RFC: remove do

[chris-morgan]

I started a mailing list thread on this topic about a week ago; in the first post I summarised the situation. I'll quote it here verbatim:

(Do I win the prize for the shortest thread name yet?)

error: last argument in do call has non-procedure type: ||

In the past three or four days I've seen at least as many enquiries in #rust about this error, and I'm sure there have been at least several others while I haven't been monitoring it. This is evidently causing quite a bit of confusion.

Here's a summary of the change. The syntax is the same as it was before:

do expr { block }
do expr |args| { block }
do expr(args) { block }
do expr(args) |args| { block }

These used to desugar to the following, respectively:

expr(|| { block })
expr(|args| { block })
expr(args, || { block })
expr(args, |args| { block })

These now desugar to the following, respectively:

expr(proc() { block })
expr(proc(args) { block })
expr(args, proc() { block })
expr(args, proc(args) { block })

The change is that it now accepts a procedure rather than a closure. No syntax change, just a semantics change which breaks a lot of code.

Closure: a stack function; used to be &fn(..) -> _, is now |..| -> _. Can be called multiple times, requires no allocations and is not Send.

Procedure: a heap function; used to be ~once fn(..) -> _, is now proc(..) -> _. Can be called once, requires heap allocation and is Send.

Procedures are good for sending cross-task; things like the task body are a good match. Still, I think there are a few problems with how things are at present (i.e. after the do semantics change):

1. do is still using the syntax of a closure (|..| { .. }), despite it now being a procedure.
2. All of a sudden, things using closures need to shift away from using do or use procedures; this is causing confusion and may cause bad design decisions where nice sugar triumphs over what is actually needed; often the best solution may not be clear. (I, for example, had not thought about the fact that proc was going to allocate; the ~once fn name was clearer about that. I'll speak about &once fn another time. Don't mention it now, this thread is just about do.)

I have two solutions that I think could answer these concerns. Leaving it as it is seems a bad idea to me.

(a) Kill do

I've had mixed feelings about do. Overall, it's pretty trivial syntax sugar, but it's sugar of a dubious sort, because it changes something that looks like a function call with N arguments to be a function call with N+1 arguments. That's just a matter of learning it.

Still, do is nice sugar in the way it gets rid of parentheses at the end. Overall, is it worth it? I don't know.

Once do is gone, there's no problem left: just remove the sugar everywhere it was used and everything works and will do for the foreseeable future.

(b) Make do support both closures and procedures

The syntax of do can be clearly seen to include the closure syntax. We could easily extend it to support both closures and procedures.

Here is a proposed do using closures once more, keeping the syntax it had last week:

do expr || { block }
do expr(args) || { block }
do expr |args| { block }
do expr(args) |args| { block }

Here is a proposed do using procedures as the current behaviour is, but with new syntax which is clearly a procedure:

do expr proc() { block }
do expr(args) proc() { block }
do expr proc(args) { block }
do expr(args) proc(args) { block }

This does leave these cases which are currently valid unclear:

do expr { block }
do expr(args) { block }

The options for this are (a) disallowing it; (b) making it always of the function types; and (c) inferring the type. I generally prefer the last solution but it is the most difficult. I'm not sure how it all fits into the function traits stuff at all.

Incidentally, all this leaves the possibility open of making do work for any argument type, where do expr1 expr2 simply desugars to expr1(expr2) and do expr1(args) expr2 to do expr1(args, expr2). I don't know if that would be a good thing or not; it's probably best to avoid discussion of that at present.

Summary

Leaving do in its present form seems to me a distinctly bad idea, with the syntax of one form of function while it uses another form of function. I think we need to redo do very soon. (I'd save this joke for later in the thread, but I'm afraid someone else might steal it. I expect all responses to indicate they're in favour of this by using the title "Re: do" :P.)

For myself, I have no preference to indicate; I am torn between the two options.

After some discussion, I think that for the moment, removing do is probably the better option. We can then see about adding it or something similar back later if we want to.

So then, concrete proposal: remove do from the language.

[liigo]

No, don't remove do before you have better syntax. Fn({}) is too ugly to accept.

[adrientetar]

I've often felt like do was not useful enough to be worth-being, admittedly.

[brson]

I think we should remove it at this point. I liked it better when it worked for stack closures.

[hatahet]

@brson Is it worth making spawn a built-in function in order to keep the nice syntax we currently get with do? It seems that this is the main point of contention for keeping do.

[chris-morgan]

@hatahet Do you mean making it a built-in keyword with syntax of its own? As it stands, it is in the prelude which means it's a built-in function. When do is removed (which I'm working on), it will just operate as

spawn(proc() {
    ...
})

That is not a horribly ugly monstrosity. Sure, it's not quite as nice-looking as do spawn { ... }, but it's not bad and is clear. I believe there will be general agreement in the core team that spawn should not be made a separate keyword (spawn { ... }); there are other ways of spawning a task (e.g. if you manually construct a task object you have task.spawn(proc() { ... }). And any syntax which could cope with all the cases would be exactly what we're looking at removing now.

[hatahet]

@chris-morgan Yep, that's what I was referring to. You're right -- I guess spawn(proc() { }); doesn't look too bad.

[brendanzab]

I've been de-do-ifying most of my code, and it it's actually not as bad as I thought. It definitely would simplify the language more to remove it, which is a plus in my books.

[brson]

Nominating.

[bytbox]

In light of the recent email, I've taken this up as a relatively-easy but apparently-helpful task.

[brson]

@hatahet I think we're more likely to prettify spawn with a macro.

spawn! {
    ...
}

Is pretty hot, though I think the braces don't work in that position right now.

[brson]

@bytbox Awesome! Ping me or anybody on #rust-internals if you need some guidance.

[hatahet]

@brson Great to hear :)