issue with sinhf

Submitted by Szabolcs Nagy on Feb. 5, 2021, 8:12 p.m.

Details

Message ID 20210205201229.GD2447252@port70.net
State New
Series "issue with sinhf"
Headers show

Commit Message

Szabolcs Nagy Feb. 5, 2021, 8:12 p.m.
* Szabolcs Nagy <nsz@port70.net> [2021-02-05 21:09:34 +0100]:

> * Paul Zimmermann <Paul.Zimmermann@inria.fr> [2021-02-05 09:01:09 +0100]:
> > $ cat test_sinh_musl.c
> > #include <stdio.h>
> > #include <stdlib.h>
> > #include <math.h>
> > 
> > int
> > main ()
> > {
> >   float x = 0x1.62e4p+6;
> >   float y = sinhf (x);
> >   printf ("x=%a y=%a\n", x, y);
> > }
> ...
> > $ gcc -fno-builtin test_sinh_musl.c $FILES
> > $ ./a.out
> > x=0x1.62e4p+6 y=-nan
> 
> this seems to be a bug, attaching a fix

that one was for the acoshf bug, the sinhf issue was a bug
in expm1f (overflowed too early, although sinhf could have
used a less tight threshold that would have prevented this
causing an issue there)

Patch hide | download patch | download mbox

From 09f8837310082aeee019828b9aa805cf4528b3d1 Mon Sep 17 00:00:00 2001
From: Szabolcs Nagy <nsz@port70.net>
Date: Fri, 5 Feb 2021 19:51:36 +0000
Subject: [PATCH] math: fix expm1f overflow threshold

the threshold was wrong so expm1f overflowed to inf a bit too early
and on most targets uint32_t compare is faster than float compare so
use that.

this also fixes sinhf incorrectly returning nan for some values where
the internal expm1f overflowed.
---
 src/math/expm1f.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/src/math/expm1f.c b/src/math/expm1f.c
index 297e0b44..09a41afe 100644
--- a/src/math/expm1f.c
+++ b/src/math/expm1f.c
@@ -16,7 +16,6 @@ 
 #include "libm.h"
 
 static const float
-o_threshold = 8.8721679688e+01, /* 0x42b17180 */
 ln2_hi      = 6.9313812256e-01, /* 0x3f317180 */
 ln2_lo      = 9.0580006145e-06, /* 0x3717f7d1 */
 invln2      = 1.4426950216e+00, /* 0x3fb8aa3b */
@@ -41,7 +40,7 @@  float expm1f(float x)
 			return x;
 		if (sign)
 			return -1;
-		if (x > o_threshold) {
+		if (hx > 0x42b17217) { /* x > log(FLT_MAX) */
 			x *= 0x1p127f;
 			return x;
 		}
-- 
2.29.2


Comments

Paul Zimmermann Feb. 6, 2021, 6:34 a.m.
Dear Szabolcs,

> that one was for the acoshf bug, the sinhf issue was a bug
> in expm1f (overflowed too early, although sinhf could have
> used a less tight threshold that would have prevented this
> causing an issue there)
> 
> 
> [2:text/x-diff Show Save:0001-math-fix-expm1f-overflow-threshold.patch (1kB)]

and that one fixes the sinhf issue:

GLIBC :
$ gcc -fno-builtin test_sinh_musl.c -lm
$ ./a.out
x=0x1.62e4p+6 y=0x1.ffe808p+126

musl-1.2.2 + above patch:
$ gcc -fno-builtin test_sinh_musl.c $FILES
$ ./a.out
x=0x1.62e4p+6 y=0x1.ffe808p+126

Thank you!
Paul