Some Haifa scheduler bugs

[meissner@cygnus.com]

Note if people are going to use -fsched-spec-load-dangerous, I think there are
bugs in the current Haifa code.  I originally enabled it in the AIX port when
I was first testing haifa years ago, and eventually removed setting it, since
it was assuming all loads could be done speculatively, when it should have
only optimized things like:

	p != NULL && *p

(either that or somebody changed the default until AIX so that page 0 was
no longer mapped by default -- it was quite some time ago).