On Thu, 07 Dec 2006 20:06:39 +0000
David Howells <[EMAIL PROTECTED]> wrote:

> Andrew Morton <[EMAIL PROTECTED]> wrote:
> 
> > and we can assume (and ensure) that a failing test_and_set_bit() will not
> > write to the affected word at all.
> 
> You may not assume that; and indeed that is not so in the generic
> spinlock-based bitops or ARM pre-v6 or PA-RISC or sparc32 or ...

Ah.  How obnoxious of them.

> Remember: if you have to put a conditional jump in there, it's going to fail
> one way or the other a certain percentage of the time, and that's going to
> cause a pipeline stall, and these ops are used quite a lot.
> 
> OTOH, I don't know that the stall would be that bad since the spin_lock and
> spin_unlock may cause a stall anyway.
> 

Yes, the branch would cost.  But in not uncommon cases that branch will save
the machine from dirtying a cacheline.

And if we add those branches, we bring those architectures' semantics in
line with all the other architectures.  And we get better semantics
overall.

So I don't think we should rule this out.
-
To unsubscribe from this list: send the line "unsubscribe linux-arch" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Reply via email to