type-parameterized functions not specialized when passed as argument

[jruderman]

Ran into this while trying to make an ivec_equal function. Here's a testcase without any vectors.

fn builtin_equal_int(&int a, &int b) -> bool {
    ret a == b;
}

fn builtin_equal[T](&T a, &T b) -> bool {
    ret a == b;
}

fn apply_equal[T](fn (&T, &T) -> bool f, &T a, &T b) -> bool {
    ret f(a, b);
}

fn main() {
    assert  builtin_equal_int(5, 5);
    assert  builtin_equal(5, 5);
    assert  apply_equal(builtin_equal_int, 5, 5);
    assert  apply_equal(builtin_equal, 5, 5); // CRASH
}

[msullivan]

This is a close relative of #488. The problem is that builtin_equal doesn't have tydescs bound to it. Doing bind builtin_equal(_, _) will fix it.

The two obvious options I can think of for fixing the bug are:

1. have the typechecker disallow this program (that is, you can't pass generic functions without explicitly binding them).
2. implicitly specialize when passing a generic function

Solution 1 is pretty lame. Solution 2 imposes a somewhat hidden performance cost, which we may want to avoid? I prefer solution 2, but would be interested in ways to make it cheaper.

[pcwalton]

I like solution #2.

[msullivan]

Another symptom of the same issue is that the following code (that marijnh posted in #488) fails in trans.

fn generic[T](&T x) -> vec[T] { ret [x]; }
fn main() {
  auto f = generic;
  log_err f(1).(0);
}

marijnh pointed out that in this case, we can create a closure with the bound tydescs statically. I think that in general we can't do that, though, since the type might be parametric at the binding site as well.

[marijnh]

I've implemented solution 2 in 022363a