On 1 March 2025 21:06:22 UTC, Jonathan Wakely <jwakely.gcc@xxxxxxxxx> wrote: >> > > It turns out these calls could be replaced with a compiler built-in (__atomic_compare_exchange_n) both for clang and gcc, which wouldn't have the above problem, and would also allow for the removal of 'stdatomic.h'. However, this didn't feel too in keeping with manpage example code. >> > >> > Well the example already relies on Linux-specific details of >> > syscall(2) and futex(2), and only GCC-compatible compilers are really >> > usable on Linux. So the GCC extensions could probably be assumed to >> > exist. >> > >> > If you were writing pure C11 code intended to be portable, you >> > wouldn't be using SYS_futex anyway. >> >> *Iff* C11 atomics were something we understand, maybe even just having >> a working example that uses it would be worth it, even if Linux-only >> stuff can do the same. > >I understand the C11 atomics fine, I don't understand how they >interact with syscall(2). > >> >> However... >> >> > > The gcc docs are here: >> > > <https://gcc.gnu.org/onlinedocs/gcc/_005f_005fatomic-Builtins.html> >> > >> > The patch assumes that you can pass _Atomic int* to the futex syscall, >> > but I'm not sure if that's correct. The syscall expects an int. >> >> Hmmm. >> >> > >> > The C standard says "NOTE The representation of atomic integer types >> > need not have the same size as their corresponding regular types." >> > >> > With GCC and Clang, _Atomic uint32_t does have the same representation >> > as unsigned, so it should work. I'm not an expert here though. >> >> Considering that the example has been broken for a long time, and I've >> asked for help in the glibc-help@ mailing list and nobody answered, and >> you don't feel comfortable with it either, the best course of action is >> to revert that commit. > >The __sync_xxx built-ins are documented as legacy and not to be used >in new code: >https://gcc.gnu.org/onlinedocs/gcc/_005f_005fsync-Builtins.html >So while reverting it seems better than having broken examples, it >would be better to not use the legacy APIs. Ok, so we should probably improve the situation rather than just revert. Perhaps using __atomic_compare_exchange_n would have been a more suitable approach than I had earlier realised.
Attachment:
signature.asc
Description: PGP signature
