perf(resolver): track per-scope extractable state at phase level#72
Merged
Conversation
`PackageVariantSlice::extractable` was a `HashSet::is_subset` call that iterated `common_fams` and tested membership in `extracted_fams` on every probe. Fresh callgrind on the rez 188-case benchmark (after #66/#67/#68/#70) put `HashSet::is_subset` at **11.8 %** of inclusive cycles — nearly half of `PackageVariantSlice::extract`'s total 25.3 %. Every `extract()` call hits this guard; the vast majority return early with "nothing left to extract." The set operation is unnecessary. `extracted_fams` is only ever populated by inserting an element of `common_fams.difference( extracted_fams)` in `extract`, and `copy_with_entries` resets it to empty. So `extracted_fams ⊆ common_fams` always holds, and under that invariant the `is_subset` check is equivalent to a length compare: common_fams.is_subset(extracted_fams) ⟺ common_fams ⊆ extracted_fams ⟺ (since extracted ⊆ common) common_fams == extracted_fams ⟺ |common| == |extracted| Replace the body with `common_fams().len() > extracted_fams.len()`. ## Benchmark (188 cases, release, same machine, two runs) | Stage | Total | Mean | vs rez | |--------------------------------------|--------:|-------:|-------:| | Baseline (main, post-#70) | 18.6 s | 99 ms | 20.5× | | + this change, run 1 | 13.1 s | 70 ms | 29.1× | | + this change, run 2 | 13.9 s | 74 ms | 27.4× | **-30 % on top of #70**, **-69 % cumulative from main** (43.0 s → 13.1 s, 8.8× rez → 29.1× rez). Differential test got the same lift: 188/188 still match rez 1:1, in 20.77 s (down from 27.67 s). Predicted gain was 5–10 %. Like #68, hidden downstream costs (the `is_subset` iterator setup/teardown, the hash lookups it performed, and the now-unnecessary `common_fams_cache` first-time computation on slices that never need extraction) made the actual gain larger. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
After #71 made `PackageVariantSlice::extractable` O(1), every `extract` call's no-op early-return is nearly free — but the function-call round-trip itself is still paid on every probe (tens of millions over the benchmark). Each iteration of the inner extract loop in `ResolvePhase::solve` walks every scope and calls `extract`, even on scopes that have already exhausted their common dependencies. A scope's extractability only changes when its entries change. Track a per-scope flag and skip the call once `extract` has returned `None`, until intersect / reduce / widen replaces the scope (in which case the flag resets to "might be extractable again"): - After `extract` returns `None` for `scope[i]`: set `non_extractable[i]`. - `ScopeIntersect::Narrowed`: clear it. - `ScopeReduce::Reduced`: clear it. - New scope added: pushed as `false`. Correctness invariant: `non_extractable[i] == true` ⇒ the latest `scope[i]` was just observed to have no common dependencies left. Any operation that would expand or rebuild the variant slice resets the flag. ## Benchmark (188 cases, release, same machine) | Stage | Total | Mean | vs rez | |------------------------------------|--------:|-------:|-------:| | Baseline (post-#71), median | ~13.5 s | 87 ms | ~28× | | + this change, run 1 | 13.1 s | 69 ms | 29.3× | | + this change, run 2 | 12.4 s | 66 ms | 30.8× | Modest, ~5 % typical. Predicted 2–4 %. The gain shrunk because #71 already collapsed the `extract` body to a length compare — there is much less function-call work left to skip than there would have been on the pre-#71 codebase. Cumulative from main: 43.0 s → 12.4 s, 8.8× → 30.8× rez. 188/188 differential still matches rez 1:1 (`cargo test … --ignored`, 20.7 s). Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Qodo reviews are paused for this user.Troubleshooting steps vary by plan Learn more → On a Teams plan? Using GitHub Enterprise Server, GitLab Self-Managed, or Bitbucket Data Center? |
doubleailes
added a commit
that referenced
this pull request
May 15, 2026
Callgrind on rez's 188-case benchmark (post-#71/#72) showed `SmallVec::extend` + `Drop` at ~4 % of cycles, almost entirely from `VersionRange::clone`. Every `Requirement::clone()` (in `extracted_request.clone()`, the per-pair `package_request.clone()` in `reduce_by`, the `req.clone()` and `package_request.clone()` in `Reduction`, etc.) deep-copies the inner `Ranges`'s `SmallVec` of `(Bound, Bound)` segments. After the rest of the perf stack (#66/#67/#68/#70/#71/#72), this is the largest non-amortised allocation cost left. Switch the inner from `Ranges<RerVersion>` to `Rc<Ranges<RerVersion>>`. `Rc<T>::clone` is a refcount bump; `Rc<T>::Hash`/`Eq` defer to the inner `T`, so the derived semantics on `VersionRange` are unchanged. Methods that build a new range (`intersection`, `union`, `complement`, `from_versions`, `span`, `split`, ...) still produce a fresh `Ranges` internally and wrap it with `Rc::new` — the win is on the read / clone path, not the construction path. `as_ranges()` still returns `&Ranges` (via `Rc::deref`). `into_ranges` now uses `Rc::unwrap_or_clone` — falls back to a clone if the `Rc` is shared, but is the consume-the-`VersionRange` API and rare in practice. ## Benchmark (188 cases, release, same machine, two runs) | Stage | Total | Mean | vs rez | |------------------------------------|--------:|-------:|-------:| | Baseline (post-#71/#72), median | ~12.7 s | 68 ms | ~30× | | + this change, run 1 | 11.2 s | 60 ms | 34.1× | | + this change, run 2 | 11.3 s | 60 ms | 33.7× | **~11 % on top of #72**, **~74 % cumulative from main** (43.0 s → 11.2 s, 8.8× rez → 34.1× rez). Differential test (`cargo test … --ignored`): 17.73 s, **188/188 still match rez 1:1**. Predicted 3–5 %. The slightly bigger gain reflects that `VersionRange::clone` cascades into a lot more than just the `SmallVec::extend` it was attributed to in the callgrind exclusive view — it also drove allocator-side work and the matching `Drop`s. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
3 tasks
doubleailes
added a commit
that referenced
this pull request
May 15, 2026
Callgrind on rez's 188-case benchmark (post-#71/#72) showed `SmallVec::extend` + `Drop` at ~4 % of cycles, almost entirely from `VersionRange::clone`. Every `Requirement::clone()` (in `extracted_request.clone()`, the per-pair `package_request.clone()` in `reduce_by`, the `req.clone()` and `package_request.clone()` in `Reduction`, etc.) deep-copies the inner `Ranges`'s `SmallVec` of `(Bound, Bound)` segments. After the rest of the perf stack (#66/#67/#68/#70/#71/#72), this is the largest non-amortised allocation cost left. Switch the inner from `Ranges<RerVersion>` to `Rc<Ranges<RerVersion>>`. `Rc<T>::clone` is a refcount bump; `Rc<T>::Hash`/`Eq` defer to the inner `T`, so the derived semantics on `VersionRange` are unchanged. Methods that build a new range (`intersection`, `union`, `complement`, `from_versions`, `span`, `split`, ...) still produce a fresh `Ranges` internally and wrap it with `Rc::new` — the win is on the read / clone path, not the construction path. `as_ranges()` still returns `&Ranges` (via `Rc::deref`). `into_ranges` now uses `Rc::unwrap_or_clone` — falls back to a clone if the `Rc` is shared, but is the consume-the-`VersionRange` API and rare in practice. ## Benchmark (188 cases, release, same machine, two runs) | Stage | Total | Mean | vs rez | |------------------------------------|--------:|-------:|-------:| | Baseline (post-#71/#72), median | ~12.7 s | 68 ms | ~30× | | + this change, run 1 | 11.2 s | 60 ms | 34.1× | | + this change, run 2 | 11.3 s | 60 ms | 33.7× | **~11 % on top of #72**, **~74 % cumulative from main** (43.0 s → 11.2 s, 8.8× rez → 34.1× rez). Differential test (`cargo test … --ignored`): 17.73 s, **188/188 still match rez 1:1**. Predicted 3–5 %. The slightly bigger gain reflects that `VersionRange::clone` cascades into a lot more than just the `SmallVec::extend` it was attributed to in the callgrind exclusive view — it also drove allocator-side work and the matching `Drop`s. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
3 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Stacks on #71. After #71 made `PackageVariantSlice::extractable` O(1), every `extract` no-op early-return is nearly free — but the function-call round-trip itself is still paid on every probe. Each iteration of the inner extract loop in `ResolvePhase::solve` walks every scope and calls `extract`, even on scopes that have already exhausted their common dependencies.
A scope's extractability only changes when its entries change. Track a per-scope `non_extractable` flag and skip the `extract()` call once it has returned `None`, until intersect / reduce / widen replaces the scope:
Benchmark (188 cases, release, same machine)
Modest, ~5 % typical. The gain shrunk vs predicted because #71 already collapsed the `extract` body to a length compare — there is much less function-call work left to skip than there would have been on the pre-#71 codebase. Honest data point: cascading wins show diminishing returns when each removes the slack the next was relying on.
Cumulative from main: 43.0 s → 12.4 s, 8.8× → 30.8× rez.
Correctness
The correctness invariant: `non_extractable[i] == true` ⇒ the latest `scope[i]` was just observed to have no common dependencies left. Any operation that would expand or rebuild the variant slice resets the flag.
Base
This PR targets `extractable-counter` (#71). When #71 merges, GitHub will retarget this to `main`.
🤖 Generated with Claude Code