Re: [PATCH v2] tools: memory-model: Make plain accesses carry dependencies

From: Boqun Feng
Date: Tue Dec 06 2022 - 15:47:15 EST


On Mon, Dec 05, 2022 at 11:18:13AM -0500, stern@xxxxxxxxxxxxxxxxxxx wrote:
> On Mon, Dec 05, 2022 at 01:42:46PM +0000, Jonas Oberhauser wrote:
> > > Besides, could you also explain a little bit why only "data;rfi" can be "carry-dep" but "ctrl;rfi" and "addr;rfi" cannot? I think it's because there are special cases when compilers can figure out a condition being true or an address being constant therefore break the dependency
> >
> > Oh, good question. A bit hard for me to write down the answer clearly
> > (which some people will claim that I don't understand it well myself,
> > but I beg to differ :) :( :) ).

Nah, I think your answer is clear to me ;-)

> >
> > In a nutshell, it's because x ->data [Plain] ->rfi y ->... z fulfils
> > the same role as storing something in a register and then using it in
> > a subsequent computation; x ->ctrl y ->... z (and ->addr) don't. So
> > it's not due to smart compilers, just the fact that the other two
> > cases seem unrelated to the problem being solved, and including them
> > might introduce some unsoundness (not that I have checked if they do).

So it's about whether a value can have a dataflow from x to y, right? In
that case registers and memory cells should be treated the same by
compilers, therefore we can extend the dependencies.
>
> More can be said here. Consider the following simple example:
>
> void P0(int *x, int *y)
> {
> int r1, r2;
> int a[10];
>
> r1 = READ_ONCE(*x);
> a[r1] = 1;
> r2 = a[r1];
> WRITE_ONCE(*y, r2);
> }
>
> There is an address dependency from the READ_ONCE to the plain store in
> a[r1]. Then there is an rfi and a data dependency to the WRITE_ONCE.
>
> But in this example, the WRITE_ONCE is _not_ ordered after the
> READ_ONCE, even though they are linked by (addr ; rfi ; data). The
> compiler knows that the value of r1 does not change between the two
> plain accesses, so it knows that it can optimize the code to be:
>
> r1 = READ_ONCE(*x);
> r2 = 1;
> WRITE_ONCE(*y, r2);
> a[r1] = r2;
>
> And then the CPU can execute the WRITE_ONCE before the READ_ONCE. This
> shows that (addr ; rfi) must not be included in the carry-deps relation.
>
> You may be able to come up with a similar argument for (ctrl ; rfi),
> although it might not be quite as clear.
>

Thank you, Alan! One question though, can a "smart" compiler optimize
out the case below, with the same logic?

void P0(int *x, int *y, int *a)
{
int r1, r2;

r1 = READ_ONCE(*x); // A

*a = r1 & 0xffff; // B

r2 = *a & 0xffff0000; // C

WRITE_ONCE(*y, r2); // D

}

I think we have A ->data B ->rfi C ->data D, however a "smart" compiler
can figure out that r2 is actually zero, right? And the code get
optimized to:

r1 = READ_ONCE(*x);
r2 = 0;
WRITE_ONCE(*y, r2);
*a = r1 & 0xffff;

and break the dependency.

I know that our memory model is actually unware of the differences of
syntatics dependencies vs semantics syntatics, so one may argue that in
the (data; rfi) example above the compiler optimization is outside the
scope of LKMM, but won't the same reasoning apply to the (addr; rfi)
example from you? The WRITE_ONCE() _syntatically_ depends on load of
a[r1], therefore even a "smart" compiler can figure out the value, LKMM
won't take that into consideration.

Am I missing something subtle here?

Regards,
Boqun

> Alan