Re: [patch] Fix DW_OP_call2 and DW_OP_call4 for max-cache-age 0

[Doug Evans <dje@google.com>]

On Thu, Sep 2, 2010 at 9:02 AM, Jan Kratochvil
<jan.kratochvil@redhat.com> wrote:
> On Mon, 23 Aug 2010 21:30:06 +0200, Doug Evans wrote:
>> On Mon, Aug 23, 2010 at 11:50 AM, Jan Kratochvil <jan.kratochvil@redhat.com> wrote:
>> > as discussed on #gdb when you set max-cache-age 0 DW_OP_call{2,4} crashed GDB.
> [...]
>> > --- a/gdb/dwarf2read.c
>> > +++ b/gdb/dwarf2read.c
>> > @@ -1636,6 +1636,11 @@ dw2_do_instantiate_symtab (struct objfile *objfile,
>> >  {
>> >   struct cleanup *back_to;
>> >
>> > +  /* Age the cache, releasing compilation units that have not been used
>> > +     recently.  Age them first so that we do not age out the requested PER_CU
>> > +     unit if DWARF2_MAX_CACHE_AGE is too low.  */
>> > +  age_cached_comp_units ();
>>
>> Aging cached units first feels weird (if not wrong at least weird); we
>> may toss out something we're about to want.
>> At the least IWBN to elaborate on why this fixes things.
>
> As otherwise we will age out what we have found (on max-cache-age 0).

Ah.
Still, dw2_do_instantiate_symtab seems like the wrong tool for the job here.
Its job is to instantiate a symtab, it currently doesn't guarantee it
will leave the CU read in when finished, and adding that guarantee
doesn't feel right.

Assuming (and I don't know dwarf2_fetch_die_location_block well) just
needs the dies and not a symtab, how about moving this bit of code to
its own function, and calling it from both dw2_do_instantiate_symtab
and dwarf2_fetch_die_location_block.

    if (per_cu->from_debug_types)
      read_signatured_type_at_offset (objfile, per_cu->offset);
    else
      load_full_comp_unit (per_cu, objfile);

I haven't thought it through (e.g. it may need a bit of glue), but it
feels like a better approach.

> One could forbid value zero for max-cache-age but that also does not seem
> right to me.

max-cache-age == 0 is defined to disable the cache.  It's a useful
test vehicle, and I don't see any reason to disallow it either.

> There is such a general cleanup moment when GDB is fully idle
> - prepare_execute_command() - shouldn't age_cached_comp_units be called there?

I don't know.  Or as a cleanup (either via a cleanup itself, or as
part of some top level thing akin to whatever you'd do in
prepare_execute_command.  making use of an existing facility
(make_cleanup) would be preferable of course, assuming it's the way to
go)? It feels better to do this at the end of a command, not before.

> But that way sooner or later we will age out every CU.  This may occur a bit
> even nowadays, the default value 5 is also very low.  max-cache-age as "how
> long" is IMO not userful to the user.  There could be more a setting "how
> many" CUs can be loaded at once.  CU age would be then just an internal
> indicator to maintain the count under the "how many" limit.

A better measure may be memory used (e.g. lots of CUs are ok if
they're all relatively small).  IWBN to find/collect stats on the
distribution of #CUs and sizes.  [e.g. can we make some reasonable
assumptions so that we don't have to track die memory usage?]

> I would change "max-cache-age" to "max-cache-size" and call it from
> prepare_execute_command() instead.  I will provide a patch if not replied.