https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123894
--- Comment #2 from Vineet Gupta <vineetg at gcc dot gnu.org> --- (In reply to Andrew Pinski from comment #1) > Confirmed. This is a specific issue with preserve_access_index. It could be > the way preserve_access_index causes: > MEM <<unnamed-signed:5>> [(void *)_6]; Sweet, for this test I didn't try this variation (I was mucking with bigger tests and toggling the attr etc). Anyhow this gives me a direction to work. > Which might be either a middle-end issue or maybe the backend should expand > preserve_access_index accesses differently. RISC-V codegen for same test (modulo attr) seems to be generating extraction so likely backend related.
