sol::object had a few reference leaks in the way it retrieved values: it now does it properly without leaving the stack at +1 item
sol::stack was drastically cleaned up, with the following key change:
* sol::stack::push now returns an integer of the number of things its pushed (usually 1, but can be more) (Thanks, @PrincessNyanara!)
* sol::stack::call now calls functions flexibly, and getting is done more reliably
* due to the innovation of stack::call and using absolute indices, we no longer have to use reverse_call style programming to deal with lua
* sol::reference::get_type is now const-correct
* sol::state and sol::table now have a cleaned up `get` implementation since it is no longer held back by the ugliness of VC++'s incapability to handle templates
* the name `sol::userdata` now belongs to a type that actually encapsualtes a void* with a pusher/getter than gets a userdata void* value (TODO: give it a template to make it static_cast to that type on get?)
* lightuserdata_t -> light_userdata, upvalue_t -> upvalue as type names (mostly details)
* pushers for various types were updated to return integers
Also a convenience type `sol::ref`, which is just an
alias to `std::reference_wrapper`, to enable easy getting with `lua.get<>`
(we can't use `lua.get<some_type>` because of necessary `Unqualified<T>` use)
now are usable as functions for userdata.
allows free functions and lambdas to provide useful operations, like operator+*-/
and other things which may not be implemented as class members.
userdata now performs lookup based on tables
userdata now has reduced number of vector tables
userdata garbage collection improved
debug.hpp - new header for debugging problems with stack, mostly for internal use
this triggered overhaul of set_function/pusher<function_t>::push(...)
both state and table reflect changes to userdata structure to make it easier to use
tests updated to account for overload resolution
some function-related traits added to make use easier -- cleaned up archaic typenames in function_types.hpp
Account for std::reference_wrapper for objects -- sol now uses copy-by-default (value-semantics) for all functors
updated tests to reflect this
Userdata now properly forwards arguments to constructor
get_call now properly has extra parameters to allow for forwarding items from the first call that are not popped
Added tests to cover new cases
c++ semantics dictate that it's accessed by index, but the tests seem to want to indicate that it should be accessed like a hashmap (or just using basic lua table semantics)
i have no idea how to make this incompatibility work in the new system...
i will ask repo master if he knows anything
`#` == __len
`[key] = value` == __newindex
`key` == __index
are not working properly for some reason, will need to investigate more deeply to find out why
on bright side all tests pass including new tests added (for take/return std::function)
fixed some things forgot to change with addition of upvalue_t for clarity (see previous commit about userdata/lightuserdata vs upvalue)
a lua upvalue can be lightuserdata, userdata, or anything else that can have its address taken (it's immediately popped of the stack and carted around with function call)
a lightuserdata can only be a pointer (void*)
a regular userdata can be anything, but is stored as void* because of "anything" semantics and C heritage of lua
upvalues deserve to use the `lua_upvalueindex(n)` macro: lightuserdata/userdata does not (must not) go through this process
get turned into getter<T>, matches pusher<T> and uses same semantics as std::allocator and other things used throughout the codebase
-----
userdata has its traits defined outside in new file of userdata to prevent errors when trying to use those typetraits in places before userdata.hpp gets included
userdata was changed to support returning itself via pointers or references.
rework of stack changes semantics based on T&, T*, and T&& (the last one tries to create a new userdata and move in data)
solves problems maybe presented in https://github.com/Rapptz/sol/issues/25
-----
container.hpp is attempt at solving original problem before going on wild tangent with userdata, stack, and get
is going to attempt to use userdata to allow transporation of containers losslessly, perhaps without copying need
-----
found out trying to return a std::function does not work -- not sure what do exactly?
perhaps should push c closure as last thing, but right now it is tied to a key value (code comes from table.hpp and set_function)
will just have to think over how stack arranges itself and learn what to do
by forward-declaring the std::function overload in `stack`, it is able to find the right function to use.
All tests are passing
the ninja file has been tweaked to make it easier to invoke a g++ build on windows
tests added to confirm userdata can be passed into C++ function types
demangle is now named lua_demangle, and the core demangle without any replacements (to fit lua) is just named demangle
Formattings fixes everywhere
Refactoring on function_types.hpp performed to slim down some of the calls: could use more refactoring
Drastically simplified userdata's binding capabilities: constructor supports both `:` and `.` syntax (but member functions DO NOT).
All tests are passing
Added pop() function to `reference`
`stack.hpp` now has many more functions to properly handle user data
`types.hpp` now has `userdata_t` and `lightuserdata_t` to faciliate `stack`'s `pop` and `push` operations
Class binding functionality built into `table` and `state`, as well as placed in `userdata` class in `userdata.hpp`
demangling detail now present for clang, gcc, and MSVC (somewhat for MSVC)
Constructor arity still needs to be handled for `userdata<T>`
tests updated and passing
But! I left it on the `sol::function` type, because it's necessary to discard returns. The .call is still there, though, just incase you need it. <3
reverse_indices_builder is also there, to make sure we can push and pop without the lua api taking our types and breaking them for the std::tuple returns.
All is at it should be~
STATEFUL FUNCTIONS!
Any stateful function now works and is properly cleaned up, thanks to some additional metatables that are associated with the function values.
This lays the ground work for class bindings, but that's a far off dream. For now, table retrieval and `operator[]` is what's for dinner.
Member functions now work as well for set_function. If performance of `new`ing a type ever becomes too large, we can create a custom allocator for the std::shared_ptr's of the types. We can also up-front allocate for the unordered_map as well.
It'll be up to @Rapptz's to make further changes to make sure GCC compiles, but the core of the functionality is there and my work is essentially done.
Class bindings are up next, but really I don't want to even touch those. :c
