MIME-Version: 1.0
References: <19435-3cc1a87f291129f1@postgresql.org> <CAHGQGwEEHFnH8DsZbwxdZeiHozm2LCRTLSgL8Qn=5MoN4450ZA@mail.gmail.com>
 <CAPpHfdsDQhDqvqi+-UxvyYRP72ASkdWzCn43e2Hj6WsY15opvA@mail.gmail.com>
 <CAHewXNmU3E0oXLgvoOtXw7kfTiAZsKu8La_hQ=pyScgxvP45iw@mail.gmail.com>
 <CAHewXN=LjuWz3PcyhjdbJAyo+Zs9MisPDRYnSZBUy4PMeKi+zA@mail.gmail.com>
 <CALdSSPj1kTTQvmV3H3HMf5P3um8ybxoH3DaTPm+XgdYAur1Q4A@mail.gmail.com>
 <CAHewXNndByMu3S+_h4LLDkXA5qrO1s=s-CE8HqUtc9vTA9yrjg@mail.gmail.com>
 <CAPpHfdv6gzSTXHJxYSgB8sULadXM4wvhgoQODaOxYCJfagKNPw@mail.gmail.com>
 <CAHewXN=7kDJjUcgEm+6qhaKOXuqzvhRqAAKdafNCRgn0yH7BGg@mail.gmail.com>
 <CAHewXNm5OOREJ8wZ1cLJdQz7O1aQ0E1RBB55S6O138K8vBdc9g@mail.gmail.com> <CAPpHfducqLJ=o3LkoPKGfZJVQuuei+P=2oUF6hX6rzHTZSxoyA@mail.gmail.com>
In-Reply-To: <CAPpHfducqLJ=o3LkoPKGfZJVQuuei+P=2oUF6hX6rzHTZSxoyA@mail.gmail.com>
From: Tender Wang <tndrwang@gmail.com>
Date: Wed, 18 Mar 2026 17:46:51 +0800
Message-ID: <CAHewXN=hG+jRgr1huE+hzdT-iLpR7r3c+2B5-4ajXBXfd+zcdQ@mail.gmail.com>
Subject: Re: BUG #19435: Error: "No relation entry for relid 2" Triggered by
 Complex Join with Self-Referencing Tables
To: Alexander Korotkov <aekorotkov@gmail.com>
Cc: Kirill Reshke <reshkekirill@gmail.com>, Fujii Masao <masao.fujii@gmail.com>, ammmkilo@163.com, 
	pgsql-bugs@lists.postgresql.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Archived-At: <https://www.postgresql.org/message-id/CAHewXN%3DhG%2BjRgr1huE%2BhzdT-iLpR7r3c%2B2B5-4ajXBXfd%2BzcdQ%40mail.gmail.com>
Precedence: bulk

Alexander Korotkov <aekorotkov@gmail.com> =E4=BA=8E2026=E5=B9=B43=E6=9C=881=
8=E6=97=A5=E5=91=A8=E4=B8=89 16:40=E5=86=99=E9=81=93=EF=BC=9A
>
> What about being more generic and call ChangeVarNodes_walker() for the
> node in ChangeVarNodesWalkExpression()?  It also works with out case
> and avoids code duplication.

Works for me.

>
> Changes in restrict_infos_logically_equal() makes me a bit uneasy.  I
> see, restictinfo's are different by their outer_relids.  Why
> outer_relids doesn't matter when required_relids is singleton?  More
> general when do outer_relids matter for add_non_redundant_clauses() if
> we're putting restictinfo's into a single list anyway?

(gdb) call nodeToString(binfo_candidates )
{RESTRICTINFO :clause {NULLTEST :arg {VAR :varno 3 :varattno 2
:vartype 16 :vartypmod -1 :varcollid 0 :varnullingrels (b)
:varlevelsup 0 :varreturningtype 0 :varnosyn 3 :varattnosyn 2
:location -1} :nulltesttype 1 :argisrow false :location -1}
:is_pushed_down true :can_join false :pseudoconstant false :has_clone
false :is_clone false :leakproof false :has_volatile 0 :security_level
0 :num_base_rels 1 :clause_relids (b 3) :required_relids (b 3)
:incompatible_relids (b) :outer_relids (b 5) :left_relids (b)
:right_relids (b) :orclause <> :rinfo_serial 3 :eval_cost.startup -1
:eval_cost.per_tuple 0 :norm_selec -1 :outer_selec -1 :mergeopfamilies
<> :left_em <> :right_em <> :outer_is_left false :hashjoinoperator 0
:left_bucketsize -1 :right_bucketsize -1 :left_mcvfreq -1
:right_mcvfreq -1 :left_hasheqoperator 0 :right_hasheqoperator 0}

{RESTRICTINFO :clause {NULLTEST :arg {VAR :varno 3 :varattno 2
:vartype 16 :vartypmod -1 :varcollid 0 :varnullingrels (b)
:varlevelsup 0 :varreturningtype 0 :varnosyn 3 :varattnosyn 2
:location -1} :nulltesttype 1 :argisrow false :location -1}
:is_pushed_down true :can_join true :pseudoconstant false :has_clone
false :is_clone false :leakproof false :has_volatile 2 :security_level
0 :num_base_rels 1 :clause_relids (b 3) :required_relids (b 3)
:incompatible_relids (b) :outer_relids (b) :left_relids (b 3)
:right_relids (b 3) :orclause <> :rinfo_serial 6 :eval_cost.startup -1
:eval_cost.per_tuple 0 :norm_selec -1 :outer_selec -1 :mergeopfamilies
<> :left_em <> :right_em <> :outer_is_left true :hashjoinoperator 91
:left_bucketsize -1 :right_bucketsize -1 :left_mcvfreq -1
:right_mcvfreq -1 :left_hasheqoperator 91 :right_hasheqoperator 91})"

The first is tom1.col_bool IS NOT NULL,  its outer_relids (b 5)  is
not empty because it's in the nullable side.
The second is added in replace_relid_callback().  Some fields do not match.

In this case, tom1.col_bool IS NOT NULL becomes a filter clause, not a
join clause, and is safe to only check rinfo->clause.
I think it's better to add is_pushed_down =3D=3D true check, for example:
...
-       if (bms_membership(a->required_relids) =3D=3D BMS_SINGLETON)
+       if (bms_membership(a->required_relids) =3D=3D BMS_SINGLETON &&
+               a->is_pushed_down &&
+               b->is_pushed_down)
...

Any thought?

--=20
Thanks,
Tender Wang