AL vs EAX optimizer bug in egcs-1.1.2 on x86

["Scott McPeak" <smcpeak at cs dot berkeley dot edu>]

(I don't know if anyone cares about tracking egcs bugs anymore.  But I
still use it, and I couldn't find a list of known bugs, so ...)

NOTE: The bug I'm reporting is *not* present in gcc-2.95.2.


Version: 
  gcc version egcs-2.91.66 19990314/Linux (egcs-1.1.2 release)
System: 
  Linux leetch 2.2.13 #2 Wed May 24 18:36:44 PDT 2000 i686 unknown


Source code:

    /* optimbug.c */
    int main()
    {
      int x = 0;     /* start with all bits 0 */

      /* little endian: sets x to 1 */
      /* big endian: sets x to a large number */
      *((char*)(&x)) = 1;

      if (x == 1) {
        exit(0);     /* success will mean little endian */
      }
      else {
        exit(1);     /* big endian */
      }
    }


If this code is compiled by
  gcc optimbug.c
then a.out will exit with status code 0 on my little-endian machine.

But if the code is compiled by
  gcc -O2 optimbug.c
then a.out will exit with status code 1!


here is assembly dump of egcs-1.1.2 optimized version:

    Dump of assembler code for function main:
    0x80483b4 <main>:       push   %ebp
    0x80483b5 <main+1>:     mov    %esp,%ebp
    0x80483b7 <main+3>:     mov    $0x1,%al     <-- the bug is 'al'
    0x80483b9 <main+5>:     cmp    $0x1,%eax
    0x80483bc <main+8>:     jne    0x80483c8 <main+20>
    0x80483be <main+10>:    push   $0x0
    0x80483c0 <main+12>:    call   0x8048308 <exit>
    0x80483c5 <main+17>:    lea    0x0(%esi),%esi
    0x80483c8 <main+20>:    push   $0x1
    0x80483ca <main+22>:    call   0x8048308 <exit>
    End of assembler dump.

and here is an assembly dump of gcc-2.95.2 optimized (where
the problem does *not* occur):

    Dump of assembler code for function main:
    0x80483f0 <main>:       push   %ebp
    0x80483f1 <main+1>:     mov    %esp,%ebp
    0x80483f3 <main+3>:     sub    $0x8,%esp
    0x80483f6 <main+6>:     mov    $0x1,%eax    <-- correctly uses 'eax'
    0x80483fb <main+11>:    cmp    $0x1,%eax
    0x80483fe <main+14>:    jne    0x8048410 <main+32>
    0x8048400 <main+16>:    add    $0xfffffff4,%esp
    0x8048403 <main+19>:    push   $0x0
    0x8048405 <main+21>:    call   0x8048308 <exit>
    0x804840a <main+26>:    lea    0x0(%esi),%esi
    0x8048410 <main+32>:    add    $0xfffffff4,%esp
    0x8048413 <main+35>:    push   $0x1
    0x8048415 <main+37>:    call   0x8048308 <exit>
    End of assembler dump.


Do what you will with this report.  I've already got a workaround
(make 'x' global).

-Scott