This flips the switch on forming reduction intrinsics in the vectorizers. The IR diffs seem ok, but that doesn't provide any info on what happens in expansion/codegen. I will see if we can expose any obvious bugs using PhaseOrdering IR tests that include the expansion pass.
A motivating example is seen in https://bugs.llvm.org/show_bug.cgi?id=43953#c2 - if we had intrinsics there, we might get CGP or InstCombine to fold them.
maybe it's time to drop the experimental bit? :)