[Bug middle-end/58742] pointer arithmetic simplification

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58742

--- Comment #21 from Richard Biener <rguenth at gcc dot gnu.org> ---
Ok, I have a patch that does the remaining (including the duplicate bug).  But
for the (p + sz) - p case it can only optimize the sizeof (*p) == 1 without
range information - for example for

__SIZE_TYPE__
fx (int *a, __SIZE_TYPE__ sz)
{
  int *b = a + sz;
  return b - a;
}

we get

fx (int * a, long unsigned int sz)
{
  int * b;
  long unsigned int _2;
  long int _7;
  long int _8;
  long unsigned int _9;

  <bb 2>:
  _2 = sz_1(D) * 4;
  _7 = (long int) _2;
  _8 = _7 /[ex] 4;
  _9 = (long unsigned int) _8;
  return _9;

}

as result which is basically (sz * 4) /[ex] 4 as we don't know whether
the multiplication by 4 overflows (well, the C language may say it
doesn't but the IL does not reflect this).  If we make 'sz' a signed
int then VRP could later optimize this (but it doesn't currently),
or forwprop could use range information.  On RTL we manage to optimize
the signed int input case.