[llvm] [LV] Convert gather loads with invariant stride into strided loads (PR #147297)

[Luke Lau via llvm-commits]

lukel97 wrote:

+1 on this approach of doing at as VPlan transformation. That way we don't need to work with the legacy cost model.

I presume we will need to handle VPWidenStridedLoadRecipes in planContainsAdditionalSimplifcations and skip the assertion?

> That said, the current patch faces an awkward situation: all existing strided access patterns in our lit tests are currently not converted to strided accesses, because the cost returned by TTI.getStridedMemoryOpCost is higher than that of a gather.

Yeah I think the current RISCVTTIImpl::getStridedMemoryOpCost is too expensive. It should definitely be at least cheaper than RISCVTTIImpl::getGatherScatterOpCost, and at the moment the cost model seems to return the same for them.

I know the spacemit-x60 doesn't have fast vlse but I don't think they're as slow as vluxei. Do we have benchmark results for these? cc) @mikhailramalho 

https://github.com/llvm/llvm-project/pull/147297