Clarification on directionless action arguments behaving like direction in

[jaehyun1ee]

The spec mentions about directionless action arguments:

For an action, a directionless parameter indicates that it is “action data”. See Section 14.1 for the meaning of action data, but its meaning includes the following possibilities:

• The parameter's value is provided in the P4 program. In this case, the parameter behaves as if the direction were in.

The phrase "behaves as if the direction were in", I believe, is not quite clear, particularly regarding compile-time known-ness.

The differences between an in parameter and a directionless parameters, aside their copy-in/out semantics, is:

• in can be implicitly cast, while directionless cannot (as per the current spec)
• directionless argument should be compile-time known, while no such restriction exists for in argument.

So, I wonder if "behaves as if the direction were in" also imply that the directionless arguments supplied to actions in a P4 program (not by the control plane) need not be a compile-time known value?

Relevant to this issue, issue473.p4 in the p4c test suite is expected to be rejected by the frontend.

action b(inout bit<32> x, bit<8> data) {
    meta.a = meta.a ^ (bit<4>) x ^ (bit<4>) data;
}
table t1 {
    key = { hdr.ethernet.srcAddr : exact; }
    actions = { b(meta.c); }
    default_action = b(meta.c, (bit<8>) meta.d);
}

issue473.p4(70): [--Werror=type-error] error: (bit<8>)meta.d: action argument must be a compile-time constant
        default_action = b(meta.c, (bit<8>) meta.d);
                                   ^^^^^^^^^^^^^^^

[jafingerhut]

Would it address your question fully if we were to replace this text of the spec:

The parameter's value is provided in the P4 program. In this case, the parameter behaves as if the direction were in.

with something like the following?

The parameter's value is provided in the P4 program. In this case, the parameter behaves as if the direction were in, with the restriction that arguments must be compile-time known.

[jaehyun1ee]

Yes, I think it addresses my question :) Would it be alright if I create a PR based on your suggestion?

[jafingerhut]

Certainly. Please do.

[jaehyun1ee]

I just found out that I overlooked the spec saying:

For an action, a directionless parameter indicates that it is "action data". See Section 14.1 for the meaning of action data, but its meaning includes the following possibilities:

• The parameter’s value is provided in the P4 program. In this case, the parameter behaves as if the direction were in. Such an argument expression need not be a compile-time known value.

So it was stated that a directionless action argument need not be compile-time known.
Seems like the test program wasn't in line with the current spec.

[jafingerhut]

It is a small enough detail that it is easy to miss (I did in my comment above -- good that you found it later). Perhaps it would be more appropriate to file a bug against p4c, then, if it does not follow the spec?

[jaehyun1ee]

I believe so. Closing in favor of p4c issue!