[llvm] [SLP] Keep reassoc on the ordered fadd reduction when all ops allow it (PR #222479)

[Alexey Bataev via llvm-commits]

https://github.com/alexey-bataev requested changes to this pull request.

I don't think it is correct. Though the parameter of the reduction is in order, unordered reduction instruction may return different result rather than ordered reduction instruction for the same operand. Unordered reductions usually result in O(log n) complexity uops, which reorder operands, while ordered reductions implement classic linear O(n) summation, which preserves the order.

https://github.com/llvm/llvm-project/pull/222479