Subject: Re: performance impact of branch prediction?
To: Hubert Feyrer <hubert@feyrer.de>
From: Nathan J. Williams <nathanw@wasabisystems.com>
List: tech-kern
Date: 04/21/2006 14:26:18
Hubert Feyrer <hubert@feyrer.de> writes:

> Branch prediction is used in many parts throught the kernel, e.g.
> 
> subr_pool.c:    if (__predict_false((pc->pc_pool->pr_flags & PR_WANTED) != 0)) {
> subr_pool.c-            goto destruct;
> subr_pool.c-    }
> 
> Does anyone know the performance impact of this prediction?
> Were any kind of measurements done? What were the results?

Note that __predict_false() isn't really about the microarchitectural
concept of branch prediction; it's more about moving code out-of-line
to reduce the memory footprint of the common path. That is, the
following code:

A
if (B)
        C
D

would typically assemble to something like:

  A
  evaluate B
  if false, goto 1f:
  C
1f:
  D

changing (B) to __predict_false(B) makes it likely to be rearranged
as:

  A
  evaluate B
  if true, goto 2f:
1f:
  D
  ...
  return
2f:
  C
  goto 1f

I doubt it helps much when the wrapped code is a single goto. 

At the microarchitectural level, dynamic branch prediction is
considered more useful these days than static compiler-generated
branch prediction anyway.

        - Nathan