[llvm] [SLP] Reject 2-element vectorization when vector inst count exceeds scalar (PR #190414)

[Simon Pilgrim via llvm-commits]

RKSimon wrote:

Would it be better to take a step back and see if we can improve instruction costs properly to handle sub-1.0 throughputs instead of adding yet another workaround to SLP? 

We've already tentatively started this with #178962 - should we consider replacing the InstructionCost::CostType int64_t with a APFixedPoint entirely - something like an 60.4 fixed point type (or even make the FixedPointSemantics target specific?) - and then the cost tables can be tweaked later to handle this.

CC @bababuck 

https://github.com/llvm/llvm-project/pull/190414