[AdaWorldAPI]

Follow-up to #1633. The first two commits were pushed after #1633 had been merged. The third adds a lab probe.

Problem

relalg lowered Or(a1..ak), And and Not through FoldDialect::call. That gives every atom and every operator a new scratch slot, so a chain of k atoms took 2k - 1 slots. A fold may use MAX_SLOTS = 8, so a disjunction of 5 atoms was refused with ScratchBufferTooSmall. Because of this, crepe-plan-probe built its semijoin mask through relalg only up to 4 centers and fell back to a scalar test above that.

What changes

• r2il: FoldDialect::call_owned. It runs the unchanged call, then releases the slots of the operands it consumed, except the result. fresh() hands released slots out before minting new ones, and the free list is cleared at every fold.

  • Contract: the caller guarantees that no copy of a consumed Val survives. The dialect cannot check this, because Val is Copy.
  • Plain call is unchanged and is still refused past MAX_SLOTS; a test pins that.

• relalg: Or, And, Not and the ne/ge/le negations inside atoms are lowered through call_owned. relalg consumes every subexpression's value exactly once, so the contract holds by construction. A chain of any width needs 3 slots; depth still costs slots.

• crepe-plan-probe: SEMIJOIN_ATOMS_MAX and the scalar fallback are removed. The relalg mask is exact for any number of centers.

• CI: the probe workflow now also runs relalg's own fmt, tests and clippy. relalg is its own workspace with a sibling ndarray patch, and no other job ran its tests. The workflow also triggers on r2il changes now.

• crepe-plan-probe, lab only (examples/walk_mask.rs): crepe's graph_walking benchmark program, run twice:

  • once through crepe!'s generated code;
  • once as one n × n bit plane per path length, with ndarray::simd mask ops.

  The tuple counts must match crepe's. The mask form uses two facts the macro does not know: node ids are dense in 0..n, and path lengths are bounded (0..=50). No production path changes.

Tests

• relalg:
  • a disjunction of 1, 4, 5, 64 and 1,000 atoms keeps exactly the right rows, as written and rewritten, in ≤ 3 slots;
  • a conjunction of negations up to 1,000 atoms gives the exact count;
  • the paired test depth_still_costs_slots checks that depth still costs slots.
• The relalg test that pinned the as-written refusal now pins the rewrite's gain: 1 pass rewritten, 10 as written.
• r2il: call_owned_reuses_consumed_slots_and_call_does_not runs a chain of 32 ranges. It takes exactly 3 slots with call_owned and is refused with call.
• Totals: relalg 14, r2il dialect 9, crepe-plan-probe 38. The crates that depend on r2il also pass: r2il-mask-abi-probe, lance-graph-nal and holo-h14d-probe.
• Disable run: with call_owned releasing nothing, 4 relalg tests and the r2il test go red.
• walk_mask: the counts are identical to crepe at n = 128, 256 and 512. Disable run: dropping length-50 paths fails the count assertion.
  • Removing the semi-naive subtraction does not change the counts. It is a performance step only: each step adds 5 to the length, so the loop reaches the same fixpoint.

Measurement (walk_mask, release, median of 9)

n	crepe bench (Fnv)	bit planes	Walk	NoWalk
128	7.6 ms	0.30 ms	21,508	10,310
256	41.2 ms	1.1 ms	85,601	41,090
512	290.6 ms	1.3–2.3 ms	341,517	163,702

The crepe column is crepe's own criterion bench. The example's in-process crepe run uses the default hasher and is slower (11.5 / 67–77 / 580 ms). The gain comes from the dense domain and the bounded length, not from SIMD. Neither macro can use it without a domain annotation.

---

Generated by Claude Code

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Warning

Billing warning: we have not been able to collect payment for this subscription for more than 72 hours. Please update the payment method or pay any pending invoices in Billing to avoid service interruption.

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-20T21:36:23.712057Z	f842d00	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[coderabbitai]

Add Carrot credits or activate Agent usage billing to use Autofix