Should `cont.new` be constant?

[tlively]

Should cont.new be allowed to appear in global initializers, element segment items, etc.? The explainer does not currently say either way. For comparison, ref.func, struct.new, and array.new are all constant. I think it would make sense to allow cont.new to be constant, too. That would allow globals and tables to be initialized with continuations. Other new instructions should not be constant because they have observable side effects.

[slindley]

That makes sense to me. Are their any potential disadvantages of making these changes?

[titzer]

It seems odd to me, considering that resume consumes a continuation, making it no longer usable afterwards.

[slindley]

Hmm... good point. Continuations are certainly not constant in that sense (they are one-shot)... but does that mean it would necessarily be unsound to allow cont.new to appear in some of these places?

[rossberg]

Well, constancy isn't transitive. Semantically, what's constant is the reference to the continuation, while it's the continuation that is consumed/mutated, not the reference to it. This is analogous to a struct.new for a mutable struct, which also is considered a constant expression.

(Or maybe we side-step the whole question by just not calling it "constant" expressions, but something else. It's not really the relevant property.)