Compress amount of hashed bytes for `isize` values in StableHasher
This is another attempt to land https://github.com/rust-lang/rust/pull/92103, this time hopefully with a correct implementation w.r.t. stable hashing guarantees. The previous PR was [reverted](https://github.com/rust-lang/rust/pull/93014) because it could produce the [same hash](https://github.com/rust-lang/rust/pull/92103#issuecomment-1014625442) for different values even in quite simple situations. I have since added a basic [test](https://github.com/rust-lang/rust/pull/93193) that should guard against that situation, I also added a new test in this PR, specialised for this optimization.
## Why this optimization helps
Since the original PR, I have tried to analyze why this optimization even helps (and why it especially helps for `clap`). I found that the vast majority of stable-hashing `i64` actually comes from hashing `isize` (which is converted to `i64` in the stable hasher). I only found a single place where is this datatype used directly in the compiler, and this place has also been showing up in traces that I used to find out when is `isize` being hashed. This place is `rustc_span::FileName::DocTest`, however, I suppose that isizes also come from other places, but they might not be so easy to find (there were some other entries in the trace). `clap` hashes about 8.5 million `isize`s, and all of them fit into a single byte, which is why this optimization has helped it [quite a lot](https://github.com/rust-lang/rust/pull/92103#issuecomment-1005711861).
Now, I'm not sure if special casing `isize` is the correct solution here, maybe something could be done with that `isize` inside `DocTest` or in other places, but that's for another discussion I suppose. In this PR, instead of hardcoding a special case inside `SipHasher128`, I instead put it into `StableHasher`, and only used it for `isize` (I tested that for `i64` it doesn't help, or at least not for `clap` and other few benchmarks that I was testing).
## New approach
Since the most common case is a single byte, I added a fast path for hashing `isize` values which positive value fits within a single byte, and a cold path for the rest of the values.
To avoid the previous correctness problem, we need to make sure that each unique `isize` value will produce a unique hash stream to the hasher. By hash stream I mean a sequence of bytes that will be hashed (a different sequence should produce a different hash, but that is of course not guaranteed).
We have to distinguish different values that produce the same bit pattern when we combine them. For example, if we just simply skipped the leading zero bytes for values that fit within a single byte, `(0xFF, 0xFFFFFFFFFFFFFFFF)` and `(0xFFFFFFFFFFFFFFFF, 0xFF)` would send the same hash stream to the hasher, which must not happen.
To avoid this situation, values `[0, 0xFE]` are hashed as a single byte. When we hash a larger (treating `isize` as `u64`) value, we first hash an additional byte `0xFF`. Since `0xFF` cannot occur when we apply the single byte optimization, we guarantee that the hash streams will be unique when hashing two values `(a, b)` and `(b, a)` if `a != b`:
1) When both `a` and `b` are within `[0, 0xFE]`, their hash streams will be different.
2) When neither `a` and `b` are within `[0, 0xFE]`, their hash streams will be different.
3) When `a` is within `[0, 0xFE]` and `b` isn't, when we hash `(a, b)`, the hash stream will definitely not begin with `0xFF`. When we hash `(b, a)`, the hash stream will definitely begin with `0xFF`. Therefore the hash streams will be different.
r? `@the8472`
Make `Fingerprint::combine_commutative` associative
The previous implementation swapped lower and upper 64-bits of a result
of modular addition, so the function was non-associative.
r? `@Aaron1011`
Add test for stable hash uniqueness of adjacent field values
This PR adds a simple test to check that stable hash will produce a different hash if the order of two values that have the same combined bit pattern changes.
r? `@the8472`
`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.
Update rayon and rustc-rayon
This updates rayon for various tools and rustc-rayon for the compiler's parallel mode.
- rayon v1.3.1 -> v1.5.1
- rayon-core v1.7.1 -> v1.9.1
- rustc-rayon v0.3.1 -> v0.3.2
- rustc-rayon-core v0.3.1 -> v0.3.2
... and indirectly, this updates all of crossbeam-* to their latest versions.
Fixes#92677 by removing crossbeam-queue, but there's still a lingering question about how tidy discovers "runtime" dependencies. None of this is truly in the standard library's dependency tree at all.
Replace usages of vec![].into_iter with [].into_iter
`[].into_iter` is idiomatic over `vec![].into_iter` because its simpler and faster (unless the vec is optimized away in which case it would be the same)
So we should change all the implementation, documentation and tests to use it.
I skipped:
* `src/tools` - Those are copied in from upstream
* `src/test/ui` - Hard to tell if `vec![].into_iter` was used intentionally or not here and not much benefit to changing it.
* any case where `vec![].into_iter` was used because we specifically needed a `Vec::IntoIter<T>`
* any case where it looked like we were intentionally using `vec![].into_iter` to test it.
Fixes#92266
In some `HashStable` impls, we use a cache to avoid re-computing
the same `Fingerprint` from the same structure (e.g. an `AdtDef`).
However, the `StableHashingContext` used can be configured to
perform hashing in different ways (e.g. skipping `Span`s). This
configuration information is not included in the cache key,
which will cause an incorrect `Fingerprint` to be used if
we hash the same structure with different `StableHashingContext`
settings.
To fix this, the configuration settings of `StableHashingContext`
are split out into a separate `HashingControls` struct. This
struct is used as part of the cache key, ensuring that our caches
always produce the correct result for the given settings.
With this in place, we now turn off `Span` hashing during the
entire process of computing the hash included in legacy symbols.
This current has no effect, but will matter when a future PR
starts hashing more `Span`s that we currently skip.
Implement StableHash for BitSet and BitMatrix via Hash
This fixes an issue where bit sets / bit matrices the same word
content but a different domain size would receive the same hash.
Avoid sorting in hash map stable hashing
Suggested by `@the8472` [here](https://github.com/rust-lang/rust/pull/89404#issuecomment-991813333). I hope that I understood it right, I replaced the sort with modular multiplication, which should be commutative.
Can I ask for a perf. run? However, locally it didn't help at all. Creating the `StableHasher` all over again is probably slowing it down quite a lot. And using `FxHasher` is not straightforward, because the keys and values only implement `HashStable` (and probably they shouldn't be just hashed via `Hash` anyway for it to actually be stable).
Maybe the `StableHash` interface could be changed somehow to better suppor these scenarios where the hasher is short-lived. Or the `StableHasher` implementation could have variants with e.g. a shorter buffer for these scenarios.
Slightly optimize hash map stable hashing
I was profiling some of the `rustc-perf` benchmarks locally and noticed that quite some time is spent inside the stable hash of hashmaps. I tried to use a `SmallVec` instead of a `Vec` there, which helped very slightly.
Then I tried to remove the sorting, which was a bottleneck, and replaced it with insertion into a binary heap. Locally, it yielded nice improvements in instruction counts and RSS in several benchmarks for incremental builds. The implementation could probably be much nicer and possibly extended to other stable hashes, but first I wanted to test the perf impact properly.
Can I ask someone to do a perf run? Thank you!
This largely avoids remapping from and to the 'real' indices, with the exception
of predecessor lookup and the final merge back, and is conceptually better.
As the paper indicates, the unprocessed vertices in the DFS tree and processed
vertices are disjoint, and we can use them in the same space, tracking only the index
of the split.
This replaces the previous implementation with the simple variant of
Lengauer-Tarjan, which performs better in the general case. Performance on the
keccak benchmark is about equivalent between the two, but we don't see
regressions (and indeed see improvements) on other benchmarks, even on a
partially optimized implementation.
The implementation here follows that of the pseudocode in "Linear-Time
Algorithms for Dominators and Related Problems" thesis by Loukas Georgiadis. The
next few commits will optimize the implementation as suggested in the thesis.
Several related works are cited in the comments within the implementation, as
well.
Implement the simple Lengauer-Tarjan algorithm
This replaces the previous implementation (from #34169), which has not been
optimized since, with the simple variant of Lengauer-Tarjan which performs
better in the general case. A previous attempt -- not kept in commit history --
attempted a replacement with a bitset-based implementation, but this led to
regressions on perf.rust-lang.org benchmarks and equivalent wins for the keccak
benchmark, so was rejected.
The implementation here follows that of the pseudocode in "Linear-Time
Algorithms for Dominators and Related Problems" thesis by Loukas Georgiadis. The
next few commits will optimize the implementation as suggested in the thesis.
Several related works are cited in the comments within the implementation, as
well.
On the keccak benchmark, we were previously spending 15% of our cycles computing
the NCA / intersect function; this function is quite expensive, especially on
modern CPUs, as it chases pointers on every iteration in a tight loop. With this
commit, we spend ~0.05% of our time in dominator computation.
There's a conversation in the tracking issue about possibly unaccepting `in_band_lifetimes`, but it's used heavily in the compiler, and thus there'd need to be a bunch of PRs like this if that were to happen.
So here's one to see how much of an impact it has.
(Oh, and I removed `nll` while I was here too, since it didn't seem needed. Let me know if I should put that back.)
