Add an infinite-loop construct

[catamorphism]

This idea has come up before, but #1905 reminded me that we never really reached agreement on it. I propose adding a new looping construct, forever [block], that has the same semantics as while (true) [block]. The reason to add a special construct would be to allow typestate to recognize that infinite loops run forever, and not yield spurious-looking errors (see #1905 for an example) about functions not returning a value. The current workaround is to do while (true) { ... }; fail "Unreachable";, which isn't terribly burdensome, but is annoying.

I realize we're trying not to add stuff, but this issue seems to be a source of confusion, and while (true) loops seem to be pretty common.

[jruderman]

It might be simpler to have typestate recognize while true as something special.

[brson]

Could possibly also make an iterator like fn until<T>(f: fn() -> option<T>) -> T that iterates until the block returns some.

That would at least cover the case in the referenced issue.

[brson]

Well, I guess there are lots of ways to structure the code in the other issue so that it isn't a while true loop. I am not in favor of adding a new language construct for this. If they have to write it in a different way anyway then it should be written with a library iterator. Otherwise, I agree with Jesse that while true should interact with typestate in the way that most people intend.

[brson]

We could also have do { } just mean 'iterate forever'. At least that doesn't add more keywords.

[brson]

Or change do to repeat, so you can have repeat { } mean 'iterate forever', and repeat { } while something; mean 'iterate until'. repeat is clearer that it's a loop.

[pcwalton]

Java's typestate knows about while (true). It's actually rather tricky, because it makes the typestate algorithm have to track booleans. Special-casing the construct while true, and no other, might work, but if it turns out to be difficult in practice it may well be worth adding a new language construct.

[nikomatsakis]

I think it'd be ok to just accept "while true" as our "infinite loop" form, personally.

[graydon]

What's wrong with a library function?

fn forever(body: fn()) -> ! { while true { f(); } core::unreachable(); }

[nikomatsakis]

for one thing, most "infinite" loops actually break... or at least ret. but even if we had break and non-local return, wouldn't type state then have to treat the "forever" library function as special, so as to know that it never returns unless the body breaks?

[graydon]

If it breaks, it's not a forever loop, and this whole bug doesn't apply: the control paths after the loop are reachable.

[nikomatsakis]

It was more ret than break that I was concerned about.

[graydon]

Oh, right. Ok, then in that case you would have to manually put in the ::unreachable() call.

I looked through the compiler and libraries and see ... a few cases, but only a handful. I guess it's a real case; and given that the cost of "solving" it is extremely low in terms of cognitive burden I won't resist further. I'd say I like brson's suggestion of making the trailing while optional, but I think it leads to grammatical ambiguity if you juxtapose it with another while loop. We could make that grammatically illegal, but that's teaching the parser an unusual amount about control flow (nothing can come after a repeat { ... } loop?) and feels like trying to make two wrongs into a right.

If we solve this "in the language" I've a slight preference for an unadorned loop { ... } construct (and hey, it's a keyword in Sather). The word forever is misleading (especially if we usually ret, which we do) and too long anyways. I agree we could get to the same place teaching typestate about while true, but suspect people will be filing followup bugs asking us why we can't tell that while a == b || a != b { ... } is also an "infinite loop", and why don't we include a general boolean solver, and what kind of amateurs are we? I don't want to play that game.

[catamorphism]

forever was just a strawperson; loop is good. Maybe we can talk about this at the next Tuesday meeting?

[catamorphism]

I'm hearing consensus in the meeting in favor of adding loop, which I will implement.

[graydon]

At meeting, has consensus to implement loop.

[catamorphism]

Implemented. Cleanup of existing code-that-uses-while-true is in 35400e1

[graydon]

Great, thanks!