Re: [PATCH] Fix profiledbootstrap ada checking failure (PR debug/79255)

[Jakub Jelinek <jakub at redhat dot com>]

On Fri, Mar 24, 2017 at 12:45:28PM +0100, Richard Biener wrote:
> On Fri, Mar 24, 2017 at 9:43 AM, Jakub Jelinek <jakub@redhat.com> wrote:
> > On Fri, Mar 24, 2017 at 09:29:00AM +0100, Richard Biener wrote:
> >> Yeah, the thing BLOCK_NONLOCALIZED_VARS wants to do is optimize generated
> >> dwarf by adding a DW_AT_abstract_origin (just to refer to the
> >> subprogram DIE) but
> >
> > Well, for FUNCTION_DECLs in BLOCK_VARS/BLOCK_NONLOCALIZED_VARS we actually don't
> > emit any further DIE and so there is no DW_AT_abstract_origin.
> > E.g. gen_subprogram_die has:
> >           /* Detect and ignore this case, where we are trying to output
> >              something we have already output.  */
> >           if (get_AT (old_die, DW_AT_low_pc)
> >               || get_AT (old_die, DW_AT_ranges))
> >             return;
> 
> Hmm, but we do want to put the function in scope?  THus
> 
> void foo () {}
> void bar ()
> {
>   int foo;
>   {
>      void foo();
>      foo();
>   }
> }
> 
> should have a DIE for foo in bar (possibly refering to the concrete instance
> for optimization).

We actually do that.  If I change your testcase so that it actually triggers
the changed part of the code (so there is inlining etc.), so:
volatile int v;
__attribute__((noinline)) void foo () { v++; }
static inline void bar ()
{
  int foo;
  {
     void foo();
     foo();
  }
}
void
baz (void)
{
  bar ();
  bar ();
}
then at -gdwarf-3 -dA -O2 the difference between vanilla GCC and my patched GCC is
following (used -gdwarf-3 so that there are no DW_FORM_flag_present that
result in DIE offset changes):
        .uleb128 0xd    # (DIE (0x117) DW_TAG_subprogram)
        .ascii "bar\0"  # DW_AT_name
        .byte   0x1     # DW_AT_decl_file (prQQ.c)
        .byte   0x3     # DW_AT_decl_line
        .byte   0x3     # DW_AT_inline
        .long   0x13c   # DW_AT_sibling
        .uleb128 0xe    # (DIE (0x123) DW_TAG_variable)
        .ascii "foo\0"  # DW_AT_name
        .byte   0x1     # DW_AT_decl_file (prQQ.c)
        .byte   0x5     # DW_AT_decl_line
        .long   0x41    # DW_AT_type
        .uleb128 0xf    # (DIE (0x12e) DW_TAG_lexical_block)
        .uleb128 0x10   # (DIE (0x12f) DW_TAG_subprogram)
        .byte   0x1     # DW_AT_external
        .ascii "foo\0"  # DW_AT_name
        .byte   0x1     # DW_AT_decl_file (prQQ.c)
        .byte   0x7     # DW_AT_decl_line
-       .byte   0       # DW_AT_inline
+       .byte   0x1     # DW_AT_declaration
        .uleb128 0x11   # (DIE (0x138) DW_TAG_unspecified_parameters)
        .byte   0       # end of children of DIE 0x12f
        .byte   0       # end of children of DIE 0x12e
        .byte   0       # end of children of DIE 0x117
        .uleb128 0x12   # (DIE (0x13c) DW_TAG_subprogram)
        .byte   0x1     # DW_AT_external
        .ascii "foo\0"  # DW_AT_name
        .byte   0x1     # DW_AT_decl_file (prQQ.c)
        .byte   0x2     # DW_AT_decl_line
        .quad   .LFB0   # DW_AT_low_pc
        .quad   .LFE0   # DW_AT_high_pc
        .byte   0x1     # DW_AT_frame_base
        .byte   0x9c    # DW_OP_call_frame_cfa
        .byte   0x1     # DW_AT_GNU_all_call_sites
        .byte   0       # end of children of DIE 0xb
and corresponding change in .debug_abbrev.
That seems to be the right change to me, inside of bar (the DW_AT_inline
one) we have a declaration of foo, and we properly use DW_AT_declaration for
it, then at the toplevel we actually have the full definition die for foo,
and then we have in two spots inside baz
        .uleb128 0x6    # (DIE (0x6d) DW_TAG_inlined_subroutine)
        .long   0x117   # DW_AT_abstract_origin
without the need to duplicate the foo declaration in there,
that is inherited through DW_AT_abstract_origin.

	Jakub