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"
31 #include "tree-dump.h"
34 #include "tree-chrec.h"
35 #include "tree-data-ref.h"
36 #include "tree-scalar-evolution.h"
37 #include "tree-pass.h"
39 #include "value-prof.h"
40 #include "pointer-set.h"
42 #include "langhooks.h"
46 #include "cloog/cloog.h"
48 #include "graphite-cloog-util.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"
55 #include "graphite-cloog-compat.h"
57 /* This flag is set when an error occurred during the translation of
59 static bool gloog_error;
61 /* Verifies properties that GRAPHITE should maintain during translation. */
64 graphite_verify (void)
66 #ifdef ENABLE_CHECKING
67 verify_loop_structure ();
68 verify_dominators (CDI_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 (clast_name_p name, htab_t index_table)
100 struct clast_name_index tmp;
104 gcc_assert (name->type == clast_expr_name);
105 tmp.name = ((const struct clast_name*) name)->name;
110 slot = htab_find_slot (index_table, &tmp, NO_INSERT);
113 return ((struct clast_name_index *) *slot)->index;
118 /* Records in INDEX_TABLE the INDEX for NAME. */
121 save_clast_name_index (htab_t index_table, const char *name, int index)
123 struct clast_name_index tmp;
127 slot = htab_find_slot (index_table, &tmp, INSERT);
134 *slot = new_clast_name_index (name, index);
138 /* Computes a hash function for database element ELT. */
140 static inline hashval_t
141 clast_name_index_elt_info (const void *elt)
143 return htab_hash_pointer (((const struct clast_name_index *) elt)->name);
146 /* Compares database elements E1 and E2. */
149 eq_clast_name_indexes (const void *e1, const void *e2)
151 const struct clast_name_index *elt1 = (const struct clast_name_index *) e1;
152 const struct clast_name_index *elt2 = (const struct clast_name_index *) e2;
154 return (elt1->name == elt2->name);
157 /* For a given scattering dimension, return the new induction variable
161 newivs_to_depth_to_newiv (VEC (tree, heap) *newivs, int depth)
163 return VEC_index (tree, newivs, depth);
168 /* Returns the tree variable from the name NAME that was given in
169 Cloog representation. */
172 clast_name_to_gcc (clast_name_p name, sese region, VEC (tree, heap) *newivs,
173 htab_t newivs_index, htab_t params_index)
176 VEC (tree, heap) *params = SESE_PARAMS (region);
178 if (params && params_index)
180 index = clast_name_to_index (name, params_index);
183 return VEC_index (tree, params, index);
186 gcc_assert (newivs && newivs_index);
187 index = clast_name_to_index (name, newivs_index);
188 gcc_assert (index >= 0);
190 return newivs_to_depth_to_newiv (newivs, index);
193 /* Returns the signed maximal precision type for expressions TYPE1 and TYPE2. */
196 max_signed_precision_type (tree type1, tree type2)
198 int p1 = TYPE_PRECISION (type1);
199 int p2 = TYPE_PRECISION (type2);
202 enum machine_mode mode;
205 precision = TYPE_UNSIGNED (type1) ? p1 * 2 : p1;
207 precision = TYPE_UNSIGNED (type2) ? p2 * 2 : p2;
209 if (precision > BITS_PER_WORD)
212 return integer_type_node;
215 mode = smallest_mode_for_size (precision, MODE_INT);
216 precision = GET_MODE_PRECISION (mode);
217 type = build_nonstandard_integer_type (precision, false);
222 return integer_type_node;
228 /* Returns the maximal precision type for expressions TYPE1 and TYPE2. */
231 max_precision_type (tree type1, tree type2)
233 if (POINTER_TYPE_P (type1))
236 if (POINTER_TYPE_P (type2))
239 if (!TYPE_UNSIGNED (type1)
240 || !TYPE_UNSIGNED (type2))
241 return max_signed_precision_type (type1, type2);
243 return TYPE_PRECISION (type1) > TYPE_PRECISION (type2) ? type1 : type2;
247 clast_to_gcc_expression (tree, struct clast_expr *, sese, VEC (tree, heap) *,
250 /* Converts a Cloog reduction expression R with reduction operation OP
251 to a GCC expression tree of type TYPE. */
254 clast_to_gcc_expression_red (tree type, enum tree_code op,
255 struct clast_reduction *r,
256 sese region, VEC (tree, heap) *newivs,
257 htab_t newivs_index, htab_t params_index)
260 tree res = clast_to_gcc_expression (type, r->elts[0], region, newivs,
261 newivs_index, params_index);
262 tree operand_type = (op == POINTER_PLUS_EXPR) ? sizetype : type;
264 for (i = 1; i < r->n; i++)
266 tree t = clast_to_gcc_expression (operand_type, r->elts[i], region,
267 newivs, newivs_index, params_index);
268 res = fold_build2 (op, type, res, t);
274 /* Converts a Cloog AST expression E back to a GCC expression tree of
278 clast_to_gcc_expression (tree type, struct clast_expr *e,
279 sese region, VEC (tree, heap) *newivs,
280 htab_t newivs_index, htab_t params_index)
284 case clast_expr_term:
286 struct clast_term *t = (struct clast_term *) e;
290 if (mpz_cmp_si (t->val, 1) == 0)
292 tree name = clast_name_to_gcc (t->var, region, newivs,
293 newivs_index, params_index);
295 if (POINTER_TYPE_P (TREE_TYPE (name)) != POINTER_TYPE_P (type))
296 name = fold_convert (sizetype, name);
298 name = fold_convert (type, name);
302 else if (mpz_cmp_si (t->val, -1) == 0)
304 tree name = clast_name_to_gcc (t->var, region, newivs,
305 newivs_index, params_index);
307 if (POINTER_TYPE_P (TREE_TYPE (name)) != POINTER_TYPE_P (type))
308 name = fold_convert (sizetype, name);
310 name = fold_convert (type, name);
312 return fold_build1 (NEGATE_EXPR, type, name);
316 tree name = clast_name_to_gcc (t->var, region, newivs,
317 newivs_index, params_index);
318 tree cst = gmp_cst_to_tree (type, t->val);
320 if (POINTER_TYPE_P (TREE_TYPE (name)) != POINTER_TYPE_P (type))
321 name = fold_convert (sizetype, name);
323 name = fold_convert (type, name);
325 if (!POINTER_TYPE_P (type))
326 return fold_build2 (MULT_EXPR, type, cst, name);
333 return gmp_cst_to_tree (type, t->val);
338 struct clast_reduction *r = (struct clast_reduction *) e;
343 return clast_to_gcc_expression_red
344 (type, POINTER_TYPE_P (type) ? POINTER_PLUS_EXPR : PLUS_EXPR,
345 r, region, newivs, newivs_index, params_index);
348 return clast_to_gcc_expression_red (type, MIN_EXPR, r, region,
349 newivs, newivs_index,
353 return clast_to_gcc_expression_red (type, MAX_EXPR, r, region,
354 newivs, newivs_index,
365 struct clast_binary *b = (struct clast_binary *) e;
366 struct clast_expr *lhs = (struct clast_expr *) b->LHS;
367 tree tl = clast_to_gcc_expression (type, lhs, region, newivs,
368 newivs_index, params_index);
369 tree tr = gmp_cst_to_tree (type, b->RHS);
374 return fold_build2 (FLOOR_DIV_EXPR, type, tl, tr);
377 return fold_build2 (CEIL_DIV_EXPR, type, tl, tr);
380 return fold_build2 (EXACT_DIV_EXPR, type, tl, tr);
383 return fold_build2 (TRUNC_MOD_EXPR, type, tl, tr);
397 /* Return the precision needed to represent the value VAL. */
400 precision_for_value (mpz_t val)
416 while (mpz_cmp (y, x) >= 0)
429 /* Return the precision needed to represent the values between LOW and
433 precision_for_interval (mpz_t low, mpz_t up)
438 gcc_assert (mpz_cmp (low, up) <= 0);
441 mpz_sub (diff, up, low);
442 precision = precision_for_value (diff);
448 /* Return a type that could represent the integer value VAL. */
451 gcc_type_for_interval (mpz_t low, mpz_t up)
453 bool unsigned_p = true;
454 int precision, prec_up, prec_int;
456 enum machine_mode mode;
458 gcc_assert (mpz_cmp (low, up) <= 0);
460 prec_up = precision_for_value (up);
461 prec_int = precision_for_interval (low, up);
462 precision = MAX (prec_up, prec_int);
464 if (precision > BITS_PER_WORD)
467 return integer_type_node;
470 if (mpz_sgn (low) <= 0)
473 else if (precision < BITS_PER_WORD)
479 mode = smallest_mode_for_size (precision, MODE_INT);
480 precision = GET_MODE_PRECISION (mode);
481 type = build_nonstandard_integer_type (precision, unsigned_p);
486 return integer_type_node;
492 /* Return a type that could represent the integer value VAL, or
493 otherwise return NULL_TREE. */
496 gcc_type_for_value (mpz_t val)
498 return gcc_type_for_interval (val, val);
501 /* Return the type for the clast_term T used in STMT. */
504 gcc_type_for_clast_term (struct clast_term *t,
505 sese region, VEC (tree, heap) *newivs,
506 htab_t newivs_index, htab_t params_index)
508 gcc_assert (t->expr.type == clast_expr_term);
511 return gcc_type_for_value (t->val);
513 return TREE_TYPE (clast_name_to_gcc (t->var, region, newivs,
514 newivs_index, params_index));
518 gcc_type_for_clast_expr (struct clast_expr *, sese,
519 VEC (tree, heap) *, htab_t, htab_t);
521 /* Return the type for the clast_reduction R used in STMT. */
524 gcc_type_for_clast_red (struct clast_reduction *r, sese region,
525 VEC (tree, heap) *newivs,
526 htab_t newivs_index, htab_t params_index)
529 tree type = NULL_TREE;
532 return gcc_type_for_clast_expr (r->elts[0], region, newivs,
533 newivs_index, params_index);
540 type = gcc_type_for_clast_expr (r->elts[0], region, newivs,
541 newivs_index, params_index);
542 for (i = 1; i < r->n; i++)
543 type = max_precision_type (type, gcc_type_for_clast_expr
544 (r->elts[i], region, newivs,
545 newivs_index, params_index));
557 /* Return the type for the clast_binary B used in STMT. */
560 gcc_type_for_clast_bin (struct clast_binary *b,
561 sese region, VEC (tree, heap) *newivs,
562 htab_t newivs_index, htab_t params_index)
564 tree l = gcc_type_for_clast_expr ((struct clast_expr *) b->LHS, region,
565 newivs, newivs_index, params_index);
566 tree r = gcc_type_for_value (b->RHS);
567 return max_signed_precision_type (l, r);
570 /* Returns the type for the CLAST expression E when used in statement
574 gcc_type_for_clast_expr (struct clast_expr *e,
575 sese region, VEC (tree, heap) *newivs,
576 htab_t newivs_index, htab_t params_index)
580 case clast_expr_term:
581 return gcc_type_for_clast_term ((struct clast_term *) e, region,
582 newivs, newivs_index, params_index);
585 return gcc_type_for_clast_red ((struct clast_reduction *) e, region,
586 newivs, newivs_index, params_index);
589 return gcc_type_for_clast_bin ((struct clast_binary *) e, region,
590 newivs, newivs_index, params_index);
599 /* Returns the type for the equation CLEQ. */
602 gcc_type_for_clast_eq (struct clast_equation *cleq,
603 sese region, VEC (tree, heap) *newivs,
604 htab_t newivs_index, htab_t params_index)
606 tree l = gcc_type_for_clast_expr (cleq->LHS, region, newivs,
607 newivs_index, params_index);
608 tree r = gcc_type_for_clast_expr (cleq->RHS, region, newivs,
609 newivs_index, params_index);
610 return max_precision_type (l, r);
613 /* Translates a clast equation CLEQ to a tree. */
616 graphite_translate_clast_equation (sese region,
617 struct clast_equation *cleq,
618 VEC (tree, heap) *newivs,
619 htab_t newivs_index, htab_t params_index)
622 tree type = gcc_type_for_clast_eq (cleq, region, newivs, newivs_index,
624 tree lhs = clast_to_gcc_expression (type, cleq->LHS, region, newivs,
625 newivs_index, params_index);
626 tree rhs = clast_to_gcc_expression (type, cleq->RHS, region, newivs,
627 newivs_index, params_index);
632 else if (cleq->sign > 0)
638 return fold_build2 (comp, boolean_type_node, lhs, rhs);
641 /* Creates the test for the condition in STMT. */
644 graphite_create_guard_cond_expr (sese region, struct clast_guard *stmt,
645 VEC (tree, heap) *newivs,
646 htab_t newivs_index, htab_t params_index)
651 for (i = 0; i < stmt->n; i++)
653 tree eq = graphite_translate_clast_equation (region, &stmt->eq[i],
654 newivs, newivs_index,
658 cond = fold_build2 (TRUTH_AND_EXPR, TREE_TYPE (eq), cond, eq);
666 /* Creates a new if region corresponding to Cloog's guard. */
669 graphite_create_new_guard (sese region, edge entry_edge,
670 struct clast_guard *stmt,
671 VEC (tree, heap) *newivs,
672 htab_t newivs_index, htab_t params_index)
674 tree cond_expr = graphite_create_guard_cond_expr (region, stmt, newivs,
675 newivs_index, params_index);
676 edge exit_edge = create_empty_if_region_on_edge (entry_edge, cond_expr);
680 /* Compute the lower bound LOW and upper bound UP for the induction
681 variable at LEVEL for the statement PBB, based on the transformed
682 scattering of PBB: T|I|G|Cst, with T the scattering transform, I
683 the iteration domain, and G the context parameters. */
686 compute_bounds_for_level (poly_bb_p pbb, int level, mpz_t low, mpz_t up)
688 ppl_Pointset_Powerset_C_Polyhedron_t ps;
689 ppl_Linear_Expression_t le;
691 combine_context_id_scat (&ps, pbb, false);
693 /* Prepare the linear expression corresponding to the level that we
694 want to maximize/minimize. */
696 ppl_dimension_type dim = pbb_nb_scattering_transform (pbb)
697 + pbb_dim_iter_domain (pbb) + pbb_nb_params (pbb);
699 ppl_new_Linear_Expression_with_dimension (&le, dim);
700 ppl_set_coef (le, 2 * level + 1, 1);
703 ppl_max_for_le_pointset (ps, le, up);
704 ppl_min_for_le_pointset (ps, le, low);
705 ppl_delete_Linear_Expression (le);
706 ppl_delete_Pointset_Powerset_C_Polyhedron (ps);
709 /* Compute the type for the induction variable at LEVEL for the
710 statement PBB, based on the transformed schedule of PBB. */
713 compute_type_for_level (poly_bb_p pbb, int level)
721 compute_bounds_for_level (pbb, level, low, up);
722 type = gcc_type_for_interval (low, up);
729 /* Walks a CLAST and returns the first statement in the body of a
732 static struct clast_user_stmt *
733 clast_get_body_of_loop (struct clast_stmt *stmt)
736 || CLAST_STMT_IS_A (stmt, stmt_user))
737 return (struct clast_user_stmt *) stmt;
739 if (CLAST_STMT_IS_A (stmt, stmt_for))
740 return clast_get_body_of_loop (((struct clast_for *) stmt)->body);
742 if (CLAST_STMT_IS_A (stmt, stmt_guard))
743 return clast_get_body_of_loop (((struct clast_guard *) stmt)->then);
745 if (CLAST_STMT_IS_A (stmt, stmt_block))
746 return clast_get_body_of_loop (((struct clast_block *) stmt)->body);
751 /* Returns the type for the induction variable for the loop translated
755 gcc_type_for_iv_of_clast_loop (struct clast_for *stmt_for, int level,
756 tree lb_type, tree ub_type)
758 struct clast_stmt *stmt = (struct clast_stmt *) stmt_for;
759 struct clast_user_stmt *body = clast_get_body_of_loop (stmt);
760 CloogStatement *cs = body->statement;
761 poly_bb_p pbb = (poly_bb_p) cloog_statement_usr (cs);
763 return max_signed_precision_type (lb_type, max_precision_type
764 (ub_type, compute_type_for_level
768 /* Creates a new LOOP corresponding to Cloog's STMT. Inserts an
769 induction variable for the new LOOP. New LOOP is attached to CFG
770 starting at ENTRY_EDGE. LOOP is inserted into the loop tree and
771 becomes the child loop of the OUTER_LOOP. NEWIVS_INDEX binds
772 CLooG's scattering name to the induction variable created for the
773 loop of STMT. The new induction variable is inserted in the NEWIVS
777 graphite_create_new_loop (sese region, edge entry_edge,
778 struct clast_for *stmt,
779 loop_p outer, VEC (tree, heap) **newivs,
780 htab_t newivs_index, htab_t params_index, int level)
782 tree lb_type = gcc_type_for_clast_expr (stmt->LB, region, *newivs,
783 newivs_index, params_index);
784 tree ub_type = gcc_type_for_clast_expr (stmt->UB, region, *newivs,
785 newivs_index, params_index);
786 tree type = gcc_type_for_iv_of_clast_loop (stmt, level, lb_type, ub_type);
787 tree lb = clast_to_gcc_expression (type, stmt->LB, region, *newivs,
788 newivs_index, params_index);
789 tree ub = clast_to_gcc_expression (type, stmt->UB, region, *newivs,
790 newivs_index, params_index);
791 tree stride = gmp_cst_to_tree (type, stmt->stride);
792 tree ivvar = create_tmp_var (type, "graphite_IV");
793 tree iv, iv_after_increment;
794 loop_p loop = create_empty_loop_on_edge
795 (entry_edge, lb, stride, ub, ivvar, &iv, &iv_after_increment,
796 outer ? outer : entry_edge->src->loop_father);
798 add_referenced_var (ivvar);
800 save_clast_name_index (newivs_index, stmt->iterator,
801 VEC_length (tree, *newivs));
802 VEC_safe_push (tree, heap, *newivs, iv);
806 /* Inserts in iv_map a tuple (OLD_LOOP->num, NEW_NAME) for the
807 induction variables of the loops around GBB in SESE. */
810 build_iv_mapping (VEC (tree, heap) *iv_map, sese region,
811 VEC (tree, heap) *newivs, htab_t newivs_index,
812 struct clast_user_stmt *user_stmt,
815 struct clast_stmt *t;
817 CloogStatement *cs = user_stmt->statement;
818 poly_bb_p pbb = (poly_bb_p) cloog_statement_usr (cs);
819 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
821 for (t = user_stmt->substitutions; t; t = t->next, depth++)
823 struct clast_expr *expr = (struct clast_expr *)
824 ((struct clast_assignment *)t)->RHS;
825 tree type = gcc_type_for_clast_expr (expr, region, newivs,
826 newivs_index, params_index);
827 tree new_name = clast_to_gcc_expression (type, expr, region, newivs,
828 newivs_index, params_index);
829 loop_p old_loop = gbb_loop_at_index (gbb, region, depth);
831 VEC_replace (tree, iv_map, old_loop->num, new_name);
835 /* Construct bb_pbb_def with BB and PBB. */
838 new_bb_pbb_def (basic_block bb, poly_bb_p pbb)
840 bb_pbb_def *bb_pbb_p;
842 bb_pbb_p = XNEW (bb_pbb_def);
849 /* Mark BB with it's relevant PBB via hashing table BB_PBB_MAPPING. */
852 mark_bb_with_pbb (poly_bb_p pbb, basic_block bb, htab_t bb_pbb_mapping)
858 x = htab_find_slot (bb_pbb_mapping, &tmp, INSERT);
861 *x = new_bb_pbb_def (bb, pbb);
864 /* Find BB's related poly_bb_p in hash table BB_PBB_MAPPING. */
867 find_pbb_via_hash (htab_t bb_pbb_mapping, basic_block bb)
873 slot = htab_find_slot (bb_pbb_mapping, &tmp, NO_INSERT);
876 return ((bb_pbb_def *) *slot)->pbb;
881 /* Check data dependency in LOOP at scattering level LEVEL.
882 BB_PBB_MAPPING is a basic_block and it's related poly_bb_p
886 dependency_in_loop_p (loop_p loop, htab_t bb_pbb_mapping, int level)
889 basic_block *bbs = get_loop_body_in_dom_order (loop);
891 for (i = 0; i < loop->num_nodes; i++)
893 poly_bb_p pbb1 = find_pbb_via_hash (bb_pbb_mapping, bbs[i]);
898 for (j = 0; j < loop->num_nodes; j++)
900 poly_bb_p pbb2 = find_pbb_via_hash (bb_pbb_mapping, bbs[j]);
905 if (dependency_between_pbbs_p (pbb1, pbb2, level))
918 /* Translates a clast user statement STMT to gimple.
920 - REGION is the sese region we used to generate the scop.
921 - NEXT_E is the edge where new generated code should be attached.
922 - CONTEXT_LOOP is the loop in which the generated code will be placed
923 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping.
924 - PARAMS_INDEX connects the cloog parameters with the gimple parameters in
927 translate_clast_user (sese region, struct clast_user_stmt *stmt, edge next_e,
928 VEC (tree, heap) **newivs,
929 htab_t newivs_index, htab_t bb_pbb_mapping,
934 poly_bb_p pbb = (poly_bb_p) cloog_statement_usr (stmt->statement);
935 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
936 VEC (tree, heap) *iv_map;
938 if (GBB_BB (gbb) == ENTRY_BLOCK_PTR)
941 nb_loops = number_of_loops ();
942 iv_map = VEC_alloc (tree, heap, nb_loops);
943 for (i = 0; i < nb_loops; i++)
944 VEC_quick_push (tree, iv_map, NULL_TREE);
946 build_iv_mapping (iv_map, region, *newivs, newivs_index, stmt, params_index);
947 next_e = copy_bb_and_scalar_dependences (GBB_BB (gbb), region,
949 VEC_free (tree, heap, iv_map);
951 new_bb = next_e->src;
952 mark_bb_with_pbb (pbb, new_bb, bb_pbb_mapping);
953 update_ssa (TODO_update_ssa);
958 /* Creates a new if region protecting the loop to be executed, if the execution
959 count is zero (lb > ub). */
962 graphite_create_new_loop_guard (sese region, edge entry_edge,
963 struct clast_for *stmt,
964 VEC (tree, heap) *newivs,
965 htab_t newivs_index, htab_t params_index)
969 tree lb_type = gcc_type_for_clast_expr (stmt->LB, region, newivs,
970 newivs_index, params_index);
971 tree ub_type = gcc_type_for_clast_expr (stmt->UB, region, newivs,
972 newivs_index, params_index);
973 tree type = max_precision_type (lb_type, ub_type);
974 tree lb = clast_to_gcc_expression (type, stmt->LB, region, newivs,
975 newivs_index, params_index);
976 tree ub = clast_to_gcc_expression (type, stmt->UB, region, newivs,
977 newivs_index, params_index);
978 tree one = POINTER_TYPE_P (type) ? size_one_node
979 : fold_convert (type, integer_one_node);
980 /* Adding +1 and using LT_EXPR helps with loop latches that have a
981 loop iteration count of "PARAMETER - 1". For PARAMETER == 0 this becomes
982 2^{32|64}, and the condition lb <= ub is true, even if we do not want this.
983 However lb < ub + 1 is false, as expected. */
984 tree ub_one = fold_build2 (POINTER_TYPE_P (type) ? POINTER_PLUS_EXPR
985 : PLUS_EXPR, type, ub, one);
987 /* When ub + 1 wraps around, use lb <= ub. */
988 if (integer_zerop (ub_one))
989 cond_expr = fold_build2 (LE_EXPR, boolean_type_node, lb, ub);
991 cond_expr = fold_build2 (LT_EXPR, boolean_type_node, lb, ub_one);
993 exit_edge = create_empty_if_region_on_edge (entry_edge, cond_expr);
999 translate_clast (sese, loop_p, struct clast_stmt *, edge,
1000 VEC (tree, heap) **, htab_t, htab_t, int, htab_t);
1002 /* Create the loop for a clast for statement.
1004 - REGION is the sese region we used to generate the scop.
1005 - NEXT_E is the edge where new generated code should be attached.
1006 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping.
1007 - PARAMS_INDEX connects the cloog parameters with the gimple parameters in
1010 translate_clast_for_loop (sese region, loop_p context_loop,
1011 struct clast_for *stmt, edge next_e,
1012 VEC (tree, heap) **newivs,
1013 htab_t newivs_index, htab_t bb_pbb_mapping,
1014 int level, htab_t params_index)
1016 struct loop *loop = graphite_create_new_loop (region, next_e, stmt,
1017 context_loop, newivs,
1018 newivs_index, params_index,
1020 edge last_e = single_exit (loop);
1021 edge to_body = single_succ_edge (loop->header);
1022 basic_block after = to_body->dest;
1024 /* Create a basic block for loop close phi nodes. */
1025 last_e = single_succ_edge (split_edge (last_e));
1027 /* Translate the body of the loop. */
1028 next_e = translate_clast (region, loop, stmt->body, to_body,
1029 newivs, newivs_index, bb_pbb_mapping, level + 1,
1031 redirect_edge_succ_nodup (next_e, after);
1032 set_immediate_dominator (CDI_DOMINATORS, next_e->dest, next_e->src);
1034 if (flag_loop_parallelize_all
1035 && !dependency_in_loop_p (loop, bb_pbb_mapping,
1036 get_scattering_level (level)))
1037 loop->can_be_parallel = true;
1042 /* Translates a clast for statement STMT to gimple. First a guard is created
1043 protecting the loop, if it is executed zero times. In this guard we create
1044 the real loop structure.
1046 - REGION is the sese region we used to generate the scop.
1047 - NEXT_E is the edge where new generated code should be attached.
1048 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping.
1049 - PARAMS_INDEX connects the cloog parameters with the gimple parameters in
1052 translate_clast_for (sese region, loop_p context_loop, struct clast_for *stmt,
1053 edge next_e, VEC (tree, heap) **newivs,
1054 htab_t newivs_index, htab_t bb_pbb_mapping, int level,
1055 htab_t params_index)
1057 edge last_e = graphite_create_new_loop_guard (region, next_e, stmt, *newivs,
1058 newivs_index, params_index);
1059 edge true_e = get_true_edge_from_guard_bb (next_e->dest);
1061 translate_clast_for_loop (region, context_loop, stmt, true_e, newivs,
1062 newivs_index, bb_pbb_mapping, level,
1067 /* Translates a clast guard statement STMT to gimple.
1069 - REGION is the sese region we used to generate the scop.
1070 - NEXT_E is the edge where new generated code should be attached.
1071 - CONTEXT_LOOP is the loop in which the generated code will be placed
1072 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping.
1073 - PARAMS_INDEX connects the cloog parameters with the gimple parameters in
1076 translate_clast_guard (sese region, loop_p context_loop,
1077 struct clast_guard *stmt, edge next_e,
1078 VEC (tree, heap) **newivs,
1079 htab_t newivs_index, htab_t bb_pbb_mapping, int level,
1080 htab_t params_index)
1082 edge last_e = graphite_create_new_guard (region, next_e, stmt, *newivs,
1083 newivs_index, params_index);
1084 edge true_e = get_true_edge_from_guard_bb (next_e->dest);
1086 translate_clast (region, context_loop, stmt->then, true_e,
1087 newivs, newivs_index, bb_pbb_mapping,
1088 level, params_index);
1092 /* Translates a CLAST statement STMT to GCC representation in the
1095 - NEXT_E is the edge where new generated code should be attached.
1096 - CONTEXT_LOOP is the loop in which the generated code will be placed
1097 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping. */
1099 translate_clast (sese region, loop_p context_loop, struct clast_stmt *stmt,
1100 edge next_e, VEC (tree, heap) **newivs,
1101 htab_t newivs_index, htab_t bb_pbb_mapping, int level,
1102 htab_t params_index)
1107 if (CLAST_STMT_IS_A (stmt, stmt_root))
1110 else if (CLAST_STMT_IS_A (stmt, stmt_user))
1111 next_e = translate_clast_user (region, (struct clast_user_stmt *) stmt,
1112 next_e, newivs, newivs_index,
1113 bb_pbb_mapping, params_index);
1115 else if (CLAST_STMT_IS_A (stmt, stmt_for))
1116 next_e = translate_clast_for (region, context_loop,
1117 (struct clast_for *) stmt, next_e,
1118 newivs, newivs_index,
1119 bb_pbb_mapping, level, params_index);
1121 else if (CLAST_STMT_IS_A (stmt, stmt_guard))
1122 next_e = translate_clast_guard (region, context_loop,
1123 (struct clast_guard *) stmt, next_e,
1124 newivs, newivs_index,
1125 bb_pbb_mapping, level, params_index);
1127 else if (CLAST_STMT_IS_A (stmt, stmt_block))
1128 next_e = translate_clast (region, context_loop,
1129 ((struct clast_block *) stmt)->body,
1130 next_e, newivs, newivs_index,
1131 bb_pbb_mapping, level, params_index);
1135 recompute_all_dominators ();
1138 return translate_clast (region, context_loop, stmt->next, next_e,
1139 newivs, newivs_index,
1140 bb_pbb_mapping, level, params_index);
1143 /* Free the SCATTERING domain list. */
1146 free_scattering (CloogScatteringList *scattering)
1150 CloogScattering *dom = cloog_scattering (scattering);
1151 CloogScatteringList *next = cloog_next_scattering (scattering);
1153 cloog_scattering_free (dom);
1159 /* Initialize Cloog's parameter names from the names used in GIMPLE.
1160 Initialize Cloog's iterator names, using 'graphite_iterator_%d'
1161 from 0 to scop_nb_loops (scop). */
1164 initialize_cloog_names (scop_p scop, CloogProgram *prog)
1166 sese region = SCOP_REGION (scop);
1168 int nb_iterators = scop_max_loop_depth (scop);
1169 int nb_scattering = cloog_program_nb_scattdims (prog);
1170 int nb_parameters = VEC_length (tree, SESE_PARAMS (region));
1171 char **iterators = XNEWVEC (char *, nb_iterators * 2);
1172 char **scattering = XNEWVEC (char *, nb_scattering);
1173 char **parameters= XNEWVEC (char *, nb_parameters);
1175 cloog_program_set_names (prog, cloog_names_malloc ());
1177 for (i = 0; i < nb_parameters; i++)
1179 tree param = VEC_index (tree, SESE_PARAMS(region), i);
1180 const char *name = get_name (param);
1186 len = strlen (name);
1188 parameters[i] = XNEWVEC (char, len + 1);
1189 snprintf (parameters[i], len, "%s_%d", name, SSA_NAME_VERSION (param));
1192 cloog_names_set_nb_parameters (cloog_program_names (prog), nb_parameters);
1193 cloog_names_set_parameters (cloog_program_names (prog), parameters);
1195 for (i = 0; i < nb_iterators; i++)
1198 iterators[i] = XNEWVEC (char, len);
1199 snprintf (iterators[i], len, "git_%d", i);
1202 cloog_names_set_nb_iterators (cloog_program_names (prog),
1204 cloog_names_set_iterators (cloog_program_names (prog),
1207 for (i = 0; i < nb_scattering; i++)
1210 scattering[i] = XNEWVEC (char, len);
1211 snprintf (scattering[i], len, "scat_%d", i);
1214 cloog_names_set_nb_scattering (cloog_program_names (prog),
1216 cloog_names_set_scattering (cloog_program_names (prog),
1220 /* Initialize a CLooG input file. */
1223 init_cloog_input_file (int scop_number)
1225 FILE *graphite_out_file;
1226 int len = strlen (dump_base_name);
1227 char *dumpname = XNEWVEC (char, len + 25);
1228 char *s_scop_number = XNEWVEC (char, 15);
1230 memcpy (dumpname, dump_base_name, len + 1);
1231 strip_off_ending (dumpname, len);
1232 sprintf (s_scop_number, ".%d", scop_number);
1233 strcat (dumpname, s_scop_number);
1234 strcat (dumpname, ".cloog");
1235 graphite_out_file = fopen (dumpname, "w+b");
1237 if (graphite_out_file == 0)
1238 fatal_error ("can%'t open %s for writing: %m", dumpname);
1242 return graphite_out_file;
1245 /* Build cloog program for SCoP. */
1248 build_cloog_prog (scop_p scop, CloogProgram *prog,
1249 CloogOptions *options, CloogState *state ATTRIBUTE_UNUSED)
1252 int max_nb_loops = scop_max_loop_depth (scop);
1254 CloogLoop *loop_list = NULL;
1255 CloogBlockList *block_list = NULL;
1256 CloogScatteringList *scattering = NULL;
1257 int nbs = 2 * max_nb_loops + 1;
1260 cloog_program_set_context
1261 (prog, new_Cloog_Domain_from_ppl_Pointset_Powerset (SCOP_CONTEXT (scop),
1262 scop_nb_params (scop), state));
1263 nbs = unify_scattering_dimensions (scop);
1264 scaldims = (int *) xmalloc (nbs * (sizeof (int)));
1265 cloog_program_set_nb_scattdims (prog, nbs);
1266 initialize_cloog_names (scop, prog);
1268 FOR_EACH_VEC_ELT (poly_bb_p, SCOP_BBS (scop), i, pbb)
1270 CloogStatement *stmt;
1274 /* Dead code elimination: when the domain of a PBB is empty,
1275 don't generate code for the PBB. */
1276 if (ppl_Pointset_Powerset_C_Polyhedron_is_empty (PBB_DOMAIN (pbb)))
1279 /* Build the new statement and its block. */
1280 stmt = cloog_statement_alloc (state, pbb_index (pbb));
1281 dom = new_Cloog_Domain_from_ppl_Pointset_Powerset (PBB_DOMAIN (pbb),
1282 scop_nb_params (scop),
1284 block = cloog_block_alloc (stmt, 0, NULL, pbb_dim_iter_domain (pbb));
1285 cloog_statement_set_usr (stmt, pbb);
1287 /* Build loop list. */
1289 CloogLoop *new_loop_list = cloog_loop_malloc (state);
1290 cloog_loop_set_next (new_loop_list, loop_list);
1291 cloog_loop_set_domain (new_loop_list, dom);
1292 cloog_loop_set_block (new_loop_list, block);
1293 loop_list = new_loop_list;
1296 /* Build block list. */
1298 CloogBlockList *new_block_list = cloog_block_list_malloc ();
1300 cloog_block_list_set_next (new_block_list, block_list);
1301 cloog_block_list_set_block (new_block_list, block);
1302 block_list = new_block_list;
1305 /* Build scattering list. */
1307 /* XXX: Replace with cloog_domain_list_alloc(), when available. */
1308 CloogScatteringList *new_scattering
1309 = (CloogScatteringList *) xmalloc (sizeof (CloogScatteringList));
1310 ppl_Polyhedron_t scat;
1311 CloogScattering *dom;
1313 scat = PBB_TRANSFORMED_SCATTERING (pbb);
1314 dom = new_Cloog_Scattering_from_ppl_Polyhedron
1315 (scat, scop_nb_params (scop), pbb_nb_scattering_transform (pbb),
1318 cloog_set_next_scattering (new_scattering, scattering);
1319 cloog_set_scattering (new_scattering, dom);
1320 scattering = new_scattering;
1324 cloog_program_set_loop (prog, loop_list);
1325 cloog_program_set_blocklist (prog, block_list);
1327 for (i = 0; i < nbs; i++)
1330 cloog_program_set_scaldims (prog, scaldims);
1332 /* Extract scalar dimensions to simplify the code generation problem. */
1333 cloog_program_extract_scalars (prog, scattering, options);
1335 /* Dump a .cloog input file, if requested. This feature is only
1336 enabled in the Graphite branch. */
1339 static size_t file_scop_number = 0;
1340 FILE *cloog_file = init_cloog_input_file (file_scop_number);
1342 cloog_program_dump_cloog (cloog_file, prog, scattering);
1346 /* Apply scattering. */
1347 cloog_program_scatter (prog, scattering, options);
1348 free_scattering (scattering);
1350 /* Iterators corresponding to scalar dimensions have to be extracted. */
1351 cloog_names_scalarize (cloog_program_names (prog), nbs,
1352 cloog_program_scaldims (prog));
1354 /* Free blocklist. */
1356 CloogBlockList *next = cloog_program_blocklist (prog);
1360 CloogBlockList *toDelete = next;
1361 next = cloog_block_list_next (next);
1362 cloog_block_list_set_next (toDelete, NULL);
1363 cloog_block_list_set_block (toDelete, NULL);
1364 cloog_block_list_free (toDelete);
1366 cloog_program_set_blocklist (prog, NULL);
1370 /* Return the options that will be used in GLOOG. */
1372 static CloogOptions *
1373 set_cloog_options (CloogState *state ATTRIBUTE_UNUSED)
1375 CloogOptions *options = cloog_options_malloc (state);
1377 /* Change cloog output language to C. If we do use FORTRAN instead, cloog
1378 will stop e.g. with "ERROR: unbounded loops not allowed in FORTRAN.", if
1379 we pass an incomplete program to cloog. */
1380 options->language = LANGUAGE_C;
1382 /* Enable complex equality spreading: removes dummy statements
1383 (assignments) in the generated code which repeats the
1384 substitution equations for statements. This is useless for
1389 /* Silence CLooG to avoid failing tests due to debug output to stderr. */
1392 /* Enable C pretty-printing mode: normalizes the substitution
1393 equations for statements. */
1397 /* Allow cloog to build strides with a stride width different to one.
1398 This example has stride = 4:
1400 for (i = 0; i < 20; i += 4)
1402 options->strides = 1;
1404 /* Disable optimizations and make cloog generate source code closer to the
1405 input. This is useful for debugging, but later we want the optimized
1408 XXX: We can not disable optimizations, as loop blocking is not working
1413 options->l = INT_MAX;
1419 /* Prints STMT to STDERR. */
1422 print_clast_stmt (FILE *file, struct clast_stmt *stmt)
1424 CloogState *state = cloog_state_malloc ();
1425 CloogOptions *options = set_cloog_options (state);
1427 clast_pprint (file, stmt, 0, options);
1428 cloog_options_free (options);
1429 cloog_state_free (state);
1432 /* Prints STMT to STDERR. */
1435 debug_clast_stmt (struct clast_stmt *stmt)
1437 print_clast_stmt (stderr, stmt);
1440 /* Translate SCOP to a CLooG program and clast. These two
1441 representations should be freed together: a clast cannot be used
1442 without a program. */
1445 scop_to_clast (scop_p scop, CloogState *state)
1447 CloogOptions *options = set_cloog_options (state);
1448 cloog_prog_clast pc;
1450 /* Connect new cloog prog generation to graphite. */
1451 pc.prog = cloog_program_malloc ();
1452 build_cloog_prog (scop, pc.prog, options, state);
1453 pc.prog = cloog_program_generate (pc.prog, options);
1454 pc.stmt = cloog_clast_create (pc.prog, options);
1456 cloog_options_free (options);
1460 /* Prints to FILE the code generated by CLooG for SCOP. */
1463 print_generated_program (FILE *file, scop_p scop)
1465 CloogState *state = cloog_state_malloc ();
1466 CloogOptions *options = set_cloog_options (state);
1468 cloog_prog_clast pc = scop_to_clast (scop, state);
1470 fprintf (file, " (prog: \n");
1471 cloog_program_print (file, pc.prog);
1472 fprintf (file, " )\n");
1474 fprintf (file, " (clast: \n");
1475 clast_pprint (file, pc.stmt, 0, options);
1476 fprintf (file, " )\n");
1478 cloog_options_free (options);
1479 cloog_clast_free (pc.stmt);
1480 cloog_program_free (pc.prog);
1483 /* Prints to STDERR the code generated by CLooG for SCOP. */
1486 debug_generated_program (scop_p scop)
1488 print_generated_program (stderr, scop);
1491 /* Add CLooG names to parameter index. The index is used to translate
1492 back from CLooG names to GCC trees. */
1495 create_params_index (htab_t index_table, CloogProgram *prog) {
1496 CloogNames* names = cloog_program_names (prog);
1497 int nb_parameters = cloog_names_nb_parameters (names);
1498 char **parameters = cloog_names_parameters (names);
1501 for (i = 0; i < nb_parameters; i++)
1502 save_clast_name_index (index_table, parameters[i], i);
1505 /* GIMPLE Loop Generator: generates loops from STMT in GIMPLE form for
1506 the given SCOP. Return true if code generation succeeded.
1507 BB_PBB_MAPPING is a basic_block and it's related poly_bb_p mapping.
1511 gloog (scop_p scop, htab_t bb_pbb_mapping)
1513 VEC (tree, heap) *newivs = VEC_alloc (tree, heap, 10);
1514 loop_p context_loop;
1515 sese region = SCOP_REGION (scop);
1516 ifsese if_region = NULL;
1517 htab_t newivs_index, params_index;
1518 cloog_prog_clast pc;
1521 state = cloog_state_malloc ();
1522 timevar_push (TV_GRAPHITE_CODE_GEN);
1523 gloog_error = false;
1525 pc = scop_to_clast (scop, state);
1527 if (dump_file && (dump_flags & TDF_DETAILS))
1529 fprintf (dump_file, "\nCLAST generated by CLooG: \n");
1530 print_clast_stmt (dump_file, pc.stmt);
1531 fprintf (dump_file, "\n");
1534 recompute_all_dominators ();
1537 if_region = move_sese_in_condition (region);
1538 sese_insert_phis_for_liveouts (region,
1539 if_region->region->exit->src,
1540 if_region->false_region->exit,
1541 if_region->true_region->exit);
1542 recompute_all_dominators ();
1545 context_loop = SESE_ENTRY (region)->src->loop_father;
1546 newivs_index = htab_create (10, clast_name_index_elt_info,
1547 eq_clast_name_indexes, free);
1548 params_index = htab_create (10, clast_name_index_elt_info,
1549 eq_clast_name_indexes, free);
1551 create_params_index (params_index, pc.prog);
1553 translate_clast (region, context_loop, pc.stmt,
1554 if_region->true_region->entry,
1555 &newivs, newivs_index,
1556 bb_pbb_mapping, 1, params_index);
1559 recompute_all_dominators ();
1563 set_ifsese_condition (if_region, integer_zero_node);
1565 free (if_region->true_region);
1566 free (if_region->region);
1569 htab_delete (newivs_index);
1570 htab_delete (params_index);
1571 VEC_free (tree, heap, newivs);
1572 cloog_clast_free (pc.stmt);
1573 cloog_program_free (pc.prog);
1574 timevar_pop (TV_GRAPHITE_CODE_GEN);
1576 if (dump_file && (dump_flags & TDF_DETAILS))
1580 int num_no_dependency = 0;
1582 FOR_EACH_LOOP (li, loop, 0)
1583 if (loop->can_be_parallel)
1584 num_no_dependency++;
1586 fprintf (dump_file, "\n%d loops carried no dependency.\n",
1590 cloog_state_free (state);
1592 return !gloog_error;