https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121599
--- Comment #3 from Alex Coplan <acoplan at gcc dot gnu.org> ---
(In reply to Jennifer Schmitz from comment #2)
> I can confirm the ICE. Thanks for the report.
> However, I think that g5289540ed58e42ae66255e31f22afe4ca0a6e15e only
> revealed a previously existing problem by folding the inner call
> (svuqadd_n_s8_z(svpfalse(), c, e)) to a zero vector.
>
> I compiled the following snippet with GCC built from the parent commit
> (g8772f37e45e9401c9a361548e00c9691424e75e0) and it has the same ICE:
>
> #include <arm_sve.h>
>
> svuint8_t f2 (svbool_t pg, svuint8_t op1)
> {
> return svsqadd_u8_z (pg, op1, svdup_s8 (0));
> }
>
> I will investigate why the insn is not recognized.
Thanks, yes it looks to be a pre-existing problem indeed. The adjusted
testcase you gave seems to ICE all the way back to the introduction of SVE in
GCC 10.
