From: Andrey Rachitskiy Date: Tue, 11 Aug 2026 19:53:33 +0500 Subject: [PATCH] Fix covar_pop/regr_sxy for Inf with a constant other input Commit 6498287696d tracks commonX/commonY so that Sxx, Syy, and Sxy stay exactly zero while a column is still constant. That fixed roundoff problems in corr() and friends. It also skipped the Youngs-Cramer update of Sxy whenever either side was still constant. If the other argument later becomes Inf or NaN, the Youngs-Cramer update we skip would have been the 0*Inf / 0*NaN product that yields NaN. Sxy then remains 0, so covar_pop(), covar_samp(), and regr_sxy() return 0 instead of NaN. Inf in the first row is still fine, because the first-input path already forces Sxy to NaN. Force Sxy to NaN in that short-circuit path when either new input is Inf or NaN, matching the first-input handling. Finite constant inputs still produce exact zero. Reported-by: Junwen An Author: Andrey Rachitskiy --- src/backend/utils/adt/float.c | 7 +++++++ src/test/regress/expected/aggregates.out | 22 ++++++++++++++++++++++ src/test/regress/sql/aggregates.sql | 12 ++++++++++++ 3 files changed, 41 insertions(+) diff --git a/src/backend/utils/adt/float.c b/src/backend/utils/adt/float.c index 4c2ccdfbf3f..aef03cfebb1 100644 --- a/src/backend/utils/adt/float.c +++ b/src/backend/utils/adt/float.c @@ -3449,6 +3449,10 @@ float8_regr_accum(PG_FUNCTION_ARGS) * etc). Updating them would just create the possibility of injecting * roundoff error, and we need exact zero results so that the final * functions will return NULL in the right cases. + * + * Inf/NaN inputs must still force Sxy to NaN when the other variable + * is constant and the Sxy update is skipped. Otherwise we leave a + * false exact-zero Sxy. The first-input path below does the same. */ if (isnan(commonX)) Sxx += tmpX * tmpX * scale; @@ -3456,6 +3460,9 @@ float8_regr_accum(PG_FUNCTION_ARGS) Syy += tmpY * tmpY * scale; if (isnan(commonX) && isnan(commonY)) Sxy += tmpX * tmpY * scale; + else if (isnan(newvalX) || isinf(newvalX) || + isnan(newvalY) || isinf(newvalY)) + Sxy = get_float8_nan(); /* * Overflow check. We only report an overflow error when finite diff --git a/src/test/regress/expected/aggregates.out b/src/test/regress/expected/aggregates.out index 5f0668382ed..82666a729cc 100644 --- a/src/test/regress/expected/aggregates.out +++ b/src/test/regress/expected/aggregates.out @@ -515,6 +515,28 @@ SELECT covar_pop(1::float8,'nan'::float8), covar_samp(3::float8,'nan'::float8); NaN | (1 row) +-- Inf/NaN not first, other arg constant: must still yield NaN, not 0 +CREATE TEMP TABLE regr_inf_pos (ord int, y float8); +INSERT INTO regr_inf_pos VALUES (1, 3), (2, 'Infinity'), (3, 4); +SELECT covar_pop(0::float8, y ORDER BY ord), + covar_pop(y, 0::float8 ORDER BY ord), + regr_sxy(0::float8, y ORDER BY ord) + FROM regr_inf_pos; + covar_pop | covar_pop | regr_sxy +-----------+-----------+---------- + NaN | NaN | NaN +(1 row) + +DELETE FROM regr_inf_pos; +INSERT INTO regr_inf_pos VALUES (1, 3), (2, 'NaN'), (3, 4); +SELECT covar_pop(0::float8, y ORDER BY ord) FROM regr_inf_pos; + covar_pop +----------- + NaN +(1 row) + +DROP TABLE regr_inf_pos; + -- check some cases that formerly had poor roundoff-error behavior -- note: regr_r2() differs from corr() for a horizontal line, per spec SELECT corr(0.09, g), regr_r2(0.09, g) diff --git a/src/test/regress/sql/aggregates.sql b/src/test/regress/sql/aggregates.sql index b788152f0c2..3c615f7dc60 100644 --- a/src/test/regress/sql/aggregates.sql +++ b/src/test/regress/sql/aggregates.sql @@ -140,6 +140,18 @@ SELECT covar_pop(1::float8,2::float8), covar_samp(3::float8,4::float8); SELECT covar_pop(1::float8,'inf'::float8), covar_samp(3::float8,'inf'::float8); SELECT covar_pop(1::float8,'nan'::float8), covar_samp(3::float8,'nan'::float8); +-- Inf/NaN not first, other arg constant: must still yield NaN, not 0 +CREATE TEMP TABLE regr_inf_pos (ord int, y float8); +INSERT INTO regr_inf_pos VALUES (1, 3), (2, 'Infinity'), (3, 4); +SELECT covar_pop(0::float8, y ORDER BY ord), + covar_pop(y, 0::float8 ORDER BY ord), + regr_sxy(0::float8, y ORDER BY ord) + FROM regr_inf_pos; +DELETE FROM regr_inf_pos; +INSERT INTO regr_inf_pos VALUES (1, 3), (2, 'NaN'), (3, 4); +SELECT covar_pop(0::float8, y ORDER BY ord) FROM regr_inf_pos; +DROP TABLE regr_inf_pos; + -- check some cases that formerly had poor roundoff-error behavior -- note: regr_r2() differs from corr() for a horizontal line, per spec SELECT corr(0.09, g), regr_r2(0.09, g) -- 2.53.0