#include "graphite-clast-to-gimple.h"
#include "graphite-dependences.h"
+/* This flag is set when an error occurred during the translation of
+ CLAST to Gimple. */
+static bool gloog_error;
+
/* Verifies properties that GRAPHITE should maintain during translation. */
static inline void
slot = htab_find_slot (index_table, &tmp, INSERT);
if (slot)
- *slot = new_clast_name_index (name, index);
+ {
+ if (*slot)
+ free (*slot);
+
+ *slot = new_clast_name_index (name, index);
+ }
}
/* Print to stderr the element ELT. */
{
tree name = clast_name_to_gcc (t->var, region, newivs,
newivs_index, params_index);
- return fold_convert (type, name);
+
+ if (POINTER_TYPE_P (TREE_TYPE (name)) != POINTER_TYPE_P (type))
+ name = fold_convert (sizetype, name);
+
+ name = fold_convert (type, name);
+ return name;
}
else if (value_mone_p (t->val))
{
tree name = clast_name_to_gcc (t->var, region, newivs,
newivs_index, params_index);
+
+ if (POINTER_TYPE_P (TREE_TYPE (name)) != POINTER_TYPE_P (type))
+ name = fold_convert (sizetype, name);
+
name = fold_convert (type, name);
+
return fold_build1 (NEGATE_EXPR, type, name);
}
else
tree name = clast_name_to_gcc (t->var, region, newivs,
newivs_index, params_index);
tree cst = gmp_cst_to_tree (type, t->val);
+
+ if (POINTER_TYPE_P (TREE_TYPE (name)) != POINTER_TYPE_P (type))
+ name = fold_convert (sizetype, name);
+
name = fold_convert (type, name);
- return fold_build2 (MULT_EXPR, type, cst, name);
+
+ if (!POINTER_TYPE_P (type))
+ return fold_build2 (MULT_EXPR, type, cst, name);
+
+ gloog_error = true;
+ return cst;
}
}
else
gcc_unreachable ();
}
-/* Given a CLOOG_IV, returns the type that it should have in GCC land.
- If the information is not available, i.e. in the case one of the
- transforms created the loop, just return integer_type_node. */
+/* Java does not initialize long_long_integer_type_node. */
+#define my_long_long (long_long_integer_type_node ? long_long_integer_type_node : ssizetype)
+
+/* Given a CLOOG_IV, return the type that CLOOG_IV should have in GCC
+ land. The selected type is big enough to include the original loop
+ iteration variable, but signed to work with the subtractions CLooG
+ may have introduced. If such a type is not available, we fail.
+
+ TODO: Do not always return long_long, but the smallest possible
+ type, that still holds the original type.
+
+ TODO: Get the types using CLooG instead. This enables further
+ optimizations, but needs CLooG support. */
static tree
gcc_type_for_cloog_iv (const char *cloog_iv, gimple_bb_p gbb)
slot = htab_find_slot (GBB_CLOOG_IV_TYPES (gbb), &tmp, NO_INSERT);
if (slot && *slot)
- return ((ivtype_map_elt) *slot)->type;
+ {
+ tree type = ((ivtype_map_elt) *slot)->type;
+ int type_precision = TYPE_PRECISION (type);
+
+ /* Find the smallest signed type possible. */
+ if (!TYPE_UNSIGNED (type))
+ {
+ if (type_precision <= TYPE_PRECISION (integer_type_node))
+ return integer_type_node;
+
+ if (type_precision <= TYPE_PRECISION (long_integer_type_node))
+ return long_integer_type_node;
+
+ if (type_precision <= TYPE_PRECISION (my_long_long))
+ return my_long_long;
+
+ gcc_unreachable ();
+ }
+
+ if (type_precision < TYPE_PRECISION (integer_type_node))
+ return integer_type_node;
- return integer_type_node;
+ if (type_precision < TYPE_PRECISION (long_integer_type_node))
+ return long_integer_type_node;
+
+ if (type_precision < TYPE_PRECISION (my_long_long))
+ return my_long_long;
+
+ /* There is no signed type available, that is large enough to hold the
+ original value. */
+ gcc_unreachable ();
+ }
+
+ return my_long_long;
}
+#undef my_long_long
+
/* Returns the induction variable for the loop that gets translated to
STMT. */
tmp.old_name = old_name;
x = htab_find_slot (res, &tmp, INSERT);
- if (!*x)
+ if (x && !*x)
*x = new_rename_map_elt (old_name, expr);
return 1;
tmp.bb = bb;
x = htab_find_slot (bb_pbb_mapping, &tmp, INSERT);
- if (!*x)
+ if (x && !*x)
*x = new_bb_pbb_def (bb, pbb);
}
}
static edge
-translate_clast (sese, struct clast_stmt *, edge, htab_t, VEC (tree, heap) **,
- htab_t, htab_t, htab_t);
+translate_clast (sese, loop_p, struct clast_stmt *, edge, htab_t,
+ VEC (tree, heap) **, htab_t, htab_t, int, htab_t);
/* Translates a clast user statement STMT to gimple.
- REGION is the sese region we used to generate the scop.
- NEXT_E is the edge where new generated code should be attached.
+ - CONTEXT_LOOP is the loop in which the generated code will be placed
- RENAME_MAP contains a set of tuples of new names associated to
the original variables names.
- BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping.
htab_t newivs_index, htab_t bb_pbb_mapping,
htab_t params_index)
{
+ gimple_bb_p gbb;
+ basic_block new_bb;
poly_bb_p pbb = (poly_bb_p) cloog_statement_usr (stmt->statement);
- gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
+ gbb = PBB_BLACK_BOX (pbb);
if (GBB_BB (gbb) == ENTRY_BLOCK_PTR)
return next_e;
params_index);
next_e = copy_bb_and_scalar_dependences (GBB_BB (gbb), region,
next_e, rename_map);
- mark_bb_with_pbb (pbb, next_e->src, bb_pbb_mapping);
+ new_bb = next_e->src;
+ mark_bb_with_pbb (pbb, new_bb, bb_pbb_mapping);
update_ssa (TODO_update_ssa);
return next_e;
}
-/* Mark a loop parallel, if the graphite dependency check cannot find any
- * dependencies. This triggers parallel code generation in the autopar pass.
- * */
-static void
-try_mark_loop_parallel (sese region, loop_p loop, htab_t bb_pbb_mapping)
+/* Creates a new if region protecting the loop to be executed, if the execution
+ count is zero (lb > ub). */
+static edge
+graphite_create_new_loop_guard (sese region, edge entry_edge,
+ struct clast_for *stmt,
+ VEC (tree, heap) *newivs,
+ htab_t newivs_index, htab_t params_index)
{
- loop_p outermost_loop = SESE_ENTRY (region)->src->loop_father;
- int level = loop_depth (loop) - loop_depth (outermost_loop);
+ tree cond_expr;
+ edge exit_edge;
+ tree type = gcc_type_for_iv_of_clast_loop (stmt);
+ tree lb = clast_to_gcc_expression (type, stmt->LB, region, newivs,
+ newivs_index, params_index);
+ tree ub = clast_to_gcc_expression (type, stmt->UB, region, newivs,
+ newivs_index, params_index);
- if (flag_loop_parallelize_all
- && !dependency_in_loop_p (loop, bb_pbb_mapping,
- get_scattering_level (level)))
- loop->can_be_parallel = true;
+ /* XXX: Adding +1 and using LT_EXPR helps with loop latches that have a
+ loop iteration count of "PARAMETER - 1". For PARAMETER == 0 this becomes
+ 2^{32|64}, and the condition lb <= ub is true, even if we do not want this.
+ However lb < ub + 1 is false, as expected.
+ There might be a problem with cases where ub is 2^32. */
+ tree one;
+ Value gmp_one;
+ value_init (gmp_one);
+ value_set_si (gmp_one, 1);
+ one = gmp_cst_to_tree (type, gmp_one);
+ value_clear (gmp_one);
+
+ ub = fold_build2 (PLUS_EXPR, type, ub, one);
+ cond_expr = fold_build2 (LT_EXPR, boolean_type_node, lb, ub);
+
+ exit_edge = create_empty_if_region_on_edge (entry_edge, cond_expr);
+
+ return exit_edge;
}
-/* Translates a clast for statement STMT to gimple.
+/* Create the loop for a clast for statement.
- REGION is the sese region we used to generate the scop.
- NEXT_E is the edge where new generated code should be attached.
- PARAMS_INDEX connects the cloog parameters with the gimple parameters in
the sese region. */
static edge
-translate_clast_for (sese region, struct clast_for *stmt, edge next_e,
- htab_t rename_map, VEC (tree, heap) **newivs,
- htab_t newivs_index, htab_t bb_pbb_mapping,
- htab_t params_index)
+translate_clast_for_loop (sese region, loop_p context_loop,
+ struct clast_for *stmt, edge next_e,
+ htab_t rename_map, VEC (tree, heap) **newivs,
+ htab_t newivs_index, htab_t bb_pbb_mapping,
+ int level, htab_t params_index)
{
- loop_p context_loop = next_e->dest->loop_father;
- loop_p loop = graphite_create_new_loop (region, next_e, stmt, context_loop,
- newivs, newivs_index, params_index);
+ struct loop *loop = graphite_create_new_loop (region, next_e, stmt,
+ context_loop, newivs,
+ newivs_index, params_index);
edge last_e = single_exit (loop);
- edge body = single_succ_edge (loop->header);
-
- next_e = translate_clast (region, stmt->body, body, rename_map, newivs,
- newivs_index, bb_pbb_mapping, params_index);
+ edge to_body = single_succ_edge (loop->header);
+ basic_block after = to_body->dest;
/* Create a basic block for loop close phi nodes. */
last_e = single_succ_edge (split_edge (last_e));
+
+ /* Translate the body of the loop. */
+ next_e = translate_clast (region, loop, stmt->body, to_body, rename_map,
+ newivs, newivs_index, bb_pbb_mapping, level + 1,
+ params_index);
+ redirect_edge_succ_nodup (next_e, after);
+ set_immediate_dominator (CDI_DOMINATORS, next_e->dest, next_e->src);
+
+ /* Remove from rename_map all the tuples containing variables
+ defined in loop's body. */
insert_loop_close_phis (rename_map, loop);
- try_mark_loop_parallel (region, loop, bb_pbb_mapping);
+ if (flag_loop_parallelize_all
+ && !dependency_in_loop_p (loop, bb_pbb_mapping,
+ get_scattering_level (level)))
+ loop->can_be_parallel = true;
+
+ return last_e;
+}
+
+/* Translates a clast for statement STMT to gimple. First a guard is created
+ protecting the loop, if it is executed zero times. In this guard we create
+ the real loop structure.
+
+ - REGION is the sese region we used to generate the scop.
+ - NEXT_E is the edge where new generated code should be attached.
+ - RENAME_MAP contains a set of tuples of new names associated to
+ the original variables names.
+ - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping.
+ - PARAMS_INDEX connects the cloog parameters with the gimple parameters in
+ the sese region. */
+static edge
+translate_clast_for (sese region, loop_p context_loop, struct clast_for *stmt,
+ edge next_e, htab_t rename_map, VEC (tree, heap) **newivs,
+ htab_t newivs_index, htab_t bb_pbb_mapping, int level,
+ htab_t params_index)
+{
+ edge last_e = graphite_create_new_loop_guard (region, next_e, stmt, *newivs,
+ newivs_index, params_index);
+
+ edge true_e = get_true_edge_from_guard_bb (next_e->dest);
+ edge false_e = get_false_edge_from_guard_bb (next_e->dest);
+ edge exit_true_e = single_succ_edge (true_e->dest);
+ edge exit_false_e = single_succ_edge (false_e->dest);
+
+ htab_t before_guard = htab_create (10, rename_map_elt_info,
+ eq_rename_map_elts, free);
+ htab_traverse (rename_map, copy_renames, before_guard);
+
+ next_e = translate_clast_for_loop (region, context_loop, stmt, true_e,
+ rename_map, newivs,
+ newivs_index, bb_pbb_mapping, level,
+ params_index);
+
+ insert_guard_phis (last_e->src, exit_true_e, exit_false_e,
+ before_guard, rename_map);
+
+ htab_delete (before_guard);
return last_e;
}
- REGION is the sese region we used to generate the scop.
- NEXT_E is the edge where new generated code should be attached.
+ - CONTEXT_LOOP is the loop in which the generated code will be placed
- RENAME_MAP contains a set of tuples of new names associated to
the original variables names.
- BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping.
- PARAMS_INDEX connects the cloog parameters with the gimple parameters in
the sese region. */
static edge
-translate_clast_guard (sese region, struct clast_guard *stmt, edge next_e,
+translate_clast_guard (sese region, loop_p context_loop,
+ struct clast_guard *stmt, edge next_e,
htab_t rename_map, VEC (tree, heap) **newivs,
- htab_t newivs_index, htab_t bb_pbb_mapping,
+ htab_t newivs_index, htab_t bb_pbb_mapping, int level,
htab_t params_index)
{
edge last_e = graphite_create_new_guard (region, next_e, stmt, *newivs,
eq_rename_map_elts, free);
htab_traverse (rename_map, copy_renames, before_guard);
- next_e = translate_clast (region, stmt->then, true_e,
+ next_e = translate_clast (region, context_loop, stmt->then, true_e,
rename_map, newivs, newivs_index, bb_pbb_mapping,
- params_index);
+ level, params_index);
insert_guard_phis (last_e->src, exit_true_e, exit_false_e,
before_guard, rename_map);
context of a SESE.
- NEXT_E is the edge where new generated code should be attached.
+ - CONTEXT_LOOP is the loop in which the generated code will be placed
- RENAME_MAP contains a set of tuples of new names associated to
the original variables names.
- BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping. */
static edge
-translate_clast (sese region, struct clast_stmt *stmt,
+translate_clast (sese region, loop_p context_loop, struct clast_stmt *stmt,
edge next_e, htab_t rename_map, VEC (tree, heap) **newivs,
- htab_t newivs_index, htab_t bb_pbb_mapping,
+ htab_t newivs_index, htab_t bb_pbb_mapping, int level,
htab_t params_index)
{
if (!stmt)
bb_pbb_mapping, params_index);
else if (CLAST_STMT_IS_A (stmt, stmt_for))
- next_e = translate_clast_for (region,
- (struct clast_for *) stmt, next_e, rename_map,
- newivs, newivs_index, bb_pbb_mapping,
- params_index);
+ next_e = translate_clast_for (region, context_loop,
+ (struct clast_for *) stmt, next_e,
+ rename_map, newivs, newivs_index,
+ bb_pbb_mapping, level, params_index);
else if (CLAST_STMT_IS_A (stmt, stmt_guard))
- next_e = translate_clast_guard (region, (struct clast_guard *) stmt, next_e,
+ next_e = translate_clast_guard (region, context_loop,
+ (struct clast_guard *) stmt, next_e,
rename_map, newivs, newivs_index,
- bb_pbb_mapping, params_index);
+ bb_pbb_mapping, level, params_index);
else if (CLAST_STMT_IS_A (stmt, stmt_block))
- next_e = translate_clast (region, ((struct clast_block *) stmt)->body,
+ next_e = translate_clast (region, context_loop,
+ ((struct clast_block *) stmt)->body,
next_e, rename_map, newivs, newivs_index,
- bb_pbb_mapping, params_index);
+ bb_pbb_mapping, level, params_index);
else
gcc_unreachable();
recompute_all_dominators ();
graphite_verify ();
- return translate_clast (region, stmt->next, next_e, rename_map, newivs,
- newivs_index, bb_pbb_mapping, params_index);
+ return translate_clast (region, context_loop, stmt->next, next_e,
+ rename_map, newivs, newivs_index,
+ bb_pbb_mapping, level, params_index);
}
/* Returns the first cloog name used in EXPR. */
find_cloog_iv_in_expr (struct clast_expr *expr)
{
struct clast_term *term = (struct clast_term *) expr;
-
- if (expr->type == expr_term
- && !term->var)
- return NULL;
+ struct clast_reduction *red;
+ int i;
if (expr->type == expr_term)
return term->var;
- if (expr->type == expr_red)
- {
- int i;
- struct clast_reduction *red = (struct clast_reduction *) expr;
+ if (expr->type != expr_red)
+ return NULL;
- for (i = 0; i < red->n; i++)
- {
- const char *res = find_cloog_iv_in_expr ((red)->elts[i]);
+ red = (struct clast_reduction *) expr;
+ for (i = 0; i < red->n; i++)
+ {
+ const char *res = find_cloog_iv_in_expr (red->elts[i]);
- if (res)
- return res;
- }
+ if (res)
+ return res;
}
return NULL;
}
-/* Build for a clast_user_stmt USER_STMT a map between the CLAST
- induction variables and the corresponding GCC old induction
- variables. This information is stored on each GRAPHITE_BB. */
+/* Build for USER_STMT a map between the CLAST induction variables and
+ the corresponding GCC old induction variables. This information is
+ stored on each GRAPHITE_BB. */
static void
compute_cloog_iv_types_1 (poly_bb_p pbb, struct clast_user_stmt *user_stmt)
slot = htab_find_slot (GBB_CLOOG_IV_TYPES (gbb), &tmp, INSERT);
- if (!*slot)
+ if (slot && !*slot)
{
tree oldiv = pbb_to_depth_to_oldiv (pbb, index);
- tree type = oldiv ? TREE_TYPE (oldiv) : integer_type_node;
+ tree type = TREE_TYPE (oldiv);
*slot = new_ivtype_map_elt (tmp.cloog_iv, type);
}
}
print_generated_program (stderr, scop);
}
-/* Add CLooG names to parameter index. The index is used to translate back from
- * CLooG names to GCC trees. */
+/* Add CLooG names to parameter index. The index is used to translate
+ back from CLooG names to GCC trees. */
static void
create_params_index (htab_t index_table, CloogProgram *prog) {
*/
bool
-gloog (scop_p scop, htab_t bb_pbb_mapping)
+gloog (scop_p scop, VEC (scop_p, heap) *scops, htab_t bb_pbb_mapping)
{
- edge new_scop_exit_edge = NULL;
VEC (tree, heap) *newivs = VEC_alloc (tree, heap, 10);
+ loop_p context_loop;
sese region = SCOP_REGION (scop);
ifsese if_region = NULL;
htab_t rename_map, newivs_index, params_index;
cloog_prog_clast pc;
+ int i;
timevar_push (TV_GRAPHITE_CODE_GEN);
+ gloog_error = false;
pc = scop_to_clast (scop);
recompute_all_dominators ();
graphite_verify ();
+ context_loop = SESE_ENTRY (region)->src->loop_father;
compute_cloog_iv_types (pc.stmt);
rename_map = htab_create (10, rename_map_elt_info, eq_rename_map_elts, free);
newivs_index = htab_create (10, clast_name_index_elt_info,
create_params_index (params_index, pc.prog);
- new_scop_exit_edge = translate_clast (region, pc.stmt,
- if_region->true_region->entry,
- rename_map, &newivs, newivs_index,
- bb_pbb_mapping, params_index);
+ translate_clast (region, context_loop, pc.stmt,
+ if_region->true_region->entry,
+ rename_map, &newivs, newivs_index,
+ bb_pbb_mapping, 1, params_index);
graphite_verify ();
sese_adjust_liveout_phis (region, rename_map,
if_region->region->exit->src,
if_region->false_region->exit,
if_region->true_region->exit);
+ scev_reset_htab ();
+ rename_nb_iterations (rename_map);
+
+ for (i = 0; VEC_iterate (scop_p, scops, i, scop); i++)
+ rename_sese_parameters (rename_map, SCOP_REGION (scop));
+
recompute_all_dominators ();
graphite_verify ();
+ if (gloog_error)
+ set_ifsese_condition (if_region, integer_zero_node);
+
free (if_region->true_region);
free (if_region->region);
free (if_region);
num_no_dependency);
}
- return true;
+ return !gloog_error;
}
#endif