I'm new to LLVM dev, but I have been working with a target with a
minimum addressable byte of 16-bits. I found that in
DAGCombiner::visitAND, EVT::isRound could create i8 loads on my 16-bit
target which are ultimately invalid. EVT::isRound appears to use a
hard-coded 8, rather than pulling the targets BitsPerByte field. Is this
a potential bug or is there a better way to address this? Hard coding a
16 in the isRound field fixes the issue for me.