[Mesa-dev] [PATCH 4/6] meson: osx doesn't have librt, so don't require it

[Emil Velikov]

On 31 January 2018 at 16:07, Jon Turney <jon.turney at dronecode.org.uk> wrote:
> On 31/01/2018 15:21, Emil Velikov wrote:
>>
>> On 30 January 2018 at 20:27, Dylan Baker <dylan at pnwbakers.com> wrote:
>>>
>>> Quoting Emil Velikov (2018-01-30 10:56:42)
>>>>
>>>> Hi Jon,
>>>>
>>>> On 28 January 2018 at 14:24, Jon Turney <jon.turney at dronecode.org.uk>
>>>> wrote:
>>>>>
>>>>> ---
>>>>>   meson.build | 2 +-
>>>>>   1 file changed, 1 insertion(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/meson.build b/meson.build
>>>>> index 7e194a9f10d..8fdbaa8b8d8 100644
>>>>> --- a/meson.build
>>>>> +++ b/meson.build
>>>>> @@ -935,7 +935,7 @@ elif with_dri_i965 and get_option('shader-cache')
>>>>>   endif
>>>>>
>>>>>   # Determine whether or not the rt library is needed for time
>>>>> functions
>>>>> -if cc.has_function('clock_gettime')
>>>>> +if cc.has_function('clock_gettime') or (host_machine.system() ==
>>>>> 'darwin')
>>>>
>>>>
>>>> Absolutely no objections against the patch - just a small question.
>>>> If the meson/autotools check fails, does this mean that the resulting
>>>> binaries are having unresolved reference wrt said symbol?
>>>>
>>>> Thanks
>>>> Emil
>>>
>>>
>>> Not for meson, it builds -Wl,--no-undefined (or it's MSVC equivalent) by
>>> default, so it shouldn't be possible to get unresolved symbols in a
>>> binary or
>>> shared library.
>>>
>> Right, the question is why does the test (has_function) fails?
>
>
> Some misunderstanding going on here, I think. It fails if the function isn't
> present.
>
> (We then go on to report the failure as being librt not found, not that we
> couldn't work out how to find clock_gettime, but perhaps that's a separate
> issue)
>
>> A few possible solutions come to mind, but only one with toolchain
>> handy can confirm.
>>   - the test is broken (regardless meson/autotools/foo implementation)
>>   - the symbol is indirectly resolved
>> The est does not pull libfoo, while the final binary does. Libfoo
>> pulls libbar with the latter providing the symbol.
>>   - the final binary is having unresolved reference
>>   - all the clock_gettime references get 'magically' removed due to the
>> linker garbage collector
>>
>>  From a quick search OSX 10.12 (or so) has clock_gettime. But details
>> are extremely sparse :-(
>
>
> This change (and the check that autoconf is currently doing) is not needed.
>
> If we're targeting OSX 10.12 or later, clock_gettime() exists.
>
> Commit 990bd49f might have made sense once, but I think now that
> clock_gettime is used in include/c11/threads_posix.h, we're going to fail to
> build for OSX prior to 10.12, even when not trying to link with librt.
>
> (I also had a patch to provide an implementation of clock_gettime if the
> target is an earlier OSX version, but I dropped it as probably not very
> useful)

Right, so in this case:
 - targeting pre 10.12 where API is missing (be that librt or elsewhere)
 - 990bd49f and this patch are off, remove the linkage and rely on the
linker GC magic

I think, one wants to distinct between presence vs linkage requirements.
ATM they're assumed to be the same thing.

-Emil