[Intel-gfx] [PATCH i-g-t v3 1/4] meson: add libatomic dependency

[Guillaume Tucker]

On 18/06/2019 15:37, Ser, Simon wrote:
> On Tue, 2019-06-18 at 14:59 +0100, Guillaume Tucker wrote:
>> On 18/06/2019 14:20, Ser, Simon wrote:
>>> On Tue, 2019-06-18 at 13:27 +0100, Guillaume Tucker wrote:
>>>> Add conditional dependency on libatomic in order to be able to use the
>>>> __atomic_* functions instead of the older __sync_* ones.  The
>>>> libatomic library is only needed when there aren't any native support
>>>> on the current architecture, so a linker test is used for this
>>>> purpose.  This enables atomic operations to be on a wider number of
>>>> architectures including MIPS.
>>>>
>>>> Signed-off-by: Guillaume Tucker <guillaume.tucker at collabora.com>
>>>> ---
>>>>
>>>> Notes:
>>>>     v2: add linker test for libatomic
>>>>     v3: use null_dep
>>>>
>>>>  meson.build | 14 ++++++++++++++
>>>>  1 file changed, 14 insertions(+)
>>>>
>>>> diff --git a/meson.build b/meson.build
>>>> index 6268c58d3634..118ad667ffb5 100644
>>>> --- a/meson.build
>>>> +++ b/meson.build
>>>> @@ -180,6 +180,20 @@ realtime = cc.find_library('rt')
>>>>  dlsym = cc.find_library('dl')
>>>>  zlib = cc.find_library('z')
>>>>  
>>>> +if cc.links('''
>>>> +#include <stdint.h>
>>>> +int main(void) {
>>>> +  uint32_t x32 = 0;
>>>> +  uint64_t x64 = 0;
>>>> +  __atomic_load_n(&x32, __ATOMIC_SEQ_CST);
>>>> +  __atomic_load_n(&x64, __ATOMIC_SEQ_CST);
>>>
>>> See my reply for v2. I've looked into this a little bit more and it
>>> looks like __atomic_* functions are a GCC implementation detail. OIn
>>> other words, the C11 standard [1] defines only atomic_* functions, and
>>> GCC implements them with __atomic_* builtins when the platform supports
>>> it, but other compilers might not expose those builtins and still
>>> support atomic_* functions without them. This also seems to be what [2]
>>> explains:
>>>
>>>> The first set of library functions are named __atomic_*. This set has
>>>> been “standardized” by GCC, and is described below. (See also GCC’s
>>>> documentation)
>>>
>>> (Notice the quotes around “standardized”, meaning they are a GCC
>>> extension)
>>
>> Quite, and while the stdatomic.h API is part of the C11 standard,
>> libatomic is part of GCC.  So this test is to determine whether
>> linking against GCC's libatomic.so is needed for its __atomic_*
>> fallback implementation.
>>
>> It raises the question of what to do with other compilers, but
>> igt has other build errors with clang on mips at the moment.
>> With a quick search, it looks like its __atomic_* functions are
>> part of libclang.so for clang.
> 
> I don't see anything in `readelf -s /usr/lib/libclang.so.8`.

Yes, well I did this:

$ for f in $(find . -name "*.so"); do strings $f | grep __atomic_load && echo $f; done
__atomic_load
__atomic_load_1
__atomic_load_2
__atomic_load_4
__atomic_load_8
./gcc/mips-linux-gnu/8/libatomic.so
__atomic_load
__atomic_load_1
__atomic_load_2
__atomic_load_4
__atomic_load_8
__atomic_load_16
./mips-linux-gnu/libLLVM-7.so

although it's true that they don't appear as proper symbols with
readelf.  It would take a bit more investigation in the LLVM
source code to get to the bottom of that, but I don't think it's
necessary to solve the problem at hand.

>> Maybe this test should only be used when the compiler name is
>> gcc?  In practice it does work with both gcc and clang though, as
>> they both use the same naming convention for atomic built-ins.
> 
> Hmm. I'm still not quite sure I understand why checking with __atomic_*
> is preferred.
> 
> - If the compiler has __atomic_* builtins: this won't link with
>   libatomic
> - If the compiler doesn't have __atomic_* builtins: this will link with
>   libatomic even if stdatomic.h works without it
> 
> What we're really interested in is stdatomic.h support, not __atomic_*.
> So I still think checking for atomic_* is better than __atomic_*. Am I
> missing something?

I think the issue is that there is no absolute relationship
between stdatomic.h and the __atomic_* functions.  So the test is
currently designed from libatomic's point of view, and it might
add libatomic dependency even if stdatomic.h doesn't use the
__atomic_* functions.  Then conversely, using the C11 atomic_*
instead in the test means that we would add dependency on
libatomic if it fails to link without being completely sure that
it is the missing library.

If you take the current test on its own, it doesn't claim to
cover stdatomic.h support but just libatomic dependency.  So
that's why I prefer it.

But in practice, both variants (__atomic_* and C11 atomic_*) can
be used in the test with existing versions of GCC and I'm not
trying to cover Clang MIPS builds in this series.  I think
there's no perfect solution here, and if you still think it makes
more sense to use the C11 atomic_* functions then fine, I can
change the test to do that instead.

Guillaume

>>> [1]: http://www.open-std.org/jtc1/sc22/wg14/www/docs/n1548.pdf
>>> [2]: https://llvm.org/docs/Atomics.html
>>>
>>>> +  return 0;
>>>> +}''', name : 'built-in atomics')
>>>> +	libatomic = null_dep
>>>> +else
>>>> +	libatomic = cc.find_library('atomic')
>>>> +endif
>>>> +
>>>>  if cc.has_header('linux/kd.h')
>>>>  	config.set('HAVE_LINUX_KD_H', 1)
>>>>  endif