Summary: Fix a relation going missing when a CTE shares its name

[aurelienbottazini]

Related PR on Ruby gem: pganalyze/pg_query#355

pg_query_summary() drops a relation when a CTE shares its name:

"WITH t AS (SELECT * FROM s.t) SELECT 1"   /* tables: []   expected: s.t */
"WITH t AS (SELECT * FROM t) SELECT 1"     /* tables: []   expected: t   */

WalkState.cte_names is a single flat set for the whole statement, and
handle_range_var() treated every range var whose relname matched one of those
names as a CTE reference. Three separate cases fall out of that:

1. A non-recursive CTE is not visible inside its own definition, so in
   WITH users AS (SELECT * FROM users) SELECT * FROM users the inner reference
   is the users table and only the outer one is the CTE. Both were skipped.
2. Only relname was compared, so a schema qualified reference was skipped too
   (the first example above).
3. A CTE cannot be the target of INSERT/UPDATE/DELETE/MERGE or DDL, so
   those always name a real relation. WITH users AS (...) UPDATE users SET ...
   reported no target table.

This is easy to hit from ORMs that name a CTE after the model's own table, and it
matters for anything using the table list for query attribution, read/write
routing or cache invalidation, because the relation goes missing silently.

Fix

• handle_with_clause() runs before descending into the statement, so the CTE names of enclosing scopes are
  known by the time a nested WITH clause is reached.
• It records the range vars inside each CTE's own definition that carry the CTE's
  own name (cte_self_reference_walker()), and handle_range_var() skips only
  those. The scan is skipped forWITH RECURSIVE (which does make a CTE visible to itself) and when the name is
  already a visible CTE from an enclosing scope, e.g. the inner a in
  WITH a AS (...) SELECT * FROM (WITH a AS (SELECT * FROM a) ...) s.
• is_cte_reference() now gathers the whole decision: CONTEXT_SELECT only,
  unqualified only, name in cte_names, and not a recorded self-reference.

Range vars are still resolved after the tree walk, and the order of
summary->tables and summary->cte_names is unchanged, so no existing
expectations move.

Tests

13 cases in test/summary_tests.c (test/summary_tests_list.c regenerated with
scripts/update_summary_tests_list.rb)

Against unpatched src/, 9 of them fail (20 assertions). With the fix,
make test is green

Performance

Measured through the Ruby binding, with these files dropped into pg_query's
ext/pg_query/ and PgQuery.summary called 20k times per query:

query	before	after
SELECT * FROM companies WHERE id = 1	1.91us	1.89us
WITH x AS (SELECT * FROM companies) SELECT * FROM x	2.08us	2.14us
WITH companies AS (SELECT * FROM companies) SELECT * FROM companies	1.90us	2.07us
3-CTE query with joins	7.23us	7.78us
SELECT * FROM t0 + 100 joins	36.0us	37.1us

Statements without a WITH clause are unaffected; statements with one pay a
single extra pass over the CTE definitions.

Notes / open questions

No CHANGELOG.md entry here, matching recent fix PRs on this branch. Glad to
add one wherever you'd like it.

[lfittl]

Thanks! I've asked @duckinator to take a look at this.

[duckinator]

This looks good to me. I'm not surprised there were a few bugs lurking in the CTE handling; it had a lot of edge cases.