Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Auto merge of #60187 - tmandry:generator-optimization, r=eddyb
Generator optimization: Overlap locals that never have storage live at the same time The specific goal of this optimization is to optimize async fns which use `await!`. Notably, `await!` has an enclosing scope around the futures it awaits ([definition](https://github.com/rust-lang/rust/blob/08bfe16129b0621bc90184f8704523d4929695ef/src/libstd/macros.rs#L365-L381)), which we rely on to implement the optimization. More generally, the optimization allows overlapping the storage of some locals which are never storage-live at the same time. **We care about storage-liveness when computing the layout, because knowing a field is `StorageDead` is the only way to prove it will not be accessed, either directly or through a reference.** To determine whether we can overlap two locals in the generator layout, we look at whether they might *both* be `StorageLive` at any point in the MIR. We use the `MaybeStorageLive` dataflow analysis for this. We iterate over every location in the MIR, and build a bitset for each local of the locals it might potentially conflict with. Next, we assign every saved local to one or more variants. The variants correspond to suspension points, and we include the set of locals live across a given suspension point in the variant. (Note that we use liveness instead of storage-liveness here; this ensures that the local has actually been initialized in each variant it has been included in. If the local is not live across a suspension point, then it doesn't need to be included in that variant.). It's important to note that the variants are a "view" into our layout. For the layout computation, we use a simplified approach. 1. Start with the set of locals assigned to only one variant. The rest are disqualified. 2. For each pair of locals which may conflict *and are not assigned to the same variant*, we pick one local to disqualify from overlapping. Disqualified locals go into a non-overlapping "prefix" at the beginning of our layout. This means they always have space reserved for them. All the locals that are allowed to overlap in each variant are then laid out after this prefix, in the "overlap zone". So, if A and B were disqualified, and X, Y, and Z were all eligible for overlap, our generator might look something like this: You can think of a generator as an enum, where some fields are shared between variants. e.g. ```rust enum Generator { Unresumed, Poisoned, Returned, Suspend0(A, B, X), Suspend1(B), Suspend2(A, Y, Z), } ``` where every mention of `A` and `B` refer to the same field, which does not move when changing variants. Note that `A` and `B` would automatically be sent to the prefix in this example. Assuming that `X` is never `StorageLive` at the same time as either `Y` or `Z`, it would be allowed to overlap with them. Note that if two locals (`Y` and `Z` in this case) are assigned to the same variant in our generator, their memory would never overlap in the layout. Thus they can both be eligible for the overlapping section, even if they are storage-live at the same time. --- Depends on: - [x] #59897 Multi-variant layouts for generators - [x] #60840 Preserve local scopes in generator MIR - [x] #61373 Emit StorageDead along unwind paths for generators Before merging: - [x] ~Wrap the types of all generator fields in `MaybeUninitialized` in layout::ty::field~ (opened #60889) - [x] Make PR description more complete (e.g. explain why storage liveness is important and why we have to check every location) - [x] Clean up TODO - [x] Fix the layout code to enforce that the same field never moves around in the generator - [x] Add tests for async/await - [x] ~Reduce # bits we store by half, since the conflict relation is symmetric~ (note: decided not to do this, for simplicity) - [x] Store liveness information for each yield point in our `GeneratorLayout`, that way we can emit more useful debuginfo AND tell miri which fields are definitely initialized for a given variant (see discussion at #59897 (comment))
- Loading branch information
