Hi,
On Wed, Mar 11, 2026 at 03:07:10PM +0100, Peter Eisentraut wrote:
> On 11.03.26 14:43, Andres Freund wrote:
> > Hi,
> >
> > On 2026-03-11 11:52:28 +0000, Bertrand Drouvot wrote:
> > > while working on a patch (not shared yet), I had issues with
> > > floating-point and
> > > realized that we don't defend against -ffast-math in meson builds. We
> > > defend
> > > against in autoconf (because we don't want fast-math optimizations [1]),
> > > so the
> > > attached does the same for meson.
> >
> > I don't find this (nor the autoconf check) particularly likely to be
> > helpful. The most likely way the flag would unintentionally be added would
> > be
> > via the cflags of some dependency - which won't be picked up by the tests.
> >
> > However we do have tests during the builds that should pick it up, in date.c
> > and timestamp.c...
>
> The existing check in configure is because certain Linux distributions used
> to compile everything with -ffast-math to be "faster", and that kept
> breaking PostgreSQL and so we wanted to stop them very early. These are
> gone, and the defenses in the code like date.c should be sufficient for any
> new attempts. I think we could remove the check in configure.
Oh, I did not see that we also have __FAST_MATH__ check in some C files. The
reason
is that I build with:
CFLAGS="-ffast-math" CC="gcc" meson setup meson_build and -Dc_args="-Og -ggdb"
and that did not produce the error at compilation time.
But I got some regression tests failing like:
--- /home/postgres/postgresql/postgres/src/test/regress/expected/numeric.out
+++
/home/postgres/postgresql/postgres/meson_build/testrun/regress/regress/results/numeric.out
@@ -1630,22 +1630,7 @@
(0, 0, 1, 2147483647),
(1, 1, 0, 2147483647)
) as sample(oper, low, high, cnt);
- oper | low | high | cnt | width_bucket
--------------+-------------+-------------+------------+--------------
- 10.5 | -1.797e+308 | 1.797e+308 | 1 | 1
- 10.5 | -1.797e+308 | 1.797e+308 | 2 | 2
- 10.5 | -1.797e+308 | 1.797e+308 | 3 | 2
- 4.4925e+307 | -8.985e+307 | 8.985e+307 | 10 | 8
- 10.5 | 1.797e+308 | -1.797e+308 | 1 | 1
- 10.5 | 1.797e+308 | -1.797e+308 | 2 | 2
- 10.5 | 1.797e+308 | -1.797e+308 | 3 | 2
- 4.4925e+307 | 8.985e+307 | -8.985e+307 | 10 | 3
- 0 | 0 | 5e-324 | 4 | 1
- 5e-324 | 0 | 5e-324 | 4 | 5
- 0 | 0 | 1 | 2147483647 | 1
- 1 | 1 | 0 | 2147483647 | 1
-(12 rows)
-
+ERROR: "0.00000000..<LOT OF 0>..000005" is out of range for type double
precision
or
--- /home/postgres/postgresql/postgres/src/test/regress/expected/money.out
+++
/home/postgres/postgresql/postgres/meson_build/testrun/regress/regress/results/money.out
@@ -536,7 +536,7 @@
SELECT '92233720368547758.07'::money * 2::float8;
ERROR: money out of range
SELECT '-1'::money / 1.175494e-38::float4;
-ERROR: money out of range
+ERROR: "0.00000000000000000000000000000000000001175494" is out of range for
type real
But if I use:
CC="gcc" meson setup meson_build and -Dc_args="-Og -ggdb -ffast-math" then the
error happens at compilation time.
The check added in the patch would also catch the CFLAGS="-ffast-math" case.
Regards,
--
Bertrand Drouvot
PostgreSQL Contributors Team
RDS Open Source Databases
Amazon Web Services: https://aws.amazon.com
