Re: Alternative implementation of the generic __ffs

[Matti Aarnio <matti.aarnio@zmailer.org>]

On Sun, Apr 20, 2008 at 10:42:21AM +0200, Alexander van Heukelum wrote:
> On Sat, 19 Apr 2008 20:06:57 -0700, "Joe Perches" <joe@perches.com>
> said:
> > On Sun, 2008-04-20 at 01:29 +0300, Matti Aarnio wrote:
> > > I am curious, why not take the code already in glibc ffs() for ARM ?
> > > That is, if the ffs() is all that important detail in kernel ?
> 
> Hi,
> 
> The glibc version is based on a table-lookup. This makes it
> behave differently in hot and cold cache situations. That's
> fine if __ffs is used in tight loops, but in the kernel such
> use of __ffs is avoided because it might be slow. I added it
> to the benchmark, but it would need testing for the cold
> cache case too.
> 
> As for the importance of __ffs in the kernel: as far as I
> know the hot-spots in the kernel using __ffs are the
> schedular (sched_find_first_bit) and the cpu mask walking
> code (for_each_cpu_mask).

Perhaps those hot-spots would benefit from more broadly
accelerable algorithms.   ARM architecture v5 introduced
a CLZ instruction -- Count Leading Zeroes.

Well, gcc's  __builtin_ffs() for ARM Arch5 and up (including
XScale) does things in a bit more interesting way:

  http://mail-index.netbsd.org/port-arm/2002/08/20/0001.html

$ cat try.c
int foo(int i)
{
        return __builtin_ffs(i);
}
$ arm-gp2x-linux-gcc -S -O -march=armv5 try.c 
$ more try.s 
        .file   "try.c"
        .text
        .align  2
        .global foo
        .type   foo, %function
foo:
        @ args = 0, pretend = 0, frame = 0
        @ frame_needed = 0, uses_anonymous_args = 0
        @ link register save eliminated.
        @ lr needed for prologue
        rsb     r3, r0, #0
        and     r3, r3, r0
        clz     r3, r3
        rsb     r0, r3, #32
        bx      lr
        .size   foo, .-foo
        .ident  "GCC: (GNU) 4.1.2 (Fedora GP2X 4.1.2-8.fc9)"


> Greetings,
>     Alexander

/Matti Aarnio