Incompatibility between gcc and glibc

[M.Schimschak]

I already submitted this report to bug-gcc@prep.ai.mit.edu
(gnu.gcc.bug), but didn't get any feedback yet. Since this bug also
applies to egcs-1.0.1 I send it to you as well. 
Here's the program which is compiled with errors by several gcc based
compilers when used with optimization, basically when inlining is
allowed (see below):

-------------------------------------------------------------------------------

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


int main(void)
{
  double     c=0;
  double     d=0;
  double     e[2];


  e[0] = 1.0;
  e[1] = 1.0;

  
  printf("%5g %5g   %10g %10g\n",e[0], e[1], c, d);

  c = atan2(e[0], e[1]);
  d = e[0] * e[0] + e[1] * e[1];

  printf("%5g %5g   %10g %10g\n",e[0], e[1], c, d);

  return 0;
}

-------------------------------------------------------------------------------
This program is supposed to produce the following output:

"    1     1                 0               0
     1     1      0.7853981634               2"

Instead it produces the following output with several compilers and
several optimization levels:

"    1     1                 0               0
     1     1      0.7853981634     1.785398163"
                                   ^^^^^^^^^^^

It seems that there is some kind of incompatibility between gcc and
glibc. I traced the bug down to the inline optimization. On a glibc
(2.0.6) based Red Hat Linux system, gcc-2.7.2.3 produces erroneous
code when used with -O2. On the same system  egcs-2.90.23 980102
(egcs-1.0.1 release) produces erroneous code when used with -O1.
In both cases the compilers produce correct code when used with
-fno-inline. In the case where the -fno-inline flag is used the
linking of libm is necessary, whereas in the case where inlining is
allowed the code for the atan2 function call is taken from
"/usr/include/__math.h" which is included by math.h. Here's the code
block from __math.h, maybe it will help (unfortunately I have no idea
what is going on there ;-)

-----------------------------------------------------------------------

__MATH_INLINE double atan2 (double __y, double __x);
__MATH_INLINE double
atan2 (double __y, double __x)
{
  register double __value;
  __asm __volatile__
    ("fpatan\n\t"
     "fldl %%st(0)"
     : "=t" (__value) : "0" (__x), "u" (__y));

  return __value;
}

----------------------------------------------------------------------

Systems that I checked so far involve 
   Red Hat 4.2 Linux system (libc 5.3.12, i486) -- no problems
       gcc version 2.7.2.1 -- no problems

  Red Hat 5.0 Linux system (glibc 2.0.6, i586) -- all with erroneous code
       gcc version 2.7.2.3
       gcc version 2.8.0
       gcc version 2.8.1
       egcs-2.90.23 980102 (egcs-1.0.1 release) (erroneous code for
                                                 -O1, save for -O0 or -O2)
       pgcc-2.91.04 980115 (gcc-2.8.0 release)

  Dec Alpha system  (Digital UNIX V4.0B, ?, alpha) -- no problems
       gcc-2.8.0

It would be interesting to know if the error also occurs on Dec Alphas 
when used with Red Hat 5.0 Linux.


Hope this helps for tracing the bug,

        Martin