Pass-by-value semantics not preserved when using __addr_of

[ironcev]

Consider the following test that is passing on v0.46.0 and the Sway latest master branch:

script;

fn main() -> u64 {
   0
}

struct A {
  a: u64,
}

struct B {
  a: A,
  x: (u64, u64),
}

#[test]
fn ptr_issue() -> u64 {
    let a = A { a: 11 };
    let mut b = B { a: a, x: (11, 11) };

    let mut ptr_b = ptr(b);

    ptr_b.write(B { a: A { a: 22 }, x: (22, 22)});

    assert(b.a.a == 22);
    assert(b.x.0 == 22);
    assert(b.x.1 == 22);

    42
}

#[inline(never)]
fn ptr<T>(t: T) -> raw_ptr {
    __addr_of(t)
}

Because of the pass-by-value semantics ptr should get the copy of b and thus a different pointer should be returned from the function. (A pointer to the local value indeed, which is an issue on its own, but different then the one described here.)

Since ptr is actually getting the pointer to the original value and no copy is made writing to ptr_b overwrites b.

Notice that if we don't use _addr_of (or references one day) to the immutable parameters, the fact that the copy is not made is a valid optimization.