Re: [PATCH RFC 0/3] Static calls

[Ingo Molnar]

* Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:

> On Fri, Nov 09, 2018 at 08:28:11AM +0100, Ingo Molnar wrote:
> > > - I'm not sure about the objtool approach.  Objtool is (currently)
> > >   x86-64 only, which means we have to use the "unoptimized" version
> > >   everywhere else.  I may experiment with a GCC plugin instead.
> > 
> > I'd prefer the objtool approach. It's a pretty reliable first-principles 
> > approach while GCC plugin would have to be replicated for Clang and any 
> > other compilers, etc.
> 
> The benefit of a plugin is that we'd only need two of them: GCC and
> Clang.  And presumably, they'd share a lot of code.
> 
> The prospect of porting objtool to all architectures is going to be much
> more of a daunting task (though we are at least already considering it
> for some arches).

Which architectures would benefit from ORC support the most?

I really think that hard reliance on GCC plugins is foolish - but maybe 
Clang's plugin infrastructure is a guarantee that it remains a sane and 
usable interface.

> > I'd be very happy with a demonstrated paravirt optimization already - 
> > i.e. seeing the before/after effect on the vmlinux with an x86 distro 
> > config.
> > 
> > All major Linux distributions enable CONFIG_PARAVIRT=y and 
> > CONFIG_PARAVIRT_XXL=y on x86 at the moment, so optimizing it away as much 
> > as possible in the 99.999% cases where it's not used is a primary 
> > concern.
> 
> For paravirt, I was thinking of it as more of a cleanup than an
> optimization.  The paravirt patching code already replaces indirect
> branches with direct ones -- see paravirt_patch_default().
> 
> Though it *would* reduce the instruction footprint a bit, as the 7-byte
> indirect calls (later patched to 5-byte direct + 2-byte nop) would
> instead be 5-byte direct calls to begin with.

Yes.

> > All other usecases are bonus, but it would certainly be interesting to 
> > investigate the impact of using these APIs for tracing: that too is a 
> > feature enabled everywhere but utilized only by a small fraction of Linux 
> > users - so literally every single cycle or instruction saved or hot-path 
> > shortened is a major win.
> 
> With retpolines, and with tracepoints enabled, it's definitely a major
> win.  Steve measured an 8.9% general slowdown on hackbench caused by
> retpolines.

How much of that slowdown is reversed?

> But with tracepoints disabled, I believe static jumps are used, which
> already minimizes the impact on hot paths.

Yeah.

Thanks,

	Ing