Optimize reachability with non-mutating global passes

[AlexanderPortland]

Context

When transforming MIR, we use both TransformPass-es that operate over a single body and GlobalPass-es that require further context thus and operate over the whole program. Since these global passes could potentially modify the bodies they operate over, we currently re-run reachability after them to ensure that any changes are reflected in our reachability output.

However, our global passes are largely gated behind unstable features (e.g. -Z uninit-checks to check for uninitialized memory) or specific user actions (e.g. compiling w/ RUSTFLAGS="--emit mir"), so this extra reachability collection is often done unnecessarily. Furthermore, some global passes (like emitting the generated MIR) do not modify the bodies they interact with at all, so re-doing collection is unneeded even when they are enabled.

Changes

This PR modifies the GlobalPass trait to return a boolean representing if the pass modified the MIR in a way that could affect reachability. We then use the boolean flags for each result to redo reachability analysis only if a GlobalPass could have affected its result.

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 and MIT licenses.

[AlexanderPortland]

For context, I intentionally implemented this as separate traits rather than having GlobalPass::transform() return a boolean as we do for TransformPass::transform() (this seemed more explicit and elegant type-wise), but I'm very open to feedback and can switch if the latter seems preferable!

[AlexanderPortland]

I actually just switched this to use the same boolean strategy as TransformPass. I love my RefCell<T>s, but this made the changes a lot cleaner without the same runtime overhead.

Different implementation now, so old review doesn't appy; I can do the re-review next week.

[carolynzech]

Thanks!