[llvm] [SLP] Improve handling of sibling loop BTC mismatch (PR #208164)

[Hari Limaye via llvm-commits]

hazzlim wrote:

> No, what's wrong with the regression - bad cost of fmul, anything else - not a fix, but a hack to avoid the right fix. The main idea of the Loop Aware cost support - to spit vectorization between the loops with the potential different costs, ut you again returning the same (wrong) behavior.

Thanks for taking a look, I appreciate that this approach is at odds with the move towards the Loop Aware cost model - I am looking into this further and will aim to help move towards a proper fix!

https://github.com/llvm/llvm-project/pull/208164