We recently discovered that the loop with a dynamic upper bound is
unexpectedly unrolled during the NVVM to PTX process. By attaching the
`llvm.loop_annotation`, we can control the unrolling behavior precisely.
This PR enables the `cf.cond_br` to retain the loop annotation of
`scf.for` after the `convert-scf-to-cf` pass. This change allows users
to have precise control over the loop behavior during backend
transformation.
---------
Co-authored-by: Xiaolei Shi <xiaoleis@nvidia.com>
fbf81e3004