Thread

Re: Row pattern recognition

Henson Choi <assam258@gmail.com> — 2026-05-01T17:11:00Z

Hi Tatsuo,


Attached are 40 incremental patches on top of v46, with all
review-driven follow-ups folded in.  The 0001-0031 portion is
unchanged from the previous send (filenames preserved).  The new
patches 0032-0040 are the delta + follow-up patches we agreed to
deliver on top of 0031, plus two cosmetic patches that surfaced
during the sweep.

As you suggested, I have not refreshed 0001-0031; the new work
is layered on top so that the existing review state on those
patches stays valid.

Numbering and subjects of 0001-0031 are unchanged.  The new
patches are:


  - 0032 Fix A1 frame optimization bypass test in
    rpr_integration

    Per your A1 review.  Switches the baseline to row_number()
    with an explicit ROWS frame so that the bypass guard is
    actually exercised.  Block comment reworded in A2/A3 style.


  - 0033 Fix imperative voice in check_rpr_nav_expr header
    comment

    Per your typo note: "Checks for" -> "Check for".


  - 0034 Fix passive voice in RPR error messages

    Per your style note.  Sweep of RPR-touched code paths
    (parse_rpr.c, parse_func.c, windowfuncs.c) for the
    "is not permitted when ... is used" / "can only be used in"
    patterns, rewritten to "cannot use ... with" /
    "cannot use ... outside".


  - 0035 Use Max()/Min() macros for RPR extremum accumulation

    Per your Max() macro suggestion.  Sweep of all hand-rolled
    "if (x > y) y = x;" sites in createplan.c, nodeWindowAgg.c,
    and execRPR.c.  No double-evaluation hazards on the call
    sites.


  - 0036 Reword RPR navigation function synopses

    Per your sgml note.  "Returns the column value at the row"
    -> "Returns value evaluated at the row" (the first argument
    can be an arbitrary expression).


  - 0037 Replace non-ASCII characters in RPR code and tests

    Per your non-ASCII note.  Sweep of RPR-touched files;
    mostly comments and rpr*.out.


  - 0038 Move RPR design notes from execRPR.c to README.rpr

    Per the README.rpr split we discussed.  Extracts the
    design block at the head of execRPR.c into
    src/backend/executor/README.rpr (plain text, matching the
    executor/README convention).  No code changes.


  - 0039 Apply pgindent canonicalization to RPR comment blocks

    Cosmetic.  pgindent sweep across RPR-touched files; labeled
    blocks wrapped in /*---------- ... *---------- to preserve
    hanging indents.  No code changes.


  - 0040 Reword RPR navigation function descriptions in
    pg_proc.dat

    Cosmetic.  Aligns the eight RPR navigation descr fields
    with the surrounding "fetch the ... row value" pattern used
    by lag / lead / first_value / last_value / nth_value.


On the items deferred to post-v47:


  - DEFINE volatile prohibition: prepared as a self-contained
    follow-up so it can be reviewed independently of v47.

  - 0009 nfaVisitedElems representation: per your decision,
    kept as-is.  Three points behind that:

      (1) Size concern.  visited is 1 bit per element, a flat
          NFA element is 16 bytes -- a fixed 128:1 ratio
          (~0.78%).  The bitmap is dwarfed by the NFA it
          tracks (worst case ~4 KiB next to ~512 KiB), so it
          is not a memory hot spot.

      (2) Bitmapset misfit.  Its invariants ("no trailing zero
          word", "empty set = NULL") map the per-advance reset
          onto pfree + palloc-on-next-add, which defeats the
          in-place / fixed-capacity reuse pattern this code
          relies on.

      (3) Future variant.  A high-water mark approach (track
          the min/max touched word index, memset only that
          span on reset) preserves the in-place pattern and
          shrinks reset cost from O(numElements/64) to
          O(span_words).  Held as a future optimization
          candidate.

  - 0007 / B7 (Recursive CTE XXX): tracked as a discussion item
    awaiting community input on the standard interpretation.


Regards,
Henson