[Bug debug/78835] [7 regression] ICE with -fdebug-types-section and member function

[rguenther at suse dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78835

--- Comment #12 from rguenther at suse dot de <rguenther at suse dot de> ---
On Tue, 17 Jan 2017, jakub at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78835
> 
> --- Comment #11 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
> (In reply to Richard Biener from comment #10)
> > (In reply to Jakub Jelinek from comment #9)
> > > Created attachment 40524 [details]
> > > gcc7-pr78835.patch
> > > 
> > > Untested patch that fixes the ICE on this testcase by marking directly
> > > called functions as needed.  Will see what will it do with the size of the
> > > debug info.
> > > Of course, it is not sufficient, with -fdebug-types-section we can still ICE
> > > unless the earlier patch is also applied 9or some more sophisticated one to
> > > construct skeleton DIEs if type inside of type unit is detected).
> > 
> > I think this is a reasonable approach (plus the hunk from comment#1 of
> > course).
> 
> For the above patch (sans #c1) I've bootstrapped/regtested it last night, and
> compared to gcc without the patch the growth of the debug info is small and
> IMHO acceptable.  On cc1plus on x86_64 comparing build without the patch and
> with the patch (in the former I've applied the patch and did make in stage3
> dir, so that it is the same source) .debug_info size grew by 0.009% and in
> libstdc++.so.6 by 0.29%.
> 
> > But we can't really create new skeleton DIEs after early-finish because with
> > LTO we do not have access to enough information to build type DIEs.
> 
> I think it is possible, e.g. considering
> namespace A { inline namespace B { namespace C { struct D { struct Bar { void
> Baz (); int i; }; };
> } } int vvv = 17; }
> A::C::D::Bar b;
> void Foo (A::C::D::Bar &t) { t.Baz (); }
> to create
>         .uleb128 0x2    # (DIE (0x25) DW_TAG_namespace)
>         .ascii "A\0"    # DW_AT_name
>                         # DW_AT_declaration
>         .uleb128 0x3    # (DIE (0x2c) DW_TAG_namespace)
>         .ascii "B\0"    # DW_AT_name
>                         # DW_AT_declaration
>         .uleb128 0x3    # (DIE (0x2f) DW_TAG_namespace)
>         .ascii "C\0"    # DW_AT_name
>                         # DW_AT_declaration
>         .uleb128 0xa    # (DIE (0x32) DW_TAG_structure_type)
>         .ascii "D\0"    # DW_AT_name
>                         # DW_AT_declaration
>         .uleb128 0x4    # (DIE (0x35) DW_TAG_structure_type)
>         .ascii "Bar\0"  # DW_AT_name
>                         # DW_AT_declaration
>         .byte   0xf0    # DW_AT_signature
>         .byte   0x8
>         .byte   0x56
>         .byte   0xd4
>         .byte   0x2d
>         .byte   0xa7
>         .byte   0x74
>         .byte   0x7c
>         .uleb128 0x8    # (DIE (0x5c) DW_TAG_subprogram)
>                         # DW_AT_external
>                         # DW_AT_declaration
>         .ascii "Baz\0"  # DW_AT_name

In the end the great plan(TM) is to no longer have the TREE_TYPEs/DECLs of
the contexts around during GIMPLE optimizations ...  I think
-fdebug-types-section always creates the skeletons so we may choose to
just never remove those as unused?

> Where the DW_AT_signature would actually point to the .debug_types signature
> and therefore the full declaration.  But I'm not going to work on that right
> now, so would prefer for now just #c9 + #c1 + testcase.

Yes, that's reasonable.