On Mon, Oct 15, 2012 at 12:00 PM, Eric Botcazou <ebotca...@adacore.com> wrote:
>> Hmm. I think this points to an issue with update_alias_info_with_stack_vars
>> instead. That is, this function should have already cared for handling
>> this case where two decls have their stack slot shared.
>
> The problem here is that mark_addressable is called _after_ the function is
> run. IOW, by the time update_alias_info_with_stack_vars is run, there are no
> aliased variables in the function.
Where is mark_addressable called? It's wrong (and generally impossible) to
do that late.
>> static bool
>> can_escape (tree expr)
>> {
>> tree base;
>> if (!expr)
>> return true;
>> base = get_base_address (expr);
>> if (DECL_P (base)
>> && !may_be_aliased (base))
>> return false;
>> return true;
>>
>> it needs to check decls_to_pointers[base] and then check
>> if any of the pointed-to decls may be aliased.
>
> That's essentially what the patch does though (except that it does it more
> efficiently), since update_alias_info_with_stack_vars correctly computes
> cfun->gimple_df->escaped.vars for partitioned decls.
No, what it does is if a decl is in ESCAPED make sure to add decls that
share the same partition also to ESCAPED. The issue is that can_escape
queries TREE_ADDRESSABLE (which is correct on the gimple level, only
things that have their address taken can escape) - that's no longer possible
as soon as we have partitions with both addressable and non-addressable
decls.
>> Now, that's not that easy because we don't have a
>> mapping from DECL UID to DECL (and the decl
>> isn't in the escaped solution if it is just used by
>> memcpy), but we could compute a bitmap of
>> all address-taken decls in update_alias_info_with_stack_vars
>> or simply treat all check decls_to_pointers[base] != NULL
>> bases as possibly having their address taken.
>
> OK, we can populate another bitmap in update_alias_info_with_stack_vars and
> update it in mark_addressable by means of decls_to_pointers and pi->pt.vars.
> That seems a bit redundant with cfun->gimple_df->escaped.vars, but why not.
If you only have memcpy then escaped will be empty. fixing escaped is
not the right solution (it may work for some reason in this case though).
The rtl code has to approximate ref_maybe_used_by_call_p in a conservative
way which it doesn't seem to do correctly (I don't remember a RTL alias.c
interface that would match this, or ref_maybe_used_by_stmt_p - maybe
we should add one?)
Thanks,
Richard.
> --
> Eric Botcazou
