[llvm-commits] [PATCH] Rotate constant folding bug

[John McCall]

On Oct 11, 2011, at 6:06 PM, Cameron McInally wrote:
> Constant folding for the rotate operators does not work currently. Here's a fix...
> 
> royale mcinally/llvm> svn diff
> Index: lib/Support/APInt.cpp
> ===================================================================
> --- lib/Support/APInt.cpp	(revision 141748)
> +++ lib/Support/APInt.cpp	(working copy)
> @@ -1334,8 +1334,8 @@
>    // Don't get too fancy, just use existing shift/or facilities
>    APInt hi(*this);
>    APInt lo(*this);
> -  hi.shl(rotateAmt);
> -  lo.lshr(BitWidth - rotateAmt);
> +  hi = hi.shl(rotateAmt);
> +  lo = lo.lshr(BitWidth - rotateAmt);
>    return hi | lo;
>  }

Nice catch!  One tweak:  the old code was written assuming that the shift
operations are destructive, so it copies 'this' twice.  Since the shift operations
aren't destructive, those copies are completely unnecessary;  please optimize
them out, like so:
  APInt hi = shl(rotateAmt);
  APInt lo = lshr(BitWidth - rotateAmt);

John.