LoopStrengthReduce.cpp

      break;
    }
  }

  // Forgo this transformation if it the increment happens to be
  // unfortunately positioned after the condition, and the condition
  // has multiple uses which prevent it from being moved immediately
  // before the branch. See
  // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
  // for an example of this situation.
  if (!Cond->hasOneUse()) {
    for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
         I != E; ++I)
      if (I == NewCmpLHS)
        return Cond;
  }

  if (NewCmpRHS) {
    // Create a new compare instruction using new stride / iv.
    ICmpInst *OldCond = Cond;
    // Insert new compare instruction.
    Cond = new ICmpInst(Predicate, NewCmpLHS, NewCmpRHS,
                        L->getHeader()->getName() + ".termcond",
                        OldCond);

    // Remove the old compare instruction. The old indvar is probably dead too.
    DeadInsts.push_back(CondUse->getOperandValToReplace());
    OldCond->replaceAllUsesWith(Cond);
    OldCond->eraseFromParent();

    IU->IVUsesByStride[*NewStride]->addUser(NewOffset, Cond, NewCmpLHS);
    CondUse = &IU->IVUsesByStride[*NewStride]->Users.back();
    CondStride = NewStride;
    ++NumEliminated;
    Changed = true;
  }

  return Cond;
}

/// OptimizeMax - Rewrite the loop's terminating condition if it uses
/// a max computation.
///
/// This is a narrow solution to a specific, but acute, problem. For loops
/// like this:
///
///   i = 0;
///   do {
///     p[i] = 0.0;
///   } while (++i < n);
///
/// the trip count isn't just 'n', because 'n' might not be positive. And
/// unfortunately this can come up even for loops 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
/// max expression, which allows it to give the loop a canonical
/// induction variable:
///
///   i = 0;
///   max = n < 1 ? 1 : n;
///   do {
///     p[i] = 0.0;
///   } while (++i != max);
///
/// 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::OptimizeMax(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;

  const SCEV* BackedgeTakenCount = SE->getBackedgeTakenCount(L);
  if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
    return Cond;
  const SCEV* One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());

  // Add one to the backedge-taken count to get the trip count.
  const SCEV* IterationCount = SE->getAddExpr(BackedgeTakenCount, One);

  // Check for a max calculation that matches the pattern.
  if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount))
    return Cond;
  const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount);
  if (Max != SE->getSCEV(Sel)) return Cond;

  // To handle a max with more than two operands, this optimization would
  // require additional checking and setup.
  if (Max->getNumOperands() != 2)
    return Cond;

  const SCEV* MaxLHS = Max->getOperand(0);
  const SCEV* MaxRHS = Max->getOperand(1);
  if (!MaxLHS || MaxLHS != One) return Cond;

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

  assert(AR->getLoop() == L &&
         "Loop condition operand is an addrec in a different loop!");

  // 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)) == MaxRHS)
    NewRHS = Sel->getOperand(1);
  else if (SE->getSCEV(Sel->getOperand(2)) == MaxRHS)
    NewRHS = Sel->getOperand(2);
  if (!NewRHS) return Cond;

  // Determine the new comparison opcode. It may be signed or unsigned,
  // and the original comparison may be either equality or inequality.
  CmpInst::Predicate Pred =
    isa<SCEVSMaxExpr>(Max) ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
  if (Cond->getPredicate() == CmpInst::ICMP_EQ)
    Pred = CmpInst::getInversePredicate(Pred);

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

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

  const SCEV* BackedgeTakenCount = SE->getBackedgeTakenCount(L);
  if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
    return;

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

    for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
           E = SI->second->Users.end(); UI != E; /* empty */) {
      ilist<IVStrideUse>::iterator CandidateUI = UI;
      ++UI;
      Instruction *ShadowUse = CandidateUI->getUser();
      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->getUser()))
        DestTy = UCast->getDestTy();
      else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
        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)SE->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;
      Constant *NewInit = Context->getConstantFP(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. */
      Constant *CFP = Context->getConstantFP(DestTy, C->getZExtValue());
      BinaryOperator *NewIncr = 
        BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
                                 Instruction::FAdd : Instruction::FSub,
                               NewPH, CFP, "IV.S.next.", Incr);

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

      /* Remove cast operation */
      ShadowUse->replaceAllUsesWith(NewPH);
      ShadowUse->eraseFromParent();
      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);
}

/// OptimizeLoopTermCond - Change loop terminating condition to use the 
/// postinc iv when possible.
void LoopStrengthReduce::OptimizeLoopTermCond(Loop *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.
  BasicBlock *LatchBlock = L->getLoopLatch();
  BasicBlock *ExitingBlock = L->getExitingBlock();
  if (!ExitingBlock)
    // Multiple exits, just look at the exit in the latch block if there is one.
    ExitingBlock = LatchBlock;
  BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
  if (!TermBr)
    return;
  if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
    return;

  // Search IVUsesByStride to find Cond's IVUse if there is one.
  IVStrideUse *CondUse = 0;
  const SCEV* const *CondStride = 0;
  ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
  if (!FindIVUserForCond(Cond, CondUse, CondStride))
    return; // setcc doesn't use the IV.

  if (ExitingBlock != LatchBlock) {
    if (!Cond->hasOneUse())
      // See below, we don't want the condition to be cloned.
      return;

    // If exiting block is the latch block, we know it's safe and profitable to
    // transform the icmp to use post-inc iv. Otherwise do so only if it would
    // not reuse another iv and its iv would be reused by other uses. We are
    // optimizing for the case where the icmp is the only use of the iv.
    IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[*CondStride];
    for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
         E = StrideUses.Users.end(); I != E; ++I) {
      if (I->getUser() == Cond)
        continue;
      if (!I->isUseOfPostIncrementedValue())
        return;
    }

    // FIXME: This is expensive, and worse still ChangeCompareStride does a
    // similar check. Can we perform all the icmp related transformations after
    // StrengthReduceStridedIVUsers?
    if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride)) {
      int64_t SInt = SC->getValue()->getSExtValue();
      for (unsigned NewStride = 0, ee = IU->StrideOrder.size(); NewStride != ee;
           ++NewStride) {
        std::map<const SCEV*, IVUsersOfOneStride *>::iterator SI =
          IU->IVUsesByStride.find(IU->StrideOrder[NewStride]);
        if (!isa<SCEVConstant>(SI->first) || SI->first == *CondStride)
          continue;
        int64_t SSInt =
          cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
        if (SSInt == SInt)
          return; // This can definitely be reused.
        if (unsigned(abs64(SSInt)) < SInt || (SSInt % SInt) != 0)
          continue;
        int64_t Scale = SSInt / SInt;
        bool AllUsesAreAddresses = true;
        bool AllUsesAreOutsideLoop = true;
        std::vector<BasedUser> UsersToProcess;
        const SCEV* CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
                                                AllUsesAreAddresses,
                                                AllUsesAreOutsideLoop,
                                                UsersToProcess);
        // Avoid rewriting the compare instruction with an iv of new stride
        // if it's likely the new stride uses will be rewritten using the
        // stride of the compare instruction.
        if (AllUsesAreAddresses &&
            ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess))
          return;
      }
    }

    StrideNoReuse.insert(*CondStride);
  }

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

  // If possible, change stride and operands of the compare instruction to
  // eliminate one stride.
  if (ExitingBlock == LatchBlock)
    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!
      IU->IVUsesByStride[*CondStride]->addUser(CondUse->getOffset(), Cond,
                                             CondUse->getOperandValToReplace());
      CondUse = &IU->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->setOffset(SE->getMinusSCEV(CondUse->getOffset(), *CondStride));
  CondUse->setIsUseOfPostIncrementedValue(true);
  Changed = true;

  ++NumLoopCond;
}

/// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for deciding
/// when to exit the loop is used only for that purpose, try to rearrange things
/// so it counts down to a test against zero.
void LoopStrengthReduce::OptimizeLoopCountIV(Loop *L) {

  // If the number of times the loop is executed isn't computable, give up.
  const SCEV* BackedgeTakenCount = SE->getBackedgeTakenCount(L);
  if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
    return;

  // Get the terminating condition for the loop if possible (this isn't
  // necessarily in the latch, or a block that's a predecessor of the header).
  if (!L->getExitBlock())
    return; // More than one loop exit blocks.

  // Okay, there is one exit block.  Try to find the condition that causes the
  // loop to be exited.
  BasicBlock *ExitingBlock = L->getExitingBlock();
  if (!ExitingBlock)
    return; // More than one block exiting!

  // Okay, we've computed the exiting block.  See what condition causes us to
  // exit.
  //
  // FIXME: we should be able to handle switch instructions (with a single exit)
  BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
  if (TermBr == 0) return;
  assert(TermBr->isConditional() && "If unconditional, it can't be in loop!");
  if (!isa<ICmpInst>(TermBr->getCondition()))
    return;
  ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());

  // Handle only tests for equality for the moment, and only stride 1.
  if (Cond->getPredicate() != CmpInst::ICMP_EQ)
    return;
  const SCEV* IV = SE->getSCEV(Cond->getOperand(0));
  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
  const SCEV* One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
  if (!AR || !AR->isAffine() || AR->getStepRecurrence(*SE) != One)
    return;
  // If the RHS of the comparison is defined inside the loop, the rewrite
  // cannot be done.
  if (Instruction *CR = dyn_cast<Instruction>(Cond->getOperand(1)))
    if (L->contains(CR->getParent()))
      return;

  // Make sure the IV is only used for counting.  Value may be preinc or
  // postinc; 2 uses in either case.
  if (!Cond->getOperand(0)->hasNUses(2))
    return;
  PHINode *phi = dyn_cast<PHINode>(Cond->getOperand(0));
  Instruction *incr;
  if (phi && phi->getParent()==L->getHeader()) {
    // value tested is preinc.  Find the increment.
    // A CmpInst is not a BinaryOperator; we depend on this.
    Instruction::use_iterator UI = phi->use_begin();
    incr = dyn_cast<BinaryOperator>(UI);
    if (!incr)
      incr = dyn_cast<BinaryOperator>(++UI);
    // 1 use for postinc value, the phi.  Unnecessarily conservative?
    if (!incr || !incr->hasOneUse() || incr->getOpcode()!=Instruction::Add)
      return;
  } else {
    // Value tested is postinc.  Find the phi node.
    incr = dyn_cast<BinaryOperator>(Cond->getOperand(0));
    if (!incr || incr->getOpcode()!=Instruction::Add)
      return;

    Instruction::use_iterator UI = Cond->getOperand(0)->use_begin();
    phi = dyn_cast<PHINode>(UI);
    if (!phi)
      phi = dyn_cast<PHINode>(++UI);
    // 1 use for preinc value, the increment.
    if (!phi || phi->getParent()!=L->getHeader() || !phi->hasOneUse())
      return;
  }

  // Replace the increment with a decrement.
  BinaryOperator *decr = 
    BinaryOperator::Create(Instruction::Sub, incr->getOperand(0),
                           incr->getOperand(1), "tmp", incr);
  incr->replaceAllUsesWith(decr);
  incr->eraseFromParent();

  // Substitute endval-startval for the original startval, and 0 for the
  // original endval.  Since we're only testing for equality this is OK even 
  // if the computation wraps around.
  BasicBlock  *Preheader = L->getLoopPreheader();
  Instruction *PreInsertPt = Preheader->getTerminator();
  int inBlock = L->contains(phi->getIncomingBlock(0)) ? 1 : 0;
  Value *startVal = phi->getIncomingValue(inBlock);
  Value *endVal = Cond->getOperand(1);
  // FIXME check for case where both are constant
  Constant* Zero = Context->getConstantInt(Cond->getOperand(1)->getType(), 0);
  BinaryOperator *NewStartVal = 
    BinaryOperator::Create(Instruction::Sub, endVal, startVal,
                           "tmp", PreInsertPt);
  phi->setIncomingValue(inBlock, NewStartVal);
  Cond->setOperand(1, Zero);

  Changed = true;
}

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

  IU = &getAnalysis<IVUsers>();
  LI = &getAnalysis<LoopInfo>();
  DT = &getAnalysis<DominatorTree>();
  SE = &getAnalysis<ScalarEvolution>();
  Changed = false;

  if (!IU->IVUsesByStride.empty()) {
#ifndef NDEBUG
    DOUT << "\nLSR on \"" << L->getHeader()->getParent()->getNameStart()
         << "\" ";
    DEBUG(L->dump());
#endif

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

    // 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);

    // Change loop terminating condition to use the postinc iv when possible
    // and optimize loop terminating compare. FIXME: Move this after
    // StrengthReduceStridedIVUsers?
    OptimizeLoopTermCond(L);

    // FIXME: We can shrink overlarge IV's here.  e.g. if the code has
    // computation in i64 values and the target doesn't support i64, demote
    // the computation to 32-bit 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.

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

    // 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 = IU->StrideOrder.size();
         Stride != e; ++Stride) {
      std::map<const SCEV*, IVUsersOfOneStride *>::iterator SI =
        IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
      assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
      // FIXME: Generalize to non-affine IV's.
      if (!SI->first->isLoopInvariant(L))
        continue;
      StrengthReduceStridedIVUsers(SI->first, *SI->second, L);
    }
  }

  // After all sharing is done, see if we can adjust the loop to test against
  // zero instead of counting up to a maximum.  This is usually faster.
  OptimizeLoopCountIV(L);

  // We're done analyzing this loop; release all the state we built up for it.
  IVsByStride.clear();
  StrideNoReuse.clear();

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

  // At this point, it is worth checking to see if any recurrence PHIs are also
  // dead, so that we can remove them as well.
  DeleteDeadPHIs(L->getHeader());

  return Changed;
}