Re: [PATCH] [gdb/symtab] Find main language without symtab expansion

[Tom de Vries <tdevries@suse.de>]

On 8/4/23 19:17, Tom Tromey wrote:
>>>>>> "Tom" == Tom de Vries via Gdb-patches <gdb-patches@sourceware.org> writes:
> 
> Tom> Reimplement the feature by adding
> Tom> cooked_index_functions::lookup_global_symbol_language.
> 
> Thanks for the patch.
> 
> Tom> +    /* Expansion of large CUs can be slow.  By returning the language of main
> Tom> +       here for C and C++, we avoid CU expansion during set_initial_language.
> Tom> +       But by doing a symbol lookup in the cooked index, we are forced to wait
> Tom> +       for finalization to complete.  See PR symtab/30174 for ideas how to
> Tom> +       bypass that as well.  */
> Tom> +    cooked_index *table = per_bfd->index_table->index_for_writing ();
> 
> I think it's more normal to use:
> 
>    cooked_index *table
>      = (gdb::checked_static_cast<cooked_index *>
>         (per_bfd->index_table.get ()));
> 
> in the body of cooked_index_functions.
> 

Ack, updated (not reposting for this), thanks.

> Tom> +    for (const cooked_index_entry *entry : table->find (name, false))
> 
> I think this is probably racy, if the intent is not to wait for
> finalization, because cooked_index_shard::find uses the canonical name.
> 

The intent is to wait for finalization, as I've tried to make clear in 
the comment quoted above.

Indeed cooked_index_shard::find uses the canonical name, and I think 
that's why it uses "wait ()" to wait for finalization.

So, I don't think this is racy.

I've also checked by building with -fsanitizer=thread and running the 
two test-cases with make-check-all.sh.

> Maybe we need to resurrect the approach of looking for "main" in C/C++
> CUs.

Um, AFAICT that's exactly what this patch does, so I suppose you mean 
something else.

Maybe you mean reverting this bit of commit 47fe57c9281 ("Fix "start" 
for D, Rust, etc"):
...
@@ -218,10 +232,6 @@ cooked_index_shard::add (sect_offset die_offset, 
enum dwarf_tag tag,
       implicit "main" discovery.  */
    if ((flags & IS_MAIN) != 0)
      m_main = result;
-  else if (per_cu->lang () != language_ada
-	   && m_main == nullptr
-	   && strcmp (name, "main") == 0)
-    m_main = result;

    return result;
  }
...

That was my initial approach, but it required a lot of reasoning about 
why it's not racy, so I like this approach better, which is not racy by 
design.

I'll commit this tomorrow unless there are further comments.

Thanks,
- Tom

> I don't really remember all the state of this right now, there's
> been a lot of flux.  I did also reimplement some of this on my
> background DWARF reading branch, but I think that's kind of invasive.
> 
> Tom