https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125905
--- Comment #2 from Jeffrey A. Law <law at gcc dot gnu.org> ---
Another example, this time derived from x264. I *think* this is representative
of what's created by the vectorizer for selecting across two different loop
variants. It's ultimately just a conditional add. It's a gimple test because
I haven't managed to trip it directly in C.
long __GIMPLE (ssa)
f (long sa, long ua)
{
long ua15;
long hi;
_Bool neg;
__BB(2):
ua15_5 = ua_3(D) + 15l;
neg_6 = sa_4(D) < 0l;
hi_7 = neg_6 ? 15l : ua15_5;
return hi_7;
}
Generates:
srli a5,a0,63 # 10 [c=4 l=4] lshrdi3
addi a1,a1,15 # 31 [c=4 l=4] *adddi3/1
li a4,15 # 30 [c=4 l=4] *movdi_64bit/1
czero.nez a0,a1,a5 # 33 [c=4 l=4] *czero.nez.didi
czero.eqz t0,a4,a5 # 34 [c=4 l=4] *czero.eqz.didi
add a0,t0,a0 # 22 [c=4 l=4] *adddi3/0
ret # 47 [c=0 l=4] simple_return
But really just one czero is needed. Select across 0/15 using that single
czero, then unconditionall add that to a1.
