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

[Doug Evans <dje at google dot com>]

On Mon, Sep 6, 2010 at 2:48 AM, Jan Kratochvil
<jan.kratochvil@redhat.com> wrote:
> On Fri, 03 Sep 2010 17:59:16 +0200, Doug Evans wrote:
>> Unless code that needs a CU reads it in as necessary, and the API into
>> the dwarf reader only ages the cache at well defined points that no
>> longer need CUs (e.g. when we're done psymtab->symtab expansion).
>> (right?)
>
> This means only one CU is guaranteed to be read in at one time. ?And when you
> hold that CU you must not call any GDB function as this function can (upon
> a change in the future) request its own CU and thus invalidate CU at the
> caller. ?I find it as a too fragile policy.
>
> Still I find it even ineffective. ?Regular aging means CUs get flushed even if
> only a few of them is now read-in.
>
>
>> > Therefore I find prepare_execute_command as the only safe place to flush any
>> > CU.
>>
>> OOC, If we did move cache aging to a higher level, is there a reason
>> it can't be done at cleanup time?
>> [For reference sake, it's actually done today for free_stack_comp_unit.]
>
> The aging currently affects all CUs read-in. ?There can be several nested
> calls each requesting its own CU and doing aging at the (inner) cleanup time.
> That means a cleanup in the inner call may age-out CU in the outer call still
> before the outer cleanup.
>
> Without any CU locking in place I still do not see a safe point other than at
> the top level idle time (that is prepare_execute_command).

How about we take a step back and enumerate the entry points into the
dwarf subsystem (at least those that need CUs).  Then we can see
if/when a CU needs to survive calls across the dwarf API.  If there's
a correctness issue here, perhaps we want to separate it from the
internal optimization detail of CU flushing.  E.g. we may find that
only flushing CUs in prepare_execute_command has issues as well.