Re: BUG #19524: NaN handling in btree_gist's float4/float8 opclasses

 Hi all,

  I have a patch for this and `make check` passes (core + all contrib, plus
  new btree_gist NaN cases covering the three scenarios in the report).
  Posting here first in case someone is already further along — if not,
  I'll send it to pgsql-hackers and register it in the open CF in a couple
  of days.

  The fix swaps the five gbt_float{4,8}{gt,ge,eq,le,lt} comparators and
  gbt_float{4,8}key_cmp for the NaN-aware float{4,8}_* helpers and
  float{4,8}_cmp_internal() from utils/float.h — same total order the
  btree opclass already uses. No on-disk change.

  Bill

  Hi,

  contrib/btree_gist's float4/float8 GiST opclasses compared keys with
  raw C operators (==, <, >). Under IEEE 754 every comparison involving
  NaN returns false, so GiST disagreed with the regular btree opclass,
  which uses float[4|8]_cmp_internal() (all NaNs equal, NaN sorts after
  every non-NaN value).

  The disagreement was reachable from SQL:

  * an index scan on a GiST index over a float column returned no rows
    for `WHERE a = 'NaN'`, while a sequential scan returned the NaN row;
  * `EXCLUDE USING gist (a WITH =)` accepted two rows whose key was NaN,
    because gbt_float8eq(NaN, NaN) was false;
  * an RLS predicate `USING (a != 'NaN')` leaked NaN rows on index scan.

  The fix swaps the five gbt_float{4,8}{gt,ge,eq,le,lt} comparators and
  the picksplit gbt_float{4,8}key_cmp for the NaN-aware
  float{4,8}_{gt,ge,eq,le,lt} / float{4,8}_cmp_internal() helpers in
  utils/float.h — same total order the btree opclass already uses. No
  on-disk format change.

  Added regression coverage in contrib/btree_gist/sql/float{4,8}.sql for
  the three scenarios above. `make check` is green: core 245/245 +
  contrib btree_gist 32/32 + all 48 contrib modules.
  Originally reported as BUG #19501 and BUG #19524:
  https://www.postgresql.org/message-id/19501-3bff3bbc97f1e7c9%40postgresql.org
  https://www.postgresql.org/message-id/19524-9559d302c8455664%40postgresql.org

  I used an LLM (Claude Code) to help with the analysis and to draft the
  patch. I reviewed and tested the diff myself before sending.

  Regards,
  Bill Kim

Symptoms are silent wrong answers + EXCLUDE bypass leans me toward back-patching despite the reindex cost.
Happy with whatever call you make.