[llvm] [DAG] Always allow folding XOR patterns to ABS pre-legalization (PR #94601)

[Matt Arsenault via llvm-commits]

https://github.com/arsenm commented:

AMDGPU only has truly legal abs for the SALU 32-bit case. We fake report it legal and then expand the VALU version in selection. It looks like for targets with packed operations, we get a bad expansion in the 16-bit vector case. e.g. v2i16 we get 2 instructions, but some complicated expansion for v4i16 instead of just splitting the vector 

https://github.com/llvm/llvm-project/pull/94601