This patch supports the syntax

    unchecked {
      ...
    }

    to disable purity checking within a block. Presumably it will only be
    used within a declared "pure fn". However, there is no checking that it
    doesn't occur elsewhere, and it would be harmless for it to do so.

    I went with Lindsey's suggestion for the syntax, but it's subject to
    change.

    This allows you to write code that uses predicates that call arbitrary
    Rust functions, but you must declare your intentions by wrapping it in
    an unchecked { ... } block. The test case run-pass/unchecked-predicates.rs
    demonstrates how to do that.

rt: Null check in walk_obj_contents. Prevents marking from crashing when calling object constructors.

rt: Factor out type param construction into from_tydesc and from_obj_shape. (The latter is unimplemented at the moment.)

This is part 1 of changing the "pred" keyword to "pure fn".
Right now, the compiler accepts both "pred" and "pure fn".

This reverts commit 629ee94a.

This wasn't a good idea after all.

This seems to be faster than creating separate ones for each block
context.

You now do

    bld::Ret(bcx, someval)

where you used to say

    bcx.build.Ret(someval)

Two fewer boxes are allocated for each block context, and build calls
no longer go through a vtable.

This is a preparation for making vectors always-on-the-heap again,
which would cause way too much malloc traffic for this idiom. I will
add an efficient std::vec::push in the future, and migrate += [x] to
that instead.

Reduces compiler code size by 3%

Closes #861

Add kind-checking for assign-op, copy, ret, be, fail exprs. Fix caught kinding-violations in rustc and libstd.

This file has suffered a lot of reformats and was looking pretty ragged