Re: Control Dependencies vs C Compilers

[Peter Zijlstra]

On Tue, Oct 06, 2020 at 10:23:24AM -0400, stern@xxxxxxxxxxxxxxxxxxx wrote:
> On Tue, Oct 06, 2020 at 03:31:15PM +0200, Peter Zijlstra wrote:

> > Only if we get the compiler people on board and have them provide means
> > are we guaranteed safe from the optimizer. Otherwise we'll just keep
> > playing whack-a-mole with fancy new optimization techniques. And given
> > how horridly painful it is to debug memory ordering problems, I feel it
> > is best to make sure we're not going to have to more than necessary.
> 
> Given that you would have to add a compiler annotation, isn't it just as 
> easy to use READ_ONCE and WRITE_ONCE?
> 
> Or are you worried that even with READ_ONCE and WRITE_ONCE, the compiler 
> might still somehow defeat the control dependency?

Yes.

Also, not all instances actually have the WRITE_ONCE() on. The one in
the perf ringbuffer for example uses local_read() for the load (which is
basically READ_ONCE()), but doesn't have WRITE_ONCE() on the inside.

Also, afaiu WRITE_ONCE() also doesn't stop the compiler from lifting it
if it thinks it can get away with it -- memory-barriers.txt has
examples.

And then there's the case where the common branch has a store, I know
ARM64 ARM isn't clear on that, but I'm thinking any actual hardware will
still have to respect it, and it's a matter of 'fixing' the ARM.


Mostly I just want the compiler people to say they'll guarantee the
behaviour if we do 'X'. If 'X' happens to be 'any dynamic branch headed
by a volatile load' that's fine by me.

If they want a new keyword or attribute, that's also fine. But I want to
have the compiler people tell me what they want and guarantee they'll
not come and wreck things.