
Enables initial non-power-of-2 support (but still requires number of elements, forming whole registers) for reductions. Enables extra vectorization for MultiSource/Benchmarks/7zip/7zip-benchmark, CINT2006/464.h264ref and CFP2017rate/526.blender_r (checked for SSE2) Reviewers: RKSimon Reviewed By: RKSimon Pull Request: https://github.com/llvm/llvm-project/pull/112361