Re: [PATCH] x86/resctrl: avoid compiler optimization in __resctrl_sched_in

[Segher Boessenkool]

Hi!

On Tue, Mar 07, 2023 at 12:35:45PM +0100, Peter Zijlstra wrote:
> So per that summary, I'm going to nit-pick and state we very much want
> CSE. CSE good. What we don't want it violating store-load ordering.

So you need to describe exactly what you *do* want.  There is no way to
forbid most otherwise valid things.  But you can express pretty much all
dependencies.

> Oh, geez, what a twisty tale that... So Linus knew back in '09 that "p"
> was icky, but it sorta was the only thing and it 'worked' -- until now
> :/

The "p" constraint is just like any other address_constraint, in most
aspects.  Since this is very specific to "p", that limits what is going
on to really just one thing.

For "p", after reload, strict_memory_address_addr_space_p is used.  That
is, targetm.addr_space.legitimate_address_p with strict set to true.  On
x86 that limits what registers can be used?  So I guess that made things
accidentally work before?

> So ideally we'd get something that respects the whole store-load
> ordering but still allows agressive CSE. And works for both toolchains.
> Small ask, I know :-)

Well, what is the ordering you need respected, *exactly*?


Segher