Re: [PATCH]: Improve data ref disambiguation

["Daniel Berlin" <dberlin at dberlin dot org>]

On 6/6/07, Zdenek Dvorak <rakdver@kam.mff.cuni.cz> wrote:
> Hello,
>
> > Right now, we currently decide that if two base objects are not
> > operand_equal_p, they must have different base objects.
> >
> > We can do slightly better.
> >
> > If the vops on the datarefs are exactly equal, they must have the same
> > base object.
> >
> > IE:
> >
> > (gdb) p debug_generic_stmt (a->stmt)
> > # VUSE <PARM_NOALIAS.4_22, PARM_NOALIAS.5_14, SMT.6_39> {
> > PARM_NOALIAS.4 PARM_NOALIAS.5 SMT.6 }
> > D.1886_24 = *D.1885_23;
> >
> > $1 = void
> > (gdb) p debug_generic_stmt (b->stmt)
> > # VUSE <PARM_NOALIAS.4_22, PARM_NOALIAS.5_14, SMT.6_39> {
> > PARM_NOALIAS.4 PARM_NOALIAS.5 SMT.6 }
> > D.1888_28 = *D.1887_27;
> >
> >
> > Dereferences of these pointers can touch exactly the same memory, so
> > they must share a base object (there is no guarantee they have the
> > same offsets, etc, only that their pointers can only touch the same
> > set of memory).
>
> is this patch really correct? The rest of code in
> initialize_data_dependence_relation, as well as
> compute_affine_dependence, assumes that DR_BASE_OBJECTs are exactly
> equal).

At worst, this will cause it to claim dependence where there is none.
It will never say that two completely unrelated accesses are related.
If you know of a place that  cares whether they are exactly equal, i'd
be happy to not commit the patch  (It actually has not been committed
yet).
I'd just wait for ptr-plus to merge so that we can separate out the
base object better

The motivating example in this case is:
void test(unsigned char * __restrict__ dst, unsigned char *
__restrict__ row, char n) {
   unsigned char *row0 = row;
   unsigned char *row1 = row + n;
   unsigned char *row2 = row + n * 2;
   unsigned char *dst1 = dst + n;
   for(char x = 0; x < n; x++ )  {
       dst[x] = row0[x] + row1[x] + row2[x];
       dst1[x] = (row1[x] + row2[x])*4;
   }
}

Without this patch, we refuse to determine the dependence between
row0, row1, and row2, because they end up with different ssa_names,
even though they have the exact same set of vuses, and can only touch
the same memory

(gdb) p debug_generic_stmt (dra->stmt)
# VUSE <SMT.4_31> { SMT.4 }
D.1886_24 = *D.1885_23;

$1 = void
(gdb) p debug_generic_stmt (drb->stmt)
# VUSE <SMT.4_31> { SMT.4 }
D.1888_28 = *D.1887_27;


--Dan