[llvm] [TTI][Vectorize] Migrate masked/gather-scatter/strided/expand-compress costing (NFCI) (PR #165532)

[Philip Reames via llvm-commits]

preames wrote:

I was initial skeptical from the review description, but reading over the code this doesn't seem unreasonable.  My main worry with the change is arranging the staging of changes to be actually NFC.  Towards that end, could we consider introducing MemIntrinsicCostAttributes in a separate patch (using the prior API names), and updating each routine in it's own commit with the update going all the way through to backend impl?  Doing that would make it easier to lean on the compiler and audit the results.  Once each was updated, we could then do a final change which does only the API name update and the dispatch in TTI.  

Please take my suggestion as only a suggestion; please don't consider the review blocked.  If someone else approves the current direction, I have no problems with that.



https://github.com/llvm/llvm-project/pull/165532