https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123870

--- Comment #30 from Robin Dapp <rdapp at gcc dot gnu.org> ---
The issue is that we don't emit a vsetvl before a spill-restore that
schedule-insn introduces (there are lots of them and I'm not sure all are
worthwhile but that's a separate discussion).

As suspected it's missing/wrong vsetvl "ratio" attribute of a full-register
load.  I'm not yet sure whether it only affects xtheadvector.

Reply via email to