1 /* Translation of CLAST (CLooG AST) to Gimple.
2 Copyright (C) 2009, 2010 Free Software Foundation, Inc.
3 Contributed by Sebastian Pop <sebastian.pop@amd.com>.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 3, or (at your option)
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
23 #include "coretypes.h"
28 #include "basic-block.h"
29 #include "diagnostic.h"
30 #include "tree-flow.h"
32 #include "tree-dump.h"
35 #include "tree-chrec.h"
36 #include "tree-data-ref.h"
37 #include "tree-scalar-evolution.h"
38 #include "tree-pass.h"
40 #include "value-prof.h"
41 #include "pointer-set.h"
43 #include "langhooks.h"
47 #include "cloog/cloog.h"
49 #include "graphite-ppl.h"
51 #include "graphite-poly.h"
52 #include "graphite-scop-detection.h"
53 #include "graphite-clast-to-gimple.h"
54 #include "graphite-dependences.h"
56 /* This flag is set when an error occurred during the translation of
58 static bool gloog_error;
60 /* Verifies properties that GRAPHITE should maintain during translation. */
63 graphite_verify (void)
65 #ifdef ENABLE_CHECKING
66 verify_loop_structure ();
67 verify_dominators (CDI_DOMINATORS);
68 verify_dominators (CDI_POST_DOMINATORS);
69 verify_loop_closed_ssa (true);
73 /* Stores the INDEX in a vector for a given clast NAME. */
75 typedef struct clast_name_index {
78 } *clast_name_index_p;
80 /* Returns a pointer to a new element of type clast_name_index_p built
81 from NAME and INDEX. */
83 static inline clast_name_index_p
84 new_clast_name_index (const char *name, int index)
86 clast_name_index_p res = XNEW (struct clast_name_index);
93 /* For a given clast NAME, returns -1 if it does not correspond to any
94 parameter, or otherwise, returns the index in the PARAMS or
95 SCATTERING_DIMENSIONS vector. */
98 clast_name_to_index (const char *name, htab_t index_table)
100 struct clast_name_index tmp;
104 slot = htab_find_slot (index_table, &tmp, NO_INSERT);
107 return ((struct clast_name_index *) *slot)->index;
112 /* Records in INDEX_TABLE the INDEX for NAME. */
115 save_clast_name_index (htab_t index_table, const char *name, int index)
117 struct clast_name_index tmp;
121 slot = htab_find_slot (index_table, &tmp, INSERT);
128 *slot = new_clast_name_index (name, index);
132 /* Print to stderr the element ELT. */
135 debug_clast_name_index (clast_name_index_p elt)
137 fprintf (stderr, "(index = %d, name = %s)\n", elt->index, elt->name);
140 /* Helper function for debug_rename_map. */
143 debug_clast_name_indexes_1 (void **slot, void *s ATTRIBUTE_UNUSED)
145 struct clast_name_index *entry = (struct clast_name_index *) *slot;
146 debug_clast_name_index (entry);
150 /* Print to stderr all the elements of MAP. */
153 debug_clast_name_indexes (htab_t map)
155 htab_traverse (map, debug_clast_name_indexes_1, NULL);
158 /* Computes a hash function for database element ELT. */
160 static inline hashval_t
161 clast_name_index_elt_info (const void *elt)
163 return htab_hash_pointer (((const struct clast_name_index *) elt)->name);
166 /* Compares database elements E1 and E2. */
169 eq_clast_name_indexes (const void *e1, const void *e2)
171 const struct clast_name_index *elt1 = (const struct clast_name_index *) e1;
172 const struct clast_name_index *elt2 = (const struct clast_name_index *) e2;
174 return (elt1->name == elt2->name);
178 /* For a given loop DEPTH in the loop nest of the original black box
179 PBB, return the old induction variable associated to that loop. */
182 pbb_to_depth_to_oldiv (poly_bb_p pbb, int depth)
184 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
185 sese region = SCOP_REGION (PBB_SCOP (pbb));
186 loop_p loop = gbb_loop_at_index (gbb, region, depth);
188 return loop->single_iv;
191 /* For a given scattering dimension, return the new induction variable
195 newivs_to_depth_to_newiv (VEC (tree, heap) *newivs, int depth)
197 return VEC_index (tree, newivs, depth);
202 /* Returns the tree variable from the name NAME that was given in
203 Cloog representation. */
206 clast_name_to_gcc (const char *name, sese region, VEC (tree, heap) *newivs,
207 htab_t newivs_index, htab_t params_index)
210 VEC (tree, heap) *params = SESE_PARAMS (region);
212 if (params && params_index)
214 index = clast_name_to_index (name, params_index);
217 return VEC_index (tree, params, index);
220 gcc_assert (newivs && newivs_index);
221 index = clast_name_to_index (name, newivs_index);
222 gcc_assert (index >= 0);
224 return newivs_to_depth_to_newiv (newivs, index);
227 /* Returns the signed maximal precision type for expressions TYPE1 and TYPE2. */
230 max_signed_precision_type (tree type1, tree type2)
232 int p1 = TYPE_PRECISION (type1);
233 int p2 = TYPE_PRECISION (type2);
234 int precision = p1 > p2 ? p1 : p2;
235 tree type = lang_hooks.types.type_for_size (precision, false);
240 return integer_type_node;
245 /* Returns the maximal precision type for expressions TYPE1 and TYPE2. */
248 max_precision_type (tree type1, tree type2)
251 if (POINTER_TYPE_P (type1))
254 if (POINTER_TYPE_P (type2))
257 if (!TYPE_UNSIGNED (type1)
258 || !TYPE_UNSIGNED (type2))
259 return max_signed_precision_type (type1, type2);
261 return TYPE_PRECISION (type1) > TYPE_PRECISION (type2) ? type1 : type2;
265 clast_to_gcc_expression (tree, struct clast_expr *, sese, VEC (tree, heap) *,
268 /* Converts a Cloog reduction expression R with reduction operation OP
269 to a GCC expression tree of type TYPE. */
272 clast_to_gcc_expression_red (tree type, enum tree_code op,
273 struct clast_reduction *r,
274 sese region, VEC (tree, heap) *newivs,
275 htab_t newivs_index, htab_t params_index)
278 tree res = clast_to_gcc_expression (type, r->elts[0], region, newivs,
279 newivs_index, params_index);
280 tree operand_type = (op == POINTER_PLUS_EXPR) ? sizetype : type;
282 for (i = 1; i < r->n; i++)
284 tree t = clast_to_gcc_expression (operand_type, r->elts[i], region,
285 newivs, newivs_index, params_index);
286 res = fold_build2 (op, type, res, t);
292 /* Converts a Cloog AST expression E back to a GCC expression tree of
296 clast_to_gcc_expression (tree type, struct clast_expr *e,
297 sese region, VEC (tree, heap) *newivs,
298 htab_t newivs_index, htab_t params_index)
304 struct clast_term *t = (struct clast_term *) e;
308 if (value_one_p (t->val))
310 tree name = clast_name_to_gcc (t->var, region, newivs,
311 newivs_index, params_index);
313 if (POINTER_TYPE_P (TREE_TYPE (name)) != POINTER_TYPE_P (type))
314 name = fold_convert (sizetype, name);
316 name = fold_convert (type, name);
320 else if (value_mone_p (t->val))
322 tree name = clast_name_to_gcc (t->var, region, newivs,
323 newivs_index, params_index);
325 if (POINTER_TYPE_P (TREE_TYPE (name)) != POINTER_TYPE_P (type))
326 name = fold_convert (sizetype, name);
328 name = fold_convert (type, name);
330 return fold_build1 (NEGATE_EXPR, type, name);
334 tree name = clast_name_to_gcc (t->var, region, newivs,
335 newivs_index, params_index);
336 tree cst = gmp_cst_to_tree (type, t->val);
338 if (POINTER_TYPE_P (TREE_TYPE (name)) != POINTER_TYPE_P (type))
339 name = fold_convert (sizetype, name);
341 name = fold_convert (type, name);
343 if (!POINTER_TYPE_P (type))
344 return fold_build2 (MULT_EXPR, type, cst, name);
351 return gmp_cst_to_tree (type, t->val);
356 struct clast_reduction *r = (struct clast_reduction *) e;
361 return clast_to_gcc_expression_red
362 (type, POINTER_TYPE_P (type) ? POINTER_PLUS_EXPR : PLUS_EXPR,
363 r, region, newivs, newivs_index, params_index);
366 return clast_to_gcc_expression_red (type, MIN_EXPR, r, region,
367 newivs, newivs_index,
371 return clast_to_gcc_expression_red (type, MAX_EXPR, r, region,
372 newivs, newivs_index,
383 struct clast_binary *b = (struct clast_binary *) e;
384 struct clast_expr *lhs = (struct clast_expr *) b->LHS;
385 tree tl = clast_to_gcc_expression (type, lhs, region, newivs,
386 newivs_index, params_index);
387 tree tr = gmp_cst_to_tree (type, b->RHS);
392 return fold_build2 (FLOOR_DIV_EXPR, type, tl, tr);
395 return fold_build2 (CEIL_DIV_EXPR, type, tl, tr);
398 return fold_build2 (EXACT_DIV_EXPR, type, tl, tr);
401 return fold_build2 (TRUNC_MOD_EXPR, type, tl, tr);
415 /* Return the precision needed to represent the value VAL. */
418 precision_for_value (Value val)
427 value_assign (y, val);
428 value_set_si (two, 2);
434 while (value_gt (y, x))
436 value_multiply (x, x, two);
447 /* Return the precision needed to represent the values between LOW and
451 precision_for_interval (Value low, Value up)
456 gcc_assert (value_le (low, up));
459 value_subtract (diff, up, low);
460 precision = precision_for_value (diff);
466 /* Return a type that could represent the integer value VAL, or
467 otherwise return NULL_TREE. */
470 gcc_type_for_interval (Value low, Value up, tree old_type)
472 bool unsigned_p = true;
473 int precision, prec_up, prec_int;
476 gcc_assert (value_le (low, up));
478 /* Preserve the signedness of the old IV. */
479 if ((old_type && !TYPE_UNSIGNED (old_type))
480 || value_neg_p (low))
483 prec_up = precision_for_value (up);
484 prec_int = precision_for_interval (low, up);
485 precision = prec_up > prec_int ? prec_up : prec_int;
487 type = lang_hooks.types.type_for_size (precision, unsigned_p);
491 return integer_type_node;
497 /* Return a type that could represent the integer value VAL, or
498 otherwise return NULL_TREE. */
501 gcc_type_for_value (Value val)
503 return gcc_type_for_interval (val, val, NULL_TREE);
506 /* Return the type for the clast_term T used in STMT. */
509 gcc_type_for_clast_term (struct clast_term *t,
510 sese region, VEC (tree, heap) *newivs,
511 htab_t newivs_index, htab_t params_index)
513 gcc_assert (t->expr.type == expr_term);
516 return gcc_type_for_value (t->val);
518 return TREE_TYPE (clast_name_to_gcc (t->var, region, newivs,
519 newivs_index, params_index));
523 gcc_type_for_clast_expr (struct clast_expr *, sese,
524 VEC (tree, heap) *, htab_t, htab_t);
526 /* Return the type for the clast_reduction R used in STMT. */
529 gcc_type_for_clast_red (struct clast_reduction *r, sese region,
530 VEC (tree, heap) *newivs,
531 htab_t newivs_index, htab_t params_index)
534 tree type = NULL_TREE;
537 return gcc_type_for_clast_expr (r->elts[0], region, newivs,
538 newivs_index, params_index);
545 type = gcc_type_for_clast_expr (r->elts[0], region, newivs,
546 newivs_index, params_index);
547 for (i = 1; i < r->n; i++)
548 type = max_precision_type (type, gcc_type_for_clast_expr
549 (r->elts[i], region, newivs,
550 newivs_index, params_index));
562 /* Return the type for the clast_binary B used in STMT. */
565 gcc_type_for_clast_bin (struct clast_binary *b,
566 sese region, VEC (tree, heap) *newivs,
567 htab_t newivs_index, htab_t params_index)
569 tree l = gcc_type_for_clast_expr ((struct clast_expr *) b->LHS, region,
570 newivs, newivs_index, params_index);
571 tree r = gcc_type_for_value (b->RHS);
572 return max_signed_precision_type (l, r);
575 /* Returns the type for the CLAST expression E when used in statement
579 gcc_type_for_clast_expr (struct clast_expr *e,
580 sese region, VEC (tree, heap) *newivs,
581 htab_t newivs_index, htab_t params_index)
586 return gcc_type_for_clast_term ((struct clast_term *) e, region,
587 newivs, newivs_index, params_index);
590 return gcc_type_for_clast_red ((struct clast_reduction *) e, region,
591 newivs, newivs_index, params_index);
594 return gcc_type_for_clast_bin ((struct clast_binary *) e, region,
595 newivs, newivs_index, params_index);
604 /* Returns the type for the equation CLEQ. */
607 gcc_type_for_clast_eq (struct clast_equation *cleq,
608 sese region, VEC (tree, heap) *newivs,
609 htab_t newivs_index, htab_t params_index)
611 tree l = gcc_type_for_clast_expr (cleq->LHS, region, newivs,
612 newivs_index, params_index);
613 tree r = gcc_type_for_clast_expr (cleq->RHS, region, newivs,
614 newivs_index, params_index);
615 return max_precision_type (l, r);
618 /* Translates a clast equation CLEQ to a tree. */
621 graphite_translate_clast_equation (sese region,
622 struct clast_equation *cleq,
623 VEC (tree, heap) *newivs,
624 htab_t newivs_index, htab_t params_index)
627 tree type = gcc_type_for_clast_eq (cleq, region, newivs, newivs_index,
629 tree lhs = clast_to_gcc_expression (type, cleq->LHS, region, newivs,
630 newivs_index, params_index);
631 tree rhs = clast_to_gcc_expression (type, cleq->RHS, region, newivs,
632 newivs_index, params_index);
637 else if (cleq->sign > 0)
643 return fold_build2 (comp, boolean_type_node, lhs, rhs);
646 /* Creates the test for the condition in STMT. */
649 graphite_create_guard_cond_expr (sese region, struct clast_guard *stmt,
650 VEC (tree, heap) *newivs,
651 htab_t newivs_index, htab_t params_index)
656 for (i = 0; i < stmt->n; i++)
658 tree eq = graphite_translate_clast_equation (region, &stmt->eq[i],
659 newivs, newivs_index,
663 cond = fold_build2 (TRUTH_AND_EXPR, TREE_TYPE (eq), cond, eq);
671 /* Creates a new if region corresponding to Cloog's guard. */
674 graphite_create_new_guard (sese region, edge entry_edge,
675 struct clast_guard *stmt,
676 VEC (tree, heap) *newivs,
677 htab_t newivs_index, htab_t params_index)
679 tree cond_expr = graphite_create_guard_cond_expr (region, stmt, newivs,
680 newivs_index, params_index);
681 edge exit_edge = create_empty_if_region_on_edge (entry_edge, cond_expr);
685 /* Compute the lower bound LOW and upper bound UP for the induction
686 variable at LEVEL for the statement PBB, based on the transformed
687 scattering of PBB: T|I|G|Cst, with T the scattering transform, I
688 the iteration domain, and G the context parameters. */
691 compute_bounds_for_level (poly_bb_p pbb, int level, Value low, Value up)
693 ppl_Pointset_Powerset_C_Polyhedron_t ps;
694 ppl_Linear_Expression_t le;
696 combine_context_id_scat (&ps, pbb, false);
698 /* Prepare the linear expression corresponding to the level that we
699 want to maximize/minimize. */
701 ppl_dimension_type dim = pbb_nb_scattering_transform (pbb)
702 + pbb_dim_iter_domain (pbb) + pbb_nb_params (pbb);
704 ppl_new_Linear_Expression_with_dimension (&le, dim);
705 ppl_set_coef (le, 2 * level + 1, 1);
708 ppl_max_for_le_pointset (ps, le, up);
709 ppl_min_for_le_pointset (ps, le, low);
712 /* Compute the type for the induction variable at LEVEL for the
713 statement PBB, based on the transformed schedule of PBB. OLD_TYPE
714 is the type of the old induction variable for that loop. */
717 compute_type_for_level_1 (poly_bb_p pbb, int level, tree old_type)
725 compute_bounds_for_level (pbb, level, low, up);
726 type = gcc_type_for_interval (low, up, old_type);
733 /* Compute the type for the induction variable at LEVEL for the
734 statement PBB, based on the transformed schedule of PBB. */
737 compute_type_for_level (poly_bb_p pbb, int level)
739 tree oldiv = pbb_to_depth_to_oldiv (pbb, level);
740 tree type = TREE_TYPE (oldiv);
742 if (type && POINTER_TYPE_P (type))
744 #ifdef ENABLE_CHECKING
745 tree ctype = compute_type_for_level_1 (pbb, level, type);
747 /* In the case of a pointer type, check that after the loop
748 transform, the lower and the upper bounds of the type fit the
749 oldiv pointer type. */
750 gcc_assert (TYPE_PRECISION (type) >= TYPE_PRECISION (ctype)
751 && integer_zerop (lower_bound_in_type (ctype, ctype)));
756 return compute_type_for_level_1 (pbb, level, type);
759 /* Walks a CLAST and returns the first statement in the body of a
762 static struct clast_user_stmt *
763 clast_get_body_of_loop (struct clast_stmt *stmt)
766 || CLAST_STMT_IS_A (stmt, stmt_user))
767 return (struct clast_user_stmt *) stmt;
769 if (CLAST_STMT_IS_A (stmt, stmt_for))
770 return clast_get_body_of_loop (((struct clast_for *) stmt)->body);
772 if (CLAST_STMT_IS_A (stmt, stmt_guard))
773 return clast_get_body_of_loop (((struct clast_guard *) stmt)->then);
775 if (CLAST_STMT_IS_A (stmt, stmt_block))
776 return clast_get_body_of_loop (((struct clast_block *) stmt)->body);
781 /* Returns the type for the induction variable for the loop translated
785 gcc_type_for_iv_of_clast_loop (struct clast_for *stmt_for, int level,
786 tree lb_type, tree ub_type)
788 struct clast_stmt *stmt = (struct clast_stmt *) stmt_for;
789 struct clast_user_stmt *body = clast_get_body_of_loop (stmt);
790 CloogStatement *cs = body->statement;
791 poly_bb_p pbb = (poly_bb_p) cloog_statement_usr (cs);
793 return max_precision_type (lb_type, max_precision_type
794 (ub_type, compute_type_for_level (pbb,
798 /* Creates a new LOOP corresponding to Cloog's STMT. Inserts an
799 induction variable for the new LOOP. New LOOP is attached to CFG
800 starting at ENTRY_EDGE. LOOP is inserted into the loop tree and
801 becomes the child loop of the OUTER_LOOP. NEWIVS_INDEX binds
802 CLooG's scattering name to the induction variable created for the
803 loop of STMT. The new induction variable is inserted in the NEWIVS
807 graphite_create_new_loop (sese region, edge entry_edge,
808 struct clast_for *stmt,
809 loop_p outer, VEC (tree, heap) **newivs,
810 htab_t newivs_index, htab_t params_index, int level)
812 tree lb_type = gcc_type_for_clast_expr (stmt->LB, region, *newivs,
813 newivs_index, params_index);
814 tree ub_type = gcc_type_for_clast_expr (stmt->UB, region, *newivs,
815 newivs_index, params_index);
816 tree type = gcc_type_for_iv_of_clast_loop (stmt, level, lb_type, ub_type);
817 tree lb = clast_to_gcc_expression (type, stmt->LB, region, *newivs,
818 newivs_index, params_index);
819 tree ub = clast_to_gcc_expression (type, stmt->UB, region, *newivs,
820 newivs_index, params_index);
821 tree stride = gmp_cst_to_tree (type, stmt->stride);
822 tree ivvar = create_tmp_var (type, "graphite_IV");
823 tree iv, iv_after_increment;
824 loop_p loop = create_empty_loop_on_edge
825 (entry_edge, lb, stride, ub, ivvar, &iv, &iv_after_increment,
826 outer ? outer : entry_edge->src->loop_father);
828 add_referenced_var (ivvar);
830 save_clast_name_index (newivs_index, stmt->iterator,
831 VEC_length (tree, *newivs));
832 VEC_safe_push (tree, heap, *newivs, iv);
836 /* Inserts in MAP a tuple (OLD_NAME, NEW_NAME) for the induction
837 variables of the loops around GBB in SESE. */
840 build_iv_mapping (htab_t map, sese region,
841 VEC (tree, heap) *newivs, htab_t newivs_index,
842 struct clast_user_stmt *user_stmt,
845 struct clast_stmt *t;
847 CloogStatement *cs = user_stmt->statement;
848 poly_bb_p pbb = (poly_bb_p) cloog_statement_usr (cs);
850 for (t = user_stmt->substitutions; t; t = t->next, index++)
852 struct clast_expr *expr = (struct clast_expr *)
853 ((struct clast_assignment *)t)->RHS;
854 tree type = gcc_type_for_clast_expr (expr, region, newivs,
855 newivs_index, params_index);
856 tree old_name = pbb_to_depth_to_oldiv (pbb, index);
857 tree e = clast_to_gcc_expression (type, expr, region, newivs,
858 newivs_index, params_index);
859 set_rename (map, old_name, e);
863 /* Helper function for htab_traverse. */
866 copy_renames (void **slot, void *s)
868 struct rename_map_elt_s *entry = (struct rename_map_elt_s *) *slot;
869 htab_t res = (htab_t) s;
870 tree old_name = entry->old_name;
871 tree expr = entry->expr;
872 struct rename_map_elt_s tmp;
875 tmp.old_name = old_name;
876 x = htab_find_slot (res, &tmp, INSERT);
879 *x = new_rename_map_elt (old_name, expr);
884 /* Construct bb_pbb_def with BB and PBB. */
887 new_bb_pbb_def (basic_block bb, poly_bb_p pbb)
889 bb_pbb_def *bb_pbb_p;
891 bb_pbb_p = XNEW (bb_pbb_def);
898 /* Mark BB with it's relevant PBB via hashing table BB_PBB_MAPPING. */
901 mark_bb_with_pbb (poly_bb_p pbb, basic_block bb, htab_t bb_pbb_mapping)
907 x = htab_find_slot (bb_pbb_mapping, &tmp, INSERT);
910 *x = new_bb_pbb_def (bb, pbb);
913 /* Find BB's related poly_bb_p in hash table BB_PBB_MAPPING. */
916 find_pbb_via_hash (htab_t bb_pbb_mapping, basic_block bb)
922 slot = htab_find_slot (bb_pbb_mapping, &tmp, NO_INSERT);
925 return ((bb_pbb_def *) *slot)->pbb;
930 /* Check data dependency in LOOP at scattering level LEVEL.
931 BB_PBB_MAPPING is a basic_block and it's related poly_bb_p
935 dependency_in_loop_p (loop_p loop, htab_t bb_pbb_mapping, int level)
938 basic_block *bbs = get_loop_body_in_dom_order (loop);
940 for (i = 0; i < loop->num_nodes; i++)
942 poly_bb_p pbb1 = find_pbb_via_hash (bb_pbb_mapping, bbs[i]);
947 for (j = 0; j < loop->num_nodes; j++)
949 poly_bb_p pbb2 = find_pbb_via_hash (bb_pbb_mapping, bbs[j]);
954 if (dependency_between_pbbs_p (pbb1, pbb2, level))
968 translate_clast (sese, loop_p, struct clast_stmt *, edge, htab_t,
969 VEC (tree, heap) **, htab_t, htab_t, int, htab_t);
971 /* Translates a clast user statement STMT to gimple.
973 - REGION is the sese region we used to generate the scop.
974 - NEXT_E is the edge where new generated code should be attached.
975 - CONTEXT_LOOP is the loop in which the generated code will be placed
976 - RENAME_MAP contains a set of tuples of new names associated to
977 the original variables names.
978 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping.
979 - PARAMS_INDEX connects the cloog parameters with the gimple parameters in
982 translate_clast_user (sese region, struct clast_user_stmt *stmt, edge next_e,
983 htab_t rename_map, VEC (tree, heap) **newivs,
984 htab_t newivs_index, htab_t bb_pbb_mapping,
989 poly_bb_p pbb = (poly_bb_p) cloog_statement_usr (stmt->statement);
990 gbb = PBB_BLACK_BOX (pbb);
992 if (GBB_BB (gbb) == ENTRY_BLOCK_PTR)
995 build_iv_mapping (rename_map, region, *newivs, newivs_index, stmt,
997 next_e = copy_bb_and_scalar_dependences (GBB_BB (gbb), region,
999 new_bb = next_e->src;
1000 mark_bb_with_pbb (pbb, new_bb, bb_pbb_mapping);
1001 update_ssa (TODO_update_ssa);
1006 /* Creates a new if region protecting the loop to be executed, if the execution
1007 count is zero (lb > ub). */
1009 graphite_create_new_loop_guard (sese region, edge entry_edge,
1010 struct clast_for *stmt,
1011 VEC (tree, heap) *newivs,
1012 htab_t newivs_index, htab_t params_index)
1016 tree lb_type = gcc_type_for_clast_expr (stmt->LB, region, newivs,
1017 newivs_index, params_index);
1018 tree ub_type = gcc_type_for_clast_expr (stmt->UB, region, newivs,
1019 newivs_index, params_index);
1020 tree type = max_precision_type (lb_type, ub_type);
1021 tree lb = clast_to_gcc_expression (type, stmt->LB, region, newivs,
1022 newivs_index, params_index);
1023 tree ub = clast_to_gcc_expression (type, stmt->UB, region, newivs,
1024 newivs_index, params_index);
1026 /* XXX: Adding +1 and using LT_EXPR helps with loop latches that have a
1027 loop iteration count of "PARAMETER - 1". For PARAMETER == 0 this becomes
1028 2^{32|64}, and the condition lb <= ub is true, even if we do not want this.
1029 However lb < ub + 1 is false, as expected.
1030 There might be a problem with cases where ub is 2^32. */
1033 value_init (gmp_one);
1034 value_set_si (gmp_one, 1);
1035 one = gmp_cst_to_tree (type, gmp_one);
1036 value_clear (gmp_one);
1038 ub = fold_build2 (POINTER_TYPE_P (type) ? POINTER_PLUS_EXPR : PLUS_EXPR,
1040 cond_expr = fold_build2 (LT_EXPR, boolean_type_node, lb, ub);
1042 exit_edge = create_empty_if_region_on_edge (entry_edge, cond_expr);
1048 /* Create the loop for a clast for statement.
1050 - REGION is the sese region we used to generate the scop.
1051 - NEXT_E is the edge where new generated code should be attached.
1052 - RENAME_MAP contains a set of tuples of new names associated to
1053 the original variables names.
1054 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping.
1055 - PARAMS_INDEX connects the cloog parameters with the gimple parameters in
1058 translate_clast_for_loop (sese region, loop_p context_loop,
1059 struct clast_for *stmt, edge next_e,
1060 htab_t rename_map, VEC (tree, heap) **newivs,
1061 htab_t newivs_index, htab_t bb_pbb_mapping,
1062 int level, htab_t params_index)
1064 struct loop *loop = graphite_create_new_loop (region, next_e, stmt,
1065 context_loop, newivs,
1066 newivs_index, params_index,
1068 edge last_e = single_exit (loop);
1069 edge to_body = single_succ_edge (loop->header);
1070 basic_block after = to_body->dest;
1072 /* Create a basic block for loop close phi nodes. */
1073 last_e = single_succ_edge (split_edge (last_e));
1075 /* Translate the body of the loop. */
1076 next_e = translate_clast (region, loop, stmt->body, to_body, rename_map,
1077 newivs, newivs_index, bb_pbb_mapping, level + 1,
1079 redirect_edge_succ_nodup (next_e, after);
1080 set_immediate_dominator (CDI_DOMINATORS, next_e->dest, next_e->src);
1082 /* Remove from rename_map all the tuples containing variables
1083 defined in loop's body. */
1084 insert_loop_close_phis (rename_map, loop);
1086 if (flag_loop_parallelize_all
1087 && !dependency_in_loop_p (loop, bb_pbb_mapping,
1088 get_scattering_level (level)))
1089 loop->can_be_parallel = true;
1094 /* Translates a clast for statement STMT to gimple. First a guard is created
1095 protecting the loop, if it is executed zero times. In this guard we create
1096 the real loop structure.
1098 - REGION is the sese region we used to generate the scop.
1099 - NEXT_E is the edge where new generated code should be attached.
1100 - RENAME_MAP contains a set of tuples of new names associated to
1101 the original variables names.
1102 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping.
1103 - PARAMS_INDEX connects the cloog parameters with the gimple parameters in
1106 translate_clast_for (sese region, loop_p context_loop, struct clast_for *stmt,
1107 edge next_e, htab_t rename_map, VEC (tree, heap) **newivs,
1108 htab_t newivs_index, htab_t bb_pbb_mapping, int level,
1109 htab_t params_index)
1111 edge last_e = graphite_create_new_loop_guard (region, next_e, stmt, *newivs,
1112 newivs_index, params_index);
1114 edge true_e = get_true_edge_from_guard_bb (next_e->dest);
1115 edge false_e = get_false_edge_from_guard_bb (next_e->dest);
1116 edge exit_true_e = single_succ_edge (true_e->dest);
1117 edge exit_false_e = single_succ_edge (false_e->dest);
1119 htab_t before_guard = htab_create (10, rename_map_elt_info,
1120 eq_rename_map_elts, free);
1121 htab_traverse (rename_map, copy_renames, before_guard);
1123 next_e = translate_clast_for_loop (region, context_loop, stmt, true_e,
1125 newivs_index, bb_pbb_mapping, level,
1128 insert_guard_phis (last_e->src, exit_true_e, exit_false_e,
1129 before_guard, rename_map);
1131 htab_delete (before_guard);
1136 /* Translates a clast guard statement STMT to gimple.
1138 - REGION is the sese region we used to generate the scop.
1139 - NEXT_E is the edge where new generated code should be attached.
1140 - CONTEXT_LOOP is the loop in which the generated code will be placed
1141 - RENAME_MAP contains a set of tuples of new names associated to
1142 the original variables names.
1143 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping.
1144 - PARAMS_INDEX connects the cloog parameters with the gimple parameters in
1147 translate_clast_guard (sese region, loop_p context_loop,
1148 struct clast_guard *stmt, edge next_e,
1149 htab_t rename_map, VEC (tree, heap) **newivs,
1150 htab_t newivs_index, htab_t bb_pbb_mapping, int level,
1151 htab_t params_index)
1153 edge last_e = graphite_create_new_guard (region, next_e, stmt, *newivs,
1154 newivs_index, params_index);
1156 edge true_e = get_true_edge_from_guard_bb (next_e->dest);
1157 edge false_e = get_false_edge_from_guard_bb (next_e->dest);
1158 edge exit_true_e = single_succ_edge (true_e->dest);
1159 edge exit_false_e = single_succ_edge (false_e->dest);
1161 htab_t before_guard = htab_create (10, rename_map_elt_info,
1162 eq_rename_map_elts, free);
1163 htab_traverse (rename_map, copy_renames, before_guard);
1165 next_e = translate_clast (region, context_loop, stmt->then, true_e,
1166 rename_map, newivs, newivs_index, bb_pbb_mapping,
1167 level, params_index);
1169 insert_guard_phis (last_e->src, exit_true_e, exit_false_e,
1170 before_guard, rename_map);
1172 htab_delete (before_guard);
1177 /* Translates a CLAST statement STMT to GCC representation in the
1180 - NEXT_E is the edge where new generated code should be attached.
1181 - CONTEXT_LOOP is the loop in which the generated code will be placed
1182 - RENAME_MAP contains a set of tuples of new names associated to
1183 the original variables names.
1184 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping. */
1186 translate_clast (sese region, loop_p context_loop, struct clast_stmt *stmt,
1187 edge next_e, htab_t rename_map, VEC (tree, heap) **newivs,
1188 htab_t newivs_index, htab_t bb_pbb_mapping, int level,
1189 htab_t params_index)
1194 if (CLAST_STMT_IS_A (stmt, stmt_root))
1197 else if (CLAST_STMT_IS_A (stmt, stmt_user))
1198 next_e = translate_clast_user (region, (struct clast_user_stmt *) stmt,
1199 next_e, rename_map, newivs, newivs_index,
1200 bb_pbb_mapping, params_index);
1202 else if (CLAST_STMT_IS_A (stmt, stmt_for))
1203 next_e = translate_clast_for (region, context_loop,
1204 (struct clast_for *) stmt, next_e,
1205 rename_map, newivs, newivs_index,
1206 bb_pbb_mapping, level, params_index);
1208 else if (CLAST_STMT_IS_A (stmt, stmt_guard))
1209 next_e = translate_clast_guard (region, context_loop,
1210 (struct clast_guard *) stmt, next_e,
1211 rename_map, newivs, newivs_index,
1212 bb_pbb_mapping, level, params_index);
1214 else if (CLAST_STMT_IS_A (stmt, stmt_block))
1215 next_e = translate_clast (region, context_loop,
1216 ((struct clast_block *) stmt)->body,
1217 next_e, rename_map, newivs, newivs_index,
1218 bb_pbb_mapping, level, params_index);
1222 recompute_all_dominators ();
1225 return translate_clast (region, context_loop, stmt->next, next_e,
1226 rename_map, newivs, newivs_index,
1227 bb_pbb_mapping, level, params_index);
1230 /* Free the SCATTERING domain list. */
1233 free_scattering (CloogDomainList *scattering)
1237 CloogDomain *dom = cloog_domain (scattering);
1238 CloogDomainList *next = cloog_next_domain (scattering);
1240 cloog_domain_free (dom);
1246 /* Initialize Cloog's parameter names from the names used in GIMPLE.
1247 Initialize Cloog's iterator names, using 'graphite_iterator_%d'
1248 from 0 to scop_nb_loops (scop). */
1251 initialize_cloog_names (scop_p scop, CloogProgram *prog)
1253 sese region = SCOP_REGION (scop);
1255 int nb_iterators = scop_max_loop_depth (scop);
1256 int nb_scattering = cloog_program_nb_scattdims (prog);
1257 int nb_parameters = VEC_length (tree, SESE_PARAMS (region));
1258 char **iterators = XNEWVEC (char *, nb_iterators * 2);
1259 char **scattering = XNEWVEC (char *, nb_scattering);
1260 char **parameters= XNEWVEC (char *, nb_parameters);
1262 cloog_program_set_names (prog, cloog_names_malloc ());
1264 for (i = 0; i < nb_parameters; i++)
1266 tree param = VEC_index (tree, SESE_PARAMS(region), i);
1267 const char *name = get_name (param);
1273 len = strlen (name);
1275 parameters[i] = XNEWVEC (char, len + 1);
1276 snprintf (parameters[i], len, "%s_%d", name, SSA_NAME_VERSION (param));
1279 cloog_names_set_nb_parameters (cloog_program_names (prog), nb_parameters);
1280 cloog_names_set_parameters (cloog_program_names (prog), parameters);
1282 for (i = 0; i < nb_iterators; i++)
1285 iterators[i] = XNEWVEC (char, len);
1286 snprintf (iterators[i], len, "git_%d", i);
1289 cloog_names_set_nb_iterators (cloog_program_names (prog),
1291 cloog_names_set_iterators (cloog_program_names (prog),
1294 for (i = 0; i < nb_scattering; i++)
1297 scattering[i] = XNEWVEC (char, len);
1298 snprintf (scattering[i], len, "scat_%d", i);
1301 cloog_names_set_nb_scattering (cloog_program_names (prog),
1303 cloog_names_set_scattering (cloog_program_names (prog),
1307 /* Build cloog program for SCoP. */
1310 build_cloog_prog (scop_p scop, CloogProgram *prog)
1313 int max_nb_loops = scop_max_loop_depth (scop);
1315 CloogLoop *loop_list = NULL;
1316 CloogBlockList *block_list = NULL;
1317 CloogDomainList *scattering = NULL;
1318 int nbs = 2 * max_nb_loops + 1;
1321 cloog_program_set_context
1322 (prog, new_Cloog_Domain_from_ppl_Pointset_Powerset (SCOP_CONTEXT (scop)));
1323 nbs = unify_scattering_dimensions (scop);
1324 scaldims = (int *) xmalloc (nbs * (sizeof (int)));
1325 cloog_program_set_nb_scattdims (prog, nbs);
1326 initialize_cloog_names (scop, prog);
1328 for (i = 0; VEC_iterate (poly_bb_p, SCOP_BBS (scop), i, pbb); i++)
1330 CloogStatement *stmt;
1333 /* Dead code elimination: when the domain of a PBB is empty,
1334 don't generate code for the PBB. */
1335 if (ppl_Pointset_Powerset_C_Polyhedron_is_empty (PBB_DOMAIN (pbb)))
1338 /* Build the new statement and its block. */
1339 stmt = cloog_statement_alloc (pbb_index (pbb));
1340 block = cloog_block_alloc (stmt, 0, NULL, pbb_dim_iter_domain (pbb));
1341 cloog_statement_set_usr (stmt, pbb);
1343 /* Build loop list. */
1345 CloogLoop *new_loop_list = cloog_loop_malloc ();
1346 cloog_loop_set_next (new_loop_list, loop_list);
1347 cloog_loop_set_domain
1349 new_Cloog_Domain_from_ppl_Pointset_Powerset (PBB_DOMAIN (pbb)));
1350 cloog_loop_set_block (new_loop_list, block);
1351 loop_list = new_loop_list;
1354 /* Build block list. */
1356 CloogBlockList *new_block_list = cloog_block_list_malloc ();
1358 cloog_block_list_set_next (new_block_list, block_list);
1359 cloog_block_list_set_block (new_block_list, block);
1360 block_list = new_block_list;
1363 /* Build scattering list. */
1365 /* XXX: Replace with cloog_domain_list_alloc(), when available. */
1366 CloogDomainList *new_scattering
1367 = (CloogDomainList *) xmalloc (sizeof (CloogDomainList));
1368 ppl_Polyhedron_t scat;
1371 scat = PBB_TRANSFORMED_SCATTERING (pbb);
1372 dom = new_Cloog_Domain_from_ppl_Polyhedron (scat);
1374 cloog_set_next_domain (new_scattering, scattering);
1375 cloog_set_domain (new_scattering, dom);
1376 scattering = new_scattering;
1380 cloog_program_set_loop (prog, loop_list);
1381 cloog_program_set_blocklist (prog, block_list);
1383 for (i = 0; i < nbs; i++)
1386 cloog_program_set_scaldims (prog, scaldims);
1388 /* Extract scalar dimensions to simplify the code generation problem. */
1389 cloog_program_extract_scalars (prog, scattering);
1391 /* Apply scattering. */
1392 cloog_program_scatter (prog, scattering);
1393 free_scattering (scattering);
1395 /* Iterators corresponding to scalar dimensions have to be extracted. */
1396 cloog_names_scalarize (cloog_program_names (prog), nbs,
1397 cloog_program_scaldims (prog));
1399 /* Free blocklist. */
1401 CloogBlockList *next = cloog_program_blocklist (prog);
1405 CloogBlockList *toDelete = next;
1406 next = cloog_block_list_next (next);
1407 cloog_block_list_set_next (toDelete, NULL);
1408 cloog_block_list_set_block (toDelete, NULL);
1409 cloog_block_list_free (toDelete);
1411 cloog_program_set_blocklist (prog, NULL);
1415 /* Return the options that will be used in GLOOG. */
1417 static CloogOptions *
1418 set_cloog_options (void)
1420 CloogOptions *options = cloog_options_malloc ();
1422 /* Change cloog output language to C. If we do use FORTRAN instead, cloog
1423 will stop e.g. with "ERROR: unbounded loops not allowed in FORTRAN.", if
1424 we pass an incomplete program to cloog. */
1425 options->language = LANGUAGE_C;
1427 /* Enable complex equality spreading: removes dummy statements
1428 (assignments) in the generated code which repeats the
1429 substitution equations for statements. This is useless for
1433 /* Enable C pretty-printing mode: normalizes the substitution
1434 equations for statements. */
1437 /* Allow cloog to build strides with a stride width different to one.
1438 This example has stride = 4:
1440 for (i = 0; i < 20; i += 4)
1442 options->strides = 1;
1444 /* Disable optimizations and make cloog generate source code closer to the
1445 input. This is useful for debugging, but later we want the optimized
1448 XXX: We can not disable optimizations, as loop blocking is not working
1453 options->l = INT_MAX;
1459 /* Prints STMT to STDERR. */
1462 print_clast_stmt (FILE *file, struct clast_stmt *stmt)
1464 CloogOptions *options = set_cloog_options ();
1466 pprint (file, stmt, 0, options);
1467 cloog_options_free (options);
1470 /* Prints STMT to STDERR. */
1473 debug_clast_stmt (struct clast_stmt *stmt)
1475 print_clast_stmt (stderr, stmt);
1478 /* Translate SCOP to a CLooG program and clast. These two
1479 representations should be freed together: a clast cannot be used
1480 without a program. */
1483 scop_to_clast (scop_p scop)
1485 CloogOptions *options = set_cloog_options ();
1486 cloog_prog_clast pc;
1488 /* Connect new cloog prog generation to graphite. */
1489 pc.prog = cloog_program_malloc ();
1490 build_cloog_prog (scop, pc.prog);
1491 pc.prog = cloog_program_generate (pc.prog, options);
1492 pc.stmt = cloog_clast_create (pc.prog, options);
1494 cloog_options_free (options);
1498 /* Prints to FILE the code generated by CLooG for SCOP. */
1501 print_generated_program (FILE *file, scop_p scop)
1503 CloogOptions *options = set_cloog_options ();
1504 cloog_prog_clast pc = scop_to_clast (scop);
1506 fprintf (file, " (prog: \n");
1507 cloog_program_print (file, pc.prog);
1508 fprintf (file, " )\n");
1510 fprintf (file, " (clast: \n");
1511 pprint (file, pc.stmt, 0, options);
1512 fprintf (file, " )\n");
1514 cloog_options_free (options);
1515 cloog_clast_free (pc.stmt);
1516 cloog_program_free (pc.prog);
1519 /* Prints to STDERR the code generated by CLooG for SCOP. */
1522 debug_generated_program (scop_p scop)
1524 print_generated_program (stderr, scop);
1527 /* Add CLooG names to parameter index. The index is used to translate
1528 back from CLooG names to GCC trees. */
1531 create_params_index (htab_t index_table, CloogProgram *prog) {
1532 CloogNames* names = cloog_program_names (prog);
1533 int nb_parameters = cloog_names_nb_parameters (names);
1534 char **parameters = cloog_names_parameters (names);
1537 for (i = 0; i < nb_parameters; i++)
1538 save_clast_name_index (index_table, parameters[i], i);
1541 /* GIMPLE Loop Generator: generates loops from STMT in GIMPLE form for
1542 the given SCOP. Return true if code generation succeeded.
1543 BB_PBB_MAPPING is a basic_block and it's related poly_bb_p mapping.
1547 gloog (scop_p scop, VEC (scop_p, heap) *scops, htab_t bb_pbb_mapping)
1549 VEC (tree, heap) *newivs = VEC_alloc (tree, heap, 10);
1550 loop_p context_loop;
1551 sese region = SCOP_REGION (scop);
1552 ifsese if_region = NULL;
1553 htab_t rename_map, newivs_index, params_index;
1554 cloog_prog_clast pc;
1557 timevar_push (TV_GRAPHITE_CODE_GEN);
1558 gloog_error = false;
1560 pc = scop_to_clast (scop);
1562 if (dump_file && (dump_flags & TDF_DETAILS))
1564 fprintf (dump_file, "\nCLAST generated by CLooG: \n");
1565 print_clast_stmt (dump_file, pc.stmt);
1566 fprintf (dump_file, "\n");
1569 recompute_all_dominators ();
1572 if_region = move_sese_in_condition (region);
1573 sese_insert_phis_for_liveouts (region,
1574 if_region->region->exit->src,
1575 if_region->false_region->exit,
1576 if_region->true_region->exit);
1577 recompute_all_dominators ();
1580 context_loop = SESE_ENTRY (region)->src->loop_father;
1581 rename_map = htab_create (10, rename_map_elt_info, eq_rename_map_elts, free);
1582 newivs_index = htab_create (10, clast_name_index_elt_info,
1583 eq_clast_name_indexes, free);
1584 params_index = htab_create (10, clast_name_index_elt_info,
1585 eq_clast_name_indexes, free);
1587 create_params_index (params_index, pc.prog);
1589 translate_clast (region, context_loop, pc.stmt,
1590 if_region->true_region->entry,
1591 rename_map, &newivs, newivs_index,
1592 bb_pbb_mapping, 1, params_index);
1594 sese_adjust_liveout_phis (region, rename_map,
1595 if_region->region->exit->src,
1596 if_region->false_region->exit,
1597 if_region->true_region->exit);
1599 rename_nb_iterations (rename_map);
1601 for (i = 0; VEC_iterate (scop_p, scops, i, scop); i++)
1602 rename_sese_parameters (rename_map, SCOP_REGION (scop));
1604 recompute_all_dominators ();
1608 set_ifsese_condition (if_region, integer_zero_node);
1610 free (if_region->true_region);
1611 free (if_region->region);
1614 htab_delete (rename_map);
1615 htab_delete (newivs_index);
1616 htab_delete (params_index);
1617 VEC_free (tree, heap, newivs);
1618 cloog_clast_free (pc.stmt);
1619 cloog_program_free (pc.prog);
1620 timevar_pop (TV_GRAPHITE_CODE_GEN);
1622 if (dump_file && (dump_flags & TDF_DETAILS))
1626 int num_no_dependency = 0;
1628 FOR_EACH_LOOP (li, loop, 0)
1629 if (loop->can_be_parallel)
1630 num_no_dependency++;
1632 fprintf (dump_file, "\n%d loops carried no dependency.\n",
1636 return !gloog_error;