On Tue, Apr 24, 2012 at 1:01 AM, Jakub Jelinek <[email protected]> wrote:
> Hi!
>
> My PR52533 patch fixed unsigned comparisons, but not signed ones
> where the maximum is different.
> Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux,
> ok for trunk?
>
> 2012-04-23 Jakub Jelinek <[email protected]>
>
> PR tree-optimization/53058
> * tree-vrp.c (register_edge_assert_for_2): Compare mask
> for LE_EXPR or GT_EXPR with TYPE_MAX_VALUE of TREE_TYPE (val)
> instead of all ones in the precision.
>
> * gcc.c-torture/compile/pr53058.c: New test.
>
> --- gcc/tree-vrp.c.jj 2012-04-23 11:11:21.000000000 +0200
> +++ gcc/tree-vrp.c 2012-04-23 16:08:43.752462433 +0200
> @@ -4565,6 +4565,7 @@ register_edge_assert_for_2 (tree name, e
> && INTEGRAL_TYPE_P (TREE_TYPE (name2))
> && IN_RANGE (tree_low_cst (cst2, 1), 1, prec - 1)
> && prec <= 2 * HOST_BITS_PER_WIDE_INT
> + && prec == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (val)))
> && live_on_edge (e, name2)
> && !has_single_use (name2))
> {
> @@ -4598,8 +4599,10 @@ register_edge_assert_for_2 (tree name, e
> new_val = val2;
> else
> {
> + double_int maxval
> + = tree_to_double_int (TYPE_MAX_VALUE (TREE_TYPE (val)));
I don't like using TYPE_MAX_VALUE in VRP - at least please use vrp_val_max.
For enums this can be not what you expect(?)
Please consider adding a double_int_max_value (unsigned prec, bool sign)
and double_int_min_value.
Thanks,
Richard.
> mask = double_int_ior (tree_to_double_int (val2), mask);
> - if (double_int_minus_one_p (double_int_sext (mask, prec)))
> + if (double_int_equal_p (mask, maxval))
> new_val = NULL_TREE;
> else
> new_val = double_int_to_tree (TREE_TYPE (val2), mask);
> --- gcc/testsuite/gcc.c-torture/compile/pr53058.c.jj 2012-04-23
> 15:53:41.489982650 +0200
> +++ gcc/testsuite/gcc.c-torture/compile/pr53058.c 2012-04-23
> 15:53:13.000000000 +0200
> @@ -0,0 +1,12 @@
> +/* PR tree-optimization/53058 */
> +
> +int a, b, c;
> +
> +void
> +foo ()
> +{
> + c = b >> 16;
> + if (c > 32767)
> + c = 0;
> + a = b;
> +}
>
> Jakub
