perf.rlo shows that running the `ConstProp` pass results in
across-the-board wins regardless of debug or opt complilation mode. As a
result, we're turning it on to get the compile time benefits.
`ConstProp` doesn't currently intern the memory used by its `Machine` so
we can't yet propagate allocations which is why
`ConstProp::should_const_prop()` checks if the value being propagated is
a scalar or not.
The `ConstProp` can cause many locals to be initialized to a constant
value and then never read from. `ConstProp` can also evaluate ZSTs into
constant values. Previously, many of these would be removed by other
parts of the MIR optimization pipeline. However, evaluating ZSTs
(especially `()`) into constant values defeated those parts of the
optimizer and so in a2e3ed5c05, I added a
hack to `ConstProp` that skips evaluating ZSTs to avoid that regression.
This commit changes `SimplifyLocals` so that it doesn't consider writes
of const values to a local to be a use of that local. In doing so,
`SimplifyLocals` is able to remove otherwise unused locals left behind
by other optimization passes (`ConstProp` in particular).
