libcxx: Implement atomic::wait/wake using _umtx_op(2) ) for 64bit arches

[kib]

Summary

Only 64bit architectures can be supported this way, because libcxx
defines __cxx_contention_t to be int64_t for FreeBSD, and 32bit arches
do not have a kind of UMTX_OP_WAIT_INT64_PRIVATE operation.

LLVM review: https://reviews.llvm.org/D142134

Tested by: emaste

[markj]

contrib/llvm-project/libcxx/src/atomic.cpp

86

It looks like the __cxx_atomic_contention_t value is a int32_t on mips, and an int64_t otherwise. Is mips the reason for using the UINT variant?

[dim]

I'm fine with this, but we should upstream it. I'll have a look if upstream already has a regression test for the atomic stuff.

contrib/llvm-project/libcxx/src/atomic.cpp

86

Note that was committed from D26774 in rSd061adc48d54ebb91b1709b16c59e8ceca895be2, as a workaround for some older gcc. Maybe it could be dropped at some point, as this change is also not upstream.

[emaste]

Upstream review is https://reviews.llvm.org/D142134

There are upstream regression tests, see my comments in that review and in D128084. Some of the atomic tests were inconsistent (timing out, etc.) before in my CI work, but they've passed in all of my attempts with this change included.

[kib]

contrib/llvm-project/libcxx/src/atomic.cpp

86

No, I simply missed the fact that contention object is always 64bit. In other words, this thing cannot work on 32bit arches, without either ABI change for libcxx, or adding a new umtx op. The later seems to be hard because some 32bit arches (ppc) do not have neccessary atomics, not to hope for their copyin version.

It seems that the mips change was not up-streamed.

[emaste]

LLVM CI passed with this change applied https://buildkite.com/llvm-project/libcxx-ci/builds/17748

[arichardson]

Could 32-bit arches just cast to an int32 pointer and ignore the remaining bits of the contention object?

[arichardson]

In D38132#867351, @arichardson wrote:

Could 32-bit arches just cast to an int32 pointer and ignore the remaining bits of the contention object?

Never mind it looks like that would require changes to more functions.

[kib]

In D38132#867363, @arichardson wrote:

In D38132#867351, @arichardson wrote:

Could 32-bit arches just cast to an int32 pointer and ignore the remaining bits of the contention object?

Never mind it looks like that would require changes to more functions

It is not a problem to cast to int32. Issue is that UMTX_OP_WAIT (or any of its variants) only compares 32bit of the destination with the 32bits of the passed value, causing unexpected wait if other 32bit of the value do not match. This would result in the hang of the waiting thread.