prerelease-3 haifa-scheduler PentiumPro Optimizer bug

[N8TM@aol.com]

Yes, I see the same problem (built as i586-pc-cygwin32), but I didn't ask for
haifa-scheduler.  With -march=i686, gcc sees that the fld1 and fcmove
instructions would be appropriate to implement  ((i == j) ? 1 : 0), but it
gets it wrong somehow.  Maybe this could go into the testsuite.  Unlike the
snapshot compilers, I have no difficulty building the compiler with options
such as -Os -march=pentiumpro.