Vec::append should swap if the lhs is empty

[Gankra]

Here's a reasonable pattern:

let mut results = Vec::new();
for ... {
  results.append(new_results);
}

On the first append, we can just swap ourselves with the rhs, avoiding an allocation and copy. This is annoying for our users to remember to do, and is basically free to support given how much work this method otherwise does.

It should perhaps only be done if lhs.capacity is 0, just because we don't want to mess up a user who has preallocated a nice big buffer. Maybe lhs.capacity < rhs.capacity?

[petertodd]

Nit: lhs.capacity() <= rhs.capacity() should be fine. :)

Though I could imagine this causing problems in conjunction with Vec::from_raw_parts() in unusual cases where the code had done something special with the allocation of the lhs.

[the8472]

I attempted implementing this in #77538 but it didn't seem to have any measurable impact on compiler performance. Is there any known codebase where we would expect this to be beneficial? If not perhaps the issue can be closed.