Free intermediate translation results as soon as possible

[dotdash]

This fixes the recently introduced peak memory usage regression by
freeing the intermediate results as soon as they're not required
anymore instead of keeping them around for the whole compilation
process.

Refs #8077

[alexcrichton]

Could you add a comment above it to explain why it looks a bit silly? Otherwise r+!

[thestinger]

@dotdash: it might be clearer to use util::ignore as an early destructor call

[dotdash]

Does that work with @-ptrs? Looks like it would just cause an inc + dec on the refcount.

[alexcrichton]

Sadly I think it does refcount business with util::ignore:

$ cat foo.rs
struct A;
impl Drop for A {
fn drop(&self) {
println("drop");
}
}
fn main() {
let a = @A;
std::util::ignore(a);
println("here");
}
$ rust run foo.rs
warning: no debug symbols in executable (-arch x86_64)
here
drop

[thestinger]

Ah, didn't realize these were managed pointers. Although, these should really be Rc rather than managed pointers which aren't going to be freed immediately when the garbage collector lands. The Rc type does have move semantics, rather than implicit copies, as any reference-counted type should.

[graydon]

Doh! I actually had it written this way but in a later commit in the wip branch I was working on this, because I was trying (and failing, due to too many @ boxes in the session) to put the earlier phases in a task and dump it (freeing the whole AST) before running LLVM passes. Thanks.