[polly] [Polly] Skip vectorize.enable for FP loops with dist=1 dependences (PR #205756)

[Karthika Devi C via llvm-commits]

kartcq wrote:

Yes. Right now there is no metadata as such to block this reassociation, There is a flag at file level only.
So for now we are neutralising the loops. Maybe we can eventually have a fine grained control on this such as blocking reassociation alone. The existing behaviour of LV allows reordering for FP ops when force vectorization is enabled. This resulted in correctness failures in our internal benchmarks with 'polly-annotate-metadata-vectorize'

https://github.com/llvm/llvm-project/pull/205756