Re: [GIT PULL] slab fixes for 3.2-rc4

[Christoph Lameter <cl@linux.com>]

On Wed, 4 Jan 2012, Linus Torvalds wrote:

> On Wed, Jan 4, 2012 at 9:00 AM, Christoph Lameter <cl@linux.com> wrote:
> >
> > xadd is 3 cycles. add is one cycle.
>
> On some uarchs. On new uarchs it can be a single cycle, I think, and
> on some uarchs it will even be microcoded and/or only go in one pipe
> because it has that odd behavior that it writes both to memory and a
> register, and thus doesn't fit the normal fastpaths.

> The point is, xadd isn't actually any faster than just doing the
> regular "add and read". It's *slower*.

Ok that assumes that both add and read are this_cpu operations that
use the segment override to relocate the address. I thought you wanted to
open code everything. So then we would have to use the following?

	this_cpu_add(var, count);
	result = this_cpu_read(var);


XADD and ADD have the same cycle count if the ADD is used to add to a
memory location. Both use 4 microops.

The MOVE from memory costs another 2 so we are at 6. Two segment overrides
for each instruction add 2 more (nothing in the optimization manual but as
far as I can recall another uop is needed for the address calculation). So
this ends up at 8.

On the other hand

	result = this_cpu_add_return(var, count)

is 4 + 1.