Re: 2.6.7-rc3-mm1

[Andrew Morton <akpm@osdl.org>]

Paul Jackson <pj@sgi.com> wrote:
>
> Do you recall why your i386-uninline-bitops.patch moves i386
>  find_next_bit() and find_next_zero_bit() out of line, but not
>  find_first_zero_bit() nor find_first_bit()?

They're the two non-leaf functions - they expand other inlines and end up
quite big.

>  Perhaps someone else has further insight to the tradeoffs here, such as
>  a 'recommended size', above which most routines should be not inlined,
>  except in special cases.

Hard call.  Lots of hand-waving is involved.

Yes, an aggregate reduction in kernel text size is a good thing, but the
main reason for uninlining things is for performance: reduction of icache
footprint.

If an inline function is expanded several times in, say, fs/dcache.c then
it's a good candidate for uninlining, because it's probably the case that
all the expanded versions are in icache simultaneously.  But if a function
is expanded once in ext2 and once in ext3 then it's less useful to uninline
it, because it is rare that two different filesystem drivers are in use
simultaneously.