[pedretti@roguewave.com] libc/2072: sscanf fails to correctly parse certain floating point numbers

[Andreas Jaeger <aj at suse dot de>]

We received the appended bug report.  First analysis shows that
strtold is broken.  The output of the program is:

ldval = 3.000000, ret = 1
ldval = 0.000000, ret = 1
ldval = 0.000000, errno: Success

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
#define _GNU_SOURCE 1
#include <errno.h>
#include <stdio.h>
#include <stdlib.h>

int
main (void)
{
  long double ldval = 0;
  int ret;
  
  ret = sscanf("30000000000000000000.00e-19", "%Lf", &ldval);
  printf("ldval = %Lf, ret = %d\n", ldval, ret);

  ret = sscanf("300000000000000000000.00e-20", "%Lf", &ldval);
  printf("ldval = %Lf, ret = %d\n", ldval, ret);

  errno = 0;
  ldval = strtold ("300000000000000000000.00e-20", NULL);
  printf("ldval = %Lf, errno: %m\n", ldval);
  
  return 0;
}
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Any ideas?

Andreas

Subject: Topics

Topics:
   libc/2072: sscanf fails to correctly parse certain floating point numbers

---

• To: libc-gnats at gnu dot org, gnats-admin at gnu dot org
• Subject: libc/2072: sscanf fails to correctly parse certain floating point numbers
• From: pedretti at roguewave dot com
• Date: Tue, 06 Feb 2001 14:26:18 -0500

>Number:         2072
>Category:       libc
>Synopsis:       sscanf fails to correctly parse certain floating point numbers
>Confidential:   yes
>Severity:       serious
>Priority:       high
>Responsible:    libc-gnats
>State:          open
>Quarter:        
>Keywords:       
>Class:          sw-bug
>Submitter-Id:   gnatsweb
>Arrival-Date:   Tue Feb 06 14:26:17 -0500 2001
>Cases:          
>Originator:     John A. Pedretti
>Release:        GNU C Library stable release version 2.1.1, by Roland McGrath et al.
>Organization:
Rogue Wave Software, Inc
Boulder CO
>Environment:
Linux 2.2.5-15smp #1 SMP Mon Apr 19 22:43:28 EDT 1999 i686 unknown
>Description:
sscanf fails to correctly parse floating point numbers of the form: a digit from 1-9, followed by N occurrences of the digit 0, followed by e-N, where N>20 (i.e., the value of the whole number is equal to the initial digit). 

float, double, and long double are all affected, but for different values of N: 

float fails for the range 31 <= N <= FLT_MIN_EXP
double fails for the range 180 <= N <= DBL_MIN_EXP
long double fails for the range 20 <= N <= LDBL_MIN_EXP

In most cases, the value scanned in is 0.0, but in some cases it is "inf".

The attached file contains a program that exercises the bug for all meaningful values of N.
>How-To-Repeat:
Example: 

#include <stdio.h>
int main() {
    long double ldval = 0;

    sscanf("30000000000000000000.00e-19", "%Lf", &ldval);
    printf("ldval = %Lf\n", ldval);

    sscanf("300000000000000000000.00e-20", "%Lf", &ldval);
    printf("ldval = %Lf\n", ldval);
}
--- output ---
ldval = 3.000000
ldval = 0.000000
--------------
>Fix:
Unknown
>Unformatted:
 

---

-- 
 Andreas Jaeger
  SuSE Labs aj@suse.de
   private aj@arthur.inka.de
    http://www.suse.de/~aj