Re: [PATCH] seqlock: Use WRITE_ONCE() when updating sequence

[Florian Weimer <fweimer@redhat.com>]

* Paul E. McKenney:

> On Wed, Dec 18, 2024 at 08:56:07PM +0100, Florian Weimer wrote:
>> * Peter Zijlstra:
>> 
>> > +linux-toolchains
>> >
>> > On Wed, Dec 18, 2024 at 08:59:47AM -0800, Paul E. McKenney wrote:
>> >
>> >> > Perhaps something like: (*(volatile unsigned int *)&s->sequence)++; ?
>> >> > I'd have to check what the compiler makes of that.
>> >> > 
>> >> > /me mucks about with godbolt for a bit...
>> >> > 
>> >> > GCC doesn't optimize that, but Clang does.
>> >> > 
>> >> > I would still very much refrain from making this change until both
>> >> > compilers can generate sane code for it.
>> >> 
>> >> Is GCC on track to do this, or do we need to encourage them?
>> >
>> > I have no clue; probably wise to offer encouragement.
>> 
>> What do you consider sane code?
>
> Peter's "(*(volatile unsigned int *)&s->sequence)++;" qualifies as sane.

I think the reference was originally to machine code.

>> Clang's choice to generate an incl instruction (on x86-64 at least) is a
>> bit surprising.  Curiously, the C11 abstract machine has a value-less
>> increment-in-place operation, so it's probably not in violation of the
>> volatile rules.  (C doesn't specify x++ in terms of ++x and x += 1.)
>
> Very good!  Should I do something like file a bug somewhere to help
> this along?

I don't know.  It seems that Clang/LLVM is cheating.  It's doing this
optimization even for

  i = i + 1;

with a volatile i.  That doesn't look like “strictly according to the
abstract machine” anymore.  A proper implementation would need explicit
representation of volatile increment/decrement in the IR.  Given that
volatile increment/decrement is deprecated, that seems quite a bit of
effort.

Thanks,
Florian