xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/MachineCSE.cpp (revision 031beb4e239bfce798af17f5fe8dba8bcaf13d99)
1 //===- MachineCSE.cpp - Machine Common Subexpression Elimination Pass -----===//
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 performs global common subexpression elimination on machine
10 // instructions using a scoped hash table based value numbering scheme. It
11 // must be run while the machine function is still in SSA form.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/ScopedHashTable.h"
17 #include "llvm/ADT/SmallPtrSet.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/AliasAnalysis.h"
22 #include "llvm/Analysis/CFG.h"
23 #include "llvm/CodeGen/MachineBasicBlock.h"
24 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
25 #include "llvm/CodeGen/MachineDominators.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineFunctionPass.h"
28 #include "llvm/CodeGen/MachineInstr.h"
29 #include "llvm/CodeGen/MachineOperand.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/CodeGen/Passes.h"
32 #include "llvm/CodeGen/TargetInstrInfo.h"
33 #include "llvm/CodeGen/TargetOpcodes.h"
34 #include "llvm/CodeGen/TargetRegisterInfo.h"
35 #include "llvm/CodeGen/TargetSubtargetInfo.h"
36 #include "llvm/InitializePasses.h"
37 #include "llvm/MC/MCRegister.h"
38 #include "llvm/MC/MCRegisterInfo.h"
39 #include "llvm/Pass.h"
40 #include "llvm/Support/Allocator.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/RecyclingAllocator.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include <cassert>
45 #include <iterator>
46 #include <utility>
47 #include <vector>
48 
49 using namespace llvm;
50 
51 #define DEBUG_TYPE "machine-cse"
52 
53 STATISTIC(NumCoalesces, "Number of copies coalesced");
54 STATISTIC(NumCSEs,      "Number of common subexpression eliminated");
55 STATISTIC(NumPREs,      "Number of partial redundant expression"
56                         " transformed to fully redundant");
57 STATISTIC(NumPhysCSEs,
58           "Number of physreg referencing common subexpr eliminated");
59 STATISTIC(NumCrossBBCSEs,
60           "Number of cross-MBB physreg referencing CS eliminated");
61 STATISTIC(NumCommutes,  "Number of copies coalesced after commuting");
62 
63 // Threshold to avoid excessive cost to compute isProfitableToCSE.
64 static cl::opt<int>
65     CSUsesThreshold("csuses-threshold", cl::Hidden, cl::init(1024),
66                     cl::desc("Threshold for the size of CSUses"));
67 
68 namespace {
69 
70   class MachineCSE : public MachineFunctionPass {
71     const TargetInstrInfo *TII;
72     const TargetRegisterInfo *TRI;
73     AliasAnalysis *AA;
74     MachineDominatorTree *DT;
75     MachineRegisterInfo *MRI;
76     MachineBlockFrequencyInfo *MBFI;
77 
78   public:
79     static char ID; // Pass identification
80 
81     MachineCSE() : MachineFunctionPass(ID) {
82       initializeMachineCSEPass(*PassRegistry::getPassRegistry());
83     }
84 
85     bool runOnMachineFunction(MachineFunction &MF) override;
86 
87     void getAnalysisUsage(AnalysisUsage &AU) const override {
88       AU.setPreservesCFG();
89       MachineFunctionPass::getAnalysisUsage(AU);
90       AU.addRequired<AAResultsWrapperPass>();
91       AU.addPreservedID(MachineLoopInfoID);
92       AU.addRequired<MachineDominatorTree>();
93       AU.addPreserved<MachineDominatorTree>();
94       AU.addRequired<MachineBlockFrequencyInfo>();
95       AU.addPreserved<MachineBlockFrequencyInfo>();
96     }
97 
98     MachineFunctionProperties getRequiredProperties() const override {
99       return MachineFunctionProperties()
100         .set(MachineFunctionProperties::Property::IsSSA);
101     }
102 
103     void releaseMemory() override {
104       ScopeMap.clear();
105       PREMap.clear();
106       Exps.clear();
107     }
108 
109   private:
110     using AllocatorTy = RecyclingAllocator<BumpPtrAllocator,
111                             ScopedHashTableVal<MachineInstr *, unsigned>>;
112     using ScopedHTType =
113         ScopedHashTable<MachineInstr *, unsigned, MachineInstrExpressionTrait,
114                         AllocatorTy>;
115     using ScopeType = ScopedHTType::ScopeTy;
116     using PhysDefVector = SmallVector<std::pair<unsigned, unsigned>, 2>;
117 
118     unsigned LookAheadLimit = 0;
119     DenseMap<MachineBasicBlock *, ScopeType *> ScopeMap;
120     DenseMap<MachineInstr *, MachineBasicBlock *, MachineInstrExpressionTrait>
121         PREMap;
122     ScopedHTType VNT;
123     SmallVector<MachineInstr *, 64> Exps;
124     unsigned CurrVN = 0;
125 
126     bool PerformTrivialCopyPropagation(MachineInstr *MI,
127                                        MachineBasicBlock *MBB);
128     bool isPhysDefTriviallyDead(MCRegister Reg,
129                                 MachineBasicBlock::const_iterator I,
130                                 MachineBasicBlock::const_iterator E) const;
131     bool hasLivePhysRegDefUses(const MachineInstr *MI,
132                                const MachineBasicBlock *MBB,
133                                SmallSet<MCRegister, 8> &PhysRefs,
134                                PhysDefVector &PhysDefs, bool &PhysUseDef) const;
135     bool PhysRegDefsReach(MachineInstr *CSMI, MachineInstr *MI,
136                           SmallSet<MCRegister, 8> &PhysRefs,
137                           PhysDefVector &PhysDefs, bool &NonLocal) const;
138     bool isCSECandidate(MachineInstr *MI);
139     bool isProfitableToCSE(Register CSReg, Register Reg,
140                            MachineBasicBlock *CSBB, MachineInstr *MI);
141     void EnterScope(MachineBasicBlock *MBB);
142     void ExitScope(MachineBasicBlock *MBB);
143     bool ProcessBlockCSE(MachineBasicBlock *MBB);
144     void ExitScopeIfDone(MachineDomTreeNode *Node,
145                          DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren);
146     bool PerformCSE(MachineDomTreeNode *Node);
147 
148     bool isPRECandidate(MachineInstr *MI, SmallSet<MCRegister, 8> &PhysRefs);
149     bool ProcessBlockPRE(MachineDominatorTree *MDT, MachineBasicBlock *MBB);
150     bool PerformSimplePRE(MachineDominatorTree *DT);
151     /// Heuristics to see if it's profitable to move common computations of MBB
152     /// and MBB1 to CandidateBB.
153     bool isProfitableToHoistInto(MachineBasicBlock *CandidateBB,
154                                  MachineBasicBlock *MBB,
155                                  MachineBasicBlock *MBB1);
156   };
157 
158 } // end anonymous namespace
159 
160 char MachineCSE::ID = 0;
161 
162 char &llvm::MachineCSEID = MachineCSE::ID;
163 
164 INITIALIZE_PASS_BEGIN(MachineCSE, DEBUG_TYPE,
165                       "Machine Common Subexpression Elimination", false, false)
166 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
167 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
168 INITIALIZE_PASS_END(MachineCSE, DEBUG_TYPE,
169                     "Machine Common Subexpression Elimination", false, false)
170 
171 /// The source register of a COPY machine instruction can be propagated to all
172 /// its users, and this propagation could increase the probability of finding
173 /// common subexpressions. If the COPY has only one user, the COPY itself can
174 /// be removed.
175 bool MachineCSE::PerformTrivialCopyPropagation(MachineInstr *MI,
176                                                MachineBasicBlock *MBB) {
177   bool Changed = false;
178   for (MachineOperand &MO : MI->operands()) {
179     if (!MO.isReg() || !MO.isUse())
180       continue;
181     Register Reg = MO.getReg();
182     if (!Reg.isVirtual())
183       continue;
184     bool OnlyOneUse = MRI->hasOneNonDBGUse(Reg);
185     MachineInstr *DefMI = MRI->getVRegDef(Reg);
186     if (!DefMI->isCopy())
187       continue;
188     Register SrcReg = DefMI->getOperand(1).getReg();
189     if (!SrcReg.isVirtual())
190       continue;
191     if (DefMI->getOperand(0).getSubReg())
192       continue;
193     // FIXME: We should trivially coalesce subregister copies to expose CSE
194     // opportunities on instructions with truncated operands (see
195     // cse-add-with-overflow.ll). This can be done here as follows:
196     // if (SrcSubReg)
197     //  RC = TRI->getMatchingSuperRegClass(MRI->getRegClass(SrcReg), RC,
198     //                                     SrcSubReg);
199     // MO.substVirtReg(SrcReg, SrcSubReg, *TRI);
200     //
201     // The 2-addr pass has been updated to handle coalesced subregs. However,
202     // some machine-specific code still can't handle it.
203     // To handle it properly we also need a way find a constrained subregister
204     // class given a super-reg class and subreg index.
205     if (DefMI->getOperand(1).getSubReg())
206       continue;
207     if (!MRI->constrainRegAttrs(SrcReg, Reg))
208       continue;
209     LLVM_DEBUG(dbgs() << "Coalescing: " << *DefMI);
210     LLVM_DEBUG(dbgs() << "***     to: " << *MI);
211 
212     // Propagate SrcReg of copies to MI.
213     MO.setReg(SrcReg);
214     MRI->clearKillFlags(SrcReg);
215     // Coalesce single use copies.
216     if (OnlyOneUse) {
217       // If (and only if) we've eliminated all uses of the copy, also
218       // copy-propagate to any debug-users of MI, or they'll be left using
219       // an undefined value.
220       DefMI->changeDebugValuesDefReg(SrcReg);
221 
222       DefMI->eraseFromParent();
223       ++NumCoalesces;
224     }
225     Changed = true;
226   }
227 
228   return Changed;
229 }
230 
231 bool MachineCSE::isPhysDefTriviallyDead(
232     MCRegister Reg, MachineBasicBlock::const_iterator I,
233     MachineBasicBlock::const_iterator E) const {
234   unsigned LookAheadLeft = LookAheadLimit;
235   while (LookAheadLeft) {
236     // Skip over dbg_value's.
237     I = skipDebugInstructionsForward(I, E);
238 
239     if (I == E)
240       // Reached end of block, we don't know if register is dead or not.
241       return false;
242 
243     bool SeenDef = false;
244     for (const MachineOperand &MO : I->operands()) {
245       if (MO.isRegMask() && MO.clobbersPhysReg(Reg))
246         SeenDef = true;
247       if (!MO.isReg() || !MO.getReg())
248         continue;
249       if (!TRI->regsOverlap(MO.getReg(), Reg))
250         continue;
251       if (MO.isUse())
252         // Found a use!
253         return false;
254       SeenDef = true;
255     }
256     if (SeenDef)
257       // See a def of Reg (or an alias) before encountering any use, it's
258       // trivially dead.
259       return true;
260 
261     --LookAheadLeft;
262     ++I;
263   }
264   return false;
265 }
266 
267 static bool isCallerPreservedOrConstPhysReg(MCRegister Reg,
268                                             const MachineOperand &MO,
269                                             const MachineFunction &MF,
270                                             const TargetRegisterInfo &TRI,
271                                             const TargetInstrInfo &TII) {
272   // MachineRegisterInfo::isConstantPhysReg directly called by
273   // MachineRegisterInfo::isCallerPreservedOrConstPhysReg expects the
274   // reserved registers to be frozen. That doesn't cause a problem  post-ISel as
275   // most (if not all) targets freeze reserved registers right after ISel.
276   //
277   // It does cause issues mid-GlobalISel, however, hence the additional
278   // reservedRegsFrozen check.
279   const MachineRegisterInfo &MRI = MF.getRegInfo();
280   return TRI.isCallerPreservedPhysReg(Reg, MF) || TII.isIgnorableUse(MO) ||
281          (MRI.reservedRegsFrozen() && MRI.isConstantPhysReg(Reg));
282 }
283 
284 /// hasLivePhysRegDefUses - Return true if the specified instruction read/write
285 /// physical registers (except for dead defs of physical registers). It also
286 /// returns the physical register def by reference if it's the only one and the
287 /// instruction does not uses a physical register.
288 bool MachineCSE::hasLivePhysRegDefUses(const MachineInstr *MI,
289                                        const MachineBasicBlock *MBB,
290                                        SmallSet<MCRegister, 8> &PhysRefs,
291                                        PhysDefVector &PhysDefs,
292                                        bool &PhysUseDef) const {
293   // First, add all uses to PhysRefs.
294   for (const MachineOperand &MO : MI->operands()) {
295     if (!MO.isReg() || MO.isDef())
296       continue;
297     Register Reg = MO.getReg();
298     if (!Reg)
299       continue;
300     if (Reg.isVirtual())
301       continue;
302     // Reading either caller preserved or constant physregs is ok.
303     if (!isCallerPreservedOrConstPhysReg(Reg.asMCReg(), MO, *MI->getMF(), *TRI,
304                                          *TII))
305       for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
306         PhysRefs.insert(*AI);
307   }
308 
309   // Next, collect all defs into PhysDefs.  If any is already in PhysRefs
310   // (which currently contains only uses), set the PhysUseDef flag.
311   PhysUseDef = false;
312   MachineBasicBlock::const_iterator I = MI; I = std::next(I);
313   for (const auto &MOP : llvm::enumerate(MI->operands())) {
314     const MachineOperand &MO = MOP.value();
315     if (!MO.isReg() || !MO.isDef())
316       continue;
317     Register Reg = MO.getReg();
318     if (!Reg)
319       continue;
320     if (Reg.isVirtual())
321       continue;
322     // Check against PhysRefs even if the def is "dead".
323     if (PhysRefs.count(Reg.asMCReg()))
324       PhysUseDef = true;
325     // If the def is dead, it's ok. But the def may not marked "dead". That's
326     // common since this pass is run before livevariables. We can scan
327     // forward a few instructions and check if it is obviously dead.
328     if (!MO.isDead() && !isPhysDefTriviallyDead(Reg.asMCReg(), I, MBB->end()))
329       PhysDefs.push_back(std::make_pair(MOP.index(), Reg));
330   }
331 
332   // Finally, add all defs to PhysRefs as well.
333   for (unsigned i = 0, e = PhysDefs.size(); i != e; ++i)
334     for (MCRegAliasIterator AI(PhysDefs[i].second, TRI, true); AI.isValid();
335          ++AI)
336       PhysRefs.insert(*AI);
337 
338   return !PhysRefs.empty();
339 }
340 
341 bool MachineCSE::PhysRegDefsReach(MachineInstr *CSMI, MachineInstr *MI,
342                                   SmallSet<MCRegister, 8> &PhysRefs,
343                                   PhysDefVector &PhysDefs,
344                                   bool &NonLocal) const {
345   // For now conservatively returns false if the common subexpression is
346   // not in the same basic block as the given instruction. The only exception
347   // is if the common subexpression is in the sole predecessor block.
348   const MachineBasicBlock *MBB = MI->getParent();
349   const MachineBasicBlock *CSMBB = CSMI->getParent();
350 
351   bool CrossMBB = false;
352   if (CSMBB != MBB) {
353     if (MBB->pred_size() != 1 || *MBB->pred_begin() != CSMBB)
354       return false;
355 
356     for (unsigned i = 0, e = PhysDefs.size(); i != e; ++i) {
357       if (MRI->isAllocatable(PhysDefs[i].second) ||
358           MRI->isReserved(PhysDefs[i].second))
359         // Avoid extending live range of physical registers if they are
360         //allocatable or reserved.
361         return false;
362     }
363     CrossMBB = true;
364   }
365   MachineBasicBlock::const_iterator I = CSMI; I = std::next(I);
366   MachineBasicBlock::const_iterator E = MI;
367   MachineBasicBlock::const_iterator EE = CSMBB->end();
368   unsigned LookAheadLeft = LookAheadLimit;
369   while (LookAheadLeft) {
370     // Skip over dbg_value's.
371     while (I != E && I != EE && I->isDebugInstr())
372       ++I;
373 
374     if (I == EE) {
375       assert(CrossMBB && "Reaching end-of-MBB without finding MI?");
376       (void)CrossMBB;
377       CrossMBB = false;
378       NonLocal = true;
379       I = MBB->begin();
380       EE = MBB->end();
381       continue;
382     }
383 
384     if (I == E)
385       return true;
386 
387     for (const MachineOperand &MO : I->operands()) {
388       // RegMasks go on instructions like calls that clobber lots of physregs.
389       // Don't attempt to CSE across such an instruction.
390       if (MO.isRegMask())
391         return false;
392       if (!MO.isReg() || !MO.isDef())
393         continue;
394       Register MOReg = MO.getReg();
395       if (MOReg.isVirtual())
396         continue;
397       if (PhysRefs.count(MOReg.asMCReg()))
398         return false;
399     }
400 
401     --LookAheadLeft;
402     ++I;
403   }
404 
405   return false;
406 }
407 
408 bool MachineCSE::isCSECandidate(MachineInstr *MI) {
409   if (MI->isPosition() || MI->isPHI() || MI->isImplicitDef() || MI->isKill() ||
410       MI->isInlineAsm() || MI->isDebugInstr())
411     return false;
412 
413   // Ignore copies.
414   if (MI->isCopyLike())
415     return false;
416 
417   // Ignore stuff that we obviously can't move.
418   if (MI->mayStore() || MI->isCall() || MI->isTerminator() ||
419       MI->mayRaiseFPException() || MI->hasUnmodeledSideEffects())
420     return false;
421 
422   if (MI->mayLoad()) {
423     // Okay, this instruction does a load. As a refinement, we allow the target
424     // to decide whether the loaded value is actually a constant. If so, we can
425     // actually use it as a load.
426     if (!MI->isDereferenceableInvariantLoad())
427       // FIXME: we should be able to hoist loads with no other side effects if
428       // there are no other instructions which can change memory in this loop.
429       // This is a trivial form of alias analysis.
430       return false;
431   }
432 
433   // Ignore stack guard loads, otherwise the register that holds CSEed value may
434   // be spilled and get loaded back with corrupted data.
435   if (MI->getOpcode() == TargetOpcode::LOAD_STACK_GUARD)
436     return false;
437 
438   return true;
439 }
440 
441 /// isProfitableToCSE - Return true if it's profitable to eliminate MI with a
442 /// common expression that defines Reg. CSBB is basic block where CSReg is
443 /// defined.
444 bool MachineCSE::isProfitableToCSE(Register CSReg, Register Reg,
445                                    MachineBasicBlock *CSBB, MachineInstr *MI) {
446   // FIXME: Heuristics that works around the lack the live range splitting.
447 
448   // If CSReg is used at all uses of Reg, CSE should not increase register
449   // pressure of CSReg.
450   bool MayIncreasePressure = true;
451   if (CSReg.isVirtual() && Reg.isVirtual()) {
452     MayIncreasePressure = false;
453     SmallPtrSet<MachineInstr*, 8> CSUses;
454     int NumOfUses = 0;
455     for (MachineInstr &MI : MRI->use_nodbg_instructions(CSReg)) {
456       CSUses.insert(&MI);
457       // Too costly to compute if NumOfUses is very large. Conservatively assume
458       // MayIncreasePressure to avoid spending too much time here.
459       if (++NumOfUses > CSUsesThreshold) {
460         MayIncreasePressure = true;
461         break;
462       }
463     }
464     if (!MayIncreasePressure)
465       for (MachineInstr &MI : MRI->use_nodbg_instructions(Reg)) {
466         if (!CSUses.count(&MI)) {
467           MayIncreasePressure = true;
468           break;
469         }
470       }
471   }
472   if (!MayIncreasePressure) return true;
473 
474   // Heuristics #1: Don't CSE "cheap" computation if the def is not local or in
475   // an immediate predecessor. We don't want to increase register pressure and
476   // end up causing other computation to be spilled.
477   if (TII->isAsCheapAsAMove(*MI)) {
478     MachineBasicBlock *BB = MI->getParent();
479     if (CSBB != BB && !CSBB->isSuccessor(BB))
480       return false;
481   }
482 
483   // Heuristics #2: If the expression doesn't not use a vr and the only use
484   // of the redundant computation are copies, do not cse.
485   bool HasVRegUse = false;
486   for (const MachineOperand &MO : MI->operands()) {
487     if (MO.isReg() && MO.isUse() && MO.getReg().isVirtual()) {
488       HasVRegUse = true;
489       break;
490     }
491   }
492   if (!HasVRegUse) {
493     bool HasNonCopyUse = false;
494     for (MachineInstr &MI : MRI->use_nodbg_instructions(Reg)) {
495       // Ignore copies.
496       if (!MI.isCopyLike()) {
497         HasNonCopyUse = true;
498         break;
499       }
500     }
501     if (!HasNonCopyUse)
502       return false;
503   }
504 
505   // Heuristics #3: If the common subexpression is used by PHIs, do not reuse
506   // it unless the defined value is already used in the BB of the new use.
507   bool HasPHI = false;
508   for (MachineInstr &UseMI : MRI->use_nodbg_instructions(CSReg)) {
509     HasPHI |= UseMI.isPHI();
510     if (UseMI.getParent() == MI->getParent())
511       return true;
512   }
513 
514   return !HasPHI;
515 }
516 
517 void MachineCSE::EnterScope(MachineBasicBlock *MBB) {
518   LLVM_DEBUG(dbgs() << "Entering: " << MBB->getName() << '\n');
519   ScopeType *Scope = new ScopeType(VNT);
520   ScopeMap[MBB] = Scope;
521 }
522 
523 void MachineCSE::ExitScope(MachineBasicBlock *MBB) {
524   LLVM_DEBUG(dbgs() << "Exiting: " << MBB->getName() << '\n');
525   DenseMap<MachineBasicBlock*, ScopeType*>::iterator SI = ScopeMap.find(MBB);
526   assert(SI != ScopeMap.end());
527   delete SI->second;
528   ScopeMap.erase(SI);
529 }
530 
531 bool MachineCSE::ProcessBlockCSE(MachineBasicBlock *MBB) {
532   bool Changed = false;
533 
534   SmallVector<std::pair<unsigned, unsigned>, 8> CSEPairs;
535   SmallVector<unsigned, 2> ImplicitDefsToUpdate;
536   SmallVector<unsigned, 2> ImplicitDefs;
537   for (MachineInstr &MI : llvm::make_early_inc_range(*MBB)) {
538     if (!isCSECandidate(&MI))
539       continue;
540 
541     bool FoundCSE = VNT.count(&MI);
542     if (!FoundCSE) {
543       // Using trivial copy propagation to find more CSE opportunities.
544       if (PerformTrivialCopyPropagation(&MI, MBB)) {
545         Changed = true;
546 
547         // After coalescing MI itself may become a copy.
548         if (MI.isCopyLike())
549           continue;
550 
551         // Try again to see if CSE is possible.
552         FoundCSE = VNT.count(&MI);
553       }
554     }
555 
556     // Commute commutable instructions.
557     bool Commuted = false;
558     if (!FoundCSE && MI.isCommutable()) {
559       if (MachineInstr *NewMI = TII->commuteInstruction(MI)) {
560         Commuted = true;
561         FoundCSE = VNT.count(NewMI);
562         if (NewMI != &MI) {
563           // New instruction. It doesn't need to be kept.
564           NewMI->eraseFromParent();
565           Changed = true;
566         } else if (!FoundCSE)
567           // MI was changed but it didn't help, commute it back!
568           (void)TII->commuteInstruction(MI);
569       }
570     }
571 
572     // If the instruction defines physical registers and the values *may* be
573     // used, then it's not safe to replace it with a common subexpression.
574     // It's also not safe if the instruction uses physical registers.
575     bool CrossMBBPhysDef = false;
576     SmallSet<MCRegister, 8> PhysRefs;
577     PhysDefVector PhysDefs;
578     bool PhysUseDef = false;
579     if (FoundCSE &&
580         hasLivePhysRegDefUses(&MI, MBB, PhysRefs, PhysDefs, PhysUseDef)) {
581       FoundCSE = false;
582 
583       // ... Unless the CS is local or is in the sole predecessor block
584       // and it also defines the physical register which is not clobbered
585       // in between and the physical register uses were not clobbered.
586       // This can never be the case if the instruction both uses and
587       // defines the same physical register, which was detected above.
588       if (!PhysUseDef) {
589         unsigned CSVN = VNT.lookup(&MI);
590         MachineInstr *CSMI = Exps[CSVN];
591         if (PhysRegDefsReach(CSMI, &MI, PhysRefs, PhysDefs, CrossMBBPhysDef))
592           FoundCSE = true;
593       }
594     }
595 
596     if (!FoundCSE) {
597       VNT.insert(&MI, CurrVN++);
598       Exps.push_back(&MI);
599       continue;
600     }
601 
602     // Found a common subexpression, eliminate it.
603     unsigned CSVN = VNT.lookup(&MI);
604     MachineInstr *CSMI = Exps[CSVN];
605     LLVM_DEBUG(dbgs() << "Examining: " << MI);
606     LLVM_DEBUG(dbgs() << "*** Found a common subexpression: " << *CSMI);
607 
608     // Prevent CSE-ing non-local convergent instructions.
609     // LLVM's current definition of `isConvergent` does not necessarily prove
610     // that non-local CSE is illegal. The following check extends the definition
611     // of `isConvergent` to assume a convergent instruction is dependent not
612     // only on additional conditions, but also on fewer conditions. LLVM does
613     // not have a MachineInstr attribute which expresses this extended
614     // definition, so it's necessary to use `isConvergent` to prevent illegally
615     // CSE-ing the subset of `isConvergent` instructions which do fall into this
616     // extended definition.
617     if (MI.isConvergent() && MI.getParent() != CSMI->getParent()) {
618       LLVM_DEBUG(dbgs() << "*** Convergent MI and subexpression exist in "
619                            "different BBs, avoid CSE!\n");
620       VNT.insert(&MI, CurrVN++);
621       Exps.push_back(&MI);
622       continue;
623     }
624 
625     // Check if it's profitable to perform this CSE.
626     bool DoCSE = true;
627     unsigned NumDefs = MI.getNumDefs();
628 
629     for (unsigned i = 0, e = MI.getNumOperands(); NumDefs && i != e; ++i) {
630       MachineOperand &MO = MI.getOperand(i);
631       if (!MO.isReg() || !MO.isDef())
632         continue;
633       Register OldReg = MO.getReg();
634       Register NewReg = CSMI->getOperand(i).getReg();
635 
636       // Go through implicit defs of CSMI and MI, if a def is not dead at MI,
637       // we should make sure it is not dead at CSMI.
638       if (MO.isImplicit() && !MO.isDead() && CSMI->getOperand(i).isDead())
639         ImplicitDefsToUpdate.push_back(i);
640 
641       // Keep track of implicit defs of CSMI and MI, to clear possibly
642       // made-redundant kill flags.
643       if (MO.isImplicit() && !MO.isDead() && OldReg == NewReg)
644         ImplicitDefs.push_back(OldReg);
645 
646       if (OldReg == NewReg) {
647         --NumDefs;
648         continue;
649       }
650 
651       assert(OldReg.isVirtual() && NewReg.isVirtual() &&
652              "Do not CSE physical register defs!");
653 
654       if (!isProfitableToCSE(NewReg, OldReg, CSMI->getParent(), &MI)) {
655         LLVM_DEBUG(dbgs() << "*** Not profitable, avoid CSE!\n");
656         DoCSE = false;
657         break;
658       }
659 
660       // Don't perform CSE if the result of the new instruction cannot exist
661       // within the constraints (register class, bank, or low-level type) of
662       // the old instruction.
663       if (!MRI->constrainRegAttrs(NewReg, OldReg)) {
664         LLVM_DEBUG(
665             dbgs() << "*** Not the same register constraints, avoid CSE!\n");
666         DoCSE = false;
667         break;
668       }
669 
670       CSEPairs.push_back(std::make_pair(OldReg, NewReg));
671       --NumDefs;
672     }
673 
674     // Actually perform the elimination.
675     if (DoCSE) {
676       for (const std::pair<unsigned, unsigned> &CSEPair : CSEPairs) {
677         unsigned OldReg = CSEPair.first;
678         unsigned NewReg = CSEPair.second;
679         // OldReg may have been unused but is used now, clear the Dead flag
680         MachineInstr *Def = MRI->getUniqueVRegDef(NewReg);
681         assert(Def != nullptr && "CSEd register has no unique definition?");
682         Def->clearRegisterDeads(NewReg);
683         // Replace with NewReg and clear kill flags which may be wrong now.
684         MRI->replaceRegWith(OldReg, NewReg);
685         MRI->clearKillFlags(NewReg);
686       }
687 
688       // Go through implicit defs of CSMI and MI, if a def is not dead at MI,
689       // we should make sure it is not dead at CSMI.
690       for (unsigned ImplicitDefToUpdate : ImplicitDefsToUpdate)
691         CSMI->getOperand(ImplicitDefToUpdate).setIsDead(false);
692       for (const auto &PhysDef : PhysDefs)
693         if (!MI.getOperand(PhysDef.first).isDead())
694           CSMI->getOperand(PhysDef.first).setIsDead(false);
695 
696       // Go through implicit defs of CSMI and MI, and clear the kill flags on
697       // their uses in all the instructions between CSMI and MI.
698       // We might have made some of the kill flags redundant, consider:
699       //   subs  ... implicit-def %nzcv    <- CSMI
700       //   csinc ... implicit killed %nzcv <- this kill flag isn't valid anymore
701       //   subs  ... implicit-def %nzcv    <- MI, to be eliminated
702       //   csinc ... implicit killed %nzcv
703       // Since we eliminated MI, and reused a register imp-def'd by CSMI
704       // (here %nzcv), that register, if it was killed before MI, should have
705       // that kill flag removed, because it's lifetime was extended.
706       if (CSMI->getParent() == MI.getParent()) {
707         for (MachineBasicBlock::iterator II = CSMI, IE = &MI; II != IE; ++II)
708           for (auto ImplicitDef : ImplicitDefs)
709             if (MachineOperand *MO = II->findRegisterUseOperand(
710                     ImplicitDef, /*isKill=*/true, TRI))
711               MO->setIsKill(false);
712       } else {
713         // If the instructions aren't in the same BB, bail out and clear the
714         // kill flag on all uses of the imp-def'd register.
715         for (auto ImplicitDef : ImplicitDefs)
716           MRI->clearKillFlags(ImplicitDef);
717       }
718 
719       if (CrossMBBPhysDef) {
720         // Add physical register defs now coming in from a predecessor to MBB
721         // livein list.
722         while (!PhysDefs.empty()) {
723           auto LiveIn = PhysDefs.pop_back_val();
724           if (!MBB->isLiveIn(LiveIn.second))
725             MBB->addLiveIn(LiveIn.second);
726         }
727         ++NumCrossBBCSEs;
728       }
729 
730       MI.eraseFromParent();
731       ++NumCSEs;
732       if (!PhysRefs.empty())
733         ++NumPhysCSEs;
734       if (Commuted)
735         ++NumCommutes;
736       Changed = true;
737     } else {
738       VNT.insert(&MI, CurrVN++);
739       Exps.push_back(&MI);
740     }
741     CSEPairs.clear();
742     ImplicitDefsToUpdate.clear();
743     ImplicitDefs.clear();
744   }
745 
746   return Changed;
747 }
748 
749 /// ExitScopeIfDone - Destroy scope for the MBB that corresponds to the given
750 /// dominator tree node if its a leaf or all of its children are done. Walk
751 /// up the dominator tree to destroy ancestors which are now done.
752 void
753 MachineCSE::ExitScopeIfDone(MachineDomTreeNode *Node,
754                         DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren) {
755   if (OpenChildren[Node])
756     return;
757 
758   // Pop scope.
759   ExitScope(Node->getBlock());
760 
761   // Now traverse upwards to pop ancestors whose offsprings are all done.
762   while (MachineDomTreeNode *Parent = Node->getIDom()) {
763     unsigned Left = --OpenChildren[Parent];
764     if (Left != 0)
765       break;
766     ExitScope(Parent->getBlock());
767     Node = Parent;
768   }
769 }
770 
771 bool MachineCSE::PerformCSE(MachineDomTreeNode *Node) {
772   SmallVector<MachineDomTreeNode*, 32> Scopes;
773   SmallVector<MachineDomTreeNode*, 8> WorkList;
774   DenseMap<MachineDomTreeNode*, unsigned> OpenChildren;
775 
776   CurrVN = 0;
777 
778   // Perform a DFS walk to determine the order of visit.
779   WorkList.push_back(Node);
780   do {
781     Node = WorkList.pop_back_val();
782     Scopes.push_back(Node);
783     OpenChildren[Node] = Node->getNumChildren();
784     append_range(WorkList, Node->children());
785   } while (!WorkList.empty());
786 
787   // Now perform CSE.
788   bool Changed = false;
789   for (MachineDomTreeNode *Node : Scopes) {
790     MachineBasicBlock *MBB = Node->getBlock();
791     EnterScope(MBB);
792     Changed |= ProcessBlockCSE(MBB);
793     // If it's a leaf node, it's done. Traverse upwards to pop ancestors.
794     ExitScopeIfDone(Node, OpenChildren);
795   }
796 
797   return Changed;
798 }
799 
800 // We use stronger checks for PRE candidate rather than for CSE ones to embrace
801 // checks inside ProcessBlockCSE(), not only inside isCSECandidate(). This helps
802 // to exclude instrs created by PRE that won't be CSEed later.
803 bool MachineCSE::isPRECandidate(MachineInstr *MI,
804                                 SmallSet<MCRegister, 8> &PhysRefs) {
805   if (!isCSECandidate(MI) ||
806       MI->isNotDuplicable() ||
807       MI->mayLoad() ||
808       TII->isAsCheapAsAMove(*MI) ||
809       MI->getNumDefs() != 1 ||
810       MI->getNumExplicitDefs() != 1)
811     return false;
812 
813   for (const MachineOperand &MO : MI->operands()) {
814     if (MO.isReg() && !MO.getReg().isVirtual()) {
815       if (MO.isDef())
816         return false;
817       else
818         PhysRefs.insert(MO.getReg());
819     }
820   }
821 
822   return true;
823 }
824 
825 bool MachineCSE::ProcessBlockPRE(MachineDominatorTree *DT,
826                                  MachineBasicBlock *MBB) {
827   bool Changed = false;
828   for (MachineInstr &MI : llvm::make_early_inc_range(*MBB)) {
829     SmallSet<MCRegister, 8> PhysRefs;
830     if (!isPRECandidate(&MI, PhysRefs))
831       continue;
832 
833     if (!PREMap.count(&MI)) {
834       PREMap[&MI] = MBB;
835       continue;
836     }
837 
838     auto MBB1 = PREMap[&MI];
839     assert(
840         !DT->properlyDominates(MBB, MBB1) &&
841         "MBB cannot properly dominate MBB1 while DFS through dominators tree!");
842     auto CMBB = DT->findNearestCommonDominator(MBB, MBB1);
843     if (!CMBB->isLegalToHoistInto())
844       continue;
845 
846     if (!isProfitableToHoistInto(CMBB, MBB, MBB1))
847       continue;
848 
849     // Two instrs are partial redundant if their basic blocks are reachable
850     // from one to another but one doesn't dominate another.
851     if (CMBB != MBB1) {
852       auto BB = MBB->getBasicBlock(), BB1 = MBB1->getBasicBlock();
853       if (BB != nullptr && BB1 != nullptr &&
854           (isPotentiallyReachable(BB1, BB) ||
855            isPotentiallyReachable(BB, BB1))) {
856         // The following check extends the definition of `isConvergent` to
857         // assume a convergent instruction is dependent not only on additional
858         // conditions, but also on fewer conditions. LLVM does not have a
859         // MachineInstr attribute which expresses this extended definition, so
860         // it's necessary to use `isConvergent` to prevent illegally PRE-ing the
861         // subset of `isConvergent` instructions which do fall into this
862         // extended definition.
863         if (MI.isConvergent() && CMBB != MBB)
864           continue;
865 
866         // If this instruction uses physical registers then we can only do PRE
867         // if it's using the value that is live at the place we're hoisting to.
868         bool NonLocal;
869         PhysDefVector PhysDefs;
870         if (!PhysRefs.empty() &&
871             !PhysRegDefsReach(&*(CMBB->getFirstTerminator()), &MI, PhysRefs,
872                               PhysDefs, NonLocal))
873           continue;
874 
875         assert(MI.getOperand(0).isDef() &&
876                "First operand of instr with one explicit def must be this def");
877         Register VReg = MI.getOperand(0).getReg();
878         Register NewReg = MRI->cloneVirtualRegister(VReg);
879         if (!isProfitableToCSE(NewReg, VReg, CMBB, &MI))
880           continue;
881         MachineInstr &NewMI =
882             TII->duplicate(*CMBB, CMBB->getFirstTerminator(), MI);
883 
884         // When hoisting, make sure we don't carry the debug location of
885         // the original instruction, as that's not correct and can cause
886         // unexpected jumps when debugging optimized code.
887         auto EmptyDL = DebugLoc();
888         NewMI.setDebugLoc(EmptyDL);
889 
890         NewMI.getOperand(0).setReg(NewReg);
891 
892         PREMap[&MI] = CMBB;
893         ++NumPREs;
894         Changed = true;
895       }
896     }
897   }
898   return Changed;
899 }
900 
901 // This simple PRE (partial redundancy elimination) pass doesn't actually
902 // eliminate partial redundancy but transforms it to full redundancy,
903 // anticipating that the next CSE step will eliminate this created redundancy.
904 // If CSE doesn't eliminate this, than created instruction will remain dead
905 // and eliminated later by Remove Dead Machine Instructions pass.
906 bool MachineCSE::PerformSimplePRE(MachineDominatorTree *DT) {
907   SmallVector<MachineDomTreeNode *, 32> BBs;
908 
909   PREMap.clear();
910   bool Changed = false;
911   BBs.push_back(DT->getRootNode());
912   do {
913     auto Node = BBs.pop_back_val();
914     append_range(BBs, Node->children());
915 
916     MachineBasicBlock *MBB = Node->getBlock();
917     Changed |= ProcessBlockPRE(DT, MBB);
918 
919   } while (!BBs.empty());
920 
921   return Changed;
922 }
923 
924 bool MachineCSE::isProfitableToHoistInto(MachineBasicBlock *CandidateBB,
925                                          MachineBasicBlock *MBB,
926                                          MachineBasicBlock *MBB1) {
927   if (CandidateBB->getParent()->getFunction().hasMinSize())
928     return true;
929   assert(DT->dominates(CandidateBB, MBB) && "CandidateBB should dominate MBB");
930   assert(DT->dominates(CandidateBB, MBB1) &&
931          "CandidateBB should dominate MBB1");
932   return MBFI->getBlockFreq(CandidateBB) <=
933          MBFI->getBlockFreq(MBB) + MBFI->getBlockFreq(MBB1);
934 }
935 
936 bool MachineCSE::runOnMachineFunction(MachineFunction &MF) {
937   if (skipFunction(MF.getFunction()))
938     return false;
939 
940   TII = MF.getSubtarget().getInstrInfo();
941   TRI = MF.getSubtarget().getRegisterInfo();
942   MRI = &MF.getRegInfo();
943   AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
944   DT = &getAnalysis<MachineDominatorTree>();
945   MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
946   LookAheadLimit = TII->getMachineCSELookAheadLimit();
947   bool ChangedPRE, ChangedCSE;
948   ChangedPRE = PerformSimplePRE(DT);
949   ChangedCSE = PerformCSE(DT->getRootNode());
950   return ChangedPRE || ChangedCSE;
951 }
952