[Bug rtl-optimization/33410] [4.2/4.3 regression] ICE in iv_analyze_expr, at loop-iv.c:934

["rth at gcc dot gnu dot org" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #13 from rth at gcc dot gnu dot org  2008-01-31 17:23 -------
Created an attachment (id=15066)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=15066&action=view)
proposed fix

There's two things we can do.  First, we could modify loop-iv.c to cope with
a CCmode REG_EQUAL note associated with an integral register as a part of a
libcall.  Certainly there's nothing about the FP comparison that's going to
interest the rest of that function.  Second, we can assume that the tree 
optimizers have done all the CSE that's really going to happen and eliminate
the libcall notes at the rtl level as obsolete.

Certainly the later option is safer, as it affects no one but Alpha, and it
eliminates a potentially confusing reg note.

This fixes the testcase for cross-compile.  I'm trying to build it natively
now, but as the machine has locked up once already I'm affeared my last Alpha
machine is on it's last legs.  Someone else should try to build it as well.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33410