Re: luajit 2.0b10 performance extremely brittle?
- From: demetri <demetri.spanos@xxxxxxxxx>
- To: luajit@xxxxxxxxxxxxx
- Date: Mon, 4 Jun 2012 08:48:47 -0700
I (think I) have a decent idea of how to write code that the JIT will like,
but the name-dependent performance thing was quite a surprise.
Is there a good way to know that one's code is not falling into whatever
memory-layout trap Mike mentioned? If there's a 25%+ differential to be had
in a tight loop by optimizing some names, that would potentially be worth
tracking down for some of my work.
Or should I interpret Mike's comments about broader inefficiencies in that
example to mean that, with suitably optimized code/data structures, the
naming problem doesn't arise?
Any thoughts would be appreciated :)
On Mon, Jun 4, 2012 at 8:27 AM, Francesco Abbate <francesco.bbt@xxxxxxxxx>wrote:
> Hi,
>
> I've actually some experience writing optimized code for LuaJIT2.
> Usually you have to respect some rules to ensure that JIT code is
> actually generated and it does not fallback to interpreter mode.
>
> Normally what you should avoid is:
> - do not use global variables inside loops
> * use local declarations to hold the variables
> * use the 'strict' module to detect unintentional use of global variables
> - use an imperative style of coding:
> * smart functional code tend to spoil optimization if closures have
> to be createad inside a loop
> - avoid objects creation inside a loop
> - do not use deeply nested tables to hold the data used inside loops
> - write clean simple code
>
> If you respect these rules LuaJIT often make a very good job.
>
> It is somewhat inconvenient that sometimes "reasonable" code does not
> get compiled at all but this is the way luajit works. It is not like a
> C compiler that always produce fast code even if the code is slightly
> suboptimal.
>
> Francesco
>
>
Other related posts:
