Caching the stable hash of Ty within itself
Instead of computing stable hashes on types as needed, we compute it during interning.
This way we can, when a hash is requested, just hash that hash, which is significantly faster than traversing the type itself.
We only do this for incremental for now, as incremental is the only frequent user of stable hashing.
As a next step we can try out
* moving the hash and TypeFlags to Interner, so projections and regions get the same benefit (tho regions are not nested, so maybe that's not a good idea? Would be nice for dedup tho)
* start comparing types via their stable hash instead of their address?
Specifically, rename the `Const` struct as `ConstS` and re-introduce `Const` as
this:
```
pub struct Const<'tcx>(&'tcx Interned<ConstS>);
```
This now matches `Ty` and `Predicate` more closely, including using
pointer-based `eq` and `hash`.
Notable changes:
- `mk_const` now takes a `ConstS`.
- `Const` was copy, despite being 48 bytes. Now `ConstS` is not, so need a
we need separate arena for it, because we can't use the `Dropless` one any
more.
- Many `&'tcx Const<'tcx>`/`&Const<'tcx>` to `Const<'tcx>` changes
- Many `ct.ty` to `ct.ty()` and `ct.val` to `ct.val()` changes.
- Lots of tedious sigil fiddling.
The variant names are exported, so we can use them directly (possibly
with a `ty::` qualifier). Lots of places already do this, this commit
just increases consistency.
Specifically, change `Region` from this:
```
pub type Region<'tcx> = &'tcx RegionKind;
```
to this:
```
pub struct Region<'tcx>(&'tcx Interned<RegionKind>);
```
This now matches `Ty` and `Predicate` more closely.
Things to note
- Regions have always been interned, but we haven't been using pointer-based
`Eq` and `Hash`. This is now happening.
- I chose to impl `Deref` for `Region` because it makes pattern matching a lot
nicer, and `Region` can be viewed as just a smart wrapper for `RegionKind`.
- Various methods are moved from `RegionKind` to `Region`.
- There is a lot of tedious sigil changes.
- A couple of types like `HighlightBuilder`, `RegionHighlightMode` now have a
`'tcx` lifetime because they hold a `Ty<'tcx>`, so they can call `mk_region`.
- A couple of test outputs change slightly, I'm not sure why, but the new
outputs are a little better.
Specifically, change `Ty` from this:
```
pub struct Predicate<'tcx> { inner: &'tcx PredicateInner<'tcx> }
```
to this:
```
pub struct Predicate<'tcx>(&'tcx Interned<PredicateS<'tcx>>)
```
where `PredicateInner` is renamed as `PredicateS`.
This (plus a few other minor changes) makes the parallels with `Ty` and
`TyS` much clearer, and makes the uniqueness more explicit.
Specifically, change `Ty` from this:
```
pub type Ty<'tcx> = &'tcx TyS<'tcx>;
```
to this
```
pub struct Ty<'tcx>(Interned<'tcx, TyS<'tcx>>);
```
There are two benefits to this.
- It's now a first class type, so we can define methods on it. This
means we can move a lot of methods away from `TyS`, leaving `TyS` as a
barely-used type, which is appropriate given that it's not meant to
be used directly.
- The uniqueness requirement is now explicit, via the `Interned` type.
E.g. the pointer-based `Eq` and `Hash` comes from `Interned`, rather
than via `TyS`, which wasn't obvious at all.
Much of this commit is boring churn. The interesting changes are in
these files:
- compiler/rustc_middle/src/arena.rs
- compiler/rustc_middle/src/mir/visit.rs
- compiler/rustc_middle/src/ty/context.rs
- compiler/rustc_middle/src/ty/mod.rs
Specifically:
- Most mentions of `TyS` are removed. It's very much a dumb struct now;
`Ty` has all the smarts.
- `TyS` now has `crate` visibility instead of `pub`.
- `TyS::make_for_test` is removed in favour of the static `BOOL_TY`,
which just works better with the new structure.
- The `Eq`/`Ord`/`Hash` impls are removed from `TyS`. `Interned`s impls
of `Eq`/`Hash` now suffice. `Ord` is now partly on `Interned`
(pointer-based, for the `Equal` case) and partly on `TyS`
(contents-based, for the other cases).
- There are many tedious sigil adjustments, i.e. adding or removing `*`
or `&`. They seem to be unavoidable.
Ensure that queries only return Copy types.
This should pervent the perf footgun of returning a result with an expensive `Clone` impl (like a `Vec` of a hash map).
I went for the stupid solution of allocating on an arena everything that was not `Copy`. Some query results could be made Copy easily, but I did not really investigate.
Refactor query system to maintain a global job id counter
This replaces the per-shard counters with a single global counter, simplifying
the JobId struct down to just a u64 and removing the need to pipe a DepKind
generic through a bunch of code. The performance implications on non-parallel
compilers are likely minimal (this switches to `Cell<u64>` as the backing
storage over a `u64`, but the latter was already inside a `RefCell` so it's not
really a significance divergence). On parallel compilers, the cost of a single
global u64 counter may be more significant: it adds a serialization point in
theory. On the other hand, we can imagine changing the counter to have a
thread-local component if it becomes worrisome or some similar structure.
The new design is sufficiently simpler that it warrants the potential for slight
changes down the line if/when we get parallel compilation to be more of a
default.
A u64 counter, instead of u32 (the old per-shard width), is chosen to avoid
possibly overflowing it and causing problems; it is effectively impossible that
we would overflow a u64 counter in this context.
This replaces the per-shard counters with a single global counter, simplifying
the JobId struct down to just a u64 and removing the need to pipe a DepKind
generic through a bunch of code. The performance implications on non-parallel
compilers are likely minimal (this switches to `Cell<u64>` as the backing
storage over a `u64`, but the latter was already inside a `RefCell` so it's not
really a significance divergence). On parallel compilers, the cost of a single
global u64 counter may be more significant: it adds a serialization point in
theory. On the other hand, we can imagine changing the counter to have a
thread-local component if it becomes worrisome or some similar structure.
The new design is sufficiently simpler that it warrants the potential for slight
changes down the line if/when we get parallel compilation to be more of a
default.
A u64 counter, instead of u32 (the old per-shard width), is chosen to avoid
possibly overflowing it and causing problems; it is effectively impossible that
we would overflow a u64 counter in this context.
Store def_id_to_hir_id as variant in hir_owner.
If hir_owner is Owner(_), the LocalDefId is pointing to an owner, so the ItemLocalId is 0.
If the HIR node does not exist, we store Phantom.
Otherwise, we store the HirId associated to the LocalDefId.
Related to #89278
r? `@oli-obk`
If hir_owner is Owner(_), the LocalDefId is pointing to an owner, so the ItemLocalId is 0.
If the HIR node does not exist, we store Phantom.
Otherwise, we store the HirId associated to the LocalDefId.
Make `Decodable` and `Decoder` infallible.
`Decoder` has two impls:
- opaque: this impl is already partly infallible, i.e. in some places it
currently panics on failure (e.g. if the input is too short, or on a
bad `Result` discriminant), and in some places it returns an error
(e.g. on a bad `Option` discriminant). The number of places where
either happens is surprisingly small, just because the binary
representation has very little redundancy and a lot of input reading
can occur even on malformed data.
- json: this impl is fully fallible, but it's only used (a) for the
`.rlink` file production, and there's a `FIXME` comment suggesting it
should change to a binary format, and (b) in a few tests in
non-fundamental ways. Indeed #85993 is open to remove it entirely.
And the top-level places in the compiler that call into decoding just
abort on error anyway. So the fallibility is providing little value, and
getting rid of it leads to some non-trivial performance improvements.
Much of this PR is pretty boring and mechanical. Some notes about
a few interesting parts:
- The commit removes `Decoder::{Error,error}`.
- `InternIteratorElement::intern_with`: the impl for `T` now has the same
optimization for small counts that the impl for `Result<T, E>` has,
because it's now much hotter.
- Decodable impls for SmallVec, LinkedList, VecDeque now all use
`collect`, which is nice; the one for `Vec` uses unsafe code, because
that gave better perf on some benchmarks.
r? `@bjorn3`
`Decoder` has two impls:
- opaque: this impl is already partly infallible, i.e. in some places it
currently panics on failure (e.g. if the input is too short, or on a
bad `Result` discriminant), and in some places it returns an error
(e.g. on a bad `Option` discriminant). The number of places where
either happens is surprisingly small, just because the binary
representation has very little redundancy and a lot of input reading
can occur even on malformed data.
- json: this impl is fully fallible, but it's only used (a) for the
`.rlink` file production, and there's a `FIXME` comment suggesting it
should change to a binary format, and (b) in a few tests in
non-fundamental ways. Indeed #85993 is open to remove it entirely.
And the top-level places in the compiler that call into decoding just
abort on error anyway. So the fallibility is providing little value, and
getting rid of it leads to some non-trivial performance improvements.
Much of this commit is pretty boring and mechanical. Some notes about
a few interesting parts:
- The commit removes `Decoder::{Error,error}`.
- `InternIteratorElement::intern_with`: the impl for `T` now has the same
optimization for small counts that the impl for `Result<T, E>` has,
because it's now much hotter.
- Decodable impls for SmallVec, LinkedList, VecDeque now all use
`collect`, which is nice; the one for `Vec` uses unsafe code, because
that gave better perf on some benchmarks.
I have found this code very confusing at times. This commit clarifies
things.
In particular, the commit explains the requirements that the `Borrow`
impls put on the `Eq` and `Hash` impls, which are non-obvious. And it
puts the `Borrow` impls first, since they force `Eq` and `Hash` to have
particular forms.
The commit also notes `TyS`'s uniqueness requirements.
Rollup of 17 pull requests
Successful merges:
- #91032 (Introduce drop range tracking to generator interior analysis)
- #92856 (Exclude "test" from doc_auto_cfg)
- #92860 (Fix errors on blanket impls by ignoring the children of generated impls)
- #93038 (Fix star handling in block doc comments)
- #93061 (Only suggest adding `!` to expressions that can be macro invocation)
- #93067 (rustdoc mobile: fix scroll offset when jumping to internal id)
- #93086 (Add tests to ensure that `let_chains` works with `if_let_guard`)
- #93087 (Fix src/test/run-make/raw-dylib-alt-calling-convention)
- #93091 (⬆ chalk to 0.76.0)
- #93094 (src/test/rustdoc-json: Check for `struct_field`s in `variant_tuple_struct.rs`)
- #93098 (Show a more informative panic message when `DefPathHash` does not exist)
- #93099 (rustdoc: auto create output directory when "--output-format json")
- #93102 (Pretty printer algorithm revamp step 3)
- #93104 (Support --bless for pp-exact pretty printer tests)
- #93114 (update comment for `ensure_monomorphic_enough`)
- #93128 (Add script to prevent point releases with same number as existing ones)
- #93136 (Backport the 1.58.1 release notes to master)
Failed merges:
r? `@ghost`
`@rustbot` modify labels: rollup
The field is also renamed from `ident` to `name. In most cases,
we don't actually need the `Span`. A new `ident` method is added
to `VariantDef` and `FieldDef`, which constructs the full `Ident`
using `tcx.def_ident_span()`. This method is used in the cases
where we actually need an `Ident`.
This makes incremental compilation properly track changes
to the `Span`, without all of the invalidations caused by storing
a `Span` directly via an `Ident`.
