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);
238 precision = TYPE_UNSIGNED (type1) ? p1 * 2 : p1;
240 precision = TYPE_UNSIGNED (type2) ? p2 * 2 : p2;
242 type = lang_hooks.types.type_for_size (precision, false);
247 return integer_type_node;
252 /* Returns the maximal precision type for expressions TYPE1 and TYPE2. */
255 max_precision_type (tree type1, tree type2)
257 if (POINTER_TYPE_P (type1))
260 if (POINTER_TYPE_P (type2))
263 if (!TYPE_UNSIGNED (type1)
264 || !TYPE_UNSIGNED (type2))
265 return max_signed_precision_type (type1, type2);
267 return TYPE_PRECISION (type1) > TYPE_PRECISION (type2) ? type1 : type2;
271 clast_to_gcc_expression (tree, struct clast_expr *, sese, VEC (tree, heap) *,
274 /* Converts a Cloog reduction expression R with reduction operation OP
275 to a GCC expression tree of type TYPE. */
278 clast_to_gcc_expression_red (tree type, enum tree_code op,
279 struct clast_reduction *r,
280 sese region, VEC (tree, heap) *newivs,
281 htab_t newivs_index, htab_t params_index)
284 tree res = clast_to_gcc_expression (type, r->elts[0], region, newivs,
285 newivs_index, params_index);
286 tree operand_type = (op == POINTER_PLUS_EXPR) ? sizetype : type;
288 for (i = 1; i < r->n; i++)
290 tree t = clast_to_gcc_expression (operand_type, r->elts[i], region,
291 newivs, newivs_index, params_index);
292 res = fold_build2 (op, type, res, t);
298 /* Converts a Cloog AST expression E back to a GCC expression tree of
302 clast_to_gcc_expression (tree type, struct clast_expr *e,
303 sese region, VEC (tree, heap) *newivs,
304 htab_t newivs_index, htab_t params_index)
310 struct clast_term *t = (struct clast_term *) e;
314 if (mpz_cmp_si (t->val, 1) == 0)
316 tree name = clast_name_to_gcc (t->var, region, newivs,
317 newivs_index, params_index);
319 if (POINTER_TYPE_P (TREE_TYPE (name)) != POINTER_TYPE_P (type))
320 name = fold_convert (sizetype, name);
322 name = fold_convert (type, name);
326 else if (mpz_cmp_si (t->val, -1) == 0)
328 tree name = clast_name_to_gcc (t->var, region, newivs,
329 newivs_index, params_index);
331 if (POINTER_TYPE_P (TREE_TYPE (name)) != POINTER_TYPE_P (type))
332 name = fold_convert (sizetype, name);
334 name = fold_convert (type, name);
336 return fold_build1 (NEGATE_EXPR, type, name);
340 tree name = clast_name_to_gcc (t->var, region, newivs,
341 newivs_index, params_index);
342 tree cst = gmp_cst_to_tree (type, t->val);
344 if (POINTER_TYPE_P (TREE_TYPE (name)) != POINTER_TYPE_P (type))
345 name = fold_convert (sizetype, name);
347 name = fold_convert (type, name);
349 if (!POINTER_TYPE_P (type))
350 return fold_build2 (MULT_EXPR, type, cst, name);
357 return gmp_cst_to_tree (type, t->val);
362 struct clast_reduction *r = (struct clast_reduction *) e;
367 return clast_to_gcc_expression_red
368 (type, POINTER_TYPE_P (type) ? POINTER_PLUS_EXPR : PLUS_EXPR,
369 r, region, newivs, newivs_index, params_index);
372 return clast_to_gcc_expression_red (type, MIN_EXPR, r, region,
373 newivs, newivs_index,
377 return clast_to_gcc_expression_red (type, MAX_EXPR, r, region,
378 newivs, newivs_index,
389 struct clast_binary *b = (struct clast_binary *) e;
390 struct clast_expr *lhs = (struct clast_expr *) b->LHS;
391 tree tl = clast_to_gcc_expression (type, lhs, region, newivs,
392 newivs_index, params_index);
393 tree tr = gmp_cst_to_tree (type, b->RHS);
398 return fold_build2 (FLOOR_DIV_EXPR, type, tl, tr);
401 return fold_build2 (CEIL_DIV_EXPR, type, tl, tr);
404 return fold_build2 (EXACT_DIV_EXPR, type, tl, tr);
407 return fold_build2 (TRUNC_MOD_EXPR, type, tl, tr);
421 /* Return the precision needed to represent the value VAL. */
424 precision_for_value (mpz_t val)
433 value_assign (y, val);
434 value_set_si (two, 2);
440 while (value_gt (y, x))
442 value_multiply (x, x, two);
453 /* Return the precision needed to represent the values between LOW and
457 precision_for_interval (mpz_t low, mpz_t up)
462 gcc_assert (value_le (low, up));
465 value_subtract (diff, up, low);
466 precision = precision_for_value (diff);
472 /* Return a type that could represent the integer value VAL. */
475 gcc_type_for_interval (mpz_t low, mpz_t up)
477 bool unsigned_p = true;
478 int precision, prec_up, prec_int;
480 enum machine_mode mode;
482 gcc_assert (value_le (low, up));
484 if (value_neg_p (low))
487 prec_up = precision_for_value (up);
488 prec_int = precision_for_interval (low, up);
489 precision = MAX (prec_up, prec_int);
491 if (precision > BITS_PER_WORD)
494 return integer_type_node;
497 mode = smallest_mode_for_size (precision, MODE_INT);
498 precision = GET_MODE_PRECISION (mode);
499 type = build_nonstandard_integer_type (precision, unsigned_p);
504 return integer_type_node;
510 /* Return a type that could represent the integer value VAL, or
511 otherwise return NULL_TREE. */
514 gcc_type_for_value (mpz_t val)
516 return gcc_type_for_interval (val, val);
519 /* Return the type for the clast_term T used in STMT. */
522 gcc_type_for_clast_term (struct clast_term *t,
523 sese region, VEC (tree, heap) *newivs,
524 htab_t newivs_index, htab_t params_index)
526 gcc_assert (t->expr.type == expr_term);
529 return gcc_type_for_value (t->val);
531 return TREE_TYPE (clast_name_to_gcc (t->var, region, newivs,
532 newivs_index, params_index));
536 gcc_type_for_clast_expr (struct clast_expr *, sese,
537 VEC (tree, heap) *, htab_t, htab_t);
539 /* Return the type for the clast_reduction R used in STMT. */
542 gcc_type_for_clast_red (struct clast_reduction *r, sese region,
543 VEC (tree, heap) *newivs,
544 htab_t newivs_index, htab_t params_index)
547 tree type = NULL_TREE;
550 return gcc_type_for_clast_expr (r->elts[0], region, newivs,
551 newivs_index, params_index);
558 type = gcc_type_for_clast_expr (r->elts[0], region, newivs,
559 newivs_index, params_index);
560 for (i = 1; i < r->n; i++)
561 type = max_precision_type (type, gcc_type_for_clast_expr
562 (r->elts[i], region, newivs,
563 newivs_index, params_index));
575 /* Return the type for the clast_binary B used in STMT. */
578 gcc_type_for_clast_bin (struct clast_binary *b,
579 sese region, VEC (tree, heap) *newivs,
580 htab_t newivs_index, htab_t params_index)
582 tree l = gcc_type_for_clast_expr ((struct clast_expr *) b->LHS, region,
583 newivs, newivs_index, params_index);
584 tree r = gcc_type_for_value (b->RHS);
585 return max_signed_precision_type (l, r);
588 /* Returns the type for the CLAST expression E when used in statement
592 gcc_type_for_clast_expr (struct clast_expr *e,
593 sese region, VEC (tree, heap) *newivs,
594 htab_t newivs_index, htab_t params_index)
599 return gcc_type_for_clast_term ((struct clast_term *) e, region,
600 newivs, newivs_index, params_index);
603 return gcc_type_for_clast_red ((struct clast_reduction *) e, region,
604 newivs, newivs_index, params_index);
607 return gcc_type_for_clast_bin ((struct clast_binary *) e, region,
608 newivs, newivs_index, params_index);
617 /* Returns the type for the equation CLEQ. */
620 gcc_type_for_clast_eq (struct clast_equation *cleq,
621 sese region, VEC (tree, heap) *newivs,
622 htab_t newivs_index, htab_t params_index)
624 tree l = gcc_type_for_clast_expr (cleq->LHS, region, newivs,
625 newivs_index, params_index);
626 tree r = gcc_type_for_clast_expr (cleq->RHS, region, newivs,
627 newivs_index, params_index);
628 return max_precision_type (l, r);
631 /* Translates a clast equation CLEQ to a tree. */
634 graphite_translate_clast_equation (sese region,
635 struct clast_equation *cleq,
636 VEC (tree, heap) *newivs,
637 htab_t newivs_index, htab_t params_index)
640 tree type = gcc_type_for_clast_eq (cleq, region, newivs, newivs_index,
642 tree lhs = clast_to_gcc_expression (type, cleq->LHS, region, newivs,
643 newivs_index, params_index);
644 tree rhs = clast_to_gcc_expression (type, cleq->RHS, region, newivs,
645 newivs_index, params_index);
650 else if (cleq->sign > 0)
656 return fold_build2 (comp, boolean_type_node, lhs, rhs);
659 /* Creates the test for the condition in STMT. */
662 graphite_create_guard_cond_expr (sese region, struct clast_guard *stmt,
663 VEC (tree, heap) *newivs,
664 htab_t newivs_index, htab_t params_index)
669 for (i = 0; i < stmt->n; i++)
671 tree eq = graphite_translate_clast_equation (region, &stmt->eq[i],
672 newivs, newivs_index,
676 cond = fold_build2 (TRUTH_AND_EXPR, TREE_TYPE (eq), cond, eq);
684 /* Creates a new if region corresponding to Cloog's guard. */
687 graphite_create_new_guard (sese region, edge entry_edge,
688 struct clast_guard *stmt,
689 VEC (tree, heap) *newivs,
690 htab_t newivs_index, htab_t params_index)
692 tree cond_expr = graphite_create_guard_cond_expr (region, stmt, newivs,
693 newivs_index, params_index);
694 edge exit_edge = create_empty_if_region_on_edge (entry_edge, cond_expr);
698 /* Compute the lower bound LOW and upper bound UP for the induction
699 variable at LEVEL for the statement PBB, based on the transformed
700 scattering of PBB: T|I|G|Cst, with T the scattering transform, I
701 the iteration domain, and G the context parameters. */
704 compute_bounds_for_level (poly_bb_p pbb, int level, mpz_t low, mpz_t up)
706 ppl_Pointset_Powerset_C_Polyhedron_t ps;
707 ppl_Linear_Expression_t le;
709 combine_context_id_scat (&ps, pbb, false);
711 /* Prepare the linear expression corresponding to the level that we
712 want to maximize/minimize. */
714 ppl_dimension_type dim = pbb_nb_scattering_transform (pbb)
715 + pbb_dim_iter_domain (pbb) + pbb_nb_params (pbb);
717 ppl_new_Linear_Expression_with_dimension (&le, dim);
718 ppl_set_coef (le, 2 * level + 1, 1);
721 ppl_max_for_le_pointset (ps, le, up);
722 ppl_min_for_le_pointset (ps, le, low);
725 /* Compute the type for the induction variable at LEVEL for the
726 statement PBB, based on the transformed schedule of PBB. */
729 compute_type_for_level (poly_bb_p pbb, int level)
737 compute_bounds_for_level (pbb, level, low, up);
738 type = gcc_type_for_interval (low, up);
745 /* Walks a CLAST and returns the first statement in the body of a
748 static struct clast_user_stmt *
749 clast_get_body_of_loop (struct clast_stmt *stmt)
752 || CLAST_STMT_IS_A (stmt, stmt_user))
753 return (struct clast_user_stmt *) stmt;
755 if (CLAST_STMT_IS_A (stmt, stmt_for))
756 return clast_get_body_of_loop (((struct clast_for *) stmt)->body);
758 if (CLAST_STMT_IS_A (stmt, stmt_guard))
759 return clast_get_body_of_loop (((struct clast_guard *) stmt)->then);
761 if (CLAST_STMT_IS_A (stmt, stmt_block))
762 return clast_get_body_of_loop (((struct clast_block *) stmt)->body);
767 /* Returns the type for the induction variable for the loop translated
771 gcc_type_for_iv_of_clast_loop (struct clast_for *stmt_for, int level,
772 tree lb_type, tree ub_type)
774 struct clast_stmt *stmt = (struct clast_stmt *) stmt_for;
775 struct clast_user_stmt *body = clast_get_body_of_loop (stmt);
776 CloogStatement *cs = body->statement;
777 poly_bb_p pbb = (poly_bb_p) cloog_statement_usr (cs);
779 return max_signed_precision_type (lb_type, max_precision_type
780 (ub_type, compute_type_for_level
784 /* Creates a new LOOP corresponding to Cloog's STMT. Inserts an
785 induction variable for the new LOOP. New LOOP is attached to CFG
786 starting at ENTRY_EDGE. LOOP is inserted into the loop tree and
787 becomes the child loop of the OUTER_LOOP. NEWIVS_INDEX binds
788 CLooG's scattering name to the induction variable created for the
789 loop of STMT. The new induction variable is inserted in the NEWIVS
793 graphite_create_new_loop (sese region, edge entry_edge,
794 struct clast_for *stmt,
795 loop_p outer, VEC (tree, heap) **newivs,
796 htab_t newivs_index, htab_t params_index, int level)
798 tree lb_type = gcc_type_for_clast_expr (stmt->LB, region, *newivs,
799 newivs_index, params_index);
800 tree ub_type = gcc_type_for_clast_expr (stmt->UB, region, *newivs,
801 newivs_index, params_index);
802 tree type = gcc_type_for_iv_of_clast_loop (stmt, level, lb_type, ub_type);
803 tree lb = clast_to_gcc_expression (type, stmt->LB, region, *newivs,
804 newivs_index, params_index);
805 tree ub = clast_to_gcc_expression (type, stmt->UB, region, *newivs,
806 newivs_index, params_index);
807 tree stride = gmp_cst_to_tree (type, stmt->stride);
808 tree ivvar = create_tmp_var (type, "graphite_IV");
809 tree iv, iv_after_increment;
810 loop_p loop = create_empty_loop_on_edge
811 (entry_edge, lb, stride, ub, ivvar, &iv, &iv_after_increment,
812 outer ? outer : entry_edge->src->loop_father);
814 add_referenced_var (ivvar);
816 save_clast_name_index (newivs_index, stmt->iterator,
817 VEC_length (tree, *newivs));
818 VEC_safe_push (tree, heap, *newivs, iv);
822 /* Inserts in MAP a tuple (OLD_NAME, NEW_NAME) for the induction
823 variables of the loops around GBB in SESE. */
826 build_iv_mapping (htab_t map, sese region,
827 VEC (tree, heap) *newivs, htab_t newivs_index,
828 struct clast_user_stmt *user_stmt,
831 struct clast_stmt *t;
833 CloogStatement *cs = user_stmt->statement;
834 poly_bb_p pbb = (poly_bb_p) cloog_statement_usr (cs);
836 for (t = user_stmt->substitutions; t; t = t->next, index++)
838 struct clast_expr *expr = (struct clast_expr *)
839 ((struct clast_assignment *)t)->RHS;
840 tree type = gcc_type_for_clast_expr (expr, region, newivs,
841 newivs_index, params_index);
842 tree old_name = pbb_to_depth_to_oldiv (pbb, index);
843 tree e = clast_to_gcc_expression (type, expr, region, newivs,
844 newivs_index, params_index);
845 set_rename (map, old_name, e);
849 /* Helper function for htab_traverse. */
852 copy_renames (void **slot, void *s)
854 struct rename_map_elt_s *entry = (struct rename_map_elt_s *) *slot;
855 htab_t res = (htab_t) s;
856 tree old_name = entry->old_name;
857 tree expr = entry->expr;
858 struct rename_map_elt_s tmp;
861 tmp.old_name = old_name;
862 x = htab_find_slot (res, &tmp, INSERT);
865 *x = new_rename_map_elt (old_name, expr);
870 /* Construct bb_pbb_def with BB and PBB. */
873 new_bb_pbb_def (basic_block bb, poly_bb_p pbb)
875 bb_pbb_def *bb_pbb_p;
877 bb_pbb_p = XNEW (bb_pbb_def);
884 /* Mark BB with it's relevant PBB via hashing table BB_PBB_MAPPING. */
887 mark_bb_with_pbb (poly_bb_p pbb, basic_block bb, htab_t bb_pbb_mapping)
893 x = htab_find_slot (bb_pbb_mapping, &tmp, INSERT);
896 *x = new_bb_pbb_def (bb, pbb);
899 /* Find BB's related poly_bb_p in hash table BB_PBB_MAPPING. */
902 find_pbb_via_hash (htab_t bb_pbb_mapping, basic_block bb)
908 slot = htab_find_slot (bb_pbb_mapping, &tmp, NO_INSERT);
911 return ((bb_pbb_def *) *slot)->pbb;
916 /* Check data dependency in LOOP at scattering level LEVEL.
917 BB_PBB_MAPPING is a basic_block and it's related poly_bb_p
921 dependency_in_loop_p (loop_p loop, htab_t bb_pbb_mapping, int level)
924 basic_block *bbs = get_loop_body_in_dom_order (loop);
926 for (i = 0; i < loop->num_nodes; i++)
928 poly_bb_p pbb1 = find_pbb_via_hash (bb_pbb_mapping, bbs[i]);
933 for (j = 0; j < loop->num_nodes; j++)
935 poly_bb_p pbb2 = find_pbb_via_hash (bb_pbb_mapping, bbs[j]);
940 if (dependency_between_pbbs_p (pbb1, pbb2, level))
954 translate_clast (sese, loop_p, struct clast_stmt *, edge, htab_t,
955 VEC (tree, heap) **, htab_t, htab_t, int, htab_t);
957 /* Translates a clast user statement STMT to gimple.
959 - REGION is the sese region we used to generate the scop.
960 - NEXT_E is the edge where new generated code should be attached.
961 - CONTEXT_LOOP is the loop in which the generated code will be placed
962 - RENAME_MAP contains a set of tuples of new names associated to
963 the original variables names.
964 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping.
965 - PARAMS_INDEX connects the cloog parameters with the gimple parameters in
968 translate_clast_user (sese region, struct clast_user_stmt *stmt, edge next_e,
969 htab_t rename_map, VEC (tree, heap) **newivs,
970 htab_t newivs_index, htab_t bb_pbb_mapping,
975 poly_bb_p pbb = (poly_bb_p) cloog_statement_usr (stmt->statement);
976 gbb = PBB_BLACK_BOX (pbb);
978 if (GBB_BB (gbb) == ENTRY_BLOCK_PTR)
981 build_iv_mapping (rename_map, region, *newivs, newivs_index, stmt,
983 next_e = copy_bb_and_scalar_dependences (GBB_BB (gbb), region,
985 new_bb = next_e->src;
986 mark_bb_with_pbb (pbb, new_bb, bb_pbb_mapping);
987 update_ssa (TODO_update_ssa);
992 /* Creates a new if region protecting the loop to be executed, if the execution
993 count is zero (lb > ub). */
995 graphite_create_new_loop_guard (sese region, edge entry_edge,
996 struct clast_for *stmt,
997 VEC (tree, heap) *newivs,
998 htab_t newivs_index, htab_t params_index)
1002 tree lb_type = gcc_type_for_clast_expr (stmt->LB, region, newivs,
1003 newivs_index, params_index);
1004 tree ub_type = gcc_type_for_clast_expr (stmt->UB, region, newivs,
1005 newivs_index, params_index);
1006 tree type = max_precision_type (lb_type, ub_type);
1007 tree lb = clast_to_gcc_expression (type, stmt->LB, region, newivs,
1008 newivs_index, params_index);
1009 tree ub = clast_to_gcc_expression (type, stmt->UB, region, newivs,
1010 newivs_index, params_index);
1013 /* Adding +1 and using LT_EXPR helps with loop latches that have a
1014 loop iteration count of "PARAMETER - 1". For PARAMETER == 0 this becomes
1015 2^{32|64}, and the condition lb <= ub is true, even if we do not want this.
1016 However lb < ub + 1 is false, as expected. */
1021 mpz_set_si (gmp_one, 1);
1022 one = gmp_cst_to_tree (type, gmp_one);
1023 mpz_clear (gmp_one);
1025 ub_one = fold_build2 (POINTER_TYPE_P (type) ? POINTER_PLUS_EXPR : PLUS_EXPR,
1028 /* When ub + 1 wraps around, use lb <= ub. */
1029 if (integer_zerop (ub_one))
1030 cond_expr = fold_build2 (LE_EXPR, boolean_type_node, lb, ub);
1032 cond_expr = fold_build2 (LT_EXPR, boolean_type_node, lb, ub_one);
1034 exit_edge = create_empty_if_region_on_edge (entry_edge, cond_expr);
1040 /* Create the loop for a clast for statement.
1042 - REGION is the sese region we used to generate the scop.
1043 - NEXT_E is the edge where new generated code should be attached.
1044 - RENAME_MAP contains a set of tuples of new names associated to
1045 the original variables names.
1046 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping.
1047 - PARAMS_INDEX connects the cloog parameters with the gimple parameters in
1050 translate_clast_for_loop (sese region, loop_p context_loop,
1051 struct clast_for *stmt, edge next_e,
1052 htab_t rename_map, VEC (tree, heap) **newivs,
1053 htab_t newivs_index, htab_t bb_pbb_mapping,
1054 int level, htab_t params_index)
1056 struct loop *loop = graphite_create_new_loop (region, next_e, stmt,
1057 context_loop, newivs,
1058 newivs_index, params_index,
1060 edge last_e = single_exit (loop);
1061 edge to_body = single_succ_edge (loop->header);
1062 basic_block after = to_body->dest;
1064 /* Create a basic block for loop close phi nodes. */
1065 last_e = single_succ_edge (split_edge (last_e));
1067 /* Translate the body of the loop. */
1068 next_e = translate_clast (region, loop, stmt->body, to_body, rename_map,
1069 newivs, newivs_index, bb_pbb_mapping, level + 1,
1071 redirect_edge_succ_nodup (next_e, after);
1072 set_immediate_dominator (CDI_DOMINATORS, next_e->dest, next_e->src);
1074 /* Remove from rename_map all the tuples containing variables
1075 defined in loop's body. */
1076 insert_loop_close_phis (rename_map, loop);
1078 if (flag_loop_parallelize_all
1079 && !dependency_in_loop_p (loop, bb_pbb_mapping,
1080 get_scattering_level (level)))
1081 loop->can_be_parallel = true;
1086 /* Translates a clast for statement STMT to gimple. First a guard is created
1087 protecting the loop, if it is executed zero times. In this guard we create
1088 the real loop structure.
1090 - REGION is the sese region we used to generate the scop.
1091 - NEXT_E is the edge where new generated code should be attached.
1092 - RENAME_MAP contains a set of tuples of new names associated to
1093 the original variables names.
1094 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping.
1095 - PARAMS_INDEX connects the cloog parameters with the gimple parameters in
1098 translate_clast_for (sese region, loop_p context_loop, struct clast_for *stmt,
1099 edge next_e, htab_t rename_map, VEC (tree, heap) **newivs,
1100 htab_t newivs_index, htab_t bb_pbb_mapping, int level,
1101 htab_t params_index)
1103 edge last_e = graphite_create_new_loop_guard (region, next_e, stmt, *newivs,
1104 newivs_index, params_index);
1106 edge true_e = get_true_edge_from_guard_bb (next_e->dest);
1107 edge false_e = get_false_edge_from_guard_bb (next_e->dest);
1108 edge exit_true_e = single_succ_edge (true_e->dest);
1109 edge exit_false_e = single_succ_edge (false_e->dest);
1111 htab_t before_guard = htab_create (10, rename_map_elt_info,
1112 eq_rename_map_elts, free);
1113 htab_traverse (rename_map, copy_renames, before_guard);
1115 next_e = translate_clast_for_loop (region, context_loop, stmt, true_e,
1117 newivs_index, bb_pbb_mapping, level,
1120 insert_guard_phis (last_e->src, exit_true_e, exit_false_e,
1121 before_guard, rename_map);
1123 htab_delete (before_guard);
1128 /* Translates a clast guard statement STMT to gimple.
1130 - REGION is the sese region we used to generate the scop.
1131 - NEXT_E is the edge where new generated code should be attached.
1132 - CONTEXT_LOOP is the loop in which the generated code will be placed
1133 - RENAME_MAP contains a set of tuples of new names associated to
1134 the original variables names.
1135 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping.
1136 - PARAMS_INDEX connects the cloog parameters with the gimple parameters in
1139 translate_clast_guard (sese region, loop_p context_loop,
1140 struct clast_guard *stmt, edge next_e,
1141 htab_t rename_map, VEC (tree, heap) **newivs,
1142 htab_t newivs_index, htab_t bb_pbb_mapping, int level,
1143 htab_t params_index)
1145 edge last_e = graphite_create_new_guard (region, next_e, stmt, *newivs,
1146 newivs_index, params_index);
1148 edge true_e = get_true_edge_from_guard_bb (next_e->dest);
1149 edge false_e = get_false_edge_from_guard_bb (next_e->dest);
1150 edge exit_true_e = single_succ_edge (true_e->dest);
1151 edge exit_false_e = single_succ_edge (false_e->dest);
1153 htab_t before_guard = htab_create (10, rename_map_elt_info,
1154 eq_rename_map_elts, free);
1155 htab_traverse (rename_map, copy_renames, before_guard);
1157 next_e = translate_clast (region, context_loop, stmt->then, true_e,
1158 rename_map, newivs, newivs_index, bb_pbb_mapping,
1159 level, params_index);
1161 insert_guard_phis (last_e->src, exit_true_e, exit_false_e,
1162 before_guard, rename_map);
1164 htab_delete (before_guard);
1169 /* Translates a CLAST statement STMT to GCC representation in the
1172 - NEXT_E is the edge where new generated code should be attached.
1173 - CONTEXT_LOOP is the loop in which the generated code will be placed
1174 - RENAME_MAP contains a set of tuples of new names associated to
1175 the original variables names.
1176 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping. */
1178 translate_clast (sese region, loop_p context_loop, struct clast_stmt *stmt,
1179 edge next_e, htab_t rename_map, VEC (tree, heap) **newivs,
1180 htab_t newivs_index, htab_t bb_pbb_mapping, int level,
1181 htab_t params_index)
1186 if (CLAST_STMT_IS_A (stmt, stmt_root))
1189 else if (CLAST_STMT_IS_A (stmt, stmt_user))
1190 next_e = translate_clast_user (region, (struct clast_user_stmt *) stmt,
1191 next_e, rename_map, newivs, newivs_index,
1192 bb_pbb_mapping, params_index);
1194 else if (CLAST_STMT_IS_A (stmt, stmt_for))
1195 next_e = translate_clast_for (region, context_loop,
1196 (struct clast_for *) stmt, next_e,
1197 rename_map, newivs, newivs_index,
1198 bb_pbb_mapping, level, params_index);
1200 else if (CLAST_STMT_IS_A (stmt, stmt_guard))
1201 next_e = translate_clast_guard (region, context_loop,
1202 (struct clast_guard *) stmt, next_e,
1203 rename_map, newivs, newivs_index,
1204 bb_pbb_mapping, level, params_index);
1206 else if (CLAST_STMT_IS_A (stmt, stmt_block))
1207 next_e = translate_clast (region, context_loop,
1208 ((struct clast_block *) stmt)->body,
1209 next_e, rename_map, newivs, newivs_index,
1210 bb_pbb_mapping, level, params_index);
1214 recompute_all_dominators ();
1217 return translate_clast (region, context_loop, stmt->next, next_e,
1218 rename_map, newivs, newivs_index,
1219 bb_pbb_mapping, level, params_index);
1222 /* Free the SCATTERING domain list. */
1225 free_scattering (CloogDomainList *scattering)
1229 CloogDomain *dom = cloog_domain (scattering);
1230 CloogDomainList *next = cloog_next_domain (scattering);
1232 cloog_domain_free (dom);
1238 /* Initialize Cloog's parameter names from the names used in GIMPLE.
1239 Initialize Cloog's iterator names, using 'graphite_iterator_%d'
1240 from 0 to scop_nb_loops (scop). */
1243 initialize_cloog_names (scop_p scop, CloogProgram *prog)
1245 sese region = SCOP_REGION (scop);
1247 int nb_iterators = scop_max_loop_depth (scop);
1248 int nb_scattering = cloog_program_nb_scattdims (prog);
1249 int nb_parameters = VEC_length (tree, SESE_PARAMS (region));
1250 char **iterators = XNEWVEC (char *, nb_iterators * 2);
1251 char **scattering = XNEWVEC (char *, nb_scattering);
1252 char **parameters= XNEWVEC (char *, nb_parameters);
1254 cloog_program_set_names (prog, cloog_names_malloc ());
1256 for (i = 0; i < nb_parameters; i++)
1258 tree param = VEC_index (tree, SESE_PARAMS(region), i);
1259 const char *name = get_name (param);
1265 len = strlen (name);
1267 parameters[i] = XNEWVEC (char, len + 1);
1268 snprintf (parameters[i], len, "%s_%d", name, SSA_NAME_VERSION (param));
1271 cloog_names_set_nb_parameters (cloog_program_names (prog), nb_parameters);
1272 cloog_names_set_parameters (cloog_program_names (prog), parameters);
1274 for (i = 0; i < nb_iterators; i++)
1277 iterators[i] = XNEWVEC (char, len);
1278 snprintf (iterators[i], len, "git_%d", i);
1281 cloog_names_set_nb_iterators (cloog_program_names (prog),
1283 cloog_names_set_iterators (cloog_program_names (prog),
1286 for (i = 0; i < nb_scattering; i++)
1289 scattering[i] = XNEWVEC (char, len);
1290 snprintf (scattering[i], len, "scat_%d", i);
1293 cloog_names_set_nb_scattering (cloog_program_names (prog),
1295 cloog_names_set_scattering (cloog_program_names (prog),
1299 /* Build cloog program for SCoP. */
1302 build_cloog_prog (scop_p scop, CloogProgram *prog)
1305 int max_nb_loops = scop_max_loop_depth (scop);
1307 CloogLoop *loop_list = NULL;
1308 CloogBlockList *block_list = NULL;
1309 CloogDomainList *scattering = NULL;
1310 int nbs = 2 * max_nb_loops + 1;
1313 cloog_program_set_context
1314 (prog, new_Cloog_Domain_from_ppl_Pointset_Powerset (SCOP_CONTEXT (scop)));
1315 nbs = unify_scattering_dimensions (scop);
1316 scaldims = (int *) xmalloc (nbs * (sizeof (int)));
1317 cloog_program_set_nb_scattdims (prog, nbs);
1318 initialize_cloog_names (scop, prog);
1320 for (i = 0; VEC_iterate (poly_bb_p, SCOP_BBS (scop), i, pbb); i++)
1322 CloogStatement *stmt;
1325 /* Dead code elimination: when the domain of a PBB is empty,
1326 don't generate code for the PBB. */
1327 if (ppl_Pointset_Powerset_C_Polyhedron_is_empty (PBB_DOMAIN (pbb)))
1330 /* Build the new statement and its block. */
1331 stmt = cloog_statement_alloc (pbb_index (pbb));
1332 block = cloog_block_alloc (stmt, 0, NULL, pbb_dim_iter_domain (pbb));
1333 cloog_statement_set_usr (stmt, pbb);
1335 /* Build loop list. */
1337 CloogLoop *new_loop_list = cloog_loop_malloc ();
1338 cloog_loop_set_next (new_loop_list, loop_list);
1339 cloog_loop_set_domain
1341 new_Cloog_Domain_from_ppl_Pointset_Powerset (PBB_DOMAIN (pbb)));
1342 cloog_loop_set_block (new_loop_list, block);
1343 loop_list = new_loop_list;
1346 /* Build block list. */
1348 CloogBlockList *new_block_list = cloog_block_list_malloc ();
1350 cloog_block_list_set_next (new_block_list, block_list);
1351 cloog_block_list_set_block (new_block_list, block);
1352 block_list = new_block_list;
1355 /* Build scattering list. */
1357 /* XXX: Replace with cloog_domain_list_alloc(), when available. */
1358 CloogDomainList *new_scattering
1359 = (CloogDomainList *) xmalloc (sizeof (CloogDomainList));
1360 ppl_Polyhedron_t scat;
1363 scat = PBB_TRANSFORMED_SCATTERING (pbb);
1364 dom = new_Cloog_Domain_from_ppl_Polyhedron (scat);
1366 cloog_set_next_domain (new_scattering, scattering);
1367 cloog_set_domain (new_scattering, dom);
1368 scattering = new_scattering;
1372 cloog_program_set_loop (prog, loop_list);
1373 cloog_program_set_blocklist (prog, block_list);
1375 for (i = 0; i < nbs; i++)
1378 cloog_program_set_scaldims (prog, scaldims);
1380 /* Extract scalar dimensions to simplify the code generation problem. */
1381 cloog_program_extract_scalars (prog, scattering);
1383 /* Apply scattering. */
1384 cloog_program_scatter (prog, scattering);
1385 free_scattering (scattering);
1387 /* Iterators corresponding to scalar dimensions have to be extracted. */
1388 cloog_names_scalarize (cloog_program_names (prog), nbs,
1389 cloog_program_scaldims (prog));
1391 /* Free blocklist. */
1393 CloogBlockList *next = cloog_program_blocklist (prog);
1397 CloogBlockList *toDelete = next;
1398 next = cloog_block_list_next (next);
1399 cloog_block_list_set_next (toDelete, NULL);
1400 cloog_block_list_set_block (toDelete, NULL);
1401 cloog_block_list_free (toDelete);
1403 cloog_program_set_blocklist (prog, NULL);
1407 /* Return the options that will be used in GLOOG. */
1409 static CloogOptions *
1410 set_cloog_options (void)
1412 CloogOptions *options = cloog_options_malloc ();
1414 /* Change cloog output language to C. If we do use FORTRAN instead, cloog
1415 will stop e.g. with "ERROR: unbounded loops not allowed in FORTRAN.", if
1416 we pass an incomplete program to cloog. */
1417 options->language = LANGUAGE_C;
1419 /* Enable complex equality spreading: removes dummy statements
1420 (assignments) in the generated code which repeats the
1421 substitution equations for statements. This is useless for
1425 /* Enable C pretty-printing mode: normalizes the substitution
1426 equations for statements. */
1429 /* Allow cloog to build strides with a stride width different to one.
1430 This example has stride = 4:
1432 for (i = 0; i < 20; i += 4)
1434 options->strides = 1;
1436 /* Disable optimizations and make cloog generate source code closer to the
1437 input. This is useful for debugging, but later we want the optimized
1440 XXX: We can not disable optimizations, as loop blocking is not working
1445 options->l = INT_MAX;
1451 /* Prints STMT to STDERR. */
1454 print_clast_stmt (FILE *file, struct clast_stmt *stmt)
1456 CloogOptions *options = set_cloog_options ();
1458 pprint (file, stmt, 0, options);
1459 cloog_options_free (options);
1462 /* Prints STMT to STDERR. */
1465 debug_clast_stmt (struct clast_stmt *stmt)
1467 print_clast_stmt (stderr, stmt);
1470 /* Translate SCOP to a CLooG program and clast. These two
1471 representations should be freed together: a clast cannot be used
1472 without a program. */
1475 scop_to_clast (scop_p scop)
1477 CloogOptions *options = set_cloog_options ();
1478 cloog_prog_clast pc;
1480 /* Connect new cloog prog generation to graphite. */
1481 pc.prog = cloog_program_malloc ();
1482 build_cloog_prog (scop, pc.prog);
1483 pc.prog = cloog_program_generate (pc.prog, options);
1484 pc.stmt = cloog_clast_create (pc.prog, options);
1486 cloog_options_free (options);
1490 /* Prints to FILE the code generated by CLooG for SCOP. */
1493 print_generated_program (FILE *file, scop_p scop)
1495 CloogOptions *options = set_cloog_options ();
1496 cloog_prog_clast pc = scop_to_clast (scop);
1498 fprintf (file, " (prog: \n");
1499 cloog_program_print (file, pc.prog);
1500 fprintf (file, " )\n");
1502 fprintf (file, " (clast: \n");
1503 pprint (file, pc.stmt, 0, options);
1504 fprintf (file, " )\n");
1506 cloog_options_free (options);
1507 cloog_clast_free (pc.stmt);
1508 cloog_program_free (pc.prog);
1511 /* Prints to STDERR the code generated by CLooG for SCOP. */
1514 debug_generated_program (scop_p scop)
1516 print_generated_program (stderr, scop);
1519 /* Add CLooG names to parameter index. The index is used to translate
1520 back from CLooG names to GCC trees. */
1523 create_params_index (htab_t index_table, CloogProgram *prog) {
1524 CloogNames* names = cloog_program_names (prog);
1525 int nb_parameters = cloog_names_nb_parameters (names);
1526 char **parameters = cloog_names_parameters (names);
1529 for (i = 0; i < nb_parameters; i++)
1530 save_clast_name_index (index_table, parameters[i], i);
1533 /* GIMPLE Loop Generator: generates loops from STMT in GIMPLE form for
1534 the given SCOP. Return true if code generation succeeded.
1535 BB_PBB_MAPPING is a basic_block and it's related poly_bb_p mapping.
1539 gloog (scop_p scop, VEC (scop_p, heap) *scops, htab_t bb_pbb_mapping)
1541 VEC (tree, heap) *newivs = VEC_alloc (tree, heap, 10);
1542 loop_p context_loop;
1543 sese region = SCOP_REGION (scop);
1544 ifsese if_region = NULL;
1545 htab_t rename_map, newivs_index, params_index;
1546 cloog_prog_clast pc;
1549 timevar_push (TV_GRAPHITE_CODE_GEN);
1550 gloog_error = false;
1552 pc = scop_to_clast (scop);
1554 if (dump_file && (dump_flags & TDF_DETAILS))
1556 fprintf (dump_file, "\nCLAST generated by CLooG: \n");
1557 print_clast_stmt (dump_file, pc.stmt);
1558 fprintf (dump_file, "\n");
1561 recompute_all_dominators ();
1564 if_region = move_sese_in_condition (region);
1565 sese_insert_phis_for_liveouts (region,
1566 if_region->region->exit->src,
1567 if_region->false_region->exit,
1568 if_region->true_region->exit);
1569 recompute_all_dominators ();
1572 context_loop = SESE_ENTRY (region)->src->loop_father;
1573 rename_map = htab_create (10, rename_map_elt_info, eq_rename_map_elts, free);
1574 newivs_index = htab_create (10, clast_name_index_elt_info,
1575 eq_clast_name_indexes, free);
1576 params_index = htab_create (10, clast_name_index_elt_info,
1577 eq_clast_name_indexes, free);
1579 create_params_index (params_index, pc.prog);
1581 translate_clast (region, context_loop, pc.stmt,
1582 if_region->true_region->entry,
1583 rename_map, &newivs, newivs_index,
1584 bb_pbb_mapping, 1, params_index);
1586 sese_adjust_liveout_phis (region, rename_map,
1587 if_region->region->exit->src,
1588 if_region->false_region->exit,
1589 if_region->true_region->exit);
1591 rename_nb_iterations (rename_map);
1593 for (i = 0; VEC_iterate (scop_p, scops, i, scop); i++)
1594 rename_sese_parameters (rename_map, SCOP_REGION (scop));
1596 recompute_all_dominators ();
1600 set_ifsese_condition (if_region, integer_zero_node);
1602 free (if_region->true_region);
1603 free (if_region->region);
1606 htab_delete (rename_map);
1607 htab_delete (newivs_index);
1608 htab_delete (params_index);
1609 VEC_free (tree, heap, newivs);
1610 cloog_clast_free (pc.stmt);
1611 cloog_program_free (pc.prog);
1612 timevar_pop (TV_GRAPHITE_CODE_GEN);
1614 if (dump_file && (dump_flags & TDF_DETAILS))
1618 int num_no_dependency = 0;
1620 FOR_EACH_LOOP (li, loop, 0)
1621 if (loop->can_be_parallel)
1622 num_no_dependency++;
1624 fprintf (dump_file, "\n%d loops carried no dependency.\n",
1628 return !gloog_error;