gcc -O2 produces bad code for function given below. Correct
code is generated for no optimization or -O.

The problem is that with -O2 the 3rd param passed to fprintf
is uninitialized garbage, not the bytes set by the assignment
to T0.

More presisely, the bad code generated is:
main:
        pushl   %ebp
        movl    %esp, %ebp
        subl    $12, %esp
        movl    stderr, %eax
        movl    -4(%ebp), %edx
        movl    $0x3fa00000, -4(%ebp)
        pushl   %edx
        pushl   $.LC0
        pushl   %eax
        call    fprintf
        movl    %ebp, %esp
        popl    %ebp
        ret

That is, %edx is set prior to the initialization of -4(%ebp).
And %edx is then passed to fprintf, causing garbage to be
printed.

The uuencoded, gzipped, preprocessed file is attached.

This error also occurs using CodeSourcery's Online Test
Compiler. The code is slightly different than what I
see using 3.0, but it is still wrong.

Release:
gcc version 3.0

Environment:
i686-pc-linux-gnu

Configured with: /export/gnu/gcc-3.0/configure --enable-languages=c++
Thread model: single

linux 2.4.4 running on rh7.1 (Pentium III)

How-To-Repeat:
#include <stdio.h>

main()
{
  float T0;
  T0=1.25;
  fprintf(stderr, "%x\n", *(unsigned int*)&T0);
}