Add overload support in lpython

[Smit-create]

Fixes#141.

[certik]

I think the symbol table visitor can do all of this, correct? Or do we need to handle this in the body visitor also?

If just symbol table visitor, then I think what you can do is to start the ASR GenericProcedure node, and just keep adding to it. I think. You can see how this is done in LFortran, I can't remember right now. Otherwise using a temporary structure like you have done in this PR also works.

[Smit-create]

ASR GenericProcedure node,

Will have a look how that is done.

[Smit-create]

On the current diff, I get:

fromltypesimportoverload, i32@overloaddeffoo(a: i32, b: i32) ->i32:
returna*b@overloaddeffoo(a: i32) ->i32:
returna**2

ASR:

(TranslationUnit (SymbolTable 1 {
__lpython_overloaded_0__foo: (Function (SymbolTable 2 {
_lpython_return_variable: (Variable 2 _lpython_return_variable ReturnVar () () Default (Integer 4 []) Source Public Required .false.), a: (Variable 2 a In () () Default (Integer 4 []) Source Public Required .false.), b: (Variable 2 b In () () Default (Integer 4 []) Source Public Required .false.)
}) __lpython_overloaded_0__foo [
(Var 2 a) (Var 2 b)] [
(= (Var 2 _lpython_return_variable) (BinOp (Var 2 a) Mul (Var 2 b) (Integer 4 []) () ()) ()) (Return)] (Var 2 _lpython_return_variable) Source Public Implementation ()), __lpython_overloaded_1__foo: (Function (SymbolTable 3 {
_lpython_return_variable: (Variable 3 _lpython_return_variable ReturnVar () () Default (Integer 4 []) Source Public Required .false.), a: (Variable 3 a In () () Default (Integer 4 []) Source Public Required .false.)
}) __lpython_overloaded_1__foo [
(Var 3 a)] [
(= (Var 3 _lpython_return_variable) (BinOp (Var 3 a) Pow (ConstantInteger 2 (Integer 4 [])) (Integer 4 []) () ()) ()) (Return)] (Var 3 _lpython_return_variable) Source Public Implementation ()), main_program: (Program (SymbolTable 4 {
}) main_program [] [])
}) [])

[Smit-create]

We will now need to add support for overload in Call

[certik]

The ASR example you posted is almost ok, but it needs to also add foo: GenericProcedure(__lpython_overloaded_0__foo, __lpython_overloaded_1__foo) into the symbol table.

[Smit-create]

The ASR example you posted is almost ok, but it needs to also add foo: GenericProcedure(__lpython_overloaded_0__foo, __lpython_overloaded_1__foo) into the symbol table.

I thought since we are changing the name of the function/subroutine, it might not be that important to use GenericProcedure.

[certik]

What is your idea/design of making this work without a GenericProcedure in ASR? By storing the list of overloaded procedures in overload_defs inside AST to ASR, but not in ASR itself?

How would that work with independent compilation of modules to "mod" files, and then loading them in a new lpython invocation?

The idea that we have followed so far is to keep ASR self-contained, so that it contains all the information that you might need later. It seems in this PR we are not storing all the information for generic subroutines in ASR, but keeping it outside of it.

[Smit-create]

The idea that we have followed so far is to keep ASR self-contained, so that it contains all the information that you might need later. It seems in this PR we are not storing all the information for generic subroutines in ASR, but keeping it outside of it.

Yeah, that makes sense to me now. I'll try it using GenericProcedure.

[Smit-create]

The PR looks all good to go once the tests passes. Just need to figure out how to get rid of the global variable ast_overload.

[certik]

The last part to do is to pass this via arguments to the constructor of the visitor classes and have it as a local variable in python_ast_to_asr. Then it will not clash when modules are imported recursively.

[Smit-create]

Then it will not clash when modules are imported recursively.

That's fine, but I guess it won't be a big issue since we are storing the memory reference of each function which will indeed be unique.

[certik]

Only if you get lucky. When you load multiple modules, the visitor gets created, then destroyed. I think the nodes are allocated using our linear allocator, so I guess they would still be unique. But then when you compile multiple stuff from the main driver and happen to create a new allocator, then it will possibly collide. So we should not rely on that and have no global state.

[Smit-create]

How about using it as a class variable and clearing it completely when a symbol table visitor is instantiated?

[certik]

That will erase it when you import other modules recursively, wouldn't it?

[Smit-create]

Ahh, yes! One probable approach can be to add an optional arguement in FunctionDef to store the index where it will be present in GenericNode.

[certik]

Otherwise I think this looks good.

[certik]

Let me know if you need any help with implementing the passing of the ast_overload into the visitor patterns via their constructors.

[certik]

I moved the ast_overload into a local variable, so it should be good to go in now. Thanks @Smit-create !

[Smit-create]

Thanks for updating the PR and merging!!