> Note that not having runtime-known stack allocations is a key piece of the puzzle in Zig's upcoming async I/O strategy because it allows the compiler to calculate upper bound stack usage for a given function call.

That’s a genuinely interesting point. I don’t think known sizes for locals are a hard requirement here, though threading this needle in a lower-level fashion than Swift would need some subtle language design.

Fundamentally, what you want to do is construct an (inevitably) runtime-sized type (the coroutine) out of (by problem statement) runtime-sized pieces (the activation frames, itself composed out of individual, possibly runtime-sized locals). It’s true that you can’t then allow the activations to perform arbitrary allocas. You can, however, allow them to do allocas whose sizes (and alignments) are known at the time the coroutine is constructed, with some bookkeeping burden morally equivalent to maintaining a frame pointer, which seems fair. (In Swift terms, you can construct a generic type if you know what type arguments are passed to it.) And that’s enough to have a local of type of unknown size pulled in from a dynamic library, for example.

Again, I’m not sure how a language could express this constraint on allocas without being Swift (and hiding the whole thing from the user completely) or C (and forcing the user to maintain the frames by hand), so thank you for drawing my attention to this question. But I’m not ready to give up on it just yet.

> At a fundamental level, runtime-known stack allocation harms code reusability.

This is an assertion, not an argument, so it doesn’t really have any points I could respond to. I guess my view is this: there are programs that can be written with alloca and can’t be written without (unless you introduce a fully general allocator, which brings fragmentation problems, or a parallel stack, which is silly but was in fact used to implement alloca historically). One other example I can give in addition to locals of dynamically-linked types is a bytecode interpreter that allocates virtual frames on the host stack. So I guess that’s the other side of being opinionated—those whose opinions don’t match are turned away.

Frankly, I don’t even know why I’m defending alloca this hard. I’m not actually happy with the status quo of just yoloing a hopefully maybe sufficiently large stack. I guess the sticking point is that you seem to think alloca is obviously the wrong thing, when it’s not even close to obvious to me what the right thing is.

Alloca is a fundmentally insecure way of doing allocations. Languages that promote alloca will find themselves stuck in a morass of security messes and buffer overflows. If Zig were to adopt alloca, it would make the catastrophic mistake that plagued C for over several decades and introduce permanently unfixable security issues for another generation of programming languages.

Out of curiosity, why is knowing the size of locals required, exactly? Because it avoids dynamic allocations for each fiber?