[Amanieu]

Depends on #163338

This PR extends the MoveElimination pass from rust-lang/rfcs#3943 to reconstruct StorageLive/StorageDead statements for all locals by using the liveness information collected by PreciseLiveness.

r? tmiasko

[Amanieu]

LLM disclosure: LLMs were used to generate the boilerplate for the MIR tests, but each test was manually reviewed and adjusted before committing. LLMs were also used to audit this pass against the rest of the compiler MIR transforms and Miri semantics.

[rustbot]

Some changes occurred to MIR optimizations

cc @rust-lang/wg-mir-opt

[dianqk]

I think this is a bit similar to #142531. I don't actually see any mir-opt pass that uses storage's UB in the real world. Should we have a dedicated StorageReconstruction pass placed later in the pipeline?
When using a local at a new location, we just remove the storage marks. StorageReconstruction will bring them back.
I think this should benefit other passes, such as Inline.

[tmiasko]

Removing storage statements and reconstructing them later, would lose information about ends of storage ranges. Consider a local which is borrowed and never moved. The best we could do, is to end its storage when function returns.

Ideally we would make decisions based on an actual measurements. We could extend perf benchmarks, e.g., by including a sum of stack adjustments from function prologues as an approximation of stack usage.

[Amanieu]

Essentially what @tmiasko said. The main issue is that place unification has deleted all storage statements for merged locals, so at that point the only record of where local lifetimes end is in the post-unification liveness matrix. This matrix must be used as the source for storage reconstruction within the pass.