> I can't approve stuff in this area, but I think this is simple enough that > you may > consider pushing it as obvious. Or you need to CC Richi, but I think it > meets the bar. > > That said I do have a question. I was copying values from an existing LEN > operation, > so are the modes on the types in GIMPLE not what we expand to with len? > > i.e. for the bias is the type in GIMPLE made form something other than QImode?
No, it's just a quirk. The bias is (should) always be created as an intQI_type_node so that's fine. I think the issue rather is that it is also always a constant, i.e. expand_normal will create a VOIDmode RTX. When expanding an input operand we want a non-VOID mode to be able to copy the value to a reg. -- Regards Robin
