[Bug middle-end/71654] [6/7 Regression] missing VRP optimization on c++ unsigned char and short expressions

["glisse at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71654

Marc Glisse <glisse at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2016-06-24
            Summary|missing VRP optimization on |[6/7 Regression] missing
                   |c++ unsigned char and short |VRP optimization on c++
                   |expressions                 |unsigned char and short
                   |                            |expressions
     Ever confirmed|0                           |1

--- Comment #1 from Marc Glisse <glisse at gcc dot gnu.org> ---
  _1 = (intD.9) i_3(D);
  if (_1 <= 3)
[...]
  _5 = i_3(D) & 4;

So we know stuff about the range of _1, but we don't transfer that information
to i_3. Far from the first time I see something like this in VRP :-(

We could optimize _1 <= 3 to i_3(D) <= 3, and in this case it should help, but
IIRC there are also cases where it hurts (if we use _1 later on and don't move
the definition of _1 after the condition...). Here the failure is because we
were inconsistent about narrowing, forwprop changed (int)i & 4 to (int)(i & 4)
but not (int)i <= 3 to i <= 3. I get return 0" if I disable the first 2
forwprop passes.