[llvm] [AArch64][CostModel] Consider the cost of const vector (PR #117539)

[David Green via llvm-commits]

davemgreen wrote:

I believe I tried this and it was crashing inside getConstantMaterializationCost. It could now do with an update.

4 sounds like too high for the cost, when using TCK_RecipThroughput. I was wondering if we should be reducing the cost of scalar and/or/xor+shift by a constant, but that might lead to other issues. Perhaps we should be disincentivising 2x vectorization in general a little, especially if it involves shuffles.

https://github.com/llvm/llvm-project/pull/117539