optimization/1762: treatment of NaNs violates IEEE 754

[reichelt at igpm dot rwth-aachen dot de]

>Number:         1762
>Category:       optimization
>Synopsis:       treatment of NaNs violates IEEE 754
>Confidential:   no
>Severity:       non-critical
>Priority:       medium
>Responsible:    unassigned
>State:          open
>Class:          wrong-code
>Submitter-Id:   net
>Arrival-Date:   Wed Jan 24 10:56:00 PST 2001
>Closed-Date:
>Last-Modified:
>Originator:     Volker Reichelt
>Release:        gcc 2.95, 2.96.20000920 on IRIX6.5
>Organization:
>Environment:
SGI O2, IRIX 6.5
CPU: MIPS R5000 Processor Chip Revision: 2.1
FPU: MIPS R5000 Floating Point Coprocessor Revision: 1.0
>Description:
The following program shows some problems with the
treatment of NaNs. According to IEEE 754 the result
of "0 < NaN" should be false. Therefore the result
of "less" and the return value of the program should
be "0". This is the case, if I compile the program
without optimization. With optimization enabled,
however, I get the result "1" instead.
(The same problem arises, when "a<b?b:a" is evaluated.)


double less(double a, double b)
{
  return a < b;
}

int main()
{
  const double NaN = 0.0/0.0;
  return (int)less(0.0,NaN);
}
>How-To-Repeat:
gcc -O1 NaN.cpp; ./a.out; echo $?
(assuming the file is called NaN.cpp)
>Fix:

>Release-Note:
>Audit-Trail:
>Unformatted: