Re: [PATCH v5 00/18] Rework READ_ONCE() to improve codegen

[Peter Zijlstra]

On Thu, May 14, 2020 at 04:20:42PM +0200, Marco Elver wrote:
> On Thu, 14 May 2020 at 16:13, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >
> > On Wed, May 13, 2020 at 03:58:30PM +0200, Marco Elver wrote:
> > > On Wed, 13 May 2020 at 15:24, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >
> > > > Also, could not this compiler instrumentation live as a kernel specific
> > > > GCC-plugin instead of being part of GCC proper? Because in that case,
> > > > we'd have much better control over it.
> > >
> > > I'd like it if we could make it a GCC-plugin for GCC, but how? I don't
> > > see a way to affect TSAN instrumentation. FWIW Clang already has
> > > distinguish-volatile support (unreleased Clang 11).
> >
> > Ah, I figured not use the built-in TSAN at all, do a complete
> > replacement of the instrumentation with a plugin. AFAIU plugins are able
> > to emit instrumentation, but this isn't something I know a lot about.
> 
> Interesting option. But it will likely not solve the no_sanitize and
> inlining problem, because those are deeply tied to the optimization
> pipelines.

So I'm imagining adding the instrumentation is done at a very late pass,
after all, all we want is to add instrumentation to any memops. I
imagine this is done right before doing register allocation and emitting
asm.

At this point we can look if the current function has a no_sanitize
attribute, no?

That is, this is done after all the optimization and inlining stages
anyway; why would we care about that?

Maybe I'm too naive of compiler internals; this really isn't my area :-)