[InstCombine] Signed saturation patterns
[llvm-complete.git] / lib / CodeGen / MachineCSE.cpp
blobd9bd32b2fbab089fd6562d4d27187fb7a39c1f39
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 // Propagate SrcReg of copies to MI.
202 MO.setReg(SrcReg);
203 MRI->clearKillFlags(SrcReg);
204 // Coalesce single use copies.
205 if (OnlyOneUse) {
206 // If (and only if) we've eliminated all uses of the copy, also
207 // copy-propagate to any debug-users of MI, or they'll be left using
208 // an undefined value.
209 DefMI->changeDebugValuesDefReg(SrcReg);
211 DefMI->eraseFromParent();
212 ++NumCoalesces;
214 Changed = true;
217 return Changed;
220 bool
221 MachineCSE::isPhysDefTriviallyDead(unsigned Reg,
222 MachineBasicBlock::const_iterator I,
223 MachineBasicBlock::const_iterator E) const {
224 unsigned LookAheadLeft = LookAheadLimit;
225 while (LookAheadLeft) {
226 // Skip over dbg_value's.
227 I = skipDebugInstructionsForward(I, E);
229 if (I == E)
230 // Reached end of block, we don't know if register is dead or not.
231 return false;
233 bool SeenDef = false;
234 for (const MachineOperand &MO : I->operands()) {
235 if (MO.isRegMask() && MO.clobbersPhysReg(Reg))
236 SeenDef = true;
237 if (!MO.isReg() || !MO.getReg())
238 continue;
239 if (!TRI->regsOverlap(MO.getReg(), Reg))
240 continue;
241 if (MO.isUse())
242 // Found a use!
243 return false;
244 SeenDef = true;
246 if (SeenDef)
247 // See a def of Reg (or an alias) before encountering any use, it's
248 // trivially dead.
249 return true;
251 --LookAheadLeft;
252 ++I;
254 return false;
257 static bool isCallerPreservedOrConstPhysReg(unsigned Reg,
258 const MachineFunction &MF,
259 const TargetRegisterInfo &TRI) {
260 // MachineRegisterInfo::isConstantPhysReg directly called by
261 // MachineRegisterInfo::isCallerPreservedOrConstPhysReg expects the
262 // reserved registers to be frozen. That doesn't cause a problem post-ISel as
263 // most (if not all) targets freeze reserved registers right after ISel.
265 // It does cause issues mid-GlobalISel, however, hence the additional
266 // reservedRegsFrozen check.
267 const MachineRegisterInfo &MRI = MF.getRegInfo();
268 return TRI.isCallerPreservedPhysReg(Reg, MF) ||
269 (MRI.reservedRegsFrozen() && MRI.isConstantPhysReg(Reg));
272 /// hasLivePhysRegDefUses - Return true if the specified instruction read/write
273 /// physical registers (except for dead defs of physical registers). It also
274 /// returns the physical register def by reference if it's the only one and the
275 /// instruction does not uses a physical register.
276 bool MachineCSE::hasLivePhysRegDefUses(const MachineInstr *MI,
277 const MachineBasicBlock *MBB,
278 SmallSet<unsigned, 8> &PhysRefs,
279 PhysDefVector &PhysDefs,
280 bool &PhysUseDef) const {
281 // First, add all uses to PhysRefs.
282 for (const MachineOperand &MO : MI->operands()) {
283 if (!MO.isReg() || MO.isDef())
284 continue;
285 Register Reg = MO.getReg();
286 if (!Reg)
287 continue;
288 if (Register::isVirtualRegister(Reg))
289 continue;
290 // Reading either caller preserved or constant physregs is ok.
291 if (!isCallerPreservedOrConstPhysReg(Reg, *MI->getMF(), *TRI))
292 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
293 PhysRefs.insert(*AI);
296 // Next, collect all defs into PhysDefs. If any is already in PhysRefs
297 // (which currently contains only uses), set the PhysUseDef flag.
298 PhysUseDef = false;
299 MachineBasicBlock::const_iterator I = MI; I = std::next(I);
300 for (const auto &MOP : llvm::enumerate(MI->operands())) {
301 const MachineOperand &MO = MOP.value();
302 if (!MO.isReg() || !MO.isDef())
303 continue;
304 Register Reg = MO.getReg();
305 if (!Reg)
306 continue;
307 if (Register::isVirtualRegister(Reg))
308 continue;
309 // Check against PhysRefs even if the def is "dead".
310 if (PhysRefs.count(Reg))
311 PhysUseDef = true;
312 // If the def is dead, it's ok. But the def may not marked "dead". That's
313 // common since this pass is run before livevariables. We can scan
314 // forward a few instructions and check if it is obviously dead.
315 if (!MO.isDead() && !isPhysDefTriviallyDead(Reg, I, MBB->end()))
316 PhysDefs.push_back(std::make_pair(MOP.index(), Reg));
319 // Finally, add all defs to PhysRefs as well.
320 for (unsigned i = 0, e = PhysDefs.size(); i != e; ++i)
321 for (MCRegAliasIterator AI(PhysDefs[i].second, TRI, true); AI.isValid();
322 ++AI)
323 PhysRefs.insert(*AI);
325 return !PhysRefs.empty();
328 bool MachineCSE::PhysRegDefsReach(MachineInstr *CSMI, MachineInstr *MI,
329 SmallSet<unsigned, 8> &PhysRefs,
330 PhysDefVector &PhysDefs,
331 bool &NonLocal) const {
332 // For now conservatively returns false if the common subexpression is
333 // not in the same basic block as the given instruction. The only exception
334 // is if the common subexpression is in the sole predecessor block.
335 const MachineBasicBlock *MBB = MI->getParent();
336 const MachineBasicBlock *CSMBB = CSMI->getParent();
338 bool CrossMBB = false;
339 if (CSMBB != MBB) {
340 if (MBB->pred_size() != 1 || *MBB->pred_begin() != CSMBB)
341 return false;
343 for (unsigned i = 0, e = PhysDefs.size(); i != e; ++i) {
344 if (MRI->isAllocatable(PhysDefs[i].second) ||
345 MRI->isReserved(PhysDefs[i].second))
346 // Avoid extending live range of physical registers if they are
347 //allocatable or reserved.
348 return false;
350 CrossMBB = true;
352 MachineBasicBlock::const_iterator I = CSMI; I = std::next(I);
353 MachineBasicBlock::const_iterator E = MI;
354 MachineBasicBlock::const_iterator EE = CSMBB->end();
355 unsigned LookAheadLeft = LookAheadLimit;
356 while (LookAheadLeft) {
357 // Skip over dbg_value's.
358 while (I != E && I != EE && I->isDebugInstr())
359 ++I;
361 if (I == EE) {
362 assert(CrossMBB && "Reaching end-of-MBB without finding MI?");
363 (void)CrossMBB;
364 CrossMBB = false;
365 NonLocal = true;
366 I = MBB->begin();
367 EE = MBB->end();
368 continue;
371 if (I == E)
372 return true;
374 for (const MachineOperand &MO : I->operands()) {
375 // RegMasks go on instructions like calls that clobber lots of physregs.
376 // Don't attempt to CSE across such an instruction.
377 if (MO.isRegMask())
378 return false;
379 if (!MO.isReg() || !MO.isDef())
380 continue;
381 Register MOReg = MO.getReg();
382 if (Register::isVirtualRegister(MOReg))
383 continue;
384 if (PhysRefs.count(MOReg))
385 return false;
388 --LookAheadLeft;
389 ++I;
392 return false;
395 bool MachineCSE::isCSECandidate(MachineInstr *MI) {
396 if (MI->isPosition() || MI->isPHI() || MI->isImplicitDef() || MI->isKill() ||
397 MI->isInlineAsm() || MI->isDebugInstr())
398 return false;
400 // Ignore copies.
401 if (MI->isCopyLike())
402 return false;
404 // Ignore stuff that we obviously can't move.
405 if (MI->mayStore() || MI->isCall() || MI->isTerminator() ||
406 MI->mayRaiseFPException() || MI->hasUnmodeledSideEffects())
407 return false;
409 if (MI->mayLoad()) {
410 // Okay, this instruction does a load. As a refinement, we allow the target
411 // to decide whether the loaded value is actually a constant. If so, we can
412 // actually use it as a load.
413 if (!MI->isDereferenceableInvariantLoad(AA))
414 // FIXME: we should be able to hoist loads with no other side effects if
415 // there are no other instructions which can change memory in this loop.
416 // This is a trivial form of alias analysis.
417 return false;
420 // Ignore stack guard loads, otherwise the register that holds CSEed value may
421 // be spilled and get loaded back with corrupted data.
422 if (MI->getOpcode() == TargetOpcode::LOAD_STACK_GUARD)
423 return false;
425 return true;
428 /// isProfitableToCSE - Return true if it's profitable to eliminate MI with a
429 /// common expression that defines Reg. CSBB is basic block where CSReg is
430 /// defined.
431 bool MachineCSE::isProfitableToCSE(unsigned CSReg, unsigned Reg,
432 MachineBasicBlock *CSBB, MachineInstr *MI) {
433 // FIXME: Heuristics that works around the lack the live range splitting.
435 // If CSReg is used at all uses of Reg, CSE should not increase register
436 // pressure of CSReg.
437 bool MayIncreasePressure = true;
438 if (Register::isVirtualRegister(CSReg) && Register::isVirtualRegister(Reg)) {
439 MayIncreasePressure = false;
440 SmallPtrSet<MachineInstr*, 8> CSUses;
441 for (MachineInstr &MI : MRI->use_nodbg_instructions(CSReg)) {
442 CSUses.insert(&MI);
444 for (MachineInstr &MI : MRI->use_nodbg_instructions(Reg)) {
445 if (!CSUses.count(&MI)) {
446 MayIncreasePressure = true;
447 break;
451 if (!MayIncreasePressure) return true;
453 // Heuristics #1: Don't CSE "cheap" computation if the def is not local or in
454 // an immediate predecessor. We don't want to increase register pressure and
455 // end up causing other computation to be spilled.
456 if (TII->isAsCheapAsAMove(*MI)) {
457 MachineBasicBlock *BB = MI->getParent();
458 if (CSBB != BB && !CSBB->isSuccessor(BB))
459 return false;
462 // Heuristics #2: If the expression doesn't not use a vr and the only use
463 // of the redundant computation are copies, do not cse.
464 bool HasVRegUse = false;
465 for (const MachineOperand &MO : MI->operands()) {
466 if (MO.isReg() && MO.isUse() && Register::isVirtualRegister(MO.getReg())) {
467 HasVRegUse = true;
468 break;
471 if (!HasVRegUse) {
472 bool HasNonCopyUse = false;
473 for (MachineInstr &MI : MRI->use_nodbg_instructions(Reg)) {
474 // Ignore copies.
475 if (!MI.isCopyLike()) {
476 HasNonCopyUse = true;
477 break;
480 if (!HasNonCopyUse)
481 return false;
484 // Heuristics #3: If the common subexpression is used by PHIs, do not reuse
485 // it unless the defined value is already used in the BB of the new use.
486 bool HasPHI = false;
487 for (MachineInstr &UseMI : MRI->use_nodbg_instructions(CSReg)) {
488 HasPHI |= UseMI.isPHI();
489 if (UseMI.getParent() == MI->getParent())
490 return true;
493 return !HasPHI;
496 void MachineCSE::EnterScope(MachineBasicBlock *MBB) {
497 LLVM_DEBUG(dbgs() << "Entering: " << MBB->getName() << '\n');
498 ScopeType *Scope = new ScopeType(VNT);
499 ScopeMap[MBB] = Scope;
502 void MachineCSE::ExitScope(MachineBasicBlock *MBB) {
503 LLVM_DEBUG(dbgs() << "Exiting: " << MBB->getName() << '\n');
504 DenseMap<MachineBasicBlock*, ScopeType*>::iterator SI = ScopeMap.find(MBB);
505 assert(SI != ScopeMap.end());
506 delete SI->second;
507 ScopeMap.erase(SI);
510 bool MachineCSE::ProcessBlockCSE(MachineBasicBlock *MBB) {
511 bool Changed = false;
513 SmallVector<std::pair<unsigned, unsigned>, 8> CSEPairs;
514 SmallVector<unsigned, 2> ImplicitDefsToUpdate;
515 SmallVector<unsigned, 2> ImplicitDefs;
516 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E; ) {
517 MachineInstr *MI = &*I;
518 ++I;
520 if (!isCSECandidate(MI))
521 continue;
523 bool FoundCSE = VNT.count(MI);
524 if (!FoundCSE) {
525 // Using trivial copy propagation to find more CSE opportunities.
526 if (PerformTrivialCopyPropagation(MI, MBB)) {
527 Changed = true;
529 // After coalescing MI itself may become a copy.
530 if (MI->isCopyLike())
531 continue;
533 // Try again to see if CSE is possible.
534 FoundCSE = VNT.count(MI);
538 // Commute commutable instructions.
539 bool Commuted = false;
540 if (!FoundCSE && MI->isCommutable()) {
541 if (MachineInstr *NewMI = TII->commuteInstruction(*MI)) {
542 Commuted = true;
543 FoundCSE = VNT.count(NewMI);
544 if (NewMI != MI) {
545 // New instruction. It doesn't need to be kept.
546 NewMI->eraseFromParent();
547 Changed = true;
548 } else if (!FoundCSE)
549 // MI was changed but it didn't help, commute it back!
550 (void)TII->commuteInstruction(*MI);
554 // If the instruction defines physical registers and the values *may* be
555 // used, then it's not safe to replace it with a common subexpression.
556 // It's also not safe if the instruction uses physical registers.
557 bool CrossMBBPhysDef = false;
558 SmallSet<unsigned, 8> PhysRefs;
559 PhysDefVector PhysDefs;
560 bool PhysUseDef = false;
561 if (FoundCSE && hasLivePhysRegDefUses(MI, MBB, PhysRefs,
562 PhysDefs, PhysUseDef)) {
563 FoundCSE = false;
565 // ... Unless the CS is local or is in the sole predecessor block
566 // and it also defines the physical register which is not clobbered
567 // in between and the physical register uses were not clobbered.
568 // This can never be the case if the instruction both uses and
569 // defines the same physical register, which was detected above.
570 if (!PhysUseDef) {
571 unsigned CSVN = VNT.lookup(MI);
572 MachineInstr *CSMI = Exps[CSVN];
573 if (PhysRegDefsReach(CSMI, MI, PhysRefs, PhysDefs, CrossMBBPhysDef))
574 FoundCSE = true;
578 if (!FoundCSE) {
579 VNT.insert(MI, CurrVN++);
580 Exps.push_back(MI);
581 continue;
584 // Found a common subexpression, eliminate it.
585 unsigned CSVN = VNT.lookup(MI);
586 MachineInstr *CSMI = Exps[CSVN];
587 LLVM_DEBUG(dbgs() << "Examining: " << *MI);
588 LLVM_DEBUG(dbgs() << "*** Found a common subexpression: " << *CSMI);
590 // Check if it's profitable to perform this CSE.
591 bool DoCSE = true;
592 unsigned NumDefs = MI->getNumDefs();
594 for (unsigned i = 0, e = MI->getNumOperands(); NumDefs && i != e; ++i) {
595 MachineOperand &MO = MI->getOperand(i);
596 if (!MO.isReg() || !MO.isDef())
597 continue;
598 Register OldReg = MO.getReg();
599 Register NewReg = CSMI->getOperand(i).getReg();
601 // Go through implicit defs of CSMI and MI, if a def is not dead at MI,
602 // we should make sure it is not dead at CSMI.
603 if (MO.isImplicit() && !MO.isDead() && CSMI->getOperand(i).isDead())
604 ImplicitDefsToUpdate.push_back(i);
606 // Keep track of implicit defs of CSMI and MI, to clear possibly
607 // made-redundant kill flags.
608 if (MO.isImplicit() && !MO.isDead() && OldReg == NewReg)
609 ImplicitDefs.push_back(OldReg);
611 if (OldReg == NewReg) {
612 --NumDefs;
613 continue;
616 assert(Register::isVirtualRegister(OldReg) &&
617 Register::isVirtualRegister(NewReg) &&
618 "Do not CSE physical register defs!");
620 if (!isProfitableToCSE(NewReg, OldReg, CSMI->getParent(), MI)) {
621 LLVM_DEBUG(dbgs() << "*** Not profitable, avoid CSE!\n");
622 DoCSE = false;
623 break;
626 // Don't perform CSE if the result of the new instruction cannot exist
627 // within the constraints (register class, bank, or low-level type) of
628 // the old instruction.
629 if (!MRI->constrainRegAttrs(NewReg, OldReg)) {
630 LLVM_DEBUG(
631 dbgs() << "*** Not the same register constraints, avoid CSE!\n");
632 DoCSE = false;
633 break;
636 CSEPairs.push_back(std::make_pair(OldReg, NewReg));
637 --NumDefs;
640 // Actually perform the elimination.
641 if (DoCSE) {
642 for (std::pair<unsigned, unsigned> &CSEPair : CSEPairs) {
643 unsigned OldReg = CSEPair.first;
644 unsigned NewReg = CSEPair.second;
645 // OldReg may have been unused but is used now, clear the Dead flag
646 MachineInstr *Def = MRI->getUniqueVRegDef(NewReg);
647 assert(Def != nullptr && "CSEd register has no unique definition?");
648 Def->clearRegisterDeads(NewReg);
649 // Replace with NewReg and clear kill flags which may be wrong now.
650 MRI->replaceRegWith(OldReg, NewReg);
651 MRI->clearKillFlags(NewReg);
654 // Go through implicit defs of CSMI and MI, if a def is not dead at MI,
655 // we should make sure it is not dead at CSMI.
656 for (unsigned ImplicitDefToUpdate : ImplicitDefsToUpdate)
657 CSMI->getOperand(ImplicitDefToUpdate).setIsDead(false);
658 for (auto PhysDef : PhysDefs)
659 if (!MI->getOperand(PhysDef.first).isDead())
660 CSMI->getOperand(PhysDef.first).setIsDead(false);
662 // Go through implicit defs of CSMI and MI, and clear the kill flags on
663 // their uses in all the instructions between CSMI and MI.
664 // We might have made some of the kill flags redundant, consider:
665 // subs ... implicit-def %nzcv <- CSMI
666 // csinc ... implicit killed %nzcv <- this kill flag isn't valid anymore
667 // subs ... implicit-def %nzcv <- MI, to be eliminated
668 // csinc ... implicit killed %nzcv
669 // Since we eliminated MI, and reused a register imp-def'd by CSMI
670 // (here %nzcv), that register, if it was killed before MI, should have
671 // that kill flag removed, because it's lifetime was extended.
672 if (CSMI->getParent() == MI->getParent()) {
673 for (MachineBasicBlock::iterator II = CSMI, IE = MI; II != IE; ++II)
674 for (auto ImplicitDef : ImplicitDefs)
675 if (MachineOperand *MO = II->findRegisterUseOperand(
676 ImplicitDef, /*isKill=*/true, TRI))
677 MO->setIsKill(false);
678 } else {
679 // If the instructions aren't in the same BB, bail out and clear the
680 // kill flag on all uses of the imp-def'd register.
681 for (auto ImplicitDef : ImplicitDefs)
682 MRI->clearKillFlags(ImplicitDef);
685 if (CrossMBBPhysDef) {
686 // Add physical register defs now coming in from a predecessor to MBB
687 // livein list.
688 while (!PhysDefs.empty()) {
689 auto LiveIn = PhysDefs.pop_back_val();
690 if (!MBB->isLiveIn(LiveIn.second))
691 MBB->addLiveIn(LiveIn.second);
693 ++NumCrossBBCSEs;
696 MI->eraseFromParent();
697 ++NumCSEs;
698 if (!PhysRefs.empty())
699 ++NumPhysCSEs;
700 if (Commuted)
701 ++NumCommutes;
702 Changed = true;
703 } else {
704 VNT.insert(MI, CurrVN++);
705 Exps.push_back(MI);
707 CSEPairs.clear();
708 ImplicitDefsToUpdate.clear();
709 ImplicitDefs.clear();
712 return Changed;
715 /// ExitScopeIfDone - Destroy scope for the MBB that corresponds to the given
716 /// dominator tree node if its a leaf or all of its children are done. Walk
717 /// up the dominator tree to destroy ancestors which are now done.
718 void
719 MachineCSE::ExitScopeIfDone(MachineDomTreeNode *Node,
720 DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren) {
721 if (OpenChildren[Node])
722 return;
724 // Pop scope.
725 ExitScope(Node->getBlock());
727 // Now traverse upwards to pop ancestors whose offsprings are all done.
728 while (MachineDomTreeNode *Parent = Node->getIDom()) {
729 unsigned Left = --OpenChildren[Parent];
730 if (Left != 0)
731 break;
732 ExitScope(Parent->getBlock());
733 Node = Parent;
737 bool MachineCSE::PerformCSE(MachineDomTreeNode *Node) {
738 SmallVector<MachineDomTreeNode*, 32> Scopes;
739 SmallVector<MachineDomTreeNode*, 8> WorkList;
740 DenseMap<MachineDomTreeNode*, unsigned> OpenChildren;
742 CurrVN = 0;
744 // Perform a DFS walk to determine the order of visit.
745 WorkList.push_back(Node);
746 do {
747 Node = WorkList.pop_back_val();
748 Scopes.push_back(Node);
749 const std::vector<MachineDomTreeNode*> &Children = Node->getChildren();
750 OpenChildren[Node] = Children.size();
751 for (MachineDomTreeNode *Child : Children)
752 WorkList.push_back(Child);
753 } while (!WorkList.empty());
755 // Now perform CSE.
756 bool Changed = false;
757 for (MachineDomTreeNode *Node : Scopes) {
758 MachineBasicBlock *MBB = Node->getBlock();
759 EnterScope(MBB);
760 Changed |= ProcessBlockCSE(MBB);
761 // If it's a leaf node, it's done. Traverse upwards to pop ancestors.
762 ExitScopeIfDone(Node, OpenChildren);
765 return Changed;
768 // We use stronger checks for PRE candidate rather than for CSE ones to embrace
769 // checks inside ProcessBlockCSE(), not only inside isCSECandidate(). This helps
770 // to exclude instrs created by PRE that won't be CSEed later.
771 bool MachineCSE::isPRECandidate(MachineInstr *MI) {
772 if (!isCSECandidate(MI) ||
773 MI->isNotDuplicable() ||
774 MI->mayLoad() ||
775 MI->isAsCheapAsAMove() ||
776 MI->getNumDefs() != 1 ||
777 MI->getNumExplicitDefs() != 1)
778 return false;
780 for (auto def : MI->defs())
781 if (!Register::isVirtualRegister(def.getReg()))
782 return false;
784 for (auto use : MI->uses())
785 if (use.isReg() && !Register::isVirtualRegister(use.getReg()))
786 return false;
788 return true;
791 bool MachineCSE::ProcessBlockPRE(MachineDominatorTree *DT,
792 MachineBasicBlock *MBB) {
793 bool Changed = false;
794 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E;) {
795 MachineInstr *MI = &*I;
796 ++I;
798 if (!isPRECandidate(MI))
799 continue;
801 if (!PREMap.count(MI)) {
802 PREMap[MI] = MBB;
803 continue;
806 auto MBB1 = PREMap[MI];
807 assert(
808 !DT->properlyDominates(MBB, MBB1) &&
809 "MBB cannot properly dominate MBB1 while DFS through dominators tree!");
810 auto CMBB = DT->findNearestCommonDominator(MBB, MBB1);
811 if (!CMBB->isLegalToHoistInto())
812 continue;
814 if (!isProfitableToHoistInto(CMBB, MBB, MBB1))
815 continue;
817 // Two instrs are partial redundant if their basic blocks are reachable
818 // from one to another but one doesn't dominate another.
819 if (CMBB != MBB1) {
820 auto BB = MBB->getBasicBlock(), BB1 = MBB1->getBasicBlock();
821 if (BB != nullptr && BB1 != nullptr &&
822 (isPotentiallyReachable(BB1, BB) ||
823 isPotentiallyReachable(BB, BB1))) {
825 assert(MI->getOperand(0).isDef() &&
826 "First operand of instr with one explicit def must be this def");
827 Register VReg = MI->getOperand(0).getReg();
828 Register NewReg = MRI->cloneVirtualRegister(VReg);
829 if (!isProfitableToCSE(NewReg, VReg, CMBB, MI))
830 continue;
831 MachineInstr &NewMI =
832 TII->duplicate(*CMBB, CMBB->getFirstTerminator(), *MI);
833 NewMI.getOperand(0).setReg(NewReg);
835 PREMap[MI] = CMBB;
836 ++NumPREs;
837 Changed = true;
841 return Changed;
844 // This simple PRE (partial redundancy elimination) pass doesn't actually
845 // eliminate partial redundancy but transforms it to full redundancy,
846 // anticipating that the next CSE step will eliminate this created redundancy.
847 // If CSE doesn't eliminate this, than created instruction will remain dead
848 // and eliminated later by Remove Dead Machine Instructions pass.
849 bool MachineCSE::PerformSimplePRE(MachineDominatorTree *DT) {
850 SmallVector<MachineDomTreeNode *, 32> BBs;
852 PREMap.clear();
853 bool Changed = false;
854 BBs.push_back(DT->getRootNode());
855 do {
856 auto Node = BBs.pop_back_val();
857 const std::vector<MachineDomTreeNode *> &Children = Node->getChildren();
858 for (MachineDomTreeNode *Child : Children)
859 BBs.push_back(Child);
861 MachineBasicBlock *MBB = Node->getBlock();
862 Changed |= ProcessBlockPRE(DT, MBB);
864 } while (!BBs.empty());
866 return Changed;
869 bool MachineCSE::isProfitableToHoistInto(MachineBasicBlock *CandidateBB,
870 MachineBasicBlock *MBB,
871 MachineBasicBlock *MBB1) {
872 if (CandidateBB->getParent()->getFunction().hasMinSize())
873 return true;
874 assert(DT->dominates(CandidateBB, MBB) && "CandidateBB should dominate MBB");
875 assert(DT->dominates(CandidateBB, MBB1) &&
876 "CandidateBB should dominate MBB1");
877 return MBFI->getBlockFreq(CandidateBB) <=
878 MBFI->getBlockFreq(MBB) + MBFI->getBlockFreq(MBB1);
881 bool MachineCSE::runOnMachineFunction(MachineFunction &MF) {
882 if (skipFunction(MF.getFunction()))
883 return false;
885 TII = MF.getSubtarget().getInstrInfo();
886 TRI = MF.getSubtarget().getRegisterInfo();
887 MRI = &MF.getRegInfo();
888 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
889 DT = &getAnalysis<MachineDominatorTree>();
890 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
891 LookAheadLimit = TII->getMachineCSELookAheadLimit();
892 bool ChangedPRE, ChangedCSE;
893 ChangedPRE = PerformSimplePRE(DT);
894 ChangedCSE = PerformCSE(DT->getRootNode());
895 return ChangedPRE || ChangedCSE;