http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53639
--- Comment #3 from Uros Bizjak <ubizjak at gmail dot com> 2012-06-12 12:21:07 UTC --- (In reply to comment #2) > Unfortunately that patch regressed pr49095.c testcase. So, either we limit > the > splitter to the paradoxical subreg that is created by the combiner when seeing > SImode and followed by zero_extend to DImode of the result (done in this > patch), or we'd need to add new peepholes for the a = mem; a &= const; mem = > a; > if (a) > cases where a &= const has been transformed into andsi_1_zext. Uros, any > preference? The splitter, since the scheduler can break interesting sequence by inserting unrelated instructions.