`comptime` not required for param with comptime-only type in function called at runtime about zig HOT 2 OPEN

while I wouldn't mind seeing this fixed and still write this way in my own code atm, I could also see an argument that it should be left as-is since it would better parallel with function body code that the comptime keyword is only used to elevate otherwise-runtime things to be comptime-known. in bodies we even have an error about using it redundantly.

from zig.

This problem is a little trickier than it seems, because we don't necessarily know if a function call is comptime until after we evaluate its parameters. Consider:

fn f(T: type, x: anytype) @TypeOf(x) {
    _ = T;
    return x;
}

If f is called in a runtime context, the call itself might still be comptime depending on whether x has a comptime-only type. That means we require a comptime annotation on T only if x has a runtime type. This might be non-trivial - I forget how the Sema logic is structured.

If that's hard, it'd be easiest to instead tighten the rule a bit and say that the comptime is only optional if the call site is already being comptime-evaluated, i.e. it's not just because the function is comptime-only. I'm not sure if this would regress use cases however, and it feels like a more complex rule.

By the way, a good separate enhancement might be to tighten the rules further by requiring the annotation on any argument which is not generic. We could check in AstGen that no identifier is a prior parameter, or a generic capture, or a decl derived from a generic capture. In that case, we could completely change the rule to the following, which would be nicer and also more generally helpful:

"A function parameter with a comptime-only type must be marked comptime if its type is comptime-only or the return type is comptime-only, and the parameter type is not generic."

from zig.