On Thu, Oct 25, 2012 at 01:54:58AM +0200, Andreas Schwab wrote:
> You are still allowing the bug to happen.

Hmmm, yes, thank you. I was trying to handle the case where sizeof
long double is less than 8, but obviously, that's not the right
approach. I have applied the following patch that just removes "inf"
from the list of valid outputs. We'll make the output target-specific
if we find that we have to.

Tested on x86_64-linux.  Checked in.

-- 
Joel