Re: [PATCH 1/3] gdb/ctf: don't create a buildsym_compunit when building partial symbols

[Wei-min Pan <weimin.pan@oracle.com>]

On 3/31/2022 2:20 PM, Simon Marchi wrote:
> I am trying to do some changes to buildsym_compunit, so I am auditing
> the current uses.  Something seems odd with this use of
> buildsym_compunit (that this patch removes).
>
> A buildsym_compunit is normally used when building a compunit_symtab.
> That is, when expanding a partial symtab into a full compunit symtab.
> In ctfread.c, a buildsym_compunit is created in ctf_start_archive, which
> is only used when creating partial symtabs.  At this moment, I don't
> see how that's useful.  ctf_start_archive creates a new
> buildsym_compunit and starts a subfile.  But that buildsym_compunit is
> never used again.  It's just overriden in ctf_start_symtab, which means
> we leak the old buildsym_compunit, I suppose.
>
> Remove ctf_start_archive completely.  Add an assert in
> ctf_start_symtab to verify that we are not overwriting an existing
> buildsym_compunit (meaning we'd leak the existing one).  This assert
> triggers without the other part of the fix.  When doing:
>
>    $ ./gdb --data-directory=data-directory /tmp/babeltrace-ctf/src/lib/.libs/libbabeltrace2.so.0.0.0
>    ...
>    (gdb) maintenance expand-symtabs
>    /home/simark/src/binutils-gdb/gdb/ctfread.c:1255: internal-error: ctf_start_symtab: Assertion `!ccp->builder' failed.
>
> Change-Id: I666d146454a019f08e7305f3a1c4a974d27b4592
> ---
>   gdb/ctfread.c | 20 +-------------------
>   1 file changed, 1 insertion(+), 19 deletions(-)
>
> diff --git a/gdb/ctfread.c b/gdb/ctfread.c
> index 8da38cc70af8..8636692e2e1b 100644
> --- a/gdb/ctfread.c
> +++ b/gdb/ctfread.c
> @@ -1531,21 +1531,6 @@ ctf_psymtab_var_cb (const char *name, ctf_id_t id, void *arg)
>     return 0;
>   }
>   
> -/* Start a subfile for CTF. FNAME is the name of the archive.  */
> -
> -static void
> -ctf_start_archive (struct ctf_context *ccx, struct objfile *of,
> -		   const char *fname)
> -{
> -  if (ccx->builder == nullptr)
> -    {
> -      ccx->builder = new buildsym_compunit (of,
> -		      of->original_name, nullptr, language_c, 0);
> -      ccx->builder->record_debugformat ("ctf");
> -    }
> -  ccx->builder->start_subfile (fname);
> -}
> -
>   /* Setup partial_symtab's describing each source file for which
>      debugging information is available.  */
>   
> @@ -1567,10 +1552,7 @@ scan_partial_symbols (ctf_dict_t *cfp, psymtab_storage *partial_symtabs,
>   
>     struct ctf_context *ccx = &pst->context;
>     if (isparent == false)
> -    {
> -      ctf_start_archive (ccx, of, fname);
> -      ccx->pst = pst;
> -    }
> +    ccx->pst = pst;
>   
>     if (ctf_type_iter (cfp, ctf_psymtab_type_cb, ccx) == CTF_ERR)
>       complaint (_("ctf_type_iter scan_partial_symbols failed - %s"),

A CTF archive contains conflicting type info and is treated as a CU.
Since one can't switch scope among CUs in CTF, not reading archives
might not be a problem. Please go ahead and make the change. It can
be revisited when we come up a test case that shows otherwise.

Thanks.

>
> base-commit: 59f837cb11e3b96c581fe23b79f6995b7c8177ee