Description
Continuing from discussion in #69716.
MaybeRequiresStorage
(introduced in #60187 as RequiresStorage
) is one of four dataflow analyses used to determine what locals are live across yield points or live at the same time as other locals during the MIR generator pass. I don't see any reason we're not using a liveness analysis exclusively for this, with an additional check against MaybeBorrowedLocals ∩ MaybeStorageLive
in cases where borrows can live across suspension points (immovable generators).
MaybeRequiresStorage
is basically MaybeStorageLive
with some additional logic that kills locals after they are moved out of. It also incorporates the MaybeBorrowedLocals
analysis as part of its transfer function, which is probably unnecessary, since we could check the results of MaybeBorrowedLocals
directly at each yield point. If there is often a large gap between the point at which a local is moved out of and the point at which it is marked StorageDead
, we could augment MaybeStorageLive
to kill locals when they are moved out of.
The liveness analysis in librustc_mir/util/liveness.rs
will implicitly handle moved-from locals as long as there are no future accesses to them, which is what causes a variable to be marked as live.
I'd like refactor the liveness parts of the MIR generator pass, but will wait for buy-in from its authors. I think that basing the current storage optimization on a combination of a few well-known dataflow analyses will make it easier to reason about and therefore fine-tune. This will involve using the liveness analysis (plus MaybeBorrowedLocals
/ MaybeStorageLive
inside immovable generators) for compute_storage_conflicts
. Does anyone see a problem with this? @tmandry @jonas-schievink