P4_14 converter and action functions

[jnfoster]

The converter from P4_14 to P4_16 allows action functions to be used as functions -- e.g., one action may invoke another. It's not clear that this is always in-line with the P4_14 spec.

For example, consider the following snippet of code from testdata/p4_14_samples/action_inline.p4 (also attached to the bottom of this issue):

action a(y0)
{ add_to_field(y0, 1); }
    
action b() {
    a(md.b);
    a(md.b);
}

According to the spec, the first argument to the add_to_field primitive should be a field. However, here it is a parameter y0, which could either be a field or action data supplied by the control plane.

In this case, since a is only invoked by b, the compiler happily generates the following:

    @name(".a") action a(inout bit<1> y0) {
        y0 = y0 + 1w1;
    }
    @name(".b") action b() {
        a(meta.md.b);
        a(meta.md.b);
    }

which is fine.

However, if we also use a in a table, then we get the following error:

action_inline_tweaked.p4(25): error: a: parameter y0 must be bound
action a(y0)
         ^^

This error disappears if we replace the body of a with different code.

Going a step further, if we change b to invoke a on another field with another width (7 bits instead of 1), we get a different error:

action_inline_tweaked.p4(29): error: cast: must be a left-value
    a(md.b);
      ^^^^

The relevant code generated by the converter is

    @name(".a") action a(inout bit<7> y0) {
        y0 = y0 + 7w1;
    }
    @name(".b") action b() {
        a((bit<7>)meta.md.b);
        a(meta.md.c);
    }
    @name(".t") table t {
        actions = {
            a;
            b;
        }
    }

I'm not sure what's intended, and I'm all for abstraction, but what we currently have seems somewhat...

See here for the P4 sources:
action_inline.txt
action_inline_tweaked.txt