[PATCH] math: fix expf(-NAN) to return -NAN instead of 0

[PATCH] math: fix expf(-NAN) to return -NAN instead of 0

[Szabolcs Nagy]

expf(-NAN) was treated as expf(-large) which unconditionally
returns +0, so special case +-NAN.

reported by Petr Hosek.
---
 src/math/expf.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/src/math/expf.c b/src/math/expf.c
index 16e9afe..4a742e4 100644
--- a/src/math/expf.c
+++ b/src/math/expf.c
@@ -39,6 +39,8 @@ float expf(float x)
 
 	/* special cases */
 	if (hx >= 0x42aeac50) {  /* if |x| >= -87.33655f or NaN */
+		if (hx > 0x7f800000) /* +-NaN */
+			return x;
 		if (hx >= 0x42b17218 && !sign) {  /* x >= 88.722839f */
 			/* overflow */
 			x *= 0x1p127f;
-- 
2.7.0

Re: [PATCH] math: fix expf(-NAN) to return -NAN instead of 0

[Szabolcs Nagy]

* Szabolcs Nagy <nsz@port70.net> [2016-03-04 22:37:53 +0100]:
> expf(-NAN) was treated as expf(-large) which unconditionally
> returns +0, so special case +-NAN.
> 
> reported by Petr Hosek.
> ---
>  src/math/expf.c | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/src/math/expf.c b/src/math/expf.c
> index 16e9afe..4a742e4 100644
> --- a/src/math/expf.c
> +++ b/src/math/expf.c
> @@ -39,6 +39,8 @@ float expf(float x)
>  
>  	/* special cases */
>  	if (hx >= 0x42aeac50) {  /* if |x| >= -87.33655f or NaN */

in the comment -87... should be 87...
but i didnt want to include that into the bug fix

> +		if (hx > 0x7f800000) /* +-NaN */
> +			return x;
>  		if (hx >= 0x42b17218 && !sign) {  /* x >= 88.722839f */
>  			/* overflow */
>  			x *= 0x1p127f;
> -- 
> 2.7.0

Re: [PATCH] math: fix expf(-NAN) to return -NAN instead of 0

[Szabolcs Nagy]

[-- Attachment #1: Type: text/plain, Size: 147 bytes --]

attached a better version with exp2f fix too.
as far as i can tell other functions handle -nan correctly,
but i will have to do some -nan testing.

[-- Attachment #2: 0001-math-fix-expf-NAN-and-exp2f-NAN-to-return-NAN-instea.patch --]
[-- Type: text/x-diff, Size: 1252 bytes --]

From d805064900932fdd47f119f0932680a93184cca6 Mon Sep 17 00:00:00 2001
From: Szabolcs Nagy <nsz@port70.net>
Date: Fri, 4 Mar 2016 21:23:33 +0000
Subject: [PATCH] math: fix expf(-NAN) and exp2f(-NAN) to return -NAN instead
 of 0

expf(-NAN) was treated as expf(-large) which unconditionally
returns +0, so special case +-NAN.
reported by Petr Hosek.
---
 src/math/exp2f.c | 2 ++
 src/math/expf.c  | 2 ++
 2 files changed, 4 insertions(+)

diff --git a/src/math/exp2f.c b/src/math/exp2f.c
index cf6126e..296b634 100644
--- a/src/math/exp2f.c
+++ b/src/math/exp2f.c
@@ -91,6 +91,8 @@ float exp2f(float x)
 	/* Filter out exceptional cases. */
 	ix = u.i & 0x7fffffff;
 	if (ix > 0x42fc0000) {  /* |x| > 126 */
+		if (ix > 0x7f800000) /* NaN */
+			return x;
 		if (u.i >= 0x43000000 && u.i < 0x80000000) {  /* x >= 128 */
 			x *= 0x1p127f;
 			return x;
diff --git a/src/math/expf.c b/src/math/expf.c
index 16e9afe..feee2b0 100644
--- a/src/math/expf.c
+++ b/src/math/expf.c
@@ -39,6 +39,8 @@ float expf(float x)
 
 	/* special cases */
 	if (hx >= 0x42aeac50) {  /* if |x| >= -87.33655f or NaN */
+		if (hx > 0x7f800000) /* NaN */
+			return x;
 		if (hx >= 0x42b17218 && !sign) {  /* x >= 88.722839f */
 			/* overflow */
 			x *= 0x1p127f;
-- 
2.7.0

Re: [PATCH] math: fix expf(-NAN) to return -NAN instead of 0

[Rich Felker]

On Fri, Mar 04, 2016 at 11:06:32PM +0100, Szabolcs Nagy wrote:
> attached a better version with exp2f fix too.
> as far as i can tell other functions handle -nan correctly,
> but i will have to do some -nan testing.

FYI I committed this.

Rich