[llvm] [LV][EVL] Generate negative strided load/store for reversed load/store (PR #123608)

[Mel Chen via llvm-commits]

Mel-Chen wrote:

> > I am wondering about the necessity of emitting vp.stride_load/store in the vectorizer stage.
> 
> One benefit of emitting strided intrinsics earlier is that it allows us to use the strided memory op TTI hook in the cost model.
> 
> Otherwise we would have to conservatively cost it as a wide load + reverse, which IIRC the reverse is still costed as a quadratic vrgather today

A more precise cost model sounds like a good reason.
But if that's the case, wouldn't it be sufficient to just modify `computeCost`, keep load + reverse, and convert it into strided accesses in CodeGenPrepare? This would also benefit hand-written LLVM IR.

Another one is [enum InstWidening](https://llvm.org/doxygen/classllvm_1_1LoopVectorizationCostModel.html#abd75237965fb9548c506a19549c4a260), which is usually where decisions on how to widen memory accesses are made. It might be helpful.


https://github.com/llvm/llvm-project/pull/123608