Re: [PATCH 1/3] softirq: Use ffs in __do_softirq

[Frederic Weisbecker <fweisbec@gmail.com>]

On Mon, Dec 09, 2013 at 10:56:46AM -0800, Joe Perches wrote:
> On Mon, 2013-12-09 at 19:44 +0100, Frederic Weisbecker wrote:
> > On Sun, Nov 17, 2013 at 01:55:10AM -0800, Joe Perches wrote:
> > > Possible speed improvement of the __do_softirq function by using ffs
> > > instead of using a while loop with an & 1 test then single bit shift.
> []
> > Perhaps using for_each_set_bit() would simplify that more?
> 
> It might simplify the appearance of the code but it
> would/could expand the amount of generated code because
> for_each_set_bit uses an address_of(unsigned long) and
> the value tested is an unsigned int.
> 
> extra dereferences, can't be in a register, etc...

I'm not sure that would matter that much. But yeah it appears that find_first_bit/find_next_bit
aren't even overriden in x86. So they are function calls. Although I guess that most
of the time only one softirq is pending at a time.

But anyway perhaps we want that path to stay very optimized, so you're
patch look ok.

Thanks.