For a function with no prototype that defines an argument as unsigned
int, GCC for powerpc64-linux with -m64 and optimization uses the
entire 64-bit argument register rather than masking off the lower
32 bits.  This problem was discoverd with 176.gcc using options
"-m64 -O2 -ftree-vectorize -maltivec -fdata-sections", which changed
the order of 32-element arrays so that accessing element 63 instead of
element 31 started getting garbage that no longer happened to work.

Here's a simple testcase that works with -O0 and fails with -O1.
It's useful to compare this with a a version whose argument is
unsigned short.  For the GCC testsuite, just remove references to
printf.

-------------------------------------------------------------------
extern int printf (const char *, ...);

int
__attribute__((noinline))
floor_log2_wide (x)
  unsigned int x;
{
  int log = -1;
  while (x != 0)
    log++,
    x >>= 1;
  return log;
}

int
main ()
{
  int m;
  unsigned int n = 4294967259U;

  m = floor_log2_wide ((int)n);
  printf ("n = %u, m = %d\n", n, m);
  if (m != 31)
    __builtin_abort ();
  return 0;
}
-------------------------------------------------------------------

elm3c105% /home/janis/tools/gcc-trunk-anonsvn/bin/gcc -m64 -O0 bug.c && ./a.out
n = 4294967259, m = 31
elm3c105% /home/janis/tools/gcc-trunk-anonsvn/bin/gcc -m64 -O1 bug.c && ./a.out
n = 4294967259, m = 63
Aborted

I've tried compilers back to GCC 4.1 and they all have the same problem;
it's probably not a regression.