[PATCH 2/3] bfd/ELF: Retain strong undefined symbols under -z dynamic-undefined-weak

[Jan Beulich]

On 31.08.2025 12:54, Hakan Candar wrote:
> Jan Beulich <jbeulich@suse.com> wrote:
>> On 16.07.2025 18:03, Hakan Candar wrote:
>> > Add `bfd_is_undef_symbol` macro for clarity and consistency across
>> > backends. Adjust symbol hiding and dynamic marking logic in elflink.c
>> > to treat undefined strong symbols the same as weak ones when applying
>> > retention rules.
>> > 
>> > This improves alignment with the gABI and LLD behavior, especially
>> > under `-z [no]dynamic-undefined-weak`. Backends that already respected
>> > this logic (e.g., x86_64, ppc) show no regressions, and can now retain
>> > strong undefined symbols as dynamic relocations via -z dynamic-undefined-weak.
>>
>> What does -z dynamic-undefined-weak have to do with strong undefined symbols?
>> (I guess there is simply some context missing here, but I&#x27;d like to understand
>> that before looking at the patch itself.)
>>
> 
> The flag name is admittedly misleading. After I first sent this out as
> RFC [1], I determined I would reuse the existing dynamic-undefined-weak
> flag, even though the behavior extends to strong undefined symbols as 
> well.
> 
> As Fangrui Song pointed out [2], LLD recently generalized the option in
> exactly this way: it governs retention of both weak and strong
> undefineds. I mirrored that model for compatibility and to avoid
> introducing yet another variant flag that controls nearly the same
> logic.
> 
> That said, if the consensus is that the current name is too narrow for
> what it now does, I can rework the patch to split the behavior or
> introduce a new option.

The present name is likely okay to keep (others may want to chime in, though)
as long as it is made very clear in doc and description that what it does
goes beyond what its name says. That is, in the doc change imo you don't
want to simply drop "weak" but emphasize that both weak and "normal"
("strong") symbols are dealt with the same way when the option is given.

Jan