Fix EXISTS with select-list window functions (row-index error)

[vishnujayvel]

Fixes: dolthub/dolt#11421

Reported by @Yibo-Dong.

Problem

A correlated EXISTS whose select list is a window expression (e.g. ROW_NUMBER() OVER ()) fails with an internal error:

unable to find field with index 3 in row of 3 columns
This is a bug. Please file an issue here: https://github.com/dolthub/dolt/issues

The same query with SELECT 1 (or any non-window projection) works. MySQL accepts the window form and returns the expected existence result.

This is not crash-only. Multi-window EXISTS bodies (two or more window functions with distinct partition schemes) return silently wrong rows on main — for example, with u2(x,y) rows (1,10),(1,20),(2,30) and t ids 1,2,3:

SELECT id FROM t
WHERE EXISTS (
  SELECT ROW_NUMBER() OVER (PARTITION BY u2.y), RANK() OVER (ORDER BY u2.y)
  FROM u2 WHERE u2.x = t.a
)
ORDER BY id;
-- main (wrong): 1, 2, 3
-- correct / MySQL: 1, 2

NOT EXISTS of the same shape returns no rows on main instead of row 3. Silent wrong results are the highest-priority maintainer concern for this fix.

Separately, the uncorrelated NOT EXISTS ... OVER () shape panics the server on main (nil-pointer SIGSEGV at sql/core.go:432 via join_iters.go:433). Panic outranks internal error.

How to reproduce (crash shape)

CREATE TABLE t(id INT PRIMARY KEY, a INT);
CREATE TABLE u(x INT);
INSERT INTO t VALUES (1,1),(2,2),(3,3);
INSERT INTO u VALUES (1),(1),(2);

SELECT id FROM t
WHERE EXISTS (SELECT ROW_NUMBER() OVER () FROM u WHERE u.x = t.a)
ORDER BY id;
-- expected: 1, 2

Root cause

simplifyPartialJoinParents in sql/analyzer/unnest_exists_subqueries.go strips nodes that do not affect existence (Project, GroupBy, Sort, Distinct, TopN, Limit) from the decorrelated EXISTS body before building the semi-join right side. It did not strip *plan.Window.

The Window node shrinks the right-side schema to the window outputs, while join-filter GetField indices still refer to pre-window column ids. Evaluation then hits an out-of-bounds GetField in sql/expression/get_field.go via the exists iterator — or, with multiple windows, can produce silently incorrect existence results.

Fix

Treat *plan.Window like *plan.Project in simplifyPartialJoinParents: strip it when simplifying the EXISTS body. Window functions are 1:1 row-preserving (empty child → empty output), so existence does not depend on select-list window outputs.

Filter is not in the strip list, so the loop halts at any surviving Filter; correlated Filters are hoisted by decorrelateOuterCols and re-attached above the simplified node (unnest_exists_subqueries.go:416-417). A Filter with zero correlated conjuncts is left untouched (:314-317 returns SameTree). Having bodies are separately rejected earlier (case *plan.Having: return nil, false). Stripping is therefore limited to pure select-list window projection wrappers.

Tests

Enginetest script in JoinScriptTests (enginetest/queries/join_queries.go):

• Control: EXISTS (SELECT 1 FROM u WHERE u.x = t.a) → rows 1, 2
• Bug (crash shape): EXISTS (SELECT ROW_NUMBER() OVER () FROM u WHERE u.x = t.a) → rows 1, 2
• NOT EXISTS variant with the same window select list → row 3
• Multi-window silent-corruption shape (distinct partition schemes on u2) → EXISTS 1, 2; NOT EXISTS 3
• Uncorrelated panic shape: NOT EXISTS (SELECT ROW_NUMBER() OVER () FROM u) → empty result (0 rows)

Test plan

export CGO_CFLAGS="-I/opt/homebrew/opt/icu4c@78/include"
export CGO_CXXFLAGS="-I/opt/homebrew/opt/icu4c@78/include"
export CGO_LDFLAGS="-L/opt/homebrew/opt/icu4c@78/lib"

go test ./sql/analyzer/... -count=1
go test ./enginetest/... -count=1

• Full suite re-run after this patch:

$ go test ./sql/analyzer/... -count=1
ok  github.com/dolthub/go-mysql-server/sql/analyzer  0.457s

$ go test ./enginetest/... -count=1
ok  github.com/dolthub/go-mysql-server/enginetest              44.573s
ok  github.com/dolthub/go-mysql-server/enginetest/mysql_harness 1.047s
ok  github.com/dolthub/go-mysql-server/enginetest/scriptgen/setup 0.205s

JoinQueries multi-window subtests (verbatim -v):

--- PASS: TestJoinQueries/EXISTS_with_select-list_window_function (0.00s)
    --- PASS: .../SELECT_id_FROM_t_WHERE_EXISTS_(SELECT_1_FROM_u_WHERE_u.x_=_t.a)_ORDER_BY_id
    --- PASS: .../SELECT_id_FROM_t_WHERE_EXISTS_(SELECT_ROW_NUMBER()_OVER_()_FROM_u_WHERE_u.x_=_t.a)_ORDER_BY_id
    --- PASS: .../SELECT_id_FROM_t_WHERE_NOT_EXISTS_(SELECT_ROW_NUMBER()_OVER_()_FROM_u_WHERE_u.x_=_t.a)_ORDER_BY_id
    --- PASS: .../SELECT_id_FROM_t_WHERE_EXISTS_(SELECT_ROW_NUMBER()_OVER_(PARTITION_BY_u2.y),_RANK()_OVER_(ORDER_BY_u2.y)_FROM_u2_WHERE_u2.x_=_t.a)_ORDER_BY_id
    --- PASS: .../SELECT_id_FROM_t_WHERE_NOT_EXISTS_(SELECT_ROW_NUMBER()_OVER_(PARTITION_BY_u2.y),_RANK()_OVER_(ORDER_BY_u2.y)_FROM_u2_WHERE_u2.x_=_t.a)_ORDER_BY_id

Limitations / follow-ups

• Fix is in go-mysql-server. Dolt consumers need a GMS dependency pin bump as a separate change.
• Only pure select-list windows are stripped. Bodies where a window feeds a HAVING above the Window still refuse unnest (Having returns false); that path was already conservative and is unchanged.
• LIMIT stripping is pre-existing (TODO at unnest_exists_subqueries.go:114-115, dolt#10493). With Window now stripped, EXISTS(...window... LIMIT 0) reaches that pre-existing hole (base hard-errors; patched returns rows where MySQL says empty) — same behavior as the existing SELECT 1 LIMIT 0 shape; nothing new introduced, noted for completeness.

Checklist

• Minimal analyzer change (*plan.Window only)
• Failing-before / passing-after enginetest for the issue query + NOT EXISTS + SELECT 1 control + multi-window silent-corruption shape
• CI green
• Dolt GMS pin bump (follow-up)