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
48using namespace llvm;
49
50#define DEBUG_TYPE "machine-cse"
51
52STATISTIC(NumCoalesces, "Number of copies coalesced");
53STATISTIC(NumCSEs, "Number of common subexpression eliminated");
54STATISTIC(NumPREs, "Number of partial redundant expression"
55 " transformed to fully redundant");
56STATISTIC(NumPhysCSEs,
57 "Number of physreg referencing common subexpr eliminated");
58STATISTIC(NumCrossBBCSEs,
59 "Number of cross-MBB physreg referencing CS eliminated");
60STATISTIC(NumCommutes, "Number of copies coalesced after commuting");
61
62// Threshold to avoid excessive cost to compute isProfitableToCSE.
63static cl::opt<int>
64 CSUsesThreshold("csuses-threshold", cl::Hidden, cl::init(Val: 1024),
65 cl::desc("Threshold for the size of CSUses"));
66
67static cl::opt<bool> AggressiveMachineCSE(
68 "aggressive-machine-cse", cl::Hidden, cl::init(Val: false),
69 cl::desc("Override the profitability heuristics for Machine CSE"));
70
71namespace {
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
84 MachineCSE() : MachineFunctionPass(ID) {
85 initializeMachineCSEPass(*PassRegistry::getPassRegistry());
86 }
87
88 bool runOnMachineFunction(MachineFunction &MF) override;
89
90 void getAnalysisUsage(AnalysisUsage &AU) const override {
91 AU.setPreservesCFG();
92 MachineFunctionPass::getAnalysisUsage(AU);
93 AU.addRequired<AAResultsWrapperPass>();
94 AU.addPreservedID(ID&: MachineLoopInfoID);
95 AU.addRequired<MachineDominatorTreeWrapperPass>();
96 AU.addPreserved<MachineDominatorTreeWrapperPass>();
97 AU.addRequired<MachineBlockFrequencyInfoWrapperPass>();
98 AU.addPreserved<MachineBlockFrequencyInfoWrapperPass>();
99 }
100
101 MachineFunctionProperties getRequiredProperties() const override {
102 return MachineFunctionProperties()
103 .set(MachineFunctionProperties::Property::IsSSA);
104 }
105
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
163char MachineCSE::ID = 0;
164
165char &llvm::MachineCSEID = MachineCSE::ID;
166
167INITIALIZE_PASS_BEGIN(MachineCSE, DEBUG_TYPE,
168 "Machine Common Subexpression Elimination", false, false)
169INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass)
170INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
171INITIALIZE_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.
178bool 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(RegNo: Reg);
186 MachineInstr *DefMI = MRI->getVRegDef(Reg);
187 if (!DefMI || !DefMI->isCopy())
188 continue;
189 Register SrcReg = DefMI->getOperand(i: 1).getReg();
190 if (!SrcReg.isVirtual())
191 continue;
192 if (DefMI->getOperand(i: 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(i: 1).getSubReg())
207 continue;
208 if (!MRI->constrainRegAttrs(Reg: SrcReg, ConstrainingReg: 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(Reg: 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(Reg: SrcReg);
222
223 DefMI->eraseFromParent();
224 ++NumCoalesces;
225 }
226 Changed = true;
227 }
228
229 return Changed;
230}
231
232bool 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(It: I, End: 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(PhysReg: Reg))
247 SeenDef = true;
248 if (!MO.isReg() || !MO.getReg())
249 continue;
250 if (!TRI->regsOverlap(RegA: MO.getReg(), RegB: 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
268static 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(PhysReg: Reg, MF) || TII.isIgnorableUse(MO) ||
282 (MRI.reservedRegsFrozen() && MRI.isConstantPhysReg(PhysReg: 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.
289bool 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: Reg.asMCReg(), MO, MF: *MI->getMF(), TRI: *TRI,
303 TII: *TII))
304 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
305 PhysRefs.insert(V: *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(x: I);
312 for (const auto &MOP : llvm::enumerate(First: 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(V: 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: Reg.asMCReg(), I, E: MBB->end()))
328 PhysDefs.push_back(Elt: std::make_pair(x: MOP.index(), y&: 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(V: *AI);
336
337 return !PhysRefs.empty();
338}
339
340bool 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(PhysReg: PhysDefs[i].second) ||
357 MRI->isReserved(PhysReg: 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(x: 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(V: MOReg.asMCReg()))
397 return false;
398 }
399
400 --LookAheadLeft;
401 ++I;
402 }
403
404 return false;
405}
406
407bool 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.
443bool 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(Reg: CSReg)) {
458 CSUses.insert(Ptr: &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(Ptr: &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: *MI)) {
480 MachineBasicBlock *BB = MI->getParent();
481 if (CSBB != BB && !CSBB->isSuccessor(MBB: 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(Reg: CSReg)) {
511 HasPHI |= UseMI.isPHI();
512 if (UseMI.getParent() == MI->getParent())
513 return true;
514 }
515
516 return !HasPHI;
517}
518
519void MachineCSE::EnterScope(MachineBasicBlock *MBB) {
520 LLVM_DEBUG(dbgs() << "Entering: " << MBB->getName() << '\n');
521 ScopeType *Scope = new ScopeType(VNT);
522 ScopeMap[MBB] = Scope;
523}
524
525void MachineCSE::ExitScope(MachineBasicBlock *MBB) {
526 LLVM_DEBUG(dbgs() << "Exiting: " << MBB->getName() << '\n');
527 DenseMap<MachineBasicBlock*, ScopeType*>::iterator SI = ScopeMap.find(Val: MBB);
528 assert(SI != ScopeMap.end());
529 delete SI->second;
530 ScopeMap.erase(I: SI);
531}
532
533bool 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(Range&: *MBB)) {
540 if (!isCSECandidate(MI: &MI))
541 continue;
542
543 bool FoundCSE = VNT.count(Key: &MI);
544 if (!FoundCSE) {
545 // Using trivial copy propagation to find more CSE opportunities.
546 if (PerformTrivialCopyPropagation(MI: &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(Key: &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(Key: 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: &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(Key: &MI);
592 MachineInstr *CSMI = Exps[CSVN];
593 if (PhysRegDefsReach(CSMI, MI: &MI, PhysRefs, PhysDefs, NonLocal&: CrossMBBPhysDef))
594 FoundCSE = true;
595 }
596 }
597
598 if (!FoundCSE) {
599 VNT.insert(Key: &MI, Val: CurrVN++);
600 Exps.push_back(Elt: &MI);
601 continue;
602 }
603
604 // Found a common subexpression, eliminate it.
605 unsigned CSVN = VNT.lookup(Key: &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(Key: &MI, Val: CurrVN++);
623 Exps.push_back(Elt: &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(Elt: 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(Elt: 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(CSReg: NewReg, Reg: OldReg, CSBB: CSMI->getParent(), MI: &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(Reg: NewReg, ConstrainingReg: 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(Elt: std::make_pair(x&: OldReg, y&: 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(Reg: NewReg);
683 assert(Def != nullptr && "CSEd register has no unique definition?");
684 Def->clearRegisterDeads(Reg: NewReg);
685 // Replace with NewReg and clear kill flags which may be wrong now.
686 MRI->replaceRegWith(FromReg: OldReg, ToReg: NewReg);
687 MRI->clearKillFlags(Reg: 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(i: ImplicitDefToUpdate).setIsDead(false);
694 for (const auto &PhysDef : PhysDefs)
695 if (!MI.getOperand(i: PhysDef.first).isDead())
696 CSMI->getOperand(i: 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 Reg: 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(Reg: 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(Reg: LiveIn.second))
727 MBB->addLiveIn(PhysReg: 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(Key: &MI, Val: CurrVN++);
741 Exps.push_back(Elt: &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.
754void
755MachineCSE::ExitScopeIfDone(MachineDomTreeNode *Node,
756 DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren) {
757 if (OpenChildren[Node])
758 return;
759
760 // Pop scope.
761 ExitScope(MBB: 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(MBB: Parent->getBlock());
769 Node = Parent;
770 }
771}
772
773bool 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(Elt: Node);
782 do {
783 Node = WorkList.pop_back_val();
784 Scopes.push_back(Elt: Node);
785 OpenChildren[Node] = Node->getNumChildren();
786 append_range(C&: WorkList, R: 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.
805bool MachineCSE::isPRECandidate(MachineInstr *MI,
806 SmallSet<MCRegister, 8> &PhysRefs) {
807 if (!isCSECandidate(MI) ||
808 MI->isNotDuplicable() ||
809 MI->mayLoad() ||
810 TII->isAsCheapAsAMove(MI: *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(V: MO.getReg());
821 }
822 }
823
824 return true;
825}
826
827bool MachineCSE::ProcessBlockPRE(MachineDominatorTree *DT,
828 MachineBasicBlock *MBB) {
829 bool Changed = false;
830 for (MachineInstr &MI : llvm::make_early_inc_range(Range&: *MBB)) {
831 SmallSet<MCRegister, 8> PhysRefs;
832 if (!isPRECandidate(MI: &MI, PhysRefs))
833 continue;
834
835 if (!PREMap.count(Val: &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(A: MBB, B: MBB1);
845 if (!CMBB->isLegalToHoistInto())
846 continue;
847
848 if (!isProfitableToHoistInto(CandidateBB: 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(From: BB1, To: BB) ||
857 isPotentiallyReachable(From: BB, To: 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(CSMI: &*(CMBB->getFirstTerminator()), MI: &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(i: 0).getReg();
880 Register NewReg = MRI->cloneVirtualRegister(VReg);
881 if (!isProfitableToCSE(CSReg: NewReg, Reg: VReg, CSBB: CMBB, MI: &MI))
882 continue;
883 MachineInstr &NewMI =
884 TII->duplicate(MBB&: *CMBB, InsertBefore: CMBB->getFirstTerminator(), Orig: 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(i: 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.
908bool MachineCSE::PerformSimplePRE(MachineDominatorTree *DT) {
909 SmallVector<MachineDomTreeNode *, 32> BBs;
910
911 PREMap.clear();
912 bool Changed = false;
913 BBs.push_back(Elt: DT->getRootNode());
914 do {
915 auto Node = BBs.pop_back_val();
916 append_range(C&: BBs, R: Node->children());
917
918 MachineBasicBlock *MBB = Node->getBlock();
919 Changed |= ProcessBlockPRE(DT, MBB);
920
921 } while (!BBs.empty());
922
923 return Changed;
924}
925
926bool 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(MBB: CandidateBB) <=
935 MBFI->getBlockFreq(MBB) + MBFI->getBlockFreq(MBB: MBB1);
936}
937
938bool MachineCSE::runOnMachineFunction(MachineFunction &MF) {
939 if (skipFunction(F: 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(Node: DT->getRootNode());
952 return ChangedPRE || ChangedCSE;
953}
954