xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/MachineCSE.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
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 
48 using namespace llvm;
49 
50 #define DEBUG_TYPE "machine-cse"
51 
52 STATISTIC(NumCoalesces, "Number of copies coalesced");
53 STATISTIC(NumCSEs,      "Number of common subexpression eliminated");
54 STATISTIC(NumPREs,      "Number of partial redundant expression"
55                         " transformed to fully redundant");
56 STATISTIC(NumPhysCSEs,
57           "Number of physreg referencing common subexpr eliminated");
58 STATISTIC(NumCrossBBCSEs,
59           "Number of cross-MBB physreg referencing CS eliminated");
60 STATISTIC(NumCommutes,  "Number of copies coalesced after commuting");
61 
62 // Threshold to avoid excessive cost to compute isProfitableToCSE.
63 static cl::opt<int>
64     CSUsesThreshold("csuses-threshold", cl::Hidden, cl::init(1024),
65                     cl::desc("Threshold for the size of CSUses"));
66 
67 static cl::opt<bool> AggressiveMachineCSE(
68     "aggressive-machine-cse", cl::Hidden, cl::init(false),
69     cl::desc("Override the profitability heuristics for Machine CSE"));
70 
71 namespace {
72 
73   class MachineCSE : public MachineFunctionPass {
74     const TargetInstrInfo *TII = nullptr;
75     const TargetRegisterInfo *TRI = nullptr;
76     AliasAnalysis *AA = nullptr;
77     MachineDominatorTree *DT = nullptr;
78     MachineRegisterInfo *MRI = nullptr;
79     MachineBlockFrequencyInfo *MBFI = nullptr;
80 
81   public:
82     static char ID; // Pass identification
83 
MachineCSE()84     MachineCSE() : MachineFunctionPass(ID) {
85       initializeMachineCSEPass(*PassRegistry::getPassRegistry());
86     }
87 
88     bool runOnMachineFunction(MachineFunction &MF) override;
89 
getAnalysisUsage(AnalysisUsage & AU) const90     void getAnalysisUsage(AnalysisUsage &AU) const override {
91       AU.setPreservesCFG();
92       MachineFunctionPass::getAnalysisUsage(AU);
93       AU.addRequired<AAResultsWrapperPass>();
94       AU.addPreservedID(MachineLoopInfoID);
95       AU.addRequired<MachineDominatorTreeWrapperPass>();
96       AU.addPreserved<MachineDominatorTreeWrapperPass>();
97       AU.addRequired<MachineBlockFrequencyInfoWrapperPass>();
98       AU.addPreserved<MachineBlockFrequencyInfoWrapperPass>();
99     }
100 
getRequiredProperties() const101     MachineFunctionProperties getRequiredProperties() const override {
102       return MachineFunctionProperties()
103         .set(MachineFunctionProperties::Property::IsSSA);
104     }
105 
releaseMemory()106     void releaseMemory() override {
107       ScopeMap.clear();
108       PREMap.clear();
109       Exps.clear();
110     }
111 
112   private:
113     using AllocatorTy = RecyclingAllocator<BumpPtrAllocator,
114                             ScopedHashTableVal<MachineInstr *, unsigned>>;
115     using ScopedHTType =
116         ScopedHashTable<MachineInstr *, unsigned, MachineInstrExpressionTrait,
117                         AllocatorTy>;
118     using ScopeType = ScopedHTType::ScopeTy;
119     using PhysDefVector = SmallVector<std::pair<unsigned, unsigned>, 2>;
120 
121     unsigned LookAheadLimit = 0;
122     DenseMap<MachineBasicBlock *, ScopeType *> ScopeMap;
123     DenseMap<MachineInstr *, MachineBasicBlock *, MachineInstrExpressionTrait>
124         PREMap;
125     ScopedHTType VNT;
126     SmallVector<MachineInstr *, 64> Exps;
127     unsigned CurrVN = 0;
128 
129     bool PerformTrivialCopyPropagation(MachineInstr *MI,
130                                        MachineBasicBlock *MBB);
131     bool isPhysDefTriviallyDead(MCRegister Reg,
132                                 MachineBasicBlock::const_iterator I,
133                                 MachineBasicBlock::const_iterator E) const;
134     bool hasLivePhysRegDefUses(const MachineInstr *MI,
135                                const MachineBasicBlock *MBB,
136                                SmallSet<MCRegister, 8> &PhysRefs,
137                                PhysDefVector &PhysDefs, bool &PhysUseDef) const;
138     bool PhysRegDefsReach(MachineInstr *CSMI, MachineInstr *MI,
139                           SmallSet<MCRegister, 8> &PhysRefs,
140                           PhysDefVector &PhysDefs, bool &NonLocal) const;
141     bool isCSECandidate(MachineInstr *MI);
142     bool isProfitableToCSE(Register CSReg, Register Reg,
143                            MachineBasicBlock *CSBB, MachineInstr *MI);
144     void EnterScope(MachineBasicBlock *MBB);
145     void ExitScope(MachineBasicBlock *MBB);
146     bool ProcessBlockCSE(MachineBasicBlock *MBB);
147     void ExitScopeIfDone(MachineDomTreeNode *Node,
148                          DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren);
149     bool PerformCSE(MachineDomTreeNode *Node);
150 
151     bool isPRECandidate(MachineInstr *MI, SmallSet<MCRegister, 8> &PhysRefs);
152     bool ProcessBlockPRE(MachineDominatorTree *MDT, MachineBasicBlock *MBB);
153     bool PerformSimplePRE(MachineDominatorTree *DT);
154     /// Heuristics to see if it's profitable to move common computations of MBB
155     /// and MBB1 to CandidateBB.
156     bool isProfitableToHoistInto(MachineBasicBlock *CandidateBB,
157                                  MachineBasicBlock *MBB,
158                                  MachineBasicBlock *MBB1);
159   };
160 
161 } // end anonymous namespace
162 
163 char MachineCSE::ID = 0;
164 
165 char &llvm::MachineCSEID = MachineCSE::ID;
166 
167 INITIALIZE_PASS_BEGIN(MachineCSE, DEBUG_TYPE,
168                       "Machine Common Subexpression Elimination", false, false)
INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass)169 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass)
170 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
171 INITIALIZE_PASS_END(MachineCSE, DEBUG_TYPE,
172                     "Machine Common Subexpression Elimination", false, false)
173 
174 /// The source register of a COPY machine instruction can be propagated to all
175 /// its users, and this propagation could increase the probability of finding
176 /// common subexpressions. If the COPY has only one user, the COPY itself can
177 /// be removed.
178 bool MachineCSE::PerformTrivialCopyPropagation(MachineInstr *MI,
179                                                MachineBasicBlock *MBB) {
180   bool Changed = false;
181   for (MachineOperand &MO : MI->all_uses()) {
182     Register Reg = MO.getReg();
183     if (!Reg.isVirtual())
184       continue;
185     bool OnlyOneUse = MRI->hasOneNonDBGUse(Reg);
186     MachineInstr *DefMI = MRI->getVRegDef(Reg);
187     if (!DefMI || !DefMI->isCopy())
188       continue;
189     Register SrcReg = DefMI->getOperand(1).getReg();
190     if (!SrcReg.isVirtual())
191       continue;
192     if (DefMI->getOperand(0).getSubReg())
193       continue;
194     // FIXME: We should trivially coalesce subregister copies to expose CSE
195     // opportunities on instructions with truncated operands (see
196     // cse-add-with-overflow.ll). This can be done here as follows:
197     // if (SrcSubReg)
198     //  RC = TRI->getMatchingSuperRegClass(MRI->getRegClass(SrcReg), RC,
199     //                                     SrcSubReg);
200     // MO.substVirtReg(SrcReg, SrcSubReg, *TRI);
201     //
202     // The 2-addr pass has been updated to handle coalesced subregs. However,
203     // some machine-specific code still can't handle it.
204     // To handle it properly we also need a way find a constrained subregister
205     // class given a super-reg class and subreg index.
206     if (DefMI->getOperand(1).getSubReg())
207       continue;
208     if (!MRI->constrainRegAttrs(SrcReg, Reg))
209       continue;
210     LLVM_DEBUG(dbgs() << "Coalescing: " << *DefMI);
211     LLVM_DEBUG(dbgs() << "***     to: " << *MI);
212 
213     // Propagate SrcReg of copies to MI.
214     MO.setReg(SrcReg);
215     MRI->clearKillFlags(SrcReg);
216     // Coalesce single use copies.
217     if (OnlyOneUse) {
218       // If (and only if) we've eliminated all uses of the copy, also
219       // copy-propagate to any debug-users of MI, or they'll be left using
220       // an undefined value.
221       DefMI->changeDebugValuesDefReg(SrcReg);
222 
223       DefMI->eraseFromParent();
224       ++NumCoalesces;
225     }
226     Changed = true;
227   }
228 
229   return Changed;
230 }
231 
isPhysDefTriviallyDead(MCRegister Reg,MachineBasicBlock::const_iterator I,MachineBasicBlock::const_iterator E) const232 bool MachineCSE::isPhysDefTriviallyDead(
233     MCRegister Reg, MachineBasicBlock::const_iterator I,
234     MachineBasicBlock::const_iterator E) const {
235   unsigned LookAheadLeft = LookAheadLimit;
236   while (LookAheadLeft) {
237     // Skip over dbg_value's.
238     I = skipDebugInstructionsForward(I, E);
239 
240     if (I == E)
241       // Reached end of block, we don't know if register is dead or not.
242       return false;
243 
244     bool SeenDef = false;
245     for (const MachineOperand &MO : I->operands()) {
246       if (MO.isRegMask() && MO.clobbersPhysReg(Reg))
247         SeenDef = true;
248       if (!MO.isReg() || !MO.getReg())
249         continue;
250       if (!TRI->regsOverlap(MO.getReg(), Reg))
251         continue;
252       if (MO.isUse())
253         // Found a use!
254         return false;
255       SeenDef = true;
256     }
257     if (SeenDef)
258       // See a def of Reg (or an alias) before encountering any use, it's
259       // trivially dead.
260       return true;
261 
262     --LookAheadLeft;
263     ++I;
264   }
265   return false;
266 }
267 
isCallerPreservedOrConstPhysReg(MCRegister Reg,const MachineOperand & MO,const MachineFunction & MF,const TargetRegisterInfo & TRI,const TargetInstrInfo & TII)268 static bool isCallerPreservedOrConstPhysReg(MCRegister Reg,
269                                             const MachineOperand &MO,
270                                             const MachineFunction &MF,
271                                             const TargetRegisterInfo &TRI,
272                                             const TargetInstrInfo &TII) {
273   // MachineRegisterInfo::isConstantPhysReg directly called by
274   // MachineRegisterInfo::isCallerPreservedOrConstPhysReg expects the
275   // reserved registers to be frozen. That doesn't cause a problem  post-ISel as
276   // most (if not all) targets freeze reserved registers right after ISel.
277   //
278   // It does cause issues mid-GlobalISel, however, hence the additional
279   // reservedRegsFrozen check.
280   const MachineRegisterInfo &MRI = MF.getRegInfo();
281   return TRI.isCallerPreservedPhysReg(Reg, MF) || TII.isIgnorableUse(MO) ||
282          (MRI.reservedRegsFrozen() && MRI.isConstantPhysReg(Reg));
283 }
284 
285 /// hasLivePhysRegDefUses - Return true if the specified instruction read/write
286 /// physical registers (except for dead defs of physical registers). It also
287 /// returns the physical register def by reference if it's the only one and the
288 /// instruction does not uses a physical register.
hasLivePhysRegDefUses(const MachineInstr * MI,const MachineBasicBlock * MBB,SmallSet<MCRegister,8> & PhysRefs,PhysDefVector & PhysDefs,bool & PhysUseDef) const289 bool MachineCSE::hasLivePhysRegDefUses(const MachineInstr *MI,
290                                        const MachineBasicBlock *MBB,
291                                        SmallSet<MCRegister, 8> &PhysRefs,
292                                        PhysDefVector &PhysDefs,
293                                        bool &PhysUseDef) const {
294   // First, add all uses to PhysRefs.
295   for (const MachineOperand &MO : MI->all_uses()) {
296     Register Reg = MO.getReg();
297     if (!Reg)
298       continue;
299     if (Reg.isVirtual())
300       continue;
301     // Reading either caller preserved or constant physregs is ok.
302     if (!isCallerPreservedOrConstPhysReg(Reg.asMCReg(), MO, *MI->getMF(), *TRI,
303                                          *TII))
304       for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
305         PhysRefs.insert(*AI);
306   }
307 
308   // Next, collect all defs into PhysDefs.  If any is already in PhysRefs
309   // (which currently contains only uses), set the PhysUseDef flag.
310   PhysUseDef = false;
311   MachineBasicBlock::const_iterator I = MI; I = std::next(I);
312   for (const auto &MOP : llvm::enumerate(MI->operands())) {
313     const MachineOperand &MO = MOP.value();
314     if (!MO.isReg() || !MO.isDef())
315       continue;
316     Register Reg = MO.getReg();
317     if (!Reg)
318       continue;
319     if (Reg.isVirtual())
320       continue;
321     // Check against PhysRefs even if the def is "dead".
322     if (PhysRefs.count(Reg.asMCReg()))
323       PhysUseDef = true;
324     // If the def is dead, it's ok. But the def may not marked "dead". That's
325     // common since this pass is run before livevariables. We can scan
326     // forward a few instructions and check if it is obviously dead.
327     if (!MO.isDead() && !isPhysDefTriviallyDead(Reg.asMCReg(), I, MBB->end()))
328       PhysDefs.push_back(std::make_pair(MOP.index(), Reg));
329   }
330 
331   // Finally, add all defs to PhysRefs as well.
332   for (unsigned i = 0, e = PhysDefs.size(); i != e; ++i)
333     for (MCRegAliasIterator AI(PhysDefs[i].second, TRI, true); AI.isValid();
334          ++AI)
335       PhysRefs.insert(*AI);
336 
337   return !PhysRefs.empty();
338 }
339 
PhysRegDefsReach(MachineInstr * CSMI,MachineInstr * MI,SmallSet<MCRegister,8> & PhysRefs,PhysDefVector & PhysDefs,bool & NonLocal) const340 bool MachineCSE::PhysRegDefsReach(MachineInstr *CSMI, MachineInstr *MI,
341                                   SmallSet<MCRegister, 8> &PhysRefs,
342                                   PhysDefVector &PhysDefs,
343                                   bool &NonLocal) const {
344   // For now conservatively returns false if the common subexpression is
345   // not in the same basic block as the given instruction. The only exception
346   // is if the common subexpression is in the sole predecessor block.
347   const MachineBasicBlock *MBB = MI->getParent();
348   const MachineBasicBlock *CSMBB = CSMI->getParent();
349 
350   bool CrossMBB = false;
351   if (CSMBB != MBB) {
352     if (MBB->pred_size() != 1 || *MBB->pred_begin() != CSMBB)
353       return false;
354 
355     for (unsigned i = 0, e = PhysDefs.size(); i != e; ++i) {
356       if (MRI->isAllocatable(PhysDefs[i].second) ||
357           MRI->isReserved(PhysDefs[i].second))
358         // Avoid extending live range of physical registers if they are
359         //allocatable or reserved.
360         return false;
361     }
362     CrossMBB = true;
363   }
364   MachineBasicBlock::const_iterator I = CSMI; I = std::next(I);
365   MachineBasicBlock::const_iterator E = MI;
366   MachineBasicBlock::const_iterator EE = CSMBB->end();
367   unsigned LookAheadLeft = LookAheadLimit;
368   while (LookAheadLeft) {
369     // Skip over dbg_value's.
370     while (I != E && I != EE && I->isDebugInstr())
371       ++I;
372 
373     if (I == EE) {
374       assert(CrossMBB && "Reaching end-of-MBB without finding MI?");
375       (void)CrossMBB;
376       CrossMBB = false;
377       NonLocal = true;
378       I = MBB->begin();
379       EE = MBB->end();
380       continue;
381     }
382 
383     if (I == E)
384       return true;
385 
386     for (const MachineOperand &MO : I->operands()) {
387       // RegMasks go on instructions like calls that clobber lots of physregs.
388       // Don't attempt to CSE across such an instruction.
389       if (MO.isRegMask())
390         return false;
391       if (!MO.isReg() || !MO.isDef())
392         continue;
393       Register MOReg = MO.getReg();
394       if (MOReg.isVirtual())
395         continue;
396       if (PhysRefs.count(MOReg.asMCReg()))
397         return false;
398     }
399 
400     --LookAheadLeft;
401     ++I;
402   }
403 
404   return false;
405 }
406 
isCSECandidate(MachineInstr * MI)407 bool MachineCSE::isCSECandidate(MachineInstr *MI) {
408   if (MI->isPosition() || MI->isPHI() || MI->isImplicitDef() || MI->isKill() ||
409       MI->isInlineAsm() || MI->isDebugInstr() || MI->isJumpTableDebugInfo())
410     return false;
411 
412   // Ignore copies.
413   if (MI->isCopyLike())
414     return false;
415 
416   // Ignore stuff that we obviously can't move.
417   if (MI->mayStore() || MI->isCall() || MI->isTerminator() ||
418       MI->mayRaiseFPException() || MI->hasUnmodeledSideEffects())
419     return false;
420 
421   if (MI->mayLoad()) {
422     // Okay, this instruction does a load. As a refinement, we allow the target
423     // to decide whether the loaded value is actually a constant. If so, we can
424     // actually use it as a load.
425     if (!MI->isDereferenceableInvariantLoad())
426       // FIXME: we should be able to hoist loads with no other side effects if
427       // there are no other instructions which can change memory in this loop.
428       // This is a trivial form of alias analysis.
429       return false;
430   }
431 
432   // Ignore stack guard loads, otherwise the register that holds CSEed value may
433   // be spilled and get loaded back with corrupted data.
434   if (MI->getOpcode() == TargetOpcode::LOAD_STACK_GUARD)
435     return false;
436 
437   return true;
438 }
439 
440 /// isProfitableToCSE - Return true if it's profitable to eliminate MI with a
441 /// common expression that defines Reg. CSBB is basic block where CSReg is
442 /// defined.
isProfitableToCSE(Register CSReg,Register Reg,MachineBasicBlock * CSBB,MachineInstr * MI)443 bool MachineCSE::isProfitableToCSE(Register CSReg, Register Reg,
444                                    MachineBasicBlock *CSBB, MachineInstr *MI) {
445   if (AggressiveMachineCSE)
446     return true;
447 
448   // FIXME: Heuristics that works around the lack the live range splitting.
449 
450   // If CSReg is used at all uses of Reg, CSE should not increase register
451   // pressure of CSReg.
452   bool MayIncreasePressure = true;
453   if (CSReg.isVirtual() && Reg.isVirtual()) {
454     MayIncreasePressure = false;
455     SmallPtrSet<MachineInstr*, 8> CSUses;
456     int NumOfUses = 0;
457     for (MachineInstr &MI : MRI->use_nodbg_instructions(CSReg)) {
458       CSUses.insert(&MI);
459       // Too costly to compute if NumOfUses is very large. Conservatively assume
460       // MayIncreasePressure to avoid spending too much time here.
461       if (++NumOfUses > CSUsesThreshold) {
462         MayIncreasePressure = true;
463         break;
464       }
465     }
466     if (!MayIncreasePressure)
467       for (MachineInstr &MI : MRI->use_nodbg_instructions(Reg)) {
468         if (!CSUses.count(&MI)) {
469           MayIncreasePressure = true;
470           break;
471         }
472       }
473   }
474   if (!MayIncreasePressure) return true;
475 
476   // Heuristics #1: Don't CSE "cheap" computation if the def is not local or in
477   // an immediate predecessor. We don't want to increase register pressure and
478   // end up causing other computation to be spilled.
479   if (TII->isAsCheapAsAMove(*MI)) {
480     MachineBasicBlock *BB = MI->getParent();
481     if (CSBB != BB && !CSBB->isSuccessor(BB))
482       return false;
483   }
484 
485   // Heuristics #2: If the expression doesn't not use a vr and the only use
486   // of the redundant computation are copies, do not cse.
487   bool HasVRegUse = false;
488   for (const MachineOperand &MO : MI->all_uses()) {
489     if (MO.getReg().isVirtual()) {
490       HasVRegUse = true;
491       break;
492     }
493   }
494   if (!HasVRegUse) {
495     bool HasNonCopyUse = false;
496     for (MachineInstr &MI : MRI->use_nodbg_instructions(Reg)) {
497       // Ignore copies.
498       if (!MI.isCopyLike()) {
499         HasNonCopyUse = true;
500         break;
501       }
502     }
503     if (!HasNonCopyUse)
504       return false;
505   }
506 
507   // Heuristics #3: If the common subexpression is used by PHIs, do not reuse
508   // it unless the defined value is already used in the BB of the new use.
509   bool HasPHI = false;
510   for (MachineInstr &UseMI : MRI->use_nodbg_instructions(CSReg)) {
511     HasPHI |= UseMI.isPHI();
512     if (UseMI.getParent() == MI->getParent())
513       return true;
514   }
515 
516   return !HasPHI;
517 }
518 
EnterScope(MachineBasicBlock * MBB)519 void MachineCSE::EnterScope(MachineBasicBlock *MBB) {
520   LLVM_DEBUG(dbgs() << "Entering: " << MBB->getName() << '\n');
521   ScopeType *Scope = new ScopeType(VNT);
522   ScopeMap[MBB] = Scope;
523 }
524 
ExitScope(MachineBasicBlock * MBB)525 void MachineCSE::ExitScope(MachineBasicBlock *MBB) {
526   LLVM_DEBUG(dbgs() << "Exiting: " << MBB->getName() << '\n');
527   DenseMap<MachineBasicBlock*, ScopeType*>::iterator SI = ScopeMap.find(MBB);
528   assert(SI != ScopeMap.end());
529   delete SI->second;
530   ScopeMap.erase(SI);
531 }
532 
ProcessBlockCSE(MachineBasicBlock * MBB)533 bool MachineCSE::ProcessBlockCSE(MachineBasicBlock *MBB) {
534   bool Changed = false;
535 
536   SmallVector<std::pair<unsigned, unsigned>, 8> CSEPairs;
537   SmallVector<unsigned, 2> ImplicitDefsToUpdate;
538   SmallVector<unsigned, 2> ImplicitDefs;
539   for (MachineInstr &MI : llvm::make_early_inc_range(*MBB)) {
540     if (!isCSECandidate(&MI))
541       continue;
542 
543     bool FoundCSE = VNT.count(&MI);
544     if (!FoundCSE) {
545       // Using trivial copy propagation to find more CSE opportunities.
546       if (PerformTrivialCopyPropagation(&MI, MBB)) {
547         Changed = true;
548 
549         // After coalescing MI itself may become a copy.
550         if (MI.isCopyLike())
551           continue;
552 
553         // Try again to see if CSE is possible.
554         FoundCSE = VNT.count(&MI);
555       }
556     }
557 
558     // Commute commutable instructions.
559     bool Commuted = false;
560     if (!FoundCSE && MI.isCommutable()) {
561       if (MachineInstr *NewMI = TII->commuteInstruction(MI)) {
562         Commuted = true;
563         FoundCSE = VNT.count(NewMI);
564         if (NewMI != &MI) {
565           // New instruction. It doesn't need to be kept.
566           NewMI->eraseFromParent();
567           Changed = true;
568         } else if (!FoundCSE)
569           // MI was changed but it didn't help, commute it back!
570           (void)TII->commuteInstruction(MI);
571       }
572     }
573 
574     // If the instruction defines physical registers and the values *may* be
575     // used, then it's not safe to replace it with a common subexpression.
576     // It's also not safe if the instruction uses physical registers.
577     bool CrossMBBPhysDef = false;
578     SmallSet<MCRegister, 8> PhysRefs;
579     PhysDefVector PhysDefs;
580     bool PhysUseDef = false;
581     if (FoundCSE &&
582         hasLivePhysRegDefUses(&MI, MBB, PhysRefs, PhysDefs, PhysUseDef)) {
583       FoundCSE = false;
584 
585       // ... Unless the CS is local or is in the sole predecessor block
586       // and it also defines the physical register which is not clobbered
587       // in between and the physical register uses were not clobbered.
588       // This can never be the case if the instruction both uses and
589       // defines the same physical register, which was detected above.
590       if (!PhysUseDef) {
591         unsigned CSVN = VNT.lookup(&MI);
592         MachineInstr *CSMI = Exps[CSVN];
593         if (PhysRegDefsReach(CSMI, &MI, PhysRefs, PhysDefs, CrossMBBPhysDef))
594           FoundCSE = true;
595       }
596     }
597 
598     if (!FoundCSE) {
599       VNT.insert(&MI, CurrVN++);
600       Exps.push_back(&MI);
601       continue;
602     }
603 
604     // Found a common subexpression, eliminate it.
605     unsigned CSVN = VNT.lookup(&MI);
606     MachineInstr *CSMI = Exps[CSVN];
607     LLVM_DEBUG(dbgs() << "Examining: " << MI);
608     LLVM_DEBUG(dbgs() << "*** Found a common subexpression: " << *CSMI);
609 
610     // Prevent CSE-ing non-local convergent instructions.
611     // LLVM's current definition of `isConvergent` does not necessarily prove
612     // that non-local CSE is illegal. The following check extends the definition
613     // of `isConvergent` to assume a convergent instruction is dependent not
614     // only on additional conditions, but also on fewer conditions. LLVM does
615     // not have a MachineInstr attribute which expresses this extended
616     // definition, so it's necessary to use `isConvergent` to prevent illegally
617     // CSE-ing the subset of `isConvergent` instructions which do fall into this
618     // extended definition.
619     if (MI.isConvergent() && MI.getParent() != CSMI->getParent()) {
620       LLVM_DEBUG(dbgs() << "*** Convergent MI and subexpression exist in "
621                            "different BBs, avoid CSE!\n");
622       VNT.insert(&MI, CurrVN++);
623       Exps.push_back(&MI);
624       continue;
625     }
626 
627     // Check if it's profitable to perform this CSE.
628     bool DoCSE = true;
629     unsigned NumDefs = MI.getNumDefs();
630 
631     for (unsigned i = 0, e = MI.getNumOperands(); NumDefs && i != e; ++i) {
632       MachineOperand &MO = MI.getOperand(i);
633       if (!MO.isReg() || !MO.isDef())
634         continue;
635       Register OldReg = MO.getReg();
636       Register NewReg = CSMI->getOperand(i).getReg();
637 
638       // Go through implicit defs of CSMI and MI, if a def is not dead at MI,
639       // we should make sure it is not dead at CSMI.
640       if (MO.isImplicit() && !MO.isDead() && CSMI->getOperand(i).isDead())
641         ImplicitDefsToUpdate.push_back(i);
642 
643       // Keep track of implicit defs of CSMI and MI, to clear possibly
644       // made-redundant kill flags.
645       if (MO.isImplicit() && !MO.isDead() && OldReg == NewReg)
646         ImplicitDefs.push_back(OldReg);
647 
648       if (OldReg == NewReg) {
649         --NumDefs;
650         continue;
651       }
652 
653       assert(OldReg.isVirtual() && NewReg.isVirtual() &&
654              "Do not CSE physical register defs!");
655 
656       if (!isProfitableToCSE(NewReg, OldReg, CSMI->getParent(), &MI)) {
657         LLVM_DEBUG(dbgs() << "*** Not profitable, avoid CSE!\n");
658         DoCSE = false;
659         break;
660       }
661 
662       // Don't perform CSE if the result of the new instruction cannot exist
663       // within the constraints (register class, bank, or low-level type) of
664       // the old instruction.
665       if (!MRI->constrainRegAttrs(NewReg, OldReg)) {
666         LLVM_DEBUG(
667             dbgs() << "*** Not the same register constraints, avoid CSE!\n");
668         DoCSE = false;
669         break;
670       }
671 
672       CSEPairs.push_back(std::make_pair(OldReg, NewReg));
673       --NumDefs;
674     }
675 
676     // Actually perform the elimination.
677     if (DoCSE) {
678       for (const std::pair<unsigned, unsigned> &CSEPair : CSEPairs) {
679         unsigned OldReg = CSEPair.first;
680         unsigned NewReg = CSEPair.second;
681         // OldReg may have been unused but is used now, clear the Dead flag
682         MachineInstr *Def = MRI->getUniqueVRegDef(NewReg);
683         assert(Def != nullptr && "CSEd register has no unique definition?");
684         Def->clearRegisterDeads(NewReg);
685         // Replace with NewReg and clear kill flags which may be wrong now.
686         MRI->replaceRegWith(OldReg, NewReg);
687         MRI->clearKillFlags(NewReg);
688       }
689 
690       // Go through implicit defs of CSMI and MI, if a def is not dead at MI,
691       // we should make sure it is not dead at CSMI.
692       for (unsigned ImplicitDefToUpdate : ImplicitDefsToUpdate)
693         CSMI->getOperand(ImplicitDefToUpdate).setIsDead(false);
694       for (const auto &PhysDef : PhysDefs)
695         if (!MI.getOperand(PhysDef.first).isDead())
696           CSMI->getOperand(PhysDef.first).setIsDead(false);
697 
698       // Go through implicit defs of CSMI and MI, and clear the kill flags on
699       // their uses in all the instructions between CSMI and MI.
700       // We might have made some of the kill flags redundant, consider:
701       //   subs  ... implicit-def %nzcv    <- CSMI
702       //   csinc ... implicit killed %nzcv <- this kill flag isn't valid anymore
703       //   subs  ... implicit-def %nzcv    <- MI, to be eliminated
704       //   csinc ... implicit killed %nzcv
705       // Since we eliminated MI, and reused a register imp-def'd by CSMI
706       // (here %nzcv), that register, if it was killed before MI, should have
707       // that kill flag removed, because it's lifetime was extended.
708       if (CSMI->getParent() == MI.getParent()) {
709         for (MachineBasicBlock::iterator II = CSMI, IE = &MI; II != IE; ++II)
710           for (auto ImplicitDef : ImplicitDefs)
711             if (MachineOperand *MO = II->findRegisterUseOperand(
712                     ImplicitDef, TRI, /*isKill=*/true))
713               MO->setIsKill(false);
714       } else {
715         // If the instructions aren't in the same BB, bail out and clear the
716         // kill flag on all uses of the imp-def'd register.
717         for (auto ImplicitDef : ImplicitDefs)
718           MRI->clearKillFlags(ImplicitDef);
719       }
720 
721       if (CrossMBBPhysDef) {
722         // Add physical register defs now coming in from a predecessor to MBB
723         // livein list.
724         while (!PhysDefs.empty()) {
725           auto LiveIn = PhysDefs.pop_back_val();
726           if (!MBB->isLiveIn(LiveIn.second))
727             MBB->addLiveIn(LiveIn.second);
728         }
729         ++NumCrossBBCSEs;
730       }
731 
732       MI.eraseFromParent();
733       ++NumCSEs;
734       if (!PhysRefs.empty())
735         ++NumPhysCSEs;
736       if (Commuted)
737         ++NumCommutes;
738       Changed = true;
739     } else {
740       VNT.insert(&MI, CurrVN++);
741       Exps.push_back(&MI);
742     }
743     CSEPairs.clear();
744     ImplicitDefsToUpdate.clear();
745     ImplicitDefs.clear();
746   }
747 
748   return Changed;
749 }
750 
751 /// ExitScopeIfDone - Destroy scope for the MBB that corresponds to the given
752 /// dominator tree node if its a leaf or all of its children are done. Walk
753 /// up the dominator tree to destroy ancestors which are now done.
754 void
ExitScopeIfDone(MachineDomTreeNode * Node,DenseMap<MachineDomTreeNode *,unsigned> & OpenChildren)755 MachineCSE::ExitScopeIfDone(MachineDomTreeNode *Node,
756                         DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren) {
757   if (OpenChildren[Node])
758     return;
759 
760   // Pop scope.
761   ExitScope(Node->getBlock());
762 
763   // Now traverse upwards to pop ancestors whose offsprings are all done.
764   while (MachineDomTreeNode *Parent = Node->getIDom()) {
765     unsigned Left = --OpenChildren[Parent];
766     if (Left != 0)
767       break;
768     ExitScope(Parent->getBlock());
769     Node = Parent;
770   }
771 }
772 
PerformCSE(MachineDomTreeNode * Node)773 bool MachineCSE::PerformCSE(MachineDomTreeNode *Node) {
774   SmallVector<MachineDomTreeNode*, 32> Scopes;
775   SmallVector<MachineDomTreeNode*, 8> WorkList;
776   DenseMap<MachineDomTreeNode*, unsigned> OpenChildren;
777 
778   CurrVN = 0;
779 
780   // Perform a DFS walk to determine the order of visit.
781   WorkList.push_back(Node);
782   do {
783     Node = WorkList.pop_back_val();
784     Scopes.push_back(Node);
785     OpenChildren[Node] = Node->getNumChildren();
786     append_range(WorkList, Node->children());
787   } while (!WorkList.empty());
788 
789   // Now perform CSE.
790   bool Changed = false;
791   for (MachineDomTreeNode *Node : Scopes) {
792     MachineBasicBlock *MBB = Node->getBlock();
793     EnterScope(MBB);
794     Changed |= ProcessBlockCSE(MBB);
795     // If it's a leaf node, it's done. Traverse upwards to pop ancestors.
796     ExitScopeIfDone(Node, OpenChildren);
797   }
798 
799   return Changed;
800 }
801 
802 // We use stronger checks for PRE candidate rather than for CSE ones to embrace
803 // checks inside ProcessBlockCSE(), not only inside isCSECandidate(). This helps
804 // to exclude instrs created by PRE that won't be CSEed later.
isPRECandidate(MachineInstr * MI,SmallSet<MCRegister,8> & PhysRefs)805 bool MachineCSE::isPRECandidate(MachineInstr *MI,
806                                 SmallSet<MCRegister, 8> &PhysRefs) {
807   if (!isCSECandidate(MI) ||
808       MI->isNotDuplicable() ||
809       MI->mayLoad() ||
810       TII->isAsCheapAsAMove(*MI) ||
811       MI->getNumDefs() != 1 ||
812       MI->getNumExplicitDefs() != 1)
813     return false;
814 
815   for (const MachineOperand &MO : MI->operands()) {
816     if (MO.isReg() && !MO.getReg().isVirtual()) {
817       if (MO.isDef())
818         return false;
819       else
820         PhysRefs.insert(MO.getReg());
821     }
822   }
823 
824   return true;
825 }
826 
ProcessBlockPRE(MachineDominatorTree * DT,MachineBasicBlock * MBB)827 bool MachineCSE::ProcessBlockPRE(MachineDominatorTree *DT,
828                                  MachineBasicBlock *MBB) {
829   bool Changed = false;
830   for (MachineInstr &MI : llvm::make_early_inc_range(*MBB)) {
831     SmallSet<MCRegister, 8> PhysRefs;
832     if (!isPRECandidate(&MI, PhysRefs))
833       continue;
834 
835     if (!PREMap.count(&MI)) {
836       PREMap[&MI] = MBB;
837       continue;
838     }
839 
840     auto MBB1 = PREMap[&MI];
841     assert(
842         !DT->properlyDominates(MBB, MBB1) &&
843         "MBB cannot properly dominate MBB1 while DFS through dominators tree!");
844     auto CMBB = DT->findNearestCommonDominator(MBB, MBB1);
845     if (!CMBB->isLegalToHoistInto())
846       continue;
847 
848     if (!isProfitableToHoistInto(CMBB, MBB, MBB1))
849       continue;
850 
851     // Two instrs are partial redundant if their basic blocks are reachable
852     // from one to another but one doesn't dominate another.
853     if (CMBB != MBB1) {
854       auto BB = MBB->getBasicBlock(), BB1 = MBB1->getBasicBlock();
855       if (BB != nullptr && BB1 != nullptr &&
856           (isPotentiallyReachable(BB1, BB) ||
857            isPotentiallyReachable(BB, BB1))) {
858         // The following check extends the definition of `isConvergent` to
859         // assume a convergent instruction is dependent not only on additional
860         // conditions, but also on fewer conditions. LLVM does not have a
861         // MachineInstr attribute which expresses this extended definition, so
862         // it's necessary to use `isConvergent` to prevent illegally PRE-ing the
863         // subset of `isConvergent` instructions which do fall into this
864         // extended definition.
865         if (MI.isConvergent() && CMBB != MBB)
866           continue;
867 
868         // If this instruction uses physical registers then we can only do PRE
869         // if it's using the value that is live at the place we're hoisting to.
870         bool NonLocal;
871         PhysDefVector PhysDefs;
872         if (!PhysRefs.empty() &&
873             !PhysRegDefsReach(&*(CMBB->getFirstTerminator()), &MI, PhysRefs,
874                               PhysDefs, NonLocal))
875           continue;
876 
877         assert(MI.getOperand(0).isDef() &&
878                "First operand of instr with one explicit def must be this def");
879         Register VReg = MI.getOperand(0).getReg();
880         Register NewReg = MRI->cloneVirtualRegister(VReg);
881         if (!isProfitableToCSE(NewReg, VReg, CMBB, &MI))
882           continue;
883         MachineInstr &NewMI =
884             TII->duplicate(*CMBB, CMBB->getFirstTerminator(), MI);
885 
886         // When hoisting, make sure we don't carry the debug location of
887         // the original instruction, as that's not correct and can cause
888         // unexpected jumps when debugging optimized code.
889         auto EmptyDL = DebugLoc();
890         NewMI.setDebugLoc(EmptyDL);
891 
892         NewMI.getOperand(0).setReg(NewReg);
893 
894         PREMap[&MI] = CMBB;
895         ++NumPREs;
896         Changed = true;
897       }
898     }
899   }
900   return Changed;
901 }
902 
903 // This simple PRE (partial redundancy elimination) pass doesn't actually
904 // eliminate partial redundancy but transforms it to full redundancy,
905 // anticipating that the next CSE step will eliminate this created redundancy.
906 // If CSE doesn't eliminate this, than created instruction will remain dead
907 // and eliminated later by Remove Dead Machine Instructions pass.
PerformSimplePRE(MachineDominatorTree * DT)908 bool MachineCSE::PerformSimplePRE(MachineDominatorTree *DT) {
909   SmallVector<MachineDomTreeNode *, 32> BBs;
910 
911   PREMap.clear();
912   bool Changed = false;
913   BBs.push_back(DT->getRootNode());
914   do {
915     auto Node = BBs.pop_back_val();
916     append_range(BBs, Node->children());
917 
918     MachineBasicBlock *MBB = Node->getBlock();
919     Changed |= ProcessBlockPRE(DT, MBB);
920 
921   } while (!BBs.empty());
922 
923   return Changed;
924 }
925 
isProfitableToHoistInto(MachineBasicBlock * CandidateBB,MachineBasicBlock * MBB,MachineBasicBlock * MBB1)926 bool MachineCSE::isProfitableToHoistInto(MachineBasicBlock *CandidateBB,
927                                          MachineBasicBlock *MBB,
928                                          MachineBasicBlock *MBB1) {
929   if (CandidateBB->getParent()->getFunction().hasMinSize())
930     return true;
931   assert(DT->dominates(CandidateBB, MBB) && "CandidateBB should dominate MBB");
932   assert(DT->dominates(CandidateBB, MBB1) &&
933          "CandidateBB should dominate MBB1");
934   return MBFI->getBlockFreq(CandidateBB) <=
935          MBFI->getBlockFreq(MBB) + MBFI->getBlockFreq(MBB1);
936 }
937 
runOnMachineFunction(MachineFunction & MF)938 bool MachineCSE::runOnMachineFunction(MachineFunction &MF) {
939   if (skipFunction(MF.getFunction()))
940     return false;
941 
942   TII = MF.getSubtarget().getInstrInfo();
943   TRI = MF.getSubtarget().getRegisterInfo();
944   MRI = &MF.getRegInfo();
945   AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
946   DT = &getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree();
947   MBFI = &getAnalysis<MachineBlockFrequencyInfoWrapperPass>().getMBFI();
948   LookAheadLimit = TII->getMachineCSELookAheadLimit();
949   bool ChangedPRE, ChangedCSE;
950   ChangedPRE = PerformSimplePRE(DT);
951   ChangedCSE = PerformCSE(DT->getRootNode());
952   return ChangedPRE || ChangedCSE;
953 }
954