LLVM API Documentation

MachineLICM.cpp

Go to the documentation of this file.
00001 //===-- MachineLICM.cpp - Machine Loop Invariant Code Motion Pass ---------===//
00002 //
00003 //                     The LLVM Compiler Infrastructure
00004 //
00005 // This file is distributed under the University of Illinois Open Source
00006 // License. See LICENSE.TXT for details.
00007 //
00008 //===----------------------------------------------------------------------===//
00009 //
00010 // This pass performs loop invariant code motion on machine instructions. We
00011 // attempt to remove as much code from the body of a loop as possible.
00012 //
00013 // This pass does not attempt to throttle itself to limit register pressure.
00014 // The register allocation phases are expected to perform rematerialization
00015 // to recover when register pressure is high.
00016 //
00017 // This pass is not intended to be a replacement or a complete alternative
00018 // for the LLVM-IR-level LICM pass. It is only designed to hoist simple
00019 // constructs that are not exposed before lowering and instruction selection.
00020 //
00021 //===----------------------------------------------------------------------===//
00022 
00023 #define DEBUG_TYPE "machine-licm"
00024 #include "llvm/CodeGen/Passes.h"
00025 #include "llvm/CodeGen/MachineDominators.h"
00026 #include "llvm/CodeGen/MachineLoopInfo.h"
00027 #include "llvm/CodeGen/MachineRegisterInfo.h"
00028 #include "llvm/Target/TargetRegisterInfo.h"
00029 #include "llvm/Target/TargetInstrInfo.h"
00030 #include "llvm/Target/TargetMachine.h"
00031 #include "llvm/ADT/DenseMap.h"
00032 #include "llvm/ADT/Statistic.h"
00033 #include "llvm/Support/Compiler.h"
00034 #include "llvm/Support/Debug.h"
00035 #include "llvm/Support/raw_ostream.h"
00036 
00037 using namespace llvm;
00038 
00039 STATISTIC(NumHoisted, "Number of machine instructions hoisted out of loops");
00040 STATISTIC(NumCSEed,   "Number of hoisted machine instructions CSEed");
00041 
00042 namespace {
00043   class VISIBILITY_HIDDEN MachineLICM : public MachineFunctionPass {
00044     const TargetMachine   *TM;
00045     const TargetInstrInfo *TII;
00046 
00047     // Various analyses that we use...
00048     MachineLoopInfo      *LI;      // Current MachineLoopInfo
00049     MachineDominatorTree *DT;      // Machine dominator tree for the cur loop
00050     MachineRegisterInfo  *RegInfo; // Machine register information
00051 
00052     // State that is updated as we process loops
00053     bool         Changed;          // True if a loop is changed.
00054     MachineLoop *CurLoop;          // The current loop we are working on.
00055     MachineBasicBlock *CurPreheader; // The preheader for CurLoop.
00056 
00057     // For each BB and opcode pair, keep a list of hoisted instructions.
00058     DenseMap<std::pair<unsigned, unsigned>,
00059       std::vector<const MachineInstr*> > CSEMap;
00060   public:
00061     static char ID; // Pass identification, replacement for typeid
00062     MachineLICM() : MachineFunctionPass(&ID) {}
00063 
00064     virtual bool runOnMachineFunction(MachineFunction &MF);
00065 
00066     const char *getPassName() const { return "Machine Instruction LICM"; }
00067 
00068     // FIXME: Loop preheaders?
00069     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
00070       AU.setPreservesCFG();
00071       AU.addRequired<MachineLoopInfo>();
00072       AU.addRequired<MachineDominatorTree>();
00073       AU.addPreserved<MachineLoopInfo>();
00074       AU.addPreserved<MachineDominatorTree>();
00075       MachineFunctionPass::getAnalysisUsage(AU);
00076     }
00077 
00078     virtual void releaseMemory() {
00079       CSEMap.clear();
00080     }
00081 
00082   private:
00083     /// IsLoopInvariantInst - Returns true if the instruction is loop
00084     /// invariant. I.e., all virtual register operands are defined outside of
00085     /// the loop, physical registers aren't accessed (explicitly or implicitly),
00086     /// and the instruction is hoistable.
00087     /// 
00088     bool IsLoopInvariantInst(MachineInstr &I);
00089 
00090     /// IsProfitableToHoist - Return true if it is potentially profitable to
00091     /// hoist the given loop invariant.
00092     bool IsProfitableToHoist(MachineInstr &MI);
00093 
00094     /// HoistRegion - Walk the specified region of the CFG (defined by all
00095     /// blocks dominated by the specified block, and that are in the current
00096     /// loop) in depth first order w.r.t the DominatorTree. This allows us to
00097     /// visit definitions before uses, allowing us to hoist a loop body in one
00098     /// pass without iteration.
00099     ///
00100     void HoistRegion(MachineDomTreeNode *N);
00101 
00102     /// Hoist - When an instruction is found to only use loop invariant operands
00103     /// that is safe to hoist, this instruction is called to do the dirty work.
00104     ///
00105     void Hoist(MachineInstr &MI);
00106   };
00107 } // end anonymous namespace
00108 
00109 char MachineLICM::ID = 0;
00110 static RegisterPass<MachineLICM>
00111 X("machinelicm", "Machine Loop Invariant Code Motion");
00112 
00113 FunctionPass *llvm::createMachineLICMPass() { return new MachineLICM(); }
00114 
00115 /// LoopIsOuterMostWithPreheader - Test if the given loop is the outer-most
00116 /// loop that has a preheader.
00117 static bool LoopIsOuterMostWithPreheader(MachineLoop *CurLoop) {
00118   for (MachineLoop *L = CurLoop->getParentLoop(); L; L = L->getParentLoop())
00119     if (L->getLoopPreheader())
00120       return false;
00121   return true;
00122 }
00123 
00124 /// Hoist expressions out of the specified loop. Note, alias info for inner loop
00125 /// is not preserved so it is not a good idea to run LICM multiple times on one
00126 /// loop.
00127 ///
00128 bool MachineLICM::runOnMachineFunction(MachineFunction &MF) {
00129   const Function *F = MF.getFunction();
00130   if (F->hasFnAttr(Attribute::OptimizeForSize))
00131     return false;
00132 
00133   DOUT << "******** Machine LICM ********\n";
00134 
00135   Changed = false;
00136   TM = &MF.getTarget();
00137   TII = TM->getInstrInfo();
00138   RegInfo = &MF.getRegInfo();
00139 
00140   // Get our Loop information...
00141   LI = &getAnalysis<MachineLoopInfo>();
00142   DT = &getAnalysis<MachineDominatorTree>();
00143 
00144   for (MachineLoopInfo::iterator
00145          I = LI->begin(), E = LI->end(); I != E; ++I) {
00146     CurLoop = *I;
00147 
00148     // Only visit outer-most preheader-sporting loops.
00149     if (!LoopIsOuterMostWithPreheader(CurLoop))
00150       continue;
00151 
00152     // Determine the block to which to hoist instructions. If we can't find a
00153     // suitable loop preheader, we can't do any hoisting.
00154     //
00155     // FIXME: We are only hoisting if the basic block coming into this loop
00156     // has only one successor. This isn't the case in general because we haven't
00157     // broken critical edges or added preheaders.
00158     CurPreheader = CurLoop->getLoopPreheader();
00159     if (!CurPreheader)
00160       continue;
00161 
00162     HoistRegion(DT->getNode(CurLoop->getHeader()));
00163   }
00164 
00165   return Changed;
00166 }
00167 
00168 /// HoistRegion - Walk the specified region of the CFG (defined by all blocks
00169 /// dominated by the specified block, and that are in the current loop) in depth
00170 /// first order w.r.t the DominatorTree. This allows us to visit definitions
00171 /// before uses, allowing us to hoist a loop body in one pass without iteration.
00172 ///
00173 void MachineLICM::HoistRegion(MachineDomTreeNode *N) {
00174   assert(N != 0 && "Null dominator tree node?");
00175   MachineBasicBlock *BB = N->getBlock();
00176 
00177   // If this subregion is not in the top level loop at all, exit.
00178   if (!CurLoop->contains(BB)) return;
00179 
00180   for (MachineBasicBlock::iterator
00181          MII = BB->begin(), E = BB->end(); MII != E; ) {
00182     MachineBasicBlock::iterator NextMII = MII; ++NextMII;
00183     MachineInstr &MI = *MII;
00184 
00185     Hoist(MI);
00186 
00187     MII = NextMII;
00188   }
00189 
00190   const std::vector<MachineDomTreeNode*> &Children = N->getChildren();
00191 
00192   for (unsigned I = 0, E = Children.size(); I != E; ++I)
00193     HoistRegion(Children[I]);
00194 }
00195 
00196 /// IsLoopInvariantInst - Returns true if the instruction is loop
00197 /// invariant. I.e., all virtual register operands are defined outside of the
00198 /// loop, physical registers aren't accessed explicitly, and there are no side
00199 /// effects that aren't captured by the operands or other flags.
00200 /// 
00201 bool MachineLICM::IsLoopInvariantInst(MachineInstr &I) {
00202   const TargetInstrDesc &TID = I.getDesc();
00203   
00204   // Ignore stuff that we obviously can't hoist.
00205   if (TID.mayStore() || TID.isCall() || TID.isTerminator() ||
00206       TID.hasUnmodeledSideEffects())
00207     return false;
00208 
00209   if (TID.mayLoad()) {
00210     // Okay, this instruction does a load. As a refinement, we allow the target
00211     // to decide whether the loaded value is actually a constant. If so, we can
00212     // actually use it as a load.
00213     if (!TII->isInvariantLoad(&I))
00214       // FIXME: we should be able to sink loads with no other side effects if
00215       // there is nothing that can change memory from here until the end of
00216       // block. This is a trivial form of alias analysis.
00217       return false;
00218   }
00219 
00220   DEBUG({
00221       DOUT << "--- Checking if we can hoist " << I;
00222       if (I.getDesc().getImplicitUses()) {
00223         DOUT << "  * Instruction has implicit uses:\n";
00224 
00225         const TargetRegisterInfo *TRI = TM->getRegisterInfo();
00226         for (const unsigned *ImpUses = I.getDesc().getImplicitUses();
00227              *ImpUses; ++ImpUses)
00228           DOUT << "      -> " << TRI->getName(*ImpUses) << "\n";
00229       }
00230 
00231       if (I.getDesc().getImplicitDefs()) {
00232         DOUT << "  * Instruction has implicit defines:\n";
00233 
00234         const TargetRegisterInfo *TRI = TM->getRegisterInfo();
00235         for (const unsigned *ImpDefs = I.getDesc().getImplicitDefs();
00236              *ImpDefs; ++ImpDefs)
00237           DOUT << "      -> " << TRI->getName(*ImpDefs) << "\n";
00238       }
00239     });
00240 
00241   if (I.getDesc().getImplicitDefs() || I.getDesc().getImplicitUses()) {
00242     DOUT << "Cannot hoist with implicit defines or uses\n";
00243     return false;
00244   }
00245 
00246   // The instruction is loop invariant if all of its operands are.
00247   for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
00248     const MachineOperand &MO = I.getOperand(i);
00249 
00250     if (!MO.isReg())
00251       continue;
00252 
00253     unsigned Reg = MO.getReg();
00254     if (Reg == 0) continue;
00255 
00256     // Don't hoist an instruction that uses or defines a physical register.
00257     if (TargetRegisterInfo::isPhysicalRegister(Reg))
00258       return false;
00259 
00260     if (!MO.isUse())
00261       continue;
00262 
00263     assert(RegInfo->getVRegDef(Reg) &&
00264            "Machine instr not mapped for this vreg?!");
00265 
00266     // If the loop contains the definition of an operand, then the instruction
00267     // isn't loop invariant.
00268     if (CurLoop->contains(RegInfo->getVRegDef(Reg)->getParent()))
00269       return false;
00270   }
00271 
00272   // If we got this far, the instruction is loop invariant!
00273   return true;
00274 }
00275 
00276 
00277 /// HasPHIUses - Return true if the specified register has any PHI use.
00278 static bool HasPHIUses(unsigned Reg, MachineRegisterInfo *RegInfo) {
00279   for (MachineRegisterInfo::use_iterator UI = RegInfo->use_begin(Reg),
00280          UE = RegInfo->use_end(); UI != UE; ++UI) {
00281     MachineInstr *UseMI = &*UI;
00282     if (UseMI->getOpcode() == TargetInstrInfo::PHI)
00283       return true;
00284   }
00285   return false;
00286 }
00287 
00288 /// IsProfitableToHoist - Return true if it is potentially profitable to hoist
00289 /// the given loop invariant.
00290 bool MachineLICM::IsProfitableToHoist(MachineInstr &MI) {
00291   if (MI.getOpcode() == TargetInstrInfo::IMPLICIT_DEF)
00292     return false;
00293 
00294   const TargetInstrDesc &TID = MI.getDesc();
00295 
00296   // FIXME: For now, only hoist re-materilizable instructions. LICM will
00297   // increase register pressure. We want to make sure it doesn't increase
00298   // spilling.
00299   if (!TID.mayLoad() && (!TID.isRematerializable() ||
00300                          !TII->isTriviallyReMaterializable(&MI)))
00301     return false;
00302 
00303   // If result(s) of this instruction is used by PHIs, then don't hoist it.
00304   // The presence of joins makes it difficult for current register allocator
00305   // implementation to perform remat.
00306   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
00307     const MachineOperand &MO = MI.getOperand(i);
00308     if (!MO.isReg() || !MO.isDef())
00309       continue;
00310     if (HasPHIUses(MO.getReg(), RegInfo))
00311       return false;
00312   }
00313 
00314   return true;
00315 }
00316 
00317 static const MachineInstr *LookForDuplicate(const MachineInstr *MI,
00318                                       std::vector<const MachineInstr*> &PrevMIs,
00319                                       MachineRegisterInfo *RegInfo) {
00320   unsigned NumOps = MI->getNumOperands();
00321   for (unsigned i = 0, e = PrevMIs.size(); i != e; ++i) {
00322     const MachineInstr *PrevMI = PrevMIs[i];
00323     unsigned NumOps2 = PrevMI->getNumOperands();
00324     if (NumOps != NumOps2)
00325       continue;
00326     bool IsSame = true;
00327     for (unsigned j = 0; j != NumOps; ++j) {
00328       const MachineOperand &MO = MI->getOperand(j);
00329       if (MO.isReg() && MO.isDef()) {
00330         if (RegInfo->getRegClass(MO.getReg()) !=
00331             RegInfo->getRegClass(PrevMI->getOperand(j).getReg())) {
00332           IsSame = false;
00333           break;
00334         }
00335         continue;
00336       }
00337       if (!MO.isIdenticalTo(PrevMI->getOperand(j))) {
00338         IsSame = false;
00339         break;
00340       }
00341     }
00342     if (IsSame)
00343       return PrevMI;
00344   }
00345   return 0;
00346 }
00347 
00348 /// Hoist - When an instruction is found to use only loop invariant operands
00349 /// that are safe to hoist, this instruction is called to do the dirty work.
00350 ///
00351 void MachineLICM::Hoist(MachineInstr &MI) {
00352   if (!IsLoopInvariantInst(MI)) return;
00353   if (!IsProfitableToHoist(MI)) return;
00354 
00355   // Now move the instructions to the predecessor, inserting it before any
00356   // terminator instructions.
00357   DEBUG({
00358       errs() << "Hoisting " << MI;
00359       if (CurPreheader->getBasicBlock())
00360         errs() << " to MachineBasicBlock "
00361                << CurPreheader->getBasicBlock()->getName();
00362       if (MI.getParent()->getBasicBlock())
00363         errs() << " from MachineBasicBlock "
00364                << MI.getParent()->getBasicBlock()->getName();
00365       errs() << "\n";
00366     });
00367 
00368   // Look for opportunity to CSE the hoisted instruction.
00369   std::pair<unsigned, unsigned> BBOpcPair =
00370     std::make_pair(CurPreheader->getNumber(), MI.getOpcode());
00371   DenseMap<std::pair<unsigned, unsigned>,
00372     std::vector<const MachineInstr*> >::iterator CI = CSEMap.find(BBOpcPair);
00373   bool DoneCSE = false;
00374   if (CI != CSEMap.end()) {
00375     const MachineInstr *Dup = LookForDuplicate(&MI, CI->second, RegInfo);
00376     if (Dup) {
00377       DOUT << "CSEing " << MI;
00378       DOUT << " with " << *Dup;
00379       for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
00380         const MachineOperand &MO = MI.getOperand(i);
00381         if (MO.isReg() && MO.isDef())
00382           RegInfo->replaceRegWith(MO.getReg(), Dup->getOperand(i).getReg());
00383       }
00384       MI.eraseFromParent();
00385       DoneCSE = true;
00386       ++NumCSEed;
00387     }
00388   }
00389 
00390   // Otherwise, splice the instruction to the preheader.
00391   if (!DoneCSE) {
00392     CurPreheader->splice(CurPreheader->getFirstTerminator(),
00393                          MI.getParent(), &MI);
00394     // Add to the CSE map.
00395     if (CI != CSEMap.end())
00396       CI->second.push_back(&MI);
00397     else {
00398       std::vector<const MachineInstr*> CSEMIs;
00399       CSEMIs.push_back(&MI);
00400       CSEMap.insert(std::make_pair(BBOpcPair, CSEMIs));
00401     }
00402   }
00403 
00404   ++NumHoisted;
00405   Changed = true;
00406 }