register is not killed in the loop.

This fixes 188.ammp on ARM where the post-ra scheduler would grab a register
that looked available but wasn't.

A testcase would be huge and fragile, sorry.

llvm-svn: 101930