Re: strtod ("nan") returns negative NaN

[Masamichi Hosoda <trueroad@trueroad.jp>]

> On Aug 14 11:56, Corinna Vinschen wrote:
>> On Aug 14 13:45, Masamichi Hosoda wrote:
>> > >From a50ee5a4747a99c70469a53fe959f3dc22d3b79a Mon Sep 17 00:00:00 2001
>> > From: Masamichi Hosoda <trueroad@trueroad.jp>
>> > Date: Tue, 14 Aug 2018 12:50:32 +0900
>> > Subject: [PATCH] Fix strtod ("nan") returns qNaN
>> > 
>> > The definition of qNaN for x86_64 and x86 was wrong.
>> > So strtod ("nan") returned sNaN instead of qNaN.
>> > 
>> > Furthermore, it was inverted the sign bit with the presence of `-` character.
>> > So strtod ("-nan") returned qNaN.
>> > 
>> > This commit fixes definition of qNaN
>> > and removes the sign bit inversion when evaluating "nan".
>> > ---
>> >  newlib/libc/stdlib/gd_qnan.h | 8 ++++----
>> >  newlib/libc/stdlib/strtod.c  | 1 +
>> >  2 files changed, 5 insertions(+), 4 deletions(-)
>> 
>> Can you please send this patch to the newlib AT sourceware DOT org
>> mailing list?  As soon as something in newlib gets changed, a lot of
>> other targets are affected and the guys working on those targets should
>> have a chance to chime in.

I'll improve the patch and send it.
It did not consider environments excluding x86 and x86_64.

> Looks like strtold is affected as well, just differently:
> 
>   printf ("strtod (\"nan\", NULL) = %f\n", strtod ("nan", NULL));
>   printf ("strtod (\"-nan\", NULL) = %f\n", strtod ("-nan", NULL));
>   printf ("strtold (\"nan\", NULL) = %Lf\n", strtold ("nan", NULL));
>   printf ("strtold (\"-nan\", NULL) = %Lf\n", strtold ("-nan", NULL));
>   printf ("nan (\"\") = %f\n", nan (""));
> 
> ==>
> 
>   strtod ("nan", NULL) = -nan
>   strtod ("-nan", NULL) = nan
>   strtold ("nan", NULL) = -nan
>   strtold ("-nan", NULL) = -nan
>   nan ("") = nan
> 
> so it prints always -nan.
> 
> With your patch, strtold looks more correct, but it still prints the
> sign of NaN:
> 
>   strtod ("nan", NULL) = nan
>   strtod ("-nan", NULL) = nan
>   strtold ("nan", NULL) = nan
>   strtold ("-nan", NULL) = -nan
>   nan ("") = nan
> 
> Question: What's wrong with that?  Wouldn't it be more correct if
> strtod returns -NaN for "-nan" as well?

In my investigate,
strtold sets sign bit when parameter has '-' character.
The wrong long double NaN definition is negative NaN that is set sign bit.
So without my patch, both strtold ("nan") and
strtold ("-nan") return negative NaN.

On the other hand, strtod inverts the sign when parameter has '-' character.
The wrong double NaN definition is negative NaN.
So without my patch, strtod ("nan") returns negative NaN
and strtod ("-nan") returns positive NaN.

My previous patch removes the sign inversion in strtod when NaN.
But I did not fix strtold.

Perhaps, the following patch removes the sign bit setting of strtold when NaN.

```
--- a/newlib/libc/stdlib/strtodg.c
+++ b/newlib/libc/stdlib/strtodg.c
@@ -585,6 +585,7 @@ _strtodg_l (struct _reent *p, const char *s00, char **se, FPI *fpi, Long *exp,
 					if (*s == '(') /*)*/
 						irv = hexnan(&s, fpi, bits);
 #endif
+					sign = 0;
 					goto infnanexp;
 					}
 			  }
```

--
Problem reports:       http://cygwin.com/problems.html
FAQ:                   http://cygwin.com/faq/
Documentation:         http://cygwin.com/docs.html
Unsubscribe info:      http://cygwin.com/ml/#unsubscribe-simple