Remove `[u64; 4]` from small version to move `Arc` to full version by konstin · Pull Request #10345 · astral-sh/uv

konstin · 2025-01-07T09:30:14Z

Cloning and dropping the version arc took a significant fraction of the time in the resolver, which is a large overhead especially for the small variant that has only 9 bytes payload.

When moving the Arc to only apply to the full variant, the small variant is too large because it stores a [u64; 4] to have a release accessor a &[u64] that's shared with the Vec<u64> of the full variant. We proxy this by first extracting the compressed version digits of the small variant to a proxy type that stores up to 4 u64 on the stack and can be deref'ed to the existing &[u64], minimizing churn.

Cloning and dropping the version arc took a significant fraction of the time in the resolver, which is a large overhead especially for the small variant that has only 9 bytes payload. When moving the `Arc` to only apply to the full variant, the small variant is too large because it stores a `[u64; 4]` to have a release accessor a `&[u64]` that's shared with the `Vec<u64>` of the full variant. We proxy this by first extracting the compressed version digits of the small variant to a proxy type that stores up to 4 u64 on the stack and can be deref'ed to the existing `&[u64]`, minimizing churn. ``` $ Benchmark 1: target/profiling/uv pip compile scripts/requirements/airflow.in Time (mean ± σ): 361.3 ms ± 2.7 ms [User: 503.4 ms, System: 174.9 ms] Range (min … max): 356.3 ms … 365.2 ms 10 runs Benchmark 2: ./uv-3 pip compile scripts/requirements/airflow.in Time (mean ± σ): 402.9 ms ± 8.5 ms [User: 571.2 ms, System: 196.6 ms] Range (min … max): 393.9 ms … 418.0 ms 10 runs Summary target/profiling/uv pip compile scripts/requirements/airflow.in ran 1.12 ± 0.03 times faster than ./uv-3 pip compile scripts/requirements/airflow.in ```

codspeed-hq · 2025-01-07T09:38:53Z

CodSpeed Performance Report

Merging #10345 will improve performances by 29.06%

_{Comparing konsti/opt-small-version-arc (c09adcf) with main (14a9008)}

Summary

⚡ 3 improvements
✅ 11 untouched benchmarks

Benchmarks breakdown

	Benchmark	`main`	`konsti/opt-small-version-arc`	Change
⚡	`resolve_warm_airflow`	1,039.2 ms	841.5 ms	+23.48%
⚡	`resolve_warm_jupyter`	95 ms	73.6 ms	+29.06%
⚡	`resolve_warm_jupyter_universal`	232.7 ms	209.6 ms	+11.02%

BurntSushi

This is awesome! I love it.

BurntSushi · 2025-01-07T13:17:59Z

crates/uv-pep440/src/version.rs

 }

+/// Lifetime and indexing workaround to allow accessing the release as `&[u64]` even though the
+/// digits may be stored in a compressed representation.


Since this is part of the public API, I'd suggest keeping the first sentence to one line, and then adding more (if necessary) in a paragraph below it. Otherwise, this will show up as an overlong summary in the type listing.

BurntSushi · 2025-01-07T13:25:22Z

crates/uv-pep440/src/version.rs

+    Small3([u64; 3]),
+    Small4([u64; 4]),
+    Full(&'a [u64]),
+}


I was wondering if there was a more compact representation than this, so I tried:

enum ReleaseInner<'a> { Small { numbers: [u64; 4], len: u8 }, Full(&'a [u64]), }

But indeed, it's the same size (which makes sense).

BurntSushi · 2025-01-07T13:27:54Z

crates/uv-pep440/src/version.rs

@@ -930,6 +939,8 @@ struct VersionSmall {
    /// places somewhat exposes internal details, since the "full" version
    /// representation would not do that.
    len: u8,


Oh man, if we could get rid of len here, this type would decrease in size by half I believe. IDK if that would make an actual difference perf-wise... And I think the only way to do it would be to take bits away from repr, which in turn means moving more cases to the Full representation. Which might be a net negative overall.

We could get rid of len by counting the zeroes ourselves. The problem is that Version doesn't become 64 byte since the Arc has only a single niche (NonNull) in rustc's view, so both variants don't fit into a 64-bit word.

We could try a tagged pointer library, or we could lean into it and make the small representation [u8; 11] or [u8; 15] to use the space we're already allocating while leaving a bit (byte) for rustc for the enum discriminant.

This problem MRE:

use std::sync::Arc; enum A { Small(()), Large(Arc<Vec<u8>>), } enum B { Small(bool), Large(Arc<Vec<u8>>), } enum C { Small([u8; 7]), Large(Arc<Vec<u8>>), } fn main() { println!("{}", size_of::<A>()); // -> 8 println!("{}", size_of::<B>()); // -> 16, but can it be 8 please? println!("{}", size_of::<C>()); // -> 16, but can it be 8 please? }

Another option would be stealing two bits somewhere else (fourth release number, or capping the first release number to 14 bits).

We could get rid of len by counting the zeroes ourselves.

Hmmm, but then I think you might lose the current property that trailing zeros are preserved. See: #10362.

Otherwise yeah, I see. I agree that going further probably means a tagged pointer of some kind.

A less thought out idea is a global interner like we use in uv-pep508 for markers. But whether that hits the right trade-offs is not at all clear to me. It could let you get a Version down to 32-bits pretty easily I think, but then pretty much every interaction with it would require consulting the global interner table. And that will probably destroy any benefits you might otherwise get.

BurntSushi · 2025-01-07T13:28:27Z

crates/uv-pep440/src/version.rs

+                        (small.repr >> 0o60) & 0xFFFF,
+                        (small.repr >> 0o50) & 0xFF,
+                        (small.repr >> 0o40) & 0xFF,
+                        (small.repr >> 0o30) & 0xFF,


Octal!?!?! -_- >_< o_0 Haha

Basically, this explicitly checks that parsing a `1.2.0` into a `Version` will roundtrip back to a `1.2.0`, and that parsing a `1.2` will roundtrip back to a `1.2`. I think this case is included in the other tests in this module, but this test makes the behavior more clearly intentional I think. Ref #10345

This happened as a result of #10345 and #10362 being merged independently. The latter used the old `Version::release` API, but the former changed the `Version::release` API. This PR tweaks the new test to use the new API (i.e., force a deref on the proxy type).

This MR contains the following updates: | Package | Update | Change | |---|---|---| | [astral-sh/uv](https://github.com/astral-sh/uv) | patch | `0.5.15` -> `0.5.22` | MR created with the help of [el-capitano/tools/renovate-bot](https://gitlab.com/el-capitano/tools/renovate-bot). **Proposed changes to behavior should be submitted there as MRs.** --- ### Release Notes <details> <summary>astral-sh/uv (astral-sh/uv)</summary> ### [`v0.5.22`](https://github.com/astral-sh/uv/blob/HEAD/CHANGELOG.md#0522) [Compare Source](astral-sh/uv@0.5.21...0.5.22) ##### Enhancements - Include version and contact information in GitHub User Agent ([#10785](astral-sh/uv#10785)) ##### Performance - Add fast-path for recursive extras in dynamic validation ([#10823](astral-sh/uv#10823)) - Fetch `pyproject.toml` from GitHub API ([#10765](astral-sh/uv#10765)) - Remove allocation in Git SHA truncation ([#10801](astral-sh/uv#10801)) - Skip GitHub fast path when full commit is already known ([#10800](astral-sh/uv#10800)) ##### Bug fixes - Add fallback to build backend when `Requires-Dist` mismatches ([#10797](astral-sh/uv#10797)) - Avoid deserialization error for paths above the root ([#10789](astral-sh/uv#10789)) - Avoid respecting preferences from other indexes ([#10782](astral-sh/uv#10782)) - Disable the distutils setuptools shim during interpreter query ([#10819](astral-sh/uv#10819)) - Omit variant when detecting compatible Python installs ([#10722](astral-sh/uv#10722)) - Remove TOCTOU errors in Git clone ([#10758](astral-sh/uv#10758)) - Validate metadata under GitHub fast path ([#10796](astral-sh/uv#10796)) - Include conflict markers in fork markers ([#10818](astral-sh/uv#10818)) ##### Error messages - Add tag incompatibility hints to sync failures ([#10739](astral-sh/uv#10739)) - Improve log when distutils is missing ([#10713](astral-sh/uv#10713)) - Show non-critical Python discovery errors if no other interpreter is found ([#10716](astral-sh/uv#10716)) - Use colors for lock errors ([#10736](astral-sh/uv#10736)) ##### Documentation - Add testing instructions to the AWS Lambda guide ([#10805](astral-sh/uv#10805)) ### [`v0.5.21`](https://github.com/astral-sh/uv/blob/HEAD/CHANGELOG.md#0521) [Compare Source](astral-sh/uv@0.5.20...0.5.21) ##### Enhancements - Avoid building dynamic versions when validating lockfile ([#10703](astral-sh/uv#10703)) ##### Configuration - Add `UV_VENV_SEED` environment variable ([#10715](astral-sh/uv#10715)) ##### Performance - Store unsupported tags in wheel filename ([#10665](astral-sh/uv#10665)) ##### Bug fixes - Avoid attempting to patch macOS dylib for non-macOS installs ([#10721](astral-sh/uv#10721)) - Avoid narrowing `requires-python` marker with disjunctions ([#10704](astral-sh/uv#10704)) - Respect environment variable credentials for indexes outside root ([#10688](astral-sh/uv#10688)) - Respect preferences for explicit index dependencies from `requirements.txt` ([#10690](astral-sh/uv#10690)) - Sort preferences by environment, then index ([#10700](astral-sh/uv#10700)) - Ignore permission errors when looking for user-level configuration file ([#10697](astral-sh/uv#10697)) ##### Documentation - Add `SyntaxWarning` compatibility note to bytecode compilation docs ([#10701](astral-sh/uv#10701)) - Add `MACOSX_DEPLOYMENT_TARGET` to the `--python-platform` documentation ([#10698](astral-sh/uv#10698)) ### [`v0.5.20`](https://github.com/astral-sh/uv/blob/HEAD/CHANGELOG.md#0520) [Compare Source](astral-sh/uv@0.5.19...0.5.20) ##### Bug fixes - Avoid failing when deserializing unknown tags ([#10655](astral-sh/uv#10655)) ### [`v0.5.19`](https://github.com/astral-sh/uv/blob/HEAD/CHANGELOG.md#0519) [Compare Source](astral-sh/uv@0.5.18...0.5.19) ##### Enhancements - Filter wheels from lockfile based on architecture ([#10584](astral-sh/uv#10584)) - Omit dynamic versions from the lockfile ([#10622](astral-sh/uv#10622)) - Add support for `pip freeze --path` ([#10488](astral-sh/uv#10488)) - Reduce verbosity of inline-metadata message when using `uv run <script.py>` ([#10588](astral-sh/uv#10588)) - Add opt-in Git LFS support ([#10335](astral-sh/uv#10335)) - Recommend `--native-tls` on SSL errors ([#10605](astral-sh/uv#10605)) - Show expected and available ABI tags in resolver errors ([#10527](astral-sh/uv#10527)) - Show target Python version in error messages ([#10582](astral-sh/uv#10582)) - Add `--output-format=json` support to `uv python list` ([#10596](astral-sh/uv#10596)) ##### Python The managed Python distributions have been updated, including: - Python 3.14 support on Windows - Python 3.14.0a4 support - 64-bit RISC-V Linux support - Bundled `libedit` updated from [`2021091`](https://github.com/astral-sh/uv/commit/20210910)-3.1 -> [`2024080`](https://github.com/astral-sh/uv/commit/20240808)-3.1 - Bundled `tcl/tk` updated from 8.6.12 -> 8.6.14 (for all Python versions on Unix, only for Python 3.14 on Windows) See the [`python-build-standalone` release notes](https://github.com/astral-sh/python-build-standalone/releases/tag/20250115) for more details. ##### Performance - Avoid allocating when stripping source distribution extension ([#10625](astral-sh/uv#10625)) - Reduce `WheelFilename` to 48 bytes ([#10583](astral-sh/uv#10583)) - Reduce distribution size to 200 bytes ([#10601](astral-sh/uv#10601)) - Remove `import re` from entrypoint wrapper scripts ([#10627](astral-sh/uv#10627)) - Shrink size of platform tag enum ([#10546](astral-sh/uv#10546)) - Use `ArcStr` in verbatim URL ([#10600](astral-sh/uv#10600)) - Use `memchr` for wheel parsing ([#10620](astral-sh/uv#10620)) ##### Bug fixes - Avoid reading symlinks during `uv python install` on Windows ([#10639](astral-sh/uv#10639)) - Correct Pyston tag format ([#10580](astral-sh/uv#10580)) - Provide `pyproject.toml` path for parse errors in `uv venv` ([#10553](astral-sh/uv#10553)) - Don't treat `setuptools` and `wheel` as seed packages in uv sync on Python 3.12 ([#10572](astral-sh/uv#10572)) - Fix git-tag cache-key reader in case of slashes ([#10467](astral-sh/uv#10467)) ([#10500](astral-sh/uv#10500)) - Include build tag in rendered wheel filenames ([#10599](astral-sh/uv#10599)) - Patch embedded install path for Python dylib on macOS during `python install` ([#10629](astral-sh/uv#10629)) - Read cached registry distributions when `--config-settings` are present ([#10578](astral-sh/uv#10578)) - Show resolver hints for packages with markers ([#10607](astral-sh/uv#10607)) ##### Documentation - Add meta titles to documents in guides, excluding integration documents ([#10539](astral-sh/uv#10539)) - Remove `build-system` from example workspace rot ([#10636](astral-sh/uv#10636)) ##### Preview features - Make build backend type annotations more generic ([#10549](astral-sh/uv#10549)) ### [`v0.5.18`](https://github.com/astral-sh/uv/blob/HEAD/CHANGELOG.md#0518) [Compare Source](astral-sh/uv@0.5.17...0.5.18) ##### Bug fixes - Avoid forking for identical markers ([#10490](astral-sh/uv#10490)) - Avoid panic in `uv remove` when only comments exist ([#10484](astral-sh/uv#10484)) - Revert "improve shell compatibility of venv activate scripts ([#10397](astral-sh/uv#10397))" ([#10497](astral-sh/uv#10497)) ### [`v0.5.17`](https://github.com/astral-sh/uv/blob/HEAD/CHANGELOG.md#0517) [Compare Source](astral-sh/uv@0.5.16...0.5.17) This release includes support for generating lockfiles from scripts based on inline metadata, as defined in PEP 723. By default, scripts remain unlocked, and must be locked explicitly with `uv lock --script /path/to/script.py`, which will generate a lockfile adjacent to the script (e.g., `script.py.lock`). Once generated, the lockfile will be respected (and updated, if necessary) across `uv run --script`, `uv add --script`, and `uv remove --script` invocations. This release also includes support for `uv export --script` and `uv tree --script`. Both commands support PEP 723 scripts with and without accompanying lockfiles. ##### Enhancements - Add support for locking PEP 723 scripts ([#10135](astral-sh/uv#10135)) - Respect PEP 723 script lockfiles in `uv run` ([#10136](astral-sh/uv#10136)) - Update PEP 723 lockfile in `uv add --script` ([#10145](astral-sh/uv#10145)) - Update PEP 723 lockfile in `uv remove --script` ([#10162](astral-sh/uv#10162)) - Add `--script` support to `uv export` for PEP 723 scripts ([#10160](astral-sh/uv#10160)) - Add `--script` support to `uv tree` for PEP 723 scripts ([#10159](astral-sh/uv#10159)) - Add `ls` alias to `uv {tool, python, pip} list` ([#10240](astral-sh/uv#10240)) - Allow reading `--with-requirements` from stdin in `uv add` and `uv run` ([#10447](astral-sh/uv#10447)) - Warn-and-ignore for unsupported `requirements.txt` options ([#10420](astral-sh/uv#10420)) ##### Preview features - Add remaining Python type annotations to build backend ([#10434](astral-sh/uv#10434)) ##### Performance - Avoid allocating for names in the PEP 508 parser ([#10476](astral-sh/uv#10476)) - Fetch concurrently for non-first-match index strategies ([#10432](astral-sh/uv#10432)) - Remove unnecessary `.to_string()` call ([#10419](astral-sh/uv#10419)) - Respect sentinels in package prioritization ([#10443](astral-sh/uv#10443)) - Use `ArcStr` for marker values ([#10453](astral-sh/uv#10453)) - Use `ArcStr` for package, extra, and group names ([#10475](astral-sh/uv#10475)) - Use `matches!` rather than `contains` in `requirements.txt` parsing ([#10423](astral-sh/uv#10423)) - Use faster disjointness check for markers ([#10439](astral-sh/uv#10439)) - Pre-compute PEP 508 markers from universal markers ([#10472](astral-sh/uv#10472)) ##### Bug fixes - Fix `UV_FIND_LINKS` delimiter to split on commas ([#10477](astral-sh/uv#10477)) - Improve `uv tool list` output when tool environment is broken ([#10409](astral-sh/uv#10409)) - Only track markers for compatible versions ([#10457](astral-sh/uv#10457)) - Respect `requires-python` when installing tools ([#10401](astral-sh/uv#10401)) - Visit proxy packages eagerly ([#10441](astral-sh/uv#10441)) - Improve shell compatibility of `venv` activate scripts ([#10397](astral-sh/uv#10397)) - Read publish username from URL ([#10469](astral-sh/uv#10469)) ##### Documentation - Add Lambda layer instructions to AWS Lambda guide ([#10411](astral-sh/uv#10411)) - Add `uv lock --script` to the docs ([#10414](astral-sh/uv#10414)) - Use Windows-specific instructions in Jupyter guide ([#10446](astral-sh/uv#10446)) ### [`v0.5.16`](https://github.com/astral-sh/uv/blob/HEAD/CHANGELOG.md#0516) [Compare Source](astral-sh/uv@0.5.15...0.5.16) ##### Enhancements - Accept full requirements in `uv remove` ([#10338](astral-sh/uv#10338)) ##### Performance - Avoid over-counting versions in batch prefetcher ([#10350](astral-sh/uv#10350)) - Deactivate tracing for version-choosing ([#10351](astral-sh/uv#10351)) - Force a niche into `VersionSmall` ([#10385](astral-sh/uv#10385)) - Optimize `requirements_for_extra` ([#10348](astral-sh/uv#10348)) - Re-enable `zlib-ng` on x86 platforms ([#10365](astral-sh/uv#10365)) - Re-enable zlib-ng on all platforms (except s390x, PowerPC, and FreeBSD) ([#10370](astral-sh/uv#10370)) - Remove `[u64; 4]` from small version to move `Arc` to full version ([#10345](astral-sh/uv#10345)) - Shrink `Dist` from 352 to 288 bytes ([#10389](astral-sh/uv#10389)) - Speed up file pins by removing nested hash map ([#10346](astral-sh/uv#10346)) - Buffer file reads in `serde_json::from_reader` ([#10341](astral-sh/uv#10341)) ##### Bug fixes - Avoid enforcing project-level required version for `uv self` ([#10374](astral-sh/uv#10374)) - Fix Ruff linting warnings from generated template files for extension modules ([#10371](astral-sh/uv#10371)) ##### Documentation - Add AWS Lambda integration guide ([#10278](astral-sh/uv#10278)) </details> --- ### Configuration 📅 **Schedule**: Branch creation - At any time (no schedule defined), Automerge - At any time (no schedule defined). 🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied. ♻ **Rebasing**: Whenever MR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this MR and you won't be reminded about this update again. --- - [ ] If you want to rebase/retry this MR, check this box --- This MR has been generated by [Renovate Bot](https://github.com/renovatebot/renovate).

konstin requested a review from BurntSushi January 7, 2025 09:30

konstin force-pushed the konsti/opt-small-version-arc branch from 4053db3 to 64a8980 Compare January 7, 2025 09:30

konstin added the performance Potential performance improvement label Jan 7, 2025

Windows clippy: Fix too large future

182975a

BurntSushi approved these changes Jan 7, 2025

View reviewed changes

BurntSushi mentioned this pull request Jan 7, 2025

uv-pep440: adds an explicit test for trailing zeros #10362

Merged

Improve docs

c09adcf

konstin enabled auto-merge (squash) January 7, 2025 14:15

konstin merged commit 373e34f into main Jan 7, 2025
64 checks passed

konstin deleted the konsti/opt-small-version-arc branch January 7, 2025 14:25

BurntSushi mentioned this pull request Jan 7, 2025

uv-pep440: fix bad merge #10368

Merged

BrewTestBot mentioned this pull request Jan 8, 2025

uv 0.5.16 Homebrew/homebrew-core#203608

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Comments

Remove `[u64; 4]` from small version to move `Arc` to full version#10345

Remove `[u64; 4]` from small version to move `Arc` to full version#10345
konstin merged 3 commits intomainfrom
konsti/opt-small-version-arc

konstin commented Jan 7, 2025

Uh oh!

codspeed-hq bot commented Jan 7, 2025 •

edited

Loading

Uh oh!

BurntSushi left a comment

Uh oh!

BurntSushi Jan 7, 2025

Uh oh!

BurntSushi Jan 7, 2025

Uh oh!

BurntSushi Jan 7, 2025

Uh oh!

konstin Jan 7, 2025

Uh oh!

konstin Jan 7, 2025

Uh oh!

BurntSushi Jan 7, 2025

Uh oh!

BurntSushi Jan 7, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Comments

Conversation

konstin commented Jan 7, 2025

Uh oh!

codspeed-hq bot commented Jan 7, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Merging #10345 will improve performances by 29.06%

Summary

Benchmarks breakdown

Uh oh!

BurntSushi left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

codspeed-hq bot commented Jan 7, 2025 •

edited

Loading