Re: DCE enhancement to delete dead vptr assignment

[Paolo Bonzini <bonzini at gnu dot org>]

On 04/16/2010 07:01 PM, Xinliang David Li wrote:
> On Fri, Apr 16, 2010 at 9:29 AM, Steven Bosscher<stevenb.gcc@gmail.com>  wrote:
> > +    {
> > +      /* Initialization of the vptr field for the same object this
> > +         destructor is operating on won't be live out of the function
> > +        if it is not used within the function, so do not treat them
> > +        as a hidden global store.  */
> > +      if (lang_hooks.decls.is_decl_dtor (current_function_decl))
> > +        {
>
> Do you want me to wrap it in a wrapper function?

No, it should be avoided altogether.

In particular:

- one langhook you add, is_polymorphic_type_p, is unused

- one, is_global_delete_fndecl_p type can be replaced with an attribute 
(if it is safe WRT redefinitions).

- one, cp_is_vptr_field_p, need not be a langhook at all as far as I 
could see as it only accesses common fields (if I'm wrong, you may be 
masking other bugs).

This leaves is_decl_dtor.  I think the front-end should add one dummy 
GIMPLE statement (e.g. a call to a builtin with a special attribute and 
that expand to nothing) that says "treat this datum as dead".  This way, 
this special case can fall in the normal dataflow scheme, as Richard 
suggested.

BTW, is_killed_by_delete is called rarely---though you never know about 
pathological cases---but it is slow!

> > We should not be calling lang hooks from the optimizers.
> >
> > How would this work with inter-language IPA?
>
> When can the tree structure be flatten out so that there is no
> language specific attributes? Until then, what is the solution?
> Inter-language IPA? When do we plan to have full support of that?

Already now having optimizations that pass with g++ but fail with LTO is 
considered bad.

Paolo