<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/146145>146145</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            [libc++] Make sure that we use the fast platform wait for all eligible types in atomic_wait
        </td>
    </tr>

    <tr>
      <th>Labels</th>
      <td>
            libc++
      </td>
    </tr>

    <tr>
      <th>Assignees</th>
      <td>
            huixie90
      </td>
    </tr>

    <tr>
      <th>Reporter</th>
      <td>
          ldionne
      </td>
    </tr>
</table>

<pre>
    While discussing about https://github.com/llvm/llvm-project/issues/109290 with @huixie90 and Apple folks, we noticed that we didn't use the underlying platform's API (e.g. `__ulock_wait` on Apple, `futex` on Linux) even in cases where we could. This is caused by the brittle way in which we distinguish between the platform-native contention state and other states (e.g. [here](https://github.com/llvm/llvm-project/blob/52040b44f6060540a0b9d56fdd2e0eb5a540e84c/libcxx/src/atomic.cpp#L191)):

```c++
_LIBCPP_EXPORTED_FROM_ABI void __cxx_atomic_notify_one(__cxx_atomic_contention_t const volatile* __location) noexcept; // fast
_LIBCPP_EXPORTED_FROM_ABI void __cxx_atomic_notify_one(void const volatile* __location) noexcept; // slow
```

We basically forward only atomics which are *exactly* of the right type to the fast path, and everything else gets the slow path. For example, on Apple `__cxx_atomic_contention_t` is `int64_t`. That means that if we try to wait on a `std::atomic<uint64_t>`, we'll end up in the slow path even though we could clearly use the underlying platform's wait mechanism, which is faster.

To resolve this, I would suggest re-thinking the ABI boundary of the synchronization library and basically forwarding (from the headers) any type that has a natural alignment and `sizeof(T) == {32,64}` to the underlying platform API. At the end of the day, we basically hand over a bag of bytes to the underlying platform wait, the value and the type of these bytes doesn't matter.
</pre>
<img width="1" height="1" alt="" src="http://email.email.llvm.org/o/eJykVU2L4zgQ_TXKpZigyHYSH3JIOh1o6GGaoWH2ZmSrbGtbkYIkJ_H8-qXkZLqH_WJ3QGBkSVWlV-89yRB0ZxE3rNgxIfpBXzWWnAnBiv1MDrF3fmOUdtbirHZq3HzrtUFQOjRDCNp2IGs3ROhjPAWWbZk4MHHodOyHet64IxMHY873z6eTd79jE5k46BAGDEwcFrwUJYeLjj2wnN9LAGkVbE8ng9A68xaYeIALgnVRN6gg9jLSXGllmVhFGAJC7BEGq9CbkSo7GRlb549MrAJsX56AiTXOuzmwJa-qwbjmrbpIHdmSg7NTMkrDlrwdIl5v_5-1Ha5MlIBntKAtNDJggEuPHqmExg1GzeG11wF0gEYOARXUYyqn9jpGg3CRIx299Lrpp7pD1LYbdOihxnhBtGn_veZPVkZ9puA2oo3aWQhRRkywuNijn-bh_VLFjipixZ6J9X9uR21czcShEDzndZ63S77kRc4lr0tVLFulBHKsC1nkHNd5QwF03VyvTByCp6mM7qibeXM6MZE9L8oFEyWNbMt4Gks-jYaJHQ2-rZ6fdg8vL9Xjby9fvr4-7qvD1y-fq-3uCc5OK6iq5nqtprgV9b0dK2eRifVPK-8IVZHgChHOzsioqZlbqCrjGknL1ELr8NrgKbJsBxM20MoQf6WYtPw_8gbjLh9xmWD6hlDLoBtpzAit8xfpFThrRphyhxuFpEdgYotX2UQzUkLXJgJ53fUR4nhCiC79oQvCScaeuE30wTP6MfYkETQBocMY0k4qKe2cw8F5wKs83iRxl8cknb9Bn_SiA23RNi7z9IN0ISMcUdowiVa3xP_oR6qP5EfBJZ0KURFdsu0Um2UPwz1Q9kgIJQtgYmUMoFUwnEhSPxU-aTT2buj6H9qExqD0ZvxXj0jVHLHppdXhmNIlsHVIIKKfT016deAxOHOmcDpZ0xNcUqowdB2GCB4_EcJvlINSEpFqN1gl_XhvVRht03tn9ffEEzC69rRMPfoTCSgQE-vWu2M63KNU6AORS9rx1nDCt5cBJFgZBy8NSKM7e0QbU1QCWX9H1zKxfqWjLNuzbA9stcsEEw_LnK321MYbdf4CKDLSOWxjWqcu3C6j5Hiz6PfS-2RWZ_QgoZYdba1H8qx_CJ8MWTyk9bM0w-R4NEtXnLIFvAVSDsPk_0cZqUEztclUmZVyhpvFqljwQqxW61m_KQVf5YtCKWzKTNZixddtWWZNu17lrazLmd4ILgq-FKtFmYliPV80ol4WnOclLlpVLFjO8Si1mZN1zp3vZukN2yzy5SIvZkbWaMLtISVzvPnc9JT6TTLceugCy7nRIYb3OFFHk57gD8eKPXyWbwhh8PjjtbsTeJL0R8iIJiBJGEZ3ujYTXIEEclMq7ZoN3mx-4amebnreiD8CAAD__7WuwCQ">