https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122297
--- Comment #8 from jchrist at linux dot ibm.com --- (In reply to Richard Biener from comment #5) > Hum. Can you point me to where we fall back to byte-based len? I don't know where this happens, but if I check the generated gimple, I see that we use byte sizes for loads/stores, but if none are present (as in nodump-vecextract-1.c), we use element sizes. This can also be seen (and I assume you found the same thing) with the different factors passed to vect_record_loop_len. I was surprised to see that this factor is stored, but in the case of a load bias completely ignored.
