[llvm] [DAGCombiner] Attempt to fold 'add' nodes to funnel-shift or rotate (PR #125612)

[Alex MacLean via llvm-commits]

AlexMaclean wrote:

> Thanks @AlexMaclean. We had a situation downstream which needed this transformation at SelectionDAG level too. I just wanted to mention that I tried following, similar to your change at IR level in [59ced72bc211f](https://github.com/llvm/llvm-project/commit/59ced72bc211f) (using the pattern matching framework at SD level).
> 
> ```
> static bool haveNoCommonBitsSetCommutative(SDValue A, SDValue B) {
>   ...
> +  SDValue V;
> +  APInt R;
> +    if (((sd_match(B, m_Shl(m_Value(), m_Sub(m_ConstInt(R), m_Value(V)))) &&
> +          sd_match(A, m_Srl(m_Value(), m_Specific(V)))) ||
> +         (sd_match(B, m_Srl(m_Value(), m_Sub(m_ConstInt(R), m_Value(V)))) &&
> +          sd_match(A, m_Shl(m_Value(), m_Specific(V))))) &&
> +        R.uge(A.getValueType().getScalarSizeInBits()))
> +      return true;
>       
>   return false;
> }
> ```
> 
> This then triggers the 'add' to 'or' transformation in [DAGCombiner](https://github.com/llvm/llvm-project/blob/dcb7a695004c49aaef02c3171343864870009961/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp#L2999), which in turn helps in the selection of the rotate instruction.

Nice, the SDAG matching is new to me and does seem quite useful. However, I think at the SelectionDAG level it makes more sense to fold 'add' to rotate directly, reusing the existing rotate matching code which looks for many different rotate idioms beyond just this simple case (ex. https://github.com/llvm/llvm-project/pull/125612/files#diff-7cb0d03bd314924f121a96a010f963383fe24079e25c71399b68ba979e8453abR140). 

https://github.com/llvm/llvm-project/pull/125612