Julia: restrict typed variable declaration to `local x::T` and `global x::T`

[StefanKarpinski]

It's confusing that x::T is an assertion in some contexts and a declaration that x is a local variable of type T in other contexts. I propose we ultimately require a declaration keyword for the typed variable declaration usage, i.e. local x::T or global x::T. The first step is to deprecate the keywordless syntax, which should be done as soon as possible if we're going to do this.

[yuyichao]

Does this mean after the deprecation period we can make x::T in statement position always mean typeassert? That'll be nice since AFAIK we you can only do that with Base.typeassert(x, T) now.

[StefanKarpinski]

Yes, that is what it would mean.

[JeffBezanson]

Seems like a good change. What should x::T = 0 mean?

[vtjnash]

I think that can still be an implicit local. Unless we want to enforce that the declaration must dominate the usages? For example, what does the following mean:

function f(x)
         if x
           T = Int
           y::(println("HI");T) = 2.0
         else
           y = 3.0
         end
         y
       end
f(false)

(answer: it prints "HI" twice then throws "UndefVarError: T not defined")

[toivoh]

I think we should probably go all in and make type assertions on the left hand side of an assignment be an error.

Or possibly a type assertion on the old value, but that seems bound to be confusing. Anyway, consider x[5]::Int = 1, that's a type assertion on the lhs that needs to be treated in another way than as a variable declaration.

[StefanKarpinski]

What should x::T = 0 mean?

Syntax error suggesting local x::T = 0 instead?

[StefanKarpinski]

Unless we want to enforce that the declaration must dominate the usages?

I think this is a good idea. That example should be required to be written as:

function f(x)
    local y::Int
    if x
        y = 2.0
    else
        y = 3.0
    end
    return y
end

[Tetralux]

local is implicit isn't it? Why would specifying a type for the variable require you specify that?

I would rather remove ::T as a shortcut for type assertions, and instead have the programmer write typeassert(x, T), or---perhaps more clearly, assert(typeof(x) === T).

[jamesonquinn]

The above PR is probably not the final word; it uses peek after processing a :: expression to see if the following operator is an assignment, which means it catches a simple a::Int = 3 but misses things like (a::Int) = 3 or a::Int, b = 3,4. To work more generally, it would have to look backwards rather than forwards; that is, pick apart the LHS expression when it encounters an assignment. I believe that's a matter of recursing down through tuples looking at the car of each item, but I'm not certain, and at any rate that would be more work. I wanted to keep my first PR simple so people could tell me if I'm doing it wrong.

[vtjnash]

I'm strongly in favor of this. Just tracked down a latent bug in the numbers test to

for T in (Int32,Int64)
    i::T, j::T = 1, 2
end

sometimes meaning (and doing) something completely different for i and j in that expression: :: means either of a type-assert or a local variable declaration in either usage (yikes!).

In observation, since the dates test used j as a global, that expression is a declaration of a local variable i with type T and assignment local i = convert(2, T)::T, vs. a typeassert that j::T followed by an assignment of j = 2.

[vtjnash]

can I nominate this deprecation in 0.5 (and volunteer myself to work on it now)? (or at least the non-breaking part that in x::Foo = 1, the :: will never mean typeassert, which is the deprecation for #964)

[tkelman]

If you want to finish this before we feature freeze I'd be okay with it. I'm not okay with putting a patch I haven't seen yet on the milestone right now.

[StefanKarpinski]

sometimes meaning (and doing) something completely different for i and j in that expression: :: means either of a type-assert or a local variable declaration in either usage (yikes!).

Can you explain this a little more, @vtjnash?

[vtjnash]

Sometimes x::T = 1 could mean: the local variable x has type T and assign it the value convert(T, 1); other times, it could mean: typeassert the global variable x was of type T, then assign it the value 1. Anyways, PR submitted to add deprecation warnings to the context-dependent usages.

[cstjean]

What about type annotations in loops and function arguments? In 0.4, the former is a type declaration, but the latter is not (function foo(x::Int) x="hello"; x*x end; foo(2) is valid).

[vtjnash]

Those already seem consistent to me (and unchanged from 0.4). Am I missing something?

julia> for x::Int = 1:10; x = 0.0; end
# declares that x has type Int

julia> function foo(x::Int) x="hello"; x*x end
# "asserts" that the first argument was passed an Int

[cstjean]

You're right, I wasn't thinking of the loop as an assignment, but it makes perfect sense.

[JeffBezanson]

Fixed by #17445