[llvm] [SLP]Initial FMA support (PR #149102)

[Alexey Bataev via llvm-commits]

alexey-bataev wrote:

> Does anyone know why we do not generally convert fast fmul+fadd to fmuladd, to make cost-modelling easier? It would prevent us from doing things like inserting a shuffle between a fmul and an fadd, but maybe it prevents reductions in some cases? It might make cost-modelling in general simpler for any target with fma.

Maybe, conservative canonical representation?

https://github.com/llvm/llvm-project/pull/149102