lib/CodeGen/MachineCSE.cpp

   1 //===-- MachineCSE.cpp - Machine Common Subexpression Elimination Pass ----===//
   2 //
   3 //                     The LLVM Compiler Infrastructure
   4 //
   5 // This file is distributed under the University of Illinois Open Source
   6 // License. See LICENSE.TXT for details.
   7 //
   8 //===----------------------------------------------------------------------===//
   9 //
  10 // This pass performs global common subexpression elimination on machine
  11 // instructions using a scoped hash table based value numbering scheme. It
  12 // must be run while the machine function is still in SSA form.
  13 //
  14 //===----------------------------------------------------------------------===//
  15
  16 #define DEBUG_TYPE "machine-cse"
  17 #include "llvm/CodeGen/Passes.h"
  18 #include "llvm/CodeGen/MachineDominators.h"
  19 #include "llvm/CodeGen/MachineInstr.h"
  20 #include "llvm/CodeGen/MachineRegisterInfo.h"
  21 #include "llvm/Analysis/AliasAnalysis.h"
  22 #include "llvm/Target/TargetInstrInfo.h"
  23 #include "llvm/ADT/DenseMap.h"
  24 #include "llvm/ADT/ScopedHashTable.h"
  25 #include "llvm/ADT/SmallSet.h"
  26 #include "llvm/ADT/Statistic.h"
  27 #include "llvm/Support/Debug.h"
  28 #include "llvm/Support/RecyclingAllocator.h"
  29
  30 using namespace llvm;
  31
  32 STATISTIC(NumCoalesces, "Number of copies coalesced");
  33 STATISTIC(NumCSEs,      "Number of common subexpression eliminated");
  34 STATISTIC(NumPhysCSEs,
  35           "Number of physreg referencing common subexpr eliminated");
  36 STATISTIC(NumCommutes,  "Number of copies coalesced after commuting");
  37
  38 namespace {
  39   class MachineCSE : public MachineFunctionPass {
  40     const TargetInstrInfo *TII;
  41     const TargetRegisterInfo *TRI;
  42     AliasAnalysis *AA;
  43     MachineDominatorTree *DT;
  44     MachineRegisterInfo *MRI;
  45   public:
  46     static char ID; // Pass identification
  47     MachineCSE() : MachineFunctionPass(ID), LookAheadLimit(5), CurrVN(0) {
  48       initializeMachineCSEPass(*PassRegistry::getPassRegistry());
  49     }
  50
  51     virtual bool runOnMachineFunction(MachineFunction &MF);
  52
  53     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
  54       AU.setPreservesCFG();
  55       MachineFunctionPass::getAnalysisUsage(AU);
  56       AU.addRequired<AliasAnalysis>();
  57       AU.addPreservedID(MachineLoopInfoID);
  58       AU.addRequired<MachineDominatorTree>();
  59       AU.addPreserved<MachineDominatorTree>();
  60     }
  61
  62     virtual void releaseMemory() {
  63       ScopeMap.clear();
  64       Exps.clear();
  65     }
  66
  67   private:
  68     const unsigned LookAheadLimit;
  69     typedef RecyclingAllocator<BumpPtrAllocator,
  70         ScopedHashTableVal<MachineInstr*, unsigned> > AllocatorTy;
  71     typedef ScopedHashTable<MachineInstr*, unsigned,
  72         MachineInstrExpressionTrait, AllocatorTy> ScopedHTType;
  73     typedef ScopedHTType::ScopeTy ScopeType;
  74     DenseMap<MachineBasicBlock*, ScopeType*> ScopeMap;
  75     ScopedHTType VNT;
  76     SmallVector<MachineInstr*, 64> Exps;
  77     unsigned CurrVN;
  78
  79     bool PerformTrivialCoalescing(MachineInstr *MI, MachineBasicBlock *MBB);
  80     bool isPhysDefTriviallyDead(unsigned Reg,
  81                                 MachineBasicBlock::const_iterator I,
  82                                 MachineBasicBlock::const_iterator E) const ;
  83     bool hasLivePhysRegDefUses(const MachineInstr *MI,
  84                                const MachineBasicBlock *MBB,
  85                                SmallSet<unsigned,8> &PhysRefs) const;
  86     bool PhysRegDefsReach(MachineInstr *CSMI, MachineInstr *MI,
  87                           SmallSet<unsigned,8> &PhysRefs) const;
  88     bool isCSECandidate(MachineInstr *MI);
  89     bool isProfitableToCSE(unsigned CSReg, unsigned Reg,
  90                            MachineInstr *CSMI, MachineInstr *MI);
  91     void EnterScope(MachineBasicBlock *MBB);
  92     void ExitScope(MachineBasicBlock *MBB);
  93     bool ProcessBlock(MachineBasicBlock *MBB);
  94     void ExitScopeIfDone(MachineDomTreeNode *Node,
  95                  DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren,
  96                  DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> &ParentMap);
  97     bool PerformCSE(MachineDomTreeNode *Node);
  98   };
  99 } // end anonymous namespace
 100
 101 char MachineCSE::ID = 0;
 102 INITIALIZE_PASS_BEGIN(MachineCSE, "machine-cse",
 103                 "Machine Common Subexpression Elimination", false, false)
 104 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
 105 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
 106 INITIALIZE_PASS_END(MachineCSE, "machine-cse",
 107                 "Machine Common Subexpression Elimination", false, false)
 108
 109 FunctionPass *llvm::createMachineCSEPass() { return new MachineCSE(); }
 110
 111 bool MachineCSE::PerformTrivialCoalescing(MachineInstr *MI,
 112                                           MachineBasicBlock *MBB) {
 113   bool Changed = false;
 114   for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
 115     MachineOperand &MO = MI->getOperand(i);
 116     if (!MO.isReg() || !MO.isUse())
 117       continue;
 118     unsigned Reg = MO.getReg();
 119     if (!TargetRegisterInfo::isVirtualRegister(Reg))
 120       continue;
 121     if (!MRI->hasOneNonDBGUse(Reg))
 122       // Only coalesce single use copies. This ensure the copy will be
 123       // deleted.
 124       continue;
 125     MachineInstr *DefMI = MRI->getVRegDef(Reg);
 126     if (DefMI->getParent() != MBB)
 127       continue;
 128     if (!DefMI->isCopy())
 129       continue;
 130     unsigned SrcReg = DefMI->getOperand(1).getReg();
 131     if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
 132       continue;
 133     if (DefMI->getOperand(0).getSubReg() || DefMI->getOperand(1).getSubReg())
 134       continue;
 135     if (!MRI->constrainRegClass(SrcReg, MRI->getRegClass(Reg)))
 136       continue;
 137     DEBUG(dbgs() << "Coalescing: " << *DefMI);
 138     DEBUG(dbgs() << "***     to: " << *MI);
 139     MO.setReg(SrcReg);
 140     MRI->clearKillFlags(SrcReg);
 141     DefMI->eraseFromParent();
 142     ++NumCoalesces;
 143     Changed = true;
 144   }
 145
 146   return Changed;
 147 }
 148
 149 bool
 150 MachineCSE::isPhysDefTriviallyDead(unsigned Reg,
 151                                    MachineBasicBlock::const_iterator I,
 152                                    MachineBasicBlock::const_iterator E) const {
 153   unsigned LookAheadLeft = LookAheadLimit;
 154   while (LookAheadLeft) {
 155     // Skip over dbg_value's.
 156     while (I != E && I->isDebugValue())
 157       ++I;
 158
 159     if (I == E)
 160       // Reached end of block, register is obviously dead.
 161       return true;
 162
 163     bool SeenDef = false;
 164     for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
 165       const MachineOperand &MO = I->getOperand(i);
 166       if (!MO.isReg() || !MO.getReg())
 167         continue;
 168       if (!TRI->regsOverlap(MO.getReg(), Reg))
 169         continue;
 170       if (MO.isUse())
 171         // Found a use!
 172         return false;
 173       SeenDef = true;
 174     }
 175     if (SeenDef)
 176       // See a def of Reg (or an alias) before encountering any use, it's
 177       // trivially dead.
 178       return true;
 179
 180     --LookAheadLeft;
 181     ++I;
 182   }
 183   return false;
 184 }
 185
 186 /// hasLivePhysRegDefUses - Return true if the specified instruction read/write
 187 /// physical registers (except for dead defs of physical registers). It also
 188 /// returns the physical register def by reference if it's the only one and the
 189 /// instruction does not uses a physical register.
 190 bool MachineCSE::hasLivePhysRegDefUses(const MachineInstr *MI,
 191                                        const MachineBasicBlock *MBB,
 192                                        SmallSet<unsigned,8> &PhysRefs) const {
 193   MachineBasicBlock::const_iterator I = MI; I = llvm::next(I);
 194   for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
 195     const MachineOperand &MO = MI->getOperand(i);
 196     if (!MO.isReg())
 197       continue;
 198     unsigned Reg = MO.getReg();
 199     if (!Reg)
 200       continue;
 201     if (TargetRegisterInfo::isVirtualRegister(Reg))
 202       continue;
 203     // If the def is dead, it's ok. But the def may not marked "dead". That's
 204     // common since this pass is run before livevariables. We can scan
 205     // forward a few instructions and check if it is obviously dead.
 206     if (MO.isDef() &&
 207         (MO.isDead() || isPhysDefTriviallyDead(Reg, I, MBB->end())))
 208       continue;
 209     PhysRefs.insert(Reg);
 210     for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias)
 211       PhysRefs.insert(*Alias);
 212   }
 213
 214   return !PhysRefs.empty();
 215 }
 216
 217 bool MachineCSE::PhysRegDefsReach(MachineInstr *CSMI, MachineInstr *MI,
 218                                   SmallSet<unsigned,8> &PhysRefs) const {
 219   // For now conservatively returns false if the common subexpression is
 220   // not in the same basic block as the given instruction.
 221   MachineBasicBlock *MBB = MI->getParent();
 222   if (CSMI->getParent() != MBB)
 223     return false;
 224   MachineBasicBlock::const_iterator I = CSMI; I = llvm::next(I);
 225   MachineBasicBlock::const_iterator E = MI;
 226   unsigned LookAheadLeft = LookAheadLimit;
 227   while (LookAheadLeft) {
 228     // Skip over dbg_value's.
 229     while (I != E && I->isDebugValue())
 230       ++I;
 231
 232     if (I == E)
 233       return true;
 234
 235     for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
 236       const MachineOperand &MO = I->getOperand(i);
 237       if (!MO.isReg() || !MO.isDef())
 238         continue;
 239       unsigned MOReg = MO.getReg();
 240       if (TargetRegisterInfo::isVirtualRegister(MOReg))
 241         continue;
 242       if (PhysRefs.count(MOReg))
 243         return false;
 244     }
 245
 246     --LookAheadLeft;
 247     ++I;
 248   }
 249
 250   return false;
 251 }
 252
 253 bool MachineCSE::isCSECandidate(MachineInstr *MI) {
 254   if (MI->isLabel() || MI->isPHI() || MI->isImplicitDef() ||
 255       MI->isKill() || MI->isInlineAsm() || MI->isDebugValue())
 256     return false;
 257
 258   // Ignore copies.
 259   if (MI->isCopyLike())
 260     return false;
 261
 262   // Ignore stuff that we obviously can't move.
 263   const MCInstrDesc &MCID = MI->getDesc();
 264   if (MCID.mayStore() || MCID.isCall() || MCID.isTerminator() ||
 265       MI->hasUnmodeledSideEffects())
 266     return false;
 267
 268   if (MCID.mayLoad()) {
 269     // Okay, this instruction does a load. As a refinement, we allow the target
 270     // to decide whether the loaded value is actually a constant. If so, we can
 271     // actually use it as a load.
 272     if (!MI->isInvariantLoad(AA))
 273       // FIXME: we should be able to hoist loads with no other side effects if
 274       // there are no other instructions which can change memory in this loop.
 275       // This is a trivial form of alias analysis.
 276       return false;
 277   }
 278   return true;
 279 }
 280
 281 /// isProfitableToCSE - Return true if it's profitable to eliminate MI with a
 282 /// common expression that defines Reg.
 283 bool MachineCSE::isProfitableToCSE(unsigned CSReg, unsigned Reg,
 284                                    MachineInstr *CSMI, MachineInstr *MI) {
 285   // FIXME: Heuristics that works around the lack the live range splitting.
 286
 287   // Heuristics #1: Don't CSE "cheap" computation if the def is not local or in
 288   // an immediate predecessor. We don't want to increase register pressure and
 289   // end up causing other computation to be spilled.
 290   if (MI->getDesc().isAsCheapAsAMove()) {
 291     MachineBasicBlock *CSBB = CSMI->getParent();
 292     MachineBasicBlock *BB = MI->getParent();
 293     if (CSBB != BB && !CSBB->isSuccessor(BB))
 294       return false;
 295   }
 296
 297   // Heuristics #2: If the expression doesn't not use a vr and the only use
 298   // of the redundant computation are copies, do not cse.
 299   bool HasVRegUse = false;
 300   for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
 301     const MachineOperand &MO = MI->getOperand(i);
 302     if (MO.isReg() && MO.isUse() &&
 303         TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
 304       HasVRegUse = true;
 305       break;
 306     }
 307   }
 308   if (!HasVRegUse) {
 309     bool HasNonCopyUse = false;
 310     for (MachineRegisterInfo::use_nodbg_iterator I =  MRI->use_nodbg_begin(Reg),
 311            E = MRI->use_nodbg_end(); I != E; ++I) {
 312       MachineInstr *Use = &*I;
 313       // Ignore copies.
 314       if (!Use->isCopyLike()) {
 315         HasNonCopyUse = true;
 316         break;
 317       }
 318     }
 319     if (!HasNonCopyUse)
 320       return false;
 321   }
 322
 323   // Heuristics #3: If the common subexpression is used by PHIs, do not reuse
 324   // it unless the defined value is already used in the BB of the new use.
 325   bool HasPHI = false;
 326   SmallPtrSet<MachineBasicBlock*, 4> CSBBs;
 327   for (MachineRegisterInfo::use_nodbg_iterator I =  MRI->use_nodbg_begin(CSReg),
 328        E = MRI->use_nodbg_end(); I != E; ++I) {
 329     MachineInstr *Use = &*I;
 330     HasPHI |= Use->isPHI();
 331     CSBBs.insert(Use->getParent());
 332   }
 333
 334   if (!HasPHI)
 335     return true;
 336   return CSBBs.count(MI->getParent());
 337 }
 338
 339 void MachineCSE::EnterScope(MachineBasicBlock *MBB) {
 340   DEBUG(dbgs() << "Entering: " << MBB->getName() << '\n');
 341   ScopeType *Scope = new ScopeType(VNT);
 342   ScopeMap[MBB] = Scope;
 343 }
 344
 345 void MachineCSE::ExitScope(MachineBasicBlock *MBB) {
 346   DEBUG(dbgs() << "Exiting: " << MBB->getName() << '\n');
 347   DenseMap<MachineBasicBlock*, ScopeType*>::iterator SI = ScopeMap.find(MBB);
 348   assert(SI != ScopeMap.end());
 349   ScopeMap.erase(SI);
 350   delete SI->second;
 351 }
 352
 353 bool MachineCSE::ProcessBlock(MachineBasicBlock *MBB) {
 354   bool Changed = false;
 355
 356   SmallVector<std::pair<unsigned, unsigned>, 8> CSEPairs;
 357   for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E; ) {
 358     MachineInstr *MI = &*I;
 359     ++I;
 360
 361     if (!isCSECandidate(MI))
 362       continue;
 363
 364     bool FoundCSE = VNT.count(MI);
 365     if (!FoundCSE) {
 366       // Look for trivial copy coalescing opportunities.
 367       if (PerformTrivialCoalescing(MI, MBB)) {
 368         Changed = true;
 369
 370         // After coalescing MI itself may become a copy.
 371         if (MI->isCopyLike())
 372           continue;
 373         FoundCSE = VNT.count(MI);
 374       }
 375     }
 376
 377     // Commute commutable instructions.
 378     bool Commuted = false;
 379     if (!FoundCSE && MI->getDesc().isCommutable()) {
 380       MachineInstr *NewMI = TII->commuteInstruction(MI);
 381       if (NewMI) {
 382         Commuted = true;
 383         FoundCSE = VNT.count(NewMI);
 384         if (NewMI != MI) {
 385           // New instruction. It doesn't need to be kept.
 386           NewMI->eraseFromParent();
 387           Changed = true;
 388         } else if (!FoundCSE)
 389           // MI was changed but it didn't help, commute it back!
 390           (void)TII->commuteInstruction(MI);
 391       }
 392     }
 393
 394     // If the instruction defines physical registers and the values *may* be
 395     // used, then it's not safe to replace it with a common subexpression.
 396     // It's also not safe if the instruction uses physical registers.
 397     SmallSet<unsigned,8> PhysRefs;
 398     if (FoundCSE && hasLivePhysRegDefUses(MI, MBB, PhysRefs)) {
 399       FoundCSE = false;
 400
 401       // ... Unless the CS is local and it also defines the physical register
 402       // which is not clobbered in between and the physical register uses
 403       // were not clobbered.
 404       unsigned CSVN = VNT.lookup(MI);
 405       MachineInstr *CSMI = Exps[CSVN];
 406       if (PhysRegDefsReach(CSMI, MI, PhysRefs))
 407         FoundCSE = true;
 408     }
 409
 410     if (!FoundCSE) {
 411       VNT.insert(MI, CurrVN++);
 412       Exps.push_back(MI);
 413       continue;
 414     }
 415
 416     // Found a common subexpression, eliminate it.
 417     unsigned CSVN = VNT.lookup(MI);
 418     MachineInstr *CSMI = Exps[CSVN];
 419     DEBUG(dbgs() << "Examining: " << *MI);
 420     DEBUG(dbgs() << "*** Found a common subexpression: " << *CSMI);
 421
 422     // Check if it's profitable to perform this CSE.
 423     bool DoCSE = true;
 424     unsigned NumDefs = MI->getDesc().getNumDefs();
 425     for (unsigned i = 0, e = MI->getNumOperands(); NumDefs && i != e; ++i) {
 426       MachineOperand &MO = MI->getOperand(i);
 427       if (!MO.isReg() || !MO.isDef())
 428         continue;
 429       unsigned OldReg = MO.getReg();
 430       unsigned NewReg = CSMI->getOperand(i).getReg();
 431       if (OldReg == NewReg)
 432         continue;
 433       assert(TargetRegisterInfo::isVirtualRegister(OldReg) &&
 434              TargetRegisterInfo::isVirtualRegister(NewReg) &&
 435              "Do not CSE physical register defs!");
 436       if (!isProfitableToCSE(NewReg, OldReg, CSMI, MI)) {
 437         DoCSE = false;
 438         break;
 439       }
 440       CSEPairs.push_back(std::make_pair(OldReg, NewReg));
 441       --NumDefs;
 442     }
 443
 444     // Actually perform the elimination.
 445     if (DoCSE) {
 446       for (unsigned i = 0, e = CSEPairs.size(); i != e; ++i) {
 447         MRI->replaceRegWith(CSEPairs[i].first, CSEPairs[i].second);
 448         MRI->clearKillFlags(CSEPairs[i].second);
 449       }
 450       MI->eraseFromParent();
 451       ++NumCSEs;
 452       if (!PhysRefs.empty())
 453         ++NumPhysCSEs;
 454       if (Commuted)
 455         ++NumCommutes;
 456       Changed = true;
 457     } else {
 458       DEBUG(dbgs() << "*** Not profitable, avoid CSE!\n");
 459       VNT.insert(MI, CurrVN++);
 460       Exps.push_back(MI);
 461     }
 462     CSEPairs.clear();
 463   }
 464
 465   return Changed;
 466 }
 467
 468 /// ExitScopeIfDone - Destroy scope for the MBB that corresponds to the given
 469 /// dominator tree node if its a leaf or all of its children are done. Walk
 470 /// up the dominator tree to destroy ancestors which are now done.
 471 void
 472 MachineCSE::ExitScopeIfDone(MachineDomTreeNode *Node,
 473                 DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren,
 474                 DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> &ParentMap) {
 475   if (OpenChildren[Node])
 476     return;
 477
 478   // Pop scope.
 479   ExitScope(Node->getBlock());
 480
 481   // Now traverse upwards to pop ancestors whose offsprings are all done.
 482   while (MachineDomTreeNode *Parent = ParentMap[Node]) {
 483     unsigned Left = --OpenChildren[Parent];
 484     if (Left != 0)
 485       break;
 486     ExitScope(Parent->getBlock());
 487     Node = Parent;
 488   }
 489 }
 490
 491 bool MachineCSE::PerformCSE(MachineDomTreeNode *Node) {
 492   SmallVector<MachineDomTreeNode*, 32> Scopes;
 493   SmallVector<MachineDomTreeNode*, 8> WorkList;
 494   DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> ParentMap;
 495   DenseMap<MachineDomTreeNode*, unsigned> OpenChildren;
 496
 497   CurrVN = 0;
 498
 499   // Perform a DFS walk to determine the order of visit.
 500   WorkList.push_back(Node);
 501   do {
 502     Node = WorkList.pop_back_val();
 503     Scopes.push_back(Node);
 504     const std::vector<MachineDomTreeNode*> &Children = Node->getChildren();
 505     unsigned NumChildren = Children.size();
 506     OpenChildren[Node] = NumChildren;
 507     for (unsigned i = 0; i != NumChildren; ++i) {
 508       MachineDomTreeNode *Child = Children[i];
 509       ParentMap[Child] = Node;
 510       WorkList.push_back(Child);
 511     }
 512   } while (!WorkList.empty());
 513
 514   // Now perform CSE.
 515   bool Changed = false;
 516   for (unsigned i = 0, e = Scopes.size(); i != e; ++i) {
 517     MachineDomTreeNode *Node = Scopes[i];
 518     MachineBasicBlock *MBB = Node->getBlock();
 519     EnterScope(MBB);
 520     Changed |= ProcessBlock(MBB);
 521     // If it's a leaf node, it's done. Traverse upwards to pop ancestors.
 522     ExitScopeIfDone(Node, OpenChildren, ParentMap);
 523   }
 524
 525   return Changed;
 526 }
 527
 528 bool MachineCSE::runOnMachineFunction(MachineFunction &MF) {
 529   TII = MF.getTarget().getInstrInfo();
 530   TRI = MF.getTarget().getRegisterInfo();
 531   MRI = &MF.getRegInfo();
 532   AA = &getAnalysis<AliasAnalysis>();
 533   DT = &getAnalysis<MachineDominatorTree>();
 534   return PerformCSE(DT->getRootNode());
 535 }