From Bugzilla Helper:
User-Agent: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.0.0) Gecko/20020618

Description of problem:
It is possible to write a trivial C program which causes a floating point
exception using one summation.

Version-Release number of selected component (if applicable):


How reproducible:
Always

Steps to Reproduce:
Compile and link these programs with "gcc -o foo foo.c".  I've tried many other
optimization and debug settings, with no change in results.

This C code works correctly:
#include <stdio.h>
int main(void)
{
  printf("sum=%g\n", 1.21678e-308 + 1.07381e-308);
  return 0;
}

This C code dies with a floating point exception:
#include <stdio.h>
int main(void)
{
  double hi, lo, sum;
  hi = 1.21678e-308;
  lo = 1.07381e-308;
  sum = (hi + lo);            /* Produces bad result. */
  printf( "sum=%g\n", sum);   /* Dies here. */
  return 0;
}


Actual Results:  The first program works correctly.  The second does not.

Expected Results:  Both should have worked correctly. =-)

Additional info:

For both of these programs, compiling with Compaq's ccc produces underflow
warnings but the executable works correctly (no fp exceptions and prints the
expected "sum=2.29059e-308".

gcc2.96 fails, as does gcc3.1 (built from sources).

We've messed with this on 3 Alpha machines.  Two have dual 667MHz 21264 cpus
(DS-20 clones from Microway, 4MB cache), and one is a single cpu Compaq ES-40
(8MB cache).

We've fought with and tried to work around this for quite a while, and have
finally distilled it to this form.  Any help (workarounds, bug fixes, etc) would
be greatly appreciated, as this has cost us a lot of time.  I'd be happy to
provide assembly dumps or run experiments as needed.