I’ve implemented my own class system and I’m having trouble with __tostring; I suspect a similar issue can happen with other metamethods, but I haven’t tried.
(Brief detour: each class has a __classDict attribute, holding all methods. It is used as the class instances’ __index. At the same time, the __classDict’s __index is the superclass’ __classDict, so methods in superclasses are authomatically looked up.)
I wanted to have a “default tostring” behavior in all instances. But it didn’t work: the “tostring” behavior doesn’t “propagate” through subclasses correctly.
I’ve done this test exemplifying my issue:
mt1 = {__tostring=function(x) return x.name or "no name" end }
mt2 = {}
setmetatable(mt2, {__index=mt1})
x = {name='x'}
y = {name='y'}
setmetatable(x, mt1)
setmetatable(y, mt2)
print(x) -- prints "x"
print(mt2.__tostring(y)) -- prints "y"
print(y) -- prints "table: 0x9e84c18" !!
I’d rather have that last line print “y”.
Lua’s “to_String” behaviour must be using the equivalent of
rawget(instance.class.__classDict, '__tostring')
instead of doing the equivalent of
instance.class.__classDict.__tostring
I suspect the same happens with all metamethods; rawget-equivalent operations are used.
I guess one thing I could do is copying all the metamethods when I do my subclassing (the equivalent on the above example would be doing mt2.__tostring = mt1.__tostring) but that is kind of inelegant.
Has anyone fought with this kind of issue? What where your solutions?
Thanks to daurnimator’s comments, I think I found a way to make metamethods “follow”
__indexas I want them to. It’s condensed on this function:I hope it is straightforward enough. When a new metatable is set, it initializes it with all metamethods (without replacing existing ones). These metamethods are prepared to “pass on” requests to “parent metatables”.
This is the simplest solution I could find. Well, I actually found a solution that used less characters and was a bit faster, but it involved black magic (it involved metatable functions de-referencing themselves inside their own bodies) and it was much less readable than this one.
If anyone finds a shorter, simpler function that does the same, I’ll gladly give him the answer.
Usage is simple: replace
setmetatablebysetindirectmetatablewhen you want it to “go up”:A little word of warning:
setindirectmetatablecreates metamethods on mt2. Changing that behavior so a copy is made, and mt2 remains unaltered, should be trivial. But letting them set up by default is actually better for my purposes.