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