Fix codegen breaking aliasing rules for functions with sret results

[dotdash]

This reverts commit a0ec902 "Avoid
unnecessary temporary on assignments".

Leaving out the temporary for the functions return value can lead to a
situation that conflicts with rust's aliasing rules.

Given this:

fn func(f: &mut Foo) -> Foo { /* ... */ }

fn bar() {
    let mut foo = Foo { /* ... */ };

    foo = func(&mut foo);
}

We effectively get two mutable references to the same variable foo at
the same time. One for the parameter f, and one for the hidden
out-pointer. So we can't just trans_into the destination directly, but
must use trans to get a new temporary slot from which the result can
be copied.

[thestinger]

@dotdash: It would be nice if this checked for an outstanding borrow and performed the optimization if there wasn't one. I don't know how hard that would be...

[dotdash]

@thestinger could you elaborate on that check? I'm not sure I understand why an outstanding mutable borrow would make the optzn valid. Also, you said something about unwinding invalidating it as well. What about that?

[thestinger]

@dotdash: Sorry, I meant if there wasn't an outstanding borrow. I don't think this optimization is problematic for unwinding if there's no aliasing &mut, which is problematic for other reasons anyway.

[alexchandel]

Optimizing returning large structs would seem to be important for Rust's constructor idiom.

[huonw]

@alexchandel, being correct is far more important; we should revert to correctness quickly, and can add optimisations, progressively, later.

[alexchandel]

@huonw Sure, no one could reasonably disagree with this revert. I think @thestinger's optimization—or a working version of what this PR reverts—would be good, and would seem to benefit every large struct implementation that has a method like fn new() -> Foo.