Re: [RFC] is-a.h supporting casting from non const to const

[Trevor Saunders <tsaunders at mozilla dot com>]

On Sun, Nov 23, 2014 at 03:12:06PM -0800, Andrew Pinski wrote:
> On Sun, Nov 23, 2014 at 12:51 PM, Trevor Saunders <tsaunders@mozilla.com> wrote:
> > On Sun, Nov 23, 2014 at 12:05:52PM -0800, Andrew Pinski wrote:
> >> Hi,
> >>   While working on a patch to change some error to inform, I changed
> >> the as_a cast in RTL_LOCATION to be as_a <const rtx_insn *>.  This
> >> falls down when you start with a rtx rather than a const_rtx..
> >
> > so, don't do that? if you have a const_rtx use as_a<const rtx_insn *>,
> > otherwise as_a<rtx_insn *>.  I can't think of a case where it is
> > necessary to cast to const rtx_insn * from rtx, instead of letting the
> > constification be implicit.
> 
> In a function, I already have a const_rtx and I want to use
> RTL_LOCATION to get a location for the rtx.  How should I handle this
> case?  In 4.9, it would just work as there was no casting needed but
> in GCC 5 I need to cast away the const first before I could use
> RTL_LOCATION.

erf, I forgot about macros.

> RTL_LOCATION does not modify at all the rtx so it should be using
> as_a<const rtx_insn *>.  Should I just add a cast to const_rtx inside
> the RTL_LOCATION macro before doing the as_a<const rtx_insn*>?

I'd say ideally it should be changed to an inline function, but casting
seems pretty reasonable for now.

Trev

> 
> Thanks,
> Andrew
> 
> >
> > Trev
> >
> >>
> >> I had to modify is-a.h to this but I really don't like the
> >> modification at all since it shows the bigger issues with templates.
> >> Can anyone think of a better way of supporting this?  Attached is my
> >> patch currently which is able to bootstrap with the modification to
> >> RTL_LOCATION being there.
> >>
> >> Thanks,
> >> Andrew Pinski
> >
> >> Index: is-a.h
> >> ===================================================================
> >> --- is-a.h    (revision 217985)
> >> +++ is-a.h    (working copy)
> >> @@ -147,10 +147,26 @@ struct is_a_helper
> >>  {
> >>    template <typename U>
> >>    static inline bool test (U *p);
> >> +  static inline bool test (T *) { return true; }
> >>    template <typename U>
> >>    static inline T cast (U *p);
> >>  };
> >>
> >> +
> >> +template <typename T>
> >> +struct is_a_helper<const T*>
> >> +{
> >> +  template <typename U>
> >> +  static inline bool test (const U *p);
> >> +  template <typename U>
> >> +  static inline bool test (U *p) {return test(const_cast<const U*>(p));}
> >> +  static inline bool test (const T *) { return true; }
> >> +  template <typename U>
> >> +  static inline const T *cast (const U *p);
> >> +  template <typename U>
> >> +  static inline const T *cast (U *p) { return cast(const_cast<const U*>(p)); }
> >> +};
> >> +
> >>  /* Note that we deliberately do not define the 'test' member template.  Not
> >>     doing so will result in a build-time error for type relationships that have
> >>     not been defined, rather than a run-time error.  See the discussion above
> >> @@ -169,6 +185,15 @@ is_a_helper <T>::cast (U *p)
> >>  }
> >>
> >>
> >> +template <typename T>
> >> +template <typename U>
> >> +inline const T*
> >> +is_a_helper <const T*>::cast (const U *p)
> >> +{
> >> +  return reinterpret_cast <const T*> (p);
> >> +}
> >> +
> >> +
> >>  /* The public interface.  */
> >>
> >>  /* A generic test for a type relationship.  See the discussion above for when
> >