Re: [PATCH] libstdc++: Remove some more unconditional uses of atomics

[Jonathan Wakely <jwakely@redhat.com>]

On Thu, 14 Sept 2023 at 09:41, Christophe Lyon
<christophe.lyon@linaro.org> wrote:
>
>
>
> On Thu, 14 Sept 2023 at 10:17, Jonathan Wakely <jwakely@redhat.com> wrote:
>>
>> On Thu, 14 Sept 2023 at 08:44, Christophe Lyon
>> <christophe.lyon@linaro.org> wrote:
>> >
>> > Hi,
>> >
>> >
>> > On Wed, 13 Sept 2023 at 14:32, Jonathan Wakely <jwakely@redhat.com> wrote:
>> >>
>> >> Tested x86_64-linux and aarch64-linux. I intend to push this to trunk.
>> >>
>> >> -- >8 --
>> >>
>> >> These atomics cause linker errors on arm4t where __sync_synchronize is
>> >> not defined. For single-threaded targets we don't need the atomics.
>> >>
>> >
>> > I ran the tests on arm-eabi default config (so, armv4t) with this patch, and here is the list of remaining UNRESOLVED tests:
>> >  29_atomics/atomic/compare_exchange_padding.cc
>> > 29_atomics/atomic/cons/value_init.cc
>> > 29_atomics/atomic_float/value_init.cc
>> > 29_atomics/atomic_integral/cons/value_init.cc
>> > 29_atomics/atomic_ref/compare_exchange_padding.cc
>> > 29_atomics/atomic_ref/generic.cc
>> > 29_atomics/atomic_ref/integral.cc
>> > 29_atomics/atomic_ref/pointer.cc
>> > experimental/polymorphic_allocator/construct_pair.cc
>> >
>> > all of them are due to undefined reference to __sync_synchronize
>> > (some also reference __atomic_compare_exchange_4, etc...)
>> >
>> >
>> > IIUC, this should not be the case for experimental/polymorphic_allocator/construct_pair.cc ?
>> > The reference for __sync_synchronize is near the beginning of test0[123]
>> > from a call to __atomic_load_n line 835 of atomic_base.h
>> > not sure where it comes from, the .loc directive indicates line 28 of the testcase which is the opening brace
>>
>> Doh, I removed the atomics from <memory_resource> but this is
>> <experimental/memory_resource>, which has a separate implementation.
>>
>> I'll make a change to <experimental/memory_resource> as well, thanks
>> for catching my silly mistake.
>>
>
> You're welcome.
> So I'll shrink my patch and add dg-require-thread-fence only to the few 29_atomics tests listed above.

Great, thanks. That's approved for trunk then.

N.B. if you'd prefer to add { dg-require-effective-target thread_fence
} instead of { dg-require-thread-fence "" } then that's fine, just
note that the effective target uses an underscore not a hyphen. The
dg-require-thread-fence proc just uses the proc that checks the
thread_fence effective target, so both forms do the same thing.