[Bug target/108659] Suboptimal 128 bit atomics codegen on AArch64 and x64

["wilco at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108659

--- Comment #8 from Wilco <wilco at gcc dot gnu.org> ---
(In reply to Niall Douglas from comment #7)
> (In reply to Andrew Pinski from comment #4)
> > (In reply to Niall Douglas from comment #3) 
> > > You may be interested in reading https://reviews.llvm.org/D110069. It wanted
> > > to have LLVM generate a 128 bit AArch64 CAS for atomics. LLVM merged that
> > > change, it'll be in the next release.
> > 
> > Using CAS for atomic load is not valid thing to do ...
> > Because atomic load from constant rodata needs to work.
> > LLVM breaks this case as they don't care about it. GCC does though.
> 
> I've heard that argument before, and I've always wondered why _Atomic128
> types couldn't have an attribute which applies attribute section to their
> static const variable incarnations to force them into r/w memory. That would
> also solve the LLVM issue. Said attribute is not unuseful in general
> actually, it would help avoid having to mess with mprotect to apply copy on
> write perms on regions in .rodata when you need to modify static const
> variable values.
> 
> I don't think that the standard *guarantees* that static const variables go
> into read only memory, and besides, before C23 128 bit integers weren't
> supported anyway so one could argue as a proprietary extension (__int128)
> you get proprietary special casing.

Yes that sounds like a reasonable approach. There will language lawyers that
say it must also work on mmap after mprotect of course, but that seems even
more unlikely in the real world...

I believe that the vast majority of developers just want 128-bit atomics to
work efficiently without locks when possible.

Currently various packages are forced to create 128-bit atomics using inline
assembler - and that seems a much worse hack than supporting lock-free atomics
in the compiler.