https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111112
Robin Dapp <rdapp at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
See Also| |https://gcc.gnu.org/bugzill
| |a/show_bug.cgi?id=122221
--- Comment #2 from Robin Dapp <rdapp at gcc dot gnu.org> ---
After looking some more, it is related to PR122221 but not a dup.
GCC's asm still looks similar, just that we use a vx variant now
vwadd.vv v1,v3,v2
vsetvli zero,zero,e16,m1,ta,ma
vadd.vx v1,v1,a3
vsetvli zero,zero,e8,mf2,ta,ma
vnsra.wi v1,v1,2
vs LLVM's
vwadd.vx v10, v8, a3
vwadd.wv v10, v10, v9
vnsrl.wi v8, v10, 2
GCC with -fno-tree-reassoc:
vsetivli zero,8,e16,m1,ta,ma
vle8.v v4,0(a1)
vmv.v.x v1,a3
vle8.v v3,0(a2)
vsetvli zero,zero,e8,mf2,ta,ma
vwadd.wv v2,v1,v4
vwadd.wv v1,v2,v3
vnsra.wi v1,v1,2
vse8.v v1,0(a0)
So there's just the vmv.v.x that requires another vsetvl. I would have
expected us to use a vwadd.wx variant here but we don't. Maybe that's due to
canonicalization?
Anyway, reassoc doesn't help.
