[PATCH] Fix overzealous folding

[Richard Guenther]

On Fri, May 16, 2008 at 1:39 PM, Eric Botcazou <ebotcazou@adacore.com> wrote:
> This is a regression present at -O2 -gnatp on mainline and 4.3 branch,
>  although the underlying problem appears to have been there for ages.
>
>  The problem is that the compiler folds
>
>   (boolean) (i & 1)
>
>  into
>
>   (boolean) i & 1
>
>  Later VRP determines that i (an integer) is always greater than 0 and thus
>  computes that the test is always true.
>
>  The problematic code is:
>
>       /* Convert (T)(x & c) into (T)x & (T)c, if c is an integer
>          constants (if x has signed type, the sign bit cannot be set
>          in c).  This folds extension into the BIT_AND_EXPR.  */
>       if (INTEGRAL_TYPE_P (type)
>           && TREE_CODE (type) != BOOLEAN_TYPE
>           && TREE_CODE (op0) == BIT_AND_EXPR
>           && TREE_CODE (TREE_OPERAND (op0, 1)) == INTEGER_CST)
>
>  Note that it is properly disabled for BOOLEAN_TYPE; but, in Ada, "boolean"
>  is an ENUMERAL_TYPE with 2 values and the transformation is thus applied.

Hm, I would say that even for BOOLEAN_TYPE this folding is valid.
Obviously my ada skills are weak...  but are you sure you are not hiding
a bug in VRP?

I'll try to have a look myself.

Richard.


>  Tested on i586-suse-linux, OK for mainline and 4.3 branch?
>
>
>  2008-05-16  Eric Botcazou  <ebotcazou@adacore.com>
>
>         * fold-const.c (fold_unary) <CASE_CONVERT>: Fold the cast
>         into a BIT_AND_EXPR only for an INTEGER_TYPE.
>
>
>  2008-05-16  Eric Botcazou  <ebotcazou@adacore.com>
>
>         * gnat.dg/bit_packed_array3.adb: New test.
>
>
>  --
>  Eric Botcazou
>