Re: [PATCH 3/4] x86: open-code register save/restore in trace_hardirqs thunks

[Borislav Petkov <bp@alien8.de>]

On Sat, Jan 10, 2015 at 01:08:33PM -0800, Linus Torvalds wrote:
> It was true for some AMD CPU's in particular. One insn/cycle vs two.

Probably on K8: Agner Fog's insn tables show reciprocal throughput of
1/2 for MOV r64/m64 vs 1 for PUSH/POP.

> I personally would be very happy to go back to push/pop sequences.
> Even without a fancy stack engine like Intel has done for a while,
> even *simple* cores can generally pair pushes and pops. I think the

I think all the modern x86 machines have stack engines now :-)

> original Pentium already had a special magic pairing logic to pair
> pushes and pops despite both instructions using %esp. It's a common
> and fairly trivial special case, and the fact that a few AMD
> microarchitectures didn't do it is likely not really a good reason to
> avoid repeated push/pop instructions.

Well, according to the optimization manual, on F15h (Bulldozer and
later) PUSH/POP are faster than MOVs and on F16h (Jaguar and later) both
MOV and PUSH/POP have latency of 1, with MOV having a 1/2 throughput vs
PUSH/POP throughput of 1. So theoretically we can do 2 MOVs per cycle
there vs 1 PUSH/POP.

-- 
Regards/Gruss,
    Boris.

Sent from a fat crate under my desk. Formatting is fine.
--