https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68823
--- Comment #13 from Richard Biener <rguenth at gcc dot gnu.org> ---
We apply blocking on
int a[256][256];
void foo (void)
{
int *p = &a[4][7];
for (int i = 0; i < 256; ++i)
for (int j = 0; j < 256; ++j)
a[j][i] = a[j][i] * (*(int(*)[1])p)[0];
}
but not when we use a plain indirection, then we get
[scop-detection-fail] Graphite cannot handle data-refs in stmt:
# VUSE <.MEM_20>
_2 = MEM[(int *)&a];
for some reason (stmt_has_simple_data_refs_p fails). Zero subscripts I guess.
That's a weird thing to look at, even __real MEM[(int *)&a] would make it
one subscript (see dr_analyze_indices).
Means graphite doesn't handle non-indexed memory reads either.
So the above might be a different bug (we're supposed to reject the data-ref?).
We're also blocking
int a[256][256];
void foo (void)
{
int *p = &a[0][0];
for (int i = 0; i < 256; ++i)
for (int j = 0; j < 256; ++j)
a[j][i] = a[j][i] * p[i];
}
can't wrap my head around whether that's valid or not.
In the end it looks like graphite doesn't handle mixed pointer / non-pointer
accesses because it relies on DR_ACCESS_FNs "matching" for all possible
may-aliases. I suppose
/* The access polyhedron contains the polyhedral space this data
reference will access.
The polyhedron contains these dimensions:
- The alias set (a):
Every memory access is classified in at least one alias set.
- The subscripts (s_0, ..., s_n):
The memory is accessed using zero or more subscript dimensions.
- The iteration domain (variables and parameters)
means that when things are in the same alias set they may still differ
in the number of subscripts? But in that case they always appear as
not aliasing?
That is, should we, in
static void
build_alias_set (scop_p scop)
{
int num_vertices = scop->drs.length ();
struct graph *g = new_graph (num_vertices);
dr_info *dr1, *dr2;
int i, j;
int *all_vertices;
FOR_EACH_VEC_ELT (scop->drs, i, dr1)
for (j = i+1; scop->drs.iterate (j, &dr2); j++)
if (dr_may_alias_p (dr1->dr, dr2->dr, true))
{
add_edge (g, i, j);
add_edge (g, j, i);
when DR_NUM_DIMENSIONS (dr1->dr) != DR_NUM_DIMENSIONS (dr2->dr) better "FAIL"?
