The pre-regalloc pass caused some regressions in both compile time and
performance of the generated code, and it did not improve performance, except
for indirect branches.  I also moved the check for single-block loops to speed
up the common case when running the taildup pass before reg allocation.

llvm-svn: 93505