This patch changes folding in fold_sign_changed_comparison in a way
that when we have pointer/non-pointer comparison, we give up folding
here. The issue is e.g. when we have (intptr_t) &MEM[(void *)&x + 4B]
== (intptr_t) &y and forwprop wants to fold this one via
fold_binary_loc--we then end up folding non-sign-changing comparison.
I don't know about the testcase though. I had something, but I think
it is too unrealiable. Regtested/bootstrapped on
x86_64-unknown-linux-gnu. Ok for trunk?
2012-10-18 Marek Polacek <pola...@redhat.com>
* fold-const.c (fold_sign_changed_comparison): Punt if folding
pointer/non-pointer comparison.
--- gcc/fold-const.c.mp 2012-10-17 15:01:34.882926243 +0200
+++ gcc/fold-const.c 2012-10-18 18:02:55.482642941 +0200
@@ -6731,12 +6731,14 @@ fold_sign_changed_comparison (location_t
&& TREE_TYPE (TREE_OPERAND (arg1, 0)) == inner_type))
return NULL_TREE;
- if ((TYPE_UNSIGNED (inner_type) != TYPE_UNSIGNED (outer_type)
- || POINTER_TYPE_P (inner_type) != POINTER_TYPE_P (outer_type))
+ if (TYPE_UNSIGNED (inner_type) != TYPE_UNSIGNED (outer_type)
&& code != NE_EXPR
&& code != EQ_EXPR)
return NULL_TREE;
+ if (POINTER_TYPE_P (inner_type) != POINTER_TYPE_P (outer_type))
+ return NULL_TREE;
+
if (TREE_CODE (arg1) == INTEGER_CST)
arg1 = force_fit_type_double (inner_type, tree_to_double_int (arg1),
0, TREE_OVERFLOW (arg1));
Marek
