Re: ways to avoid (or speed up) luaL_checkudata() in LuaJIT2?

[Duncan Cross <duncan.cross@...>]

On Sat, Nov 6, 2010 at 9:05 AM, Stefan Behnel <stefan_ml@behnel.de> wrote:
> Hi,
>
> I'm currently optimising an iteration wrapper that I have written for Lupa
> [1]. It basically maps Python's iterator protocol to the one in Lua, so that
> Lua programs can efficiently iterate over Python objects.
>
> To this end, I need to keep a reference to the Python iterator around, which
> I store as userdata in the Lua iterator state object (-> iterator, state,
> control variable) so that Lua garbage collects the Python iterator when the
> state object goes out of scope. To get the reference back at each iteration
> step, I use luaL_checkudata().
>
> Now, the callgrind profiler tells me that almost 25% of the time to advance
> the iterator is spent in the call to luaL_checkudata(). From previous
> experience, I learned to doubt this number, as valgrind tends to
> overestimate the impact of memory allocations. However, the problem is that
> this function actually needs to create a new string in order to do the user
> data name check at all. That seems to account for almost half of the
> runtime, plus another half which is spent in other places of that function.
> So this really appears to be a bottleneck.
>
> From looking at the function in LuaJIT2 (2.0.0-beta5), it basically only
> calls internal functions, so it's not clear to me what I can do to avoid the
> call.
>
> Any ideas what I could try to improve this?
>
> Stefan
>
>
> [1] Lupa: http://pypi.python.org/pypi/lupa
>
>
>

It's not LuaJIT specific, but you could try holding the metatable you
want to check as an upvalue of the function that's doing the check
(see lua_pushcclosure, or the equivalent in Lupa) and then check it
with something like this:

 // check argument 1 has 'MyType' metatable in upvalue 1
 lua_getmetatable(L, 1);
 luaL_argcheck(L, 1,
   lua_rawequal(L, -1, lua_upvalueindex(1)),
   "expecting a 'MyType' value");
 lua_pop(L, 1);

It's a bit more complicated but might help because it doesn't need to
push any strings or do a table lookup.

-Duncan