>> Hmm, we try to use smaller offset types first, generally. >> If a 64-bit offset is used that's likely due to a double array being used >> just >> like in SPECfp 2017's lbm? How do you treat 64-bit offsets when the hardware >> can only handle 32-bits? Maybe we could use ranger range information >> in vect_truncate_gather_scatter_offset? > > Simplifying a bit, we have two gather/scatter instruction formats. One > takes a 64-bit base address and a vector of 32-bit offsets, and the > other takes a vector of 64-bit absolute addresses (optionally with a > 12-bit scalar offset).
Ah, I see, thanks. I presume the lbm issue is not reproducible with SPEC2017's lbm as we don't vectorize it at all currently (on my list, though)? If you have time a small example of where we go wrong would be appreciated so we can track it and see if something can be done. Not urgent of course. -- Regards Robin
