Re: [binutils] libctf: Remove undefined functions from ver. map

[Nick Alcock <nick.alcock@oracle.com>]

On 1 Mar 2024, Nicholas Vinson spake thusly:

>
>
> On 2/28/24 13:14, Nick Alcock wrote:
>> On 28 Feb 2024, Nicholas Vinson told this:
>> 
>>> On 2/27/24 10:05, Nick Alcock wrote:
>>>> On 27 Feb 2024, Sam James said:
>>>>
>>>>> Nicholas Vinson <nvinson234@gmail.com> writes:
>>>>>
>>>>>> The functions ctf_label_set(), ctf_label_get(), ctf_arc_open(), ctf_fdopen(),
>>>>>> ctf_open(), ctf_bfdopen(), and ctf_bfdopen_ctfsect() are not defined. Their
>>>>>> inclusion in libctf/libctf.ver causes clang/llvm LTO optimizatiosn to fail with
>>>>>> error messages similar to
>>>> This is definitely not right. They *are* defined, but only for some
>>>> libraries built from this version script. You can't just take them out.
>>>
>>> Could you point me to the definitions for ctf_label_set() and ctf_label_get() ?
>> Those two are long dead and gone and should indeed be removed from the
>> .ver script. (I wonder why Solaris's linker, which is just as picky
>> about version scripts referencing only symbols that actually exist,
>> never warned me about this.)
>> 
>>> I do find definitions for the other so the other symbols, so I'll find
>>> a different way to handle those symbols in the version file.
>> Commenting them as /* libctf only. */ should be enough, I think. Only
>> one symbol seems to be missing, ctf_arc_open. The others are already
>> properly marked and should already be being excluded from the .ver
>> script for libctf_nobfd.so.)
>> 
>
> I see what you're talking about in libctf/configure.ac. However, it's protected by a `test -n "$decommented_version_script"` check.
> Because ld.lld does not understand `-B local`, that variable remains empty and the lines in the version file containing /* libctf
> only.  */ are not removed.

Aha! GNU ld doesn't support that option either. This test was meant to
detect Solaris ld, which supports version scripts with similar "no
undefined symbols" limitations to what lld seems to here. I'd suggest
that configure test needs augmenting to catch lld as well when this flag
is provided (so specifying $LDFLAGS in whatever test you carry out).

-- 
NULL && (void)