https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101445
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
Simplified testcase, fails at -O2 -ftree-loop-vectorize
int a[35] = {1, 1, 3};
void __attribute__((noipa))
foo ()
{
for (int b = 4; b >= 0; b--)
{
int tem = a[b * 5 + 3 + 1];
a[b * 5 + 3] = tem;
a[b * 5 + 2] = tem;
a[b * 5 + 1] = tem;
a[b * 5 + 0] = tem;
}
}
int main()
{
foo ();
for (int d = 0; d < 25; d++)
if (a[d] != 0)
__builtin_abort ();
return 0;
}
the load is vectorized in an odd way, but "correct" - but the final IV
update(s)
are bogus.
<bb 2>
vectp_a.7_34 = &a + 84; // &a[21]
<bb 3>
# vectp_a.6_35 = PHI <vectp_a.6_36(5), vectp_a.7_34(2)>
...
vect_tem_9.8_37 = MEM <vector(4) int> [(int *)vectp_a.6_35];
vect_tem_9.9_38 = VEC_PERM_EXPR <vect_tem_9.8_37, vect_tem_9.8_37, { 3, 2, 1,
0 }>;
vectp_a.6_39 = vectp_a.6_35 + 18446744073709551600; // -16
vect_tem_9.10_40 = MEM <vector(4) int> [(int *)vectp_a.6_39];
vect_tem_9.11_41 = VEC_PERM_EXPR <vect_tem_9.10_40, vect_tem_9.10_40, { 3, 2,
1, 0 }>;
vectp_a.6_42 = vectp_a.6_39 + 18446744073709551604; // -12
vect_tem_9.12_43 = VEC_PERM_EXPR <vect_tem_9.9_38, vect_tem_9.9_38, { 0, 0,
0, 0 }>;
...
vectp_a.6_36 = vectp_a.6_42 + 18446744073709551600; // -16
we're doing VMAT_CONTIGUOUS_REVERSE but the group has gaps and we fail to
account for the reverse when computing group_gap_adj (which should have been
+12, not -12).
