Re: [PATCH] make light userdata a little bit heavier

[Philipp Janda <siffiejoe@...>]

Am 22.04.2014 08:44 schröbte steve donovan:
> The trouble is that it's very common to wrap a C/C++ pointer in a
> small full userdata type (pointer to pointer) which is wasteful if
> you're really worried about all these extra objects needing to be
> GC'd.  (My question would be: yes, but it does it make much difference
> in real projects?)

I use interning for such userdatas (and for enums/flags): You store a 
mapping of the wrapped C pointer to the full userdata in a weak-valued 
table. Whenever you want to create a new userdata you reuse the userdata 
in the cache if possible. This should reduce all those extra memory 
allocations. For enums/flags you need one cache per enum type (I usually 
put it somewhere in the metatable).

Which brings me to another topic: I'd rather have different integer 
types (for typesafe enums/flags, only bit operators, no arithmetic 
operators) than different lightuserdata types. One 32-bit subtag value 
and a 32-bit enum value would fit comfortably in those 64 bits we have 
in a Lua value ...

Philipp