Обсуждение: [HACKERS] Do we need the gcc feature "__builtin_expect" to promote the branches prediction?

On 02/06/2017 12:50, Craig Ringer wrote:
> 
> 
> On 2 Jun. 2017 16:42, "Hao Lee" <mixtrue@gmail.com
> <mailto:mixtrue@gmail.com>> wrote:
> 
>     Hi all, 
>            There is a lot of "if statement" in system, and GCC provides
>     a feature,"__builtin_expect", which  let compilers know which branch
>     is mostly run.
> 
> 
> Compilers and CPUs are really good at guessing this.
> 
> Humans are wrong about it more than we'd like too.

+1

> 
> It's surprisingly rarely s good idea to use branch prediction hints.
> 
> See the vsrious Linux kernel discussions about this.
> 
> If you find concrete sites of frequent or costly branch mis-prediction
> please point them out, with benchmarks.

And also see this thread:
https://www.postgresql.org/message-id/CAKJS1f8yqRW3qx2CO9r4bqqvA2Vx68=3awbh8CJWTP9zXeoHMw@mail.gmail.com

BTW Andres added support for likely/unlikely in aa3ca5e3dd6, but its
usage is still really limited.

-- 
Julien Rouhaud
http://dalibo.com - http://dalibo.org

Hi,

On 2017-06-02 16:40:56 +0800, Hao Lee wrote:
> Hi all,
>        There is a lot of "if statement" in system, and GCC provides a
> feature,"__builtin_expect", which  let compilers know which branch is
> mostly run. as we known, miss-prediction will lead the performance
> lost(because the CPU will thrown away some instructions, and re-fetch some
> new instructions). so that we can tell GCC how produce more efficient code.
> for example as following.
> It will gain performance promotion i think. As i know, the in Linux kernel,
> this feature is also applied already.
> 
>  #define likely(cond)     __builtin_expect(cond,true)
> #define unlikely(cond)  __builtin_expect(cond,false)
> 
> if (likely(cond)) {
> //most likely run.
>    xxxx
> } else //otherwise.
> {
>    xxxx
> }

We already do this in a few cases, that are performance critical enough
to matter.  But in most cases the CPUs branch predictor does a good
enough job on its own.

- Andres

On 6 June 2017 at 10:44, Hao Lee <mixtrue@gmail.com> wrote:
> and another things come into my mind, in some scenario, as we know that the
> native library is not the most effective way to do that, such as, allocation
> a large amount of memories by using "alloc()"... and "memmove()", so on. As
> the SIMD instruction became the standard in CPU, therefore, we can that to
> do something more effectively. for example, using SIMD to impl
> 'alloc',"memmove" and so on, or the other database operations. and it seems
> that there some one have done this job.

Compilers are already pretty good at this.

https://en.wikipedia.org/wiki/Program_optimization


-- Craig Ringer                   http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services