perf(table): collapse branches in MergeIterator.fix

[shaunpatterson]

Summary

Reworks MergeIterator.fix to:

1. Hoist the "other" node pointer once. The previous code called mi.bigger() (which does if mi.small == &mi.left { return &mi.right } else { return &mi.left }) up to three times per invocation. Compute it once at the top.

2. Inline the swap. swapSmall() was a 2-arm conditional that flipped mi.small between &mi.left and &mi.right. With the hoisted big pointer in hand, the swap is just mi.small = big — a straight store, no branch.

3. Collapse the cmp<0 / cmp>0 cases. The previous switch had two arms whose direction-dependence is symmetric:

   • forward + cmp > 0 → swap to big
   • reverse + cmp < 0 → swap to big
   • otherwise → no-op (small is already correct)

   This is exactly (cmp < 0) == mi.reverse, so we collapse the two arms plus the no-op //nolint:staticcheck arm into a single conditional store.

bigger() and swapSmall() were the only callers of each other and of the inlined logic; with both gone, the helpers become dead code and are removed in the same commit.

Why this is safe

The cmp == 0 arm is unchanged: same-key handling still advances mi.right.next() and then swaps to left if right was small. The two non-equal arms now reduce to: if (cmp < 0) == mi.reverse { mi.small = big }, which is identical to the original switch by case analysis.

Measurement

Composite (74-benchmark stable subset): -0.62% ns/op, median of 3 runs.

Test plan

• go test -short -race ./table/ — all merge iterator tests pass, including:
  • TestMergeMore (forward/reverse × with/without duplicates)
  • TestMergeIteratorDuplicate (forward/reverse)
  • TestMergeDuplicates (forward/reverse)
  • TestMergeIteratorSeek* (forward/reverse/invalid)
  • TestMergeIteratorNested
• go vet ./...
• CI

MergeIterator.fix reaches 100% statement coverage under the existing test suite (verified with -coverprofile).

Note: an earlier draft of an unrelated Next() optimization claimed the other child could not equal curKey after a cmp==0 advance — that claim is false (a child can hold multiple copies of the same user-key with different versions) and was caught by TestMergeDuplicates. This PR does not make that change; it only restructures the branch logic in fix(), preserving the duplicate-handling path verbatim.

🤖 Generated with Claude Code