/* Conversion of SESE regions to Polyhedra.
- Copyright (C) 2009 Free Software Foundation, Inc.
+ Copyright (C) 2009, 2010 Free Software Foundation, Inc.
Contributed by Sebastian Pop <sebastian.pop@amd.com>.
This file is part of GCC.
static bool
var_used_in_not_loop_header_phi_node (tree var)
{
-
imm_use_iterator imm_iter;
gimple stmt;
bool result = false;
if (simple_copy_phi_p (phi))
{
- /* FIXME: PRE introduces phi nodes like these, for an example,
+ /* PRE introduces phi nodes like these, for an example,
see id-5.f in the fortran graphite testsuite:
# prephitmp.85_265 = PHI <prephitmp.85_258(33), prephitmp.85_265(18)>
reductions. */
if (simple_iv (loop, loop, res, &iv, true))
{
- gsi_next (psi);
+ if (integer_zerop (iv.step))
+ remove_invariant_phi (region, psi);
+ else
+ gsi_next (psi);
+
return false;
}
switch (gimple_code (stmt))
{
+ case GIMPLE_DEBUG:
/* Control flow expressions can be ignored, as they are
represented in the iteration domains and will be
regenerated by graphite. */
GBB_DATA_REFS (gbb) = drs;
GBB_CONDITIONS (gbb) = NULL;
GBB_CONDITION_CASES (gbb) = NULL;
- GBB_CLOOG_IV_TYPES (gbb) = NULL;
return gbb;
}
+static void
+free_data_refs_aux (VEC (data_reference_p, heap) *datarefs)
+{
+ unsigned int i;
+ struct data_reference *dr;
+
+ for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++)
+ if (dr->aux)
+ {
+ base_alias_pair *bap = (base_alias_pair *)(dr->aux);
+
+ if (bap->alias_set)
+ free (bap->alias_set);
+
+ free (bap);
+ dr->aux = NULL;
+ }
+}
/* Frees GBB. */
static void
free_gimple_bb (struct gimple_bb *gbb)
{
- if (GBB_CLOOG_IV_TYPES (gbb))
- htab_delete (GBB_CLOOG_IV_TYPES (gbb));
-
+ free_data_refs_aux (GBB_DATA_REFS (gbb));
free_data_refs (GBB_DATA_REFS (gbb));
VEC_free (gimple, heap, GBB_CONDITIONS (gbb));
information. */
static void
-try_generate_gimple_bb (scop_p scop, basic_block bb)
+try_generate_gimple_bb (scop_p scop, basic_block bb, sbitmap reductions)
{
VEC (data_reference_p, heap) *drs = VEC_alloc (data_reference_p, heap, 5);
loop_p nest = outermost_loop_in_sese (SCOP_REGION (scop), bb);
gimple_stmt_iterator gsi;
for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
- graphite_find_data_references_in_stmt (nest, gsi_stmt (gsi), &drs);
+ {
+ gimple stmt = gsi_stmt (gsi);
+ if (!is_gimple_debug (stmt))
+ graphite_find_data_references_in_stmt (nest, stmt, &drs);
+ }
if (!graphite_stmt_p (SCOP_REGION (scop), bb, drs))
free_data_refs (drs);
else
- new_poly_bb (scop, new_gimple_bb (bb, drs));
+ new_poly_bb (scop, new_gimple_bb (bb, drs), TEST_BIT (reductions,
+ bb->index));
}
/* Returns true if all predecessors of BB, that are not dominated by BB, are
/* Recursive helper function for build_scops_bbs. */
static void
-build_scop_bbs_1 (scop_p scop, sbitmap visited, basic_block bb)
+build_scop_bbs_1 (scop_p scop, sbitmap visited, basic_block bb, sbitmap reductions)
{
sese region = SCOP_REGION (scop);
VEC (basic_block, heap) *dom;
|| !bb_in_sese_p (bb, region))
return;
- try_generate_gimple_bb (scop, bb);
+ try_generate_gimple_bb (scop, bb, reductions);
SET_BIT (visited, bb->index);
dom = get_dominated_by (CDI_DOMINATORS, bb);
for (i = 0; VEC_iterate (basic_block, dom, i, dom_bb); i++)
if (all_non_dominated_preds_marked_p (dom_bb, visited))
{
- build_scop_bbs_1 (scop, visited, dom_bb);
+ build_scop_bbs_1 (scop, visited, dom_bb, reductions);
VEC_unordered_remove (basic_block, dom, i);
break;
}
/* Gather the basic blocks belonging to the SCOP. */
-void
-build_scop_bbs (scop_p scop)
+static void
+build_scop_bbs (scop_p scop, sbitmap reductions)
{
sbitmap visited = sbitmap_alloc (last_basic_block);
sese region = SCOP_REGION (scop);
sbitmap_zero (visited);
- build_scop_bbs_1 (scop, visited, SESE_ENTRY_BB (region));
-
+ build_scop_bbs_1 (scop, visited, SESE_ENTRY_BB (region), reductions);
sbitmap_free (visited);
}
value_init (v);
ppl_new_Coefficient (&c);
+ PBB_TRANSFORMED (pbb) = poly_scattering_new ();
ppl_new_C_Polyhedron_from_space_dimension
(&PBB_TRANSFORMED_SCATTERING (pbb), dim, 0);
value_clear (v);
ppl_delete_Coefficient (c);
- ppl_new_C_Polyhedron_from_C_Polyhedron (&PBB_ORIGINAL_SCATTERING (pbb),
- PBB_TRANSFORMED_SCATTERING (pbb));
+ PBB_ORIGINAL (pbb) = poly_scattering_copy (PBB_TRANSFORMED (pbb));
}
/* Build for BB the static schedule.
ppl_delete_Coefficient (coef);
}
-/* Saves in NV at index I a new name for variable P. */
-
-static void
-save_var_name (char **nv, int i, tree p)
-{
- const char *name = get_name (SSA_NAME_VAR (p));
-
- if (name)
- {
- int len = strlen (name) + 16;
- nv[i] = XNEWVEC (char, len);
- snprintf (nv[i], len, "%s_%d", name, SSA_NAME_VERSION (p));
- }
- else
- {
- nv[i] = XNEWVEC (char, 16);
- snprintf (nv[i], 2 + 16, "T_%d", SSA_NAME_VERSION (p));
- }
-}
-
/* When parameter NAME is in REGION, returns its index in SESE_PARAMS.
Otherwise returns -1. */
gcc_assert (SESE_ADD_PARAMS (region));
i = VEC_length (tree, SESE_PARAMS (region));
- save_var_name (SESE_PARAMS_NAMES (region), i, name);
- save_clast_name_index (SESE_PARAMS_INDEX (region),
- SESE_PARAMS_NAMES (region)[i], i);
VEC_safe_push (tree, heap, SESE_PARAMS (region), name);
return i;
}
}
}
-/* Data structure for idx_record_params. */
-
-struct irp_data
-{
- struct loop *loop;
- sese region;
-};
-
-/* For a data reference with an ARRAY_REF as its BASE, record the
- parameters occurring in IDX. DTA is passed in as complementary
- information, and is used by the automatic walker function. This
- function is a callback for for_each_index. */
-
-static bool
-idx_record_params (tree base, tree *idx, void *dta)
-{
- struct irp_data *data = (struct irp_data *) dta;
-
- if (TREE_CODE (base) != ARRAY_REF)
- return true;
-
- if (TREE_CODE (*idx) == SSA_NAME)
- {
- tree scev;
- sese region = data->region;
- struct loop *loop = data->loop;
- Value one;
-
- scev = scalar_evolution_in_region (region, loop, *idx);
-
- value_init (one);
- value_set_si (one, 1);
- scan_tree_for_params (region, scev, NULL, one);
- value_clear (one);
- }
-
- return true;
-}
-
/* Find parameters with respect to REGION in BB. We are looking in memory
access functions, conditions and loop bounds. */
find_params_in_bb (sese region, gimple_bb_p gbb)
{
int i;
+ unsigned j;
data_reference_p dr;
gimple stmt;
loop_p loop = GBB_BB (gbb)->loop_father;
+ Value one;
- for (i = 0; VEC_iterate (data_reference_p, GBB_DATA_REFS (gbb), i, dr); i++)
- {
- struct irp_data irp;
+ value_init (one);
+ value_set_si (one, 1);
- irp.loop = loop;
- irp.region = region;
- for_each_index (&dr->ref, idx_record_params, &irp);
- }
+ /* Find parameters in the access functions of data references. */
+ for (i = 0; VEC_iterate (data_reference_p, GBB_DATA_REFS (gbb), i, dr); i++)
+ for (j = 0; j < DR_NUM_DIMENSIONS (dr); j++)
+ scan_tree_for_params (region, DR_ACCESS_FN (dr, j), NULL, one);
/* Find parameters in conditional statements. */
for (i = 0; VEC_iterate (gimple, GBB_CONDITIONS (gbb), i, stmt); i++)
{
- Value one;
tree lhs = scalar_evolution_in_region (region, loop,
gimple_cond_lhs (stmt));
tree rhs = scalar_evolution_in_region (region, loop,
gimple_cond_rhs (stmt));
- value_init (one);
- value_set_si (one, 1);
scan_tree_for_params (region, lhs, NULL, one);
scan_tree_for_params (region, rhs, NULL, one);
- value_clear (one);
}
+
+ value_clear (one);
}
/* Record the parameters used in the SCOP. A variable is a parameter
scop_set_nb_params (scop, sese_nb_params (region));
SESE_ADD_PARAMS (region) = false;
+
+ ppl_new_Pointset_Powerset_C_Polyhedron_from_space_dimension
+ (&SCOP_CONTEXT (scop), scop_nb_params (scop), 0);
}
/* Returns a gimple_bb from BB. */
return (gimple_bb_p) bb->aux;
}
+/* Insert in the SCOP context constraints from the estimation of the
+ number of iterations. UB_EXPR is a linear expression describing
+ the number of iterations in a loop. This expression is bounded by
+ the estimation NIT. */
+
+static void
+add_upper_bounds_from_estimated_nit (scop_p scop, double_int nit,
+ ppl_dimension_type dim,
+ ppl_Linear_Expression_t ub_expr)
+{
+ Value val;
+ ppl_Linear_Expression_t nb_iters_le;
+ ppl_Polyhedron_t pol;
+ ppl_Coefficient_t coef;
+ ppl_Constraint_t ub;
+
+ ppl_new_Linear_Expression_with_dimension (&ub_expr, dim);
+ ppl_new_C_Polyhedron_from_space_dimension (&pol, dim, 0);
+ ppl_new_Linear_Expression_from_Linear_Expression (&nb_iters_le,
+ ub_expr);
+
+ /* Construct the negated number of last iteration in VAL. */
+ value_init (val);
+ mpz_set_double_int (val, nit, false);
+ value_sub_int (val, val, 1);
+ value_oppose (val, val);
+
+ /* NB_ITERS_LE holds the number of last iteration in
+ parametrical form. Subtract estimated number of last
+ iteration and assert that result is not positive. */
+ ppl_new_Coefficient_from_mpz_t (&coef, val);
+ ppl_Linear_Expression_add_to_inhomogeneous (nb_iters_le, coef);
+ ppl_delete_Coefficient (coef);
+ ppl_new_Constraint (&ub, nb_iters_le,
+ PPL_CONSTRAINT_TYPE_LESS_OR_EQUAL);
+ ppl_Polyhedron_add_constraint (pol, ub);
+
+ /* Remove all but last GDIM dimensions from POL to obtain
+ only the constraints on the parameters. */
+ {
+ graphite_dim_t gdim = scop_nb_params (scop);
+ ppl_dimension_type *dims = XNEWVEC (ppl_dimension_type, dim - gdim);
+ graphite_dim_t i;
+
+ for (i = 0; i < dim - gdim; i++)
+ dims[i] = i;
+
+ ppl_Polyhedron_remove_space_dimensions (pol, dims, dim - gdim);
+ XDELETEVEC (dims);
+ }
+
+ /* Add the constraints on the parameters to the SCoP context. */
+ {
+ ppl_Pointset_Powerset_C_Polyhedron_t constraints_ps;
+
+ ppl_new_Pointset_Powerset_C_Polyhedron_from_C_Polyhedron
+ (&constraints_ps, pol);
+ ppl_Pointset_Powerset_C_Polyhedron_intersection_assign
+ (SCOP_CONTEXT (scop), constraints_ps);
+ ppl_delete_Pointset_Powerset_C_Polyhedron (constraints_ps);
+ }
+
+ ppl_delete_Polyhedron (pol);
+ ppl_delete_Linear_Expression (nb_iters_le);
+ ppl_delete_Constraint (ub);
+ value_clear (val);
+}
+
/* Builds the constraint polyhedra for LOOP in SCOP. OUTER_PH gives
the constraints for the surrounding loops. */
static void
build_loop_iteration_domains (scop_p scop, struct loop *loop,
- ppl_Polyhedron_t outer_ph, int nb)
-
+ ppl_Polyhedron_t outer_ph, int nb,
+ ppl_Pointset_Powerset_C_Polyhedron_t *domains)
{
int i;
ppl_Polyhedron_t ph;
Value one;
ppl_Constraint_t ub;
ppl_Linear_Expression_t ub_expr;
+ double_int nit;
value_init (one);
value_set_si (one, 1);
scan_tree_for_params (SCOP_REGION (scop), nb_iters, ub_expr, one);
value_clear (one);
+ if (estimated_loop_iterations (loop, true, &nit))
+ add_upper_bounds_from_estimated_nit (scop, nit, dim, ub_expr);
+
/* loop_i <= expr_nb_iters */
ppl_set_coef (ub_expr, nb, -1);
ppl_new_Constraint (&ub, ub_expr, PPL_CONSTRAINT_TYPE_GREATER_OR_EQUAL);
gcc_unreachable ();
if (loop->inner && loop_in_sese_p (loop->inner, region))
- build_loop_iteration_domains (scop, loop->inner, ph, nb + 1);
+ build_loop_iteration_domains (scop, loop->inner, ph, nb + 1, domains);
if (nb != 0
&& loop->next
&& loop_in_sese_p (loop->next, region))
- build_loop_iteration_domains (scop, loop->next, outer_ph, nb);
+ build_loop_iteration_domains (scop, loop->next, outer_ph, nb, domains);
ppl_new_Pointset_Powerset_C_Polyhedron_from_C_Polyhedron
- ((ppl_Pointset_Powerset_C_Polyhedron_t *) &loop->aux, ph);
+ (&domains[loop->num], ph);
ppl_delete_Polyhedron (ph);
}
ppl_Linear_Expression_t res;
ppl_dimension_type dim;
sese region = SCOP_REGION (PBB_SCOP (pbb));
- loop_p loop = GBB_BB (PBB_BLACK_BOX (pbb))->loop_father;
+ loop_p loop = pbb_loop (pbb);
dim = pbb_dim_iter_domain (pbb) + pbb_nb_params (pbb);
ppl_new_Linear_Expression_with_dimension (&res, dim);
ppl_Linear_Expression_t le;
tree parameter = VEC_index (tree, SESE_PARAMS (SCOP_REGION (scop)), p);
tree type = TREE_TYPE (parameter);
- tree lb, ub;
-
- /* Disabled until we fix CPU2006. */
- return;
+ tree lb = NULL_TREE;
+ tree ub = NULL_TREE;
- if (!INTEGRAL_TYPE_P (type))
- return;
+ if (POINTER_TYPE_P (type) || !TYPE_MIN_VALUE (type))
+ lb = lower_bound_in_type (type, type);
+ else
+ lb = TYPE_MIN_VALUE (type);
- lb = TYPE_MIN_VALUE (type);
- ub = TYPE_MAX_VALUE (type);
+ if (POINTER_TYPE_P (type) || !TYPE_MAX_VALUE (type))
+ ub = upper_bound_in_type (type, type);
+ else
+ ub = TYPE_MAX_VALUE (type);
if (lb)
{
build_scop_context (scop_p scop)
{
ppl_Polyhedron_t context;
+ ppl_Pointset_Powerset_C_Polyhedron_t ps;
graphite_dim_t p, n = scop_nb_params (scop);
ppl_new_C_Polyhedron_from_space_dimension (&context, n, 0);
add_param_constraints (scop, context, p);
ppl_new_Pointset_Powerset_C_Polyhedron_from_C_Polyhedron
- (&SCOP_CONTEXT (scop), context);
+ (&ps, context);
+ ppl_Pointset_Powerset_C_Polyhedron_intersection_assign
+ (SCOP_CONTEXT (scop), ps);
+ ppl_delete_Pointset_Powerset_C_Polyhedron (ps);
ppl_delete_Polyhedron (context);
}
int i;
ppl_Polyhedron_t ph;
poly_bb_p pbb;
+ int nb_loops = number_of_loops ();
+ ppl_Pointset_Powerset_C_Polyhedron_t *domains
+ = XNEWVEC (ppl_Pointset_Powerset_C_Polyhedron_t, nb_loops);
+
+ for (i = 0; i < nb_loops; i++)
+ domains[i] = NULL;
ppl_new_C_Polyhedron_from_space_dimension (&ph, scop_nb_params (scop), 0);
for (i = 0; VEC_iterate (loop_p, SESE_LOOP_NEST (region), i, loop); i++)
if (!loop_in_sese_p (loop_outer (loop), region))
- build_loop_iteration_domains (scop, loop, ph, 0);
+ build_loop_iteration_domains (scop, loop, ph, 0, domains);
for (i = 0; VEC_iterate (poly_bb_p, SCOP_BBS (scop), i, pbb); i++)
- if (gbb_loop (PBB_BLACK_BOX (pbb))->aux)
+ if (domains[gbb_loop (PBB_BLACK_BOX (pbb))->num])
ppl_new_Pointset_Powerset_C_Polyhedron_from_Pointset_Powerset_C_Polyhedron
(&PBB_DOMAIN (pbb), (ppl_const_Pointset_Powerset_C_Polyhedron_t)
- gbb_loop (PBB_BLACK_BOX (pbb))->aux);
+ domains[gbb_loop (PBB_BLACK_BOX (pbb))->num]);
else
ppl_new_Pointset_Powerset_C_Polyhedron_from_C_Polyhedron
(&PBB_DOMAIN (pbb), ph);
- for (i = 0; VEC_iterate (loop_p, SESE_LOOP_NEST (region), i, loop); i++)
- if (loop->aux)
- {
- ppl_delete_Pointset_Powerset_C_Polyhedron
- ((ppl_Pointset_Powerset_C_Polyhedron_t) loop->aux);
- loop->aux = NULL;
- }
+ for (i = 0; i < nb_loops; i++)
+ if (domains[i])
+ ppl_delete_Pointset_Powerset_C_Polyhedron (domains[i]);
ppl_delete_Polyhedron (ph);
+ free (domains);
}
/* Add a constrain to the ACCESSES polyhedron for the alias set of
ppl_Linear_Expression_t alias;
ppl_Constraint_t cstr;
int alias_set_num = 0;
+ base_alias_pair *bap = (base_alias_pair *)(dr->aux);
- if (dr->aux != NULL)
- {
- alias_set_num = *((int *)(dr->aux));
- free (dr->aux);
- dr->aux = NULL;
- }
+ if (bap && bap->alias_set)
+ alias_set_num = *(bap->alias_set);
ppl_new_Linear_Expression_with_dimension (&alias, accessp_nb_dims);
}
/* Add constrains representing the size of the accessed data to the
- DATA_CONTAINER polyhedron. ACCESSP_NB_DIMS is the dimension of the
- DATA_CONTAINER polyhedron, DOM_NB_DIMS is the dimension of the iteration
+ ACCESSES polyhedron. ACCESSP_NB_DIMS is the dimension of the
+ ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration
domain. */
static void
-pdr_add_data_dimensions (ppl_Polyhedron_t data_container, data_reference_p dr,
+pdr_add_data_dimensions (ppl_Polyhedron_t accesses, data_reference_p dr,
ppl_dimension_type accessp_nb_dims,
ppl_dimension_type dom_nb_dims)
{
tree ref = DR_REF (dr);
int i, nb_subscripts = DR_NUM_DIMENSIONS (dr);
- tree array_size;
- HOST_WIDE_INT elt_size;
- array_size = TYPE_SIZE (TREE_TYPE (ref));
- if (array_size == NULL_TREE
- || TREE_CODE (array_size) != INTEGER_CST)
- return;
-
- elt_size = int_cst_value (array_size);
-
- for (i = nb_subscripts - 1; i >= 0; i--)
+ for (i = nb_subscripts - 1; i >= 0; i--, ref = TREE_OPERAND (ref, 0))
{
ppl_Linear_Expression_t expr;
ppl_Constraint_t cstr;
ppl_dimension_type subscript = dom_nb_dims + 1 + i;
+ tree low, high;
- /* 0 <= subscript */
- ppl_new_Linear_Expression_with_dimension (&expr, accessp_nb_dims);
- ppl_set_coef (expr, subscript, 1);
- ppl_new_Constraint (&cstr, expr, PPL_CONSTRAINT_TYPE_GREATER_OR_EQUAL);
- ppl_Polyhedron_add_constraint (data_container, cstr);
- ppl_delete_Linear_Expression (expr);
- ppl_delete_Constraint (cstr);
-
- ref = TREE_OPERAND (ref, 0);
- array_size = TYPE_SIZE (TREE_TYPE (ref));
- if (array_size == NULL_TREE
- || TREE_CODE (array_size) != INTEGER_CST)
+ if (TREE_CODE (ref) != ARRAY_REF)
break;
- /* subscript <= array_size */
- ppl_new_Linear_Expression_with_dimension (&expr, accessp_nb_dims);
- ppl_set_coef (expr, subscript, -1);
+ low = array_ref_low_bound (ref);
- if (elt_size)
- ppl_set_inhomogeneous (expr, int_cst_value (array_size) / elt_size);
+ /* subscript - low >= 0 */
+ if (host_integerp (low, 0))
+ {
+ ppl_new_Linear_Expression_with_dimension (&expr, accessp_nb_dims);
+ ppl_set_coef (expr, subscript, 1);
- ppl_new_Constraint (&cstr, expr, PPL_CONSTRAINT_TYPE_GREATER_OR_EQUAL);
- ppl_Polyhedron_add_constraint (data_container, cstr);
- ppl_delete_Linear_Expression (expr);
- ppl_delete_Constraint (cstr);
+ ppl_set_inhomogeneous (expr, -int_cst_value (low));
+
+ ppl_new_Constraint (&cstr, expr, PPL_CONSTRAINT_TYPE_GREATER_OR_EQUAL);
+ ppl_Polyhedron_add_constraint (accesses, cstr);
+ ppl_delete_Linear_Expression (expr);
+ ppl_delete_Constraint (cstr);
+ }
+
+ high = array_ref_up_bound (ref);
- elt_size = int_cst_value (array_size);
+ /* high - subscript >= 0 */
+ if (high && host_integerp (high, 0)
+ /* 1-element arrays at end of structures may extend over
+ their declared size. */
+ && !(array_at_struct_end_p (ref)
+ && operand_equal_p (low, high, 0)))
+ {
+ ppl_new_Linear_Expression_with_dimension (&expr, accessp_nb_dims);
+ ppl_set_coef (expr, subscript, -1);
+
+ ppl_set_inhomogeneous (expr, int_cst_value (high));
+
+ ppl_new_Constraint (&cstr, expr, PPL_CONSTRAINT_TYPE_GREATER_OR_EQUAL);
+ ppl_Polyhedron_add_constraint (accesses, cstr);
+ ppl_delete_Linear_Expression (expr);
+ ppl_delete_Constraint (cstr);
+ }
}
}
static void
build_poly_dr (data_reference_p dr, poly_bb_p pbb)
{
- ppl_Polyhedron_t accesses, data_container;
- ppl_Pointset_Powerset_C_Polyhedron_t accesses_ps, data_container_ps;
+ ppl_Polyhedron_t accesses;
+ ppl_Pointset_Powerset_C_Polyhedron_t accesses_ps;
ppl_dimension_type dom_nb_dims;
ppl_dimension_type accessp_nb_dims;
+ int dr_base_object_set;
ppl_Pointset_Powerset_C_Polyhedron_space_dimension (PBB_DOMAIN (pbb),
&dom_nb_dims);
accessp_nb_dims = dom_nb_dims + 1 + DR_NUM_DIMENSIONS (dr);
ppl_new_C_Polyhedron_from_space_dimension (&accesses, accessp_nb_dims, 0);
- ppl_new_C_Polyhedron_from_space_dimension (&data_container,
- accessp_nb_dims, 0);
pdr_add_alias_set (accesses, dr, accessp_nb_dims, dom_nb_dims);
pdr_add_memory_accesses (accesses, dr, accessp_nb_dims, dom_nb_dims, pbb);
- pdr_add_data_dimensions (data_container, dr, accessp_nb_dims, dom_nb_dims);
+ pdr_add_data_dimensions (accesses, dr, accessp_nb_dims, dom_nb_dims);
ppl_new_Pointset_Powerset_C_Polyhedron_from_C_Polyhedron (&accesses_ps,
accesses);
- ppl_new_Pointset_Powerset_C_Polyhedron_from_C_Polyhedron (&data_container_ps,
- data_container);
ppl_delete_Polyhedron (accesses);
- ppl_delete_Polyhedron (data_container);
- new_poly_dr (pbb, accesses_ps, data_container_ps,
- DR_IS_READ (dr) ? PDR_READ : PDR_WRITE, dr);
+
+ if (dr->aux)
+ dr_base_object_set = ((base_alias_pair *)(dr->aux))->base_obj_set;
+
+ new_poly_dr (pbb, dr_base_object_set, accesses_ps, DR_IS_READ (dr) ? PDR_READ : PDR_WRITE,
+ dr, DR_NUM_DIMENSIONS (dr));
+}
+
+/* Write to FILE the alias graph of data references in DIMACS format. */
+
+static inline bool
+write_alias_graph_to_ascii_dimacs (FILE *file, char *comment,
+ VEC (data_reference_p, heap) *drs)
+{
+ int num_vertex = VEC_length (data_reference_p, drs);
+ int edge_num = 0;
+ data_reference_p dr1, dr2;
+ int i, j;
+
+ if (num_vertex == 0)
+ return true;
+
+ for (i = 0; VEC_iterate (data_reference_p, drs, i, dr1); i++)
+ for (j = i + 1; VEC_iterate (data_reference_p, drs, j, dr2); j++)
+ if (dr_may_alias_p (dr1, dr2))
+ edge_num++;
+
+ fprintf (file, "$\n");
+
+ if (comment)
+ fprintf (file, "c %s\n", comment);
+
+ fprintf (file, "p edge %d %d\n", num_vertex, edge_num);
+
+ for (i = 0; VEC_iterate (data_reference_p, drs, i, dr1); i++)
+ for (j = i + 1; VEC_iterate (data_reference_p, drs, j, dr2); j++)
+ if (dr_may_alias_p (dr1, dr2))
+ fprintf (file, "e %d %d\n", i + 1, j + 1);
+
+ return true;
+}
+
+/* Write to FILE the alias graph of data references in DOT format. */
+
+static inline bool
+write_alias_graph_to_ascii_dot (FILE *file, char *comment,
+ VEC (data_reference_p, heap) *drs)
+{
+ int num_vertex = VEC_length (data_reference_p, drs);
+ data_reference_p dr1, dr2;
+ int i, j;
+
+ if (num_vertex == 0)
+ return true;
+
+ fprintf (file, "$\n");
+
+ if (comment)
+ fprintf (file, "c %s\n", comment);
+
+ /* First print all the vertices. */
+ for (i = 0; VEC_iterate (data_reference_p, drs, i, dr1); i++)
+ fprintf (file, "n%d;\n", i);
+
+ for (i = 0; VEC_iterate (data_reference_p, drs, i, dr1); i++)
+ for (j = i + 1; VEC_iterate (data_reference_p, drs, j, dr2); j++)
+ if (dr_may_alias_p (dr1, dr2))
+ fprintf (file, "n%d n%d\n", i, j);
+
+ return true;
+}
+
+/* Write to FILE the alias graph of data references in ECC format. */
+
+static inline bool
+write_alias_graph_to_ascii_ecc (FILE *file, char *comment,
+ VEC (data_reference_p, heap) *drs)
+{
+ int num_vertex = VEC_length (data_reference_p, drs);
+ data_reference_p dr1, dr2;
+ int i, j;
+
+ if (num_vertex == 0)
+ return true;
+
+ fprintf (file, "$\n");
+
+ if (comment)
+ fprintf (file, "c %s\n", comment);
+
+ for (i = 0; VEC_iterate (data_reference_p, drs, i, dr1); i++)
+ for (j = i + 1; VEC_iterate (data_reference_p, drs, j, dr2); j++)
+ if (dr_may_alias_p (dr1, dr2))
+ fprintf (file, "%d %d\n", i, j);
+
+ return true;
+}
+
+/* Check if DR1 and DR2 are in the same object set. */
+
+static bool
+dr_same_base_object_p (const struct data_reference *dr1,
+ const struct data_reference *dr2)
+{
+ return operand_equal_p (DR_BASE_OBJECT (dr1), DR_BASE_OBJECT (dr2), 0);
+}
+
+/* Uses DFS component number as representative of alias-sets. Also tests for
+ optimality by verifying if every connected component is a clique. Returns
+ true (1) if the above test is true, and false (0) otherwise. */
+
+static int
+build_alias_set_optimal_p (VEC (data_reference_p, heap) *drs)
+{
+ int num_vertices = VEC_length (data_reference_p, drs);
+ struct graph *g = new_graph (num_vertices);
+ data_reference_p dr1, dr2;
+ int i, j;
+ int num_connected_components;
+ int v_indx1, v_indx2, num_vertices_in_component;
+ int *all_vertices;
+ int *vertices;
+ struct graph_edge *e;
+ int this_component_is_clique;
+ int all_components_are_cliques = 1;
+
+ for (i = 0; VEC_iterate (data_reference_p, drs, i, dr1); i++)
+ for (j = i+1; VEC_iterate (data_reference_p, drs, j, dr2); j++)
+ if (dr_may_alias_p (dr1, dr2))
+ {
+ add_edge (g, i, j);
+ add_edge (g, j, i);
+ }
+
+ all_vertices = XNEWVEC (int, num_vertices);
+ vertices = XNEWVEC (int, num_vertices);
+ for (i = 0; i < num_vertices; i++)
+ all_vertices[i] = i;
+
+ num_connected_components = graphds_dfs (g, all_vertices, num_vertices,
+ NULL, true, NULL);
+ for (i = 0; i < g->n_vertices; i++)
+ {
+ data_reference_p dr = VEC_index (data_reference_p, drs, i);
+ base_alias_pair *bap;
+
+ if (dr->aux)
+ bap = (base_alias_pair *)(dr->aux);
+
+ bap->alias_set = XNEW (int);
+ *(bap->alias_set) = g->vertices[i].component + 1;
+ }
+
+ /* Verify if the DFS numbering results in optimal solution. */
+ for (i = 0; i < num_connected_components; i++)
+ {
+ num_vertices_in_component = 0;
+ /* Get all vertices whose DFS component number is the same as i. */
+ for (j = 0; j < num_vertices; j++)
+ if (g->vertices[j].component == i)
+ vertices[num_vertices_in_component++] = j;
+
+ /* Now test if the vertices in 'vertices' form a clique, by testing
+ for edges among each pair. */
+ this_component_is_clique = 1;
+ for (v_indx1 = 0; v_indx1 < num_vertices_in_component; v_indx1++)
+ {
+ for (v_indx2 = v_indx1+1; v_indx2 < num_vertices_in_component; v_indx2++)
+ {
+ /* Check if the two vertices are connected by iterating
+ through all the edges which have one of these are source. */
+ e = g->vertices[vertices[v_indx2]].pred;
+ while (e)
+ {
+ if (e->src == vertices[v_indx1])
+ break;
+ e = e->pred_next;
+ }
+ if (!e)
+ {
+ this_component_is_clique = 0;
+ break;
+ }
+ }
+ if (!this_component_is_clique)
+ all_components_are_cliques = 0;
+ }
+ }
+
+ free (all_vertices);
+ free (vertices);
+ free_graph (g);
+ return all_components_are_cliques;
}
-/* Group each data reference in DRS with it's alias set num. */
+/* Group each data reference in DRS with it's base object set num. */
static void
-build_alias_set_for_drs (VEC (data_reference_p, heap) **drs)
+build_base_obj_set_for_drs (VEC (data_reference_p, heap) *drs)
{
- int num_vertex = VEC_length (data_reference_p, *drs);
+ int num_vertex = VEC_length (data_reference_p, drs);
struct graph *g = new_graph (num_vertex);
data_reference_p dr1, dr2;
int i, j;
- int num_component;
int *queue;
- for (i = 0; VEC_iterate (data_reference_p, *drs, i, dr1); i++)
- for (j = i+1; VEC_iterate (data_reference_p, *drs, j, dr2); j++)
- if (dr_may_alias_p (dr1, dr2))
+ for (i = 0; VEC_iterate (data_reference_p, drs, i, dr1); i++)
+ for (j = i + 1; VEC_iterate (data_reference_p, drs, j, dr2); j++)
+ if (dr_same_base_object_p (dr1, dr2))
{
add_edge (g, i, j);
add_edge (g, j, i);
for (i = 0; i < num_vertex; i++)
queue[i] = i;
- num_component = graphds_dfs (g, queue, num_vertex, NULL, true, NULL);
+ graphds_dfs (g, queue, num_vertex, NULL, true, NULL);
for (i = 0; i < g->n_vertices; i++)
{
- data_reference_p dr = VEC_index (data_reference_p, *drs, i);
- dr->aux = XNEW (int);
- *((int *)(dr->aux)) = g->vertices[i].component + 1;
+ data_reference_p dr = VEC_index (data_reference_p, drs, i);
+ base_alias_pair *bap;
+
+ if (dr->aux)
+ bap = (base_alias_pair *)(dr->aux);
+
+ bap->base_obj_set = g->vertices[i].component + 1;
}
free (queue);
data_reference_p dr;
VEC (data_reference_p, heap) *gbb_drs = GBB_DATA_REFS (PBB_BLACK_BOX (pbb));
- build_alias_set_for_drs (&gbb_drs);
-
for (j = 0; VEC_iterate (data_reference_p, gbb_drs, j, dr); j++)
build_poly_dr (dr, pbb);
}
+/* Dump to file the alias graphs for the data references in DRS. */
+
+static void
+dump_alias_graphs (VEC (data_reference_p, heap) *drs)
+{
+ char comment[100];
+ FILE *file_dimacs, *file_ecc, *file_dot;
+
+ file_dimacs = fopen ("/tmp/dr_alias_graph_dimacs", "ab");
+ if (file_dimacs)
+ {
+ snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
+ current_function_name ());
+ write_alias_graph_to_ascii_dimacs (file_dimacs, comment, drs);
+ fclose (file_dimacs);
+ }
+
+ file_ecc = fopen ("/tmp/dr_alias_graph_ecc", "ab");
+ if (file_ecc)
+ {
+ snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
+ current_function_name ());
+ write_alias_graph_to_ascii_ecc (file_ecc, comment, drs);
+ fclose (file_ecc);
+ }
+
+ file_dot = fopen ("/tmp/dr_alias_graph_dot", "ab");
+ if (file_dot)
+ {
+ snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
+ current_function_name ());
+ write_alias_graph_to_ascii_dot (file_dot, comment, drs);
+ fclose (file_dot);
+ }
+}
+
/* Build data references in SCOP. */
static void
build_scop_drs (scop_p scop)
{
- int i;
+ int i, j;
poly_bb_p pbb;
+ data_reference_p dr;
+ VEC (data_reference_p, heap) *drs = VEC_alloc (data_reference_p, heap, 3);
+
+ for (i = 0; VEC_iterate (poly_bb_p, SCOP_BBS (scop), i, pbb); i++)
+ for (j = 0; VEC_iterate (data_reference_p,
+ GBB_DATA_REFS (PBB_BLACK_BOX (pbb)), j, dr); j++)
+ VEC_safe_push (data_reference_p, heap, drs, dr);
+
+ for (i = 0; VEC_iterate (data_reference_p, drs, i, dr); i++)
+ dr->aux = XNEW (base_alias_pair);
+
+ if (!build_alias_set_optimal_p (drs))
+ {
+ /* TODO: Add support when building alias set is not optimal. */
+ ;
+ }
+
+ build_base_obj_set_for_drs (drs);
+
+ /* When debugging, enable the following code. This cannot be used
+ in production compilers. */
+ if (0)
+ dump_alias_graphs (drs);
+
+ VEC_free (data_reference_p, heap, drs);
for (i = 0; VEC_iterate (poly_bb_p, SCOP_BBS (scop), i, pbb); i++)
build_pbb_drs (pbb);
}
-/* Return a gsi at the position of the VAR definition. */
+/* Return a gsi at the position of the phi node STMT. */
static gimple_stmt_iterator
-gsi_for_ssa_name_def (tree var)
+gsi_for_phi_node (gimple stmt)
{
- gimple stmt;
- basic_block bb;
- gimple_stmt_iterator gsi;
gimple_stmt_iterator psi;
-
- gcc_assert (TREE_CODE (var) == SSA_NAME);
-
- stmt = SSA_NAME_DEF_STMT (var);
- bb = gimple_bb (stmt);
+ basic_block bb = gimple_bb (stmt);
for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
if (stmt == gsi_stmt (psi))
- return gsi_after_labels (bb);
-
- for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
- if (stmt == gsi_stmt (gsi))
- {
- gsi_next (&gsi);
- return gsi;
- }
+ return psi;
gcc_unreachable ();
- return gsi;
+ return psi;
}
/* Insert the assignment "RES := VAR" just after the definition of VAR. */
static void
insert_out_of_ssa_copy (tree res, tree var)
{
- gimple_stmt_iterator gsi = gsi_for_ssa_name_def (var);
gimple stmt;
gimple_seq stmts;
gimple_stmt_iterator si;
+ gimple_stmt_iterator gsi;
var = force_gimple_operand (var, &stmts, true, NULL_TREE);
stmt = gimple_build_assign (res, var);
stmts = gimple_seq_alloc ();
si = gsi_last (stmts);
gsi_insert_after (&si, stmt, GSI_NEW_STMT);
- gsi_insert_seq_before (&gsi, stmts, GSI_NEW_STMT);
+
+ stmt = SSA_NAME_DEF_STMT (var);
+ if (gimple_code (stmt) == GIMPLE_PHI)
+ {
+ gsi = gsi_after_labels (gimple_bb (stmt));
+ gsi_insert_seq_before (&gsi, stmts, GSI_NEW_STMT);
+ }
+ else
+ {
+ gsi = gsi_for_stmt (stmt);
+ gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT);
+ }
}
/* Insert on edge E the assignment "RES := EXPR". */
/* Creates a zero dimension array of the same type as VAR. */
static tree
-create_zero_dim_array (tree var)
+create_zero_dim_array (tree var, const char *base_name)
{
tree index_type = build_index_type (integer_zero_node);
tree elt_type = TREE_TYPE (var);
tree array_type = build_array_type (elt_type, index_type);
- tree base = create_tmp_var (array_type, "Red");
+ tree base = create_tmp_var (array_type, base_name);
add_referenced_var (base);
static bool
scalar_close_phi_node_p (gimple phi)
{
- gcc_assert (gimple_code (phi) == GIMPLE_PHI);
-
- if (!is_gimple_reg (gimple_phi_result (phi)))
+ if (gimple_code (phi) != GIMPLE_PHI
+ || !is_gimple_reg (gimple_phi_result (phi)))
return false;
+ /* Note that loop close phi nodes should have a single argument
+ because we translated the representation into a canonical form
+ before Graphite: see canonicalize_loop_closed_ssa_form. */
return (gimple_phi_num_args (phi) == 1);
}
gimple phi = gsi_stmt (*psi);
tree res = gimple_phi_result (phi);
tree var = SSA_NAME_VAR (res);
- tree zero_dim_array = create_zero_dim_array (var);
+ tree zero_dim_array = create_zero_dim_array (var, "Close_Phi");
gimple_stmt_iterator gsi = gsi_after_labels (gimple_bb (phi));
gimple stmt = gimple_build_assign (res, zero_dim_array);
tree arg = gimple_phi_arg_def (phi, 0);
- insert_out_of_ssa_copy (zero_dim_array, arg);
+ /* Note that loop close phi nodes should have a single argument
+ because we translated the representation into a canonical form
+ before Graphite: see canonicalize_loop_closed_ssa_form. */
+ gcc_assert (gimple_phi_num_args (phi) == 1);
+
+ if (TREE_CODE (arg) == SSA_NAME
+ && !SSA_NAME_IS_DEFAULT_DEF (arg))
+ insert_out_of_ssa_copy (zero_dim_array, arg);
+ else
+ insert_out_of_ssa_copy_on_edge (single_pred_edge (gimple_bb (phi)),
+ zero_dim_array, arg);
remove_phi_node (psi, false);
gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
basic_block bb = gimple_bb (phi);
tree res = gimple_phi_result (phi);
tree var = SSA_NAME_VAR (res);
- tree zero_dim_array = create_zero_dim_array (var);
+ tree zero_dim_array = create_zero_dim_array (var, "General_Reduction");
gimple_stmt_iterator gsi;
gimple stmt;
gimple_seq stmts;
| end_2
| end_1
- whereas inserting the copy on the incomming edge is correct
+ whereas inserting the copy on the incoming edge is correct
| a = ...
| loop_1
gsi_insert_seq_before (&gsi, stmts, GSI_NEW_STMT);
}
+/* Return true when DEF can be analyzed in REGION by the scalar
+ evolution analyzer. */
+
+static bool
+scev_analyzable_p (tree def, sese region)
+{
+ gimple stmt = SSA_NAME_DEF_STMT (def);
+ loop_p loop = loop_containing_stmt (stmt);
+ tree scev = scalar_evolution_in_region (region, loop, def);
+
+ return !chrec_contains_undetermined (scev);
+}
+
+/* Rewrite the scalar dependence of DEF used in USE_STMT with a memory
+ read from ZERO_DIM_ARRAY. */
+
+static void
+rewrite_cross_bb_scalar_dependence (tree zero_dim_array, tree def, gimple use_stmt)
+{
+ tree var = SSA_NAME_VAR (def);
+ gimple name_stmt = gimple_build_assign (var, zero_dim_array);
+ tree name = make_ssa_name (var, name_stmt);
+ ssa_op_iter iter;
+ use_operand_p use_p;
+ gimple_stmt_iterator gsi;
+
+ gcc_assert (gimple_code (use_stmt) != GIMPLE_PHI);
+
+ gimple_assign_set_lhs (name_stmt, name);
+
+ gsi = gsi_for_stmt (use_stmt);
+ gsi_insert_before (&gsi, name_stmt, GSI_NEW_STMT);
+
+ FOR_EACH_SSA_USE_OPERAND (use_p, use_stmt, iter, SSA_OP_ALL_USES)
+ if (operand_equal_p (def, USE_FROM_PTR (use_p), 0))
+ replace_exp (use_p, name);
+
+ update_stmt (use_stmt);
+}
+
+/* Rewrite the scalar dependences crossing the boundary of the BB
+ containing STMT with an array. */
+
+static void
+rewrite_cross_bb_scalar_deps (sese region, gimple_stmt_iterator *gsi)
+{
+ gimple stmt = gsi_stmt (*gsi);
+ imm_use_iterator imm_iter;
+ tree def;
+ basic_block def_bb;
+ tree zero_dim_array = NULL_TREE;
+ gimple use_stmt;
+
+ if (gimple_code (stmt) != GIMPLE_ASSIGN)
+ return;
+
+ def = gimple_assign_lhs (stmt);
+ if (!is_gimple_reg (def)
+ || scev_analyzable_p (def, region))
+ return;
+
+ def_bb = gimple_bb (stmt);
+
+ FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
+ if (def_bb != gimple_bb (use_stmt)
+ && gimple_code (use_stmt) != GIMPLE_PHI
+ && !is_gimple_debug (use_stmt))
+ {
+ if (!zero_dim_array)
+ {
+ zero_dim_array = create_zero_dim_array
+ (SSA_NAME_VAR (def), "Cross_BB_scalar_dependence");
+ insert_out_of_ssa_copy (zero_dim_array, def);
+ gsi_next (gsi);
+ }
+
+ rewrite_cross_bb_scalar_dependence (zero_dim_array, def, use_stmt);
+ }
+}
+
/* Rewrite out of SSA all the reduction phi nodes of SCOP. */
static void
sese region = SCOP_REGION (scop);
FOR_EACH_BB (bb)
- if (bb_in_region (bb, SESE_ENTRY_BB (region), SESE_EXIT_BB (region)))
+ if (bb_in_sese_p (bb, region))
for (psi = gsi_start_phis (bb); !gsi_end_p (psi);)
{
if (scalar_close_phi_node_p (gsi_stmt (psi)))
update_ssa (TODO_update_ssa);
#ifdef ENABLE_CHECKING
- verify_ssa (false);
- verify_loop_closed_ssa ();
+ verify_loop_closed_ssa (true);
+#endif
+
+ FOR_EACH_BB (bb)
+ if (bb_in_sese_p (bb, region))
+ for (psi = gsi_start_bb (bb); !gsi_end_p (psi); gsi_next (&psi))
+ rewrite_cross_bb_scalar_deps (region, &psi);
+
+ update_ssa (TODO_update_ssa);
+#ifdef ENABLE_CHECKING
+ verify_loop_closed_ssa (true);
#endif
}
return res;
}
+/* Return the number of data references in BB that write in
+ memory. */
+
+static int
+nb_data_writes_in_bb (basic_block bb)
+{
+ int res = 0;
+ gimple_stmt_iterator gsi;
+
+ for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
+ if (gimple_vdef (gsi_stmt (gsi)))
+ res++;
+
+ return res;
+}
+
+/* Splits STMT out of its current BB. */
+
+static basic_block
+split_reduction_stmt (gimple stmt)
+{
+ gimple_stmt_iterator gsi;
+ basic_block bb = gimple_bb (stmt);
+ edge e;
+
+ /* Do not split basic blocks with no writes to memory: the reduction
+ will be the only write to memory. */
+ if (nb_data_writes_in_bb (bb) == 0)
+ return bb;
+
+ split_block (bb, stmt);
+
+ if (gsi_one_before_end_p (gsi_start_nondebug_bb (bb)))
+ return bb;
+
+ gsi = gsi_last_bb (bb);
+ gsi_prev (&gsi);
+ e = split_block (bb, gsi_stmt (gsi));
+
+ return e->dest;
+}
+
+/* Return true when stmt is a reduction operation. */
+
+static inline bool
+is_reduction_operation_p (gimple stmt)
+{
+ enum tree_code code;
+
+ gcc_assert (is_gimple_assign (stmt));
+ code = gimple_assign_rhs_code (stmt);
+
+ return flag_associative_math
+ && commutative_tree_code (code)
+ && associative_tree_code (code);
+}
+
+/* Returns true when PHI contains an argument ARG. */
+
+static bool
+phi_contains_arg (gimple phi, tree arg)
+{
+ size_t i;
+
+ for (i = 0; i < gimple_phi_num_args (phi); i++)
+ if (operand_equal_p (arg, gimple_phi_arg_def (phi, i), 0))
+ return true;
+
+ return false;
+}
+
+/* Return a loop phi node that corresponds to a reduction containing LHS. */
+
+static gimple
+follow_ssa_with_commutative_ops (tree arg, tree lhs)
+{
+ gimple stmt;
+
+ if (TREE_CODE (arg) != SSA_NAME)
+ return NULL;
+
+ stmt = SSA_NAME_DEF_STMT (arg);
+
+ if (gimple_code (stmt) == GIMPLE_NOP
+ || gimple_code (stmt) == GIMPLE_CALL)
+ return NULL;
+
+ if (gimple_code (stmt) == GIMPLE_PHI)
+ {
+ if (phi_contains_arg (stmt, lhs))
+ return stmt;
+ return NULL;
+ }
+
+ if (!is_gimple_assign (stmt))
+ return NULL;
+
+ if (gimple_num_ops (stmt) == 2)
+ return follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
+
+ if (is_reduction_operation_p (stmt))
+ {
+ gimple res = follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
+
+ return res ? res :
+ follow_ssa_with_commutative_ops (gimple_assign_rhs2 (stmt), lhs);
+ }
+
+ return NULL;
+}
+
+/* Detect commutative and associative scalar reductions starting at
+ the STMT. Return the phi node of the reduction cycle, or NULL. */
+
+static gimple
+detect_commutative_reduction_arg (tree lhs, gimple stmt, tree arg,
+ VEC (gimple, heap) **in,
+ VEC (gimple, heap) **out)
+{
+ gimple phi = follow_ssa_with_commutative_ops (arg, lhs);
+
+ if (!phi)
+ return NULL;
+
+ VEC_safe_push (gimple, heap, *in, stmt);
+ VEC_safe_push (gimple, heap, *out, stmt);
+ return phi;
+}
+
+/* Detect commutative and associative scalar reductions starting at
+ the STMT. Return the phi node of the reduction cycle, or NULL. */
+
+static gimple
+detect_commutative_reduction_assign (gimple stmt, VEC (gimple, heap) **in,
+ VEC (gimple, heap) **out)
+{
+ tree lhs = gimple_assign_lhs (stmt);
+
+ if (gimple_num_ops (stmt) == 2)
+ return detect_commutative_reduction_arg (lhs, stmt,
+ gimple_assign_rhs1 (stmt),
+ in, out);
+
+ if (is_reduction_operation_p (stmt))
+ {
+ gimple res = detect_commutative_reduction_arg (lhs, stmt,
+ gimple_assign_rhs1 (stmt),
+ in, out);
+ return res ? res
+ : detect_commutative_reduction_arg (lhs, stmt,
+ gimple_assign_rhs2 (stmt),
+ in, out);
+ }
+
+ return NULL;
+}
+
+/* Return a loop phi node that corresponds to a reduction containing LHS. */
+
+static gimple
+follow_inital_value_to_phi (tree arg, tree lhs)
+{
+ gimple stmt;
+
+ if (!arg || TREE_CODE (arg) != SSA_NAME)
+ return NULL;
+
+ stmt = SSA_NAME_DEF_STMT (arg);
+
+ if (gimple_code (stmt) == GIMPLE_PHI
+ && phi_contains_arg (stmt, lhs))
+ return stmt;
+
+ return NULL;
+}
+
+
+/* Return the argument of the loop PHI that is the inital value coming
+ from outside the loop. */
+
+static edge
+edge_initial_value_for_loop_phi (gimple phi)
+{
+ size_t i;
+
+ for (i = 0; i < gimple_phi_num_args (phi); i++)
+ {
+ edge e = gimple_phi_arg_edge (phi, i);
+
+ if (loop_depth (e->src->loop_father)
+ < loop_depth (e->dest->loop_father))
+ return e;
+ }
+
+ return NULL;
+}
+
+/* Return the argument of the loop PHI that is the inital value coming
+ from outside the loop. */
+
+static tree
+initial_value_for_loop_phi (gimple phi)
+{
+ size_t i;
+
+ for (i = 0; i < gimple_phi_num_args (phi); i++)
+ {
+ edge e = gimple_phi_arg_edge (phi, i);
+
+ if (loop_depth (e->src->loop_father)
+ < loop_depth (e->dest->loop_father))
+ return gimple_phi_arg_def (phi, i);
+ }
+
+ return NULL_TREE;
+}
+
+/* Detect commutative and associative scalar reductions starting at
+ the loop closed phi node CLOSE_PHI. Return the phi node of the
+ reduction cycle, or NULL. */
+
+static gimple
+detect_commutative_reduction (gimple stmt, VEC (gimple, heap) **in,
+ VEC (gimple, heap) **out)
+{
+ if (scalar_close_phi_node_p (stmt))
+ {
+ tree arg = gimple_phi_arg_def (stmt, 0);
+ gimple def, loop_phi;
+
+ if (TREE_CODE (arg) != SSA_NAME)
+ return NULL;
+
+ /* Note that loop close phi nodes should have a single argument
+ because we translated the representation into a canonical form
+ before Graphite: see canonicalize_loop_closed_ssa_form. */
+ gcc_assert (gimple_phi_num_args (stmt) == 1);
+
+ def = SSA_NAME_DEF_STMT (arg);
+ loop_phi = detect_commutative_reduction (def, in, out);
+
+ if (loop_phi)
+ {
+ tree lhs = gimple_phi_result (stmt);
+ tree init = initial_value_for_loop_phi (loop_phi);
+ gimple phi = follow_inital_value_to_phi (init, lhs);
+
+ VEC_safe_push (gimple, heap, *in, loop_phi);
+ VEC_safe_push (gimple, heap, *out, stmt);
+ return phi;
+ }
+ else
+ return NULL;
+ }
+
+ if (gimple_code (stmt) == GIMPLE_ASSIGN)
+ return detect_commutative_reduction_assign (stmt, in, out);
+
+ return NULL;
+}
+
+/* Translate the scalar reduction statement STMT to an array RED
+ knowing that its recursive phi node is LOOP_PHI. */
+
+static void
+translate_scalar_reduction_to_array_for_stmt (tree red, gimple stmt,
+ gimple loop_phi)
+{
+ gimple_stmt_iterator insert_gsi = gsi_after_labels (gimple_bb (loop_phi));
+ tree res = gimple_phi_result (loop_phi);
+ gimple assign = gimple_build_assign (res, red);
+
+ gsi_insert_before (&insert_gsi, assign, GSI_SAME_STMT);
+
+ insert_gsi = gsi_after_labels (gimple_bb (stmt));
+ assign = gimple_build_assign (red, gimple_assign_lhs (stmt));
+ insert_gsi = gsi_for_stmt (stmt);
+ gsi_insert_after (&insert_gsi, assign, GSI_SAME_STMT);
+}
+
+/* Insert the assignment "result (CLOSE_PHI) = RED". */
+
+static void
+insert_copyout (tree red, gimple close_phi)
+{
+ tree res = gimple_phi_result (close_phi);
+ basic_block bb = gimple_bb (close_phi);
+ gimple_stmt_iterator insert_gsi = gsi_after_labels (bb);
+ gimple assign = gimple_build_assign (res, red);
+
+ gsi_insert_before (&insert_gsi, assign, GSI_SAME_STMT);
+}
+
+/* Insert the assignment "RED = initial_value (LOOP_PHI)". */
+
+static void
+insert_copyin (tree red, gimple loop_phi)
+{
+ gimple_seq stmts;
+ tree init = initial_value_for_loop_phi (loop_phi);
+ tree expr = build2 (MODIFY_EXPR, TREE_TYPE (init), red, init);
+
+ force_gimple_operand (expr, &stmts, true, NULL);
+ gsi_insert_seq_on_edge (edge_initial_value_for_loop_phi (loop_phi), stmts);
+}
+
+/* Removes the PHI node and resets all the debug stmts that are using
+ the PHI_RESULT. */
+
+static void
+remove_phi (gimple phi)
+{
+ imm_use_iterator imm_iter;
+ tree def;
+ use_operand_p use_p;
+ gimple_stmt_iterator gsi;
+ VEC (gimple, heap) *update = VEC_alloc (gimple, heap, 3);
+ unsigned int i;
+ gimple stmt;
+
+ def = PHI_RESULT (phi);
+ FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
+ {
+ stmt = USE_STMT (use_p);
+
+ if (is_gimple_debug (stmt))
+ {
+ gimple_debug_bind_reset_value (stmt);
+ VEC_safe_push (gimple, heap, update, stmt);
+ }
+ }
+
+ for (i = 0; VEC_iterate (gimple, update, i, stmt); i++)
+ update_stmt (stmt);
+
+ VEC_free (gimple, heap, update);
+
+ gsi = gsi_for_phi_node (phi);
+ remove_phi_node (&gsi, false);
+}
+
+/* Rewrite out of SSA the reduction described by the loop phi nodes
+ IN, and the close phi nodes OUT. IN and OUT are structured by loop
+ levels like this:
+
+ IN: stmt, loop_n, ..., loop_0
+ OUT: stmt, close_n, ..., close_0
+
+ the first element is the reduction statement, and the next elements
+ are the loop and close phi nodes of each of the outer loops. */
+
+static void
+translate_scalar_reduction_to_array (VEC (gimple, heap) *in,
+ VEC (gimple, heap) *out,
+ sbitmap reductions)
+{
+ unsigned int i;
+ gimple loop_phi;
+ tree red = NULL_TREE;
+
+ for (i = 0; VEC_iterate (gimple, in, i, loop_phi); i++)
+ {
+ gimple close_phi = VEC_index (gimple, out, i);
+
+ if (i == 0)
+ {
+ gimple stmt = loop_phi;
+ basic_block bb = split_reduction_stmt (stmt);
+
+ SET_BIT (reductions, bb->index);
+ gcc_assert (close_phi == loop_phi);
+
+ red = create_zero_dim_array
+ (gimple_assign_lhs (stmt), "Commutative_Associative_Reduction");
+ translate_scalar_reduction_to_array_for_stmt
+ (red, stmt, VEC_index (gimple, in, 1));
+ continue;
+ }
+
+ if (i == VEC_length (gimple, in) - 1)
+ {
+ insert_copyout (red, close_phi);
+ insert_copyin (red, loop_phi);
+ }
+
+ remove_phi (loop_phi);
+ remove_phi (close_phi);
+ }
+}
+
+/* Rewrites out of SSA a commutative reduction at CLOSE_PHI. */
+
+static void
+rewrite_commutative_reductions_out_of_ssa_close_phi (gimple close_phi,
+ sbitmap reductions)
+{
+ VEC (gimple, heap) *in = VEC_alloc (gimple, heap, 10);
+ VEC (gimple, heap) *out = VEC_alloc (gimple, heap, 10);
+
+ detect_commutative_reduction (close_phi, &in, &out);
+ if (VEC_length (gimple, in) > 0)
+ translate_scalar_reduction_to_array (in, out, reductions);
+
+ VEC_free (gimple, heap, in);
+ VEC_free (gimple, heap, out);
+}
+
+/* Rewrites all the commutative reductions from LOOP out of SSA. */
+
+static void
+rewrite_commutative_reductions_out_of_ssa_loop (loop_p loop,
+ sbitmap reductions)
+{
+ gimple_stmt_iterator gsi;
+ edge exit = single_exit (loop);
+
+ if (!exit)
+ return;
+
+ for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi))
+ rewrite_commutative_reductions_out_of_ssa_close_phi (gsi_stmt (gsi),
+ reductions);
+}
+
+/* Rewrites all the commutative reductions from SCOP out of SSA. */
+
+static void
+rewrite_commutative_reductions_out_of_ssa (sese region, sbitmap reductions)
+{
+ loop_iterator li;
+ loop_p loop;
+
+ FOR_EACH_LOOP (li, loop, 0)
+ if (loop_in_sese_p (loop, region))
+ rewrite_commutative_reductions_out_of_ssa_loop (loop, reductions);
+
+ gsi_commit_edge_inserts ();
+ update_ssa (TODO_update_ssa);
+#ifdef ENABLE_CHECKING
+ verify_loop_closed_ssa (true);
+#endif
+}
+
+/* A LOOP is in normal form for Graphite when it contains only one
+ scalar phi node that defines the main induction variable of the
+ loop, only one increment of the IV, and only one exit condition. */
+
+static void
+graphite_loop_normal_form (loop_p loop)
+{
+ struct tree_niter_desc niter;
+ tree nit;
+ gimple_seq stmts;
+ edge exit = single_dom_exit (loop);
+
+ bool known_niter = number_of_iterations_exit (loop, exit, &niter, false);
+
+ /* At this point we should know the number of iterations. */
+ gcc_assert (known_niter);
+
+ nit = force_gimple_operand (unshare_expr (niter.niter), &stmts, true,
+ NULL_TREE);
+ if (stmts)
+ gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
+
+ loop->single_iv = canonicalize_loop_ivs (loop, &nit, false);
+}
+
+/* Rewrite all the loops of SCOP in normal form: one induction
+ variable per loop. */
+
+static void
+scop_canonicalize_loops (scop_p scop)
+{
+ loop_iterator li;
+ loop_p loop;
+
+ FOR_EACH_LOOP (li, loop, 0)
+ if (loop_in_sese_p (loop, SCOP_REGION (scop)))
+ graphite_loop_normal_form (loop);
+}
+
+/* Java does not initialize long_long_integer_type_node. */
+#define my_long_long (long_long_integer_type_node ? long_long_integer_type_node : ssizetype)
+
+/* Can all ivs be represented by a signed integer?
+ As CLooG might generate negative values in its expressions, signed loop ivs
+ are required in the backend. */
+static bool
+scop_ivs_can_be_represented (scop_p scop)
+{
+ loop_iterator li;
+ loop_p loop;
+
+ FOR_EACH_LOOP (li, loop, 0)
+ {
+ tree type;
+ int precision;
+
+ if (!loop_in_sese_p (loop, SCOP_REGION (scop)))
+ continue;
+
+ if (!loop->single_iv)
+ continue;
+
+ type = TREE_TYPE(loop->single_iv);
+ precision = TYPE_PRECISION (type);
+
+ if (TYPE_UNSIGNED (type)
+ && precision >= TYPE_PRECISION (my_long_long))
+ return false;
+ }
+
+ return true;
+}
+
+#undef my_long_long
+
/* Builds the polyhedral representation for a SESE region. */
-bool
+void
build_poly_scop (scop_p scop)
{
sese region = SCOP_REGION (scop);
+ sbitmap reductions = sbitmap_alloc (last_basic_block * 2);
+ graphite_dim_t max_dim;
+
+ sbitmap_zero (reductions);
+ rewrite_commutative_reductions_out_of_ssa (region, reductions);
rewrite_reductions_out_of_ssa (scop);
- build_scop_bbs (scop);
+ build_scop_bbs (scop, reductions);
+ sbitmap_free (reductions);
/* FIXME: This restriction is needed to avoid a problem in CLooG.
Once CLooG is fixed, remove this guard. Anyways, it makes no
sense to optimize a scop containing only PBBs that do not belong
to any loops. */
if (nb_pbbs_in_loops (scop) == 0)
- return false;
+ return;
+
+ scop_canonicalize_loops (scop);
+ if (!scop_ivs_can_be_represented (scop))
+ return;
build_sese_loop_nests (region);
build_sese_conditions (region);
find_scop_parameters (scop);
+ max_dim = PARAM_VALUE (PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS);
+ if (scop_nb_params (scop) > max_dim)
+ return;
+
build_scop_iteration_domain (scop);
build_scop_context (scop);
add_conditions_to_constraints (scop);
+ scop_to_lst (scop);
build_scop_scattering (scop);
build_scop_drs (scop);
- return true;
+ /* This SCoP has been translated to the polyhedral
+ representation. */
+ POLY_SCOP_P (scop) = true;
}
/* Always return false. Exercise the scop_to_clast function. */
void
-check_poly_representation (scop_p scop)
+check_poly_representation (scop_p scop ATTRIBUTE_UNUSED)
{
#ifdef ENABLE_CHECKING
cloog_prog_clast pc = scop_to_clast (scop);