Optimize seen Predicate filtering. #50932
Conversation
This speeds up a few rustc-perf benchmark runs, most notably ones involving 'coercions', the best by 2%.
|
r? @eddyb (rust_highfive has picked a reviewer for you, use r? to override) |
rust-highfive
added
the
S-waiting-on-review
| // | ||
| // This code is hot enough that it's worth avoiding the allocation | ||
| // required for the FxHashSet when possible. Special-casing lengths 0, | ||
| // 1 and 2 covers roughly 75--80% of the cases. |
There was a problem hiding this comment.
What about using sort + dedup? That would avoid the allocation, and also quickly exit on small numbers of elements.
There was a problem hiding this comment.
That changes the order, which may break the behavior.
Also, deriving Ord is a pain.
There was a problem hiding this comment.
And Predicate is a large type, so moving the values around might be moderately expensive.
There was a problem hiding this comment.
Oh, I see, the set is used with retain. But you need to clone & move the values into a set anyway. If you want to avoid the cost of data movement, the set could be created using references into the original Vec instead. Except that doesn't work because of the retain...
|
@bors r+ |
|
📌 Commit 2ff6324 has been approved by |
bors
added
S-waiting-on-bors
Optimize seen Predicate filtering. This speeds up a few rustc-perf benchmark runs, most notably ones involving 'coercions', the best by 2%.
Rollup of 15 pull requests Successful merges: - #50846 (Add E0665) - #50849 (CheckLoopVisitor: also visit closure arguments) - #50863 (Make `[T]::len` and `str::len` const fn) - #50875 (rustdoc: use "short form" doc(cfg) printing even when combined with other conditionals) - #50913 (Fix typo in cell.rs) - #50914 (Issue #50636: Improve error diagnostic with missing commas after struct fields.) - #50931 (Inline `try_get`.) - #50932 (Optimize seen Predicate filtering.) - #50945 (Stabilize feature from_ref) - #50946 (rustc: Fix procedural macros generating lifetime tokens) - #50947 (rustdoc: set tab width in rust source blocks) - #50952 (Add the 2018 edition of the book to doc.rust-lang.org) - #50958 (Micro-optimization on PR#50697) - #50961 (Fix FileCheck finding with MSVC) - #50963 (Right-size the `VecDeque` in `coerce_unsized`.) Failed merges:
This speeds up a few rustc-perf benchmark runs, most notably ones
involving 'coercions', the best by 2%.