[Bug middle-end/33779] [4.3 Regression] folds unsigned multiplication == 0 to true

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

------- Comment #5 from tromey at gcc dot gnu dot org  2007-10-27 16:42 -------
I looked at this a little.

This test is a little bit funny because the '+' has undefined
overflow but the '*' does not.  Move the cast to make the '+'
have defined overflow, and it works.

What happens is that we call fold_sign_changed_comparison
on the '==' expression (in the very first test case).
This decides that the expression can be simplified by
removing the casts to unsigned.  Some earlier call has already
hoisted the cast to unsigned to apply to the '*' tree as a
whole, so stripping this means we are optimizing a signed
multiply.

Offhand I would say that it is not valid to convert an unsigned
operation to a signed operation, because that changes overflow
from defined to undefined.  Going the other direction should
generally be ok -- though, I suppose, sometimes miss an optimization.

Removing the special case for EQ_EXPR and NE_EXPR from
fold_sign_changed_comparison should fix this bug, but I don't know
at what cost.  Or, perhaps this test could be changed to look at
whether arg0_inner is now a signed binary expression.  Some advice
from a fold expert would be appreciated.


-- 

tromey at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |tromey at gcc dot gnu dot
                   |                            |org
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
   Last reconfirmed|0000-00-00 00:00:00         |2007-10-27 16:42:51
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33779