> On Sep 17, 2024, at 07:05, Jakub Jelinek <[email protected]> wrote:
>
> On Tue, Sep 17, 2024 at 11:13:09AM +0200, Jakub Jelinek wrote:
>> So maybe better
>> tree arg = e_p->value;
>> tree f;
>> if ((in_typeof || in_alignof)
>> && TREE_CODE (arg) == COMPONENT_REF
>> && (f = TREE_OPERAND (arg, 1))
>> && TREE_CODE (f) == FIELD_DECL
>> && c_flexible_array_member_type_p (TREE_TYPE (f))
>> && lookup_attribute ("counted_by", DECL_ATTRIBUTES (f))
>> && DECL_NAME (f))
>
> Of course with
> {
> auto save_in_typeof = in_typeof;
> auto save_in_alignof = in_alignof;
> in_typeof = 0;
> in_alignof = 0;
>> arg = build_component_ref (EXPR_LOCATION (arg),
>> TREE_OPERAND (arg, 0),
>> DECL_NAME (f), UNKNOWN_LOCATION,
>> UNKNOWN_LOCATION, true);
> in_typeof = save_in_typeof;
> in_alignof = save_in_alignof;
> }
Yeah, the “save_in_typeof/alignof” is needed. -:)
> Why don't we have the sentinel classes C++ FE has in the C FE?
Sorry for the stupid question: could you please explain the above a little bit
more? What’s the sentinel classes? And how should we use it here?
>
> Or outline the counted_by specific part of build_component_ref into
> a separate function and call just that.
Yeah, this looks like a good idea to me, I will try to split the counted_by
specific part of build_component_ref into a separate function, that might make
the code cleaner and also get rid of the unnecessary global variable
“in_builtin_counted_by_ref”.
>> if (has_counted_by_object (arg))
>> expr.value = get_counted_by_ref (arg);
>> else
>> expr.value = null_pointer_node;
>
> BTW, concerning just counted_by attribute, how does it play together
> with _Atomic on the size member?
> Perhaps it should be disallowed.
Do you mean for the following:
struct t {
_Atomic int size;
char array[] __attribute__ ((counted_by (size))
};
We should emit error and disallow it at this moment?
Thanks.
Qing
>
> The thing is that _Atomic access lowering is done in the FE, so too
> early for tree-object-size, which would either need to reimplement it by
> hand (sure, not that difficult, it needs to just atomically load).
>
> Jakub
>
