Re: [PATCH 0/2] rbree: inline rb_first() and rb_last()

[Kuan-Wei Chiu <visitorckw@gmail.com>]

On Sun, Nov 16, 2025 at 10:41:31AM -0800, Eric Dumazet wrote:
> On Sun, Nov 16, 2025 at 10:00 AM Kuan-Wei Chiu <visitorckw@gmail.com> wrote:
> >
> > Hi Eric,
> >
> > On Fri, Nov 14, 2025 at 02:06:44PM +0000, Eric Dumazet wrote:
> > > Inline these two small helpers, heavily used in TCP and FQ packet scheduler,
> > > and in many other places.
> > >
> > > This reduces kernel text size, and brings an 1.5 % improvement on network
> > > TCP stress test.
> >
> > Thanks for the patch!
> >
> > Just out of curiosity, do you think rb_first() and rb_last() would be
> > worth marking with __always_inline?
> 
> I have not seen any difference, what compilers are you using that
> would not inline this ?

I haven't specifically measured the difference. I was curious if using
__always_inline would be better to guarantee the compiler inlines the
function, given that the inline keyword is only a suggestion, and this
optimization is important for performance in a hot path.

Also, FWIW, I tried building an arm64 defconfig with LLVM both with and
without __always_inline. I got the same results from size vmlinux, but
scripts/bloat-o-meter output is as follows:

add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-8 (-8)
Function                                     old     new   delta
__aarch32_sigret_code_end                     24      16      -8
Total: Before=129127208516005098430, After=129127208516005098422, chg -0.00%

Regards,
Kuan-Wei