Re: [PATCH] Don't call double rint from float powf.

[Craig Howland <howland at LGSInnovations dot com>]

On 12/12/2017 01:00 PM, Jim Wilson wrote:
> On 12/11/2017 04:37 PM, Craig Howland wrote:
> > On 12/11/2017 07:08 PM, Jim Wilson wrote:
> > > - if(x<0.0&&rint(y)!=y) exc.retval = -HUGE_VAL;
> > > +               if(x<0.0&&rintf(y)!=y) exc.retval = -HUGE_VAL;
> > >               }
> > >               if (_LIB_VERSION == _POSIX_)
> > >                   errno = ERANGE;
> > To be most rigorous, "0.0" on the same lines as the rintf() should be "0.0F" 
> > (as otherwise the comparison strictly should be done in double).  (Not 
> > addressing the exact change you are making, but is the same class of fix and 
> > are on the same line of source code.)
>
> Any reasonable compiler will get this right.  GCC does a float compare here 
> even if you compile without optimization.  So no fix here is required, though 
> I can redo the patch if you want.
If you don't mind, it would be best that way, as it keeps the source compliant 
with the C standard in terms of requiring a float comparison.  But it also could 
be cleaned up later on, as I expect that there are some others that are the same 
way that could also use it.

(It is an interesting thing to consider to decide if GCC is doing the right 
thing here, or not.  While it knows that it is comparing against 0 and that 
converting it from double 0 to float 0 has no runtime side effects, the user 
could have purposely done it this way because they wanted to avoid a float 
comparison.  I do agree that making it a float comparison is a good choice, just 
that it could potentially be considered to be not in compliance with the standard.)