return res;
}
+/* Builds a partial difference equations and inserts them
+ into pointset powerset polyhedron P. Polyhedron is assumed
+ to have the format: T|I|T'|I'|G|S|S'|l1|l2.
+
+ TIME_DEPTH is the time dimension w.r.t. which we are
+ differentiating.
+ OFFSET represents the number of dimensions between
+ columns t_{time_depth} and t'_{time_depth}.
+ DIM_SCTR is the number of scattering dimensions. It is
+ essentially the dimensionality of the T vector.
+
+ The following equations are inserted into the polyhedron P:
+ | t_1 = t_1'
+ | ...
+ | t_{time_depth-1} = t'_{time_depth-1}
+ | t_{time_depth} = t'_{time_depth} + 1
+ | t_{time_depth+1} = t'_{time_depth + 1}
+ | ...
+ | t_{dim_sctr} = t'_{dim_sctr}. */
+
+static void
+build_partial_difference (ppl_Pointset_Powerset_C_Polyhedron_t *p,
+ ppl_dimension_type time_depth,
+ ppl_dimension_type offset,
+ ppl_dimension_type dim_sctr)
+{
+ ppl_Constraint_t new_cstr;
+ ppl_Linear_Expression_t le;
+ ppl_dimension_type i;
+ ppl_dimension_type dim;
+ ppl_Pointset_Powerset_C_Polyhedron_t temp;
+
+ /* Add the equality: t_{time_depth} = t'_{time_depth} + 1.
+ This is the core part of this alogrithm, since this
+ constraint asks for the memory access stride (difference)
+ between two consecutive points in time dimensions. */
+
+ ppl_Pointset_Powerset_C_Polyhedron_space_dimension (*p, &dim);
+ ppl_new_Linear_Expression_with_dimension (&le, dim);
+ ppl_set_coef (le, time_depth, 1);
+ ppl_set_coef (le, time_depth + offset, -1);
+ ppl_set_inhomogeneous (le, 1);
+ ppl_new_Constraint (&new_cstr, le, PPL_CONSTRAINT_TYPE_EQUAL);
+ ppl_Pointset_Powerset_C_Polyhedron_add_constraint (*p, new_cstr);
+ ppl_delete_Linear_Expression (le);
+ ppl_delete_Constraint (new_cstr);
+
+ /* Add equalities:
+ | t1 = t1'
+ | ...
+ | t_{time_depth-1} = t'_{time_depth-1}
+ | t_{time_depth+1} = t'_{time_depth+1}
+ | ...
+ | t_{dim_sctr} = t'_{dim_sctr}
+
+ This means that all the time dimensions are equal except for
+ time_depth, where the constraint is t_{depth} = t'_{depth} + 1
+ step. More to this: we should be carefull not to add equalities
+ to the 'coupled' dimensions, which happens when the one dimension
+ is stripmined dimension, and the other dimension corresponds
+ to the point loop inside stripmined dimension. */
+
+ ppl_new_Pointset_Powerset_C_Polyhedron_from_Pointset_Powerset_C_Polyhedron (&temp, *p);
+
+ for (i = 0; i < dim_sctr; i++)
+ if (i != time_depth)
+ {
+ ppl_new_Linear_Expression_with_dimension (&le, dim);
+ ppl_set_coef (le, i, 1);
+ ppl_set_coef (le, i + offset, -1);
+ ppl_new_Constraint (&new_cstr, le, PPL_CONSTRAINT_TYPE_EQUAL);
+ ppl_Pointset_Powerset_C_Polyhedron_add_constraint (temp, new_cstr);
+
+ if (ppl_Pointset_Powerset_C_Polyhedron_is_empty (temp))
+ {
+ ppl_delete_Pointset_Powerset_C_Polyhedron (temp);
+ ppl_new_Pointset_Powerset_C_Polyhedron_from_Pointset_Powerset_C_Polyhedron (&temp, *p);
+ }
+ else
+ ppl_Pointset_Powerset_C_Polyhedron_add_constraint (*p, new_cstr);
+ ppl_delete_Linear_Expression (le);
+ ppl_delete_Constraint (new_cstr);
+ }
+
+ ppl_delete_Pointset_Powerset_C_Polyhedron (temp);
+}
+
+
/* Set STRIDE to the stride of PDR in memory by advancing by one in
- time dimension DEPTH. */
+ the loop at DEPTH. */
static void
memory_stride_in_loop (Value stride, graphite_dim_t depth, poly_dr_p pdr)
ppl_set_coef (lma, dim_L1, -1);
ppl_new_Constraint (&new_cstr, lma, PPL_CONSTRAINT_TYPE_EQUAL);
ppl_Pointset_Powerset_C_Polyhedron_add_constraint (p1, new_cstr);
+ ppl_delete_Linear_Expression (lma);
+ ppl_delete_Constraint (new_cstr);
}
/* Now intersect all the parts to get the polyhedron P1:
free (map);
}
- /* Add equalities:
- | t1 = t1'
- | ...
- | t_{depth-1} = t'_{depth-1}
- | t_{depth+1} = t'_{depth+1}
- | ...
- | t_{dim_sctr} = t'_{dim_sctr}
-
- This means that all the time dimensions are equal except for
- depth, where we will add t_{depth} = t'_{depth} + 1 in the next
- step. */
-
time_depth = psct_dynamic_dim (pbb, depth);
- for (i = 0; i < dim_sctr; i++)
- if (i != time_depth)
- {
- ppl_new_Linear_Expression_with_dimension (&le, new_dim);
- ppl_set_coef (le, i, 1);
- ppl_set_coef (le, i + offset, -1);
- ppl_new_Constraint (&new_cstr, le, PPL_CONSTRAINT_TYPE_EQUAL);
- ppl_Pointset_Powerset_C_Polyhedron_add_constraint (p2, new_cstr);
- ppl_delete_Linear_Expression (le);
- ppl_delete_Constraint (new_cstr);
- }
-
- /* Add equality : t_{depth} = t'_{depth} + 1.
- This is the core part of this alogrithm, since this
- constraint asks for the memory access stride (difference)
- between two consecutive points in time dimensions. */
- {
- ppl_new_Linear_Expression_with_dimension (&le, new_dim);
- ppl_set_coef (le, time_depth, 1);
- ppl_set_coef (le, time_depth + offset, -1);
- ppl_set_inhomogeneous (le, 1);
- ppl_new_Constraint (&new_cstr, le, PPL_CONSTRAINT_TYPE_EQUAL);
- ppl_Pointset_Powerset_C_Polyhedron_add_constraint (p2, new_cstr);
- ppl_delete_Linear_Expression (le);
- ppl_delete_Constraint (new_cstr);
- }
/* P1 = P1 inter P2. */
- {
- ppl_Pointset_Powerset_C_Polyhedron_intersection_assign (p1, p2);
- ppl_delete_Pointset_Powerset_C_Polyhedron (p2);
- }
+ ppl_Pointset_Powerset_C_Polyhedron_intersection_assign (p1, p2);
+ build_partial_difference (&p1, time_depth, offset, dim_sctr);
/* Maximise the expression L2 - L1. */
{
ppl_set_coef (le, dim_L2, 1);
ppl_set_coef (le, dim_L1, -1);
ppl_max_for_le_pointset (p1, le, stride);
- ppl_delete_Linear_Expression (le);
}
+
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, "\nStride in BB_%d, DR_%d, depth %d:",
+ pbb_index (pbb), PDR_ID (pdr), (int) depth);
+ value_print (dump_file, " %s ", stride);
+ }
+
+ ppl_delete_Pointset_Powerset_C_Polyhedron (p1);
+ ppl_delete_Pointset_Powerset_C_Polyhedron (p2);
+ ppl_delete_Linear_Expression (le);
+}
+
+/* Sets STRIDES to the sum of all the strides of the data references accessed */
+
+static void
+memory_strides_in_loop_depth (poly_bb_p pbb, graphite_dim_t depth, Value strides)
+{
+ int i;
+ poly_dr_p pdr;
+ Value s, n;
+
+ value_set_si (strides, 0);
+ value_init (s);
+ value_init (n);
+
+ for (i = 0; VEC_iterate (poly_dr_p, PBB_DRS (pbb), i, pdr); i++)
+ {
+ value_set_si (n, PDR_NB_REFS (pdr));
+
+ memory_stride_in_loop (s, depth, pdr);
+ value_multiply (s, s, n);
+ value_addto (strides, strides, s);
+ }
+
+ value_clear (s);
+ value_clear (n);
}
/* Returns true when it is profitable to interchange time dimensions DEPTH1
Next, to measure the impact of iterating once in loop "i", we build
a maximization problem: first, we add to DR accesses the dimensions
- k, s2, s3, L1 = 100 * s0 + s1, L2, and D1: polyhedron P1.
+ k, s2, s3, L1 = 100 * s0 + s1, L2, and D1: this is the polyhedron P1.
+ L1 and L2 are the linearized memory access functions.
| i j N a s0 s1 k s2 s3 L1 L2 D1 1
| 0 0 0 1 0 0 0 0 0 0 0 0 -5 = 0 alias = 5
pbb_interchange_profitable_p (graphite_dim_t depth1, graphite_dim_t depth2,
poly_bb_p pbb)
{
- int i;
- poly_dr_p pdr;
- Value d1, d2, s, n;
+ Value d1, d2;
bool res;
gcc_assert (depth1 < depth2);
value_init (d1);
- value_set_si (d1, 0);
value_init (d2);
- value_set_si (d2, 0);
- value_init (s);
- value_init (n);
-
- for (i = 0; VEC_iterate (poly_dr_p, PBB_DRS (pbb), i, pdr); i++)
- {
- value_set_si (n, PDR_NB_REFS (pdr));
-
- memory_stride_in_loop (s, depth1, pdr);
- value_multiply (s, s, n);
- value_addto (d1, d1, s);
- memory_stride_in_loop (s, depth2, pdr);
- value_multiply (s, s, n);
- value_addto (d2, d2, s);
- }
+ memory_strides_in_loop_depth (pbb, depth1, d1);
+ memory_strides_in_loop_depth (pbb, depth2, d2);
res = value_lt (d1, d2);
value_clear (d1);
value_clear (d2);
- value_clear (s);
- value_clear (n);
return res;
}
return pbb_interchange_profitable_p (depth1, depth2, LST_PBB (lst));
}
+/* Return true when the nest starting at LOOP1 and ending on LOOP2 is
+ perfect: i.e. there are no sequence of statements. */
+
+static bool
+lst_perfectly_nested_p (lst_p loop1, lst_p loop2)
+{
+ if (loop1 == loop2)
+ return true;
+
+ if (!LST_LOOP_P (loop1))
+ return false;
+
+ return VEC_length (lst_p, LST_SEQ (loop1)) == 1
+ && lst_perfectly_nested_p (VEC_index (lst_p, LST_SEQ (loop1), 0), loop2);
+}
+
+/* Transform the loop nest between LOOP1 and LOOP2 into a perfect
+ nest. To continue the naming tradition, this function is called
+ after perfect_nestify. NEST is set to the perfectly nested loop
+ that is created. BEFORE/AFTER are set to the loops distributed
+ before/after the loop NEST. */
+
+static void
+lst_perfect_nestify (lst_p loop1, lst_p loop2, lst_p *before,
+ lst_p *nest, lst_p *after)
+{
+ poly_bb_p first, last;
+
+ gcc_assert (loop1 && loop2
+ && loop1 != loop2
+ && LST_LOOP_P (loop1) && LST_LOOP_P (loop2));
+
+ first = LST_PBB (lst_find_first_pbb (loop2));
+ last = LST_PBB (lst_find_last_pbb (loop2));
+
+ *before = copy_lst (loop1);
+ *nest = copy_lst (loop1);
+ *after = copy_lst (loop1);
+
+ lst_remove_all_before_including_pbb (*before, first, false);
+ lst_remove_all_before_including_pbb (*after, last, true);
+
+ lst_remove_all_before_excluding_pbb (*nest, first, true);
+ lst_remove_all_before_excluding_pbb (*nest, last, false);
+}
/* Try to interchange LOOP1 with LOOP2 for all the statements of the
body of LOOP2. LOOP1 contains LOOP2. Return true if it did the
- interchange. */
+ interchange. CREATED_LOOP_BEFORE/CREATED_LOOP_AFTER are set to
+ true if the loop distribution created a loop before/after LOOP1. */
static bool
-lst_try_interchange_loops (scop_p scop, lst_p loop1, lst_p loop2)
+lst_try_interchange_loops (scop_p scop, lst_p loop1, lst_p loop2,
+ lst_p *before, lst_p *nest, lst_p *after)
{
int depth1 = lst_depth (loop1);
int depth2 = lst_depth (loop2);
+ lst_p transformed;
+
+ *before = NULL;
+ *after = NULL;
+ *nest = NULL;
if (!lst_interchange_profitable_p (loop2, depth1, depth2))
return false;
+ if (!lst_perfectly_nested_p (loop1, loop2))
+ lst_perfect_nestify (loop1, loop2, before, nest, after);
+
lst_apply_interchange (loop2, depth1, depth2);
+ /* Sync the transformed LST information and the PBB scatterings
+ before using the scatterings in the data dependence analysis. */
+ if (*before || *nest || *after)
+ {
+ transformed = lst_substitute_3 (SCOP_TRANSFORMED_SCHEDULE (scop), loop1,
+ *before, *nest, *after);
+ lst_update_scattering (transformed);
+ free_lst (transformed);
+ }
+
if (graphite_legal_transform (scop))
{
if (dump_file && (dump_flags & TDF_DETAILS))
"Loops at depths %d and %d will be interchanged.\n",
depth1, depth2);
+ /* Transform the SCOP_TRANSFORMED_SCHEDULE of the SCOP. */
+ lst_insert_in_sequence (*before, loop1, true);
+ lst_insert_in_sequence (*after, loop1, false);
+
+ if (*nest)
+ {
+ lst_replace (loop1, *nest);
+ free_lst (loop1);
+ }
+
return true;
}
/* Undo the transform. */
lst_apply_interchange (loop2, depth2, depth1);
+ *before = NULL;
+ *after = NULL;
+ *nest = NULL;
return false;
}
+static bool lst_do_interchange_1 (scop_p, lst_p, int *);
+
/* Try to interchange LOOP with all the loops contained in the body of
- LST. Return true if it did interchanged some loops. */
+ LST. Return true if it did interchanged some loops. INDEX points
+ to the next element to be processed by lst_do_interchange. */
static bool
-lst_try_interchange (scop_p scop, lst_p loop, lst_p lst)
+lst_try_interchange (scop_p scop, lst_p loop, lst_p lst, int *index)
{
- if (!lst)
+ int i;
+ lst_p l;
+ lst_p before, nest, after;
+ bool res;
+
+ if (!lst || !LST_LOOP_P (lst))
return false;
- if (LST_LOOP_P (lst))
+ res = lst_try_interchange_loops (scop, loop, lst, &before, &nest, &after);
+
+ if (before)
{
- int i;
- lst_p l;
- bool res = lst_try_interchange_loops (scop, loop, lst);
+ res |= lst_do_interchange_1 (scop, before, index);
+ (*index)++;
+ }
- for (i = 0; VEC_iterate (lst_p, LST_SEQ (lst), i, l); i++)
- res |= lst_try_interchange (scop, loop, l);
+ if (nest)
+ res |= lst_do_interchange_1 (scop, nest, index);
+ else
+ for (i = 0; VEC_iterate (lst_p, LST_SEQ (lst), i, l); i++)
+ res |= lst_try_interchange (scop, loop, l, index);
- return res;
+ if (after)
+ {
+ res |= lst_do_interchange_1 (scop, after, index);
+ (*index)++;
}
- return false;
+ (*index)++;
+ return res;
}
-/* Interchanges all the loops of LST that are considered profitable to
- interchange. Return true if it did interchanged some loops. */
+/* Interchanges all the loops of LOOP that are considered profitable
+ to interchange. Return true if it did interchanged some loops.
+ INDEX points to the next element to be processed by
+ lst_do_interchange. */
static bool
-lst_do_interchange (scop_p scop, lst_p lst)
+lst_do_interchange_1 (scop_p scop, lst_p loop, int *index)
{
- if (!lst)
+ int i;
+ lst_p l;
+ bool res = false;
+
+ if (!loop || !LST_LOOP_P (loop))
return false;
- if (LST_LOOP_P (lst))
- {
- int i;
- lst_p l;
- bool res = false;
+ for (i = 0; VEC_iterate (lst_p, LST_SEQ (loop), i, l); i++)
+ res |= lst_try_interchange (scop, loop, l, index);
- if (lst_depth (lst) >= 0)
- for (i = 0; VEC_iterate (lst_p, LST_SEQ (lst), i, l); i++)
- res |= lst_try_interchange (scop, lst, l);
+ return res;
+}
- for (i = 0; VEC_iterate (lst_p, LST_SEQ (lst), i, l); i++)
- res |= lst_do_interchange (scop, l);
+/* Interchanges all the loops of LOOP and the loops of its body that
+ are considered profitable to interchange. Return true if it did
+ interchanged some loops. INDEX points to the next element to be
+ processed in the LST_SEQ (LOOP) vector. */
- return res;
- }
+static bool
+lst_do_interchange (scop_p scop, lst_p loop, int *index)
+{
+ lst_p l;
+ bool res = false;
- return false;
+ if (!loop || !LST_LOOP_P (loop))
+ return false;
+
+ if (lst_depth (loop) >= 0)
+ res = lst_do_interchange_1 (scop, loop, index);
+
+ while (VEC_iterate (lst_p, LST_SEQ (loop), *index, l))
+ if (LST_LOOP_P (l))
+ res |= lst_do_interchange (scop, l, index);
+ else
+ (*index)++;
+
+ (*index)++;
+ return res;
}
/* Interchanges all the loop depths that are considered profitable for SCOP. */
bool
scop_do_interchange (scop_p scop)
{
- bool transform_done = false;
-
- store_scattering (scop);
-
- transform_done = lst_do_interchange (scop, SCOP_TRANSFORMED_SCHEDULE (scop));
+ int i = 0;
+ bool res = lst_do_interchange (scop, SCOP_TRANSFORMED_SCHEDULE (scop), &i);
- if (!transform_done)
- return false;
+ lst_update_scattering (SCOP_TRANSFORMED_SCHEDULE (scop));
- if (!graphite_legal_transform (scop))
- {
- restore_scattering (scop);
- return false;
- }
-
- return transform_done;
+ return res;
}