[llvm] [SLP] Improve cost model for i1 select-as-or/and patterns (PR #188572)

[Alexey Bataev via llvm-commits]

alexey-bataev wrote:

> Should this be done by the backends themselves? We already implicitly do a lot of transformations there, why is this case different? For instance, in X86 the conversion of UDIV/UREM by power of 2 is handled like so:
> 
> ```
>   // Vector unsigned division/remainder will be simplified to shifts/masks.
>   if ((ISD == ISD::UDIV || ISD == ISD::UREM) &&
>       Op2Info.isConstant() && Op2Info.isPowerOf2()) {
>     if (ISD == ISD::UDIV)
>       return getArithmeticInstrCost(Instruction::LShr, Ty, CostKind,
>                                     Op1Info.getNoProps(), Op2Info.getNoProps());
>     // UREM
>     return getArithmeticInstrCost(Instruction::And, Ty, CostKind,
>                                   Op1Info.getNoProps(), Op2Info.getNoProps());
>   }
> ```

Looks like this transformation is target independent, i.e. performed for all targets. Generally speaking, we can support it in TargetTransformInfo for all targets, or do it only in SLPVectorizer.

https://github.com/llvm/llvm-project/pull/188572