According to the man page, -mpreferred-stack-boundary allows one to specify what the stack is aligned at.
"Attempt to keep the stack boundary aligned to a 2 raised to num byte boundary.  If -mpreferred-stack-boundary is not specified, the default is 4 (16 bytes or 128 bits)."

Example:
# gcc -mpreferred-stack-boundary=7  -mstackrealign testme.c -o ttt3.s -S
In ttt3.s you will see:
main:
        leal    4(%esp), %ecx
        andl    $-16, %esp
        pushl   -4(%ecx)

Expected:
main:
        leal    4(%esp), %ecx
        andl    $-128, %esp
        pushl   -4(%ecx)

No matter which value is passed to -mpreferred-stack-boundary (if it's larger than 4), it will always align the stack to 16 bytes.

Fix (diff according to svn):
diff -urN gcc/gcc/config/i386/i386.c stack-boundary-fix/gcc/config/i386/i386.c
--- gcc/gcc/config/i386/i386.c  2007-09-19 00:37:00.000000000 +0200
+++ stack-boundary-fix/gcc/config/i386/i386.c   2007-09-19 01:06:57.000000000 +0200
@@ -6242,7 +6242,7 @@
 
       /* Align the stack.  */
       emit_insn (gen_andsi3 (stack_pointer_rtx, stack_pointer_rtx,
-                            GEN_INT (-16)));
+                            GEN_INT (-(cfun->preferred_stack_boundary/BITS_PER_UNIT))));
 
       /* And here we cheat like madmen with the unwind info.  We force the
         cfa register back to sp+4, which is exactly what it was at the