Hi Richard,
On Tue, Jul 14, 2009 at 4:54 PM, Richard
Guenther<[email protected]> wrote:
> On Tue, Jul 14, 2009 at 8:01 AM, Li Feng<[email protected]> wrote:
>> Hi,
>>
>> I'm now working on Graphite branch and need to know
>> the alias set information for each data_reference_p, which
>> would be an integer (or alias_set_type) stands for which
>> alias set it is in.
>>
>> I tried to get the alias set information with get_alias_set (tree)
>> (I've no idea how this function works, just a experimental
>> trying), for my testcase, it returns 2 for all the
>> data_reference_p->ref, which I think is unreasonable.
>> So I think I may go wrong somewhere.
>>
>> The question will be: how could I get it's relevant
>> alias set information from data_reference_p?
>>
>> p.s. :
>> In Graphite, the data reference was built for each
>> gimple stmt with:
>> get_references_in_stmt (stmt, &references);
>> then for each ref in references, data reference is created with:
>> dr = create_data_ref (nest, *ref->pos, stmt, ref->is_read);
>
> get_alias_set (dr->ref) is the correct thing. Why do you think it
> is unreasonable to return 2 for all of them? Why do you need
> alias-set information anyway?
The testcase looks like this, where I assume that
p and a in the same alias set, while q and Q in another, and so on.
But between them, the alias set number should be different, otherwise
maybe I misunderstood somewhere about the alias set.
int Q[10];
int foo()
{
int i;
int a[10], b[20];
int *p = a;
int *q = Q;
for (i = 0; i < 10; i++)
{
p[i] = i;
a[i]= i - 5;
b[i] = i*i;
q[i] = 5;
}
return Q[3]*a[2]*b[10];
}
For you question,
We are using this information before dependency checking under
polyhedron, if 2 data references get different dimensions and they
are not in the same alias set, we could conclude that they
takes no dependency.
Li
