Experimental control-flow operator

[jsiek]

This adds "delimited continuations" with delimit, yield and resume features that provide a kind of resumable exception with first-class continuations. The point of adding this feature is not to say that we want delimited continuations in Carbon, but this represents a place-holder for other powerful control-flow features that might eventually be in Carbon, such as coroutines, threads, exceptions, etc. As we refactor the executable semantics, having this feature in place will keep us honest and prevent us from accidentally simplifying the interpreter to the point where it can't handle features like this one.

Instead of delimited continuations, we could have instead done regular continuations with callcc. However, there seems to be a consensus amongst the experts that delimited continuations are better than regular ones.

This implementation needs more testing. However, the examples from the literature that I'd like to use as tests all use lambda, so I'm going to do an experimental lambda and then circle back to add more tests for the delimited continuations.

Fixes #349

[dabrahams]

@jsiek Note my edit to the description that cross-references the issue.

[jsiek]

Thanks Dave!

[dabrahams]

I know it's a hassle, but comments on all the new entities is basically a prerequisite for getting a useful review out of me. Those statements are how I evaluate whether a thing does/is what it's supposed to be. Please do that at a minimum and ask me to re-review.

Thanks!

[jsiek]

Thanks for the feedback and reminders Dave!
I've added comments, especially for the hairy chunks of code.

[geoffromer]

I think this will be a lot easier to review (and a lot easier for future maintainers to read) if there's some kind of prose description of this feature from the point of view of the language user. It doesn't have to be super detailed, since the whole point of this code is to act as a specification, but just a quick introduction to the core vocabulary (e.g. "continuation values"), syntax, and the broad strokes of the semantics. The best place would probably be executable_semantics/README.md, where it can also act as a "this is a stress-test of the executable_semantics architecture, not a planned Carbon feature" disclaimer.

[jsiek]

@geoffromer good idea, I'll add an introduction to this in the README. Thanks!

[jsiek]

I've implemented an imperative variant of the shift/reset style of continuations.
The README file is updated with a short description and there are three new test cases.
So far so good. I think this holds together OK and gives us a peek at what coroutines might
look like for Carbon.
The next thing on my todo list for this is to add more test cases to really kick the tires.
BTW, this design is still amenable to value semantics... we just need to tweak what happens
when a continuation object is copied. (Actually, I'm not sure what happens there at the moment.)

[jsiek]

Copying a continuation now does a deep copy of its stack.

[jsiek]

After more thinking and more test cases, I flipped back to shallow copy for continuations. Some of the behavior from allowing copies is just too weird, such as double deletes of a local variable (the now deleted test testdata/experimental_continuation9.6c).