[llvm] Reland "[LV]: Teach LV to recursively (de)interleave." (PR #125094)

[Hassnaa Hamdi via llvm-commits]

hassnaaHamdi wrote:

> I think I may be missing something, I tried to track down the other divergence in the cost model besides the one mentioned in #125008.
> 
> But when I tried something like #125434, there seem to be no other differences between the cost models with this patch.
> 
> Are there perhaps other cases where computing the cost for binary VPInstructions with underlying instructions isn't enough?

Thanks for 

> I think I may be missing something, I tried to track down the other divergence in the cost model besides the one mentioned in #125008.
> 
> But when I tried something like #125434, there seem to be no other differences between the cost models with this patch.
> 
> Are there perhaps other cases where computing the cost for binary VPInstructions with underlying instructions isn't enough?

What you are saying is correct, but I did this patch because the opcode that is used by vplan-based cost model is 'Add' while the underlying instruction used in the legacy cost model is 'or'. As they are eventually different instructions, so I can't guarantee that I will get same costs from all backends.
And given that the comparison between legacy model and vpla-based model is temporary thing, I think this current solution is better.

https://github.com/llvm/llvm-project/pull/125094