#367 fpclassify/isnormal can't detect denormalised floating point numbers

[Peter Bex]

A simple program which prints the value of isnormal(4.9406564584124654417657e-324) shows 1 (also if that value is assigned to a variable declared "double" and that variable is tested):

#include <stdio.h>
#include <math.h>

int main(void) {
  double x = 4.9406564584124654417657e-324;
  printf("Normal? %d\n" isnormal(x));
  printf("Normal? %d\n" isnormal(4.9406564584124654417657e-324));
}

The program should show "Normal? 0" twice. The denormalised number has been taken from the ATF tests for fpclassify/isnormal from the NetBSD source tree.

More denormalised numbers from that test suite:

1.1125369292536006915451e-308
-5.5626846462680034577256e-309
4.9406564584124654417657e-324

This has been reported for MingW as well, and hasn't been fixed there yet either: http://sourceforge.net/p/mingw/bugs/1927/

[Kai Tietz]

I tested your testcase on linux using glibc (libm), and it produces same result as we do on mingw-w64's trunk.
In both cases isnormal is 0 (as to be expected).

[Kai Tietz]

IMHO is testcase invalid.

[Kai Tietz]

So, found the issue. It affects 32-bit only.
For 32-bits result is 1, for 64-bit result is 0. So the bug seems to be on 32-bit side.

[FX]

The 32-bit code, for double, is in effect the following:

#include <stdio.h>

#define FP_NAN      0x0100
#define FP_NORMAL   0x0400
#define FP_ZERO     0x4000
int __fpclassify (double x)
{
  unsigned short sw;
  __asm__ __volatile__ ("fxam; fstsw %%ax;" : "=a" (sw): "t" (x));
  return sw & (FP_NAN | FP_NORMAL | FP_ZERO );
}

int main (void)
{
  double x = 4.9406564584124654417657e-324;
  printf("%d\n", __fpclassify(x) == FP_NORMAL);
}

It might make sense to use (when available) GCC's type-generic __builtin_fpclassify (https://gcc.gnu.org/onlinedocs/gcc/Other-Builtins.html) instead of the current __fpclassify in mingw-w64.

[Sisyphus]

With my 32-bit mingw-w64 port of gcc-8.1.0, I still experience this issue iff optimization is enabled.
There's no problem with the 64-bit port.

Here's the demo :

#include <stdio.h>
#include <math.h>

int main(void) {
  double x = 1e-320;
  if(isnormal(x)) printf("wtf\n");
  else printf("ok\n");
  printf("%x\n", fpclassify(x));
  printf("%x\n", __fpclassify(x));
/* printf("%x\n", __builtin_fpclassify(x)); *//* error */
}

When built without optimization it correctly outputs:

ok
4400
4400

But when built with O1 (or higher) optimization it outputs:

wtf
400
400

    I was curious to see how __builtin_fpclassify fared but apparently
    it's not a drop-in replacement for fpclassify.
    The error produced was "too few arguments to function __builtin_fpclassify".

Cheers,
Rob