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

--- Comment #8 from Robin Dapp <rdapp at gcc dot gnu.org> ---
(In reply to Ilya Kurdyukov from comment #7)
> So, XuanTie was supposed to maintain this target, but they stopped doing it?
> 
> I think support for this RISC-V extension is in poor condition, it can't be
> used because there's no guarantee the code will work correctly. Even if I
> got it working correctly, it could easily break if I use other compiler
> flags, or make small changes to the code.

>From what I can tell the instructions you were trying to use just don't exist
in the 0.7.1 extension so even if we showed a nicer error message instead of an
ICE it still wouldn't work.  GCC 16 should be better in that regard but
xtheadvector just has much less test coverage "by design".  Still, there are
packages out there (ncnn for example) with xtheadvector code paths that at
least compile properly.

The vmv2r thing is a (likely fixed) bug that hasn't been backported. The
limited interest in xtheadvector of course also influenced our decision whether
or not to backport a specific fix.

I could have a look for __riscv_vfncvt but if you don't intend to continue with
your endeavor, it might not help all that much.  If you have a specific piece
of code that compiles with GCC 16 but not with 15 and would like a backport,
feel free to say so.  If it's easy to do, we might still consider it, but
obviously there's no guarantee - this extension is just very low on our list of
priorities.

Reply via email to