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

--- Comment #33 from Ilya Kurdyukov <jpegqs at gmail dot com> ---
(In reply to Jeffrey A. Law from comment #32)
> So what I'd recommend is we get a distinct bug for each issue

1. There are an unknown number of intrinsics for RVV 1.0 that are not in
xtheadvector, causing crashes with ICE (vzext, vsext) or failure to compile
with assembler (vfncvt_rtz).
2. GCC 15.2.0 -O2, crashes with ICE "validate_change_or_fail"
3. GCC 14.3.0 -O2, uses invalid th.vmv2r instruction for xtheadvector
4. GCC 15.2.0 -O1, produces code that crashes with a SIGILL error at runtime.

GDB shows this location:

   0x2aaaaad104:        addi    a3,sp,16
   0x2aaaaad106:        th.vsetvli      zero,s8,e32,m2,d1
=> 0x2aaaaad10a:        th.vlse.v       v7,(a3),zero
   0x2aaaaad10e:        th.vmv.v.i      v10,0
   0x2aaaaad112:        th.vsetvli      zero,zero,e32,m2,d1

5. GCC 16.0.0 -O2, produces code that returns incorrect results (this is
related to the "vsetvl-strategy", as we found out).

Which of these would you agree to fix?

The last one definitely needs to be fixed, the rest is less important or
unimportant.

Reply via email to