LoopStrengthReduce.cpp

/// where the user didn't use a C do-while loop. For example, seemingly
/// well-behaved top-test loops will commonly be lowered like this:
//
///   if (n > 0) {
///     i = 0;
///     do {
///       p[i] = 0.0;
///     } while (++i < n);
///   }
///
/// and then it's possible for subsequent optimization to obscure the if
/// test in such a way that indvars can't find it.
///
/// When indvars can't find the if test in loops like this, it creates a
/// signed-max expression, which allows it to give the loop a canonical
/// induction variable:
///
///   i = 0;
///   smax = n < 1 ? 1 : n;
///   do {
///     p[i] = 0.0;
///   } while (++i != smax);
///
/// Canonical induction variables are necessary because the loop passes
/// are designed around them. The most obvious example of this is the
/// LoopInfo analysis, which doesn't remember trip count values. It
/// expects to be able to rediscover the trip count each time it is
/// needed, and it does this using a simple analyis that only succeeds if
/// the loop has a canonical induction variable.
///
/// However, when it comes time to generate code, the maximum operation
/// can be quite costly, especially if it's inside of an outer loop.
///
/// This function solves this problem by detecting this type of loop and
/// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
/// the instructions for the maximum computation.
///
ICmpInst *LoopStrengthReduce::OptimizeSMax(Loop *L, ICmpInst *Cond,
                                           IVStrideUse* &CondUse) {
  // Check that the loop matches the pattern we're looking for.
  if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
      Cond->getPredicate() != CmpInst::ICMP_NE)
    return Cond;

  SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
  if (!Sel || !Sel->hasOneUse()) return Cond;

  SCEVHandle IterationCount = SE->getIterationCount(L);
  if (isa<SCEVCouldNotCompute>(IterationCount))
    return Cond;
  SCEVHandle One = SE->getIntegerSCEV(1, IterationCount->getType());

  // Adjust for an annoying getIterationCount quirk.
  IterationCount = SE->getAddExpr(IterationCount, One);

  // Check for a max calculation that matches the pattern.
  SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(IterationCount);
  if (!SMax || SMax != SE->getSCEV(Sel)) return Cond;

  SCEVHandle SMaxLHS = SMax->getOperand(0);
  SCEVHandle SMaxRHS = SMax->getOperand(1);
  if (!SMaxLHS || SMaxLHS != One) return Cond;

  // Check the relevant induction variable for conformance to
  // the pattern.
  SCEVHandle IV = SE->getSCEV(Cond->getOperand(0));
  SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
  if (!AR || !AR->isAffine() ||
      AR->getStart() != One ||
      AR->getStepRecurrence(*SE) != One)
    return Cond;

  // Check the right operand of the select, and remember it, as it will
  // be used in the new comparison instruction.
  Value *NewRHS = 0;
  if (SE->getSCEV(Sel->getOperand(1)) == SMaxRHS)
    NewRHS = Sel->getOperand(1);
  else if (SE->getSCEV(Sel->getOperand(2)) == SMaxRHS)
    NewRHS = Sel->getOperand(2);
  if (!NewRHS) return Cond;

  // Ok, everything looks ok to change the condition into an SLT or SGE and
  // delete the max calculation.
  ICmpInst *NewCond =
    new ICmpInst(Cond->getPredicate() == CmpInst::ICMP_NE ?
                   CmpInst::ICMP_SLT :
                   CmpInst::ICMP_SGE,
                 Cond->getOperand(0), NewRHS, "scmp", Cond);

  // Delete the max calculation instructions.
  SE->deleteValueFromRecords(Cond);
  Cond->replaceAllUsesWith(NewCond);
  Cond->eraseFromParent();
  Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
  SE->deleteValueFromRecords(Sel);
  Sel->eraseFromParent();
  if (Cmp->use_empty()) {
    SE->deleteValueFromRecords(Cmp);
    Cmp->eraseFromParent();
  }
  CondUse->User = NewCond;
  return NewCond;
}

/// OptimizeShadowIV - If IV is used in a int-to-float cast
/// inside the loop then try to eliminate the cast opeation.
void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {

  SCEVHandle IterationCount = SE->getIterationCount(L);
  if (isa<SCEVCouldNotCompute>(IterationCount))
    return;

  for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e;
       ++Stride) {
    std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI = 
      IVUsesByStride.find(StrideOrder[Stride]);
    assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
    if (!isa<SCEVConstant>(SI->first))
      continue;

    for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
           E = SI->second.Users.end(); UI != E; /* empty */) {
      std::vector<IVStrideUse>::iterator CandidateUI = UI;
      ++UI;
      Instruction *ShadowUse = CandidateUI->User;
      const Type *DestTy = NULL;

      /* If shadow use is a int->float cast then insert a second IV
         to eliminate this cast.

           for (unsigned i = 0; i < n; ++i) 
             foo((double)i);

         is transformed into

           double d = 0.0;
           for (unsigned i = 0; i < n; ++i, ++d) 
             foo(d);
      */
      if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->User))
        DestTy = UCast->getDestTy();
      else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->User))
        DestTy = SCast->getDestTy();
      if (!DestTy) continue;

      if (TLI) {
        /* If target does not support DestTy natively then do not apply
           this transformation. */
        MVT DVT = TLI->getValueType(DestTy);
        if (!TLI->isTypeLegal(DVT)) continue;
      }

      PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
      if (!PH) continue;
      if (PH->getNumIncomingValues() != 2) continue;

      const Type *SrcTy = PH->getType();
      int Mantissa = DestTy->getFPMantissaWidth();
      if (Mantissa == -1) continue; 
      if ((int)TD->getTypeSizeInBits(SrcTy) > Mantissa)
        continue;

      unsigned Entry, Latch;
      if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
        Entry = 0;
        Latch = 1;
      } else {
        Entry = 1;
        Latch = 0;
      }
        
      ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
      if (!Init) continue;
      ConstantFP *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());

      BinaryOperator *Incr = 
        dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
      if (!Incr) continue;
      if (Incr->getOpcode() != Instruction::Add
          && Incr->getOpcode() != Instruction::Sub)
        continue;

      /* Initialize new IV, double d = 0.0 in above example. */
      ConstantInt *C = NULL;
      if (Incr->getOperand(0) == PH)
        C = dyn_cast<ConstantInt>(Incr->getOperand(1));
      else if (Incr->getOperand(1) == PH)
        C = dyn_cast<ConstantInt>(Incr->getOperand(0));
      else
        continue;

      if (!C) continue;

      /* Add new PHINode. */
      PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);

      /* create new increment. '++d' in above example. */
      ConstantFP *CFP = ConstantFP::get(DestTy, C->getZExtValue());
      BinaryOperator *NewIncr = 
        BinaryOperator::Create(Incr->getOpcode(),
                               NewPH, CFP, "IV.S.next.", Incr);

      NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
      NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));

      /* Remove cast operation */
      SE->deleteValueFromRecords(ShadowUse);
      ShadowUse->replaceAllUsesWith(NewPH);
      ShadowUse->eraseFromParent();
      SI->second.Users.erase(CandidateUI);
      NumShadow++;
      break;
    }
  }
}

// OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
// uses in the loop, look to see if we can eliminate some, in favor of using
// common indvars for the different uses.
void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
  // TODO: implement optzns here.

  OptimizeShadowIV(L);

  // Finally, get the terminating condition for the loop if possible.  If we
  // can, we want to change it to use a post-incremented version of its
  // induction variable, to allow coalescing the live ranges for the IV into
  // one register value.
  PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
  BasicBlock  *Preheader = L->getLoopPreheader();
  BasicBlock *LatchBlock =
   SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
  BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
  if (!TermBr || TermBr->isUnconditional() || 
      !isa<ICmpInst>(TermBr->getCondition()))
    return;
  ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());

  // Search IVUsesByStride to find Cond's IVUse if there is one.
  IVStrideUse *CondUse = 0;
  const SCEVHandle *CondStride = 0;

  if (!FindIVUserForCond(Cond, CondUse, CondStride))
    return; // setcc doesn't use the IV.

  // If the trip count is computed in terms of an smax (due to ScalarEvolution
  // being unable to find a sufficient guard, for example), change the loop
  // comparison to use SLT instead of NE.
  Cond = OptimizeSMax(L, Cond, CondUse);

  // If possible, change stride and operands of the compare instruction to
  // eliminate one stride.
  Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);

  // It's possible for the setcc instruction to be anywhere in the loop, and
  // possible for it to have multiple users.  If it is not immediately before
  // the latch block branch, move it.
  if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
    if (Cond->hasOneUse()) {   // Condition has a single use, just move it.
      Cond->moveBefore(TermBr);
    } else {
      // Otherwise, clone the terminating condition and insert into the loopend.
      Cond = cast<ICmpInst>(Cond->clone());
      Cond->setName(L->getHeader()->getName() + ".termcond");
      LatchBlock->getInstList().insert(TermBr, Cond);
      
      // Clone the IVUse, as the old use still exists!
      IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
                                         CondUse->OperandValToReplace);
      CondUse = &IVUsesByStride[*CondStride].Users.back();
    }
  }

  // If we get to here, we know that we can transform the setcc instruction to
  // use the post-incremented version of the IV, allowing us to coalesce the
  // live ranges for the IV correctly.
  CondUse->Offset = SE->getMinusSCEV(CondUse->Offset, *CondStride);
  CondUse->isUseOfPostIncrementedValue = true;
  Changed = true;
}

bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {

  LI = &getAnalysis<LoopInfo>();
  DT = &getAnalysis<DominatorTree>();
  SE = &getAnalysis<ScalarEvolution>();
  TD = &getAnalysis<TargetData>();
  UIntPtrTy = TD->getIntPtrType();
  Changed = false;

  // Find all uses of induction variables in this loop, and categorize
  // them by stride.  Start by finding all of the PHI nodes in the header for
  // this loop.  If they are induction variables, inspect their uses.
  SmallPtrSet<Instruction*,16> Processed;   // Don't reprocess instructions.
  for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
    AddUsersIfInteresting(I, L, Processed);

  if (!IVUsesByStride.empty()) {
    // Optimize induction variables.  Some indvar uses can be transformed to use
    // strides that will be needed for other purposes.  A common example of this
    // is the exit test for the loop, which can often be rewritten to use the
    // computation of some other indvar to decide when to terminate the loop.
    OptimizeIndvars(L);

    // FIXME: We can widen subreg IV's here for RISC targets.  e.g. instead of
    // doing computation in byte values, promote to 32-bit values if safe.

    // FIXME: Attempt to reuse values across multiple IV's.  In particular, we
    // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
    // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
    // Need to be careful that IV's are all the same type.  Only works for
    // intptr_t indvars.

    // If we only have one stride, we can more aggressively eliminate some
    // things.
    bool HasOneStride = IVUsesByStride.size() == 1;

#ifndef NDEBUG
    DOUT << "\nLSR on ";
    DEBUG(L->dump());
#endif

    // IVsByStride keeps IVs for one particular loop.
    assert(IVsByStride.empty() && "Stale entries in IVsByStride?");

    // Sort the StrideOrder so we process larger strides first.
    std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());

    // Note: this processes each stride/type pair individually.  All users
    // passed into StrengthReduceStridedIVUsers have the same type AND stride.
    // Also, note that we iterate over IVUsesByStride indirectly by using
    // StrideOrder. This extra layer of indirection makes the ordering of
    // strides deterministic - not dependent on map order.
    for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
      std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI = 
        IVUsesByStride.find(StrideOrder[Stride]);
      assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
      StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
    }
  }

  // We're done analyzing this loop; release all the state we built up for it.
  CastedPointers.clear();
  IVUsesByStride.clear();
  IVsByStride.clear();
  StrideOrder.clear();
  for (unsigned i=0; i<GEPlist.size(); i++)
    SE->deleteValueFromRecords(GEPlist[i]);
  GEPlist.clear();  

  // Clean up after ourselves
  if (!DeadInsts.empty()) {
    DeleteTriviallyDeadInstructions();

    BasicBlock::iterator I = L->getHeader()->begin();
    while (PHINode *PN = dyn_cast<PHINode>(I++)) {
      // At this point, we know that we have killed one or more IV users.
      // It is worth checking to see if the cannonical indvar is also
      // dead, so that we can remove it as well.
      //
      // We can remove a PHI if it is on a cycle in the def-use graph
      // where each node in the cycle has degree one, i.e. only one use,
      // and is an instruction with no side effects.
      //
      // FIXME: this needs to eliminate an induction variable even if it's being
      // compared against some value to decide loop termination.
      if (!PN->hasOneUse())
        continue;
      
      SmallPtrSet<PHINode *, 4> PHIs;
      for (Instruction *J = dyn_cast<Instruction>(*PN->use_begin());
           J && J->hasOneUse() && !J->mayWriteToMemory();
           J = dyn_cast<Instruction>(*J->use_begin())) {
        // If we find the original PHI, we've discovered a cycle.
        if (J == PN) {
          // Break the cycle and mark the PHI for deletion.
          SE->deleteValueFromRecords(PN);
          PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
          DeadInsts.push_back(PN);
          Changed = true;
          break;
        }
        // If we find a PHI more than once, we're on a cycle that
        // won't prove fruitful.
        if (isa<PHINode>(J) && !PHIs.insert(cast<PHINode>(J)))
          break;
      }
    }
    DeleteTriviallyDeadInstructions();
  }
  return Changed;
}