https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118669
Bug ID: 118669
Summary: Misaligned store after after vectorization
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: kristerw at gcc dot gnu.org
Target Milestone: ---
Vectorization creates misaligned stores for the function below when compiled
for AArch64 with "-O3 -march=armv9.4":
void
foo (int i, int *a)
{
for (int j = 0; j < i; j++)
a[j] = j;
}
The vectorized code start by writing 2*vector_length elements at a time, as
long as there are more than 2*vector_length elements remaining:
<bb 3> [local count: 860067200]:
# vect_vec_iv_.6_37 = PHI <_38(6), { 0, 1, 2, ... }(14)>
# vectp_a.7_40 = PHI <vectp_a.7_41(6), a_8(D)(14)>
# ivtmp_46 = PHI <ivtmp_47(6), 0(14)>
_38 = vect_vec_iv_.6_37 + { POLY_INT_CST [8, 8], ... };
_39 = vect_vec_iv_.6_37 + { POLY_INT_CST [4, 4], ... };
MEM <vector([4,4]) int> [(int *)vectp_a.7_40] = vect_vec_iv_.6_37;
vectp_a.7_43 = vectp_a.7_40 + POLY_INT_CST [16, 16];
MEM <vector([4,4]) int> [(int *)vectp_a.7_43] = _39;
vectp_a.7_41 = vectp_a.7_43 + POLY_INT_CST [16, 16];
ivtmp_47 = ivtmp_46 + POLY_INT_CST [8, 8];
if (ivtmp_47 <= _45)
goto <bb 6>; [89.00%]
else
goto <bb 10>; [11.00%]
<bb 6> [local count: 765459809]:
goto <bb 3>; [100.00%]
The store
MEM <vector([4,4]) int> [(int *)vectp_a.7_40] = vect_vec_iv_.6_37;
may be misaligned as TYPE_ALIGN for "vector([4,4])" is 128 bits, while the
pointer in the first iteration is the "int *a" function argument which which
only guarantees 32-bit alignment.
