https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126159
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |rguenth at gcc dot gnu.org
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
The issue is
t.c:11:15: note: === vect_prune_runtime_alias_test_list ===
t.c:11:15: note: can tell at compile time that _3->f and _3->f alias
t.c:8:16: missed: not vectorized: compilation time alias: _4 = _3->f;
_3->f = _10;
not the load/store-lane. Changing the source to
void cellUnion_f(RtreeCoord *a1, RtreeCoord *a2, int n) {
int ii = 0;
do {
float tem1 = MIN(a1[ii].f, a2[ii].f);
float tem2 = MAX(a1[ii+1].f, a2[ii+1].f);
a1[ii].f = tem1;
a1[ii+1].f = tem2;
ii += 2;
} while (ii < n);
}
allows vectorizing with a VF == 1 (V2SF). Placing __restrict on a2 also
helps in a similar way but we seem to be still confused about even
(compute_affine_dependence
ref_a: *_12.i, stmt_a: _13 = *_12.i;
ref_b: *_12.i, stmt_b: *_12.i = _14;
) -> dependence analysis failed
Creating dr for *_12.i
analyze_innermost: success.
base_address: a1_20(D)
offset from base address: 0
constant offset from base address: 4
step: 8
base alignment: 4
base misalignment: 0
offset alignment: 512
step alignment: 8
base_object: *_12.i
Creating dr for *_12.i
analyze_innermost: success.
base_address: a1_20(D)
offset from base address: 0
constant offset from base address: 4
step: 8
base alignment: 4
base misalignment: 0
offset alignment: 512
step alignment: 8
base_object: *_12.i
and with
Creating dr for *_5.i
analyze_innermost: success.
base_address: a1_20(D)
offset from base address: 0
constant offset from base address: 0
step: 8
base alignment: 4
base misalignment: 0
offset alignment: 512
step alignment: 8
base_object: *_5.i
having different base objects makes things difficult.
And this is:
else if (TREE_CODE (ref) == COMPONENT_REF
&& TREE_CODE (TREE_TYPE (TREE_OPERAND (ref, 0))) == RECORD_TYPE)
{
/* For COMPONENT_REFs of records (but not unions!) use the
FIELD_DECL offset as constant access function so we can
disambiguate a[i].f1 and a[i].f2. */
tree off = component_ref_field_offset (ref);
off = size_binop (PLUS_EXPR,
size_binop (MULT_EXPR,
from dr_analyze_innermost, so the reason is you have a union here and
we refrain from touching those (for fear? I don't exactly remember,
but most definitely if you'd had two different records in the union
we'd get confused big time - it's possibly fine for "scalar" elements
and of elements of same size).
Is this practically relevant?
