Re: [C++ PATCH] Re: gcc trunk rev. 127646 breaks bootstrap during stage 3

["Richard Guenther" <richard dot guenther at gmail dot com>]

On 8/20/07, Jakub Jelinek <jakub@redhat.com> wrote:
> On Mon, Aug 20, 2007 at 07:26:02AM -0700, Andrew Pinski wrote:
> > > The statement which we are gimplifying:
> > > <<<change_dynamic_type (struct __recursive_mutex *) &fake_mutex)>>>
> > >
> > > 5887            case CHANGE_DYNAMIC_TYPE_EXPR:
> > > 5888              ret = gimplify_expr (&CHANGE_DYNAMIC_TYPE_LOCATION (*expr_p),
> > > 5889                                   pre_p, post_p, is_gimple_reg, fb_lvalue);
> > >
> > >
> > > So why fb_lvalue?
> >
> > Actually this is correct, we need a decl here as we need to mark that
> > for pointing to anything.  Jakub, how did you test your patch again?
>
> I'm sorry, I bet I forgot to make in libstdc++-v3 this time, only tested
> make -C check-g++ and make -C x86_64*/libstdc++-v3.
>
> The problem is that save_expr on constant arg (which &fake_mutex is)
> does nothing at all and we really need the SAVE_EXPR around, so that
> during gimplification a temporary is created (but we can't create it
> right away in the FE, it must be initialized in the correct spot).
>
> Attached are two possible fixes, one is to force creation of SAVE_EXPR
> when it wasn't added by save_expr, the other one is to use a TARGET_EXPR
> instead (force creation thereof).
> I have briefly tested the former, the latter was tested with make
> in libstdc++-v3, make check-g++ and make check in libstdc++-v3.
>
> Ok for trunk (which one)?

I think the target expr one is better though I cannot approve the patch.

Richard.