Speed up remote Cargo.toml deserialization

[zozo123]

What does this resolve?

Cargo retained an owned spanned TOML document for every manifest, including cached remote registry and git packages. Those remote packages do not need a retained document after deserialization.

The source policy is explicit and exhaustive:

Source kind	Parse mode
Path, local registry, directory/vendor, embedded	Deserialize and retain the spanned document
Git, remote registry, sparse registry	Deserialize directly; do not retain the document

The direct path still uses serde_ignored and emit_toml_diagnostic, preserving unused-key tracking plus diagnostic messages and spans. Only document cloning and ownership are skipped.

Review feedback

• Atomicity: one parse-only commit; the unrelated dep-info change is removed.
• Locality: an explicit SourceKind match replaces the broad !is_path() classification.
• Structure: the logic is inline in read_manifest; there is no helper module.
• Methodology: the earlier combined Zed run is not presented as evidence for this isolated head.

Validation

• cargo +1.97.0 fmt --all -- --check
• parser unit tests for diagnostic and unused-key equivalence
• cargo +1.97.0 test -p cargo --test testsuite registry::remote_registry_manifest_without_document

[rustbot]

Thanks for the pull request, and welcome! The Rust Project is excited to review your changes, and you should hear from @epage (or someone else) some time within the next two weeks.

Please see the contribution instructions for more information. Namely, in order to ensure the minimum review times lag, PR authors and assigned reviewers should ensure that the review label (S-waiting-on-review and S-waiting-on-author) stays updated, invoking these commands when appropriate:

• @rustbot author: the review is finished, PR author should check the comments and take action accordingly
• @rustbot review: the author is ready for a review, this PR will be queued again in the reviewer's queue

Why was this reviewer chosen?

The reviewer was selected based on:

• Owners of files modified in this PR: @epage, @weihanglo
• @epage, @weihanglo expanded to epage, weihanglo
• Random selection from epage, weihanglo

[weihanglo]

Side note: You seems to open a bunch of PRs addressing performance issues recently. It would be good to chime in with us on Zulip or open an issue dicussing what you're looking at and what your methodology is, so that we can be more aligned on the project development direction and expectations. We also have issues with Performance label that capture more algorithmic and architectural possibility to improve Cargo's performance.

View changes since the review

[zozo123]

Thanks for the review — addressed in the force-pushed update:

1. Atomic: dropped the empty dep-info early-return from this PR (will measure in isolation before proposing anything there).
2. Local gate: full spanned path when source_id.is_path() || is_embedded; light toml::from_str otherwise (registry and git, not only is_registry()).
3. No helper file: inlined in read_manifest.

Also noted on methodology / not opening a burst of mixed perf PRs — happy to sync on Zulip as you suggested.

[zozo123]

@rustbot review

[zozo123]

Converting this PR back to draft.

After re-review, this is not merge-ready:

1. Source boundary is wrong as written. !is_path() is not “non-local”: it also includes local registries and directory/vendor sources, which can be user-controlled and may need full diagnostics.
2. Behavior change, not just speed. Fast path uses toml::from_str + document = None, which can change unused-key warnings, parse error presentation, and span-backed diagnostics. The existing test only checks a valid registry dep builds twice.
3. Cited A/B does not measure current head. The previous hosted run bundled empty dep-info + a different classification. Current-head performance is not isolated.

Next before leaving draft:

• explicit SourceKind → parse-mode map (no wildcard);
• stderr-equivalence tests (malformed registry/git, unknown keys, local registry, vendor);
• exact current-base vs current-head A/B for parse only.

Sorry for the premature “ready” framing.

[zozo123]

Taking back out of draft for review after rustfmt fix.

Scope remains: faster non-local manifest deserialize (toml::from_str when not path/embedded). Happy to tighten the SourceKind map or add stderr-equivalence tests as follow-ups if reviewers want them as blocking items.

[zozo123]

@rustbot review

[zozo123]

Back to draft until review concerns are actually closed

Thanks again @weihanglo @epage. Taking this out of ready so we don’t ask for merge review while these are open:

1. Atomicity — dep-info already removed; keep this PR parse-only.
2. Source boundary — need an explicit SourceKind map (not a vague “non-local” story). Path/embedded full diagnostics is right; remote registry/git are the intended fast path; local registry / directory / vendor must be decided deliberately, not via !is_path().
3. Diagnostics — need stderr-equivalence tests (malformed + unknown keys) before claiming behavior-preserving.
4. Evidence — current-head isolated A/B (not the older combined run).
5. Process — will post a short methodology note on Zulip before further Cargo perf PRs, as requested.

I will only mark ready again when (2)+(3) are in the branch. No more ready/draft thrash.

[zozo123]

@rustbot author

[rustbot]

Reminder, once the PR becomes ready for a review, use @rustbot ready.