perf(spectrum): build the backward RC query in a stack buffer

[nh13]

v0.3 perf series — PR-C (C-alloc), part 1 of the net-new alloc trims (the forward-routing commits are subsumed by #23's O(log) rewrite; see triage below).

What

mem_search_backward built a fresh Vec<u8> for the reverse-complement query on every call — and the consumer fires ~50 backward calls per read. This builds it into a 256-byte stack buffer for the common read length, falling back to the heap only for reads longer than the buffer.

Byte-identical

Same (b & 0x3) ^ 0x3 bytes, same q_key, same forward_maximal_len input — output unchanged. Guarded by the existing mem_search_backward_equals_maximal_backward_step oracle proptest (passes unchanged). Pure hot-path alloc removal.

Why it's clean

This is the exact rc_stack/rc_heap pattern already merged in mem_search_backward_truncated_span_rc (#27) and _interval_rc (#30) — mem_search_backward was the one RC-build path that still allocated. So it's also a sibling-consistency fix: the three RC-build sites now match.

PR-C triage (recorded for the series)

The carving plan flagged PR-C (forward micro-opts) as needing a triage vs merged #23. Result:

• Subsumed (drop): the 7 forward-routing commits (authored against pre-O(log) mem_search; perf: O(log) cold one-shot maximal exact match (forward + backward) #23's rewrite already does that work), 2efa208 (THP for the .sa mmap — already merged as perf(sa): hint transparent huge pages for the .sa mmap (Linux) #25), and 144632a (hoist keyed_compare_mask — main's comparison path already hoists it; shares_prefix/ref_less are 6-arg and don't take the mask).
• Net-new (ship): this (fa85f82) and 17d506b (a combined position+key entry() read — a separate follow-up).

Pre-PR /coderabbitai-review: 0 findings.

Summary by CodeRabbit

• Refactor
  • Optimized memory allocation strategy in search operations to reduce heap allocations and improve performance for common query sizes.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 873121c1-7f35-433e-95e3-524f9f483fac

📥 Commits

Reviewing files that changed from the base of the PR and between b658018 and 5d61f2f.

📒 Files selected for processing (1)

• prmi/src/index/spectrum.rs

---

📝 Walkthrough

Walkthrough

In LearnedIndex::mem_search_backward, the reverse-complement construction for ambiguous anchors switches from always allocating a heap Vec<u8> to a stack-first approach. A fixed 256-byte stack buffer handles common read lengths; only longer sequences trigger heap allocation. The mask-and-complement transformation (b & 0x3) ^ 0x3 maintains the 2-bit representation.

Changes

Reverse-complement memory allocation

Layer / File(s)	Summary
Stack buffer with conditional heap fallback for reverse-complement prmi/src/index/spectrum.rs	Reverse-complement construction uses a 256-byte stack buffer for common read lengths; mask-and-complement transformation (b & 0x3) ^ 0x3 preserves 2-bit encoding while avoiding underflow; the resulting slice is passed to forward_maximal_len to compute match length.

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~20 minutes

Possibly related PRs

• fg-labs/prmi#23: Both PRs modify reverse-complement query construction in mem_search_backward's ambiguous-anchor path, with different approaches to optimization.
• fg-labs/prmi#17: Both PRs modify the backward spectrum search implementation in mem_search_backward with changes to reverse-complement handling.

Poem

🐰 A buffer on the stack so fleet,
No heap allocation, oh what a treat!
Bits flip and dance in masks of grace,
Memory saved in every race. 🚀

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately describes the main change: optimizing RC query construction in mem_search_backward by using a stack buffer instead of heap allocation.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

---

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 and usage tips.}