Type inference succeeds when it shouldn't

[jaehyun1ee]

The p4c frontend regards the following program as valid, when it is actually not.

struct a_t { bool f; }
struct b_t { bit<32> f; }

control foo(inout a_t one, inout a_t two) { apply {} }
control bar(inout b_t one, inout b_t two) { apply {} }

control C<H, M>(inout H one, inout M two);

package P<H, M>(C<H, M> c1 = foo(), C<H, M> c2 = bar());

P<_, _>() main;

Here, we do not supply any argument to the instantiation of package main, so the default values of c1 and c2 take place.
However, c1 implies H = a_t and M = a_t while c2 implies that H = b_t and M = b_t.
So in this case, the compiler shouldn't be able to infer the correct type argument for the don't care types (_) in P<_, _>.

Also, if we change the program as follows, where we instantiate P() main; instead of P<_, _>() main;,

struct a_t { bool f; }
struct b_t { bit<32> f; }

control foo(inout a_t one, inout a_t two) { apply {} }
control bar(inout b_t one, inout b_t two) { apply {} }

control C<H, M>(inout H one, inout M two);

package P<H, M>(C<H, M> c1 = foo(), C<H, M> c2 = bar());

P() main;

then the compiler rejects the program (as expected) saying:

default-package-argument-reduced.p4(11): [--Werror=type-error] error: main: could not infer a type for variable H
P() main;
    ^^^^
default-package-argument-reduced.p4(9)
package P<H, M>(C<H, M> c1 = foo(), C<H, M> c2 = bar());

I believe it suggests some possibility that the frontend overlooks some case when inferring concrete types for don't care types _.