Hi,
I will fix the free count issue - my internet access in China was bit sporadic
and I hoped you will know why the count seem to disagree between PPC64 and Linux
builds. In meantime PPC64 stopped to build for me.
>
> (early) dwarf2out is a good place, likewise verifying right after
> free-lang-data. I agree that LTO type merging is also a good place.
>
> I hope we get early dwarf2out merged soon and can enable free-lang-data
> also for non-LTO compilation.
>
> > LTO-Boostrapped/regtested x86_64-linux without Ada. I am doing full run on
> > PPC64, but it fails on unrelated libgmp miscomplation I proably need to
> > track down
> > first.
> >
> > OK if testing passes?
>
> Please make sure to test _all_ languages (all,ada,obj-c++,go) and also
> include multilibs in testing (thus -m64/-m32).
>
> You don't verify that TYPE_CANONICAL is consistent nor that TYPE_CANONICAL
> doesn't form a tree (TYPE_CANONICAL (t) == TYPE_CANONICAL (TYPE_CANONICAL
> (t))).
Yep, I have this one in queue, but of course it fires, thus it is not in the
initial patch.
> > +#ifdef ENABLE_CHECKING
> > + verify_type (expr);
> > +#endif
>
> As said I think that doing this after free-lang-data is better. Like
> here:
>
> /* Traverse every type found freeing its language data. */
> FOR_EACH_VEC_ELT (fld.types, i, t)
> free_lang_data_in_type (t);
>
> do another loop verifying types.
OK, we however build new types in middle end (ipa-icf/ipa-split/ipa-sra),
so perhaps we can check on both places?
>
> > stream_write_tree (ob, TYPE_SIZE (expr), ref_p);
> > stream_write_tree (ob, TYPE_SIZE_UNIT (expr), ref_p);
> > stream_write_tree (ob, TYPE_ATTRIBUTES (expr), ref_p);
> > Index: dwarf2out.c
> > ===================================================================
> > --- dwarf2out.c (revision 222391)
> > +++ dwarf2out.c (working copy)
> > @@ -21264,6 +21264,11 @@ dwarf2out_decl (tree decl)
> > {
> > dw_die_ref context_die = comp_unit_die ();
> >
> > +#ifdef ENABLE_CHECKING
> > + if (TREE_TYPE (decl))
> > + verify_type (TREE_TYPE (decl));
> > +#endif
> > +
>
> Hmm, this has the chance to verify types multiple times, no? Wouldn't
> gen_type_die_with_usage be a better place to verify the type we create
> the DIE for?
That looks better indeed. The checks are not paritcularly expensive though.
Honza
