------- Comment #3 from ian at airs dot com 2007-04-27 04:52 ------- Created an attachment (id=13453) --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=13453&action=view) Proposed patch
This is an interesting test case. It turns out that VRP has always consistently assumed that TREE_TYPE (TYPE_MIN_VALUE (t)) == t. The assumption turns up in == comparisons against TYPE_MIN_VALUE (and TYPE_MAX_VALUE). Those == comparisons are normally valid because we hash integer constants so that they are all the same. The hash includes the type. So when TREE_TYPE (TYPE_MIN_VALUE (t)) != t, the equality will actually never hold, and VRP will do the wrong thing. This patch simply ensures that the equality holds. I'm testing it now. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31710
