[Bug middle-end/19304] New: wrong code for spec test from emit_move_change_mode

["janis187 at us dot ibm dot com" <gcc-bugzilla at gcc dot gnu dot org>]

Benchmark 197.parser from SPEC CPU2000 fails on powerpc64-linux when
compiled with "-m32 -O2 -maltivec" or "-m32 -O2 -funroll-loops",
producing incorrect results.  The failures start with this patch:
                                                                                
2004-12-12  Richard Henderson  <rth@redhat.com>
                                                                                
        * expr.c (emit_move_change_mode): New.
        (emit_move_via_alt_mode): Use it.
                                                                                
I have no idea what's going on here but noticed that the original code
handled more conditions than does the new code.  If simplify_gen_subreg,
rather than simplify_subreg, is called when reload_in_progress is true
but MEM_P(x) is false, the test case passes again.  I'll attach the
diffs.
                                                                                
If this looks worth testing I'll do a bootstrap and testsuite run and
run the CPU2000 tests with the short test input using a variety of
compiler options.

-- 
           Summary: wrong code for spec test from emit_move_change_mode
           Product: gcc
           Version: 4.0.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: middle-end
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: janis187 at us dot ibm dot com
                CC: gcc-bugs at gcc dot gnu dot org,rth at gcc dot gnu dot
                    org
GCC target triplet: powerpc64-unknown-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19304