Re: Obvious expand_builtin_mathfn improvement installed

[Jan Hubicka <jh at suse dot cz>]

> 
> Hi Zack,
> > > Could you be a little clearer about this bug, and why your patch
> > > affects it?
> >
> > THere is already Roger's fix waiting in the queue to be reviewed, but
> > apparently Richard is taking a break...
> 
> I believe Jan owes Richard an apology :>  The patch to avoid revaluation

:) It wasn't meant in bad way.  After all tree is moving fast because
Richard does a lot of work.

> of the arguments to built-in functions has already been reviewed and
> applied to mainline CVS.
> 
> http://gcc.gnu.org/ml/gcc-patches/2003-06/msg02916.html
> http://gcc.gnu.org/ml/gcc-patches/2003-06/msg02933.html

Didn't noticed that, sorry.
> 
> This is why probably why Jan writes "Recently the tree is changing
> rapidly", and also why a recent patch of Jan's that included an early
> version of my fix to builtins.c failed to apply cleanly.
> 
> 
> As for the patch itself, its no longer required to fix any of Jan's
> regressions, but it is a useful optimization.  I didn't include the
> CODE_FOR_nothing optimization (as suggested by Jan) in my patch,
> because its unclear whether optabs.c may try implementing a mathematical
> function by widening the mode, to use an intrinsic available only in a
> wider mode.  For example, implementing sqrtsf2 using sqrtdf2 as in
> (float)sqrt((double)x).  If widening of floating point modes is
> allowed/advantageous, checking whether the builtin intrinsic is
> available in the original mode may miss cases where a platform's
> floating point hardware only provides instructions for doubles, etc...

I think you are caling with OPTAB_DIRECT so optabs won't do that trick.
You may want to allow widenting as it definitly makes sense, then it
would probably make sense to also update the test to check for wider
modes.  Still I tend to believe that the check makes sense as it will
fail for majority of targets...

Honza
> 
> [In theory, x86 only provides math functions on long doubles, but
> we currently lie to optabs, by duplicating these functions for all
> floating point modes].
> 
> Roger
> --