v48-0005-Row-pattern-recognition-patch-executor-and-comma.patch
application/octet-stream
Filename: v48-0005-Row-pattern-recognition-patch-executor-and-comma.patch
Type: application/octet-stream
Part: 4
Message:
Re: Row pattern recognition
Patch
Same data as JSON:
GET /api/v1/attachments/:id/patch
the parsed metadata as JSON — format, series position, per-file stats; never the diff bytes.
API reference →
Format: format-patch
Series: patch v48-0005
Subject: Row pattern recognition patch (executor and commands).
| File | + | − |
|---|---|---|
| src/backend/commands/explain.c | 484 | 0 |
| src/backend/executor/execExpr.c | 92 | 0 |
| src/backend/executor/execExprInterp.c | 267 | 0 |
| src/backend/executor/execRPR.c | 1979 | 0 |
| src/backend/executor/Makefile | 1 | 0 |
| src/backend/executor/meson.build | 1 | 0 |
| src/backend/executor/nodeWindowAgg.c | 1274 | 216 |
| src/backend/jit/llvm/llvmjit_expr.c | 76 | 2 |
| src/backend/jit/llvm/llvmjit_types.c | 2 | 0 |
| src/backend/utils/adt/windowfuncs.c | 118 | 1 |
| src/include/catalog/pg_proc.dat | 24 | 0 |
| src/include/executor/execExpr.h | 20 | 0 |
| src/include/executor/execRPR.h | 40 | 0 |
| src/include/executor/nodeWindowAgg.h | 3 | 0 |
| src/include/nodes/execnodes.h | 132 | 0 |
From 126cde00883d0f734b94fce0f8ffeeb45905ada9 Mon Sep 17 00:00:00 2001
From: Tatsuo Ishii <ishii@postgresql.org>
Date: Sat, 13 Jun 2026 16:46:30 +0900
Subject: [PATCH v48 5/9] Row pattern recognition patch (executor and
commands).
---
src/backend/commands/explain.c | 484 ++++++
src/backend/executor/Makefile | 1 +
src/backend/executor/execExpr.c | 92 ++
src/backend/executor/execExprInterp.c | 267 ++++
src/backend/executor/execRPR.c | 1979 +++++++++++++++++++++++++
src/backend/executor/meson.build | 1 +
src/backend/executor/nodeWindowAgg.c | 1490 ++++++++++++++++---
src/backend/jit/llvm/llvmjit_expr.c | 78 +-
src/backend/jit/llvm/llvmjit_types.c | 2 +
src/backend/utils/adt/windowfuncs.c | 119 +-
src/include/catalog/pg_proc.dat | 24 +
src/include/executor/execExpr.h | 20 +
src/include/executor/execRPR.h | 40 +
src/include/executor/nodeWindowAgg.h | 3 +
src/include/nodes/execnodes.h | 132 ++
15 files changed, 4513 insertions(+), 219 deletions(-)
create mode 100644 src/backend/executor/execRPR.c
create mode 100644 src/include/executor/execRPR.h
diff --git a/src/backend/commands/explain.c b/src/backend/commands/explain.c
index 112c17b0d64..70fd7f386a0 100644
--- a/src/backend/commands/explain.c
+++ b/src/backend/commands/explain.c
@@ -30,6 +30,7 @@
#include "nodes/extensible.h"
#include "nodes/makefuncs.h"
#include "nodes/nodeFuncs.h"
+#include "optimizer/rpr.h"
#include "parser/analyze.h"
#include "parser/parsetree.h"
#include "rewrite/rewriteHandler.h"
@@ -119,6 +120,15 @@ static void show_window_def(WindowAggState *planstate,
static void show_window_keys(StringInfo buf, PlanState *planstate,
int nkeys, AttrNumber *keycols,
List *ancestors, ExplainState *es);
+static void append_rpr_quantifier(StringInfo buf, RPRPatternElement *elem);
+static char *deparse_rpr_pattern(RPRPattern *pattern);
+static int deparse_rpr_seq(RPRPattern *pattern, int start, int limit,
+ StringInfo buf);
+static int deparse_rpr_node(RPRPattern *pattern, int i, int limit,
+ StringInfo buf);
+static int rpr_match_end(RPRPattern *pattern, int beginIdx);
+static int rpr_alt_scope_end(RPRPattern *pattern, int i);
+static int rpr_next_branch(RPRPattern *pattern, int b, int altEnd);
static void show_storage_info(char *maxStorageType, int64 maxSpaceUsed,
ExplainState *es);
static void show_tablesample(TableSampleClause *tsc, PlanState *planstate,
@@ -129,6 +139,7 @@ static void show_incremental_sort_info(IncrementalSortState *incrsortstate,
static void show_hash_info(HashState *hashstate, ExplainState *es);
static void show_material_info(MaterialState *mstate, ExplainState *es);
static void show_windowagg_info(WindowAggState *winstate, ExplainState *es);
+static void show_rpr_nfa_stats(WindowAggState *winstate, ExplainState *es);
static void show_ctescan_info(CteScanState *ctescanstate, ExplainState *es);
static void show_table_func_scan_info(TableFuncScanState *tscanstate,
ExplainState *es);
@@ -2898,6 +2909,247 @@ show_sortorder_options(StringInfo buf, Node *sortexpr,
}
}
+/*
+ * Append quantifier suffix for a pattern element.
+ */
+static void
+append_rpr_quantifier(StringInfo buf, RPRPatternElement *elem)
+{
+ /* Append quantifier if not {1,1} */
+ if (elem->min == 0 && elem->max == RPR_QUANTITY_INF)
+ appendStringInfoChar(buf, '*');
+ else if (elem->min == 1 && elem->max == RPR_QUANTITY_INF)
+ appendStringInfoChar(buf, '+');
+ else if (elem->min == 0 && elem->max == 1)
+ appendStringInfoChar(buf, '?');
+ else if (elem->max == RPR_QUANTITY_INF)
+ appendStringInfo(buf, "{%d,}", elem->min);
+ else if (elem->min == elem->max && elem->min != 1)
+ appendStringInfo(buf, "{%d}", elem->min);
+ else if (elem->min != 1 || elem->max != 1)
+ appendStringInfo(buf, "{%d,%d}", elem->min, elem->max);
+
+ if (RPRElemIsReluctant(elem))
+ {
+ if (elem->min == 1 && elem->max == 1)
+ appendStringInfoString(buf, "{1}"); /* make reluctant ?
+ * unambiguous */
+ appendStringInfoChar(buf, '?');
+ }
+
+ /* Append absorption markers: " for judgment point, ' for branch only */
+ if (RPRElemIsAbsorbable(elem))
+ {
+ Assert(elem->max == RPR_QUANTITY_INF);
+ appendStringInfoChar(buf, '"');
+ }
+ else if (RPRElemIsAbsorbableBranch(elem))
+ appendStringInfoChar(buf, '\'');
+}
+
+/*
+ * Deparse a compiled RPRPattern (bytecode) back to a pattern string.
+ *
+ * The flat RPRPatternElement[] array is walked by recursive descent. Each
+ * construct is deparsed within an inherited [start, limit) window: the parent
+ * passes the boundary down, so each construct's extent is fixed by its caller.
+ * Three signals drive the walk:
+ *
+ * - scope ends (where an ALT or GROUP body finishes) come from depth, via
+ * rpr_alt_scope_end() and rpr_match_end().
+ * - branch boundaries (where a "|" goes) come from a branch-start jump,
+ * confirmed by the relative test elem[j-1].next != j, via
+ * rpr_next_branch().
+ * - parentheses come from structure (a BEGIN group, an ALT) plus a one-step
+ * lookahead for a group that wraps a lone ALT.
+ *
+ * depth and the relative next test are stable across the next/jump values the
+ * compiler assigns to branch tails and nested alternations, which is what makes
+ * them suitable to anchor scope and branch boundaries.
+ *
+ * EXPLAIN parenthesizes every ALT on its own, so a top-level "A | B" deparses
+ * as "(a | b)". This self-consistent EXPLAIN form is the correctness oracle
+ * here; pg_get_viewdef differs, as its parens come only from an enclosing
+ * GROUP. Absorption markers (' ") are orthogonal and handled by
+ * append_rpr_quantifier().
+ *
+ * Two compiler invariants hold throughout: {1,1} groups are unwrapped before
+ * bytecode generation (so every BEGIN/END group carries a non-trivial
+ * quantifier, and a lone ALT inside a group always spans to the group's END),
+ * and a group's quantifier is read from its END element (the BEGIN copy is
+ * ignored).
+ */
+static char *
+deparse_rpr_pattern(RPRPattern *pattern)
+{
+ StringInfoData buf;
+
+ Assert(pattern != NULL && pattern->numElements >= 2);
+
+ initStringInfo(&buf);
+ deparse_rpr_seq(pattern, 0, pattern->numElements, &buf);
+ return buf.data;
+}
+
+/*
+ * Deparse a run of sibling elements in [start, limit), separated by spaces.
+ *
+ * Stops at limit or at the FIN terminator (top-level call passes limit =
+ * numElements, where the last element is FIN). Returns the index reached.
+ */
+static int
+deparse_rpr_seq(RPRPattern *pattern, int start, int limit, StringInfo buf)
+{
+ int i = start;
+ bool first = true;
+
+ while (i < limit && !RPRElemIsFin(&pattern->elements[i]))
+ {
+ if (!first)
+ appendStringInfoChar(buf, ' ');
+ first = false;
+ i = deparse_rpr_node(pattern, i, limit, buf);
+ }
+ return i;
+}
+
+/*
+ * Deparse the single construct starting at index i, bounded by the inherited
+ * limit. Returns the index just past the construct.
+ *
+ * A VAR is its name plus quantifier. A BEGIN opens a group spanning to its
+ * matching END (rpr_match_end); when the group's sole child is an ALT that
+ * runs to the END, the ALT supplies the parentheses and the group only adds
+ * the quantifier, otherwise the group body is wrapped in its own "( )". An
+ * ALT runs to its depth-determined scope end (capped by the inherited limit)
+ * and emits "( b1 | b2 | ... )", each branch deparsed within the boundary
+ * handed down by rpr_next_branch.
+ */
+static int
+deparse_rpr_node(RPRPattern *pattern, int i, int limit, StringInfo buf)
+{
+ RPRPatternElement *elem = &pattern->elements[i];
+
+ if (RPRElemIsVar(elem))
+ {
+ Assert(elem->varId < pattern->numVars);
+ appendStringInfoString(buf,
+ quote_identifier(pattern->varNames[elem->varId]));
+ append_rpr_quantifier(buf, elem);
+ return i + 1;
+ }
+
+ if (RPRElemIsBegin(elem))
+ {
+ int end = rpr_match_end(pattern, i);
+ bool loneAlt;
+
+ loneAlt = (i + 1 < end &&
+ RPRElemIsAlt(&pattern->elements[i + 1]) &&
+ rpr_alt_scope_end(pattern, i + 1) == end);
+
+ if (loneAlt)
+ {
+ /* The ALT child already parenthesizes the whole group body. */
+ (void) deparse_rpr_node(pattern, i + 1, end, buf);
+ }
+ else
+ {
+ appendStringInfoChar(buf, '(');
+ (void) deparse_rpr_seq(pattern, i + 1, end, buf);
+ appendStringInfoChar(buf, ')');
+ }
+ append_rpr_quantifier(buf, &pattern->elements[end]);
+ return end + 1;
+ }
+
+ Assert(RPRElemIsAlt(elem));
+ {
+ int altEnd = rpr_alt_scope_end(pattern, i);
+ int b;
+ bool first = true;
+
+ if (altEnd > limit)
+ altEnd = limit;
+
+ appendStringInfoChar(buf, '(');
+ b = i + 1;
+ while (b < altEnd)
+ {
+ int nb = rpr_next_branch(pattern, b, altEnd);
+
+ if (!first)
+ appendStringInfoString(buf, " | ");
+ first = false;
+ (void) deparse_rpr_seq(pattern, b, nb, buf);
+ b = nb;
+ }
+ appendStringInfoChar(buf, ')');
+ return altEnd;
+ }
+}
+
+/*
+ * Find the END that closes the group opened by the BEGIN at beginIdx: the
+ * first END at the same depth scanning forward.
+ */
+static int
+rpr_match_end(RPRPattern *pattern, int beginIdx)
+{
+ RPRDepth d = pattern->elements[beginIdx].depth;
+ int j;
+
+ for (j = beginIdx + 1; j < pattern->numElements; j++)
+ {
+ RPRPatternElement *e = &pattern->elements[j];
+
+ if (RPRElemIsEnd(e) && e->depth == d)
+ return j;
+ }
+ pg_unreachable(); /* a BEGIN always has a matching END */
+}
+
+/*
+ * Scope end of the construct at index i: the first following element whose
+ * depth is no greater than i's own. For an ALT marker this is the index just
+ * past its last branch, since depth stays constant across branch boundaries.
+ * FIN sits at depth 0, so a top-level ALT stops there.
+ */
+static int
+rpr_alt_scope_end(RPRPattern *pattern, int i)
+{
+ RPRDepth d = pattern->elements[i].depth;
+ int k;
+
+ for (k = i + 1; k < pattern->numElements; k++)
+ {
+ if (pattern->elements[k].depth <= d)
+ return k;
+ }
+ return pattern->numElements;
+}
+
+/*
+ * Boundary of the alternation branch starting at b (i.e. the start of the next
+ * branch, or altEnd if b is the last branch).
+ *
+ * The branch-start element's jump points at the next branch when this is not
+ * the last branch. jump is overloaded (a group BEGIN also uses it for its
+ * skip path), so confirm a real branch boundary with the relative test
+ * elem[j-1].next != j: at a true boundary the preceding branch's tail has its
+ * next redirected past the alternation, so it does not point at j.
+ */
+static int
+rpr_next_branch(RPRPattern *pattern, int b, int altEnd)
+{
+ int j = pattern->elements[b].jump;
+
+ if (j != RPR_ELEMIDX_INVALID && j < altEnd &&
+ pattern->elements[j - 1].next != j)
+ return j;
+ return altEnd;
+}
+
/*
* Show the window definition for a WindowAgg node.
*/
@@ -2956,6 +3208,83 @@ show_window_def(WindowAggState *planstate, List *ancestors, ExplainState *es)
appendStringInfoChar(&wbuf, ')');
ExplainPropertyText("Window", wbuf.data, es);
pfree(wbuf.data);
+
+ /* Show Row Pattern Recognition pattern if present */
+ if (wagg->rpPattern != NULL)
+ {
+ char *patternStr = deparse_rpr_pattern(wagg->rpPattern);
+
+ if (patternStr != NULL)
+ {
+ ExplainPropertyText("Pattern", patternStr, es);
+ pfree(patternStr);
+ }
+
+ /*
+ * Show navigation offsets for tuplestore trim. For EXPLAIN ANALYZE,
+ * use the executor-resolved values (which may differ from the plan
+ * when NEEDS_EVAL was resolved to FIXED or RETAIN_ALL at init).
+ */
+ {
+ RPRNavOffsetKind maxKind = wagg->navMaxOffsetKind;
+ int64 maxOffset = wagg->navMaxOffset;
+ RPRNavOffsetKind firstKind = wagg->navFirstOffsetKind;
+ int64 firstOffset = wagg->navFirstOffset;
+
+ if (es->analyze)
+ {
+ maxKind = planstate->navMaxOffsetKind;
+ maxOffset = planstate->navMaxOffset;
+ firstKind = planstate->navFirstOffsetKind;
+ firstOffset = planstate->navFirstOffset;
+ }
+
+ switch (maxKind)
+ {
+ case RPR_NAV_OFFSET_NEEDS_EVAL:
+ ExplainPropertyText("Nav Mark Lookback", "runtime", es);
+ break;
+ case RPR_NAV_OFFSET_RETAIN_ALL:
+ ExplainPropertyText("Nav Mark Lookback", "retain all", es);
+ break;
+ case RPR_NAV_OFFSET_FIXED:
+ ExplainPropertyInteger("Nav Mark Lookback", NULL,
+ maxOffset, es);
+ break;
+ default:
+ elog(ERROR, "unrecognized RPR nav offset kind: %d",
+ maxKind);
+ break;
+ }
+
+ if (wagg->hasFirstNav)
+ {
+ switch (firstKind)
+ {
+ case RPR_NAV_OFFSET_NEEDS_EVAL:
+ ExplainPropertyText("Nav Mark Lookahead", "runtime",
+ es);
+ break;
+ case RPR_NAV_OFFSET_RETAIN_ALL:
+ ExplainPropertyText("Nav Mark Lookahead", "retain all",
+ es);
+ break;
+ case RPR_NAV_OFFSET_FIXED:
+ if (firstOffset == PG_INT64_MAX)
+ ExplainPropertyText("Nav Mark Lookahead", "infinite",
+ es);
+ else
+ ExplainPropertyInteger("Nav Mark Lookahead", NULL,
+ firstOffset, es);
+ break;
+ default:
+ elog(ERROR, "unrecognized RPR nav offset kind: %d",
+ firstKind);
+ break;
+ }
+ }
+ }
+ }
}
/*
@@ -3508,6 +3837,7 @@ show_windowagg_info(WindowAggState *winstate, ExplainState *es)
{
char *maxStorageType;
int64 maxSpaceUsed;
+ WindowAgg *wagg = (WindowAgg *) winstate->ss.ps.plan;
Tuplestorestate *tupstore = winstate->buffer;
@@ -3520,6 +3850,160 @@ show_windowagg_info(WindowAggState *winstate, ExplainState *es)
tuplestore_get_stats(tupstore, &maxStorageType, &maxSpaceUsed);
show_storage_info(maxStorageType, maxSpaceUsed, es);
+
+ /* Show NFA statistics for Row Pattern Recognition */
+ if (wagg->rpPattern != NULL)
+ show_rpr_nfa_stats(winstate, es);
+}
+
+/*
+ * Show NFA statistics for Row Pattern Recognition on WindowAgg node.
+ */
+static void
+show_rpr_nfa_stats(WindowAggState *winstate, ExplainState *es)
+{
+ if (es->format != EXPLAIN_FORMAT_TEXT)
+ {
+ /* State and context counters */
+ ExplainPropertyInteger("NFA States Peak", NULL, winstate->nfaStatesMax, es);
+ ExplainPropertyInteger("NFA States Total", NULL, winstate->nfaStatesTotalCreated, es);
+ ExplainPropertyInteger("NFA States Merged", NULL, winstate->nfaStatesMerged, es);
+ ExplainPropertyInteger("NFA Contexts Peak", NULL, winstate->nfaContextsMax, es);
+ ExplainPropertyInteger("NFA Contexts Total", NULL, winstate->nfaContextsTotalCreated, es);
+ ExplainPropertyInteger("NFA Contexts Absorbed", NULL, winstate->nfaContextsAbsorbed, es);
+ ExplainPropertyInteger("NFA Contexts Skipped", NULL, winstate->nfaContextsSkipped, es);
+ ExplainPropertyInteger("NFA Contexts Pruned", NULL, winstate->nfaContextsPruned, es);
+
+ /* Match/mismatch counts and length statistics */
+ ExplainPropertyInteger("NFA Matched", NULL, winstate->nfaMatchesSucceeded, es);
+ ExplainPropertyInteger("NFA Mismatched", NULL, winstate->nfaMatchesFailed, es);
+ if (winstate->nfaMatchesSucceeded > 0)
+ {
+ ExplainPropertyInteger("NFA Match Length Min", NULL, winstate->nfaMatchLen.min, es);
+ ExplainPropertyInteger("NFA Match Length Max", NULL, winstate->nfaMatchLen.max, es);
+ ExplainPropertyFloat("NFA Match Length Avg", NULL,
+ (double) winstate->nfaMatchLen.total / winstate->nfaMatchesSucceeded, 1,
+ es);
+ }
+ if (winstate->nfaMatchesFailed > 0)
+ {
+ ExplainPropertyInteger("NFA Mismatch Length Min", NULL, winstate->nfaFailLen.min, es);
+ ExplainPropertyInteger("NFA Mismatch Length Max", NULL, winstate->nfaFailLen.max, es);
+ ExplainPropertyFloat("NFA Mismatch Length Avg", NULL,
+ (double) winstate->nfaFailLen.total / winstate->nfaMatchesFailed, 1,
+ es);
+ }
+
+ /* Absorbed/skipped context length statistics */
+ if (winstate->nfaContextsAbsorbed > 0)
+ {
+ ExplainPropertyInteger("NFA Absorbed Length Min", NULL, winstate->nfaAbsorbedLen.min, es);
+ ExplainPropertyInteger("NFA Absorbed Length Max", NULL, winstate->nfaAbsorbedLen.max, es);
+ ExplainPropertyFloat("NFA Absorbed Length Avg", NULL,
+ (double) winstate->nfaAbsorbedLen.total / winstate->nfaContextsAbsorbed, 1,
+ es);
+ }
+ if (winstate->nfaContextsSkipped > 0)
+ {
+ ExplainPropertyInteger("NFA Skipped Length Min", NULL, winstate->nfaSkippedLen.min, es);
+ ExplainPropertyInteger("NFA Skipped Length Max", NULL, winstate->nfaSkippedLen.max, es);
+ ExplainPropertyFloat("NFA Skipped Length Avg", NULL,
+ (double) winstate->nfaSkippedLen.total / winstate->nfaContextsSkipped, 1,
+ es);
+ }
+ }
+ else
+ {
+ /* State and context counters */
+ ExplainIndentText(es);
+ appendStringInfo(es->str,
+ "NFA States: " INT64_FORMAT " peak, " INT64_FORMAT " total, " INT64_FORMAT " merged\n",
+ winstate->nfaStatesMax,
+ winstate->nfaStatesTotalCreated,
+ winstate->nfaStatesMerged);
+ ExplainIndentText(es);
+ appendStringInfo(es->str,
+ "NFA Contexts: " INT64_FORMAT " peak, " INT64_FORMAT " total, " INT64_FORMAT " pruned\n",
+ winstate->nfaContextsMax,
+ winstate->nfaContextsTotalCreated,
+ winstate->nfaContextsPruned);
+
+ /* Match/mismatch counts with length min/max/avg */
+ ExplainIndentText(es);
+ appendStringInfoString(es->str, "NFA: ");
+ if (winstate->nfaMatchesSucceeded > 0)
+ {
+ double avgLen = (double) winstate->nfaMatchLen.total / winstate->nfaMatchesSucceeded;
+
+ appendStringInfo(es->str,
+ INT64_FORMAT " matched (len " INT64_FORMAT "/" INT64_FORMAT "/%.1f)",
+ winstate->nfaMatchesSucceeded,
+ winstate->nfaMatchLen.min,
+ winstate->nfaMatchLen.max,
+ avgLen);
+ }
+ else
+ {
+ appendStringInfoString(es->str, "0 matched");
+ }
+ if (winstate->nfaMatchesFailed > 0)
+ {
+ double avgFail = (double) winstate->nfaFailLen.total / winstate->nfaMatchesFailed;
+
+ appendStringInfo(es->str,
+ ", " INT64_FORMAT " mismatched (len " INT64_FORMAT "/" INT64_FORMAT "/%.1f)",
+ winstate->nfaMatchesFailed,
+ winstate->nfaFailLen.min,
+ winstate->nfaFailLen.max,
+ avgFail);
+ }
+ else
+ {
+ appendStringInfoString(es->str, ", 0 mismatched");
+ }
+ appendStringInfoChar(es->str, '\n');
+
+ /* Absorbed/skipped context length statistics */
+ if (winstate->nfaContextsAbsorbed > 0 || winstate->nfaContextsSkipped > 0)
+ {
+ ExplainIndentText(es);
+ appendStringInfoString(es->str, "NFA: ");
+
+ if (winstate->nfaContextsAbsorbed > 0)
+ {
+ double avgAbsorbed = (double) winstate->nfaAbsorbedLen.total / winstate->nfaContextsAbsorbed;
+
+ appendStringInfo(es->str,
+ INT64_FORMAT " absorbed (len " INT64_FORMAT "/" INT64_FORMAT "/%.1f)",
+ winstate->nfaContextsAbsorbed,
+ winstate->nfaAbsorbedLen.min,
+ winstate->nfaAbsorbedLen.max,
+ avgAbsorbed);
+ }
+ else
+ {
+ appendStringInfoString(es->str, "0 absorbed");
+ }
+
+ if (winstate->nfaContextsSkipped > 0)
+ {
+ double avgSkipped = (double) winstate->nfaSkippedLen.total / winstate->nfaContextsSkipped;
+
+ appendStringInfo(es->str,
+ ", " INT64_FORMAT " skipped (len " INT64_FORMAT "/" INT64_FORMAT "/%.1f)",
+ winstate->nfaContextsSkipped,
+ winstate->nfaSkippedLen.min,
+ winstate->nfaSkippedLen.max,
+ avgSkipped);
+ }
+ else
+ {
+ appendStringInfoString(es->str, ", 0 skipped");
+ }
+
+ appendStringInfoChar(es->str, '\n');
+ }
+ }
}
/*
diff --git a/src/backend/executor/Makefile b/src/backend/executor/Makefile
index 11118d0ce02..2b257427795 100644
--- a/src/backend/executor/Makefile
+++ b/src/backend/executor/Makefile
@@ -25,6 +25,7 @@ OBJS = \
execParallel.o \
execPartition.o \
execProcnode.o \
+ execRPR.o \
execReplication.o \
execSRF.o \
execScan.o \
diff --git a/src/backend/executor/execExpr.c b/src/backend/executor/execExpr.c
index cfea7e160c2..30367825033 100644
--- a/src/backend/executor/execExpr.c
+++ b/src/backend/executor/execExpr.c
@@ -1189,6 +1189,98 @@ ExecInitExprRec(Expr *node, ExprState *state,
break;
}
+ case T_RPRNavExpr:
+ {
+ /*
+ * RPR navigation functions (PREV/NEXT/FIRST/LAST) are
+ * compiled into EEOP_RPR_NAV_SET / EEOP_RPR_NAV_RESTORE
+ * opcodes instead of a normal function call. The SET opcode
+ * swaps ecxt_outertuple to the target row, the argument
+ * expression is compiled normally (reads from the swapped
+ * slot), and the RESTORE opcode restores the original slot.
+ *
+ * Default offset when offset_arg is NULL: PREV/NEXT: 1
+ * (physical offset from currentpos) FIRST/LAST: 0 (logical
+ * offset from match boundary)
+ */
+ RPRNavExpr *nav = (RPRNavExpr *) node;
+ WindowAggState *winstate;
+
+ Assert(state->parent && IsA(state->parent, WindowAggState));
+ winstate = (WindowAggState *) state->parent;
+
+ /* Emit SET opcode: swap slot to target row */
+ scratch.opcode = EEOP_RPR_NAV_SET;
+ scratch.d.rpr_nav.winstate = winstate;
+ scratch.d.rpr_nav.kind = nav->kind;
+
+ if (nav->kind >= RPR_NAV_PREV_FIRST)
+ {
+ /*
+ * Compound navigation: allocate array of 2 for inner [0]
+ * and outer [1] offsets.
+ */
+ Datum *offset_values = palloc_array(Datum, 2);
+ bool *offset_isnulls = palloc_array(bool, 2);
+
+ /* Inner offset (default 0 for FIRST/LAST) */
+ if (nav->offset_arg != NULL)
+ ExecInitExprRec(nav->offset_arg, state,
+ &offset_values[0], &offset_isnulls[0]);
+ else
+ {
+ offset_values[0] = Int64GetDatum(0);
+ offset_isnulls[0] = false;
+ }
+
+ /* Outer offset (default 1 for PREV/NEXT) */
+ if (nav->compound_offset_arg != NULL)
+ ExecInitExprRec(nav->compound_offset_arg, state,
+ &offset_values[1], &offset_isnulls[1]);
+ else
+ {
+ offset_values[1] = Int64GetDatum(1);
+ offset_isnulls[1] = false;
+ }
+
+ scratch.d.rpr_nav.offset_value = offset_values;
+ scratch.d.rpr_nav.offset_isnull = offset_isnulls;
+ }
+ else if (nav->offset_arg != NULL)
+ {
+ /* Simple navigation with explicit offset */
+ Datum *offset_value = palloc_object(Datum);
+ bool *offset_isnull = palloc_object(bool);
+
+ ExecInitExprRec(nav->offset_arg, state,
+ offset_value, offset_isnull);
+ scratch.d.rpr_nav.offset_value = offset_value;
+ scratch.d.rpr_nav.offset_isnull = offset_isnull;
+ }
+ else
+ {
+ /* Simple navigation with default offset */
+ scratch.d.rpr_nav.offset_value = NULL;
+ scratch.d.rpr_nav.offset_isnull = NULL;
+ }
+
+ ExprEvalPushStep(state, &scratch);
+
+ /* Compile the argument expression normally */
+ ExecInitExprRec(nav->arg, state, resv, resnull);
+
+ /* Emit RESTORE opcode: restore original slot */
+ scratch.opcode = EEOP_RPR_NAV_RESTORE;
+ scratch.resvalue = resv;
+ scratch.resnull = resnull;
+ scratch.d.rpr_nav.winstate = winstate;
+ get_typlenbyval(nav->resulttype,
+ &scratch.d.rpr_nav.resulttyplen,
+ &scratch.d.rpr_nav.resulttypbyval);
+ ExprEvalPushStep(state, &scratch);
+ break;
+ }
+
case T_FuncExpr:
{
FuncExpr *func = (FuncExpr *) node;
diff --git a/src/backend/executor/execExprInterp.c b/src/backend/executor/execExprInterp.c
index 0634af964a9..805c8583fb2 100644
--- a/src/backend/executor/execExprInterp.c
+++ b/src/backend/executor/execExprInterp.c
@@ -57,11 +57,13 @@
#include "postgres.h"
#include "access/heaptoast.h"
+#include "common/int.h"
#include "access/tupconvert.h"
#include "catalog/pg_type.h"
#include "commands/sequence.h"
#include "executor/execExpr.h"
#include "executor/nodeSubplan.h"
+#include "executor/nodeWindowAgg.h"
#include "funcapi.h"
#include "miscadmin.h"
#include "nodes/miscnodes.h"
@@ -586,6 +588,8 @@ ExecInterpExpr(ExprState *state, ExprContext *econtext, bool *isnull)
&&CASE_EEOP_WINDOW_FUNC,
&&CASE_EEOP_MERGE_SUPPORT_FUNC,
&&CASE_EEOP_SUBPLAN,
+ &&CASE_EEOP_RPR_NAV_SET,
+ &&CASE_EEOP_RPR_NAV_RESTORE,
&&CASE_EEOP_AGG_STRICT_DESERIALIZE,
&&CASE_EEOP_AGG_DESERIALIZE,
&&CASE_EEOP_AGG_STRICT_INPUT_CHECK_ARGS,
@@ -2013,6 +2017,24 @@ ExecInterpExpr(ExprState *state, ExprContext *econtext, bool *isnull)
EEO_NEXT();
}
+ /* RPR navigation: swap slot to target row */
+ EEO_CASE(EEOP_RPR_NAV_SET)
+ {
+ ExecEvalRPRNavSet(state, op, econtext);
+ outerslot = econtext->ecxt_outertuple;
+
+ EEO_NEXT();
+ }
+
+ /* RPR navigation: restore slot to original row */
+ EEO_CASE(EEOP_RPR_NAV_RESTORE)
+ {
+ ExecEvalRPRNavRestore(state, op, econtext);
+ outerslot = econtext->ecxt_outertuple;
+
+ EEO_NEXT();
+ }
+
/* evaluate a strict aggregate deserialization function */
EEO_CASE(EEOP_AGG_STRICT_DESERIALIZE)
{
@@ -5988,3 +6010,248 @@ ExecAggPlainTransByRef(AggState *aggstate, AggStatePerTrans pertrans,
MemoryContextSwitchTo(oldContext);
}
+
+/*
+ * Extract compound (outer) offset from step data.
+ * For compound nav, offset_value is an array: [0]=inner, [1]=outer.
+ * Returns the outer offset; errors on NULL or negative.
+ * Default is 1 (like PREV/NEXT implicit offset).
+ */
+static int64
+rpr_nav_get_compound_offset(ExprEvalStep *op)
+{
+ int64 val;
+
+ Assert(op->d.rpr_nav.offset_value != NULL);
+
+ if (op->d.rpr_nav.offset_isnull[1])
+ ereport(ERROR,
+ errcode(ERRCODE_NULL_VALUE_NOT_ALLOWED),
+ errmsg("row pattern navigation offset must not be null"));
+
+ val = DatumGetInt64(op->d.rpr_nav.offset_value[1]);
+
+ if (val < 0)
+ ereport(ERROR,
+ errcode(ERRCODE_INVALID_PARAMETER_VALUE),
+ errmsg("row pattern navigation offset must not be negative"));
+
+ return val;
+}
+
+/*
+ * Evaluate RPR navigation (PREV/NEXT/FIRST/LAST): swap slot to target row.
+ *
+ * Saves the current outertuple into winstate for later restore, computes
+ * the target row position, fetches the corresponding slot from the
+ * tuplestore, and replaces econtext->ecxt_outertuple with it.
+ *
+ * This is called both from the interpreter inline handler and from
+ * JIT-compiled expressions via build_EvalXFunc.
+ */
+void
+ExecEvalRPRNavSet(ExprState *state, ExprEvalStep *op, ExprContext *econtext)
+{
+ WindowAggState *winstate = op->d.rpr_nav.winstate;
+ int64 offset;
+ int64 target_pos;
+ TupleTableSlot *target_slot;
+
+ /* Save current slot for later restore */
+ winstate->nav_saved_outertuple = econtext->ecxt_outertuple;
+
+ /*
+ * Determine the inner offset. NULL or negative offsets are errors per
+ * the SQL standard.
+ *
+ * Default offset when offset_arg is NULL: PREV/NEXT: 1 (standard 5.6.2)
+ * FIRST/LAST and compound: 0 for inner, 1 for outer
+ */
+ if (op->d.rpr_nav.offset_value != NULL)
+ {
+ if (*op->d.rpr_nav.offset_isnull)
+ ereport(ERROR,
+ errcode(ERRCODE_NULL_VALUE_NOT_ALLOWED),
+ errmsg("row pattern navigation offset must not be null"));
+
+ offset = DatumGetInt64(*op->d.rpr_nav.offset_value);
+
+ if (offset < 0)
+ ereport(ERROR,
+ errcode(ERRCODE_INVALID_PARAMETER_VALUE),
+ errmsg("row pattern navigation offset must not be negative"));
+ }
+ else
+ {
+ /* Default offset: 1 for simple PREV/NEXT, 0 otherwise */
+ if (op->d.rpr_nav.kind == RPR_NAV_PREV ||
+ op->d.rpr_nav.kind == RPR_NAV_NEXT)
+ offset = 1;
+ else
+ offset = 0;
+ }
+
+ /*
+ * Calculate target position based on navigation direction. On overflow,
+ * use -1 so that ExecRPRNavGetSlot treats it as out of range.
+ */
+ switch (op->d.rpr_nav.kind)
+ {
+ case RPR_NAV_PREV:
+ if (pg_sub_s64_overflow(winstate->currentpos, offset, &target_pos))
+ target_pos = -1;
+ break;
+ case RPR_NAV_NEXT:
+ if (pg_add_s64_overflow(winstate->currentpos, offset, &target_pos))
+ target_pos = -1;
+ break;
+ case RPR_NAV_FIRST:
+ /* FIRST: offset from match_start, clamped to currentpos */
+ if (pg_add_s64_overflow(winstate->nav_match_start, offset, &target_pos))
+ target_pos = -1;
+ else if (target_pos > winstate->currentpos)
+ target_pos = -1; /* beyond current match range */
+ break;
+ case RPR_NAV_LAST:
+ /* LAST: offset backward from currentpos, clamped to match_start */
+ if (pg_sub_s64_overflow(winstate->currentpos, offset, &target_pos))
+ target_pos = -1;
+ else if (target_pos < winstate->nav_match_start)
+ target_pos = -1; /* before match_start */
+ break;
+
+ case RPR_NAV_PREV_FIRST:
+ case RPR_NAV_NEXT_FIRST:
+ {
+ int64 compound_offset;
+ int64 inner_pos;
+
+ /* Inner: match_start + offset */
+ if (pg_add_s64_overflow(winstate->nav_match_start, offset, &inner_pos))
+ {
+ target_pos = -1;
+ break;
+ }
+ if (inner_pos > winstate->currentpos || inner_pos < 0)
+ {
+ target_pos = -1;
+ break;
+ }
+
+ /* Outer offset */
+ compound_offset = rpr_nav_get_compound_offset(op);
+
+ /* Apply outer: PREV subtracts, NEXT adds */
+ if (op->d.rpr_nav.kind == RPR_NAV_PREV_FIRST)
+ {
+ if (pg_sub_s64_overflow(inner_pos, compound_offset, &target_pos))
+ target_pos = -1;
+ }
+ else
+ {
+ if (pg_add_s64_overflow(inner_pos, compound_offset, &target_pos))
+ target_pos = -1;
+ }
+ }
+ break;
+
+ case RPR_NAV_PREV_LAST:
+ case RPR_NAV_NEXT_LAST:
+ {
+ int64 compound_offset;
+ int64 inner_pos;
+
+ /* Inner: currentpos - offset */
+ if (pg_sub_s64_overflow(winstate->currentpos, offset, &inner_pos))
+ {
+ target_pos = -1;
+ break;
+ }
+ if (inner_pos < winstate->nav_match_start)
+ {
+ target_pos = -1;
+ break;
+ }
+
+ /* Outer offset */
+ compound_offset = rpr_nav_get_compound_offset(op);
+
+ /* Apply outer: PREV subtracts, NEXT adds */
+ if (op->d.rpr_nav.kind == RPR_NAV_PREV_LAST)
+ {
+ if (pg_sub_s64_overflow(inner_pos, compound_offset, &target_pos))
+ target_pos = -1;
+ }
+ else
+ {
+ if (pg_add_s64_overflow(inner_pos, compound_offset, &target_pos))
+ target_pos = -1;
+ }
+ }
+ break;
+ default:
+ elog(ERROR, "unrecognized RPR navigation kind: %d",
+ op->d.rpr_nav.kind);
+ break;
+ }
+
+ /*
+ * Slot swap elision: if target_pos is the current row, skip the
+ * tuplestore fetch and slot swap entirely. This benefits LAST(expr),
+ * PREV(expr, 0), NEXT(expr, 0), and similar cases.
+ *
+ * We must still set nav_saved_outertuple (done above) so that
+ * EEOP_RPR_NAV_RESTORE is a harmless no-op.
+ */
+ if (target_pos == winstate->currentpos)
+ return;
+
+ /* Fetch target row slot (returns nav_null_slot if out of range) */
+ target_slot = ExecRPRNavGetSlot(winstate, target_pos);
+
+ /*
+ * Update econtext to point to the target slot. Also decompress the new
+ * slot's attributes since FETCHSOME already ran for the original slot.
+ * The caller (interpreter or JIT) is responsible for updating any local
+ * slot cache (e.g. outerslot) from econtext after we return.
+ */
+ slot_getallattrs(target_slot);
+ econtext->ecxt_outertuple = target_slot;
+}
+
+/*
+ * Evaluate RPR navigation: restore slot to original row.
+ *
+ * Restores econtext->ecxt_outertuple from the saved slot in winstate.
+ * When slot swap was elided (target == currentpos), this is a harmless
+ * no-op since saved and current slots are identical.
+ * The caller is responsible for updating any local slot cache.
+ *
+ * For pass-by-reference result types, the result datum points into
+ * nav_slot's tuple memory. If a subsequent navigation in the same
+ * expression re-fetches nav_slot for a different position, the old
+ * tuple is freed, leaving a dangling pointer. We prevent this by
+ * copying pass-by-ref results into per-tuple memory, which survives
+ * until the next ResetExprContext.
+ */
+void
+ExecEvalRPRNavRestore(ExprState *state, ExprEvalStep *op,
+ ExprContext *econtext)
+{
+ WindowAggState *winstate = op->d.rpr_nav.winstate;
+
+ econtext->ecxt_outertuple = winstate->nav_saved_outertuple;
+
+ /* Stabilize pass-by-ref result against nav_slot re-fetch */
+ if (!op->d.rpr_nav.resulttypbyval &&
+ !*op->resnull)
+ {
+ MemoryContext oldContext;
+
+ oldContext = MemoryContextSwitchTo(econtext->ecxt_per_tuple_memory);
+ *op->resvalue = datumCopy(*op->resvalue,
+ false,
+ op->d.rpr_nav.resulttyplen);
+ MemoryContextSwitchTo(oldContext);
+ }
+}
diff --git a/src/backend/executor/execRPR.c b/src/backend/executor/execRPR.c
new file mode 100644
index 00000000000..b326a58bbf5
--- /dev/null
+++ b/src/backend/executor/execRPR.c
@@ -0,0 +1,1979 @@
+/*-------------------------------------------------------------------------
+ *
+ * execRPR.c
+ * NFA-based Row Pattern Recognition engine for window functions.
+ *
+ * This file implements the NFA execution engine for the ROWS BETWEEN
+ * PATTERN clause (SQL Standard Feature R020: Row Pattern Recognition in
+ * Window Functions).
+ *
+ * The engine executes the compiled RPRPattern structure directly, avoiding
+ * regex compilation overhead. It is called by nodeWindowAgg.c and exposes
+ * the interface declared in executor/execRPR.h.
+ *
+ *
+ * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
+ * Portions Copyright (c) 1994, Regents of the University of California
+ *
+ * IDENTIFICATION
+ * src/backend/executor/execRPR.c
+ *
+ *-------------------------------------------------------------------------
+ */
+#include "postgres.h"
+
+#include "common/int.h"
+#include "executor/execRPR.h"
+#include "executor/executor.h"
+#include "miscadmin.h"
+#include "optimizer/rpr.h"
+#include "utils/memutils.h"
+
+/*
+ * For the design and execution model of the NFA engine implemented
+ * in this file, see src/backend/executor/README.rpr.
+ */
+
+/* Bitmap macros for NFA cycle detection (cf. bitmapset.c, tidbitmap.c) */
+#define WORDNUM(x) ((x) / BITS_PER_BITMAPWORD)
+#define BITNUM(x) ((x) % BITS_PER_BITMAPWORD)
+
+/*
+ * Set the visited bit for elemIdx and update the high-water marks
+ * (nfaVisitedMin/MaxWord) so that the next reset only has to clear
+ * the touched range instead of the full nfaVisitedElems bitmap.
+ */
+static inline void
+nfa_mark_visited(WindowAggState *winstate, int16 elemIdx)
+{
+ int16 w = WORDNUM(elemIdx);
+
+ winstate->nfaVisitedElems[w] |= ((bitmapword) 1 << BITNUM(elemIdx));
+ winstate->nfaVisitedMinWord = Min(winstate->nfaVisitedMinWord, w);
+ winstate->nfaVisitedMaxWord = Max(winstate->nfaVisitedMaxWord, w);
+}
+
+/* Forward declarations */
+static RPRNFAState *nfa_state_make(WindowAggState *winstate);
+static void nfa_state_free(WindowAggState *winstate, RPRNFAState *state);
+static void nfa_state_free_list(WindowAggState *winstate, RPRNFAState *list);
+static RPRNFAState *nfa_state_clone(WindowAggState *winstate, int16 elemIdx,
+ int32 *counts, bool sourceAbsorbable);
+static bool nfa_states_equal(WindowAggState *winstate, RPRNFAState *s1,
+ RPRNFAState *s2);
+static void nfa_add_state_unique(WindowAggState *winstate, RPRNFAContext *ctx,
+ RPRNFAState *state);
+static void nfa_add_matched_state(WindowAggState *winstate, RPRNFAContext *ctx,
+ RPRNFAState *state, int64 matchEndRow);
+
+static RPRNFAContext *nfa_context_make(WindowAggState *winstate);
+static void nfa_unlink_context(WindowAggState *winstate, RPRNFAContext *ctx);
+
+static void nfa_update_length_stats(int64 count, NFALengthStats *stats, int64 newLen);
+static void nfa_record_context_skipped(WindowAggState *winstate, int64 skippedLen);
+static void nfa_record_context_absorbed(WindowAggState *winstate, int64 absorbedLen);
+
+static void nfa_update_absorption_flags(RPRNFAContext *ctx);
+static bool nfa_states_covered(RPRPattern *pattern, RPRNFAContext *older,
+ RPRNFAContext *newer);
+static void nfa_try_absorb_context(WindowAggState *winstate, RPRNFAContext *ctx);
+static void nfa_absorb_contexts(WindowAggState *winstate);
+
+static bool nfa_eval_var_match(WindowAggState *winstate,
+ RPRPatternElement *elem, bool *varMatched);
+static void nfa_match(WindowAggState *winstate, RPRNFAContext *ctx,
+ bool *varMatched);
+static void nfa_route_to_elem(WindowAggState *winstate, RPRNFAContext *ctx,
+ RPRNFAState *state, RPRPatternElement *nextElem,
+ int64 currentPos);
+static void nfa_advance_alt(WindowAggState *winstate, RPRNFAContext *ctx,
+ RPRNFAState *state, RPRPatternElement *elem,
+ int64 currentPos);
+static void nfa_advance_begin(WindowAggState *winstate, RPRNFAContext *ctx,
+ RPRNFAState *state, RPRPatternElement *elem,
+ int64 currentPos);
+static void nfa_advance_end(WindowAggState *winstate, RPRNFAContext *ctx,
+ RPRNFAState *state, RPRPatternElement *elem,
+ int64 currentPos);
+static void nfa_advance_var(WindowAggState *winstate, RPRNFAContext *ctx,
+ RPRNFAState *state, RPRPatternElement *elem,
+ int64 currentPos);
+static void nfa_advance_state(WindowAggState *winstate, RPRNFAContext *ctx,
+ RPRNFAState *state, int64 currentPos);
+static void nfa_advance(WindowAggState *winstate, RPRNFAContext *ctx,
+ int64 currentPos);
+
+static void nfa_reevaluate_dependent_vars(WindowAggState *winstate,
+ RPRNFAContext *ctx,
+ int64 currentPos);
+
+/*
+ * NFA-based pattern matching implementation
+ *
+ * These functions implement direct NFA execution using the compiled
+ * RPRPattern structure, avoiding regex compilation overhead.
+ *
+ * Execution Flow: match -> absorb -> advance
+ * -----------------------------------------
+ * The NFA execution follows a three-phase cycle for each row:
+ *
+ * 1. MATCH (convergence): Evaluate all waiting states against current row.
+ * States on VAR elements are checked against their defining conditions.
+ * Failed matches are removed, successful ones may transition forward.
+ * This is a "convergence" phase - the number of states tends to decrease.
+ *
+ * 2. ABSORB: After matching, check if any context can absorb another.
+ * Context absorption is an optimization that merges equivalent contexts.
+ * A context can only be absorbed if ALL its states are absorbable.
+ *
+ * 3. ADVANCE (divergence): Expand states through epsilon transitions.
+ * States advance through ALT (alternation), END (group end), and
+ * optional elements until reaching VAR or FIN elements where they wait.
+ * This is a "divergence" phase - ALT creates multiple branch states.
+ *
+ * Key Design Decisions:
+ * ---------------------
+ * - VAR->END transition in match phase: When a simple VAR (max=1) matches
+ * and the next element is END, we transition immediately in the match
+ * phase rather than waiting for advance. This is necessary for correct
+ * absorption: states must be at END to be marked absorbable before the
+ * absorption check occurs.
+ *
+ * - Optional VAR skip paths: When advance lands on a VAR with min=0,
+ * we create both a waiting state AND a skip state (like ALT branches).
+ * This ensures patterns like "A B? C" work correctly - we need a state
+ * waiting for B AND a state that has already skipped to C.
+ *
+ * - END->END count increment: When transitioning from one END to another
+ * END within advance, we must increment the outer END's count. This
+ * handles nested groups like "((A|B)+)+" correctly - exiting the inner
+ * group counts as one iteration of the outer group.
+ *
+ * - Empty match handling: The initial advance uses currentPos =
+ * startPos - 1 (before any row is consumed). If FIN is reached via
+ * epsilon transitions alone, matchEndRow = startPos - 1 < matchStartRow.
+ * If matchedState is set (FIN was reached), this is an empty match
+ * (RF_EMPTY_MATCH); otherwise it is unmatched (RF_UNMATCHED).
+ * For reluctant min=0 patterns (A*?, A??), the skip path reaches
+ * FIN first and early termination prunes enter paths, yielding an
+ * immediate empty match result. For greedy patterns (A*), the enter
+ * path adds VAR states first, then the skip FIN is recorded but VAR
+ * states survive for later matching.
+ *
+ * Context Absorption Runtime:
+ * ---------------------------
+ * Absorption uses flags computed at planning time (in rpr.c) and two
+ * context-level flags maintained at runtime:
+ *
+ * State-level:
+ * state.isAbsorbable: true if state is in the absorbable region.
+ * - Set at creation: elem->flags & RPR_ELEM_ABSORBABLE_BRANCH
+ * - At transition: prevAbsorbable && (newElem->flags & ABSORBABLE_BRANCH)
+ * - Monotonic: once false, stays false forever
+ *
+ * Context-level:
+ * ctx.hasAbsorbableState: can this context absorb others?
+ * - True if at least one state has isAbsorbable=true
+ * - Monotonic: true->false only (optimization: skip recalc when false)
+ *
+ * ctx.allStatesAbsorbable: can this context be absorbed?
+ * - True if ALL states have isAbsorbable=true
+ * - Dynamic: can change false->true (when non-absorbable states die)
+ *
+ * Absorption Algorithm:
+ * For each pair (older Ctx1, newer Ctx2):
+ * 1. Pre-check: Ctx1.hasAbsorbableState && Ctx2.allStatesAbsorbable
+ * -> If false, skip (fast filter)
+ * 2. Coverage check: For each Ctx2 state with isAbsorbable=true,
+ * find Ctx1 state with same elemIdx and count >= Ctx2.count
+ * 3. If all Ctx2 absorbable states are covered, absorb Ctx2
+ *
+ * Example: Pattern A+ B
+ * Row 1: Ctx1 at A (count=1)
+ * Row 2: Ctx1 at A (count=2), Ctx2 at A (count=1)
+ * -> Both at same elemIdx (A), Ctx1.count >= Ctx2.count
+ * -> Ctx2 absorbed
+ *
+ * The asymmetric design (Ctx1 needs hasAbsorbable, Ctx2 needs allAbsorbable)
+ * allows absorption even when Ctx1 has extra non-absorbable states.
+ */
+
+/*
+ * nfa_state_make
+ *
+ * Allocate an NFA state, reusing from freeList if available.
+ * freeList is stored in WindowAggState for reuse across match attempts.
+ * Uses flexible array member for counts[].
+ */
+static RPRNFAState *
+nfa_state_make(WindowAggState *winstate)
+{
+ RPRNFAState *state;
+
+ /* Try to reuse from free list first */
+ if (winstate->nfaStateFree != NULL)
+ {
+ state = winstate->nfaStateFree;
+ winstate->nfaStateFree = state->next;
+ }
+ else
+ {
+ /* Allocate in partition context for proper lifetime */
+ state = MemoryContextAlloc(winstate->partcontext, winstate->nfaStateSize);
+ }
+
+ /* Initialize entire state to zero */
+ memset(state, 0, winstate->nfaStateSize);
+
+ /* Update statistics */
+ winstate->nfaStatesActive++;
+ winstate->nfaStatesTotalCreated++;
+ winstate->nfaStatesMax = Max(winstate->nfaStatesMax,
+ winstate->nfaStatesActive);
+
+ return state;
+}
+
+/*
+ * nfa_state_free
+ *
+ * Return a state to the free list for later reuse.
+ */
+static void
+nfa_state_free(WindowAggState *winstate, RPRNFAState *state)
+{
+ winstate->nfaStatesActive--;
+ state->next = winstate->nfaStateFree;
+ winstate->nfaStateFree = state;
+}
+
+/*
+ * nfa_state_free_list
+ *
+ * Return all states in a list to the free list.
+ */
+static void
+nfa_state_free_list(WindowAggState *winstate, RPRNFAState *list)
+{
+ RPRNFAState *next;
+
+ for (; list != NULL; list = next)
+ {
+ next = list->next;
+ nfa_state_free(winstate, list);
+ }
+}
+
+/*
+ * nfa_state_clone
+ *
+ * Clone a state from the given elemIdx and counts.
+ * isAbsorbable is computed immediately: inherited AND new element's flag.
+ * Monotonic property: once false, stays false through all transitions.
+ *
+ * Caller is responsible for linking the returned state.
+ */
+static RPRNFAState *
+nfa_state_clone(WindowAggState *winstate, int16 elemIdx,
+ int32 *counts, bool sourceAbsorbable)
+{
+ RPRPattern *pattern = winstate->rpPattern;
+ int maxDepth = pattern->maxDepth;
+ RPRNFAState *state = nfa_state_make(winstate);
+ RPRPatternElement *elem = &pattern->elements[elemIdx];
+
+ state->elemIdx = elemIdx;
+ /* Every reachable caller passes a live state's counts; maxDepth >= 1. */
+ Assert(counts != NULL && maxDepth > 0);
+ memcpy(state->counts, counts, sizeof(int32) * maxDepth);
+
+ /*
+ * Compute isAbsorbable immediately at transition time. isAbsorbable =
+ * sourceAbsorbable && (elem->flags & ABSORBABLE_BRANCH) Monotonic: once
+ * false, stays false (can't re-enter absorbable region).
+ */
+ state->isAbsorbable = sourceAbsorbable && RPRElemIsAbsorbableBranch(elem);
+
+ return state;
+}
+
+/*
+ * nfa_states_equal
+ *
+ * Check if two states are equivalent (same elemIdx and counts).
+ */
+static bool
+nfa_states_equal(WindowAggState *winstate, RPRNFAState *s1, RPRNFAState *s2)
+{
+ RPRPattern *pattern = winstate->rpPattern;
+ RPRPatternElement *elem;
+ int compareDepth;
+
+ if (s1->elemIdx != s2->elemIdx)
+ return false;
+
+ /*
+ * Compare counts up to current element's depth. Two states sharing
+ * elemIdx are equivalent iff every enclosing-or-current depth count
+ * matches.
+ *
+ * The +1 is the slot arithmetic: comparing through depth N requires
+ * counts[0..N], i.e., N+1 entries. Deeper slots (counts[d] with d >
+ * elem->depth) are excluded because they hold scratch state from inner
+ * groups. Per the count-clear policy such a slot is zeroed when its
+ * owning element exits (see nfa_advance_var and the inline fast path in
+ * nfa_match), so it must not participate in equivalence judgment.
+ */
+ elem = &pattern->elements[s1->elemIdx];
+ compareDepth = elem->depth + 1;
+
+ if (memcmp(s1->counts, s2->counts, sizeof(int32) * compareDepth) != 0)
+ return false;
+
+ return true;
+}
+
+/*
+ * nfa_add_state_unique
+ *
+ * Add the state to the end of the ctx->states linked list, but only if a
+ * duplicate state is not already present.
+ * Earlier states have better lexical order (DFS traversal order), so existing
+ * wins; the new state is freed when a duplicate is found.
+ */
+static void
+nfa_add_state_unique(WindowAggState *winstate, RPRNFAContext *ctx, RPRNFAState *state)
+{
+ RPRNFAState *s;
+ RPRNFAState *tail = NULL;
+
+ /*
+ * Mark VAR in visited before duplicate check to prevent DFS loops. This
+ * is the deferred half of the asymmetric visited-marking scheme; see
+ * nfa_advance_state for the non-VAR (END/ALT/BEGIN/FIN) half and the
+ * rationale for the asymmetry.
+ */
+ nfa_mark_visited(winstate, state->elemIdx);
+
+ /* Check for duplicate and find tail */
+ for (s = ctx->states; s != NULL; s = s->next)
+ {
+ CHECK_FOR_INTERRUPTS();
+
+ if (nfa_states_equal(winstate, s, state))
+ {
+ /*
+ * Duplicate found - existing has better lexical order, discard
+ * new
+ */
+ nfa_state_free(winstate, state);
+ winstate->nfaStatesMerged++;
+ return;
+ }
+ tail = s;
+ }
+
+ /* No duplicate, add at end */
+ state->next = NULL;
+ if (tail == NULL)
+ ctx->states = state;
+ else
+ tail->next = state;
+}
+
+/*
+ * nfa_add_matched_state
+ *
+ * Record a state that reached FIN, replacing any previous match.
+ *
+ * For SKIP PAST LAST ROW, also prune subsequent contexts whose start row
+ * falls within the match range, as they cannot produce output rows.
+ */
+static void
+nfa_add_matched_state(WindowAggState *winstate, RPRNFAContext *ctx,
+ RPRNFAState *state, int64 matchEndRow)
+{
+ if (ctx->matchedState != NULL)
+ nfa_state_free(winstate, ctx->matchedState);
+
+ ctx->matchedState = state;
+ state->next = NULL;
+ ctx->matchEndRow = matchEndRow;
+
+ /* Prune contexts that started within this match's range */
+ if (winstate->rpSkipTo == ST_PAST_LAST_ROW)
+ {
+ int64 skippedLen;
+
+ while (ctx->next != NULL &&
+ ctx->next->matchStartRow <= matchEndRow)
+ {
+ RPRNFAContext *nextCtx = ctx->next;
+
+ Assert(nextCtx->lastProcessedRow >= nextCtx->matchStartRow);
+ skippedLen = nextCtx->lastProcessedRow - nextCtx->matchStartRow + 1;
+ nfa_record_context_skipped(winstate, skippedLen);
+
+ ExecRPRFreeContext(winstate, nextCtx);
+ }
+ }
+}
+
+/*
+ * nfa_context_make
+ *
+ * Allocate an NFA context, reusing from free list if available.
+ */
+static RPRNFAContext *
+nfa_context_make(WindowAggState *winstate)
+{
+ RPRNFAContext *ctx;
+
+ if (winstate->nfaContextFree != NULL)
+ {
+ ctx = winstate->nfaContextFree;
+ winstate->nfaContextFree = ctx->next;
+ }
+ else
+ {
+ /* Allocate in partition context for proper lifetime */
+ ctx = MemoryContextAlloc(winstate->partcontext, sizeof(RPRNFAContext));
+ }
+
+ ctx->next = NULL;
+ ctx->prev = NULL;
+ ctx->states = NULL;
+ ctx->matchStartRow = -1;
+ ctx->matchEndRow = -1;
+ ctx->lastProcessedRow = -1;
+ ctx->matchedState = NULL;
+
+ /* Initialize two-flag absorption design based on pattern */
+ ctx->hasAbsorbableState = winstate->rpPattern->isAbsorbable;
+ ctx->allStatesAbsorbable = winstate->rpPattern->isAbsorbable;
+
+ /* Update statistics */
+ winstate->nfaContextsActive++;
+ winstate->nfaContextsTotalCreated++;
+ winstate->nfaContextsMax = Max(winstate->nfaContextsMax,
+ winstate->nfaContextsActive);
+
+ return ctx;
+}
+
+/*
+ * nfa_unlink_context
+ *
+ * Remove a context from the doubly-linked active context list.
+ * Updates head (nfaContext) and tail (nfaContextTail) as needed.
+ */
+static void
+nfa_unlink_context(WindowAggState *winstate, RPRNFAContext *ctx)
+{
+ if (ctx->prev != NULL)
+ ctx->prev->next = ctx->next;
+ else
+ winstate->nfaContext = ctx->next; /* was head */
+
+ if (ctx->next != NULL)
+ ctx->next->prev = ctx->prev;
+ else
+ winstate->nfaContextTail = ctx->prev; /* was tail */
+
+ ctx->next = NULL;
+ ctx->prev = NULL;
+}
+
+/*
+ * nfa_update_length_stats
+ *
+ * Helper function to update min/max/total length statistics.
+ * Called when tracking match/mismatch/absorbed/skipped lengths.
+ */
+static void
+nfa_update_length_stats(int64 count, NFALengthStats *stats, int64 newLen)
+{
+ if (count == 1)
+ {
+ stats->min = newLen;
+ stats->max = newLen;
+ }
+ else
+ {
+ stats->min = Min(stats->min, newLen);
+ stats->max = Max(stats->max, newLen);
+ }
+ stats->total += newLen;
+}
+
+/*
+ * nfa_record_context_skipped
+ *
+ * Record a skipped context in statistics.
+ */
+static void
+nfa_record_context_skipped(WindowAggState *winstate, int64 skippedLen)
+{
+ winstate->nfaContextsSkipped++;
+ nfa_update_length_stats(winstate->nfaContextsSkipped,
+ &winstate->nfaSkippedLen,
+ skippedLen);
+}
+
+/*
+ * nfa_record_context_absorbed
+ *
+ * Record an absorbed context in statistics.
+ */
+static void
+nfa_record_context_absorbed(WindowAggState *winstate, int64 absorbedLen)
+{
+ winstate->nfaContextsAbsorbed++;
+ nfa_update_length_stats(winstate->nfaContextsAbsorbed,
+ &winstate->nfaAbsorbedLen,
+ absorbedLen);
+}
+
+/*
+ * nfa_update_absorption_flags
+ *
+ * Update context's absorption flags after state changes.
+ *
+ * Two flags control absorption behavior:
+ * hasAbsorbableState: true if context has at least one absorbable state.
+ * This flag is monotonic (true -> false only). Once all absorbable states
+ * die, no new absorbable states can be created through transitions.
+ * allStatesAbsorbable: true if ALL states in context are absorbable.
+ * This flag is dynamic and can change false -> true when non-absorbable
+ * states die off.
+ *
+ * Optimization: Once hasAbsorbableState becomes false, both flags remain false
+ * permanently, so we skip recalculation.
+ */
+static void
+nfa_update_absorption_flags(RPRNFAContext *ctx)
+{
+ RPRNFAState *state;
+ bool hasAbsorbable = false;
+ bool allAbsorbable = true;
+
+ /*
+ * Optimization: Once hasAbsorbableState becomes false, it stays false. No
+ * need to recalculate - both flags remain false permanently.
+ */
+ if (!ctx->hasAbsorbableState)
+ {
+ ctx->allStatesAbsorbable = false;
+ return;
+ }
+
+ /* No states means no absorbable states */
+ if (ctx->states == NULL)
+ {
+ ctx->hasAbsorbableState = false;
+ ctx->allStatesAbsorbable = false;
+ return;
+ }
+
+ /*
+ * Iterate through all states to check absorption status. Uses
+ * state->isAbsorbable which tracks if state is in absorbable region. This
+ * is different from RPRElemIsAbsorbable(elem) which checks judgment
+ * point.
+ */
+ for (state = ctx->states; state != NULL; state = state->next)
+ {
+ CHECK_FOR_INTERRUPTS();
+
+ if (state->isAbsorbable)
+ hasAbsorbable = true;
+ else
+ allAbsorbable = false;
+ }
+
+ ctx->hasAbsorbableState = hasAbsorbable;
+ ctx->allStatesAbsorbable = allAbsorbable;
+}
+
+/*
+ * nfa_states_covered
+ *
+ * Check if all states in newer context are "covered" by older context.
+ *
+ * A newer state is covered when older context has an absorbable state at the
+ * same pattern element (elemIdx) with count >= newer's count at that depth.
+ * The covering state must be absorbable because only absorbable states can
+ * guarantee to produce superset matches.
+ *
+ * If all newer states are covered, newer context's eventual matches will be
+ * a subset of older context's matches, making newer redundant.
+ */
+static bool
+nfa_states_covered(RPRPattern *pattern, RPRNFAContext *older, RPRNFAContext *newer)
+{
+ RPRNFAState *newerState;
+
+ for (newerState = newer->states; newerState != NULL; newerState = newerState->next)
+ {
+ RPRNFAState *olderState;
+ RPRPatternElement *elem;
+ int depth;
+ bool found = false;
+
+ /* All states are absorbable (caller checks allStatesAbsorbable) */
+ elem = &pattern->elements[newerState->elemIdx];
+ depth = elem->depth;
+
+ /*
+ * Only compare at absorption judgment points (RPR_ELEM_ABSORBABLE).
+ * Judgment points are where count-dominance guarantees the newer
+ * context's future matches are a subset of the older's.
+ */
+ if (!RPRElemIsAbsorbable(elem))
+ return false;
+
+ for (olderState = older->states; olderState != NULL; olderState = olderState->next)
+ {
+ CHECK_FOR_INTERRUPTS();
+
+ /* Covering state must also be absorbable */
+ if (olderState->isAbsorbable &&
+ olderState->elemIdx == newerState->elemIdx &&
+ olderState->counts[depth] >= newerState->counts[depth])
+ {
+ found = true;
+ break;
+ }
+ }
+
+ if (!found)
+ return false;
+ }
+
+ return true;
+}
+
+/*
+ * nfa_try_absorb_context
+ *
+ * Try to absorb ctx (newer) into an older in-progress context.
+ * Returns true if ctx was absorbed and freed.
+ *
+ * Absorption requires three conditions:
+ * 1. ctx must have all states absorbable (allStatesAbsorbable).
+ * If ctx has any non-absorbable state, it may produce unique matches.
+ * 2. older must have at least one absorbable state (hasAbsorbableState).
+ * Without absorbable states, older cannot cover newer's states.
+ * 3. All ctx states must be covered by older's absorbable states.
+ * This ensures older will produce all matches that ctx would produce.
+ *
+ * Context list is ordered by creation time (oldest first via prev chain).
+ * Each row creates at most one context, so earlier contexts have smaller
+ * matchStartRow values.
+ */
+static void
+nfa_try_absorb_context(WindowAggState *winstate, RPRNFAContext *ctx)
+{
+ RPRPattern *pattern = winstate->rpPattern;
+ RPRNFAContext *older;
+
+ /* Early exit: ctx must have all states absorbable */
+ if (!ctx->allStatesAbsorbable)
+ return;
+
+ for (older = ctx->prev; older != NULL; older = older->prev)
+ {
+ CHECK_FOR_INTERRUPTS();
+
+ /*
+ * By invariant: ctx->prev chain is in creation order (oldest first),
+ * and each row creates at most one context. So all contexts in this
+ * chain have matchStartRow < ctx->matchStartRow.
+ */
+
+ /* Older must also be in-progress */
+ if (older->states == NULL)
+ continue;
+
+ /* Older must have at least one absorbable state */
+ if (!older->hasAbsorbableState)
+ continue;
+
+ /* Check if all newer states are covered by older */
+ if (nfa_states_covered(pattern, older, ctx))
+ {
+ int64 absorbedLen = ctx->lastProcessedRow - ctx->matchStartRow + 1;
+
+ ExecRPRFreeContext(winstate, ctx);
+ nfa_record_context_absorbed(winstate, absorbedLen);
+ return;
+ }
+ }
+}
+
+/*
+ * nfa_absorb_contexts
+ *
+ * Absorb redundant contexts to reduce memory usage and computation.
+ *
+ * For patterns like A+, newer contexts starting later will produce subset
+ * matches of older contexts with higher counts. By absorbing these redundant
+ * contexts early, we avoid duplicate work.
+ *
+ * Iterates from tail (newest) toward head (oldest) via prev chain.
+ * Only in-progress contexts (states != NULL) are candidates for absorption;
+ * completed contexts represent valid match results.
+ */
+static void
+nfa_absorb_contexts(WindowAggState *winstate)
+{
+ RPRNFAContext *ctx;
+ RPRNFAContext *nextCtx;
+
+ for (ctx = winstate->nfaContextTail; ctx != NULL; ctx = nextCtx)
+ {
+ nextCtx = ctx->prev;
+
+ /*
+ * Only absorb in-progress contexts; completed contexts are valid
+ * results
+ */
+ if (ctx->states != NULL)
+ nfa_try_absorb_context(winstate, ctx);
+ }
+}
+
+/*
+ * nfa_eval_var_match
+ *
+ * Evaluate if a VAR element matches the current row.
+ *
+ * varMatched is a pre-evaluated boolean array indexed by varId, computed
+ * once per row by evaluating all DEFINE expressions. NULL means no DEFINE
+ * clauses exist (only possible during early development/testing).
+ *
+ * Per ISO/IEC 19075-5 Feature R020, pattern variables not listed in DEFINE
+ * are implicitly TRUE -- they match every row. This is checked via
+ * varId >= list_length.
+ */
+static bool
+nfa_eval_var_match(WindowAggState *winstate, RPRPatternElement *elem,
+ bool *varMatched)
+{
+ /* This function should only be called for VAR elements */
+ Assert(RPRElemIsVar(elem));
+
+ if (varMatched == NULL)
+ return false;
+ if (elem->varId >= list_length(winstate->defineVariableList))
+ return true;
+ return varMatched[elem->varId];
+}
+
+/*
+ * nfa_match
+ *
+ * Match phase (convergence): evaluate VAR elements against current row.
+ * Only updates counts and removes dead states. Minimal transitions.
+ *
+ * For VAR elements:
+ * - matched: count++, keep state (unless count > max)
+ * - not matched: remove state (exit alternatives already exist from
+ * previous advance when count >= min was satisfied)
+ *
+ * For VARs that reached max count followed by END:
+ * - Advance through the END-element chain to the absorption judgment point
+ * - Only deterministic exits (count >= max, max != INF) are handled
+ * - Chains through END elements while count >= max (must-exit path)
+ *
+ * Non-VAR elements (ALT, END, FIN) are kept as-is for advance phase.
+ */
+static void
+nfa_match(WindowAggState *winstate, RPRNFAContext *ctx, bool *varMatched)
+{
+ RPRPattern *pattern = winstate->rpPattern;
+ RPRPatternElement *elements = pattern->elements;
+ RPRNFAState **prevPtr = &ctx->states;
+ RPRNFAState *state;
+ RPRNFAState *nextState;
+
+ /*
+ * Evaluate VAR elements against current row. For VARs that reach max
+ * count with END next, advance through the chain of END elements inline
+ * so absorb phase can compare states at judgment points.
+ */
+ for (state = ctx->states; state != NULL; state = nextState)
+ {
+ RPRPatternElement *elem = &elements[state->elemIdx];
+
+ CHECK_FOR_INTERRUPTS();
+
+ nextState = state->next;
+
+ if (RPRElemIsVar(elem))
+ {
+ bool matched;
+ int depth = elem->depth;
+ int32 count = state->counts[depth];
+
+ matched = nfa_eval_var_match(winstate, elem, varMatched);
+
+ if (matched)
+ {
+ /* Increment count */
+ if (count < RPR_COUNT_MAX)
+ count++;
+
+ /* Max constraint should not be exceeded */
+ Assert(elem->max == RPR_QUANTITY_INF || count <= elem->max);
+
+ state->counts[depth] = count;
+
+ /*
+ * For VAR at max count with END next, advance through END
+ * chain to reach the absorption judgment point. Only
+ * deterministic exits (count >= max, max finite) are handled;
+ * unbounded VARs stay for advance phase.
+ *
+ * In nested patterns like ((A B){2}){3}, a VAR reaching its
+ * max triggers an exit cascade: inner END increments inner
+ * group count, which may itself reach max, requiring an exit
+ * to the next outer END. The loop below walks this chain.
+ *
+ * ABSORBABLE_BRANCH marks elements inside the absorbable
+ * region; ABSORBABLE marks the outermost judgment point where
+ * count-dominance is evaluated. We chain through BRANCH
+ * elements until reaching the ABSORBABLE point or an element
+ * that can still loop (count < max).
+ */
+ if (RPRElemIsAbsorbableBranch(elem) &&
+ !RPRElemIsAbsorbable(elem) &&
+ count >= elem->max &&
+ RPRElemIsEnd(&elements[elem->next]))
+ {
+ RPRPatternElement *endElem = &elements[elem->next];
+ int endDepth = endElem->depth;
+ int32 endCount = state->counts[endDepth];
+
+ /* Increment group count */
+ if (endCount < RPR_COUNT_MAX)
+ endCount++;
+ Assert(endElem->max == RPR_QUANTITY_INF ||
+ endCount <= endElem->max);
+
+ state->elemIdx = elem->next;
+ state->counts[endDepth] = endCount;
+
+ /*
+ * Leaf VAR exited (reached max): clear its own count so
+ * the next occupant enters with zero, as nfa_advance_var
+ * does on exit (this inline path replaces that exit).
+ * depth > endDepth, so this leaves the group count just
+ * written intact.
+ */
+ Assert(endDepth < depth);
+ state->counts[depth] = 0;
+
+ /*
+ * Chain through END elements within the absorbable region
+ * (ABSORBABLE_BRANCH) until reaching the judgment point
+ * (ABSORBABLE). Continue only on must-exit path (count
+ * >= max) with END next.
+ */
+ while (RPRElemIsAbsorbableBranch(endElem) &&
+ !RPRElemIsAbsorbable(endElem) &&
+ endCount >= endElem->max &&
+ RPRElemIsEnd(&elements[endElem->next]))
+ {
+ RPRPatternElement *outerEnd = &elements[endElem->next];
+ int outerDepth = outerEnd->depth;
+ int32 outerCount = state->counts[outerDepth];
+
+ /*
+ * Exit this intermediate group: clear its own count
+ * (count-clear policy). It sits below the absorbable
+ * judgment point, so it is excluded from the
+ * dominance comparison; the judgment point where the
+ * chain stops keeps its count.
+ */
+ state->counts[endDepth] = 0;
+
+ /* Increment outer group count */
+ if (outerCount < RPR_COUNT_MAX)
+ outerCount++;
+ Assert(outerEnd->max == RPR_QUANTITY_INF ||
+ outerCount <= outerEnd->max);
+
+ state->elemIdx = endElem->next;
+ state->counts[outerDepth] = outerCount;
+
+ /* Advance to next END in chain */
+ endElem = outerEnd;
+ endDepth = outerDepth;
+ endCount = outerCount;
+ }
+ }
+ /* else: stay at VAR for advance phase */
+ }
+ else
+ {
+ /*
+ * Not matched - remove state. Exit alternatives were already
+ * created by advance phase when count >= min was satisfied.
+ */
+ *prevPtr = nextState;
+ nfa_state_free(winstate, state);
+ continue;
+ }
+ }
+ /* Non-VAR elements: keep as-is for advance phase */
+
+ prevPtr = &state->next;
+ }
+}
+
+/*
+ * nfa_route_to_elem
+ *
+ * Route state to next element. If VAR, add to ctx->states and process
+ * skip path if optional. Otherwise, continue epsilon expansion via recursion.
+ */
+static void
+nfa_route_to_elem(WindowAggState *winstate, RPRNFAContext *ctx,
+ RPRNFAState *state, RPRPatternElement *nextElem,
+ int64 currentPos)
+{
+ if (RPRElemIsVar(nextElem))
+ {
+ RPRNFAState *skipState = NULL;
+
+ /*
+ * Entry-side check of the count-clear policy: a VAR is always routed
+ * to with a clean slot. Each element zeroes its own count on exit,
+ * so a nonzero count here would be a leak from an earlier element
+ * (see nfa_advance_var / nfa_advance_end exit handling and the inline
+ * fast path in nfa_match).
+ */
+ Assert(state->counts[nextElem->depth] == 0);
+
+ /* Create skip state before add_unique, which may free state */
+ if (RPRElemCanSkip(nextElem))
+ skipState = nfa_state_clone(winstate, nextElem->next,
+ state->counts, state->isAbsorbable);
+
+ if (skipState != NULL && RPRElemIsReluctant(nextElem))
+ {
+ RPRNFAState *savedMatch = ctx->matchedState;
+
+ /*
+ * Reluctant optional VAR: prefer skipping. Explore the skip path
+ * first so it outranks the enter (match) path; if it reaches FIN
+ * the shortest match is found and the enter state is dropped.
+ * This mirrors the reluctant branch of nfa_advance_begin used by
+ * the leading-position and optional-group paths.
+ */
+ nfa_advance_state(winstate, ctx, skipState, currentPos);
+
+ if (ctx->matchedState != savedMatch)
+ {
+ nfa_state_free(winstate, state);
+ return;
+ }
+
+ nfa_add_state_unique(winstate, ctx, state);
+ }
+ else
+ {
+ /* Greedy (or non-skippable): enter first, then skip */
+ nfa_add_state_unique(winstate, ctx, state);
+
+ if (skipState != NULL)
+ nfa_advance_state(winstate, ctx, skipState, currentPos);
+ }
+ }
+ else
+ {
+ nfa_advance_state(winstate, ctx, state, currentPos);
+ }
+}
+
+/*
+ * nfa_advance_alt
+ *
+ * Handle ALT element: expand all branches in lexical order via DFS.
+ */
+static void
+nfa_advance_alt(WindowAggState *winstate, RPRNFAContext *ctx,
+ RPRNFAState *state, RPRPatternElement *elem,
+ int64 currentPos)
+{
+ RPRPattern *pattern = winstate->rpPattern;
+ RPRPatternElement *elements = pattern->elements;
+ RPRElemIdx altIdx = elem->next;
+
+ while (altIdx >= 0)
+ {
+ RPRPatternElement *altElem;
+ RPRNFAState *newState;
+
+ /* Branch jump/next links are always -1 or a valid index */
+ Assert(altIdx < pattern->numElements);
+ altElem = &elements[altIdx];
+
+ /*
+ * Stop if element is outside ALT scope (not a branch). The check
+ * fires when the last branch is a quantified group whose BEGIN.jump
+ * (set by fillRPRPatternGroup) is preserved -- not overridden by
+ * fillRPRPatternAlt, which only links non-last branch heads -- and
+ * leads to a post-ALT element. Other branch shapes terminate the
+ * walk earlier via altIdx = RPR_ELEMIDX_INVALID. Use <=, not <: the
+ * post-ALT element may sit at the same depth as the ALT when the ALT
+ * has a sibling at that level.
+ */
+ if (altElem->depth <= elem->depth)
+ break;
+
+ /* Create independent state for each branch */
+ newState = nfa_state_clone(winstate, altIdx,
+ state->counts, state->isAbsorbable);
+
+ /* Recursively process this branch before next */
+ nfa_advance_state(winstate, ctx, newState, currentPos);
+ altIdx = altElem->jump;
+ }
+
+ nfa_state_free(winstate, state);
+}
+
+/*
+ * nfa_advance_begin
+ *
+ * Handle BEGIN element: group entry logic.
+ * BEGIN is only visited at initial group entry; loop-back from END goes
+ * directly to first child, bypassing BEGIN. Per the count-clear policy the
+ * group's own count slot is therefore already zero on entry (asserted below).
+ * If min=0, creates a skip path past the group.
+ */
+static void
+nfa_advance_begin(WindowAggState *winstate, RPRNFAContext *ctx,
+ RPRNFAState *state, RPRPatternElement *elem,
+ int64 currentPos)
+{
+ RPRPattern *pattern = winstate->rpPattern;
+ RPRPatternElement *elements = pattern->elements;
+ RPRNFAState *skipState = NULL;
+
+ /*
+ * Entry-side check of the count-clear policy: the group's own count slot
+ * is already zero here. BEGIN is only visited at initial group entry,
+ * and the previous occupant of this depth slot cleared it on exit.
+ */
+ Assert(state->counts[elem->depth] == 0);
+
+ /* Optional group: create skip path (but don't route yet) */
+ if (elem->min == 0)
+ {
+ skipState = nfa_state_clone(winstate, elem->jump,
+ state->counts, state->isAbsorbable);
+ }
+
+ if (skipState != NULL && RPRElemIsReluctant(elem))
+ {
+ RPRNFAState *savedMatch = ctx->matchedState;
+
+ /* Reluctant: skip first (prefer fewer iterations), enter second */
+ nfa_route_to_elem(winstate, ctx, skipState,
+ &elements[elem->jump], currentPos);
+
+ /*
+ * If skip path reached FIN, shortest match is found. Skip group entry
+ * to prevent longer matches.
+ */
+ if (ctx->matchedState != savedMatch)
+ {
+ nfa_state_free(winstate, state);
+ return;
+ }
+
+ state->elemIdx = elem->next;
+ nfa_route_to_elem(winstate, ctx, state,
+ &elements[state->elemIdx], currentPos);
+ }
+ else
+ {
+ /*
+ * Greedy-or-non-nullable: route to the first child. For optional
+ * groups (skipState != NULL, greedy min=0) additionally create the
+ * skip path; for non-nullable groups (skipState == NULL, min>0) the
+ * skip-path action is suppressed by the guard below.
+ */
+ state->elemIdx = elem->next;
+ nfa_route_to_elem(winstate, ctx, state,
+ &elements[state->elemIdx], currentPos);
+
+ if (skipState != NULL)
+ {
+ nfa_route_to_elem(winstate, ctx, skipState,
+ &elements[elem->jump], currentPos);
+ }
+ }
+}
+
+/*
+ * nfa_advance_end
+ *
+ * Handle END element: group repetition logic.
+ * Decides whether to loop back or exit based on count vs min/max.
+ */
+static void
+nfa_advance_end(WindowAggState *winstate, RPRNFAContext *ctx,
+ RPRNFAState *state, RPRPatternElement *elem,
+ int64 currentPos)
+{
+ RPRPattern *pattern = winstate->rpPattern;
+ RPRPatternElement *elements = pattern->elements;
+ int depth = elem->depth;
+ int32 count = state->counts[depth];
+
+ if (count < elem->min)
+ {
+ RPRPatternElement *jumpElem;
+ RPRNFAState *ffState = NULL;
+ RPRPatternElement *nextElem = NULL;
+
+ /*----------
+ * Two paths are explored when the group body is nullable
+ * (RPR_ELEM_EMPTY_LOOP):
+ *
+ * 1. Loop-back path: attempt real matches in the next iteration
+ * (state, modified below).
+ *
+ * 2. Fast-forward path: skip directly to after the group, treating
+ * all remaining required iterations as empty matches (ffState).
+ * Route to elem->next (not nfa_advance_end) to avoid creating
+ * competing greedy/reluctant loop states.
+ *
+ * Greedy prefers the loop-back first (more iterations); reluctant
+ * prefers the fast-forward (exit) first and, if it reaches FIN, drops
+ * the loop-back so a longer match cannot replace the shortest one --
+ * mirroring the min<=count<max branch below. The ffState snapshot is
+ * taken BEFORE modifying state, since both paths diverge from here.
+ *----------
+ */
+ if (RPRElemCanEmptyLoop(elem))
+ {
+ ffState = nfa_state_clone(winstate, state->elemIdx,
+ state->counts, state->isAbsorbable);
+
+ /* Exit the group: clear its own count (count-clear policy) */
+ ffState->counts[depth] = 0;
+ ffState->elemIdx = elem->next;
+ nextElem = &elements[ffState->elemIdx];
+
+ /*
+ * Unlike the must-exit path, no isAbsorbable update is needed:
+ * the fast-forward path runs only for EMPTY_LOOP (nullable)
+ * groups, which are never inside an absorbable region, so
+ * isAbsorbable is already false here.
+ */
+
+ /* END->END: increment outer END's count */
+ if (RPRElemIsEnd(nextElem) &&
+ ffState->counts[nextElem->depth] < RPR_COUNT_MAX)
+ ffState->counts[nextElem->depth]++;
+ }
+
+ /* Prepare the loop-back state */
+ state->elemIdx = elem->jump;
+ jumpElem = &elements[state->elemIdx];
+
+ if (ffState != NULL && RPRElemIsReluctant(elem))
+ {
+ RPRNFAState *savedMatch = ctx->matchedState;
+
+ /* Reluctant: take the fast-forward (exit) first */
+ nfa_route_to_elem(winstate, ctx, ffState, nextElem,
+ currentPos);
+
+ /*
+ * If the exit reached FIN, the shortest match is found. Skip the
+ * loop-back to prevent longer matches from replacing it.
+ */
+ if (ctx->matchedState != savedMatch)
+ {
+ nfa_state_free(winstate, state);
+ return;
+ }
+
+ /* Loop-back second */
+ nfa_route_to_elem(winstate, ctx, state, jumpElem,
+ currentPos);
+ }
+ else
+ {
+ /* Greedy (or non-nullable): loop-back first, fast-forward second */
+ nfa_route_to_elem(winstate, ctx, state, jumpElem,
+ currentPos);
+ if (ffState != NULL)
+ nfa_route_to_elem(winstate, ctx, ffState, nextElem,
+ currentPos);
+ }
+ }
+ else if (elem->max != RPR_QUANTITY_INF && count >= elem->max)
+ {
+ /* Must exit: reached max iterations. */
+ RPRPatternElement *nextElem;
+
+ /* Exit: clear the group's own count (count-clear policy) */
+ state->counts[depth] = 0;
+ state->elemIdx = elem->next;
+ nextElem = &elements[state->elemIdx];
+
+ /* Update isAbsorbable for target element (monotonic) */
+ state->isAbsorbable = state->isAbsorbable &&
+ RPRElemIsAbsorbableBranch(nextElem);
+
+ /* END->END: increment outer END's count */
+ if (RPRElemIsEnd(nextElem) && state->counts[nextElem->depth] < RPR_COUNT_MAX)
+ state->counts[nextElem->depth]++;
+
+ nfa_route_to_elem(winstate, ctx, state, nextElem, currentPos);
+ }
+ else
+ {
+ /*
+ * Between min and max (with at least one iteration) - can exit or
+ * loop. Greedy: loop first (prefer more iterations). Reluctant: exit
+ * first (prefer fewer iterations).
+ */
+ RPRNFAState *exitState;
+ RPRPatternElement *jumpElem;
+ RPRPatternElement *nextElem;
+
+ /*
+ * Create exit state first (need original counts before modifying
+ * state)
+ */
+ exitState = nfa_state_clone(winstate, elem->next,
+ state->counts, state->isAbsorbable);
+ /* Exit branch: clear the group's own count (count-clear policy) */
+ exitState->counts[depth] = 0;
+ nextElem = &elements[exitState->elemIdx];
+
+ /* END->END: increment outer END's count */
+ if (RPRElemIsEnd(nextElem) && exitState->counts[nextElem->depth] < RPR_COUNT_MAX)
+ exitState->counts[nextElem->depth]++;
+
+ /* Prepare loop state */
+ state->elemIdx = elem->jump;
+ jumpElem = &elements[state->elemIdx];
+
+ if (RPRElemIsReluctant(elem))
+ {
+ RPRNFAState *savedMatch = ctx->matchedState;
+
+ /* Exit first (preferred for reluctant) */
+ nfa_route_to_elem(winstate, ctx, exitState, nextElem,
+ currentPos);
+
+ /*
+ * If exit path reached FIN, shortest match is found. Skip loop to
+ * prevent longer matches from replacing it.
+ */
+ if (ctx->matchedState != savedMatch)
+ {
+ nfa_state_free(winstate, state);
+ return;
+ }
+
+ /* Loop second */
+ nfa_route_to_elem(winstate, ctx, state, jumpElem,
+ currentPos);
+ }
+ else
+ {
+ /* Loop first (preferred for greedy) */
+ nfa_route_to_elem(winstate, ctx, state, jumpElem,
+ currentPos);
+ /* Exit second */
+ nfa_route_to_elem(winstate, ctx, exitState, nextElem,
+ currentPos);
+ }
+ }
+}
+
+/*
+ * nfa_advance_var
+ *
+ * Handle VAR element: loop/exit transitions.
+ * After match phase, all VAR states have matched - decide next action.
+ */
+static void
+nfa_advance_var(WindowAggState *winstate, RPRNFAContext *ctx,
+ RPRNFAState *state, RPRPatternElement *elem,
+ int64 currentPos)
+{
+ RPRPattern *pattern = winstate->rpPattern;
+ RPRPatternElement *elements = pattern->elements;
+ int depth = elem->depth;
+ int32 count = state->counts[depth];
+ bool canLoop = (elem->max == RPR_QUANTITY_INF || count < elem->max);
+ bool canExit = (count >= elem->min);
+
+ /* min <= max, so !canExit (count < min) implies canLoop (count < max) */
+ Assert(canLoop || canExit);
+
+ /* elem->next must be a valid index for any reachable VAR */
+ Assert(elem->next >= 0 && elem->next < pattern->numElements);
+
+ if (canLoop && canExit)
+ {
+ /*
+ * Both loop and exit possible. Greedy: loop first (prefer longer
+ * match). Reluctant: exit first (prefer shorter match).
+ */
+ RPRNFAState *cloneState;
+ RPRPatternElement *nextElem;
+ bool reluctant = RPRElemIsReluctant(elem);
+
+ /*
+ * Clone state for the first-priority path. For greedy, clone is the
+ * loop state; for reluctant, clone is the exit state.
+ */
+ if (reluctant)
+ {
+ RPRNFAState *savedMatch = ctx->matchedState;
+
+ /* Clone for exit, original stays for loop */
+ cloneState = nfa_state_clone(winstate, elem->next,
+ state->counts, state->isAbsorbable);
+ /* Exit: clear the VAR's own count (count-clear policy) */
+ cloneState->counts[depth] = 0;
+ nextElem = &elements[cloneState->elemIdx];
+
+ /* When exiting directly to an outer END, increment its count */
+ if (RPRElemIsEnd(nextElem))
+ {
+ if (cloneState->counts[nextElem->depth] < RPR_COUNT_MAX)
+ cloneState->counts[nextElem->depth]++;
+ }
+
+ /* Exit first (preferred for reluctant) */
+ nfa_route_to_elem(winstate, ctx, cloneState, nextElem,
+ currentPos);
+
+ /*
+ * If exit path reached FIN, the shortest match is found. Skip
+ * loop state to prevent longer matches from replacing it.
+ */
+ if (ctx->matchedState != savedMatch)
+ {
+ nfa_state_free(winstate, state);
+ return;
+ }
+
+ /* Loop second */
+ nfa_add_state_unique(winstate, ctx, state);
+ }
+ else
+ {
+ /* Clone for loop, original used for exit */
+ cloneState = nfa_state_clone(winstate, state->elemIdx,
+ state->counts, state->isAbsorbable);
+
+ /* Loop first (preferred for greedy) */
+ nfa_add_state_unique(winstate, ctx, cloneState);
+
+ /* Exit second: clear the VAR's own count (count-clear policy) */
+ state->counts[depth] = 0;
+ state->elemIdx = elem->next;
+ nextElem = &elements[state->elemIdx];
+
+ /*
+ * Update isAbsorbable for target element (monotonic: AND
+ * preserves false)
+ */
+ state->isAbsorbable = state->isAbsorbable &&
+ RPRElemIsAbsorbableBranch(nextElem);
+
+ /*
+ * When exiting directly to an outer END, increment its iteration
+ * count. Simple VARs (min=max=1) handle this via inline advance
+ * in nfa_match, but quantified VARs bypass that path.
+ */
+ if (RPRElemIsEnd(nextElem))
+ {
+ if (state->counts[nextElem->depth] < RPR_COUNT_MAX)
+ state->counts[nextElem->depth]++;
+ }
+
+ nfa_route_to_elem(winstate, ctx, state, nextElem,
+ currentPos);
+ }
+ }
+ else if (canLoop)
+ {
+ /* Loop only: keep state as-is */
+ nfa_add_state_unique(winstate, ctx, state);
+ }
+ else
+ {
+ /* Exit only: advance to next element (canExit necessarily true) */
+ RPRPatternElement *nextElem;
+
+ Assert(canExit);
+ /* Exit: clear the VAR's own count (count-clear policy) */
+ state->counts[depth] = 0;
+ state->elemIdx = elem->next;
+ nextElem = &elements[state->elemIdx];
+
+ /*
+ * Update isAbsorbable for target element (monotonic: AND preserves
+ * false)
+ */
+ state->isAbsorbable = state->isAbsorbable &&
+ RPRElemIsAbsorbableBranch(nextElem);
+
+ /* See comment above: increment outer END count for quantified VARs */
+ if (RPRElemIsEnd(nextElem))
+ {
+ if (state->counts[nextElem->depth] < RPR_COUNT_MAX)
+ state->counts[nextElem->depth]++;
+ }
+
+ nfa_route_to_elem(winstate, ctx, state, nextElem, currentPos);
+ }
+}
+
+/*
+ * nfa_advance_state
+ *
+ * Recursively process a single state through epsilon transitions.
+ * DFS traversal ensures states are added to ctx->states in lexical order.
+ */
+static void
+nfa_advance_state(WindowAggState *winstate, RPRNFAContext *ctx,
+ RPRNFAState *state, int64 currentPos)
+{
+ RPRPattern *pattern = winstate->rpPattern;
+ RPRPatternElement *elem;
+
+ Assert(state->elemIdx >= 0 && state->elemIdx < pattern->numElements);
+
+ /* Protect against stack overflow for deeply complex patterns */
+ check_stack_depth();
+
+ /* Cycle detection: if this elemIdx was already visited in this DFS, bail */
+ if (winstate->nfaVisitedElems[WORDNUM(state->elemIdx)] &
+ ((bitmapword) 1 << BITNUM(state->elemIdx)))
+ {
+ nfa_state_free(winstate, state);
+ return;
+ }
+
+ elem = &pattern->elements[state->elemIdx];
+
+ /*
+ * Mark epsilon elements (END, ALT, BEGIN, FIN) in visited to prevent
+ * infinite epsilon cycles. VAR elements are marked later when added to
+ * the state list (nfa_add_state_unique), allowing legitimate loop-back to
+ * the same VAR in a new iteration.
+ */
+ if (!RPRElemIsVar(elem))
+ nfa_mark_visited(winstate, state->elemIdx);
+
+ switch (elem->varId)
+ {
+ case RPR_VARID_FIN:
+ /* FIN: record match */
+ nfa_add_matched_state(winstate, ctx, state, currentPos);
+ break;
+
+ case RPR_VARID_ALT:
+ nfa_advance_alt(winstate, ctx, state, elem, currentPos);
+ break;
+
+ case RPR_VARID_BEGIN:
+ nfa_advance_begin(winstate, ctx, state, elem, currentPos);
+ break;
+
+ case RPR_VARID_END:
+ nfa_advance_end(winstate, ctx, state, elem, currentPos);
+ break;
+
+ default:
+ /* VAR element */
+ nfa_advance_var(winstate, ctx, state, elem, currentPos);
+ break;
+ }
+}
+
+/*
+ * nfa_advance
+ *
+ * Advance phase (divergence): transition from all surviving states.
+ * Called after match phase with matched VAR states, or at context creation
+ * for initial epsilon expansion (with currentPos = startPos - 1).
+ *
+ * Processes states in order, using recursive DFS to maintain lexical order.
+ */
+static void
+nfa_advance(WindowAggState *winstate, RPRNFAContext *ctx, int64 currentPos)
+{
+ RPRNFAState *states = ctx->states;
+ RPRNFAState *state;
+ RPRNFAState *savedMatchedState;
+
+ ctx->states = NULL; /* Will rebuild */
+
+ /* Process each state in lexical order (DFS order from previous advance) */
+ while (states != NULL)
+ {
+ CHECK_FOR_INTERRUPTS();
+ savedMatchedState = ctx->matchedState;
+
+ /*
+ * Clear visited bitmap before each state's DFS expansion. Only the
+ * range touched since the previous reset (tracked via the high-water
+ * marks updated in nfa_mark_visited) needs to be cleared; for small
+ * NFAs this is the whole array, but for large NFAs whose DFS only
+ * reaches a few elements per advance it avoids walking the full
+ * bitmap.
+ */
+ if (winstate->nfaVisitedMaxWord >= winstate->nfaVisitedMinWord)
+ {
+ memset(&winstate->nfaVisitedElems[winstate->nfaVisitedMinWord], 0,
+ sizeof(bitmapword) *
+ (winstate->nfaVisitedMaxWord -
+ winstate->nfaVisitedMinWord + 1));
+ winstate->nfaVisitedMinWord = PG_INT16_MAX;
+ winstate->nfaVisitedMaxWord = -1;
+ }
+
+ state = states;
+ states = states->next;
+
+ /*
+ * Boundary contract: state->next is reset to NULL here, before
+ * crossing into nfa_advance_state's epsilon-expansion DFS. The inner
+ * branches (nfa_advance_var, nfa_advance_begin/end/alt) treat
+ * state->next as already-NULL and don't reset it themselves; the
+ * other linking site is nfa_add_state_unique, which sets it when
+ * appending to ctx->states.
+ */
+ state->next = NULL;
+
+ nfa_advance_state(winstate, ctx, state, currentPos);
+
+ /*
+ * Early termination: if a FIN was newly reached in this advance,
+ * remaining old states have worse lexical order and can be pruned.
+ * Only check for new FIN arrivals (not ones from previous rows).
+ */
+ if (ctx->matchedState != savedMatchedState && states != NULL)
+ {
+ nfa_state_free_list(winstate, states);
+ break;
+ }
+ }
+}
+
+/*
+ * nfa_reevaluate_dependent_vars
+ * Re-evaluate match_start-dependent DEFINE variables for a specific
+ * context whose matchStartRow differs from the shared evaluation's
+ * nav_match_start.
+ *
+ * Only variables in defineMatchStartDependent are re-evaluated. The
+ * current row's slot (ecxt_outertuple) must already be set up by
+ * nfa_evaluate_row().
+ */
+static void
+nfa_reevaluate_dependent_vars(WindowAggState *winstate, RPRNFAContext *ctx,
+ int64 currentPos)
+{
+ ExprContext *econtext = winstate->ss.ps.ps_ExprContext;
+ int64 saved_match_start = winstate->nav_match_start;
+ int64 saved_pos = winstate->currentpos;
+
+ /* Temporarily set nav_match_start and currentpos for FIRST/LAST */
+ winstate->nav_match_start = ctx->matchStartRow;
+ winstate->currentpos = currentPos;
+
+ /* Invalidate nav_slot cache since match_start changed */
+ winstate->nav_slot_pos = -1;
+
+ foreach_ptr(ExprState, exprState, winstate->defineClauseList)
+ {
+ int varIdx = foreach_current_index(exprState);
+
+ if (bms_is_member(varIdx, winstate->defineMatchStartDependent))
+ {
+ Datum result;
+ bool isnull;
+
+ result = ExecEvalExpr(exprState, econtext, &isnull);
+ winstate->nfaVarMatched[varIdx] = (!isnull && DatumGetBool(result));
+ }
+
+ if (varIdx + 1 >= list_length(winstate->defineVariableList))
+ break;
+ }
+
+ /* Restore original match_start, currentpos, and invalidate cache */
+ winstate->nav_match_start = saved_match_start;
+ winstate->currentpos = saved_pos;
+ winstate->nav_slot_pos = -1;
+}
+
+
+/***********************************************************************
+ * API exposed to nodeWindowAgg.c
+ ***********************************************************************/
+
+/*
+ * ExecRPRStartContext
+ *
+ * Start a new match context at given position.
+ * Initializes context, state absorption flags, and performs initial advance
+ * to expand epsilon transitions (ALT branches, optional elements).
+ * Adds context to the tail of winstate->nfaContext list.
+ */
+RPRNFAContext *
+ExecRPRStartContext(WindowAggState *winstate, int64 startPos)
+{
+ RPRNFAContext *ctx;
+ RPRPattern *pattern = winstate->rpPattern;
+ RPRPatternElement *elem;
+
+ ctx = nfa_context_make(winstate);
+ ctx->matchStartRow = startPos;
+ ctx->states = nfa_state_make(winstate); /* initial state at elem 0 */
+
+ elem = &pattern->elements[0];
+
+ if (RPRElemIsAbsorbableBranch(elem))
+ {
+ ctx->states->isAbsorbable = true;
+ }
+ else
+ {
+ ctx->hasAbsorbableState = false;
+ ctx->allStatesAbsorbable = false;
+ ctx->states->isAbsorbable = false;
+ }
+
+ /* Add to tail of active context list (doubly-linked, oldest-first) */
+ ctx->prev = winstate->nfaContextTail;
+ ctx->next = NULL;
+ if (winstate->nfaContextTail != NULL)
+ winstate->nfaContextTail->next = ctx;
+ else
+ winstate->nfaContext = ctx; /* first context becomes head */
+ winstate->nfaContextTail = ctx;
+
+ /*
+ * Initial advance (divergence): expand ALT branches and create exit
+ * states for VAR elements with min=0. This prepares the context for the
+ * first row's match phase.
+ *
+ * Use startPos - 1 as currentPos since no row has been consumed yet. If
+ * FIN is reached via epsilon transitions, matchEndRow = startPos - 1
+ * which is less than matchStartRow, resulting in UNMATCHED treatment.
+ */
+ nfa_advance(winstate, ctx, startPos - 1);
+
+ return ctx;
+}
+
+/*
+ * ExecRPRGetHeadContext
+ *
+ * Return the head context if its start position matches pos.
+ * Returns NULL if no context exists or head doesn't match pos.
+ */
+RPRNFAContext *
+ExecRPRGetHeadContext(WindowAggState *winstate, int64 pos)
+{
+ RPRNFAContext *ctx = winstate->nfaContext;
+
+ /*
+ * Contexts are sorted by matchStartRow ascending. If the head context
+ * doesn't match pos, no context exists for this position.
+ */
+ if (ctx == NULL || ctx->matchStartRow != pos)
+ return NULL;
+
+ return ctx;
+}
+
+/*
+ * ExecRPRFreeContext
+ *
+ * Unlink context from active list and return it to free list.
+ * Also frees any states in the context.
+ */
+void
+ExecRPRFreeContext(WindowAggState *winstate, RPRNFAContext *ctx)
+{
+ /* Unlink from active list first */
+ nfa_unlink_context(winstate, ctx);
+
+ /* Update statistics */
+ winstate->nfaContextsActive--;
+
+ if (ctx->states != NULL)
+ nfa_state_free_list(winstate, ctx->states);
+ if (ctx->matchedState != NULL)
+ nfa_state_free(winstate, ctx->matchedState);
+
+ ctx->states = NULL;
+ ctx->matchedState = NULL;
+ ctx->next = winstate->nfaContextFree;
+ winstate->nfaContextFree = ctx;
+}
+
+/*
+ * ExecRPRRecordContextSuccess
+ *
+ * Record a successful context in statistics.
+ */
+void
+ExecRPRRecordContextSuccess(WindowAggState *winstate, int64 matchLen)
+{
+ winstate->nfaMatchesSucceeded++;
+ nfa_update_length_stats(winstate->nfaMatchesSucceeded,
+ &winstate->nfaMatchLen,
+ matchLen);
+}
+
+/*
+ * ExecRPRRecordContextFailure
+ *
+ * Record a failed context in statistics.
+ * If failedLen == 1, count as pruned (failed on first row).
+ * If failedLen > 1, count as mismatched and update length stats.
+ */
+void
+ExecRPRRecordContextFailure(WindowAggState *winstate, int64 failedLen)
+{
+ if (failedLen == 1)
+ {
+ winstate->nfaContextsPruned++;
+ }
+ else
+ {
+ winstate->nfaMatchesFailed++;
+ nfa_update_length_stats(winstate->nfaMatchesFailed,
+ &winstate->nfaFailLen,
+ failedLen);
+ }
+}
+
+/*
+ * ExecRPRProcessRow
+ *
+ * Process all contexts for one row:
+ * 1. Match all contexts (convergence) - evaluate VARs, prune dead states
+ * 2. Absorb redundant contexts - ideal timing after convergence
+ * 3. Advance all contexts (divergence) - create new states for next row
+ */
+void
+ExecRPRProcessRow(WindowAggState *winstate, int64 currentPos,
+ bool hasLimitedFrame, int64 frameOffset)
+{
+ RPRNFAContext *ctx;
+ bool *varMatched = winstate->nfaVarMatched;
+ bool hasDependent = !bms_is_empty(winstate->defineMatchStartDependent);
+
+ /* Allow query cancellation once per row for simple/low-state patterns */
+ CHECK_FOR_INTERRUPTS();
+
+ /*
+ * Phase 1: Match all contexts (convergence). Evaluate VAR elements,
+ * update counts, remove dead states.
+ */
+ for (ctx = winstate->nfaContext; ctx != NULL; ctx = ctx->next)
+ {
+ if (ctx->states == NULL)
+ continue;
+
+ /* Check frame boundary - finalize the context when it is reached */
+ if (hasLimitedFrame)
+ {
+ int64 ctxFrameEnd;
+
+ /*
+ * Clamp to PG_INT64_MAX on overflow. frameOffset can be as large
+ * as PG_INT64_MAX (e.g. "ROWS <huge> FOLLOWING"), so add the
+ * offset and the trailing +1 in two separately checked steps to
+ * avoid signed-integer overflow in the "frameOffset + 1"
+ * subexpression.
+ */
+ if (pg_add_s64_overflow(ctx->matchStartRow, frameOffset,
+ &ctxFrameEnd) ||
+ pg_add_s64_overflow(ctxFrameEnd, 1, &ctxFrameEnd))
+ ctxFrameEnd = PG_INT64_MAX;
+
+ /*
+ * currentPos advances by exactly one per call, and a finalized
+ * context is skipped by the states == NULL guard above, so it can
+ * only ever reach ctxFrameEnd, never overshoot it. The Assert
+ * turns a future change that broke that invariant into an
+ * immediate failure rather than a silent slip past the boundary.
+ */
+ Assert(currentPos <= ctxFrameEnd);
+
+ if (currentPos == ctxFrameEnd)
+ {
+ /* Frame boundary reached: force mismatch */
+ nfa_match(winstate, ctx, NULL);
+ continue;
+ }
+ }
+
+ /*
+ * If this context has a different matchStartRow than the one used in
+ * the shared evaluation, re-evaluate match_start-dependent variables
+ * with this context's matchStartRow.
+ */
+ if (hasDependent && ctx->matchStartRow != winstate->nav_match_start)
+ nfa_reevaluate_dependent_vars(winstate, ctx, currentPos);
+ nfa_match(winstate, ctx, varMatched);
+ ctx->lastProcessedRow = currentPos;
+ }
+
+ /*
+ * Phase 2: Absorb redundant contexts. After match phase, states have
+ * converged - ideal for absorption. First update absorption flags that
+ * may have changed due to state removal.
+ */
+ if (winstate->rpPattern->isAbsorbable)
+ {
+ for (ctx = winstate->nfaContext; ctx != NULL; ctx = ctx->next)
+ nfa_update_absorption_flags(ctx);
+
+ nfa_absorb_contexts(winstate);
+ }
+
+ /*
+ * Phase 3: Advance all contexts (divergence). Create new states
+ * (loop/exit) from surviving matched states.
+ */
+ for (ctx = winstate->nfaContext; ctx != NULL; ctx = ctx->next)
+ {
+ if (ctx->states == NULL)
+ continue;
+
+ /*
+ * Phase 1 already handled frame boundary exceeded contexts by forcing
+ * mismatch (nfa_match with NULL), which removes all states (all
+ * states are at VAR positions after advance). So any surviving
+ * context here must be within its frame boundary.
+ *
+ * Compute the (clamped) frame end the same way as Phase 1, using two
+ * separately checked adds so that "frameOffset + 1" cannot overflow
+ * when frameOffset is near PG_INT64_MAX.
+ */
+#ifdef USE_ASSERT_CHECKING
+ if (hasLimitedFrame)
+ {
+ int64 ctxFrameEnd;
+
+ if (pg_add_s64_overflow(ctx->matchStartRow, frameOffset,
+ &ctxFrameEnd) ||
+ pg_add_s64_overflow(ctxFrameEnd, 1, &ctxFrameEnd))
+ ctxFrameEnd = PG_INT64_MAX;
+ Assert(currentPos < ctxFrameEnd);
+ }
+#endif
+
+ nfa_advance(winstate, ctx, currentPos);
+ }
+}
+
+/*
+ * ExecRPRCleanupDeadContexts
+ *
+ * Remove contexts that have failed (no active states and no match).
+ * These are contexts that failed during normal processing and should be
+ * counted as pruned (if length 1) or mismatched (if length > 1).
+ */
+void
+ExecRPRCleanupDeadContexts(WindowAggState *winstate, RPRNFAContext *excludeCtx)
+{
+ RPRNFAContext *ctx;
+ RPRNFAContext *next;
+
+ for (ctx = winstate->nfaContext; ctx != NULL; ctx = next)
+ {
+ CHECK_FOR_INTERRUPTS();
+
+ next = ctx->next;
+
+ /* Skip the target context and contexts still processing */
+ if (ctx == excludeCtx || ctx->states != NULL)
+ continue;
+
+ /* Skip successfully matched contexts (will be handled by SKIP logic) */
+ if (ctx->matchEndRow >= ctx->matchStartRow)
+ continue;
+
+ /*
+ * Failed context: always removed below. Only record the failure
+ * statistic if it actually processed its start row; contexts created
+ * for beyond-partition rows are removed without being counted.
+ */
+ if (ctx->lastProcessedRow >= ctx->matchStartRow)
+ {
+ int64 failedLen = ctx->lastProcessedRow - ctx->matchStartRow + 1;
+
+ ExecRPRRecordContextFailure(winstate, failedLen);
+ }
+
+ ExecRPRFreeContext(winstate, ctx);
+ }
+}
+
+/*
+ * ExecRPRFinalizeAllContexts
+ *
+ * Partition-end classification policy: kill any VAR states still pursuing
+ * when rows run out, so cleanup sees a uniform ctx->states == NULL across
+ * every context. By the time this runs, all genuine FIN reaches have
+ * already been recorded in-flight; three shapes survive here:
+ *
+ * - Pure pursuit (matchedState == NULL): VAR states waiting for input
+ * that never arrives (e.g., A+ B mid-pattern at partition end).
+ * - Empty-match candidate + pursuit (matchedState != NULL,
+ * matchEndRow < matchStartRow): initial-advance FIN-via-skip recorded
+ * an empty match while VAR states are still chasing a longer one
+ * (e.g., greedy A*).
+ * - Real match + pursuit (matchedState != NULL,
+ * matchEndRow >= matchStartRow): a match has been recorded and VAR
+ * states are still looping for a longer one.
+ *
+ * Killing the VAR reclassifies the first two as failures in cleanup
+ * (otherwise they linger without contributing to stats). The third is
+ * stat-neutral -- cleanup skips it either way -- but goes through the
+ * same uniform path so partition-end classification stays centralized.
+ *
+ * Implementation: nfa_match with NULL forces VAR mismatch; nfa_advance
+ * then drains any remaining epsilon transitions.
+ */
+void
+ExecRPRFinalizeAllContexts(WindowAggState *winstate, int64 lastPos)
+{
+ RPRNFAContext *ctx;
+
+ for (ctx = winstate->nfaContext; ctx != NULL; ctx = ctx->next)
+ {
+ CHECK_FOR_INTERRUPTS();
+
+ if (ctx->states != NULL)
+ {
+ nfa_match(winstate, ctx, NULL);
+ nfa_advance(winstate, ctx, lastPos);
+ }
+ }
+}
diff --git a/src/backend/executor/meson.build b/src/backend/executor/meson.build
index dc45be0b2ce..0ff4a5b1d83 100644
--- a/src/backend/executor/meson.build
+++ b/src/backend/executor/meson.build
@@ -13,6 +13,7 @@ backend_sources += files(
'execParallel.c',
'execPartition.c',
'execProcnode.c',
+ 'execRPR.c',
'execReplication.c',
'execSRF.c',
'execScan.c',
diff --git a/src/backend/executor/nodeWindowAgg.c b/src/backend/executor/nodeWindowAgg.c
index f1c524d00df..cb6a484b7de 100644
--- a/src/backend/executor/nodeWindowAgg.c
+++ b/src/backend/executor/nodeWindowAgg.c
@@ -39,12 +39,15 @@
#include "catalog/pg_proc.h"
#include "common/int.h"
#include "executor/executor.h"
+#include "executor/execRPR.h"
#include "executor/instrument.h"
#include "executor/nodeWindowAgg.h"
#include "miscadmin.h"
#include "nodes/nodeFuncs.h"
+#include "nodes/plannodes.h"
#include "optimizer/clauses.h"
#include "optimizer/optimizer.h"
+#include "optimizer/rpr.h"
#include "parser/parse_agg.h"
#include "parser/parse_coerce.h"
#include "utils/acl.h"
@@ -174,6 +177,7 @@ typedef struct WindowStatePerAggData
bool restart; /* need to restart this agg in this cycle? */
} WindowStatePerAggData;
+
static void initialize_windowaggregate(WindowAggState *winstate,
WindowStatePerFunc perfuncstate,
WindowStatePerAgg peraggstate);
@@ -210,6 +214,9 @@ static Datum GetAggInitVal(Datum textInitVal, Oid transtype);
static bool are_peers(WindowAggState *winstate, TupleTableSlot *slot1,
TupleTableSlot *slot2);
+static int WinGetSlotInFrame(WindowObject winobj, TupleTableSlot *slot,
+ int relpos, int seektype, bool set_mark,
+ bool *isnull, bool *isout);
static bool window_gettupleslot(WindowObject winobj, int64 pos,
TupleTableSlot *slot);
@@ -228,6 +235,18 @@ static uint8 get_notnull_info(WindowObject winobj,
int64 pos, int argno);
static void put_notnull_info(WindowObject winobj,
int64 pos, int argno, bool isnull);
+static bool rpr_is_defined(WindowAggState *winstate);
+static int64 row_is_in_reduced_frame(WindowObject winobj, int64 pos);
+
+static void clear_reduced_frame(WindowAggState *winstate);
+static int get_reduced_frame_status(WindowAggState *winstate, int64 pos);
+static void update_reduced_frame(WindowObject winobj, int64 pos);
+
+/* Forward declarations - NFA row evaluation */
+static bool nfa_evaluate_row(WindowObject winobj, int64 pos, bool *varMatched);
+
+/* Forward declarations - navigation offset evaluation */
+static void eval_define_offsets(WindowAggState *winstate, List *defineClause);
/*
* Not null info bit array consists of 2-bit items
@@ -821,6 +840,9 @@ eval_windowaggregates(WindowAggState *winstate)
* transition function, or
* - we have an EXCLUSION clause, or
* - if the new frame doesn't overlap the old one
+ * - if RPR (Row Pattern Recognition) is enabled, because the reduced
+ * frame depends on pattern matching results which can differ entirely
+ * from row to row, making inverse transition optimization inapplicable
*
* Note that we don't strictly need to restart in the last case, but if
* we're going to remove all rows from the aggregation anyway, a restart
@@ -835,7 +857,8 @@ eval_windowaggregates(WindowAggState *winstate)
(winstate->aggregatedbase != winstate->frameheadpos &&
!OidIsValid(peraggstate->invtransfn_oid)) ||
(winstate->frameOptions & FRAMEOPTION_EXCLUSION) ||
- winstate->aggregatedupto <= winstate->frameheadpos)
+ winstate->aggregatedupto <= winstate->frameheadpos ||
+ rpr_is_defined(winstate))
{
peraggstate->restart = true;
numaggs_restart++;
@@ -964,6 +987,14 @@ eval_windowaggregates(WindowAggState *winstate)
{
winstate->aggregatedupto = winstate->frameheadpos;
ExecClearTuple(agg_row_slot);
+
+ /*
+ * If RPR is defined, we do not use aggregatedupto_nonrestarted. To
+ * avoid assertion failure below, we reset aggregatedupto_nonrestarted
+ * to frameheadpos.
+ */
+ if (rpr_is_defined(winstate))
+ aggregatedupto_nonrestarted = winstate->frameheadpos;
}
/*
@@ -975,7 +1006,7 @@ eval_windowaggregates(WindowAggState *winstate)
*/
for (;;)
{
- int ret;
+ int64 ret;
/* Fetch next row if we didn't already */
if (TupIsNull(agg_row_slot))
@@ -993,9 +1024,40 @@ eval_windowaggregates(WindowAggState *winstate)
agg_row_slot, false);
if (ret < 0)
break;
+
if (ret == 0)
goto next_tuple;
+ if (rpr_is_defined(winstate))
+ {
+ /*
+ * If currentpos is already decided but aggregatedupto is not yet
+ * determined, we've passed the last reduced frame.
+ */
+ if (get_reduced_frame_status(winstate, winstate->currentpos)
+ != RF_NOT_DETERMINED &&
+ get_reduced_frame_status(winstate, winstate->aggregatedupto)
+ == RF_NOT_DETERMINED)
+ break;
+
+ /*
+ * Calculate the reduced frame for aggregatedupto.
+ */
+ ret = row_is_in_reduced_frame(winstate->agg_winobj,
+ winstate->aggregatedupto);
+ if (ret == -1) /* unmatched row */
+ break;
+
+ /*
+ * Check if current row is inside a match but not the head
+ * (skipped), and it's the base row for aggregation.
+ */
+ if (get_reduced_frame_status(winstate,
+ winstate->aggregatedupto) == RF_SKIPPED &&
+ winstate->aggregatedupto == winstate->aggregatedbase)
+ break;
+ }
+
/* Set tuple context for evaluation of aggregate arguments */
winstate->tmpcontext->ecxt_outertuple = agg_row_slot;
@@ -1024,6 +1086,7 @@ next_tuple:
ExecClearTuple(agg_row_slot);
}
+
/* The frame's end is not supposed to move backwards, ever */
Assert(aggregatedupto_nonrestarted <= winstate->aggregatedupto);
@@ -1191,6 +1254,28 @@ prepare_tuplestore(WindowAggState *winstate)
}
}
+ /* Create read/mark pointers for RPR navigation if needed */
+ if (winstate->nav_winobj)
+ {
+ /*
+ * Allocate mark and read pointers for RPR navigation.
+ *
+ * If navMaxOffsetKind == RPR_NAV_OFFSET_FIXED, we advance the mark
+ * based on (currentpos - navMaxOffset) and optionally
+ * (nfaContext->matchStartRow + navFirstOffset), allowing
+ * tuplestore_trim() to free rows that are no longer reachable.
+ *
+ * RPR_NAV_OFFSET_NEEDS_EVAL is resolved at executor init; by this
+ * point it is either FIXED or RETAIN_ALL.
+ */
+ winstate->nav_winobj->markptr =
+ tuplestore_alloc_read_pointer(winstate->buffer, 0);
+ winstate->nav_winobj->readptr =
+ tuplestore_alloc_read_pointer(winstate->buffer,
+ EXEC_FLAG_BACKWARD);
+ winstate->nav_winobj->markpos = 0;
+ }
+
/*
* If we are in RANGE or GROUPS mode, then determining frame boundaries
* requires physical access to the frame endpoint rows, except in certain
@@ -1247,6 +1332,8 @@ begin_partition(WindowAggState *winstate)
winstate->framehead_valid = false;
winstate->frametail_valid = false;
winstate->grouptail_valid = false;
+ if (rpr_is_defined(winstate))
+ clear_reduced_frame(winstate);
winstate->spooled_rows = 0;
winstate->currentpos = 0;
winstate->frameheadpos = 0;
@@ -1300,6 +1387,13 @@ begin_partition(WindowAggState *winstate)
winstate->aggregatedupto = 0;
}
+ /* reset mark and seek positions for RPR navigation */
+ if (winstate->nav_winobj)
+ {
+ winstate->nav_winobj->markpos = -1;
+ winstate->nav_winobj->seekpos = -1;
+ }
+
/* reset mark and seek positions for each real window function */
for (int i = 0; i < numfuncs; i++)
{
@@ -1468,6 +1562,21 @@ release_partition(WindowAggState *winstate)
tuplestore_clear(winstate->buffer);
winstate->partition_spooled = false;
winstate->next_partition = true;
+
+ /* Reset RPR match results */
+ clear_reduced_frame(winstate);
+
+ /* Reset NFA state for new partition */
+ winstate->nfaContext = NULL;
+ winstate->nfaContextTail = NULL;
+ winstate->nfaContextFree = NULL;
+ winstate->nfaStateFree = NULL;
+ winstate->nfaLastProcessedRow = -1;
+ winstate->nfaStatesActive = 0;
+ winstate->nfaContextsActive = 0;
+
+ /* Invalidate the nav slot position cache for the new partition. */
+ winstate->nav_slot_pos = -1;
}
/*
@@ -2263,6 +2372,16 @@ calculate_frame_offsets(PlanState *pstate)
ereport(ERROR,
(errcode(ERRCODE_INVALID_PRECEDING_OR_FOLLOWING_SIZE),
errmsg("frame ending offset must not be negative")));
+
+ /*
+ * Row pattern recognition forbids a zero-length frame end;
+ * checked here so a non-constant offset (e.g. a bind parameter)
+ * is caught, not just a literal 0.
+ */
+ if (winstate->rpPattern != NULL && offset == 0)
+ ereport(ERROR,
+ errcode(ERRCODE_WINDOWING_ERROR),
+ errmsg("frame ending offset must be positive with row pattern recognition"));
}
}
winstate->all_first = false;
@@ -2396,6 +2515,16 @@ ExecWindowAgg(PlanState *pstate)
/* don't evaluate the window functions when we're in pass-through mode */
if (winstate->status == WINDOWAGG_RUN)
{
+ /*
+ * If RPR is defined and skip mode is next row, clear the current
+ * match so the next row triggers re-evaluation.
+ */
+ if (rpr_is_defined(winstate))
+ {
+ if (winstate->rpSkipTo == ST_NEXT_ROW)
+ clear_reduced_frame(winstate);
+ }
+
/*
* Evaluate true window functions
*/
@@ -2435,6 +2564,43 @@ ExecWindowAgg(PlanState *pstate)
if (winstate->grouptail_ptr >= 0)
update_grouptailpos(winstate);
+ /*
+ * Advance RPR navigation mark pointer if possible, so that
+ * tuplestore_trim() can free rows no longer reachable by navigation.
+ */
+ if (winstate->nav_winobj &&
+ winstate->rpPattern != NULL &&
+ winstate->navMaxOffsetKind == RPR_NAV_OFFSET_FIXED)
+ {
+ int64 navmarkpos;
+
+ /* Backward reach from PREV/LAST/compound PREV_LAST/NEXT_LAST */
+ if (winstate->currentpos > winstate->navMaxOffset)
+ navmarkpos = winstate->currentpos - winstate->navMaxOffset;
+ else
+ navmarkpos = 0;
+
+ /*
+ * If FIRST is used, also consider match_start + navFirstOffset.
+ * The oldest active context (nfaContext) has the smallest
+ * matchStartRow.
+ */
+ if (winstate->hasFirstNav &&
+ winstate->navFirstOffsetKind == RPR_NAV_OFFSET_FIXED &&
+ winstate->nfaContext != NULL)
+ {
+ int64 firstreach;
+
+ if (!pg_add_s64_overflow(winstate->nfaContext->matchStartRow,
+ winstate->navFirstOffset,
+ &firstreach))
+ navmarkpos = Min(navmarkpos, Max(firstreach, 0));
+ }
+
+ if (navmarkpos > winstate->nav_winobj->markpos)
+ WinSetMarkPosition(winstate->nav_winobj, navmarkpos);
+ }
+
/*
* Truncate any no-longer-needed rows from the tuplestore.
*/
@@ -2660,6 +2826,20 @@ ExecInitWindowAgg(WindowAgg *node, EState *estate, int eflags)
winstate->temp_slot_2 = ExecInitExtraTupleSlot(estate, scanDesc,
&TTSOpsMinimalTuple);
+ if (node->rpPattern != NULL)
+ {
+ winstate->nav_slot = ExecInitExtraTupleSlot(estate, scanDesc,
+ &TTSOpsMinimalTuple);
+ winstate->nav_slot_pos = -1;
+
+ winstate->nav_null_slot = ExecInitExtraTupleSlot(estate, scanDesc,
+ &TTSOpsMinimalTuple);
+ winstate->nav_null_slot = ExecStoreAllNullTuple(winstate->nav_null_slot);
+
+ winstate->nav_saved_outertuple = NULL;
+ winstate->nav_match_start = 0;
+ }
+
/*
* create frame head and tail slots only if needed (must create slots in
* exactly the same cases that update_frameheadpos and update_frametailpos
@@ -2828,6 +3008,23 @@ ExecInitWindowAgg(WindowAgg *node, EState *estate, int eflags)
winstate->agg_winobj = agg_winobj;
}
+ /*
+ * Set up WindowObject for RPR navigation opcodes. This is separate from
+ * agg_winobj because it needs its own read pointer to avoid interfering
+ * with aggregate processing.
+ */
+ if (node->rpPattern != NULL)
+ {
+ WindowObject nav_winobj = makeNode(WindowObjectData);
+
+ nav_winobj->winstate = winstate;
+ nav_winobj->argstates = NIL;
+ nav_winobj->localmem = NULL;
+ nav_winobj->markptr = -1;
+ nav_winobj->readptr = -1;
+ winstate->nav_winobj = nav_winobj;
+ }
+
/* Set the status to running */
winstate->status = WINDOWAGG_RUN;
@@ -2846,6 +3043,81 @@ ExecInitWindowAgg(WindowAgg *node, EState *estate, int eflags)
winstate->inRangeAsc = node->inRangeAsc;
winstate->inRangeNullsFirst = node->inRangeNullsFirst;
+ /* Set up SKIP TO type */
+ winstate->rpSkipTo = node->rpSkipTo;
+ /* Set up row pattern recognition PATTERN clause (compiled NFA) */
+ winstate->rpPattern = node->rpPattern;
+ /* Set up nav offsets for tuplestore trim; resolve any NEEDS_EVAL kinds */
+ winstate->navMaxOffsetKind = node->navMaxOffsetKind;
+ winstate->navMaxOffset = node->navMaxOffset;
+ winstate->hasFirstNav = node->hasFirstNav;
+ winstate->navFirstOffsetKind = node->navFirstOffsetKind;
+ winstate->navFirstOffset = node->navFirstOffset;
+ eval_define_offsets(winstate, node->defineClause);
+
+ /* Copy match_start dependency bitmapset for per-context evaluation */
+ winstate->defineMatchStartDependent = bms_copy(node->defineMatchStartDependent);
+
+ /* Calculate NFA state size and allocate cycle detection bitmap */
+ if (node->rpPattern != NULL)
+ {
+ int nfaVisitedNWords;
+
+ winstate->nfaStateSize = offsetof(RPRNFAState, counts) +
+ sizeof(int32) * node->rpPattern->maxDepth;
+ nfaVisitedNWords =
+ (node->rpPattern->numElements - 1) / BITS_PER_BITMAPWORD + 1;
+ winstate->nfaVisitedElems = palloc0(sizeof(bitmapword) *
+ nfaVisitedNWords);
+ /* High-water mark sentinels: no bits set yet. */
+ winstate->nfaVisitedMinWord = PG_INT16_MAX;
+ winstate->nfaVisitedMaxWord = -1;
+ }
+
+ /* Set up row pattern recognition DEFINE clause */
+ winstate->defineVariableList = NIL;
+ winstate->defineClauseList = NIL;
+ if (node->defineClause != NIL)
+ {
+ /*
+ * Compile DEFINE clause expressions. PREV/NEXT navigation is handled
+ * by EEOP_RPR_NAV_SET/RESTORE opcodes emitted during ExecInitExpr, so
+ * no varno rewriting is needed here.
+ */
+ foreach_node(TargetEntry, te, node->defineClause)
+ {
+ char *name = te->resname;
+ Expr *expr = te->expr;
+ ExprState *exps;
+
+ winstate->defineVariableList =
+ lappend(winstate->defineVariableList,
+ makeString(pstrdup(name)));
+ exps = ExecInitExpr(expr, (PlanState *) winstate);
+ winstate->defineClauseList =
+ lappend(winstate->defineClauseList, exps);
+ }
+ }
+
+ /* Initialize NFA free lists for row pattern matching */
+ winstate->nfaContext = NULL;
+ winstate->nfaContextTail = NULL;
+ winstate->nfaContextFree = NULL;
+ winstate->nfaStateFree = NULL;
+ winstate->nfaLastProcessedRow = -1;
+ winstate->nfaStatesActive = 0;
+ winstate->nfaContextsActive = 0;
+
+ /*
+ * Allocate varMatched array for NFA evaluation. With the new varNames
+ * ordering (DEFINE order first), varId == defineIdx for all defined
+ * variables, so no mapping is needed.
+ */
+ if (list_length(winstate->defineVariableList) > 0)
+ winstate->nfaVarMatched = palloc0(sizeof(bool) *
+ list_length(winstate->defineVariableList));
+ else
+ winstate->nfaVarMatched = NULL;
winstate->all_first = true;
winstate->partition_spooled = false;
winstate->more_partitions = false;
@@ -2854,6 +3126,42 @@ ExecInitWindowAgg(WindowAgg *node, EState *estate, int eflags)
return winstate;
}
+/*
+ * ExecRPRNavGetSlot
+ *
+ * Fetch tuple at given position for RPR navigation opcodes.
+ * Returns nav_slot with the tuple loaded, or nav_null_slot if out of range.
+ */
+TupleTableSlot *
+ExecRPRNavGetSlot(WindowAggState *winstate, int64 pos)
+{
+ WindowObject winobj = winstate->nav_winobj;
+ TupleTableSlot *slot = winstate->nav_slot;
+
+ if (pos < 0)
+ return winstate->nav_null_slot;
+
+ /*
+ * If nav_slot already holds this position, return it without re-fetching.
+ * This is critical when multiple PREV/NEXT calls in the same expression
+ * navigate to the same row, because re-fetching would free the slot's
+ * tuple memory and invalidate any pass-by-ref Datum pointers from earlier
+ * navigation results.
+ */
+ if (winstate->nav_slot_pos == pos)
+ return slot;
+
+ if (!window_gettupleslot(winobj, pos, slot))
+ {
+ winstate->nav_slot_pos = -1;
+ return winstate->nav_null_slot;
+ }
+
+ winstate->nav_slot_pos = pos;
+ return slot;
+}
+
+
/* -----------------
* ExecEndWindowAgg
* -----------------
@@ -2911,6 +3219,8 @@ ExecReScanWindowAgg(WindowAggState *node)
ExecClearTuple(node->agg_row_slot);
ExecClearTuple(node->temp_slot_1);
ExecClearTuple(node->temp_slot_2);
+ if (node->nav_slot)
+ ExecClearTuple(node->nav_slot);
if (node->framehead_slot)
ExecClearTuple(node->framehead_slot);
if (node->frametail_slot)
@@ -3271,7 +3581,8 @@ window_gettupleslot(WindowObject winobj, int64 pos, TupleTableSlot *slot)
return false;
if (pos < winobj->markpos)
- elog(ERROR, "cannot fetch row before WindowObject's mark position");
+ elog(ERROR, "cannot fetch row: " INT64_FORMAT " before WindowObject's mark position: " INT64_FORMAT,
+ pos, winobj->markpos);
oldcontext = MemoryContextSwitchTo(winstate->ss.ps.ps_ExprContext->ecxt_per_query_memory);
@@ -3389,6 +3700,7 @@ ignorenulls_getfuncarginframe(WindowObject winobj, int argno,
int notnull_offset;
int notnull_relpos;
int forward;
+ int64 num_reduced_frame;
Assert(WindowObjectIsValid(winobj));
winstate = winobj->winstate;
@@ -3417,6 +3729,13 @@ ignorenulls_getfuncarginframe(WindowObject winobj, int argno,
/* rejecting relpos > 0 is easy and simplifies code below */
if (relpos > 0)
goto out_of_frame;
+
+ /*
+ * RPR cares about frame head pos. Need to call
+ * update_frameheadpos
+ */
+ update_frameheadpos(winstate);
+
update_frametailpos(winstate);
abs_pos = winstate->frametailpos - 1;
mark_pos = 0; /* keep compiler quiet */
@@ -3432,6 +3751,35 @@ ignorenulls_getfuncarginframe(WindowObject winobj, int argno,
* Get the next nonnull value in the frame, moving forward or backward
* until we find a value or reach the frame's end.
*/
+
+ /*
+ * Check whether current row is in reduced frame.
+ */
+ num_reduced_frame = row_is_in_reduced_frame(winobj, winstate->frameheadpos);
+ if (num_reduced_frame < 0) /* unmatched or skipped row */
+ goto out_of_frame;
+ else if (num_reduced_frame > 0) /* the first row of the reduced frame */
+ {
+ /*
+ * Early check if row could be out of reduced frame. When RPR is
+ * enabled, EXCLUDE clause cannot be specified and the frame is always
+ * contiguous. So we can safely perform the following checks. Note,
+ * however, it is possible that a row is out of reduced frame if
+ * there's a NULL in the middle. So we need to check it in the
+ * following do loop.
+ */
+ if (seektype == WINDOW_SEEK_HEAD && relpos >= num_reduced_frame)
+ goto out_of_frame;
+ if (seektype == WINDOW_SEEK_TAIL)
+ {
+ if (notnull_relpos >= num_reduced_frame)
+ goto out_of_frame;
+
+ /* not out of reduced frame. Set abspos as a starting point */
+ abs_pos = winstate->frameheadpos + num_reduced_frame - 1;
+ }
+ }
+
do
{
int inframe;
@@ -3493,6 +3841,16 @@ ignorenulls_getfuncarginframe(WindowObject winobj, int argno,
}
advance:
abs_pos += forward;
+ if (rpr_is_defined(winstate))
+ {
+ /*
+ * Check whether we are still in the reduced frame. (also check
+ * if we succeeded in getting the target row).
+ */
+ num_reduced_frame--;
+ if (num_reduced_frame <= 0 && notnull_offset <= notnull_relpos)
+ goto out_of_frame;
+ }
} while (notnull_offset <= notnull_relpos);
if (set_mark)
@@ -3635,87 +3993,940 @@ put_notnull_info(WindowObject winobj, int64 pos, int argno, bool isnull)
mbp[bpos] = mb;
}
-/***********************************************************************
- * API exposed to window functions
- ***********************************************************************/
+/*
+ * eval_nav_offset_helper
+ * Evaluate an offset expression at executor init time for trim
+ * optimization. Returns the offset value, or 0 for NULL/negative
+ * (these will cause a runtime error during actual navigation, so the
+ * trim value is irrelevant).
+ */
+static int64
+eval_nav_offset_helper(WindowAggState *winstate, Expr *offset_expr,
+ int64 defaultOffset)
+{
+ ExprContext *econtext = winstate->ss.ps.ps_ExprContext;
+ ExprState *estate;
+ Datum val;
+ bool isnull;
+ int64 offset;
+
+ if (offset_expr == NULL)
+ return defaultOffset;
+
+ estate = ExecInitExpr(offset_expr, (PlanState *) winstate);
+ val = ExecEvalExprSwitchContext(estate, econtext, &isnull);
+ if (isnull)
+ return 0;
+
+ offset = DatumGetInt64(val);
+ if (offset < 0)
+ return 0;
+
+ return offset;
+}
+
+typedef struct
+{
+ WindowAggState *winstate;
+ int64 maxOffset; /* max backward-reach offset across all nav
+ * exprs */
+ bool maxOverflow; /* true if backward-reach overflow detected */
+ int64 minFirstOffset; /* min forward-from-match_start offset; may be
+ * negative (PREV_FIRST: inner - outer < 0) */
+} EvalDefineOffsetsContext;
/*
- * WinCheckAndInitializeNullTreatment
- * Check null treatment clause and sets ignore_nulls
+ * visit_nav_exec
+ * nav_traversal_walker callback (NavVisitFn) for the executor side.
+ * At each RPRNavExpr, evaluates the nav's offset expression(s) at
+ * runtime via eval_nav_offset_helper and accumulates:
*
- * Window functions should call this to check if they are being called with
- * a null treatment clause when they don't allow it, or to set ignore_nulls.
+ * - maxOffset (backward reach): PREV, LAST-with-offset, compound
+ * PREV_LAST (sets maxOverflow on int64 overflow), compound
+ * NEXT_LAST (= max(inner - outer, 0))
+ * - minFirstOffset (forward reach from match_start): FIRST,
+ * compound PREV_FIRST (= inner - outer, may be negative),
+ * compound NEXT_FIRST (= inner + outer, clamped to PG_INT64_MAX on
+ * overflow; always >= 0 so never updates minFirstOffset in practice)
+ *
+ * Counterpart of visit_nav_plan but using runtime evaluation instead of
+ * Const folding; runs only for offsets the planner marked NEEDS_EVAL.
+ * Match-start dependency is not recomputed here -- the planner's bitmapset
+ * is reused via winstate->defineMatchStartDependent.
*/
-void
-WinCheckAndInitializeNullTreatment(WindowObject winobj,
- bool allowNullTreatment,
- FunctionCallInfo fcinfo)
+static void
+visit_nav_exec(NavTraversal *t, RPRNavExpr *nav)
{
- Assert(WindowObjectIsValid(winobj));
- if (winobj->ignore_nulls != NO_NULLTREATMENT && !allowNullTreatment)
+ EvalDefineOffsetsContext *context = (EvalDefineOffsetsContext *) t->data;
+
+ /*
+ * Parser guarantee (mirrors visit_nav_plan): nav's direct children are
+ * never RPRNavExpr -- compound nesting is flattened in place and any
+ * other nesting is rejected. Outer-kind dispatch is sufficient.
+ */
+ Assert(nav->arg == NULL || !IsA(nav->arg, RPRNavExpr));
+ Assert(nav->offset_arg == NULL || !IsA(nav->offset_arg, RPRNavExpr));
+ Assert(nav->compound_offset_arg == NULL ||
+ !IsA(nav->compound_offset_arg, RPRNavExpr));
+
+ /* Backward reach: PREV, LAST-with-offset */
+ if (!context->maxOverflow)
{
- const char *funcname = get_func_name(fcinfo->flinfo->fn_oid);
+ int64 reach = 0;
+ bool gotReach = false;
- if (!funcname)
- elog(ERROR, "could not get function name");
- ereport(ERROR,
- (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
- errmsg("function %s does not allow RESPECT/IGNORE NULLS",
- funcname)));
+ if (nav->kind == RPR_NAV_PREV)
+ {
+ reach = eval_nav_offset_helper(context->winstate,
+ nav->offset_arg, 1);
+ gotReach = true;
+ }
+ else if (nav->kind == RPR_NAV_LAST && nav->offset_arg != NULL)
+ {
+ reach = eval_nav_offset_helper(context->winstate,
+ nav->offset_arg, 0);
+ gotReach = true;
+ }
+ else if (nav->kind == RPR_NAV_PREV_LAST ||
+ nav->kind == RPR_NAV_NEXT_LAST)
+ {
+ int64 inner = eval_nav_offset_helper(context->winstate,
+ nav->offset_arg, 0);
+ int64 outer = eval_nav_offset_helper(context->winstate,
+ nav->compound_offset_arg, 1);
+
+ if (nav->kind == RPR_NAV_PREV_LAST)
+ {
+ if (pg_add_s64_overflow(inner, outer, &reach))
+ context->maxOverflow = true;
+ else
+ gotReach = true;
+ }
+ else
+ {
+ reach = Max(inner - outer, 0);
+ gotReach = true;
+ }
+ }
+
+ if (gotReach)
+ context->maxOffset = Max(context->maxOffset, reach);
+ }
+
+ /* Forward reach from match_start: FIRST, compound PREV_FIRST/NEXT_FIRST */
+ if (nav->kind == RPR_NAV_FIRST)
+ {
+ int64 reach;
+
+ reach = eval_nav_offset_helper(context->winstate,
+ nav->offset_arg, 0);
+ context->minFirstOffset = Min(context->minFirstOffset, reach);
+ }
+ else if (nav->kind == RPR_NAV_PREV_FIRST ||
+ nav->kind == RPR_NAV_NEXT_FIRST)
+ {
+ int64 inner = eval_nav_offset_helper(context->winstate,
+ nav->offset_arg, 0);
+ int64 outer = eval_nav_offset_helper(context->winstate,
+ nav->compound_offset_arg, 1);
+ int64 reach;
+
+ if (nav->kind == RPR_NAV_PREV_FIRST)
+ {
+ /*
+ * reach = inner - outer. Both are non-negative, so the result >=
+ * -PG_INT64_MAX, which cannot underflow int64.
+ */
+ reach = inner - outer;
+ }
+ else
+ {
+ /*
+ * NEXT_FIRST: reach = inner + outer. This can overflow, but the
+ * result is always >= 0, so it never updates minFirstOffset
+ * (which tracks the minimum). Clamp to PG_INT64_MAX on overflow.
+ */
+ if (pg_add_s64_overflow(inner, outer, &reach))
+ reach = PG_INT64_MAX;
+ }
+ context->minFirstOffset = Min(context->minFirstOffset, reach);
}
- else if (winobj->ignore_nulls == PARSER_IGNORE_NULLS)
- winobj->ignore_nulls = IGNORE_NULLS;
}
/*
- * WinGetPartitionLocalMemory
- * Get working memory that lives till end of partition processing
+ * eval_define_offsets
+ * Evaluate non-constant nav offsets at executor init time.
*
- * On first call within a given partition, this allocates and zeroes the
- * requested amount of space. Subsequent calls just return the same chunk.
+ * Called when the planner set navMaxOffsetKind and/or navFirstOffsetKind
+ * to RPR_NAV_OFFSET_NEEDS_EVAL because some offset contains a parameter
+ * or non-foldable expression. Updates only the fields whose kind was
+ * NEEDS_EVAL; FIXED kinds are left unchanged.
*
- * Memory obtained this way is normally used to hold state that should be
- * automatically reset for each new partition. If a window function wants
- * to hold state across the whole query, fcinfo->fn_extra can be used in the
- * usual way for that.
+ * On backward-reach overflow, sets navMaxOffsetKind to
+ * RPR_NAV_OFFSET_RETAIN_ALL so that tuplestore trim is disabled for
+ * backward navigation.
*/
-void *
-WinGetPartitionLocalMemory(WindowObject winobj, Size sz)
+static void
+eval_define_offsets(WindowAggState *winstate, List *defineClause)
{
- Assert(WindowObjectIsValid(winobj));
- if (winobj->localmem == NULL)
- winobj->localmem =
- MemoryContextAllocZero(winobj->winstate->partcontext, sz);
- return winobj->localmem;
+ EvalDefineOffsetsContext ctx;
+ NavTraversal trav;
+ bool needsMax = (winstate->navMaxOffsetKind == RPR_NAV_OFFSET_NEEDS_EVAL);
+ bool needsFirst = (winstate->hasFirstNav &&
+ winstate->navFirstOffsetKind == RPR_NAV_OFFSET_NEEDS_EVAL);
+
+ if (!needsMax && !needsFirst)
+ return;
+
+ ctx.winstate = winstate;
+ ctx.maxOffset = 0;
+ ctx.maxOverflow = false;
+ ctx.minFirstOffset = PG_INT64_MAX;
+
+ trav.visit = visit_nav_exec;
+ trav.data = &ctx;
+
+ foreach_node(TargetEntry, te, defineClause)
+ {
+ nav_traversal_walker((Node *) te->expr, &trav);
+ }
+
+ if (needsMax)
+ {
+ if (ctx.maxOverflow)
+ {
+ winstate->navMaxOffsetKind = RPR_NAV_OFFSET_RETAIN_ALL;
+ winstate->navMaxOffset = 0;
+ }
+ else
+ {
+ winstate->navMaxOffsetKind = RPR_NAV_OFFSET_FIXED;
+ winstate->navMaxOffset = ctx.maxOffset;
+ }
+ }
+
+ if (needsFirst)
+ {
+ winstate->navFirstOffsetKind = RPR_NAV_OFFSET_FIXED;
+ if (ctx.minFirstOffset < PG_INT64_MAX)
+ winstate->navFirstOffset = ctx.minFirstOffset;
+ else
+ winstate->navFirstOffset = PG_INT64_MAX;
+ }
}
/*
- * WinGetCurrentPosition
- * Return the current row's position (counting from 0) within the current
- * partition.
+ * rpr_is_defined
+ * Return true if row pattern recognition is defined.
*/
-int64
-WinGetCurrentPosition(WindowObject winobj)
+static bool
+rpr_is_defined(WindowAggState *winstate)
{
- Assert(WindowObjectIsValid(winobj));
- return winobj->winstate->currentpos;
+ return winstate->rpPattern != NULL;
}
/*
- * WinGetPartitionRowCount
- * Return total number of rows contained in the current partition.
+ * -----------------
+ * row_is_in_reduced_frame
+ * Determine whether a row is in the current row's reduced window frame
+ * according to row pattern matching
*
- * Note: this is a relatively expensive operation because it forces the
- * whole partition to be "spooled" into the tuplestore at once. Once
- * executed, however, additional calls within the same partition are cheap.
+ * The row must have already been determined to be in a full window frame
+ * and fetched into the slot.
+ *
+ * Returns:
+ * = 0, RPR is not defined.
+ * >0, if the row is the first in the reduced frame. Return the number of rows
+ * in the reduced frame.
+ * -1, if the row is an unmatched row
+ * -2, if the row is in the reduced frame but needed to be skipped because of
+ * AFTER MATCH SKIP PAST LAST ROW
+ * -----------------
*/
-int64
-WinGetPartitionRowCount(WindowObject winobj)
+static int64
+row_is_in_reduced_frame(WindowObject winobj, int64 pos)
{
- Assert(WindowObjectIsValid(winobj));
- spool_tuples(winobj->winstate, -1);
- return winobj->winstate->spooled_rows;
+ WindowAggState *winstate = winobj->winstate;
+ int state;
+ int64 rtn;
+
+ if (!rpr_is_defined(winstate))
+ {
+ /*
+ * RPR is not defined. Assume that we are always in the reduced window
+ * frame.
+ */
+ rtn = 0;
+ return rtn;
+ }
+
+ state = get_reduced_frame_status(winstate, pos);
+
+ if (state == RF_NOT_DETERMINED)
+ {
+ update_frameheadpos(winstate);
+ update_reduced_frame(winobj, pos);
+ }
+
+ state = get_reduced_frame_status(winstate, pos);
+
+ switch (state)
+ {
+ case RF_FRAME_HEAD:
+ rtn = winstate->rpr_match_length;
+ break;
+
+ case RF_SKIPPED:
+ rtn = -2;
+ break;
+
+ case RF_UNMATCHED:
+ case RF_EMPTY_MATCH:
+ rtn = -1;
+ break;
+
+ default:
+ elog(ERROR, "unrecognized state: %d at: " INT64_FORMAT,
+ state, pos);
+ break;
+ }
+
+ return rtn;
+}
+
+/*
+ * clear_reduced_frame
+ * Clear reduced frame status
+ */
+static void
+clear_reduced_frame(WindowAggState *winstate)
+{
+ winstate->rpr_match_valid = false;
+ winstate->rpr_match_matched = false;
+ winstate->rpr_match_start = -1;
+ winstate->rpr_match_length = 0;
+}
+
+/*
+ * get_reduced_frame_status
+ * Look up a position against the current match.
+ *
+ * Returns one of the RF_* constants:
+ * RF_NOT_DETERMINED pos has not been processed yet
+ * RF_FRAME_HEAD pos is the start of the current match
+ * RF_SKIPPED pos is inside the current match but not the start
+ * RF_UNMATCHED pos is processed but not part of any match
+ * RF_EMPTY_MATCH pos is the start of an empty (zero-length) match
+ *
+ * update_reduced_frame() records the current match as exactly one of three
+ * (rpr_match_matched, rpr_match_length) shapes: (false, 1) for unmatched,
+ * (true, 0) for an empty match, and (true, >= 1) for a real match. The
+ * tests below form a cascade with early returns: each is a minimal check
+ * that relies on the negations the preceding returns have already
+ * established, so their order is significant. The "by here" notes spell
+ * out the running invariant; reordering a test would misclassify one of
+ * the three shapes.
+ */
+static int
+get_reduced_frame_status(WindowAggState *winstate, int64 pos)
+{
+ int64 start = winstate->rpr_match_start;
+ int64 length = winstate->rpr_match_length;
+
+ if (!winstate->rpr_match_valid)
+ return RF_NOT_DETERMINED;
+
+ /*
+ * By here the record is valid and holds one of the three shapes above.
+ *
+ * The empty match (true, 0) must be classified first: it has length 0, so
+ * the range test below would compute start + length == start and reject
+ * its own start position as out of range.
+ */
+ if (pos == start && winstate->rpr_match_matched && length == 0)
+ return RF_EMPTY_MATCH;
+
+ /*
+ * By here length >= 1 -- the only zero-length record, the empty match,
+ * has been handled -- so [start, start + length) is a well-formed range.
+ */
+ if (pos < start || pos >= start + length)
+ return RF_NOT_DETERMINED;
+
+ /*
+ * By here pos lies within [start, start + length). An unmatched record
+ * is (false, 1), so this returns for its single in-range position.
+ */
+ if (!winstate->rpr_match_matched)
+ return RF_UNMATCHED;
+
+ /* By here the match is real (true, >= 1) and pos is one of its rows. */
+ if (pos == start)
+ return RF_FRAME_HEAD;
+
+ return RF_SKIPPED;
+}
+
+/*
+ * update_reduced_frame
+ * Update reduced frame info using multi-context NFA pattern matching.
+ *
+ * Maintains multiple NFA contexts simultaneously, one for each potential
+ * match start position. This allows sharing row evaluations across contexts,
+ * avoiding redundant DEFINE clause evaluations when rewinding for SKIP TO
+ * NEXT ROW mode.
+ *
+ * Key optimizations:
+ * - Row evaluations (expensive DEFINE clauses) happen only once per row
+ * - All active contexts share the same evaluation results
+ * - Contexts persist across calls, enabling O(n) DEFINE evaluations
+ */
+static void
+update_reduced_frame(WindowObject winobj, int64 pos)
+{
+ WindowAggState *winstate = winobj->winstate;
+ RPRNFAContext *targetCtx;
+ int64 currentPos;
+ int64 startPos;
+ int frameOptions = winstate->frameOptions;
+ bool hasLimitedFrame;
+ int64 frameOffset = 0;
+ int64 matchLen;
+
+ /*
+ * Check if we have a limited frame (ROWS ... N FOLLOWING). Each context
+ * needs its own frame end based on matchStartRow + offset.
+ */
+ hasLimitedFrame = (frameOptions & FRAMEOPTION_ROWS) &&
+ !(frameOptions & FRAMEOPTION_END_UNBOUNDED_FOLLOWING);
+ if (hasLimitedFrame)
+ frameOffset = DatumGetInt64(winstate->endOffsetValue);
+
+ /*
+ * Case 1: pos is before any existing context's start position. This means
+ * the position was already processed and determined unmatched. Head is
+ * the oldest context (lowest matchStartRow) since contexts are added at
+ * tail with increasing positions.
+ */
+ if (winstate->nfaContext != NULL &&
+ pos < winstate->nfaContext->matchStartRow)
+ {
+ /* already processed, unmatched */
+ winstate->rpr_match_valid = true;
+ winstate->rpr_match_matched = false;
+ winstate->rpr_match_start = pos;
+ winstate->rpr_match_length = 1;
+ return;
+ }
+
+ /*
+ * Case 2: Find existing context for this pos, or create new one.
+ */
+ targetCtx = ExecRPRGetHeadContext(winstate, pos);
+ if (targetCtx == NULL)
+ {
+ /*
+ * No context exists. If pos is already processed, it means this row
+ * was already determined to be unmatched or skipped - no need to
+ * reprocess.
+ */
+ if (pos <= winstate->nfaLastProcessedRow)
+ {
+ /* already processed, unmatched */
+ winstate->rpr_match_valid = true;
+ winstate->rpr_match_matched = false;
+ winstate->rpr_match_start = pos;
+ winstate->rpr_match_length = 1;
+ return;
+ }
+ /* Not yet processed - create new context and start fresh */
+ targetCtx = ExecRPRStartContext(winstate, pos);
+ }
+ else if (targetCtx->states == NULL)
+ {
+ /*
+ * The head context already completed in an earlier call. Reachable
+ * under SKIP TO NEXT ROW, where overlapping contexts let one reach
+ * FIN -- recording its result -- before the call for its own start
+ * row arrives. Register that result.
+ */
+ goto register_result;
+ }
+
+ /*
+ * Determine where to start processing. Usually nfaLastProcessedRow+1 >=
+ * pos since contexts are created at currentPos+1 during processing.
+ * However, pos can exceed this when rows are skipped (e.g., unmatched
+ * rows don't update nfaLastProcessedRow).
+ */
+ startPos = Max(pos, winstate->nfaLastProcessedRow + 1);
+
+ /*
+ * Process rows until target context completes or we hit boundaries. Each
+ * row evaluation is shared across all active contexts.
+ */
+ for (currentPos = startPos; targetCtx->states != NULL; currentPos++)
+ {
+ bool rowExists;
+
+ /*
+ * Evaluate variables for this row - done only once, shared by all
+ * contexts.
+ *
+ * Set nav_match_start to the head context's matchStartRow for
+ * FIRST/LAST navigation. Match_start-dependent variables (FIRST,
+ * LAST-with-offset) are re-evaluated per-context in ExecRPRProcessRow
+ * when matchStartRow differs.
+ */
+ winstate->nav_match_start = targetCtx->matchStartRow;
+ rowExists = nfa_evaluate_row(winobj, currentPos, winstate->nfaVarMatched);
+
+ /* No more rows in partition? Finalize all contexts */
+ if (!rowExists)
+ {
+ ExecRPRFinalizeAllContexts(winstate, currentPos - 1);
+ /* Clean up dead contexts from finalization */
+ ExecRPRCleanupDeadContexts(winstate, targetCtx);
+ break;
+ }
+
+ /* Update last processed row */
+ winstate->nfaLastProcessedRow = currentPos;
+
+ /*--------------------------
+ * Process all contexts for this row:
+ * 1. Match all (convergence)
+ * 2. Absorb redundant
+ * 3. Advance all (divergence)
+ */
+ ExecRPRProcessRow(winstate, currentPos, hasLimitedFrame, frameOffset);
+
+ /*
+ * Create a new context for the next potential start position. This
+ * enables overlapping match detection for SKIP TO NEXT ROW.
+ */
+ ExecRPRStartContext(winstate, currentPos + 1);
+
+ /*
+ * Clean up dead contexts (failed with no active states and no match).
+ * This removes contexts that failed during processing and counts them
+ * appropriately as pruned or mismatched.
+ */
+ ExecRPRCleanupDeadContexts(winstate, targetCtx);
+ }
+
+register_result:
+ Assert(pos == targetCtx->matchStartRow);
+
+ /*
+ * Record match result.
+ */
+ winstate->rpr_match_valid = true;
+ winstate->rpr_match_start = targetCtx->matchStartRow;
+
+ if (targetCtx->matchEndRow < targetCtx->matchStartRow)
+ {
+ matchLen = targetCtx->lastProcessedRow - targetCtx->matchStartRow + 1;
+
+ if (targetCtx->matchedState != NULL)
+ {
+ /* Empty match: FIN reached but 0 rows consumed */
+ winstate->rpr_match_matched = true;
+ winstate->rpr_match_length = 0;
+ ExecRPRRecordContextSuccess(winstate, 0);
+ }
+ else
+ {
+ /* No match */
+ winstate->rpr_match_matched = false;
+ winstate->rpr_match_length = 1;
+ ExecRPRRecordContextFailure(winstate, matchLen);
+ }
+ ExecRPRFreeContext(winstate, targetCtx);
+ return;
+ }
+
+ /* Match succeeded */
+ matchLen = targetCtx->matchEndRow - targetCtx->matchStartRow + 1;
+
+ winstate->rpr_match_matched = true;
+ winstate->rpr_match_length = matchLen;
+ ExecRPRRecordContextSuccess(winstate, matchLen);
+
+ /* Remove the matched context */
+ ExecRPRFreeContext(winstate, targetCtx);
+}
+
+/*
+ * nfa_evaluate_row
+ *
+ * Evaluate all DEFINE variables for current row.
+ * Returns true if the row exists, false if out of partition.
+ * If row exists, fills varMatched array.
+ * varMatched[i] = true if variable i matched at current row.
+ *
+ * Uses 1-slot model: only ecxt_outertuple is set to the current row.
+ * PREV/NEXT/FIRST/LAST navigation is handled by EEOP_RPR_NAV_SET/RESTORE
+ * opcodes during expression evaluation, which temporarily swap the slot.
+ */
+static bool
+nfa_evaluate_row(WindowObject winobj, int64 pos, bool *varMatched)
+{
+ WindowAggState *winstate = winobj->winstate;
+ ExprContext *econtext = winstate->ss.ps.ps_ExprContext;
+ int numDefineVars = list_length(winstate->defineVariableList);
+ int varIdx = 0;
+ TupleTableSlot *slot;
+ int64 saved_pos;
+
+ /* Fetch current row into temp_slot_1 */
+ slot = winstate->temp_slot_1;
+ if (!window_gettupleslot(winobj, pos, slot))
+ return false; /* No row exists */
+
+ /* Set up 1-slot context: only ecxt_outertuple */
+ econtext->ecxt_outertuple = slot;
+
+ /*
+ * Save and set currentpos so that EEOP_RPR_NAV_SET opcodes can calculate
+ * target positions (currentpos +/- offset).
+ */
+ saved_pos = winstate->currentpos;
+ winstate->currentpos = pos;
+
+ /* Invalidate nav_slot cache so PREV/NEXT re-fetch for new row */
+ winstate->nav_slot_pos = -1;
+
+ foreach_ptr(ExprState, exprState, winstate->defineClauseList)
+ {
+ Datum result;
+ bool isnull;
+
+ /* Evaluate DEFINE expression */
+ result = ExecEvalExpr(exprState, econtext, &isnull);
+
+ varMatched[varIdx] = (!isnull && DatumGetBool(result));
+
+ varIdx++;
+ if (varIdx >= numDefineVars)
+ break;
+ }
+
+ winstate->currentpos = saved_pos;
+
+ return true; /* Row exists */
+}
+
+/*
+ * WinGetSlotInFrame
+ * slot: TupleTableSlot to store the result
+ * relpos: signed rowcount offset from the seek position
+ * seektype: WINDOW_SEEK_HEAD or WINDOW_SEEK_TAIL
+ * set_mark: If the row is found/in frame and set_mark is true, the mark is
+ * moved to the row as a side-effect.
+ * isnull: output argument, receives isnull status of result
+ * isout: output argument, set to indicate whether target row position
+ * is out of frame (can pass NULL if caller doesn't care about this)
+ *
+ * Returns 0 if we successfully got the slot, or nonzero if out of frame.
+ * (isout is also set in the latter case.)
+ */
+static int
+WinGetSlotInFrame(WindowObject winobj, TupleTableSlot *slot,
+ int relpos, int seektype, bool set_mark,
+ bool *isnull, bool *isout)
+{
+ WindowAggState *winstate;
+ int64 abs_pos;
+ int64 mark_pos;
+ int64 num_reduced_frame;
+
+ Assert(WindowObjectIsValid(winobj));
+ winstate = winobj->winstate;
+
+ switch (seektype)
+ {
+ case WINDOW_SEEK_CURRENT:
+ elog(ERROR, "WINDOW_SEEK_CURRENT is not supported for WinGetFuncArgInFrame");
+ abs_pos = mark_pos = 0; /* keep compiler quiet */
+ break;
+ case WINDOW_SEEK_HEAD:
+ /* rejecting relpos < 0 is easy and simplifies code below */
+ if (relpos < 0)
+ goto out_of_frame;
+ update_frameheadpos(winstate);
+ abs_pos = winstate->frameheadpos + relpos;
+ mark_pos = abs_pos;
+
+ /*
+ * Account for exclusion option if one is active, but advance only
+ * abs_pos not mark_pos. This prevents changes of the current
+ * row's peer group from resulting in trying to fetch a row before
+ * some previous mark position.
+ *
+ * Note that in some corner cases such as current row being
+ * outside frame, these calculations are theoretically too simple,
+ * but it doesn't matter because we'll end up deciding the row is
+ * out of frame. We do not attempt to avoid fetching rows past
+ * end of frame; that would happen in some cases anyway.
+ */
+ switch (winstate->frameOptions & FRAMEOPTION_EXCLUSION)
+ {
+ case 0:
+ /* no adjustment needed */
+ break;
+ case FRAMEOPTION_EXCLUDE_CURRENT_ROW:
+ if (abs_pos >= winstate->currentpos &&
+ winstate->currentpos >= winstate->frameheadpos)
+ abs_pos++;
+ break;
+ case FRAMEOPTION_EXCLUDE_GROUP:
+ update_grouptailpos(winstate);
+ if (abs_pos >= winstate->groupheadpos &&
+ winstate->grouptailpos > winstate->frameheadpos)
+ {
+ int64 overlapstart = Max(winstate->groupheadpos,
+ winstate->frameheadpos);
+
+ abs_pos += winstate->grouptailpos - overlapstart;
+ }
+ break;
+ case FRAMEOPTION_EXCLUDE_TIES:
+ update_grouptailpos(winstate);
+ if (abs_pos >= winstate->groupheadpos &&
+ winstate->grouptailpos > winstate->frameheadpos)
+ {
+ int64 overlapstart = Max(winstate->groupheadpos,
+ winstate->frameheadpos);
+
+ if (abs_pos == overlapstart)
+ abs_pos = winstate->currentpos;
+ else
+ abs_pos += winstate->grouptailpos - overlapstart - 1;
+ }
+ break;
+ default:
+ elog(ERROR, "unrecognized frame option state: 0x%x",
+ winstate->frameOptions);
+ break;
+ }
+ num_reduced_frame = row_is_in_reduced_frame(winobj,
+ winstate->frameheadpos);
+ if (num_reduced_frame < 0)
+ goto out_of_frame;
+ else if (num_reduced_frame > 0)
+ if (relpos >= num_reduced_frame)
+ goto out_of_frame;
+ break;
+ case WINDOW_SEEK_TAIL:
+ /* rejecting relpos > 0 is easy and simplifies code below */
+ if (relpos > 0)
+ goto out_of_frame;
+
+ /*
+ * RPR cares about frame head pos. Need to call
+ * update_frameheadpos
+ */
+ update_frameheadpos(winstate);
+
+ update_frametailpos(winstate);
+ abs_pos = winstate->frametailpos - 1 + relpos;
+
+ /*
+ * Account for exclusion option if one is active. If there is no
+ * exclusion, we can safely set the mark at the accessed row. But
+ * if there is, we can only mark the frame start, because we can't
+ * be sure how far back in the frame the exclusion might cause us
+ * to fetch in future. Furthermore, we have to actually check
+ * against frameheadpos here, since it's unsafe to try to fetch a
+ * row before frame start if the mark might be there already.
+ */
+ switch (winstate->frameOptions & FRAMEOPTION_EXCLUSION)
+ {
+ case 0:
+ /* no adjustment needed */
+ mark_pos = abs_pos;
+ break;
+ case FRAMEOPTION_EXCLUDE_CURRENT_ROW:
+ if (abs_pos <= winstate->currentpos &&
+ winstate->currentpos < winstate->frametailpos)
+ abs_pos--;
+ update_frameheadpos(winstate);
+ if (abs_pos < winstate->frameheadpos)
+ goto out_of_frame;
+ mark_pos = winstate->frameheadpos;
+ break;
+ case FRAMEOPTION_EXCLUDE_GROUP:
+ update_grouptailpos(winstate);
+ if (abs_pos < winstate->grouptailpos &&
+ winstate->groupheadpos < winstate->frametailpos)
+ {
+ int64 overlapend = Min(winstate->grouptailpos,
+ winstate->frametailpos);
+
+ abs_pos -= overlapend - winstate->groupheadpos;
+ }
+ update_frameheadpos(winstate);
+ if (abs_pos < winstate->frameheadpos)
+ goto out_of_frame;
+ mark_pos = winstate->frameheadpos;
+ break;
+ case FRAMEOPTION_EXCLUDE_TIES:
+ update_grouptailpos(winstate);
+ if (abs_pos < winstate->grouptailpos &&
+ winstate->groupheadpos < winstate->frametailpos)
+ {
+ int64 overlapend = Min(winstate->grouptailpos,
+ winstate->frametailpos);
+
+ if (abs_pos == overlapend - 1)
+ abs_pos = winstate->currentpos;
+ else
+ abs_pos -= overlapend - 1 - winstate->groupheadpos;
+ }
+ update_frameheadpos(winstate);
+ if (abs_pos < winstate->frameheadpos)
+ goto out_of_frame;
+ mark_pos = winstate->frameheadpos;
+ break;
+ default:
+ elog(ERROR, "unrecognized frame option state: 0x%x",
+ winstate->frameOptions);
+ mark_pos = 0; /* keep compiler quiet */
+ break;
+ }
+
+ num_reduced_frame = row_is_in_reduced_frame(winobj,
+ winstate->frameheadpos);
+ if (num_reduced_frame < 0)
+ goto out_of_frame;
+ else if (num_reduced_frame > 0)
+ {
+ if (-relpos >= num_reduced_frame)
+ goto out_of_frame;
+ abs_pos = winstate->frameheadpos + relpos +
+ num_reduced_frame - 1;
+ }
+ break;
+ default:
+ elog(ERROR, "unrecognized window seek type: %d", seektype);
+ abs_pos = mark_pos = 0; /* keep compiler quiet */
+ break;
+ }
+
+ if (!window_gettupleslot(winobj, abs_pos, slot))
+ goto out_of_frame;
+
+ /* The code above does not detect all out-of-frame cases, so check */
+ if (row_is_in_frame(winobj, abs_pos, slot, false) <= 0)
+ goto out_of_frame;
+
+ if (isout)
+ *isout = false;
+ if (set_mark)
+ {
+ /*
+ * If RPR is enabled and seek type is WINDOW_SEEK_TAIL, we set the
+ * mark position unconditionally to frameheadpos. In this case the
+ * frame always starts at CURRENT_ROW and never goes back, thus
+ * setting the mark at the position is safe.
+ */
+ if (winstate->rpPattern != NULL && seektype == WINDOW_SEEK_TAIL)
+ mark_pos = winstate->frameheadpos;
+ WinSetMarkPosition(winobj, mark_pos);
+ }
+ return 0;
+
+out_of_frame:
+ if (isout)
+ *isout = true;
+ *isnull = true;
+ return -1;
+}
+
+
+/***********************************************************************
+ * API exposed to window functions
+ ***********************************************************************/
+
+
+/*
+ * WinCheckAndInitializeNullTreatment
+ * Check null treatment clause and sets ignore_nulls
+ *
+ * Window functions should call this to check if they are being called with
+ * a null treatment clause when they don't allow it, or to set ignore_nulls.
+ */
+void
+WinCheckAndInitializeNullTreatment(WindowObject winobj,
+ bool allowNullTreatment,
+ FunctionCallInfo fcinfo)
+{
+ Assert(WindowObjectIsValid(winobj));
+ if (winobj->ignore_nulls != NO_NULLTREATMENT && !allowNullTreatment)
+ {
+ const char *funcname = get_func_name(fcinfo->flinfo->fn_oid);
+
+ if (!funcname)
+ elog(ERROR, "could not get function name");
+ ereport(ERROR,
+ (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+ errmsg("function %s does not allow RESPECT/IGNORE NULLS",
+ funcname)));
+ }
+ else if (winobj->ignore_nulls == PARSER_IGNORE_NULLS)
+ winobj->ignore_nulls = IGNORE_NULLS;
+}
+
+/*
+ * WinGetPartitionLocalMemory
+ * Get working memory that lives till end of partition processing
+ *
+ * On first call within a given partition, this allocates and zeroes the
+ * requested amount of space. Subsequent calls just return the same chunk.
+ *
+ * Memory obtained this way is normally used to hold state that should be
+ * automatically reset for each new partition. If a window function wants
+ * to hold state across the whole query, fcinfo->fn_extra can be used in the
+ * usual way for that.
+ */
+void *
+WinGetPartitionLocalMemory(WindowObject winobj, Size sz)
+{
+ Assert(WindowObjectIsValid(winobj));
+ if (winobj->localmem == NULL)
+ winobj->localmem =
+ MemoryContextAllocZero(winobj->winstate->partcontext, sz);
+ return winobj->localmem;
+}
+
+/*
+ * WinGetCurrentPosition
+ * Return the current row's position (counting from 0) within the current
+ * partition.
+ */
+int64
+WinGetCurrentPosition(WindowObject winobj)
+{
+ Assert(WindowObjectIsValid(winobj));
+ return winobj->winstate->currentpos;
+}
+
+/*
+ * WinGetPartitionRowCount
+ * Return total number of rows contained in the current partition.
+ *
+ * Note: this is a relatively expensive operation because it forces the
+ * whole partition to be "spooled" into the tuplestore at once. Once
+ * executed, however, additional calls within the same partition are cheap.
+ */
+int64
+WinGetPartitionRowCount(WindowObject winobj)
+{
+ Assert(WindowObjectIsValid(winobj));
+ spool_tuples(winobj->winstate, -1);
+ return winobj->winstate->spooled_rows;
}
/*
@@ -4001,8 +5212,6 @@ WinGetFuncArgInFrame(WindowObject winobj, int argno,
WindowAggState *winstate;
ExprContext *econtext;
TupleTableSlot *slot;
- int64 abs_pos;
- int64 mark_pos;
Assert(WindowObjectIsValid(winobj));
winstate = winobj->winstate;
@@ -4013,166 +5222,15 @@ WinGetFuncArgInFrame(WindowObject winobj, int argno,
return ignorenulls_getfuncarginframe(winobj, argno, relpos, seektype,
set_mark, isnull, isout);
- switch (seektype)
+ if (WinGetSlotInFrame(winobj, slot,
+ relpos, seektype, set_mark,
+ isnull, isout) == 0)
{
- case WINDOW_SEEK_CURRENT:
- elog(ERROR, "WINDOW_SEEK_CURRENT is not supported for WinGetFuncArgInFrame");
- abs_pos = mark_pos = 0; /* keep compiler quiet */
- break;
- case WINDOW_SEEK_HEAD:
- /* rejecting relpos < 0 is easy and simplifies code below */
- if (relpos < 0)
- goto out_of_frame;
- update_frameheadpos(winstate);
- abs_pos = winstate->frameheadpos + relpos;
- mark_pos = abs_pos;
-
- /*
- * Account for exclusion option if one is active, but advance only
- * abs_pos not mark_pos. This prevents changes of the current
- * row's peer group from resulting in trying to fetch a row before
- * some previous mark position.
- *
- * Note that in some corner cases such as current row being
- * outside frame, these calculations are theoretically too simple,
- * but it doesn't matter because we'll end up deciding the row is
- * out of frame. We do not attempt to avoid fetching rows past
- * end of frame; that would happen in some cases anyway.
- */
- switch (winstate->frameOptions & FRAMEOPTION_EXCLUSION)
- {
- case 0:
- /* no adjustment needed */
- break;
- case FRAMEOPTION_EXCLUDE_CURRENT_ROW:
- if (abs_pos >= winstate->currentpos &&
- winstate->currentpos >= winstate->frameheadpos)
- abs_pos++;
- break;
- case FRAMEOPTION_EXCLUDE_GROUP:
- update_grouptailpos(winstate);
- if (abs_pos >= winstate->groupheadpos &&
- winstate->grouptailpos > winstate->frameheadpos)
- {
- int64 overlapstart = Max(winstate->groupheadpos,
- winstate->frameheadpos);
-
- abs_pos += winstate->grouptailpos - overlapstart;
- }
- break;
- case FRAMEOPTION_EXCLUDE_TIES:
- update_grouptailpos(winstate);
- if (abs_pos >= winstate->groupheadpos &&
- winstate->grouptailpos > winstate->frameheadpos)
- {
- int64 overlapstart = Max(winstate->groupheadpos,
- winstate->frameheadpos);
-
- if (abs_pos == overlapstart)
- abs_pos = winstate->currentpos;
- else
- abs_pos += winstate->grouptailpos - overlapstart - 1;
- }
- break;
- default:
- elog(ERROR, "unrecognized frame option state: 0x%x",
- winstate->frameOptions);
- break;
- }
- break;
- case WINDOW_SEEK_TAIL:
- /* rejecting relpos > 0 is easy and simplifies code below */
- if (relpos > 0)
- goto out_of_frame;
- update_frametailpos(winstate);
- abs_pos = winstate->frametailpos - 1 + relpos;
-
- /*
- * Account for exclusion option if one is active. If there is no
- * exclusion, we can safely set the mark at the accessed row. But
- * if there is, we can only mark the frame start, because we can't
- * be sure how far back in the frame the exclusion might cause us
- * to fetch in future. Furthermore, we have to actually check
- * against frameheadpos here, since it's unsafe to try to fetch a
- * row before frame start if the mark might be there already.
- */
- switch (winstate->frameOptions & FRAMEOPTION_EXCLUSION)
- {
- case 0:
- /* no adjustment needed */
- mark_pos = abs_pos;
- break;
- case FRAMEOPTION_EXCLUDE_CURRENT_ROW:
- if (abs_pos <= winstate->currentpos &&
- winstate->currentpos < winstate->frametailpos)
- abs_pos--;
- update_frameheadpos(winstate);
- if (abs_pos < winstate->frameheadpos)
- goto out_of_frame;
- mark_pos = winstate->frameheadpos;
- break;
- case FRAMEOPTION_EXCLUDE_GROUP:
- update_grouptailpos(winstate);
- if (abs_pos < winstate->grouptailpos &&
- winstate->groupheadpos < winstate->frametailpos)
- {
- int64 overlapend = Min(winstate->grouptailpos,
- winstate->frametailpos);
-
- abs_pos -= overlapend - winstate->groupheadpos;
- }
- update_frameheadpos(winstate);
- if (abs_pos < winstate->frameheadpos)
- goto out_of_frame;
- mark_pos = winstate->frameheadpos;
- break;
- case FRAMEOPTION_EXCLUDE_TIES:
- update_grouptailpos(winstate);
- if (abs_pos < winstate->grouptailpos &&
- winstate->groupheadpos < winstate->frametailpos)
- {
- int64 overlapend = Min(winstate->grouptailpos,
- winstate->frametailpos);
-
- if (abs_pos == overlapend - 1)
- abs_pos = winstate->currentpos;
- else
- abs_pos -= overlapend - 1 - winstate->groupheadpos;
- }
- update_frameheadpos(winstate);
- if (abs_pos < winstate->frameheadpos)
- goto out_of_frame;
- mark_pos = winstate->frameheadpos;
- break;
- default:
- elog(ERROR, "unrecognized frame option state: 0x%x",
- winstate->frameOptions);
- mark_pos = 0; /* keep compiler quiet */
- break;
- }
- break;
- default:
- elog(ERROR, "unrecognized window seek type: %d", seektype);
- abs_pos = mark_pos = 0; /* keep compiler quiet */
- break;
+ econtext->ecxt_outertuple = slot;
+ return ExecEvalExpr((ExprState *) list_nth(winobj->argstates, argno),
+ econtext, isnull);
}
- if (!window_gettupleslot(winobj, abs_pos, slot))
- goto out_of_frame;
-
- /* The code above does not detect all out-of-frame cases, so check */
- if (row_is_in_frame(winobj, abs_pos, slot, false) <= 0)
- goto out_of_frame;
-
- if (isout)
- *isout = false;
- if (set_mark)
- WinSetMarkPosition(winobj, mark_pos);
- econtext->ecxt_outertuple = slot;
- return ExecEvalExpr((ExprState *) list_nth(winobj->argstates, argno),
- econtext, isnull);
-
-out_of_frame:
if (isout)
*isout = true;
*isnull = true;
diff --git a/src/backend/jit/llvm/llvmjit_expr.c b/src/backend/jit/llvm/llvmjit_expr.c
index 0e160b8502c..e42c5d65bb6 100644
--- a/src/backend/jit/llvm/llvmjit_expr.c
+++ b/src/backend/jit/llvm/llvmjit_expr.c
@@ -129,6 +129,9 @@ llvm_compile_expr(ExprState *state)
LLVMValueRef v_aggvalues;
LLVMValueRef v_aggnulls;
+ /* RPR navigation: when true, EEOP_OUTER_VAR reloads from econtext */
+ bool has_rpr_nav;
+
instr_time starttime;
instr_time deform_starttime;
instr_time endtime;
@@ -298,6 +301,36 @@ llvm_compile_expr(ExprState *state)
FIELDNO_EXPRCONTEXT_AGGNULLS,
"v.econtext.aggnulls");
+ /*
+ * RPR navigation opcodes (PREV/NEXT) swap ecxt_outertuple to a different
+ * row mid-expression. The JIT code loads v_outervalues and v_outernulls
+ * once in the entry block and reuses them for all EEOP_OUTER_VAR steps.
+ * After a slot swap, these cached pointers become stale because the new
+ * slot has its own tts_values/tts_isnull arrays.
+ *
+ * When RPR navigation opcodes are present, EEOP_OUTER_VAR reloads the
+ * slot pointer from econtext->ecxt_outertuple on every access instead of
+ * using the cached entry-block values. This avoids the SSA/PHI
+ * complexity while keeping the rest of the expression JIT-compiled.
+ * Expressions without RPR navigation use the cached values as before.
+ */
+ has_rpr_nav = false;
+ if (parent && IsA(parent, WindowAggState) &&
+ ((WindowAgg *) parent->plan)->rpPattern != NULL)
+ {
+ for (int opno = 0; opno < state->steps_len; opno++)
+ {
+ ExprEvalOp opcode = ExecEvalStepOp(state, &state->steps[opno]);
+
+ if (opcode == EEOP_RPR_NAV_SET ||
+ opcode == EEOP_RPR_NAV_RESTORE)
+ {
+ has_rpr_nav = true;
+ break;
+ }
+ }
+ }
+
/* allocate blocks for each op upfront, so we can do jumps easily */
opblocks = palloc_array(LLVMBasicBlockRef, state->steps_len);
for (int opno = 0; opno < state->steps_len; opno++)
@@ -460,8 +493,37 @@ llvm_compile_expr(ExprState *state)
}
else if (opcode == EEOP_OUTER_VAR)
{
- v_values = v_outervalues;
- v_nulls = v_outernulls;
+ if (has_rpr_nav)
+ {
+ /*
+ * RPR navigation swaps ecxt_outertuple
+ * mid-expression. Reload slot pointer from
+ * econtext on every access so we read from the
+ * current (possibly swapped) slot.
+ */
+ LLVMValueRef v_tmpslot;
+
+ v_tmpslot = l_load_struct_gep(b,
+ StructExprContext,
+ v_econtext,
+ FIELDNO_EXPRCONTEXT_OUTERTUPLE,
+ "v_outerslot_reload");
+ v_values = l_load_struct_gep(b,
+ StructTupleTableSlot,
+ v_tmpslot,
+ FIELDNO_TUPLETABLESLOT_VALUES,
+ "v_outervalues_reload");
+ v_nulls = l_load_struct_gep(b,
+ StructTupleTableSlot,
+ v_tmpslot,
+ FIELDNO_TUPLETABLESLOT_ISNULL,
+ "v_outernulls_reload");
+ }
+ else
+ {
+ v_values = v_outervalues;
+ v_nulls = v_outernulls;
+ }
}
else if (opcode == EEOP_SCAN_VAR)
{
@@ -2434,6 +2496,18 @@ llvm_compile_expr(ExprState *state)
LLVMBuildBr(b, opblocks[opno + 1]);
break;
+ case EEOP_RPR_NAV_SET:
+ build_EvalXFunc(b, mod, "ExecEvalRPRNavSet",
+ v_state, op, v_econtext);
+ LLVMBuildBr(b, opblocks[opno + 1]);
+ break;
+
+ case EEOP_RPR_NAV_RESTORE:
+ build_EvalXFunc(b, mod, "ExecEvalRPRNavRestore",
+ v_state, op, v_econtext);
+ LLVMBuildBr(b, opblocks[opno + 1]);
+ break;
+
case EEOP_AGG_STRICT_DESERIALIZE:
case EEOP_AGG_DESERIALIZE:
{
diff --git a/src/backend/jit/llvm/llvmjit_types.c b/src/backend/jit/llvm/llvmjit_types.c
index c8a1f841293..e78b31d775f 100644
--- a/src/backend/jit/llvm/llvmjit_types.c
+++ b/src/backend/jit/llvm/llvmjit_types.c
@@ -168,6 +168,8 @@ void *referenced_functions[] =
ExecEvalScalarArrayOp,
ExecEvalHashedScalarArrayOp,
ExecEvalSubPlan,
+ ExecEvalRPRNavSet,
+ ExecEvalRPRNavRestore,
ExecEvalSysVar,
ExecEvalWholeRowVar,
ExecEvalXmlExpr,
diff --git a/src/backend/utils/adt/windowfuncs.c b/src/backend/utils/adt/windowfuncs.c
index 78b7f05aba2..3869f6c8994 100644
--- a/src/backend/utils/adt/windowfuncs.c
+++ b/src/backend/utils/adt/windowfuncs.c
@@ -41,7 +41,6 @@ static bool rank_up(WindowObject winobj);
static Datum leadlag_common(FunctionCallInfo fcinfo,
bool forward, bool withoffset, bool withdefault);
-
/*
* utility routine for *_rank functions.
*/
@@ -724,3 +723,121 @@ window_nth_value(PG_FUNCTION_ARGS)
PG_RETURN_DATUM(result);
}
+
+/*
+ * prev
+ * Catalog placeholder for RPR's PREV navigation operator.
+ *
+ * The parser transforms prev() calls inside DEFINE into RPRNavExpr nodes,
+ * so this function is never reached during normal RPR execution. It exists
+ * only so that the parser can resolve the function name from pg_proc.
+ * Calls outside DEFINE are rejected by parse_func.c (EXPR_KIND_RPR_DEFINE
+ * check). The error below is a defensive measure in case that check is
+ * bypassed (e.g., direct C-level function invocation).
+ */
+Datum
+window_prev(PG_FUNCTION_ARGS)
+{
+ ereport(ERROR,
+ errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+ errmsg("cannot use PREV() outside a DEFINE clause"));
+ PG_RETURN_NULL(); /* not reached */
+}
+
+/*
+ * next
+ * Catalog placeholder for RPR's NEXT navigation operator.
+ * See window_prev() for details.
+ */
+Datum
+window_next(PG_FUNCTION_ARGS)
+{
+ ereport(ERROR,
+ errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+ errmsg("cannot use NEXT() outside a DEFINE clause"));
+ PG_RETURN_NULL(); /* not reached */
+}
+
+/*
+ * prev(value, offset)
+ * Catalog placeholder for RPR's PREV navigation operator with offset.
+ * See window_prev() for details.
+ */
+Datum
+window_prev_offset(PG_FUNCTION_ARGS)
+{
+ ereport(ERROR,
+ errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+ errmsg("cannot use PREV() outside a DEFINE clause"));
+ PG_RETURN_NULL(); /* not reached */
+}
+
+/*
+ * next(value, offset)
+ * Catalog placeholder for RPR's NEXT navigation operator with offset.
+ * See window_prev() for details.
+ */
+Datum
+window_next_offset(PG_FUNCTION_ARGS)
+{
+ ereport(ERROR,
+ errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+ errmsg("cannot use NEXT() outside a DEFINE clause"));
+ PG_RETURN_NULL(); /* not reached */
+}
+
+/*
+ * first
+ * Catalog placeholder for RPR's FIRST navigation operator.
+ * See window_prev() for details.
+ */
+Datum
+window_first(PG_FUNCTION_ARGS)
+{
+ ereport(ERROR,
+ errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+ errmsg("cannot use FIRST() outside a DEFINE clause"));
+ PG_RETURN_NULL(); /* not reached */
+}
+
+/*
+ * last
+ * Catalog placeholder for RPR's LAST navigation operator.
+ * See window_prev() for details.
+ */
+Datum
+window_last(PG_FUNCTION_ARGS)
+{
+ ereport(ERROR,
+ errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+ errmsg("cannot use LAST() outside a DEFINE clause"));
+ PG_RETURN_NULL(); /* not reached */
+}
+
+/*
+ * first(value, offset)
+ * Catalog placeholder for RPR's FIRST navigation operator with offset.
+ * See window_prev() for details.
+ */
+Datum
+window_first_offset(PG_FUNCTION_ARGS)
+{
+ ereport(ERROR,
+ errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+ errmsg("cannot use FIRST() outside a DEFINE clause"));
+ PG_RETURN_NULL(); /* not reached */
+}
+
+/*
+ * last(value, offset)
+ * Catalog placeholder for RPR's LAST navigation operator with offset.
+ * See window_prev() for details.
+ */
+Datum
+window_last_offset(PG_FUNCTION_ARGS)
+{
+ ereport(ERROR,
+ errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+ errmsg("cannot use LAST() outside a DEFINE clause"));
+ PG_RETURN_NULL(); /* not reached */
+}
diff --git a/src/include/catalog/pg_proc.dat b/src/include/catalog/pg_proc.dat
index be157a5fbe9..b3aa42fc66e 100644
--- a/src/include/catalog/pg_proc.dat
+++ b/src/include/catalog/pg_proc.dat
@@ -10967,6 +10967,30 @@
{ oid => '3114', descr => 'fetch the Nth row value',
proname => 'nth_value', prokind => 'w', prorettype => 'anyelement',
proargtypes => 'anyelement int4', prosrc => 'window_nth_value' },
+{ oid => '8126', descr => 'fetch the preceding row value',
+ proname => 'prev', provolatile => 's', prorettype => 'anyelement',
+ proargtypes => 'anyelement', prosrc => 'window_prev' },
+{ oid => '8128', descr => 'fetch the Nth preceding row value',
+ proname => 'prev', provolatile => 's', proisstrict => 'f', prorettype => 'anyelement',
+ proargtypes => 'anyelement int8', prosrc => 'window_prev_offset' },
+{ oid => '8127', descr => 'fetch the following row value',
+ proname => 'next', provolatile => 's', prorettype => 'anyelement',
+ proargtypes => 'anyelement', prosrc => 'window_next' },
+{ oid => '8129', descr => 'fetch the Nth following row value',
+ proname => 'next', provolatile => 's', proisstrict => 'f', prorettype => 'anyelement',
+ proargtypes => 'anyelement int8', prosrc => 'window_next_offset' },
+{ oid => '8130', descr => 'fetch the first row value within match',
+ proname => 'first', provolatile => 's', prorettype => 'anyelement',
+ proargtypes => 'anyelement', prosrc => 'window_first' },
+{ oid => '8132', descr => 'fetch the Nth row value within match',
+ proname => 'first', provolatile => 's', proisstrict => 'f', prorettype => 'anyelement',
+ proargtypes => 'anyelement int8', prosrc => 'window_first_offset' },
+{ oid => '8131', descr => 'fetch the last row value within match',
+ proname => 'last', provolatile => 's', prorettype => 'anyelement',
+ proargtypes => 'anyelement', prosrc => 'window_last' },
+{ oid => '8133', descr => 'fetch the Nth-from-last row value within match',
+ proname => 'last', provolatile => 's', proisstrict => 'f', prorettype => 'anyelement',
+ proargtypes => 'anyelement int8', prosrc => 'window_last_offset' },
# functions for range types
{ oid => '3832', descr => 'I/O',
diff --git a/src/include/executor/execExpr.h b/src/include/executor/execExpr.h
index c61b3d624d5..db66ebe313c 100644
--- a/src/include/executor/execExpr.h
+++ b/src/include/executor/execExpr.h
@@ -274,6 +274,10 @@ typedef enum ExprEvalOp
EEOP_MERGE_SUPPORT_FUNC,
EEOP_SUBPLAN,
+ /* row pattern navigation (RPR PREV/NEXT) */
+ EEOP_RPR_NAV_SET,
+ EEOP_RPR_NAV_RESTORE,
+
/* aggregation related nodes */
EEOP_AGG_STRICT_DESERIALIZE,
EEOP_AGG_DESERIALIZE,
@@ -695,6 +699,18 @@ typedef struct ExprEvalStep
SubPlanState *sstate;
} subplan;
+ /* for EEOP_RPR_NAV_SET / EEOP_RPR_NAV_RESTORE */
+ struct
+ {
+ WindowAggState *winstate;
+ RPRNavKind kind; /* navigation kind (simple or compound) */
+ Datum *offset_value; /* offset value(s), or NULL */
+ bool *offset_isnull; /* offset null flag(s) */
+ /* For compound nav: offset_value[0] = inner, [1] = outer */
+ int16 resulttyplen; /* RESTORE: result type length */
+ bool resulttypbyval; /* RESTORE: result pass-by-value? */
+ } rpr_nav;
+
/* for EEOP_AGG_*DESERIALIZE */
struct
{
@@ -902,6 +918,10 @@ extern void ExecEvalMergeSupportFunc(ExprState *state, ExprEvalStep *op,
ExprContext *econtext);
extern void ExecEvalSubPlan(ExprState *state, ExprEvalStep *op,
ExprContext *econtext);
+extern void ExecEvalRPRNavSet(ExprState *state, ExprEvalStep *op,
+ ExprContext *econtext);
+extern void ExecEvalRPRNavRestore(ExprState *state, ExprEvalStep *op,
+ ExprContext *econtext);
extern void ExecEvalWholeRowVar(ExprState *state, ExprEvalStep *op,
ExprContext *econtext);
extern void ExecEvalSysVar(ExprState *state, ExprEvalStep *op,
diff --git a/src/include/executor/execRPR.h b/src/include/executor/execRPR.h
new file mode 100644
index 00000000000..7b2b0febb76
--- /dev/null
+++ b/src/include/executor/execRPR.h
@@ -0,0 +1,40 @@
+/*-------------------------------------------------------------------------
+ *
+ * execRPR.h
+ * prototypes for execRPR.c (NFA-based Row Pattern Recognition engine)
+ *
+ *
+ * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
+ * Portions Copyright (c) 1994, Regents of the University of California
+ *
+ * src/include/executor/execRPR.h
+ *
+ *-------------------------------------------------------------------------
+ */
+#ifndef EXECRPR_H
+#define EXECRPR_H
+
+#include "nodes/execnodes.h"
+#include "windowapi.h"
+
+/* NFA context management */
+extern RPRNFAContext *ExecRPRStartContext(WindowAggState *winstate,
+ int64 startPos);
+extern RPRNFAContext *ExecRPRGetHeadContext(WindowAggState *winstate,
+ int64 pos);
+extern void ExecRPRFreeContext(WindowAggState *winstate, RPRNFAContext *ctx);
+
+/* NFA processing */
+extern void ExecRPRProcessRow(WindowAggState *winstate, int64 currentPos,
+ bool hasLimitedFrame, int64 frameOffset);
+extern void ExecRPRCleanupDeadContexts(WindowAggState *winstate,
+ RPRNFAContext *excludeCtx);
+extern void ExecRPRFinalizeAllContexts(WindowAggState *winstate, int64 lastPos);
+
+/* NFA statistics */
+extern void ExecRPRRecordContextSuccess(WindowAggState *winstate,
+ int64 matchLen);
+extern void ExecRPRRecordContextFailure(WindowAggState *winstate,
+ int64 failedLen);
+
+#endif /* EXECRPR_H */
diff --git a/src/include/executor/nodeWindowAgg.h b/src/include/executor/nodeWindowAgg.h
index ada4a1c458c..f6f6645131c 100644
--- a/src/include/executor/nodeWindowAgg.h
+++ b/src/include/executor/nodeWindowAgg.h
@@ -20,4 +20,7 @@ extern WindowAggState *ExecInitWindowAgg(WindowAgg *node, EState *estate, int ef
extern void ExecEndWindowAgg(WindowAggState *node);
extern void ExecReScanWindowAgg(WindowAggState *node);
+/* RPR navigation support for expression evaluation opcodes */
+extern TupleTableSlot *ExecRPRNavGetSlot(WindowAggState *winstate, int64 pos);
+
#endif /* NODEWINDOWAGG_H */
diff --git a/src/include/nodes/execnodes.h b/src/include/nodes/execnodes.h
index 53c138310db..792aa3f0d05 100644
--- a/src/include/nodes/execnodes.h
+++ b/src/include/nodes/execnodes.h
@@ -38,6 +38,9 @@
#include "nodes/plannodes.h"
#include "partitioning/partdefs.h"
#include "storage/buf.h"
+#include "storage/condition_variable.h"
+#include "utils/hsearch.h"
+#include "utils/queryenvironment.h"
#include "utils/reltrigger.h"
#include "utils/typcache.h"
@@ -2525,6 +2528,71 @@ typedef enum WindowAggStatus
* tuples during spool */
} WindowAggStatus;
+/* RPR reduced frame states returned by get_reduced_frame_status() */
+#define RF_NOT_DETERMINED 0 /* not yet processed */
+#define RF_FRAME_HEAD 1 /* start row of a match */
+#define RF_SKIPPED 2 /* interior row of a match */
+#define RF_UNMATCHED 3 /* no match at this row */
+#define RF_EMPTY_MATCH 4 /* empty match (0 rows); treated as unmatched */
+
+/*
+ * RPRNFAState - single NFA state for pattern matching
+ *
+ * counts[] tracks repetition counts at each nesting depth.
+ *
+ * isAbsorbable tracks if state is in absorbable region (ABSORBABLE_BRANCH).
+ * Monotonic property: once false, stays false (can't re-enter region).
+ */
+typedef struct RPRNFAState
+{
+ struct RPRNFAState *next; /* next state in linked list */
+ int16 elemIdx; /* current pattern element index */
+ bool isAbsorbable; /* true if state is in absorbable region */
+ int32 counts[FLEXIBLE_ARRAY_MEMBER]; /* repetition counts by depth */
+} RPRNFAState;
+
+/*
+ * RPRNFAContext - context for NFA pattern matching execution
+ *
+ * Two-flag absorption design:
+ * hasAbsorbableState: can this context absorb others? (>=1 absorbable state)
+ * - Monotonic: true->false only, cannot recover once false
+ * - Used to skip absorption attempts once all absorbable states are gone
+ * allStatesAbsorbable: can this context be absorbed? (ALL states absorbable)
+ * - Dynamic: can change false->true (when non-absorbable states die)
+ * - Used to determine if this context is eligible for absorption
+ */
+typedef struct RPRNFAContext
+{
+ struct RPRNFAContext *next; /* next context in linked list */
+ struct RPRNFAContext *prev; /* previous context (for reverse traversal) */
+ RPRNFAState *states; /* active states (linked list) */
+
+ int64 matchStartRow; /* row where match started */
+ int64 matchEndRow; /* row where match ended (-1 = no match) */
+ int64 lastProcessedRow; /* last row processed (for fail depth) */
+ RPRNFAState *matchedState; /* FIN state for greedy fallback (cloned) */
+
+ /* Two-flag absorption optimization */
+ bool hasAbsorbableState; /* can absorb others (>=1 absorbable
+ * state) */
+ bool allStatesAbsorbable; /* can be absorbed (ALL states
+ * absorbable) */
+} RPRNFAContext;
+
+/*
+ * NFALengthStats
+ *
+ * Statistics for length measurements (min/max/total) used for computing
+ * average lengths in EXPLAIN ANALYZE output.
+ */
+typedef struct NFALengthStats
+{
+ int64 min; /* minimum length */
+ int64 max; /* maximum length */
+ int64 total; /* total length (for computing average) */
+} NFALengthStats;
+
typedef struct WindowAggState
{
ScanState ss; /* its first field is NodeTag */
@@ -2584,6 +2652,51 @@ typedef struct WindowAggState
int64 groupheadpos; /* current row's peer group head position */
int64 grouptailpos; /* " " " " tail position (group end+1) */
+ /* these fields are used in Row pattern recognition: */
+ RPSkipTo rpSkipTo; /* Row Pattern Skip To type */
+ struct RPRPattern *rpPattern; /* compiled pattern for NFA execution */
+ List *defineVariableList; /* list of row pattern definition
+ * variables (list of String) */
+ List *defineClauseList; /* expression for row pattern definition
+ * search conditions ExprState list */
+ RPRNFAContext *nfaContext; /* active matching contexts (head) */
+ RPRNFAContext *nfaContextTail; /* tail of active contexts (for reverse
+ * traversal) */
+ RPRNFAContext *nfaContextFree; /* recycled NFA context nodes */
+ RPRNFAState *nfaStateFree; /* recycled NFA state nodes */
+ Size nfaStateSize; /* pre-calculated RPRNFAState size */
+ bool *nfaVarMatched; /* per-row cache: varMatched[varId] for varId
+ * < numDefines */
+ Bitmapset *defineMatchStartDependent; /* DEFINE vars needing per-context
+ * evaluation
+ * (match_start-dependent) */
+ bitmapword *nfaVisitedElems; /* elemIdx visited bitmap for cycle
+ * detection */
+ int16 nfaVisitedMinWord; /* lowest bitmapword index touched since
+ * last reset (PG_INT16_MAX = none) */
+ int16 nfaVisitedMaxWord; /* highest bitmapword index touched since
+ * last reset (-1 = none) */
+ int64 nfaLastProcessedRow; /* last row processed by NFA (-1 =
+ * none) */
+
+ /* NFA statistics for EXPLAIN ANALYZE */
+ int64 nfaStatesActive; /* current active states (internal) */
+ int64 nfaStatesMax; /* peak active states */
+ int64 nfaStatesTotalCreated; /* total states allocated */
+ int64 nfaStatesMerged; /* states merged (deduplicated) */
+ int64 nfaContextsActive; /* current active contexts (internal) */
+ int64 nfaContextsMax; /* peak active contexts */
+ int64 nfaContextsTotalCreated; /* total contexts allocated */
+ int64 nfaContextsAbsorbed; /* contexts absorbed (optimization) */
+ int64 nfaContextsSkipped; /* contexts skipped (SKIP PAST LAST ROW) */
+ int64 nfaContextsPruned; /* contexts pruned on first row */
+ int64 nfaMatchesSucceeded; /* successful pattern matches */
+ int64 nfaMatchesFailed; /* failed pattern matches */
+ NFALengthStats nfaMatchLen; /* successful match length stats */
+ NFALengthStats nfaFailLen; /* mismatch length stats */
+ NFALengthStats nfaAbsorbedLen; /* absorbed context length stats */
+ NFALengthStats nfaSkippedLen; /* skipped context length stats */
+
MemoryContext partcontext; /* context for partition-lifespan data */
MemoryContext aggcontext; /* shared context for aggregate working data */
MemoryContext curaggcontext; /* current aggregate's working data */
@@ -2611,6 +2724,25 @@ typedef struct WindowAggState
TupleTableSlot *agg_row_slot;
TupleTableSlot *temp_slot_1;
TupleTableSlot *temp_slot_2;
+
+ /* RPR navigation */
+ RPRNavOffsetKind navMaxOffsetKind; /* status of navMaxOffset */
+ int64 navMaxOffset; /* max backward nav offset (when FIXED) */
+ bool hasFirstNav; /* FIRST() present in DEFINE */
+ RPRNavOffsetKind navFirstOffsetKind; /* status of navFirstOffset */
+ int64 navFirstOffset; /* min FIRST() offset (when FIXED) */
+ struct WindowObjectData *nav_winobj; /* winobj for RPR nav fetch */
+ int64 nav_slot_pos; /* position cached in nav_slot, or -1 */
+ TupleTableSlot *nav_slot; /* slot for PREV/NEXT/FIRST/LAST target row */
+ TupleTableSlot *nav_saved_outertuple; /* saved slot during nav swap */
+ TupleTableSlot *nav_null_slot; /* all NULL slot */
+ int64 nav_match_start; /* match_start for FIRST/LAST nav */
+
+ /* RPR current match result */
+ bool rpr_match_valid; /* true if a match result is set */
+ bool rpr_match_matched; /* true if the result was a match */
+ int64 rpr_match_start; /* start position of the match result */
+ int64 rpr_match_length; /* number of rows matched (0 = empty) */
} WindowAggState;
/* ----------------
--
2.43.0