https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121126
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
The following fails at -O2:
unsigned char a;
unsigned b;
int r[11];
static void __attribute__((noipa))
c(int e, unsigned s[][11][11])
{
for (int u = -(e ? 2000424973 : 0) - 2294542319; u < 7; u += 4)
for (int x = 0; x < 300000011; x += 4)
for (int y = 0; y < (0 < s[u][4][1]) + 11; y += 3) {
a = a ?: 1;
b = r[2];
}
}
long long ab;
int e = 1;
unsigned s[11][11][11];
int main()
{
for (int t = 0; t < 11; ++t)
r[t] = 308100;
c(e,s);
ab = b;
if (ab != 308100)
__builtin_abort ();
}
t.c:8:23: note: ***** Choosing vector mode RVVM1QI
t.c:8:23: optimized: loop vectorized using variable length vectors
t.c:8:23: note: OUTER LOOP VECTORIZED
which is interesting. There seems to be quite some confusion in the code
generation about the conditional reduction - and we end up with
using .VEC_EXTRACT ({ 1, ...}, _46) in the end rather than sth based
on the inner loop vect__*a_lsm.
