Re: void* in Lua_State for Application?

[Tim Hill <drtimhill@...>]

On Sep 5, 2013, at 7:31 AM, Roberto Ierusalimschy <roberto@inf.puc-rio.br> wrote:

>> On Wed, Sep 4, 2013 at 8:33 PM, Tom N Harris <telliamed@whoopdedo.org> wrote:
>>> On Monday, September 02, 2013 03:14:08 PM Emil Renner Berthing wrote:
>>>> +LUA_API void  (lua_setstatedata) (lua_State *L, void *sd);
>>>> +LUA_API void *(lua_getstatedata) (lua_State *L);
>>> 
>>> What's the problem with storing a lightuserdata in the registry?
>> 
>> Performance, mostly. That's a table lookup worth of overhead for every
>> function that needs to fetch it. A C function accessing a field in the
>> lua_State is substantially faster.
> 
> Did you benchmark it for any real case (that is, a case that will waste
> some time doing something useful with the retrieved pointer)?
> 
> -- Roberto
> 

Yes I did. The C API is a mix of heavy-ish functions that do a  chunk of work, and some very light operation that just get/set values from a structure; it's these light ones that concerned me.

On an x86 system compiled for 64-bit code (cannot recall the CPU, but was recent pedigree), we saw about a 20-30% performance drop when using the registry compared to using a pointer "hidden" in a custom memory allocator for the light APIs. Unsurprisingly there was no measurable difference when the C API did a lot of work. On an ARM we saw about the same overhead, but there was a greater spread, which I was unable to account for (possibly a result of smaller CPU caches).

--Tim