[llvm] r279568 - [InstCombine] remove icmp shr folds that are already handled by InstSimplify

[Friedman, Eli via llvm-commits]

On 8/23/2016 2:01 PM, Sanjay Patel via llvm-commits wrote:
> Author: spatel
> Date: Tue Aug 23 16:01:35 2016
> New Revision: 279568
>
> -  // If we are comparing against bits always shifted out, the
> -  // comparison cannot succeed.
> -  APInt Comp = CmpRHSV << ShAmtVal;
> -  ConstantInt *ShiftedCmpRHS = Builder->getInt(Comp);
> -  if (Shr->getOpcode() == Instruction::LShr)
> -    Comp = Comp.lshr(ShAmtVal);
> -  else
> -    Comp = Comp.ashr(ShAmtVal);
> -
> -  if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
> -    bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
> -    Constant *Cst = Builder->getInt1(IsICMP_NE);
> -    return replaceInstUsesWith(ICI, Cst);
> -  }
> -
> -  // Otherwise, check to see if the bits shifted out are known to be zero.
> -  // If so, we can compare against the unshifted value:
> +  // Check if the bits shifted out are known to be zero. If so, we can compare
> +  // against the unshifted value:

It's not obvious that it's actually safe to do this... I mean, you don't 
need the same optimization in both InstSimplify and Instcombine, but 
"Comp != CmpRHSV" also acts as a safety check here.  At the very least, 
there needs to be a comment explaining why this is safe.

-Eli

>     //  (X & 4) >> 1 == 2  --> (X & 4) == 4.
> +  ConstantInt *ShiftedCmpRHS = Builder->getInt(CmpRHSV << ShAmtVal);
>     if (Shr->hasOneUse() && Shr->isExact())
>       return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
>   
>
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits


-- 
Employee of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project