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

[David Green via llvm-commits]

davemgreen wrote:

> > OK, sounds good. I'm not against this, I was just thinking of things like what the stride was in a StridedMemoryOp (constants / small values could be cheaper on some hypothetical architecture). GatherScatters could do with better cost modelling, including possible whether the addresses are base + vector offset or vector-base + constant offset, etc. This combined interface might make passing that information harder.
> 
> I agree that a constant stride helps the cost model. In that cases I’d extend getMemIntrinsicInstrCost to take a MemAttr bag which has a constructor specific for target intrinsic. Callers would look like:
> 
> * getMemIntrinsicInstrCost(Intrinsic::masked_load, MemAttr(...), K)
> * getMemIntrinsicInstrCost(Intrinsic::vp_strided_load, MemAttr(..., Stride, ...), K)
> 
> This way just resembles getIntrinsicInstrCost. And how to fold MemAttr(...) into IntrinsicCostAttributes is a open question as noted in the reply above to @lukel97.

The constant stride to StridedMemoryOp was just a hypothetical - I don't have a real place that would be useful, so don't worry about it too much. It was more of a general question about higher-level information being passed through to each of the functions and whether this is really better.

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