[llvm] [X86] Prefer 128-bit vectors for Bulldozer CPUs (PR #225838)

[Simon Pilgrim via llvm-commits]

RKSimon wrote:

> I have access to bulldozer hardware in which I can test. However, as we know bdver2 has 128-bit native data path and 256 data is double pumped. So, this is given.

Double pumping isn't always a problem with 256-bit instructions (the lower total instruction decode cost can be useful, unless it saturates the schedulers or 256-bit memory traffic causes stalls somehow), Bulldozer problems tends to come from the 2 CMT hwthreads competitively sharing the FPU pipes - a 256-bit instruction on one hwthread can block FPU use entirely by the other :/

Someday I'd love llvm-mca to be able to model SMT/CMT behaviours, but we haven't even managed decode behaviour yet :) 

https://github.com/llvm/llvm-project/pull/225838