[llvm-objdump] - Remove one overload of reportError. NFCI.
[llvm-complete.git] / lib / CodeGen / MachineCSE.cpp
blob58d73f06599a3c432b5b034cd5367c1dfdac97a0
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.
13 //===----------------------------------------------------------------------===//
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/MC/MCInstrDesc.h"
37 #include "llvm/MC/MCRegisterInfo.h"
38 #include "llvm/Pass.h"
39 #include "llvm/Support/Allocator.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/RecyclingAllocator.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include <cassert>
44 #include <iterator>
45 #include <utility>
46 #include <vector>
48 using namespace llvm;
50 #define DEBUG_TYPE "machine-cse"
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");
62 namespace {
64 class MachineCSE : public MachineFunctionPass {
65 const TargetInstrInfo *TII;
66 const TargetRegisterInfo *TRI;
67 AliasAnalysis *AA;
68 MachineDominatorTree *DT;
69 MachineRegisterInfo *MRI;
70 MachineBlockFrequencyInfo *MBFI;
72 public:
73 static char ID; // Pass identification
75 MachineCSE() : MachineFunctionPass(ID) {
76 initializeMachineCSEPass(*PassRegistry::getPassRegistry());
79 bool runOnMachineFunction(MachineFunction &MF) override;
81 void getAnalysisUsage(AnalysisUsage &AU) const override {
82 AU.setPreservesCFG();
83 MachineFunctionPass::getAnalysisUsage(AU);
84 AU.addRequired<AAResultsWrapperPass>();
85 AU.addPreservedID(MachineLoopInfoID);
86 AU.addRequired<MachineDominatorTree>();
87 AU.addPreserved<MachineDominatorTree>();
88 AU.addRequired<MachineBlockFrequencyInfo>();
89 AU.addPreserved<MachineBlockFrequencyInfo>();
92 void releaseMemory() override {
93 ScopeMap.clear();
94 PREMap.clear();
95 Exps.clear();
98 private:
99 using AllocatorTy = RecyclingAllocator<BumpPtrAllocator,
100 ScopedHashTableVal<MachineInstr *, unsigned>>;
101 using ScopedHTType =
102 ScopedHashTable<MachineInstr *, unsigned, MachineInstrExpressionTrait,
103 AllocatorTy>;
104 using ScopeType = ScopedHTType::ScopeTy;
105 using PhysDefVector = SmallVector<std::pair<unsigned, unsigned>, 2>;
107 unsigned LookAheadLimit = 0;
108 DenseMap<MachineBasicBlock *, ScopeType *> ScopeMap;
109 DenseMap<MachineInstr *, MachineBasicBlock *, MachineInstrExpressionTrait>
110 PREMap;
111 ScopedHTType VNT;
112 SmallVector<MachineInstr *, 64> Exps;
113 unsigned CurrVN = 0;
115 bool PerformTrivialCopyPropagation(MachineInstr *MI,
116 MachineBasicBlock *MBB);
117 bool isPhysDefTriviallyDead(unsigned Reg,
118 MachineBasicBlock::const_iterator I,
119 MachineBasicBlock::const_iterator E) const;
120 bool hasLivePhysRegDefUses(const MachineInstr *MI,
121 const MachineBasicBlock *MBB,
122 SmallSet<unsigned, 8> &PhysRefs,
123 PhysDefVector &PhysDefs, bool &PhysUseDef) const;
124 bool PhysRegDefsReach(MachineInstr *CSMI, MachineInstr *MI,
125 SmallSet<unsigned, 8> &PhysRefs,
126 PhysDefVector &PhysDefs, bool &NonLocal) const;
127 bool isCSECandidate(MachineInstr *MI);
128 bool isProfitableToCSE(unsigned CSReg, unsigned Reg,
129 MachineBasicBlock *CSBB, MachineInstr *MI);
130 void EnterScope(MachineBasicBlock *MBB);
131 void ExitScope(MachineBasicBlock *MBB);
132 bool ProcessBlockCSE(MachineBasicBlock *MBB);
133 void ExitScopeIfDone(MachineDomTreeNode *Node,
134 DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren);
135 bool PerformCSE(MachineDomTreeNode *Node);
137 bool isPRECandidate(MachineInstr *MI);
138 bool ProcessBlockPRE(MachineDominatorTree *MDT, MachineBasicBlock *MBB);
139 bool PerformSimplePRE(MachineDominatorTree *DT);
140 /// Heuristics to see if it's profitable to move common computations of MBB
141 /// and MBB1 to CandidateBB.
142 bool isProfitableToHoistInto(MachineBasicBlock *CandidateBB,
143 MachineBasicBlock *MBB,
144 MachineBasicBlock *MBB1);
147 } // end anonymous namespace
149 char MachineCSE::ID = 0;
151 char &llvm::MachineCSEID = MachineCSE::ID;
153 INITIALIZE_PASS_BEGIN(MachineCSE, DEBUG_TYPE,
154 "Machine Common Subexpression Elimination", false, false)
155 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
156 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
157 INITIALIZE_PASS_END(MachineCSE, DEBUG_TYPE,
158 "Machine Common Subexpression Elimination", false, false)
160 /// The source register of a COPY machine instruction can be propagated to all
161 /// its users, and this propagation could increase the probability of finding
162 /// common subexpressions. If the COPY has only one user, the COPY itself can
163 /// be removed.
164 bool MachineCSE::PerformTrivialCopyPropagation(MachineInstr *MI,
165 MachineBasicBlock *MBB) {
166 bool Changed = false;
167 for (MachineOperand &MO : MI->operands()) {
168 if (!MO.isReg() || !MO.isUse())
169 continue;
170 Register Reg = MO.getReg();
171 if (!Register::isVirtualRegister(Reg))
172 continue;
173 bool OnlyOneUse = MRI->hasOneNonDBGUse(Reg);
174 MachineInstr *DefMI = MRI->getVRegDef(Reg);
175 if (!DefMI->isCopy())
176 continue;
177 Register SrcReg = DefMI->getOperand(1).getReg();
178 if (!Register::isVirtualRegister(SrcReg))
179 continue;
180 if (DefMI->getOperand(0).getSubReg())
181 continue;
182 // FIXME: We should trivially coalesce subregister copies to expose CSE
183 // opportunities on instructions with truncated operands (see
184 // cse-add-with-overflow.ll). This can be done here as follows:
185 // if (SrcSubReg)
186 // RC = TRI->getMatchingSuperRegClass(MRI->getRegClass(SrcReg), RC,
187 // SrcSubReg);
188 // MO.substVirtReg(SrcReg, SrcSubReg, *TRI);
190 // The 2-addr pass has been updated to handle coalesced subregs. However,
191 // some machine-specific code still can't handle it.
192 // To handle it properly we also need a way find a constrained subregister
193 // class given a super-reg class and subreg index.
194 if (DefMI->getOperand(1).getSubReg())
195 continue;
196 if (!MRI->constrainRegAttrs(SrcReg, Reg))
197 continue;
198 LLVM_DEBUG(dbgs() << "Coalescing: " << *DefMI);
199 LLVM_DEBUG(dbgs() << "*** to: " << *MI);
201 // Update matching debug values.
202 DefMI->changeDebugValuesDefReg(SrcReg);
204 // Propagate SrcReg of copies to MI.
205 MO.setReg(SrcReg);
206 MRI->clearKillFlags(SrcReg);
207 // Coalesce single use copies.
208 if (OnlyOneUse) {
209 DefMI->eraseFromParent();
210 ++NumCoalesces;
212 Changed = true;
215 return Changed;
218 bool
219 MachineCSE::isPhysDefTriviallyDead(unsigned Reg,
220 MachineBasicBlock::const_iterator I,
221 MachineBasicBlock::const_iterator E) const {
222 unsigned LookAheadLeft = LookAheadLimit;
223 while (LookAheadLeft) {
224 // Skip over dbg_value's.
225 I = skipDebugInstructionsForward(I, E);
227 if (I == E)
228 // Reached end of block, we don't know if register is dead or not.
229 return false;
231 bool SeenDef = false;
232 for (const MachineOperand &MO : I->operands()) {
233 if (MO.isRegMask() && MO.clobbersPhysReg(Reg))
234 SeenDef = true;
235 if (!MO.isReg() || !MO.getReg())
236 continue;
237 if (!TRI->regsOverlap(MO.getReg(), Reg))
238 continue;
239 if (MO.isUse())
240 // Found a use!
241 return false;
242 SeenDef = true;
244 if (SeenDef)
245 // See a def of Reg (or an alias) before encountering any use, it's
246 // trivially dead.
247 return true;
249 --LookAheadLeft;
250 ++I;
252 return false;
255 static bool isCallerPreservedOrConstPhysReg(unsigned Reg,
256 const MachineFunction &MF,
257 const TargetRegisterInfo &TRI) {
258 // MachineRegisterInfo::isConstantPhysReg directly called by
259 // MachineRegisterInfo::isCallerPreservedOrConstPhysReg expects the
260 // reserved registers to be frozen. That doesn't cause a problem post-ISel as
261 // most (if not all) targets freeze reserved registers right after ISel.
263 // It does cause issues mid-GlobalISel, however, hence the additional
264 // reservedRegsFrozen check.
265 const MachineRegisterInfo &MRI = MF.getRegInfo();
266 return TRI.isCallerPreservedPhysReg(Reg, MF) ||
267 (MRI.reservedRegsFrozen() && MRI.isConstantPhysReg(Reg));
270 /// hasLivePhysRegDefUses - Return true if the specified instruction read/write
271 /// physical registers (except for dead defs of physical registers). It also
272 /// returns the physical register def by reference if it's the only one and the
273 /// instruction does not uses a physical register.
274 bool MachineCSE::hasLivePhysRegDefUses(const MachineInstr *MI,
275 const MachineBasicBlock *MBB,
276 SmallSet<unsigned, 8> &PhysRefs,
277 PhysDefVector &PhysDefs,
278 bool &PhysUseDef) const {
279 // First, add all uses to PhysRefs.
280 for (const MachineOperand &MO : MI->operands()) {
281 if (!MO.isReg() || MO.isDef())
282 continue;
283 Register Reg = MO.getReg();
284 if (!Reg)
285 continue;
286 if (Register::isVirtualRegister(Reg))
287 continue;
288 // Reading either caller preserved or constant physregs is ok.
289 if (!isCallerPreservedOrConstPhysReg(Reg, *MI->getMF(), *TRI))
290 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
291 PhysRefs.insert(*AI);
294 // Next, collect all defs into PhysDefs. If any is already in PhysRefs
295 // (which currently contains only uses), set the PhysUseDef flag.
296 PhysUseDef = false;
297 MachineBasicBlock::const_iterator I = MI; I = std::next(I);
298 for (const auto &MOP : llvm::enumerate(MI->operands())) {
299 const MachineOperand &MO = MOP.value();
300 if (!MO.isReg() || !MO.isDef())
301 continue;
302 Register Reg = MO.getReg();
303 if (!Reg)
304 continue;
305 if (Register::isVirtualRegister(Reg))
306 continue;
307 // Check against PhysRefs even if the def is "dead".
308 if (PhysRefs.count(Reg))
309 PhysUseDef = true;
310 // If the def is dead, it's ok. But the def may not marked "dead". That's
311 // common since this pass is run before livevariables. We can scan
312 // forward a few instructions and check if it is obviously dead.
313 if (!MO.isDead() && !isPhysDefTriviallyDead(Reg, I, MBB->end()))
314 PhysDefs.push_back(std::make_pair(MOP.index(), Reg));
317 // Finally, add all defs to PhysRefs as well.
318 for (unsigned i = 0, e = PhysDefs.size(); i != e; ++i)
319 for (MCRegAliasIterator AI(PhysDefs[i].second, TRI, true); AI.isValid();
320 ++AI)
321 PhysRefs.insert(*AI);
323 return !PhysRefs.empty();
326 bool MachineCSE::PhysRegDefsReach(MachineInstr *CSMI, MachineInstr *MI,
327 SmallSet<unsigned, 8> &PhysRefs,
328 PhysDefVector &PhysDefs,
329 bool &NonLocal) const {
330 // For now conservatively returns false if the common subexpression is
331 // not in the same basic block as the given instruction. The only exception
332 // is if the common subexpression is in the sole predecessor block.
333 const MachineBasicBlock *MBB = MI->getParent();
334 const MachineBasicBlock *CSMBB = CSMI->getParent();
336 bool CrossMBB = false;
337 if (CSMBB != MBB) {
338 if (MBB->pred_size() != 1 || *MBB->pred_begin() != CSMBB)
339 return false;
341 for (unsigned i = 0, e = PhysDefs.size(); i != e; ++i) {
342 if (MRI->isAllocatable(PhysDefs[i].second) ||
343 MRI->isReserved(PhysDefs[i].second))
344 // Avoid extending live range of physical registers if they are
345 //allocatable or reserved.
346 return false;
348 CrossMBB = true;
350 MachineBasicBlock::const_iterator I = CSMI; I = std::next(I);
351 MachineBasicBlock::const_iterator E = MI;
352 MachineBasicBlock::const_iterator EE = CSMBB->end();
353 unsigned LookAheadLeft = LookAheadLimit;
354 while (LookAheadLeft) {
355 // Skip over dbg_value's.
356 while (I != E && I != EE && I->isDebugInstr())
357 ++I;
359 if (I == EE) {
360 assert(CrossMBB && "Reaching end-of-MBB without finding MI?");
361 (void)CrossMBB;
362 CrossMBB = false;
363 NonLocal = true;
364 I = MBB->begin();
365 EE = MBB->end();
366 continue;
369 if (I == E)
370 return true;
372 for (const MachineOperand &MO : I->operands()) {
373 // RegMasks go on instructions like calls that clobber lots of physregs.
374 // Don't attempt to CSE across such an instruction.
375 if (MO.isRegMask())
376 return false;
377 if (!MO.isReg() || !MO.isDef())
378 continue;
379 Register MOReg = MO.getReg();
380 if (Register::isVirtualRegister(MOReg))
381 continue;
382 if (PhysRefs.count(MOReg))
383 return false;
386 --LookAheadLeft;
387 ++I;
390 return false;
393 bool MachineCSE::isCSECandidate(MachineInstr *MI) {
394 if (MI->isPosition() || MI->isPHI() || MI->isImplicitDef() || MI->isKill() ||
395 MI->isInlineAsm() || MI->isDebugInstr())
396 return false;
398 // Ignore copies.
399 if (MI->isCopyLike())
400 return false;
402 // Ignore stuff that we obviously can't move.
403 if (MI->mayStore() || MI->isCall() || MI->isTerminator() ||
404 MI->mayRaiseFPException() || MI->hasUnmodeledSideEffects())
405 return false;
407 if (MI->mayLoad()) {
408 // Okay, this instruction does a load. As a refinement, we allow the target
409 // to decide whether the loaded value is actually a constant. If so, we can
410 // actually use it as a load.
411 if (!MI->isDereferenceableInvariantLoad(AA))
412 // FIXME: we should be able to hoist loads with no other side effects if
413 // there are no other instructions which can change memory in this loop.
414 // This is a trivial form of alias analysis.
415 return false;
418 // Ignore stack guard loads, otherwise the register that holds CSEed value may
419 // be spilled and get loaded back with corrupted data.
420 if (MI->getOpcode() == TargetOpcode::LOAD_STACK_GUARD)
421 return false;
423 return true;
426 /// isProfitableToCSE - Return true if it's profitable to eliminate MI with a
427 /// common expression that defines Reg. CSBB is basic block where CSReg is
428 /// defined.
429 bool MachineCSE::isProfitableToCSE(unsigned CSReg, unsigned Reg,
430 MachineBasicBlock *CSBB, MachineInstr *MI) {
431 // FIXME: Heuristics that works around the lack the live range splitting.
433 // If CSReg is used at all uses of Reg, CSE should not increase register
434 // pressure of CSReg.
435 bool MayIncreasePressure = true;
436 if (Register::isVirtualRegister(CSReg) && Register::isVirtualRegister(Reg)) {
437 MayIncreasePressure = false;
438 SmallPtrSet<MachineInstr*, 8> CSUses;
439 for (MachineInstr &MI : MRI->use_nodbg_instructions(CSReg)) {
440 CSUses.insert(&MI);
442 for (MachineInstr &MI : MRI->use_nodbg_instructions(Reg)) {
443 if (!CSUses.count(&MI)) {
444 MayIncreasePressure = true;
445 break;
449 if (!MayIncreasePressure) return true;
451 // Heuristics #1: Don't CSE "cheap" computation if the def is not local or in
452 // an immediate predecessor. We don't want to increase register pressure and
453 // end up causing other computation to be spilled.
454 if (TII->isAsCheapAsAMove(*MI)) {
455 MachineBasicBlock *BB = MI->getParent();
456 if (CSBB != BB && !CSBB->isSuccessor(BB))
457 return false;
460 // Heuristics #2: If the expression doesn't not use a vr and the only use
461 // of the redundant computation are copies, do not cse.
462 bool HasVRegUse = false;
463 for (const MachineOperand &MO : MI->operands()) {
464 if (MO.isReg() && MO.isUse() && Register::isVirtualRegister(MO.getReg())) {
465 HasVRegUse = true;
466 break;
469 if (!HasVRegUse) {
470 bool HasNonCopyUse = false;
471 for (MachineInstr &MI : MRI->use_nodbg_instructions(Reg)) {
472 // Ignore copies.
473 if (!MI.isCopyLike()) {
474 HasNonCopyUse = true;
475 break;
478 if (!HasNonCopyUse)
479 return false;
482 // Heuristics #3: If the common subexpression is used by PHIs, do not reuse
483 // it unless the defined value is already used in the BB of the new use.
484 bool HasPHI = false;
485 for (MachineInstr &UseMI : MRI->use_nodbg_instructions(CSReg)) {
486 HasPHI |= UseMI.isPHI();
487 if (UseMI.getParent() == MI->getParent())
488 return true;
491 return !HasPHI;
494 void MachineCSE::EnterScope(MachineBasicBlock *MBB) {
495 LLVM_DEBUG(dbgs() << "Entering: " << MBB->getName() << '\n');
496 ScopeType *Scope = new ScopeType(VNT);
497 ScopeMap[MBB] = Scope;
500 void MachineCSE::ExitScope(MachineBasicBlock *MBB) {
501 LLVM_DEBUG(dbgs() << "Exiting: " << MBB->getName() << '\n');
502 DenseMap<MachineBasicBlock*, ScopeType*>::iterator SI = ScopeMap.find(MBB);
503 assert(SI != ScopeMap.end());
504 delete SI->second;
505 ScopeMap.erase(SI);
508 bool MachineCSE::ProcessBlockCSE(MachineBasicBlock *MBB) {
509 bool Changed = false;
511 SmallVector<std::pair<unsigned, unsigned>, 8> CSEPairs;
512 SmallVector<unsigned, 2> ImplicitDefsToUpdate;
513 SmallVector<unsigned, 2> ImplicitDefs;
514 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E; ) {
515 MachineInstr *MI = &*I;
516 ++I;
518 if (!isCSECandidate(MI))
519 continue;
521 bool FoundCSE = VNT.count(MI);
522 if (!FoundCSE) {
523 // Using trivial copy propagation to find more CSE opportunities.
524 if (PerformTrivialCopyPropagation(MI, MBB)) {
525 Changed = true;
527 // After coalescing MI itself may become a copy.
528 if (MI->isCopyLike())
529 continue;
531 // Try again to see if CSE is possible.
532 FoundCSE = VNT.count(MI);
536 // Commute commutable instructions.
537 bool Commuted = false;
538 if (!FoundCSE && MI->isCommutable()) {
539 if (MachineInstr *NewMI = TII->commuteInstruction(*MI)) {
540 Commuted = true;
541 FoundCSE = VNT.count(NewMI);
542 if (NewMI != MI) {
543 // New instruction. It doesn't need to be kept.
544 NewMI->eraseFromParent();
545 Changed = true;
546 } else if (!FoundCSE)
547 // MI was changed but it didn't help, commute it back!
548 (void)TII->commuteInstruction(*MI);
552 // If the instruction defines physical registers and the values *may* be
553 // used, then it's not safe to replace it with a common subexpression.
554 // It's also not safe if the instruction uses physical registers.
555 bool CrossMBBPhysDef = false;
556 SmallSet<unsigned, 8> PhysRefs;
557 PhysDefVector PhysDefs;
558 bool PhysUseDef = false;
559 if (FoundCSE && hasLivePhysRegDefUses(MI, MBB, PhysRefs,
560 PhysDefs, PhysUseDef)) {
561 FoundCSE = false;
563 // ... Unless the CS is local or is in the sole predecessor block
564 // and it also defines the physical register which is not clobbered
565 // in between and the physical register uses were not clobbered.
566 // This can never be the case if the instruction both uses and
567 // defines the same physical register, which was detected above.
568 if (!PhysUseDef) {
569 unsigned CSVN = VNT.lookup(MI);
570 MachineInstr *CSMI = Exps[CSVN];
571 if (PhysRegDefsReach(CSMI, MI, PhysRefs, PhysDefs, CrossMBBPhysDef))
572 FoundCSE = true;
576 if (!FoundCSE) {
577 VNT.insert(MI, CurrVN++);
578 Exps.push_back(MI);
579 continue;
582 // Found a common subexpression, eliminate it.
583 unsigned CSVN = VNT.lookup(MI);
584 MachineInstr *CSMI = Exps[CSVN];
585 LLVM_DEBUG(dbgs() << "Examining: " << *MI);
586 LLVM_DEBUG(dbgs() << "*** Found a common subexpression: " << *CSMI);
588 // Check if it's profitable to perform this CSE.
589 bool DoCSE = true;
590 unsigned NumDefs = MI->getNumDefs();
592 for (unsigned i = 0, e = MI->getNumOperands(); NumDefs && i != e; ++i) {
593 MachineOperand &MO = MI->getOperand(i);
594 if (!MO.isReg() || !MO.isDef())
595 continue;
596 Register OldReg = MO.getReg();
597 Register NewReg = CSMI->getOperand(i).getReg();
599 // Go through implicit defs of CSMI and MI, if a def is not dead at MI,
600 // we should make sure it is not dead at CSMI.
601 if (MO.isImplicit() && !MO.isDead() && CSMI->getOperand(i).isDead())
602 ImplicitDefsToUpdate.push_back(i);
604 // Keep track of implicit defs of CSMI and MI, to clear possibly
605 // made-redundant kill flags.
606 if (MO.isImplicit() && !MO.isDead() && OldReg == NewReg)
607 ImplicitDefs.push_back(OldReg);
609 if (OldReg == NewReg) {
610 --NumDefs;
611 continue;
614 assert(Register::isVirtualRegister(OldReg) &&
615 Register::isVirtualRegister(NewReg) &&
616 "Do not CSE physical register defs!");
618 if (!isProfitableToCSE(NewReg, OldReg, CSMI->getParent(), MI)) {
619 LLVM_DEBUG(dbgs() << "*** Not profitable, avoid CSE!\n");
620 DoCSE = false;
621 break;
624 // Don't perform CSE if the result of the new instruction cannot exist
625 // within the constraints (register class, bank, or low-level type) of
626 // the old instruction.
627 if (!MRI->constrainRegAttrs(NewReg, OldReg)) {
628 LLVM_DEBUG(
629 dbgs() << "*** Not the same register constraints, avoid CSE!\n");
630 DoCSE = false;
631 break;
634 CSEPairs.push_back(std::make_pair(OldReg, NewReg));
635 --NumDefs;
638 // Actually perform the elimination.
639 if (DoCSE) {
640 for (std::pair<unsigned, unsigned> &CSEPair : CSEPairs) {
641 unsigned OldReg = CSEPair.first;
642 unsigned NewReg = CSEPair.second;
643 // OldReg may have been unused but is used now, clear the Dead flag
644 MachineInstr *Def = MRI->getUniqueVRegDef(NewReg);
645 assert(Def != nullptr && "CSEd register has no unique definition?");
646 Def->clearRegisterDeads(NewReg);
647 // Replace with NewReg and clear kill flags which may be wrong now.
648 MRI->replaceRegWith(OldReg, NewReg);
649 MRI->clearKillFlags(NewReg);
652 // Go through implicit defs of CSMI and MI, if a def is not dead at MI,
653 // we should make sure it is not dead at CSMI.
654 for (unsigned ImplicitDefToUpdate : ImplicitDefsToUpdate)
655 CSMI->getOperand(ImplicitDefToUpdate).setIsDead(false);
656 for (auto PhysDef : PhysDefs)
657 if (!MI->getOperand(PhysDef.first).isDead())
658 CSMI->getOperand(PhysDef.first).setIsDead(false);
660 // Go through implicit defs of CSMI and MI, and clear the kill flags on
661 // their uses in all the instructions between CSMI and MI.
662 // We might have made some of the kill flags redundant, consider:
663 // subs ... implicit-def %nzcv <- CSMI
664 // csinc ... implicit killed %nzcv <- this kill flag isn't valid anymore
665 // subs ... implicit-def %nzcv <- MI, to be eliminated
666 // csinc ... implicit killed %nzcv
667 // Since we eliminated MI, and reused a register imp-def'd by CSMI
668 // (here %nzcv), that register, if it was killed before MI, should have
669 // that kill flag removed, because it's lifetime was extended.
670 if (CSMI->getParent() == MI->getParent()) {
671 for (MachineBasicBlock::iterator II = CSMI, IE = MI; II != IE; ++II)
672 for (auto ImplicitDef : ImplicitDefs)
673 if (MachineOperand *MO = II->findRegisterUseOperand(
674 ImplicitDef, /*isKill=*/true, TRI))
675 MO->setIsKill(false);
676 } else {
677 // If the instructions aren't in the same BB, bail out and clear the
678 // kill flag on all uses of the imp-def'd register.
679 for (auto ImplicitDef : ImplicitDefs)
680 MRI->clearKillFlags(ImplicitDef);
683 if (CrossMBBPhysDef) {
684 // Add physical register defs now coming in from a predecessor to MBB
685 // livein list.
686 while (!PhysDefs.empty()) {
687 auto LiveIn = PhysDefs.pop_back_val();
688 if (!MBB->isLiveIn(LiveIn.second))
689 MBB->addLiveIn(LiveIn.second);
691 ++NumCrossBBCSEs;
694 MI->eraseFromParent();
695 ++NumCSEs;
696 if (!PhysRefs.empty())
697 ++NumPhysCSEs;
698 if (Commuted)
699 ++NumCommutes;
700 Changed = true;
701 } else {
702 VNT.insert(MI, CurrVN++);
703 Exps.push_back(MI);
705 CSEPairs.clear();
706 ImplicitDefsToUpdate.clear();
707 ImplicitDefs.clear();
710 return Changed;
713 /// ExitScopeIfDone - Destroy scope for the MBB that corresponds to the given
714 /// dominator tree node if its a leaf or all of its children are done. Walk
715 /// up the dominator tree to destroy ancestors which are now done.
716 void
717 MachineCSE::ExitScopeIfDone(MachineDomTreeNode *Node,
718 DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren) {
719 if (OpenChildren[Node])
720 return;
722 // Pop scope.
723 ExitScope(Node->getBlock());
725 // Now traverse upwards to pop ancestors whose offsprings are all done.
726 while (MachineDomTreeNode *Parent = Node->getIDom()) {
727 unsigned Left = --OpenChildren[Parent];
728 if (Left != 0)
729 break;
730 ExitScope(Parent->getBlock());
731 Node = Parent;
735 bool MachineCSE::PerformCSE(MachineDomTreeNode *Node) {
736 SmallVector<MachineDomTreeNode*, 32> Scopes;
737 SmallVector<MachineDomTreeNode*, 8> WorkList;
738 DenseMap<MachineDomTreeNode*, unsigned> OpenChildren;
740 CurrVN = 0;
742 // Perform a DFS walk to determine the order of visit.
743 WorkList.push_back(Node);
744 do {
745 Node = WorkList.pop_back_val();
746 Scopes.push_back(Node);
747 const std::vector<MachineDomTreeNode*> &Children = Node->getChildren();
748 OpenChildren[Node] = Children.size();
749 for (MachineDomTreeNode *Child : Children)
750 WorkList.push_back(Child);
751 } while (!WorkList.empty());
753 // Now perform CSE.
754 bool Changed = false;
755 for (MachineDomTreeNode *Node : Scopes) {
756 MachineBasicBlock *MBB = Node->getBlock();
757 EnterScope(MBB);
758 Changed |= ProcessBlockCSE(MBB);
759 // If it's a leaf node, it's done. Traverse upwards to pop ancestors.
760 ExitScopeIfDone(Node, OpenChildren);
763 return Changed;
766 // We use stronger checks for PRE candidate rather than for CSE ones to embrace
767 // checks inside ProcessBlockCSE(), not only inside isCSECandidate(). This helps
768 // to exclude instrs created by PRE that won't be CSEed later.
769 bool MachineCSE::isPRECandidate(MachineInstr *MI) {
770 if (!isCSECandidate(MI) ||
771 MI->isNotDuplicable() ||
772 MI->mayLoad() ||
773 MI->isAsCheapAsAMove() ||
774 MI->getNumDefs() != 1 ||
775 MI->getNumExplicitDefs() != 1)
776 return false;
778 for (auto def : MI->defs())
779 if (!Register::isVirtualRegister(def.getReg()))
780 return false;
782 for (auto use : MI->uses())
783 if (use.isReg() && !Register::isVirtualRegister(use.getReg()))
784 return false;
786 return true;
789 bool MachineCSE::ProcessBlockPRE(MachineDominatorTree *DT,
790 MachineBasicBlock *MBB) {
791 bool Changed = false;
792 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E;) {
793 MachineInstr *MI = &*I;
794 ++I;
796 if (!isPRECandidate(MI))
797 continue;
799 if (!PREMap.count(MI)) {
800 PREMap[MI] = MBB;
801 continue;
804 auto MBB1 = PREMap[MI];
805 assert(
806 !DT->properlyDominates(MBB, MBB1) &&
807 "MBB cannot properly dominate MBB1 while DFS through dominators tree!");
808 auto CMBB = DT->findNearestCommonDominator(MBB, MBB1);
809 if (!CMBB->isLegalToHoistInto())
810 continue;
812 if (!isProfitableToHoistInto(CMBB, MBB, MBB1))
813 continue;
815 // Two instrs are partial redundant if their basic blocks are reachable
816 // from one to another but one doesn't dominate another.
817 if (CMBB != MBB1) {
818 auto BB = MBB->getBasicBlock(), BB1 = MBB1->getBasicBlock();
819 if (BB != nullptr && BB1 != nullptr &&
820 (isPotentiallyReachable(BB1, BB) ||
821 isPotentiallyReachable(BB, BB1))) {
823 assert(MI->getOperand(0).isDef() &&
824 "First operand of instr with one explicit def must be this def");
825 Register VReg = MI->getOperand(0).getReg();
826 Register NewReg = MRI->cloneVirtualRegister(VReg);
827 if (!isProfitableToCSE(NewReg, VReg, CMBB, MI))
828 continue;
829 MachineInstr &NewMI =
830 TII->duplicate(*CMBB, CMBB->getFirstTerminator(), *MI);
831 NewMI.getOperand(0).setReg(NewReg);
833 PREMap[MI] = CMBB;
834 ++NumPREs;
835 Changed = true;
839 return Changed;
842 // This simple PRE (partial redundancy elimination) pass doesn't actually
843 // eliminate partial redundancy but transforms it to full redundancy,
844 // anticipating that the next CSE step will eliminate this created redundancy.
845 // If CSE doesn't eliminate this, than created instruction will remain dead
846 // and eliminated later by Remove Dead Machine Instructions pass.
847 bool MachineCSE::PerformSimplePRE(MachineDominatorTree *DT) {
848 SmallVector<MachineDomTreeNode *, 32> BBs;
850 PREMap.clear();
851 bool Changed = false;
852 BBs.push_back(DT->getRootNode());
853 do {
854 auto Node = BBs.pop_back_val();
855 const std::vector<MachineDomTreeNode *> &Children = Node->getChildren();
856 for (MachineDomTreeNode *Child : Children)
857 BBs.push_back(Child);
859 MachineBasicBlock *MBB = Node->getBlock();
860 Changed |= ProcessBlockPRE(DT, MBB);
862 } while (!BBs.empty());
864 return Changed;
867 bool MachineCSE::isProfitableToHoistInto(MachineBasicBlock *CandidateBB,
868 MachineBasicBlock *MBB,
869 MachineBasicBlock *MBB1) {
870 if (CandidateBB->getParent()->getFunction().hasMinSize())
871 return true;
872 assert(DT->dominates(CandidateBB, MBB) && "CandidateBB should dominate MBB");
873 assert(DT->dominates(CandidateBB, MBB1) &&
874 "CandidateBB should dominate MBB1");
875 return MBFI->getBlockFreq(CandidateBB) <=
876 MBFI->getBlockFreq(MBB) + MBFI->getBlockFreq(MBB1);
879 bool MachineCSE::runOnMachineFunction(MachineFunction &MF) {
880 if (skipFunction(MF.getFunction()))
881 return false;
883 TII = MF.getSubtarget().getInstrInfo();
884 TRI = MF.getSubtarget().getRegisterInfo();
885 MRI = &MF.getRegInfo();
886 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
887 DT = &getAnalysis<MachineDominatorTree>();
888 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
889 LookAheadLimit = TII->getMachineCSELookAheadLimit();
890 bool ChangedPRE, ChangedCSE;
891 ChangedPRE = PerformSimplePRE(DT);
892 ChangedCSE = PerformCSE(DT->getRootNode());
893 return ChangedPRE || ChangedCSE;