     if (MI.isDebugInstr() || MI.isPseudoProbe())
1068       continue;
1069     RegisterOperands RegOpers;
1070     RegOpers.collect(MI, *TRI, *MRI, false, false);
1071     RPTracker.recedeSkipDebugValues();
1072     assert(&*RPTracker.getPos() == &MI && "RPTracker sync error!");
1073     RPTracker.recede(RegOpers);
1074   }
1075 
1076   RPTracker.closeRegion();
1077   auto It = CachedRegisterPressure.insert(
1078       std::make_pair(&MBB, RPTracker.getPressure().MaxSetPressure));
1079   return It.first->second;
1080 }
1081 
1082 bool MachineSinking::registerPressureSetExceedsLimit(
1083     unsigned NRegs, const TargetRegisterClass *RC,
1084     const MachineBasicBlock &MBB) {
1085   unsigned Weight = NRegs * TRI->getRegClassWeight(RC).RegWeight;
1086   const int *PS = TRI->getRegClassPressureSets(RC);
1087   std::vector<unsigned> BBRegisterPressure = getBBRegisterPressure(MBB);
1088   for (; *PS != -1; PS++)
1089     if (Weight + BBRegisterPressure[*PS] >=
1090         TRI->getRegPressureSetLimit(*MBB.getParent(), *PS))
1091       return true;
1092   return false;
1093 }
1094 
1095 /// isProfitableToSinkTo - Return true if it is profitable to sink MI.
1096 bool MachineSinking::isProfitableToSinkTo(Register Reg, MachineInstr &MI,
1097                                           MachineBasicBlock *MBB,
1098                                           MachineBasicBlock *SuccToSinkTo,
1099                                           AllSuccsCache &AllSuccessors) {
1100   assert (SuccToSinkTo && "Invalid SinkTo Candidate BB");
1101 
1102   if (MBB == SuccToSinkTo)
1103     return false;
1104 
1105   // It is profitable if SuccToSinkTo does not post dominate current block.
1106   if (!PDT->dominates(SuccToSinkTo, MBB))
1107     return true;
1108 
1109   // It is profitable to sink an instruction from a deeper cycle to a shallower
1110   // cycle, even if the latter post-dominates the former (PR21115).
1111   if (CI->getCycleDepth(MBB) > CI->getCycleDepth(SuccToSinkTo))
1112     return true;
1113 
1114   // Check if only use in post dominated block is PHI instruction.
1115   bool NonPHIUse = false;
1116   for (MachineInstr &UseInst : MRI->use_nodbg_instructions(Reg)) {
1117     MachineBasicBlock *UseBlock = UseInst.getParent();
1118     if (UseBlock == SuccToSinkTo && !UseInst.isPHI())
1119       NonPHIUse = true;
1120   }
1121   if (!NonPHIUse)
1122     return true;
1123 
1124   // If SuccToSinkTo post dominates then also it may be profitable if MI
1125   // can further profitably sinked into another block in next round.
1126   bool BreakPHIEdge = false;
1127   // FIXME - If finding successor is compile time expensive then cache results.
1128   if (MachineBasicBlock *MBB2 =
1129           FindSuccToSinkTo(MI, SuccToSinkTo, BreakPHIEdge, AllSuccessors))
1130     return isProfitableToSinkTo(Reg, MI, SuccToSinkTo, MBB2, AllSuccessors);
1131 
1132   MachineCycle *MCycle = CI->getCycle(MBB);
1133 
1134   // If the instruction is not inside a cycle, it is not profitable to sink MI to
1135   // a post dominate block SuccToSinkTo.
1136   if (!MCycle)
1137     return false;
1138 
1139   // If this instruction is inside a Cycle and sinking this instruction can make
1140   // more registers live range shorten, it is still prifitable.
1141   for (const MachineOperand &MO : MI.operands()) {
1142     // Ignore non-register operands.
1143     if (!MO.isReg())
1144       continue;
1145     Register Reg = MO.getReg();
1146     if (Reg == 0)
1147       continue;
1148 
1149     if (Reg.isPhysical()) {
1150       // Don't handle non-constant and non-ignorable physical register uses.
1151       if (MO.isUse() && !MRI->isConstantPhysReg(Reg) && !TII->isIgnorableUse(MO))
1152         return false;
1153       continue;
1154     }
1155 
1156     // Users for the defs are all dominated by SuccToSinkTo.
1157     if (MO.isDef()) {
1158       // This def register's live range is shortened after sinking.
1159       bool LocalUse = false;
1160       if (!AllUsesDominatedByBlock(Reg, SuccToSinkTo, MBB, BreakPHIEdge,
1161                                    LocalUse))
1162         return false;
1163     } else {
1164       MachineInstr *DefMI = MRI->getVRegDef(Reg);
1165       if (!DefMI)
1166         continue;
1167       MachineCycle *Cycle = CI->getCycle(DefMI->getParent());
1168       // DefMI is defined outside of cycle. There should be no live range
1169       // impact for this operand. Defination outside of cycle means:
1170       // 1: defination is outside of cycle.
1171       // 2: defination is in this cycle, but it is a PHI in the cycle header.
1172       if (Cycle != MCycle || (DefMI->isPHI() && Cycle && Cycle->isReducible() &&
1173                               Cycle->getHeader() == DefMI->getParent()))
1174         continue;
1175       // The DefMI is defined inside the cycle.
1176       // If sinking this operand makes some register pressure set exceed limit,
1177       // it is not profitable.
1178       if (registerPressureSetExceedsLimit(1, MRI->getRegClass(Reg),
1179                                           *SuccToSinkTo)) {
1180         LLVM_DEBUG(dbgs() << "register pressure exceed limit, not profitable.");
1181         return false;
1182       }
1183     }
1184   }
1185 
1186   // If MI is in cycle and all its operands are alive across the whole cycle or
1187   // if no operand sinking make register pressure set exceed limit, it is
1188   // profitable to sink MI.
1189   return true;
1190 }
1191 
1192 /// Get the sorted sequence of successors for this MachineBasicBlock, possibly
1193 /// computing it if it was not already cached.
1194 SmallVector<MachineBasicBlock *, 4> &
1195 MachineSinking::GetAllSortedSuccessors(MachineInstr &MI, MachineBasicBlock *MBB,
1196                                        AllSuccsCache &AllSuccessors) const {
1197   // Do we have the sorted successors in cache ?
1198   auto Succs = AllSuccessors.find(MBB);
1199   if (Succs != AllSuccessors.end())
1200     return Succs->second;
1201 
1202   SmallVector<MachineBasicBlock *, 4> AllSuccs(MBB->successors());
1203 
1204   // Handle cases where sinking can happen but where the sink point isn't a
1205   // successor. For example:
1206   //
1207   //   x = computation
1208   //   if () {} else {}
1209   //   use x
1210   //
1211   for (MachineDomTreeNode *DTChild : DT->getNode(MBB)->children()) {
1212     // DomTree children of MBB that have MBB as immediate dominator are added.
1213     if (DTChild->getIDom()->getBlock() == MI.getParent() &&
1214         // Skip MBBs already added to the AllSuccs vector above.
1215         !MBB->isSuccessor(DTChild->getBlock()))
1216       AllSuccs.push_back(DTChild->getBlock());
1217   }
1218 
1219   // Sort Successors according to their cycle depth or block frequency info.
1220   llvm::stable_sort(
1221       AllSuccs, [this](const MachineBasicBlock *L, const MachineBasicBlock *R) {
1222         uint64_t LHSFreq = MBFI ? MBFI->getBlockFreq(L).getFrequency() : 0;
1223         uint64_t RHSFreq = MBFI ? MBFI->getBlockFreq(R).getFrequency() : 0;
1224         bool HasBlockFreq = LHSFreq != 0 || RHSFreq != 0;
1225         return HasBlockFreq ? LHSFreq < RHSFreq
1226                             : CI->getCycleDepth(L) < CI->getCycleDepth(R);
1227       });
1228 
1229   auto it = AllSuccessors.insert(std::make_pair(MBB, AllSuccs));
1230 
1231   return it.first->second;
1232 }
1233 
1234 /// FindSuccToSinkTo - Find a successor to sink this instruction to.
1235 MachineBasicBlock *
1236 MachineSinking::FindSuccToSinkTo(MachineInstr &MI, MachineBasicBlock *MBB,
1237                                  bool &BreakPHIEdge,
1238                                  AllSuccsCache &AllSuccessors) {
1239   assert (MBB && "Invalid MachineBasicBlock!");
1240 
1241   // loop over all the operands of the specified instruction.  If there is
1242   // anything we can't handle, bail out.
1243 
1244   // SuccToSinkTo - This is the successor to sink this instruction to, once we
1245   // decide.
1246   MachineBasicBlock *SuccToSinkTo = nullptr;
1247   for (const MachineOperand &MO : MI.operands()) {
1248     if (!MO.isReg()) continue;  // Ignore non-register operands.
1249 
1250     Register Reg = MO.getReg();
1251     if (Reg == 0) continue;
1252 
1253     if (Reg.isPhysical()) {
1254       if (MO.isUse()) {
1255         // If the physreg has no defs anywhere, it's just an ambient register
1256         // and we can freely move its uses. Alternatively, if it's allocatable,
1257         // it could get allocated to something with a def during allocation.
1258         if (!MRI->isConstantPhysReg(Reg) && !TII->isIgnorableUse(MO))
1259           return nullptr;
1260       } else if (!MO.isDead()) {
1261         // A def that isn't dead. We can't move it.
1262         return nullptr;
1263       }
1264     } else {
1265       // Virtual register uses are always safe to sink.
1266       if (MO.isUse()) continue;
1267 
1268       // If it's not safe to move defs of the register class, then abort.
1269       if (!TII->isSafeToMoveRegClassDefs(MRI->getRegClass(Reg)))
1270         return nullptr;
1271 
1272       // Virtual register defs can only be sunk if all their uses are in blocks
1273       // dominated by one of the successors.
1274       if (SuccToSinkTo) {
1275         // If a previous operand picked a block to sink to, then this operand
1276         // must be sinkable to the same block.
1277         bool LocalUse = false;
1278         if (!AllUsesDominatedByBlock(Reg, SuccToSinkTo, MBB,
1279                                      BreakPHIEdge, LocalUse))
1280           return nullptr;
1281 
1282         continue;
1283       }
1284 
1285       // Otherwise, we should look at all the successors and decide which one
1286       // we should sink to. If we have reliable block frequency information
1287       // (frequency != 0) available, give successors with smaller frequencies
1288       // higher priority, otherwise prioritize smaller cycle depths.
1289       for (MachineBasicBlock *SuccBlock :
1290            GetAllSortedSuccessors(MI, MBB, AllSuccessors)) {
1291         bool LocalUse = false;
1292         if (AllUsesDominatedByBlock(Reg, SuccBlock, MBB,
1293                                     BreakPHIEdge, LocalUse)) {
1294           SuccToSinkTo = SuccBlock;
1295           break;
1296         }
1297         if (LocalUse)
1298           // Def is used locally, it's never safe to move this def.
1299           return nullptr;
1300       }
1301 
1302       // If we couldn't find a block to sink to, ignore this instruction.
1303       if (!SuccToSinkTo)
1304         return nullptr;
1305       if (!isProfitableToSinkTo(Reg, MI, MBB, SuccToSinkTo, AllSuccessors))
1306         return nullptr;
1307     }
1308   }
1309 
1310   // It is not possible to sink an instruction into its own block.  This can
1311   // happen with cycles.
1312   if (MBB == SuccToSinkTo)
1313     return nullptr;
1314 
1315   // It's not safe to sink instructions to EH landing pad. Control flow into
1316   // landing pad is implicitly defined.
1317   if (SuccToSinkTo && SuccToSinkTo->isEHPad())
1318     return nullptr;
1319 
1320   // It ought to be okay to sink instructions into an INLINEASM_BR target, but
1321   // only if we make sure that MI occurs _before_ an INLINEASM_BR instruction in
1322   // the source block (which this code does not yet do). So for now, forbid
1323   // doing so.
1324   if (SuccToSinkTo && SuccToSinkTo->isInlineAsmBrIndirectTarget())
1325     return nullptr;
1326 
1327   if (SuccToSinkTo && !TII->isSafeToSink(MI, SuccToSinkTo, CI))
1328     return nullptr;
1329 
1330   return SuccToSinkTo;
1331 }
1332 
1333 /// Return true if MI is likely to be usable as a memory operation by the
1334 /// implicit null check optimization.
1335 ///
1336 /// This is a "best effort" heuristic, and should not be relied upon for
1337 /// correctness.  This returning true does not guarantee that the implicit null
1338 /// check optimization is legal over MI, and this returning false does not
1339 /// guarantee MI cannot possibly be used to do a null check.
1340 static bool SinkingPreventsImplicitNullCheck(MachineInstr &MI,
1341                                              const TargetInstrInfo *TII,
1342                                              const TargetRegisterInfo *TRI) {
1343   using MachineBranchPredicate = TargetInstrInfo::MachineBranchPredicate;
1344 
1345   auto *MBB = MI.getParent();
1346   if (MBB->pred_size() != 1)
1347     return false;
1348 
1349   auto *PredMBB = *MBB->pred_begin();
1350   auto *PredBB = PredMBB->getBasicBlock();
1351 
1352   // Frontends that don't use implicit null checks have no reason to emit
1353   // branches with make.implicit metadata, and this function should always
1354   // return false for them.
1355   if (!PredBB ||
1356       !PredBB->getTerminator()->getMetadata(LLVMContext::MD_make_implicit))
1357     return false;
1358 
1359   const MachineOperand *BaseOp;
1360   int64_t Offset;
1361   bool OffsetIsScalable;
1362   if (!TII->getMemOperandWithOffset(MI, BaseOp, Offset, OffsetIsScalable, TRI))
1363     return false;
1364 
1365   if (!BaseOp->isReg())
1366     return false;
1367 
1368   if (!(MI.mayLoad() && !MI.isPredicable()))
1369     return false;
1370 
1371   MachineBranchPredicate MBP;
1372   if (TII->analyzeBranchPredicate(*PredMBB, MBP, false))
1373     return false;
1374 
1375   return MBP.LHS.isReg() && MBP.RHS.isImm() && MBP.RHS.getImm() == 0 &&
1376          (MBP.Predicate == MachineBranchPredicate::PRED_NE ||
1377           MBP.Predicate == MachineBranchPredicate::PRED_EQ) &&
1378          MBP.LHS.getReg() == BaseOp->getReg();
1379 }
1380 
1381 /// If the sunk instruction is a copy, try to forward the copy instead of
1382 /// leaving an 'undef' DBG_VALUE in the original location. Don't do this if
1383 /// there's any subregister weirdness involved. Returns true if copy
1384 /// propagation occurred.
1385 static bool attemptDebugCopyProp(MachineInstr &SinkInst, MachineInstr &DbgMI,
1386                                  Register Reg) {
1387   const MachineRegisterInfo &MRI = SinkInst.getMF()->getRegInfo();
1388   const TargetInstrInfo &TII = *SinkInst.getMF()->getSubtarget().getInstrInfo();
1389 
1390   // Copy DBG_VALUE operand and set the original to undef. We then check to
1391   // see whether this is something that can be copy-forwarded. If it isn't,
1392   // continue around the loop.
1393 
1394   const MachineOperand *SrcMO = nullptr, *DstMO = nullptr;
1395   auto CopyOperands = TII.isCopyInstr(SinkInst);
1396   if (!CopyOperands)
1397     return false;
1398   SrcMO = CopyOperands->Source;
1399   DstMO = CopyOperands->Destination;
1400 
1401   // Check validity of forwarding this copy.
1402   bool PostRA = MRI.getNumVirtRegs() == 0;
1403 
1404   // Trying to forward between physical and virtual registers is too hard.
1405   if (Reg.isVirtual() != SrcMO->getReg().isVirtual())
1406     return false;
1407 
1408   // Only try virtual register copy-forwarding before regalloc, and physical
1409   // register copy-forwarding after regalloc.
1410   bool arePhysRegs = !Reg.isVirtual();
1411   if (arePhysRegs != PostRA)
1412     return false;
1413 
1414   // Pre-regalloc, only forward if all subregisters agree (or there are no
1415   // subregs at all). More analysis might recover some forwardable copies.
1416   if (!PostRA)
1417     for (auto &DbgMO : DbgMI.getDebugOperandsForReg(Reg))
1418       if (DbgMO.getSubReg() != SrcMO->getSubReg() ||
1419           DbgMO.getSubReg() != DstMO->getSubReg())
1420         return false;
1421 
1422   // Post-regalloc, we may be sinking a DBG_VALUE of a sub or super-register
1423   // of this copy. Only forward the copy if the DBG_VALUE operand exactly
1424   // matches the copy destination.
1425   if (PostRA && Reg != DstMO->getReg())
1426     return false;
1427 
1428   for (auto &DbgMO : DbgMI.getDebugOperandsForReg(Reg)) {
1429     DbgMO.setReg(SrcMO->getReg());
1430     DbgMO.setSubReg(SrcMO->getSubReg());
1431   }
1432   return true;
1433 }
1434 
1435 using MIRegs = std::pair<MachineInstr *, SmallVector<unsigned, 2>>;
1436 /// Sink an instruction and its associated debug instructions.
1437 static void performSink(MachineInstr &MI, MachineBasicBlock &SuccToSinkTo,
1438                         MachineBasicBlock::iterator InsertPos,
1439                         ArrayRef<MIRegs> DbgValuesToSink) {
1440   // If we cannot find a location to use (merge with), then we erase the debug
1441   // location to prevent debug-info driven tools from potentially reporting
1442   // wrong location information.
1443   if (!SuccToSinkTo.empty() && InsertPos != SuccToSinkTo.end())
1444     MI.setDebugLoc(DILocation::getMergedLocation(MI.getDebugLoc(),
1445                                                  InsertPos->getDebugLoc()));
1446   else
1447     MI.setDebugLoc(DebugLoc());
1448 
1449   // Move the instruction.
1450   MachineBasicBlock *ParentBlock = MI.getParent();
1451   SuccToSinkTo.splice(InsertPos, ParentBlock, MI,
1452                       ++MachineBasicBlock::iterator(MI));
1453 
1454   // Sink a copy of debug users to the insert position. Mark the original
1455   // DBG_VALUE location as 'undef', indicating that any earlier variable
1456   // location should be terminated as we've optimised away the value at this
1457   // point.
1458   for (const auto &DbgValueToSink : DbgValuesToSink) {
1459     MachineInstr *DbgMI = DbgValueToSink.first;
1460     MachineInstr *NewDbgMI = DbgMI->getMF()->CloneMachineInstr(DbgMI);
1461     SuccToSinkTo.insert(InsertPos, NewDbgMI);
1462 
1463     bool PropagatedAllSunkOps = true;
1464     for (unsigned Reg : DbgValueToSink.second) {
1465       if (DbgMI->hasDebugOperandForReg(Reg)) {
1466         if (!attemptDebugCopyProp(MI, *DbgMI, Reg)) {
1467           PropagatedAllSunkOps = false;
1468           break;
1469         }
1470       }
1471     }
1472     if (!PropagatedAllSunkOps)
1473       DbgMI->setDebugValueUndef();
1474   }
1475 }
1476 
1477 /// hasStoreBetween - check if there is store betweeen straight line blocks From
1478 /// and To.
1479 bool MachineSinking::hasStoreBetween(MachineBasicBlock *From,
1480                                      MachineBasicBlock *To, MachineInstr &MI) {
1481   // Make sure From and To are in straight line which means From dominates To
1482   // and To post dominates From.
1483   if (!DT->dominates(From, To) || !PDT->dominates(To, From))
1484     return true;
1485 
1486   auto BlockPair = std::make_pair(From, To);
1487 
1488   // Does these two blocks pair be queried before and have a definite cached
1489   // result?
1490   if (auto It = HasStoreCache.find(BlockPair); It != HasStoreCache.end())
1491     return It->second;
1492 
1493   if (auto It = StoreInstrCache.find(BlockPair); It != StoreInstrCache.end())
1494     return llvm::any_of(It->second, [&](MachineInstr *I) {
1495       return I->mayAlias(AA, MI, false);
1496     });
1497 
1498   bool SawStore = false;
1499   bool HasAliasedStore = false;
1500   DenseSet<MachineBasicBlock *> HandledBlocks;
1501   DenseSet<MachineBasicBlock *> HandledDomBlocks;
1502   // Go through all reachable blocks from From.
1503   for (MachineBasicBlock *BB : depth_first(From)) {
1504     // We insert the instruction at the start of block To, so no need to worry
1505     // about stores inside To.
1506     // Store in block From should be already considered when just enter function
1507     // SinkInstruction.
1508     if (BB == To || BB == From)
1509       continue;
1510 
1511     // We already handle this BB in previous iteration.
1512     if (HandledBlocks.count(BB))
1513       continue;
1514 
1515     HandledBlocks.insert(BB);
1516     // To post dominates BB, it must be a path from block From.
1517     if (PDT->dominates(To, BB)) {
1518       if (!HandledDomBlocks.count(BB))
1519         HandledDomBlocks.insert(BB);
1520 
1521       // If this BB is too big or the block number in straight line between From
1522       // and To is too big, stop searching to save compiling time.
1523       if (BB->sizeWithoutDebugLargerThan(SinkLoadInstsPerBlockThreshold) ||
1524           HandledDomBlocks.size() > SinkLoadBlocksThreshold) {
1525         for (auto *DomBB : HandledDomBlocks) {
1526           if (DomBB != BB && DT->dominates(DomBB, BB))
1527             HasStoreCache[std::make_pair(DomBB, To)] = true;
1528           else if(DomBB != BB && DT->dominates(BB, DomBB))
1529             HasStoreCache[std::make_pair(From, DomBB)] = true;
1530         }
1531         HasStoreCache[BlockPair] = true;
1532         return true;
1533       }
1534 
1535       for (MachineInstr &I : *BB) {
1536         // Treat as alias conservatively for a call or an ordered memory
1537         // operation.
1538         if (I.isCall() || I.hasOrderedMemoryRef()) {
1539           for (auto *DomBB : HandledDomBlocks) {
1540             if (DomBB != BB && DT->dominates(DomBB, BB))
1541               HasStoreCache[std::make_pair(DomBB, To)] = true;
1542             else if(DomBB != BB && DT->dominates(BB, DomBB))
1543               HasStoreCache[std::make_pair(From, DomBB)] = true;
1544           }
1545           HasStoreCache[BlockPair] = true;
1546           return true;
1547         }
1548 
1549         if (I.mayStore()) {
1550           SawStore = true;
1551           // We still have chance to sink MI if all stores between are not
1552           // aliased to MI.
1553           // Cache all store instructions, so that we don't need to go through
1554           // all From reachable blocks for next load instruction.
1555           if (I.mayAlias(AA, MI, false))
1556             HasAliasedStore = true;
1557           StoreInstrCache[BlockPair].push_back(&I);
1558         }
1559       }
1560     }
1561   }
1562   // If there is no store at all, cache the result.
1563   if (!SawStore)
1564     HasStoreCache[BlockPair] = false;
1565   return HasAliasedStore;
1566 }
1567 
1568 /// Sink instructions into cycles if profitable. This especially tries to
1569 /// prevent register spills caused by register pressure if there is little to no
1570 /// overhead moving instructions into cycles.
1571 bool MachineSinking::SinkIntoCycle(MachineCycle *Cycle, MachineInstr &I) {
1572   LLVM_DEBUG(dbgs() << "CycleSink: Finding sink block for: " << I);
1573   MachineBasicBlock *Preheader = Cycle->getCyclePreheader();
1574   assert(Preheader && "Cycle sink needs a preheader block");
1575   MachineBasicBlock *SinkBlock = nullptr;
1576   bool CanSink = true;
1577   const MachineOperand &MO = I.getOperand(0);
1578 
1579   for (MachineInstr &MI : MRI->use_instructions(MO.getReg())) {
1580     LLVM_DEBUG(dbgs() << "CycleSink:   Analysing use: " << MI);
1581     if (!Cycle->contains(MI.getParent())) {
1582       LLVM_DEBUG(dbgs() << "CycleSink:   Use not in cycle, can't sink.\n");
1583       CanSink = false;
1584       break;
1585     }
1586 
1587     // FIXME: Come up with a proper cost model that estimates whether sinking
1588     // the instruction (and thus possibly executing it on every cycle
1589     // iteration) is more expensive than a register.
1590     // For now assumes that copies are cheap and thus almost always worth it.
1591     if (!MI.isCopy()) {
1592       LLVM_DEBUG(dbgs() << "CycleSink:   Use is not a copy\n");
1593       CanSink = false;
1594       break;
1595     }
1596     if (!SinkBlock) {
1597       SinkBlock = MI.getParent();
1598       LLVM_DEBUG(dbgs() << "CycleSink:   Setting sink block to: "
1599                         << printMBBReference(*SinkBlock) << "\n");
1600       continue;
1601     }
1602     SinkBlock = DT->findNearestCommonDominator(SinkBlock, MI.getParent());
1603     if (!SinkBlock) {
1604       LLVM_DEBUG(dbgs() << "CycleSink:   Can't find nearest dominator\n");
1605       CanSink = false;
1606       break;
1607     }
1608     LLVM_DEBUG(dbgs() << "CycleSink:   Setting nearest common dom block: " <<
1609                printMBBReference(*SinkBlock) << "\n");
1610   }
1611 
1612   if (!CanSink) {
1613     LLVM_DEBUG(dbgs() << "CycleSink: Can't sink instruction.\n");
1614     return false;
1615   }
1616   if (!SinkBlock) {
1617     LLVM_DEBUG(dbgs() << "CycleSink: Not sinking, can't find sink block.\n");
1618     return false;
1619   }
1620   if (SinkBlock == Preheader) {
1621     LLVM_DEBUG(
1622         dbgs() << "CycleSink: Not sinking, sink block is the preheader\n");
1623     return false;
1624   }
1625   if (SinkBlock->sizeWithoutDebugLargerThan(SinkLoadInstsPerBlockThreshold)) {
1626     LLVM_DEBUG(
1627         dbgs() << "CycleSink: Not Sinking, block too large to analyse.\n");
1628     return false;
1629   }
1630 
1631   LLVM_DEBUG(dbgs() << "CycleSink: Sinking instruction!\n");
1632   SinkBlock->splice(SinkBlock->SkipPHIsAndLabels(SinkBlock->begin()), Preheader,
1633                     I);
1634 
1635   // Conservatively clear any kill flags on uses of sunk instruction
1636   for (MachineOperand &MO : I.operands()) {
1637     if (MO.isReg() && MO.readsReg())
1638       RegsToClearKillFlags.insert(MO.getReg());
1639   }
1640 
1641   // The instruction is moved from its basic block, so do not retain the
1642   // debug information.
1643   assert(!I.isDebugInstr() && "Should not sink debug inst");
1644   I.setDebugLoc(DebugLoc());
1645   return true;
1646 }
1647 
1648 /// SinkInstruction - Determine whether it is safe to sink the specified machine
1649 /// instruction out of its current block into a successor.
1650 bool MachineSinking::SinkInstruction(MachineInstr &MI, bool &SawStore,
1651                                      AllSuccsCache &AllSuccessors) {
1652   // Don't sink instructions that the target prefers not to sink.
1653   if (!TII->shouldSink(MI))
1654     return false;
1655 
1656   // Check if it's safe to move the instruction.
1657   if (!MI.isSafeToMove(AA, SawStore))
1658     return false;
1659 
1660   // Convergent operations may not be made control-dependent on additional
1661   // values.
1662   if (MI.isConvergent())
1663     return false;
1664 
1665   // Don't break implicit null checks.  This is a performance heuristic, and not
1666   // required for correctness.
1667   if (SinkingPreventsImplicitNullCheck(MI, TII, TRI))
1668     return false;
1669 
1670   // FIXME: This should include support for sinking instructions within the
1671   // block they are currently in to shorten the live ranges.  We often get
1672   // instructions sunk into the top of a large block, but it would be better to
1673   // also sink them down before their first use in the block.  This xform has to
1674   // be careful not to *increase* register pressure though, e.g. sinking
1675   // "x = y + z" down if it kills y and z would increase the live ranges of y
1676   // and z and only shrink the live range of x.
1677 
1678   bool BreakPHIEdge = false;
1679   MachineBasicBlock *ParentBlock = MI.getParent();
1680   MachineBasicBlock *SuccToSinkTo =
1681       FindSuccToSinkTo(MI, ParentBlock, BreakPHIEdge, AllSuccessors);
1682 
1683   // If there are no outputs, it must have side-effects.
1684   if (!SuccToSinkTo)
1685     return false;
1686 
1687   // If the instruction to move defines a dead physical register which is live
1688   // when leaving the basic block, don't move it because it could turn into a
1689   // "zombie" define of that preg. E.g., EFLAGS.
1690   for (const MachineOperand &MO : MI.all_defs()) {
1691     Register Reg = MO.getReg();
1692     if (Reg == 0 || !Reg.isPhysical())
1693       continue;
1694     if (SuccToSinkTo->isLiveIn(Reg))
1695       return false;
1696   }
1697 
1698   LLVM_DEBUG(dbgs() << "Sink instr " << MI << "\tinto block " << *SuccToSinkTo);
1699 
1700   // If the block has multiple predecessors, this is a critical edge.
1701   // Decide if we can sink along it or need to break the edge.
1702   if (SuccToSinkTo->pred_size() > 1) {
1703     // We cannot sink a load across a critical edge - there may be stores in
1704     // other code paths.
1705     bool TryBreak = false;
1706     bool Store =
1707         MI.mayLoad() ? hasStoreBetween(ParentBlock, SuccToSinkTo, MI) : true;
1708     if (!MI.isSafeToMove(AA, Store)) {
1709       LLVM_DEBUG(dbgs() << " *** NOTE: Won't sink load along critical edge.\n");
1710       TryBreak = true;
1711     }
1712 
1713     // We don't want to sink across a critical edge if we don't dominate the
1714     // successor. We could be introducing calculations to new code paths.
1715     if (!TryBreak && !DT->dominates(ParentBlock, SuccToSinkTo)) {
1716       LLVM_DEBUG(dbgs() << " *** NOTE: Critical edge found\n");
1717       TryBreak = true;
1718     }
1719 
1720     // Don't sink instructions into a cycle.
1721     if (!TryBreak && CI->getCycle(SuccToSinkTo) &&
1722         (!CI->getCycle(SuccToSinkTo)->isReducible() ||
1723          CI->getCycle(SuccToSinkTo)->getHeader() == SuccToSinkTo)) {
1724       LLVM_DEBUG(dbgs() << " *** NOTE: cycle header found\n");
1725       TryBreak = true;
1726     }
1727 
1728     // Otherwise we are OK with sinking along a critical edge.
1729     if (!TryBreak)
1730       LLVM_DEBUG(dbgs() << "Sinking along critical edge.\n");
1731     else {
1732       // Mark this edge as to be split.
1733       // If the edge can actually be split, the next iteration of the main loop
1734       // will sink MI in the newly created block.
1735       bool Status =
1736         PostponeSplitCriticalEdge(MI, ParentBlock, SuccToSinkTo, BreakPHIEdge);
1737       if (!Status)
1738         LLVM_DEBUG(dbgs() << " *** PUNTING: Not legal or profitable to "
1739                              "break critical edge\n");
1740       // The instruction will not be sunk this time.
1741       return false;
1742     }
1743   }
1744 
1745   if (BreakPHIEdge) {
1746     // BreakPHIEdge is true if all the uses are in the successor MBB being
1747     // sunken into and they are all PHI nodes. In this case, machine-sink must
1748     // break the critical edge first.
1749     bool Status = PostponeSplitCriticalEdge(MI, ParentBlock,
1750                                             SuccToSinkTo, BreakPHIEdge);
1751     if (!Status)
1752       LLVM_DEBUG(dbgs() << " *** PUNTING: Not legal or profitable to "
1753                            "break critical edge\n");
1754     // The instruction will not be sunk this time.
1755     return false;
1756   }
1757 
1758   // Determine where to insert into. Skip phi nodes.
1759   MachineBasicBlock::iterator InsertPos =
1760       SuccToSinkTo->SkipPHIsAndLabels(SuccToSinkTo->begin());
1761   if (blockPrologueInterferes(SuccToSinkTo, InsertPos, MI, TRI, TII, MRI)) {
1762     LLVM_DEBUG(dbgs() << " *** Not sinking: prologue interference\n");
1763     return false;
1764   }
1765 
1766   // Collect debug users of any vreg that this inst defines.
1767   SmallVector<MIRegs, 4> DbgUsersToSink;
1768   for (auto &MO : MI.all_defs()) {
1769     if (!MO.getReg().isVirtual())
1770       continue;
1771     if (!SeenDbgUsers.count(MO.getReg()))
1772       continue;
1773 
1774     // Sink any users that don't pass any other DBG_VALUEs for this variable.
1775     auto &Users = SeenDbgUsers[MO.getReg()];
1776     for (auto &User : Users) {
1777       MachineInstr *DbgMI = User.getPointer();
1778       if (User.getInt()) {
1779         // This DBG_VALUE would re-order assignments. If we can't copy-propagate
1780         // it, it can't be recovered. Set it undef.
1781         if (!attemptDebugCopyProp(MI, *DbgMI, MO.getReg()))
1782           DbgMI->setDebugValueUndef();
1783       } else {
1784         DbgUsersToSink.push_back(
1785             {DbgMI, SmallVector<unsigned, 2>(1, MO.getReg())});
1786       }
1787     }
1788   }
1789 
1790   // After sinking, some debug users may not be dominated any more. If possible,
1791   // copy-propagate their operands. As it's expensive, don't do this if there's
1792   // no debuginfo in the program.
1793   if (MI.getMF()->getFunction().getSubprogram() && MI.isCopy())
1794     SalvageUnsunkDebugUsersOfCopy(MI, SuccToSinkTo);
1795 
1796   performSink(MI, *SuccToSinkTo, InsertPos, DbgUsersToSink);
1797 
1798   // Conservatively, clear any kill flags, since it's possible that they are no
1799   // longer correct.
1800   // Note that we have to clear the kill flags for any register this instruction
1801   // uses as we may sink over another instruction which currently kills the
1802   // used registers.
1803   for (MachineOperand &MO : MI.all_uses())
1804     RegsToClearKillFlags.insert(MO.getReg()); // Remember to clear kill flags.
1805 
1806   return true;
1807 }
1808 
1809 void MachineSinking::SalvageUnsunkDebugUsersOfCopy(
1810     MachineInstr &MI, MachineBasicBlock *TargetBlock) {
1811   assert(MI.isCopy());
1812   assert(MI.getOperand(1).isReg());
1813 
1814   // Enumerate all users of vreg operands that are def'd. Skip those that will
1815   // be sunk. For the rest, if they are not dominated by the block we will sink
1816   // MI into, propagate the copy source to them.
1817   SmallVector<MachineInstr *, 4> DbgDefUsers;
1818   SmallVector<Register, 4> DbgUseRegs;
1819   const MachineRegisterInfo &MRI = MI.getMF()->getRegInfo();
1820   for (auto &MO : MI.all_defs()) {
1821     if (!MO.getReg().isVirtual())
1822       continue;
1823     DbgUseRegs.push_back(MO.getReg());
1824     for (auto &User : MRI.use_instructions(MO.getReg())) {
1825       if (!User.isDebugValue() || DT->dominates(TargetBlock, User.getParent()))
1826         continue;
1827 
1828       // If is in same block, will either sink or be use-before-def.
1829       if (User.getParent() == MI.getParent())
1830         continue;
1831 
1832       assert(User.hasDebugOperandForReg(MO.getReg()) &&
1833              "DBG_VALUE user of vreg, but has no operand for it?");
1834       DbgDefUsers.push_back(&User);
1835     }
1836   }
1837 
1838   // Point the users of this copy that are no longer dominated, at the source
1839   // of the copy.
1840   for (auto *User : DbgDefUsers) {
1841     for (auto &Reg : DbgUseRegs) {
1842       for (auto &DbgOp : User->getDebugOperandsForReg(Reg)) {
1843         DbgOp.setReg(MI.getOperand(1).getReg());
1844         DbgOp.setSubReg(MI.getOperand(1).getSubReg());
1845       }
1846     }
1847   }
1848 }
1849 
1850 //===----------------------------------------------------------------------===//
1851 // This pass is not intended to be a replacement or a complete alternative
1852 // for the pre-ra machine sink pass. It is only designed to sink COPY
1853 // instructions which should be handled after RA.
1854 //
1855 // This pass sinks COPY instructions into a successor block, if the COPY is not
1856 // used in the current block and the COPY is live-in to a single successor
1857 // (i.e., doesn't require the COPY to be duplicated).  This avoids executing the
1858 // copy on paths where their results aren't needed.  This also exposes
1859 // additional opportunites for dead copy elimination and shrink wrapping.
1860 //
1861 // These copies were either not handled by or are inserted after the MachineSink
1862 // pass. As an example of the former case, the MachineSink pass cannot sink
1863 // COPY instructions with allocatable source registers; for AArch64 these type
1864 // of copy instructions are frequently used to move function parameters (PhyReg)
1865 // into virtual registers in the entry block.
1866 //
1867 // For the machine IR below, this pass will sink %w19 in the entry into its
1868 // successor (%bb.1) because %w19 is only live-in in %bb.1.
1869 // %bb.0:
1870 //   %wzr = SUBSWri %w1, 1
1871 //   %w19 = COPY %w0
1872 //   Bcc 11, %bb.2
1873 // %bb.1:
1874 //   Live Ins: %w19
1875 //   BL @fun
1876 //   %w0 = ADDWrr %w0, %w19
1877 //   RET %w0
1878 // %bb.2:
1879 //   %w0 = COPY %wzr
1880 //   RET %w0
1881 // As we sink %w19 (CSR in AArch64) into %bb.1, the shrink-wrapping pass will be
1882 // able to see %bb.0 as a candidate.
1883 //===----------------------------------------------------------------------===//
1884 namespace {
1885 
1886 class PostRAMachineSinking : public MachineFunctionPass {
1887 public:
1888   bool runOnMachineFunction(MachineFunction &MF) override;
1889 
1890   static char ID;
1891   PostRAMachineSinking() : MachineFunctionPass(ID) {}
1892   StringRef getPassName() const override { return "PostRA Machine Sink"; }
1893 
1894   void getAnalysisUsage(AnalysisUsage &AU) const override {
1895     AU.setPreservesCFG();
1896     MachineFunctionPass::getAnalysisUsage(AU);
1897   }
1898 
1899   MachineFunctionProperties getRequiredProperties() const override {
1900     return MachineFunctionProperties().set(
1901         MachineFunctionProperties::Property::NoVRegs);
1902   }
1903 
1904 private:
1905   /// Track which register units have been modified and used.
1906   LiveRegUnits ModifiedRegUnits, UsedRegUnits;
1907 
1908   /// Track DBG_VALUEs of (unmodified) register units. Each DBG_VALUE has an
1909   /// entry in this map for each unit it touches. The DBG_VALUE's entry
1910   /// consists of a pointer to the instruction itself, and a vector of registers
1911   /// referred to by the instruction that overlap the key register unit.
1912   DenseMap<unsigned, SmallVector<MIRegs, 2>> SeenDbgInstrs;
1913 
1914   /// Sink Copy instructions unused in the same block close to their uses in
1915   /// successors.
1916   bool tryToSinkCopy(MachineBasicBlock &BB, MachineFunction &MF,
1917                      const TargetRegisterInfo *TRI, const TargetInstrInfo *TII);
1918 };
1919 } // namespace
1920 
1921 char PostRAMachineSinking::ID = 0;
1922 char &llvm::PostRAMachineSinkingID = PostRAMachineSinking::ID;
1923 
1924 INITIALIZE_PASS(PostRAMachineSinking, "postra-machine-sink",
1925                 "PostRA Machine Sink", false, false)
1926 
1927 static bool aliasWithRegsInLiveIn(MachineBasicBlock &MBB, unsigned Reg,
1928                                   const TargetRegisterInfo *TRI) {
1929   LiveRegUnits LiveInRegUnits(*TRI);
1930   LiveInRegUnits.addLiveIns(MBB);
1931   return !LiveInRegUnits.available(Reg);
1932 }
1933 
1934 static MachineBasicBlock *
1935 getSingleLiveInSuccBB(MachineBasicBlock &CurBB,
1936                       const SmallPtrSetImpl<MachineBasicBlock *> &SinkableBBs,
1937                       unsigned Reg, const TargetRegisterInfo *TRI) {
1938   // Try to find a single sinkable successor in which Reg is live-in.
1939   MachineBasicBlock *BB = nullptr;
1940   for (auto *SI : SinkableBBs) {
1941     if (aliasWithRegsInLiveIn(*SI, Reg, TRI)) {
1942       // If BB is set here, Reg is live-in to at least two sinkable successors,
1943       // so quit.
1944       if (BB)
1945         return nullptr;
1946       BB = SI;
1947     }
1948   }
1949   // Reg is not live-in to any sinkable successors.
1950   if (!BB)
1951     return nullptr;
1952 
1953   // Check if any register aliased with Reg is live-in in other successors.
1954   for (auto *SI : CurBB.successors()) {
1955     if (!SinkableBBs.count(SI) && aliasWithRegsInLiveIn(*SI, Reg, TRI))
1956       return nullptr;
1957   }
1958   return BB;
1959 }
1960 
1961 static MachineBasicBlock *
1962 getSingleLiveInSuccBB(MachineBasicBlock &CurBB,
1963                       const SmallPtrSetImpl<MachineBasicBlock *> &SinkableBBs,
1964                       ArrayRef<unsigned> DefedRegsInCopy,
1965                       const TargetRegisterInfo *TRI) {
1966   MachineBasicBlock *SingleBB = nullptr;
1967   for (auto DefReg : DefedRegsInCopy) {
1968     MachineBasicBlock *BB =
1969         getSingleLiveInSuccBB(CurBB, SinkableBBs, DefReg, TRI);
1970     if (!BB || (SingleBB && SingleBB != BB))
1971       return nullptr;
1972     SingleBB = BB;
1973   }
1974   return SingleBB;
1975 }
1976 
1977 static void clearKillFlags(MachineInstr *MI, MachineBasicBlock &CurBB,
1978                            SmallVectorImpl<unsigned> &UsedOpsInCopy,
1979                            LiveRegUnits &UsedRegUnits,
1980                            const TargetRegisterInfo *TRI) {
1981   for (auto U : UsedOpsInCopy) {
1982     MachineOperand &MO = MI->getOperand(U);
1983     Register SrcReg = MO.getReg();
1984     if (!UsedRegUnits.available(SrcReg)) {
1985       MachineBasicBlock::iterator NI = std::next(MI->getIterator());
1986       for (MachineInstr &UI : make_range(NI, CurBB.end())) {
1987         if (UI.killsRegister(SrcReg, TRI)) {
1988           UI.clearRegisterKills(SrcReg, TRI);
1989           MO.setIsKill(true);
1990           break;
1991         }
1992       }
1993     }
1994   }
1995 }
1996 
1997 static void updateLiveIn(MachineInstr *MI, MachineBasicBlock *SuccBB,
1998                          SmallVectorImpl<unsigned> &UsedOpsInCopy,
1999                          SmallVectorImpl<unsigned> &DefedRegsInCopy) {
2000   MachineFunction &MF = *SuccBB->getParent();
2001   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
2002   for (unsigned DefReg : DefedRegsInCopy)
2003     for (MCPhysReg S : TRI->subregs_inclusive(DefReg))
2004       SuccBB->removeLiveIn(S);
2005   for (auto U : UsedOpsInCopy)
2006     SuccBB->addLiveIn(MI->getOperand(U).getReg());
2007   SuccBB->sortUniqueLiveIns();
2008 }
2009 
2010 static bool hasRegisterDependency(MachineInstr *MI,
2011                                   SmallVectorImpl<unsigned> &UsedOpsInCopy,
2012                                   SmallVectorImpl<unsigned> &DefedRegsInCopy,
2013                                   LiveRegUnits &ModifiedRegUnits,
2014                                   LiveRegUnits &UsedRegUnits) {
2015   bool HasRegDependency = false;
2016   for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2017     MachineOperand &MO = MI->getOperand(i);
2018     if (!MO.isReg())
2019       continue;
2020     Register Reg = MO.getReg();
2021     if (!Reg)
2022       continue;
2023     if (MO.isDef()) {
2024       if (!ModifiedRegUnits.available(Reg) || !UsedRegUnits.available(Reg)) {
2025         HasRegDependency = true;
2026         break;
2027       }
2028       DefedRegsInCopy.push_back(Reg);
2029 
2030       // FIXME: instead of isUse(), readsReg() would be a better fix here,
2031       // For example, we can ignore modifications in reg with undef. However,
2032       // it's not perfectly clear if skipping the internal read is safe in all
2033       // other targets.
2034     } else if (MO.isUse()) {
2035       if (!ModifiedRegUnits.available(Reg)) {
2036         HasRegDependency = true;
2037         break;
2038       }
2039       UsedOpsInCopy.push_back(i);
2040     }
2041   }
2042   return HasRegDependency;
2043 }
2044 
2045 bool PostRAMachineSinking::tryToSinkCopy(MachineBasicBlock &CurBB,
2046                                          MachineFunction &MF,
2047                                          const TargetRegisterInfo *TRI,
2048                                          const TargetInstrInfo *TII) {
2049   SmallPtrSet<MachineBasicBlock *, 2> SinkableBBs;
2050   // FIXME: For now, we sink only to a successor which has a single predecessor
2051   // so that we can directly sink COPY instructions to the successor without
2052   // adding any new block or branch instruction.
2053   for (MachineBasicBlock *SI : CurBB.successors())
2054     if (!SI->livein_empty() && SI->pred_size() == 1)
2055       SinkableBBs.insert(SI);
2056 
2057   if (SinkableBBs.empty())
2058     return false;
2059 
2060   bool Changed = false;
2061 
2062   // Track which registers have been modified and used between the end of the
2063   // block and the current instruction.
2064   ModifiedRegUnits.clear();
2065   UsedRegUnits.clear();
2066   SeenDbgInstrs.clear();
2067 
2068   for (MachineInstr &MI : llvm::make_early_inc_range(llvm::reverse(CurBB))) {
2069     // Track the operand index for use in Copy.
2070     SmallVector<unsigned, 2> UsedOpsInCopy;
2071     // Track the register number defed in Copy.
2072     SmallVector<unsigned, 2> DefedRegsInCopy;
2073 
2074     // We must sink this DBG_VALUE if its operand is sunk. To avoid searching
2075     // for DBG_VALUEs later, record them when they're encountered.
2076     if (MI.isDebugValue() && !MI.isDebugRef()) {
2077       SmallDenseMap<MCRegister, SmallVector<unsigned, 2>, 4> MIUnits;
2078       bool IsValid = true;
2079       for (MachineOperand &MO : MI.debug_operands()) {
2080         if (MO.isReg() && MO.getReg().isPhysical()) {
2081           // Bail if we can already tell the sink would be rejected, rather
2082           // than needlessly accumulating lots of DBG_VALUEs.
2083           if (hasRegisterDependency(&MI, UsedOpsInCopy, DefedRegsInCopy,
2084                                     ModifiedRegUnits, UsedRegUnits)) {
2085             IsValid = false;
2086             break;
2087           }
2088 
2089           // Record debug use of each reg unit.
2090           for (MCRegUnit Unit : TRI->regunits(MO.getReg()))
2091             MIUnits[Unit].push_back(MO.getReg());
2092         }
2093       }
2094       if (IsValid) {
2095         for (auto &RegOps : MIUnits)
2096           SeenDbgInstrs[RegOps.first].emplace_back(&MI,
2097                                                    std::move(RegOps.second));
2098       }
2099       continue;
2100     }
2101 
2102     if (MI.isDebugOrPseudoInstr())
2103       continue;
2104 
2105     // Do not move any instruction across function call.
2106     if (MI.isCall())
2107       return false;
2108 
2109     if (!MI.isCopy() || !MI.getOperand(0).isRenamable()) {
2110       LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits,
2111                                         TRI);
2112       continue;
2113     }
2114 
2115     // Don't sink the COPY if it would violate a register dependency.
2116     if (hasRegisterDependency(&MI, UsedOpsInCopy, DefedRegsInCopy,
2117                               ModifiedRegUnits, UsedRegUnits)) {
2118       LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits,
2119                                         TRI);
2120       continue;
2121     }
2122     assert((!UsedOpsInCopy.empty() && !DefedRegsInCopy.empty()) &&
2123            "Unexpect SrcReg or DefReg");
2124     MachineBasicBlock *SuccBB =
2125         getSingleLiveInSuccBB(CurBB, SinkableBBs, DefedRegsInCopy, TRI);
2126     // Don't sink if we cannot find a single sinkable successor in which Reg
2127     // is live-in.
2128     if (!SuccBB) {
2129       LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits,
2130                                         TRI);
2131       continue;
2132     }
2133     assert((SuccBB->pred_size() == 1 && *SuccBB->pred_begin() == &CurBB) &&
2134            "Unexpected predecessor");
2135 
2136     // Collect DBG_VALUEs that must sink with this copy. We've previously
2137     // recorded which reg units that DBG_VALUEs read, if this instruction
2138     // writes any of those units then the corresponding DBG_VALUEs must sink.
2139     MapVector<MachineInstr *, MIRegs::second_type> DbgValsToSinkMap;
2140     for (auto &MO : MI.all_defs()) {
2141       for (MCRegUnit Unit : TRI->regunits(MO.getReg())) {
2142         for (const auto &MIRegs : SeenDbgInstrs.lookup(Unit)) {
2143           auto &Regs = DbgValsToSinkMap[MIRegs.first];
2144           for (unsigned Reg : MIRegs.second)
2145             Regs.push_back(Reg);
2146         }
2147       }
2148     }
2149     auto DbgValsToSink = DbgValsToSinkMap.takeVector();
2150 
2151     LLVM_DEBUG(dbgs() << "Sink instr " << MI << "\tinto block " << *SuccBB);
2152 
2153     MachineBasicBlock::iterator InsertPos =
2154         SuccBB->SkipPHIsAndLabels(SuccBB->begin());
2155     if (blockPrologueInterferes(SuccBB, InsertPos, MI, TRI, TII, nullptr)) {
2156       LLVM_DEBUG(
2157           dbgs() << " *** Not sinking: prologue interference\n");
2158       continue;
2159     }
2160 
2161     // Clear the kill flag if SrcReg is killed between MI and the end of the
2162     // block.
2163     clearKillFlags(&MI, CurBB, UsedOpsInCopy, UsedRegUnits, TRI);
2164     performSink(MI, *SuccBB, InsertPos, DbgValsToSink);
2165     updateLiveIn(&MI, SuccBB, UsedOpsInCopy, DefedRegsInCopy);
2166 
2167     Changed = true;
2168     ++NumPostRACopySink;
2169   }
2170   return Changed;
2171 }
2172 
2173 bool PostRAMachineSinking::runOnMachineFunction(MachineFunction &MF) {
2174   if (skipFunction(MF.getFunction()))
2175     return false;
2176 
2177   bool Changed = false;
2178   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
2179   const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo();
2180 
2181   ModifiedRegUnits.init(*TRI);
2182   UsedRegUnits.init(*TRI);
2183   for (auto &BB : MF)
2184     Changed |= tryToSinkCopy(BB, MF, TRI, TII);
2185 
2186   return Changed;
2187 }
2188