On Wed, Nov 02, 2016 at 10:19:26AM +0100, Richard Biener wrote:
> On Tue, Nov 1, 2016 at 1:09 AM, Jakub Jelinek <ja...@redhat.com> wrote:
> > On Mon, Oct 31, 2016 at 05:28:42PM -0500, Bill Schmidt wrote:
> >> The PowerPC back end loses performance on vector intrinsics, because
> >> currently
> >> all of them are treated as calls throughout the middle-end phases and only
> >> expanded when they reach RTL. Our version of altivec.h currently defines
> >> the
> >> public names of overloaded functions (like vec_add) to be #defines for
> >> hidden
> >> functions (like __builtin_vec_add), which are recognized in the parser as
> >> requiring special back-end support. Tables in rs6000-c.c handle dispatch
> >> of
> >> the overloaded functions to specific function calls appropriate to the
> >> argument
> >> types.
> >
> > This doesn't look very nice. If all you care is that the builtins like
> > __builtin_altivec_vaddubm etc. that __builtin_vec_add overloads into fold
> > into generic vector operations under certain conditions, just fold those
> > into whatever you want in targetm.gimple_fold_builtin (gsi).
>
> Note that traditionally "overloading" with GCC "builtins" is done by
> using varargs
> and the "type generic" attribute. That doesn't scale to return type
> overloading
> though for which we usually added direct support to the parser (for example
> for __builtin_shuffle).
My understanding is that rs6000 already does that, it hooks into
resolve_overloaded_builtin which already handles the fully type generic
builtins where not just the arguments, but also the return type can be
picked up. But it resolves the overloaded builtins into calls to other
builtins that are not type-generic.
So, either that function instead of returning the specific md builtin calls
in some cases already returns trees with the generic behavior of the
builtin, or it returns what it does now and then in the gimple fold builtin
target hook (note, the normal fold builtin target hook is not right for
that, because it is mostly used for folding builtins into constant - callers
will usually throw away other results) fold those specific md builtins
into whatever GIMPLE you want. If we want to decrease amount of folding in
the FEs, the gimple fold builtin hook is probably better.
Jakub
