https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126587
Bug ID: 126587
Summary: [15/16/17 Regression] Wrong code with vectorisation
(vect_preserves_scalar_order_p ?)
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
Target Milestone: ---
/* Wrong code with -O3 on aarch64.
vect_preserves_scalar_order_p (gcc/tree-vect-data-refs.cc:297-306) returns
true at the "Single statements are always kept in their original order"
early exit before it reaches the loop-invariant-read guard. For the pair
(invariant load *p, ungrouped store b[i]) both refs are ungrouped, so the
guard is skipped and vect_prune_runtime_alias_test_list drops the runtime
alias check under the VF == 1 rule. vectorizable_load then hoists the
VMAT_INVARIANT load into the preheader with nothing guarding the aliasing.
The vector(1) __int128 store keeps the SLP unroll factor at 1 while staying
ungrouped, and the int store group provides an SLP root with more than one
lane so that vectorisation is still attempted.
Expected (scalar): c[0] == 5, c[1..7] == 77.
-O3 gives c[0..7] == 5. */
typedef __int128 i128;
__attribute__((noipa))
void h (int *a, i128 *b, i128 *c, i128 *p, i128 *q, int n)
{
for (int i = 0; i < n; i++)
{
a[4*i] = 1; a[4*i+1] = 2; a[4*i+2] = 3; a[4*i+3] = 4;
c[i] = *p;
b[i] = *q;
}
}
int a[32];
i128 b[8], c[8];
i128 qv = 77;
int main (void)
{
for (int i = 0; i < 8; i++) { b[i] = 5 + i; c[i] = -1; }
h (a, b, c, &b[0], &qv, 8);
if ((int) c[0] != 5)
__builtin_abort ();
for (int i = 1; i < 8; i++)
if ((int) c[i] != 77)
__builtin_abort ();
return 0;
}
Aborts on aarch64 at -O3 and passes at -O1
