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"
24 #include "diagnostic-core.h"
25 #include "tree-flow.h"
26 #include "tree-dump.h"
28 #include "tree-chrec.h"
29 #include "tree-data-ref.h"
30 #include "tree-scalar-evolution.h"
34 #include "cloog/cloog.h"
36 #include "graphite-cloog-util.h"
37 #include "graphite-ppl.h"
38 #include "graphite-poly.h"
39 #include "graphite-clast-to-gimple.h"
40 #include "graphite-dependences.h"
41 #include "graphite-cloog-compat.h"
43 /* This flag is set when an error occurred during the translation of
45 static bool gloog_error
;
47 /* Verifies properties that GRAPHITE should maintain during translation. */
50 graphite_verify (void)
52 #ifdef ENABLE_CHECKING
53 verify_loop_structure ();
54 verify_dominators (CDI_DOMINATORS
);
55 verify_loop_closed_ssa (true);
59 /* Stores the INDEX in a vector for a given clast NAME. */
61 typedef struct clast_name_index
{
64 } *clast_name_index_p
;
66 /* Returns a pointer to a new element of type clast_name_index_p built
67 from NAME and INDEX. */
69 static inline clast_name_index_p
70 new_clast_name_index (const char *name
, int index
)
72 clast_name_index_p res
= XNEW (struct clast_name_index
);
79 /* For a given clast NAME, returns -1 if it does not correspond to any
80 parameter, or otherwise, returns the index in the PARAMS or
81 SCATTERING_DIMENSIONS vector. */
84 clast_name_to_index (clast_name_p name
, htab_t index_table
)
86 struct clast_name_index tmp
;
90 gcc_assert (name
->type
== clast_expr_name
);
91 tmp
.name
= ((const struct clast_name
*) name
)->name
;
96 slot
= htab_find_slot (index_table
, &tmp
, NO_INSERT
);
99 return ((struct clast_name_index
*) *slot
)->index
;
104 /* Records in INDEX_TABLE the INDEX for NAME. */
107 save_clast_name_index (htab_t index_table
, const char *name
, int index
)
109 struct clast_name_index tmp
;
113 slot
= htab_find_slot (index_table
, &tmp
, INSERT
);
119 *slot
= new_clast_name_index (name
, index
);
123 /* Computes a hash function for database element ELT. */
125 static inline hashval_t
126 clast_name_index_elt_info (const void *elt
)
128 return htab_hash_pointer (((const struct clast_name_index
*) elt
)->name
);
131 /* Compares database elements E1 and E2. */
134 eq_clast_name_indexes (const void *e1
, const void *e2
)
136 const struct clast_name_index
*elt1
= (const struct clast_name_index
*) e1
;
137 const struct clast_name_index
*elt2
= (const struct clast_name_index
*) e2
;
139 return (elt1
->name
== elt2
->name
);
142 /* For a given scattering dimension, return the new induction variable
146 newivs_to_depth_to_newiv (VEC (tree
, heap
) *newivs
, int depth
)
148 return VEC_index (tree
, newivs
, depth
);
153 /* Returns the tree variable from the name NAME that was given in
154 Cloog representation. */
157 clast_name_to_gcc (clast_name_p name
, sese region
, VEC (tree
, heap
) *newivs
,
158 htab_t newivs_index
, htab_t params_index
)
161 VEC (tree
, heap
) *params
= SESE_PARAMS (region
);
163 if (params
&& params_index
)
165 index
= clast_name_to_index (name
, params_index
);
168 return VEC_index (tree
, params
, index
);
171 gcc_assert (newivs
&& newivs_index
);
172 index
= clast_name_to_index (name
, newivs_index
);
173 gcc_assert (index
>= 0);
175 return newivs_to_depth_to_newiv (newivs
, index
);
178 /* Returns the signed maximal precision type for expressions TYPE1 and TYPE2. */
181 max_signed_precision_type (tree type1
, tree type2
)
183 int p1
= TYPE_PRECISION (type1
);
184 int p2
= TYPE_PRECISION (type2
);
187 enum machine_mode mode
;
190 precision
= TYPE_UNSIGNED (type1
) ? p1
* 2 : p1
;
192 precision
= TYPE_UNSIGNED (type2
) ? p2
* 2 : p2
;
194 if (precision
> BITS_PER_WORD
)
197 return integer_type_node
;
200 mode
= smallest_mode_for_size (precision
, MODE_INT
);
201 precision
= GET_MODE_PRECISION (mode
);
202 type
= build_nonstandard_integer_type (precision
, false);
207 return integer_type_node
;
213 /* Returns the maximal precision type for expressions TYPE1 and TYPE2. */
216 max_precision_type (tree type1
, tree type2
)
218 if (POINTER_TYPE_P (type1
))
221 if (POINTER_TYPE_P (type2
))
224 if (!TYPE_UNSIGNED (type1
)
225 || !TYPE_UNSIGNED (type2
))
226 return max_signed_precision_type (type1
, type2
);
228 return TYPE_PRECISION (type1
) > TYPE_PRECISION (type2
) ? type1
: type2
;
232 clast_to_gcc_expression (tree
, struct clast_expr
*, sese
, VEC (tree
, heap
) *,
235 /* Converts a Cloog reduction expression R with reduction operation OP
236 to a GCC expression tree of type TYPE. */
239 clast_to_gcc_expression_red (tree type
, enum tree_code op
,
240 struct clast_reduction
*r
,
241 sese region
, VEC (tree
, heap
) *newivs
,
242 htab_t newivs_index
, htab_t params_index
)
245 tree res
= clast_to_gcc_expression (type
, r
->elts
[0], region
, newivs
,
246 newivs_index
, params_index
);
247 tree operand_type
= (op
== POINTER_PLUS_EXPR
) ? sizetype
: type
;
249 for (i
= 1; i
< r
->n
; i
++)
251 tree t
= clast_to_gcc_expression (operand_type
, r
->elts
[i
], region
,
252 newivs
, newivs_index
, params_index
);
253 res
= fold_build2 (op
, type
, res
, t
);
259 /* Converts a Cloog AST expression E back to a GCC expression tree of
263 clast_to_gcc_expression (tree type
, struct clast_expr
*e
,
264 sese region
, VEC (tree
, heap
) *newivs
,
265 htab_t newivs_index
, htab_t params_index
)
269 case clast_expr_term
:
271 struct clast_term
*t
= (struct clast_term
*) e
;
275 if (mpz_cmp_si (t
->val
, 1) == 0)
277 tree name
= clast_name_to_gcc (t
->var
, region
, newivs
,
278 newivs_index
, params_index
);
280 if (POINTER_TYPE_P (TREE_TYPE (name
)) != POINTER_TYPE_P (type
))
281 name
= fold_convert (sizetype
, name
);
283 name
= fold_convert (type
, name
);
287 else if (mpz_cmp_si (t
->val
, -1) == 0)
289 tree name
= clast_name_to_gcc (t
->var
, region
, newivs
,
290 newivs_index
, params_index
);
292 if (POINTER_TYPE_P (TREE_TYPE (name
)) != POINTER_TYPE_P (type
))
293 name
= fold_convert (sizetype
, name
);
295 name
= fold_convert (type
, name
);
297 return fold_build1 (NEGATE_EXPR
, type
, name
);
301 tree name
= clast_name_to_gcc (t
->var
, region
, newivs
,
302 newivs_index
, params_index
);
303 tree cst
= gmp_cst_to_tree (type
, t
->val
);
305 if (POINTER_TYPE_P (TREE_TYPE (name
)) != POINTER_TYPE_P (type
))
306 name
= fold_convert (sizetype
, name
);
308 name
= fold_convert (type
, name
);
310 if (!POINTER_TYPE_P (type
))
311 return fold_build2 (MULT_EXPR
, type
, cst
, name
);
318 return gmp_cst_to_tree (type
, t
->val
);
323 struct clast_reduction
*r
= (struct clast_reduction
*) e
;
328 return clast_to_gcc_expression_red
329 (type
, POINTER_TYPE_P (type
) ? POINTER_PLUS_EXPR
: PLUS_EXPR
,
330 r
, region
, newivs
, newivs_index
, params_index
);
333 return clast_to_gcc_expression_red (type
, MIN_EXPR
, r
, region
,
334 newivs
, newivs_index
,
338 return clast_to_gcc_expression_red (type
, MAX_EXPR
, r
, region
,
339 newivs
, newivs_index
,
350 struct clast_binary
*b
= (struct clast_binary
*) e
;
351 struct clast_expr
*lhs
= (struct clast_expr
*) b
->LHS
;
352 tree tl
= clast_to_gcc_expression (type
, lhs
, region
, newivs
,
353 newivs_index
, params_index
);
354 tree tr
= gmp_cst_to_tree (type
, b
->RHS
);
359 return fold_build2 (FLOOR_DIV_EXPR
, type
, tl
, tr
);
362 return fold_build2 (CEIL_DIV_EXPR
, type
, tl
, tr
);
365 return fold_build2 (EXACT_DIV_EXPR
, type
, tl
, tr
);
368 return fold_build2 (TRUNC_MOD_EXPR
, type
, tl
, tr
);
382 /* Return the precision needed to represent the value VAL. */
385 precision_for_value (mpz_t val
)
401 while (mpz_cmp (y
, x
) >= 0)
414 /* Return the precision needed to represent the values between LOW and
418 precision_for_interval (mpz_t low
, mpz_t up
)
423 gcc_assert (mpz_cmp (low
, up
) <= 0);
426 mpz_sub (diff
, up
, low
);
427 precision
= precision_for_value (diff
);
433 /* Return a type that could represent the integer value VAL. */
436 gcc_type_for_interval (mpz_t low
, mpz_t up
)
438 bool unsigned_p
= true;
439 int precision
, prec_up
, prec_int
;
441 enum machine_mode mode
;
443 gcc_assert (mpz_cmp (low
, up
) <= 0);
445 prec_up
= precision_for_value (up
);
446 prec_int
= precision_for_interval (low
, up
);
447 precision
= MAX (prec_up
, prec_int
);
449 if (precision
> BITS_PER_WORD
)
452 return integer_type_node
;
455 if (mpz_sgn (low
) <= 0)
458 else if (precision
< BITS_PER_WORD
)
464 mode
= smallest_mode_for_size (precision
, MODE_INT
);
465 precision
= GET_MODE_PRECISION (mode
);
466 type
= build_nonstandard_integer_type (precision
, unsigned_p
);
471 return integer_type_node
;
477 /* Return a type that could represent the integer value VAL, or
478 otherwise return NULL_TREE. */
481 gcc_type_for_value (mpz_t val
)
483 return gcc_type_for_interval (val
, val
);
486 /* Return the type for the clast_term T used in STMT. */
489 gcc_type_for_clast_term (struct clast_term
*t
,
490 sese region
, VEC (tree
, heap
) *newivs
,
491 htab_t newivs_index
, htab_t params_index
)
493 gcc_assert (t
->expr
.type
== clast_expr_term
);
496 return gcc_type_for_value (t
->val
);
498 return TREE_TYPE (clast_name_to_gcc (t
->var
, region
, newivs
,
499 newivs_index
, params_index
));
503 gcc_type_for_clast_expr (struct clast_expr
*, sese
,
504 VEC (tree
, heap
) *, htab_t
, htab_t
);
506 /* Return the type for the clast_reduction R used in STMT. */
509 gcc_type_for_clast_red (struct clast_reduction
*r
, sese region
,
510 VEC (tree
, heap
) *newivs
,
511 htab_t newivs_index
, htab_t params_index
)
514 tree type
= NULL_TREE
;
517 return gcc_type_for_clast_expr (r
->elts
[0], region
, newivs
,
518 newivs_index
, params_index
);
525 type
= gcc_type_for_clast_expr (r
->elts
[0], region
, newivs
,
526 newivs_index
, params_index
);
527 for (i
= 1; i
< r
->n
; i
++)
528 type
= max_precision_type (type
, gcc_type_for_clast_expr
529 (r
->elts
[i
], region
, newivs
,
530 newivs_index
, params_index
));
542 /* Return the type for the clast_binary B used in STMT. */
545 gcc_type_for_clast_bin (struct clast_binary
*b
,
546 sese region
, VEC (tree
, heap
) *newivs
,
547 htab_t newivs_index
, htab_t params_index
)
549 tree l
= gcc_type_for_clast_expr ((struct clast_expr
*) b
->LHS
, region
,
550 newivs
, newivs_index
, params_index
);
551 tree r
= gcc_type_for_value (b
->RHS
);
552 return max_signed_precision_type (l
, r
);
555 /* Returns the type for the CLAST expression E when used in statement
559 gcc_type_for_clast_expr (struct clast_expr
*e
,
560 sese region
, VEC (tree
, heap
) *newivs
,
561 htab_t newivs_index
, htab_t params_index
)
565 case clast_expr_term
:
566 return gcc_type_for_clast_term ((struct clast_term
*) e
, region
,
567 newivs
, newivs_index
, params_index
);
570 return gcc_type_for_clast_red ((struct clast_reduction
*) e
, region
,
571 newivs
, newivs_index
, params_index
);
574 return gcc_type_for_clast_bin ((struct clast_binary
*) e
, region
,
575 newivs
, newivs_index
, params_index
);
584 /* Returns the type for the equation CLEQ. */
587 gcc_type_for_clast_eq (struct clast_equation
*cleq
,
588 sese region
, VEC (tree
, heap
) *newivs
,
589 htab_t newivs_index
, htab_t params_index
)
591 tree l
= gcc_type_for_clast_expr (cleq
->LHS
, region
, newivs
,
592 newivs_index
, params_index
);
593 tree r
= gcc_type_for_clast_expr (cleq
->RHS
, region
, newivs
,
594 newivs_index
, params_index
);
595 return max_precision_type (l
, r
);
598 /* Translates a clast equation CLEQ to a tree. */
601 graphite_translate_clast_equation (sese region
,
602 struct clast_equation
*cleq
,
603 VEC (tree
, heap
) *newivs
,
604 htab_t newivs_index
, htab_t params_index
)
607 tree type
= gcc_type_for_clast_eq (cleq
, region
, newivs
, newivs_index
,
609 tree lhs
= clast_to_gcc_expression (type
, cleq
->LHS
, region
, newivs
,
610 newivs_index
, params_index
);
611 tree rhs
= clast_to_gcc_expression (type
, cleq
->RHS
, region
, newivs
,
612 newivs_index
, params_index
);
617 else if (cleq
->sign
> 0)
623 return fold_build2 (comp
, boolean_type_node
, lhs
, rhs
);
626 /* Creates the test for the condition in STMT. */
629 graphite_create_guard_cond_expr (sese region
, struct clast_guard
*stmt
,
630 VEC (tree
, heap
) *newivs
,
631 htab_t newivs_index
, htab_t params_index
)
636 for (i
= 0; i
< stmt
->n
; i
++)
638 tree eq
= graphite_translate_clast_equation (region
, &stmt
->eq
[i
],
639 newivs
, newivs_index
,
643 cond
= fold_build2 (TRUTH_AND_EXPR
, TREE_TYPE (eq
), cond
, eq
);
651 /* Creates a new if region corresponding to Cloog's guard. */
654 graphite_create_new_guard (sese region
, edge entry_edge
,
655 struct clast_guard
*stmt
,
656 VEC (tree
, heap
) *newivs
,
657 htab_t newivs_index
, htab_t params_index
)
659 tree cond_expr
= graphite_create_guard_cond_expr (region
, stmt
, newivs
,
660 newivs_index
, params_index
);
661 edge exit_edge
= create_empty_if_region_on_edge (entry_edge
, cond_expr
);
665 /* Compute the lower bound LOW and upper bound UP for the induction
666 variable at LEVEL for the statement PBB, based on the transformed
667 scattering of PBB: T|I|G|Cst, with T the scattering transform, I
668 the iteration domain, and G the context parameters. */
671 compute_bounds_for_level (poly_bb_p pbb
, int level
, mpz_t low
, mpz_t up
)
673 ppl_Pointset_Powerset_C_Polyhedron_t ps
;
674 ppl_Linear_Expression_t le
;
676 combine_context_id_scat (&ps
, pbb
, false);
678 /* Prepare the linear expression corresponding to the level that we
679 want to maximize/minimize. */
681 ppl_dimension_type dim
= pbb_nb_scattering_transform (pbb
)
682 + pbb_dim_iter_domain (pbb
) + pbb_nb_params (pbb
);
684 ppl_new_Linear_Expression_with_dimension (&le
, dim
);
685 ppl_set_coef (le
, 2 * level
+ 1, 1);
688 ppl_max_for_le_pointset (ps
, le
, up
);
689 ppl_min_for_le_pointset (ps
, le
, low
);
690 ppl_delete_Linear_Expression (le
);
691 ppl_delete_Pointset_Powerset_C_Polyhedron (ps
);
694 /* Compute the type for the induction variable at LEVEL for the
695 statement PBB, based on the transformed schedule of PBB. */
698 compute_type_for_level (poly_bb_p pbb
, int level
)
706 compute_bounds_for_level (pbb
, level
, low
, up
);
707 type
= gcc_type_for_interval (low
, up
);
714 /* Walks a CLAST and returns the first statement in the body of a
717 static struct clast_user_stmt
*
718 clast_get_body_of_loop (struct clast_stmt
*stmt
)
721 || CLAST_STMT_IS_A (stmt
, stmt_user
))
722 return (struct clast_user_stmt
*) stmt
;
724 if (CLAST_STMT_IS_A (stmt
, stmt_for
))
725 return clast_get_body_of_loop (((struct clast_for
*) stmt
)->body
);
727 if (CLAST_STMT_IS_A (stmt
, stmt_guard
))
728 return clast_get_body_of_loop (((struct clast_guard
*) stmt
)->then
);
730 if (CLAST_STMT_IS_A (stmt
, stmt_block
))
731 return clast_get_body_of_loop (((struct clast_block
*) stmt
)->body
);
736 /* Returns the type for the induction variable for the loop translated
740 gcc_type_for_iv_of_clast_loop (struct clast_for
*stmt_for
, int level
,
741 tree lb_type
, tree ub_type
)
743 struct clast_stmt
*stmt
= (struct clast_stmt
*) stmt_for
;
744 struct clast_user_stmt
*body
= clast_get_body_of_loop (stmt
);
745 CloogStatement
*cs
= body
->statement
;
746 poly_bb_p pbb
= (poly_bb_p
) cloog_statement_usr (cs
);
748 return max_signed_precision_type (lb_type
, max_precision_type
749 (ub_type
, compute_type_for_level
753 /* Creates a new LOOP corresponding to Cloog's STMT. Inserts an
754 induction variable for the new LOOP. New LOOP is attached to CFG
755 starting at ENTRY_EDGE. LOOP is inserted into the loop tree and
756 becomes the child loop of the OUTER_LOOP. NEWIVS_INDEX binds
757 CLooG's scattering name to the induction variable created for the
758 loop of STMT. The new induction variable is inserted in the NEWIVS
762 graphite_create_new_loop (sese region
, edge entry_edge
,
763 struct clast_for
*stmt
,
764 loop_p outer
, VEC (tree
, heap
) **newivs
,
765 htab_t newivs_index
, htab_t params_index
, int level
)
767 tree lb_type
= gcc_type_for_clast_expr (stmt
->LB
, region
, *newivs
,
768 newivs_index
, params_index
);
769 tree ub_type
= gcc_type_for_clast_expr (stmt
->UB
, region
, *newivs
,
770 newivs_index
, params_index
);
771 tree type
= gcc_type_for_iv_of_clast_loop (stmt
, level
, lb_type
, ub_type
);
772 tree lb
= clast_to_gcc_expression (type
, stmt
->LB
, region
, *newivs
,
773 newivs_index
, params_index
);
774 tree ub
= clast_to_gcc_expression (type
, stmt
->UB
, region
, *newivs
,
775 newivs_index
, params_index
);
776 tree stride
= gmp_cst_to_tree (type
, stmt
->stride
);
777 tree ivvar
= create_tmp_var (type
, "graphite_IV");
778 tree iv
, iv_after_increment
;
779 loop_p loop
= create_empty_loop_on_edge
780 (entry_edge
, lb
, stride
, ub
, ivvar
, &iv
, &iv_after_increment
,
781 outer
? outer
: entry_edge
->src
->loop_father
);
783 add_referenced_var (ivvar
);
785 save_clast_name_index (newivs_index
, stmt
->iterator
,
786 VEC_length (tree
, *newivs
));
787 VEC_safe_push (tree
, heap
, *newivs
, iv
);
791 /* Inserts in iv_map a tuple (OLD_LOOP->num, NEW_NAME) for the
792 induction variables of the loops around GBB in SESE. */
795 build_iv_mapping (VEC (tree
, heap
) *iv_map
, sese region
,
796 VEC (tree
, heap
) *newivs
, htab_t newivs_index
,
797 struct clast_user_stmt
*user_stmt
,
800 struct clast_stmt
*t
;
802 CloogStatement
*cs
= user_stmt
->statement
;
803 poly_bb_p pbb
= (poly_bb_p
) cloog_statement_usr (cs
);
804 gimple_bb_p gbb
= PBB_BLACK_BOX (pbb
);
806 for (t
= user_stmt
->substitutions
; t
; t
= t
->next
, depth
++)
808 struct clast_expr
*expr
= (struct clast_expr
*)
809 ((struct clast_assignment
*)t
)->RHS
;
810 tree type
= gcc_type_for_clast_expr (expr
, region
, newivs
,
811 newivs_index
, params_index
);
812 tree new_name
= clast_to_gcc_expression (type
, expr
, region
, newivs
,
813 newivs_index
, params_index
);
814 loop_p old_loop
= gbb_loop_at_index (gbb
, region
, depth
);
816 VEC_replace (tree
, iv_map
, old_loop
->num
, new_name
);
820 /* Construct bb_pbb_def with BB and PBB. */
823 new_bb_pbb_def (basic_block bb
, poly_bb_p pbb
)
825 bb_pbb_def
*bb_pbb_p
;
827 bb_pbb_p
= XNEW (bb_pbb_def
);
834 /* Mark BB with it's relevant PBB via hashing table BB_PBB_MAPPING. */
837 mark_bb_with_pbb (poly_bb_p pbb
, basic_block bb
, htab_t bb_pbb_mapping
)
843 x
= htab_find_slot (bb_pbb_mapping
, &tmp
, INSERT
);
846 *x
= new_bb_pbb_def (bb
, pbb
);
849 /* Find BB's related poly_bb_p in hash table BB_PBB_MAPPING. */
852 find_pbb_via_hash (htab_t bb_pbb_mapping
, basic_block bb
)
858 slot
= htab_find_slot (bb_pbb_mapping
, &tmp
, NO_INSERT
);
861 return ((bb_pbb_def
*) *slot
)->pbb
;
866 /* Check data dependency in LOOP at scattering level LEVEL.
867 BB_PBB_MAPPING is a basic_block and it's related poly_bb_p
871 dependency_in_loop_p (loop_p loop
, htab_t bb_pbb_mapping
, int level
)
874 basic_block
*bbs
= get_loop_body_in_dom_order (loop
);
876 for (i
= 0; i
< loop
->num_nodes
; i
++)
878 poly_bb_p pbb1
= find_pbb_via_hash (bb_pbb_mapping
, bbs
[i
]);
883 for (j
= 0; j
< loop
->num_nodes
; j
++)
885 poly_bb_p pbb2
= find_pbb_via_hash (bb_pbb_mapping
, bbs
[j
]);
890 if (dependency_between_pbbs_p (pbb1
, pbb2
, level
))
903 /* Translates a clast user statement STMT to gimple.
905 - REGION is the sese region we used to generate the scop.
906 - NEXT_E is the edge where new generated code should be attached.
907 - CONTEXT_LOOP is the loop in which the generated code will be placed
908 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping.
909 - PARAMS_INDEX connects the cloog parameters with the gimple parameters in
912 translate_clast_user (sese region
, struct clast_user_stmt
*stmt
, edge next_e
,
913 VEC (tree
, heap
) **newivs
,
914 htab_t newivs_index
, htab_t bb_pbb_mapping
,
919 poly_bb_p pbb
= (poly_bb_p
) cloog_statement_usr (stmt
->statement
);
920 gimple_bb_p gbb
= PBB_BLACK_BOX (pbb
);
921 VEC (tree
, heap
) *iv_map
;
923 if (GBB_BB (gbb
) == ENTRY_BLOCK_PTR
)
926 nb_loops
= number_of_loops ();
927 iv_map
= VEC_alloc (tree
, heap
, nb_loops
);
928 for (i
= 0; i
< nb_loops
; i
++)
929 VEC_quick_push (tree
, iv_map
, NULL_TREE
);
931 build_iv_mapping (iv_map
, region
, *newivs
, newivs_index
, stmt
, params_index
);
932 next_e
= copy_bb_and_scalar_dependences (GBB_BB (gbb
), region
,
934 VEC_free (tree
, heap
, iv_map
);
936 new_bb
= next_e
->src
;
937 mark_bb_with_pbb (pbb
, new_bb
, bb_pbb_mapping
);
938 update_ssa (TODO_update_ssa
);
943 /* Creates a new if region protecting the loop to be executed, if the execution
944 count is zero (lb > ub). */
947 graphite_create_new_loop_guard (sese region
, edge entry_edge
,
948 struct clast_for
*stmt
,
949 VEC (tree
, heap
) *newivs
,
950 htab_t newivs_index
, htab_t params_index
)
954 tree lb_type
= gcc_type_for_clast_expr (stmt
->LB
, region
, newivs
,
955 newivs_index
, params_index
);
956 tree ub_type
= gcc_type_for_clast_expr (stmt
->UB
, region
, newivs
,
957 newivs_index
, params_index
);
958 tree type
= max_precision_type (lb_type
, ub_type
);
959 tree lb
= clast_to_gcc_expression (type
, stmt
->LB
, region
, newivs
,
960 newivs_index
, params_index
);
961 tree ub
= clast_to_gcc_expression (type
, stmt
->UB
, region
, newivs
,
962 newivs_index
, params_index
);
963 /* When ub is simply a constant or a parameter, use lb <= ub. */
964 if (TREE_CODE (ub
) == INTEGER_CST
|| TREE_CODE (ub
) == SSA_NAME
)
965 cond_expr
= fold_build2 (LE_EXPR
, boolean_type_node
, lb
, ub
);
968 tree one
= (POINTER_TYPE_P (type
)
970 : fold_convert (type
, integer_one_node
));
971 /* Adding +1 and using LT_EXPR helps with loop latches that have a
972 loop iteration count of "PARAMETER - 1". For PARAMETER == 0 this becomes
973 2^k-1 due to integer overflow, and the condition lb <= ub is true,
974 even if we do not want this. However lb < ub + 1 is false, as
976 tree ub_one
= fold_build2 (POINTER_TYPE_P (type
) ? POINTER_PLUS_EXPR
977 : PLUS_EXPR
, type
, ub
, one
);
979 cond_expr
= fold_build2 (LT_EXPR
, boolean_type_node
, lb
, ub_one
);
982 exit_edge
= create_empty_if_region_on_edge (entry_edge
, cond_expr
);
988 translate_clast (sese
, loop_p
, struct clast_stmt
*, edge
,
989 VEC (tree
, heap
) **, htab_t
, htab_t
, int, htab_t
);
991 /* Create the loop for a clast for statement.
993 - REGION is the sese region we used to generate the scop.
994 - NEXT_E is the edge where new generated code should be attached.
995 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping.
996 - PARAMS_INDEX connects the cloog parameters with the gimple parameters in
999 translate_clast_for_loop (sese region
, loop_p context_loop
,
1000 struct clast_for
*stmt
, edge next_e
,
1001 VEC (tree
, heap
) **newivs
,
1002 htab_t newivs_index
, htab_t bb_pbb_mapping
,
1003 int level
, htab_t params_index
)
1005 struct loop
*loop
= graphite_create_new_loop (region
, next_e
, stmt
,
1006 context_loop
, newivs
,
1007 newivs_index
, params_index
,
1009 edge last_e
= single_exit (loop
);
1010 edge to_body
= single_succ_edge (loop
->header
);
1011 basic_block after
= to_body
->dest
;
1013 /* Create a basic block for loop close phi nodes. */
1014 last_e
= single_succ_edge (split_edge (last_e
));
1016 /* Translate the body of the loop. */
1017 next_e
= translate_clast (region
, loop
, stmt
->body
, to_body
,
1018 newivs
, newivs_index
, bb_pbb_mapping
, level
+ 1,
1020 redirect_edge_succ_nodup (next_e
, after
);
1021 set_immediate_dominator (CDI_DOMINATORS
, next_e
->dest
, next_e
->src
);
1023 if (flag_loop_parallelize_all
1024 && !dependency_in_loop_p (loop
, bb_pbb_mapping
,
1025 get_scattering_level (level
)))
1026 loop
->can_be_parallel
= true;
1031 /* Translates a clast for statement STMT to gimple. First a guard is created
1032 protecting the loop, if it is executed zero times. In this guard we create
1033 the real loop structure.
1035 - REGION is the sese region we used to generate the scop.
1036 - NEXT_E is the edge where new generated code should be attached.
1037 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping.
1038 - PARAMS_INDEX connects the cloog parameters with the gimple parameters in
1041 translate_clast_for (sese region
, loop_p context_loop
, struct clast_for
*stmt
,
1042 edge next_e
, VEC (tree
, heap
) **newivs
,
1043 htab_t newivs_index
, htab_t bb_pbb_mapping
, int level
,
1044 htab_t params_index
)
1046 edge last_e
= graphite_create_new_loop_guard (region
, next_e
, stmt
, *newivs
,
1047 newivs_index
, params_index
);
1048 edge true_e
= get_true_edge_from_guard_bb (next_e
->dest
);
1050 translate_clast_for_loop (region
, context_loop
, stmt
, true_e
, newivs
,
1051 newivs_index
, bb_pbb_mapping
, level
,
1056 /* Translates a clast guard statement STMT to gimple.
1058 - REGION is the sese region we used to generate the scop.
1059 - NEXT_E is the edge where new generated code should be attached.
1060 - CONTEXT_LOOP is the loop in which the generated code will be placed
1061 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping.
1062 - PARAMS_INDEX connects the cloog parameters with the gimple parameters in
1065 translate_clast_guard (sese region
, loop_p context_loop
,
1066 struct clast_guard
*stmt
, edge next_e
,
1067 VEC (tree
, heap
) **newivs
,
1068 htab_t newivs_index
, htab_t bb_pbb_mapping
, int level
,
1069 htab_t params_index
)
1071 edge last_e
= graphite_create_new_guard (region
, next_e
, stmt
, *newivs
,
1072 newivs_index
, params_index
);
1073 edge true_e
= get_true_edge_from_guard_bb (next_e
->dest
);
1075 translate_clast (region
, context_loop
, stmt
->then
, true_e
,
1076 newivs
, newivs_index
, bb_pbb_mapping
,
1077 level
, params_index
);
1081 /* Translates a CLAST statement STMT to GCC representation in the
1084 - NEXT_E is the edge where new generated code should be attached.
1085 - CONTEXT_LOOP is the loop in which the generated code will be placed
1086 - BB_PBB_MAPPING is is a basic_block and it's related poly_bb_p mapping. */
1088 translate_clast (sese region
, loop_p context_loop
, struct clast_stmt
*stmt
,
1089 edge next_e
, VEC (tree
, heap
) **newivs
,
1090 htab_t newivs_index
, htab_t bb_pbb_mapping
, int level
,
1091 htab_t params_index
)
1096 if (CLAST_STMT_IS_A (stmt
, stmt_root
))
1099 else if (CLAST_STMT_IS_A (stmt
, stmt_user
))
1100 next_e
= translate_clast_user (region
, (struct clast_user_stmt
*) stmt
,
1101 next_e
, newivs
, newivs_index
,
1102 bb_pbb_mapping
, params_index
);
1104 else if (CLAST_STMT_IS_A (stmt
, stmt_for
))
1105 next_e
= translate_clast_for (region
, context_loop
,
1106 (struct clast_for
*) stmt
, next_e
,
1107 newivs
, newivs_index
,
1108 bb_pbb_mapping
, level
, params_index
);
1110 else if (CLAST_STMT_IS_A (stmt
, stmt_guard
))
1111 next_e
= translate_clast_guard (region
, context_loop
,
1112 (struct clast_guard
*) stmt
, next_e
,
1113 newivs
, newivs_index
,
1114 bb_pbb_mapping
, level
, params_index
);
1116 else if (CLAST_STMT_IS_A (stmt
, stmt_block
))
1117 next_e
= translate_clast (region
, context_loop
,
1118 ((struct clast_block
*) stmt
)->body
,
1119 next_e
, newivs
, newivs_index
,
1120 bb_pbb_mapping
, level
, params_index
);
1124 recompute_all_dominators ();
1127 return translate_clast (region
, context_loop
, stmt
->next
, next_e
,
1128 newivs
, newivs_index
,
1129 bb_pbb_mapping
, level
, params_index
);
1132 /* Free the SCATTERING domain list. */
1135 free_scattering (CloogScatteringList
*scattering
)
1139 CloogScattering
*dom
= cloog_scattering (scattering
);
1140 CloogScatteringList
*next
= cloog_next_scattering (scattering
);
1142 cloog_scattering_free (dom
);
1148 /* Initialize Cloog's parameter names from the names used in GIMPLE.
1149 Initialize Cloog's iterator names, using 'graphite_iterator_%d'
1150 from 0 to scop_nb_loops (scop). */
1153 initialize_cloog_names (scop_p scop
, CloogProgram
*prog
)
1155 sese region
= SCOP_REGION (scop
);
1157 int nb_iterators
= scop_max_loop_depth (scop
);
1158 int nb_scattering
= cloog_program_nb_scattdims (prog
);
1159 int nb_parameters
= VEC_length (tree
, SESE_PARAMS (region
));
1160 char **iterators
= XNEWVEC (char *, nb_iterators
* 2);
1161 char **scattering
= XNEWVEC (char *, nb_scattering
);
1162 char **parameters
= XNEWVEC (char *, nb_parameters
);
1164 cloog_program_set_names (prog
, cloog_names_malloc ());
1166 for (i
= 0; i
< nb_parameters
; i
++)
1168 tree param
= VEC_index (tree
, SESE_PARAMS(region
), i
);
1169 const char *name
= get_name (param
);
1175 len
= strlen (name
);
1177 parameters
[i
] = XNEWVEC (char, len
+ 1);
1178 snprintf (parameters
[i
], len
, "%s_%d", name
, SSA_NAME_VERSION (param
));
1181 cloog_names_set_nb_parameters (cloog_program_names (prog
), nb_parameters
);
1182 cloog_names_set_parameters (cloog_program_names (prog
), parameters
);
1184 for (i
= 0; i
< nb_iterators
; i
++)
1187 iterators
[i
] = XNEWVEC (char, len
);
1188 snprintf (iterators
[i
], len
, "git_%d", i
);
1191 cloog_names_set_nb_iterators (cloog_program_names (prog
),
1193 cloog_names_set_iterators (cloog_program_names (prog
),
1196 for (i
= 0; i
< nb_scattering
; i
++)
1199 scattering
[i
] = XNEWVEC (char, len
);
1200 snprintf (scattering
[i
], len
, "scat_%d", i
);
1203 cloog_names_set_nb_scattering (cloog_program_names (prog
),
1205 cloog_names_set_scattering (cloog_program_names (prog
),
1209 /* Initialize a CLooG input file. */
1212 init_cloog_input_file (int scop_number
)
1214 FILE *graphite_out_file
;
1215 int len
= strlen (dump_base_name
);
1216 char *dumpname
= XNEWVEC (char, len
+ 25);
1217 char *s_scop_number
= XNEWVEC (char, 15);
1219 memcpy (dumpname
, dump_base_name
, len
+ 1);
1220 strip_off_ending (dumpname
, len
);
1221 sprintf (s_scop_number
, ".%d", scop_number
);
1222 strcat (dumpname
, s_scop_number
);
1223 strcat (dumpname
, ".cloog");
1224 graphite_out_file
= fopen (dumpname
, "w+b");
1226 if (graphite_out_file
== 0)
1227 fatal_error ("can%'t open %s for writing: %m", dumpname
);
1231 return graphite_out_file
;
1234 /* Build cloog program for SCoP. */
1237 build_cloog_prog (scop_p scop
, CloogProgram
*prog
,
1238 CloogOptions
*options
)
1241 int max_nb_loops
= scop_max_loop_depth (scop
);
1243 CloogLoop
*loop_list
= NULL
;
1244 CloogBlockList
*block_list
= NULL
;
1245 CloogScatteringList
*scattering
= NULL
;
1246 int nbs
= 2 * max_nb_loops
+ 1;
1249 cloog_program_set_context
1250 (prog
, new_Cloog_Domain_from_ppl_Pointset_Powerset (SCOP_CONTEXT (scop
),
1251 scop_nb_params (scop
), cloog_state
));
1252 nbs
= unify_scattering_dimensions (scop
);
1253 scaldims
= (int *) xmalloc (nbs
* (sizeof (int)));
1254 cloog_program_set_nb_scattdims (prog
, nbs
);
1255 initialize_cloog_names (scop
, prog
);
1257 FOR_EACH_VEC_ELT (poly_bb_p
, SCOP_BBS (scop
), i
, pbb
)
1259 CloogStatement
*stmt
;
1263 /* Dead code elimination: when the domain of a PBB is empty,
1264 don't generate code for the PBB. */
1265 if (ppl_Pointset_Powerset_C_Polyhedron_is_empty (PBB_DOMAIN (pbb
)))
1268 /* Build the new statement and its block. */
1269 stmt
= cloog_statement_alloc (cloog_state
, pbb_index (pbb
));
1270 dom
= new_Cloog_Domain_from_ppl_Pointset_Powerset (PBB_DOMAIN (pbb
),
1271 scop_nb_params (scop
),
1273 block
= cloog_block_alloc (stmt
, 0, NULL
, pbb_dim_iter_domain (pbb
));
1274 cloog_statement_set_usr (stmt
, pbb
);
1276 /* Build loop list. */
1278 CloogLoop
*new_loop_list
= cloog_loop_malloc (cloog_state
);
1279 cloog_loop_set_next (new_loop_list
, loop_list
);
1280 cloog_loop_set_domain (new_loop_list
, dom
);
1281 cloog_loop_set_block (new_loop_list
, block
);
1282 loop_list
= new_loop_list
;
1285 /* Build block list. */
1287 CloogBlockList
*new_block_list
= cloog_block_list_malloc ();
1289 cloog_block_list_set_next (new_block_list
, block_list
);
1290 cloog_block_list_set_block (new_block_list
, block
);
1291 block_list
= new_block_list
;
1294 /* Build scattering list. */
1296 /* XXX: Replace with cloog_domain_list_alloc(), when available. */
1297 CloogScatteringList
*new_scattering
1298 = (CloogScatteringList
*) xmalloc (sizeof (CloogScatteringList
));
1299 ppl_Polyhedron_t scat
;
1300 CloogScattering
*dom
;
1302 scat
= PBB_TRANSFORMED_SCATTERING (pbb
);
1303 dom
= new_Cloog_Scattering_from_ppl_Polyhedron
1304 (scat
, scop_nb_params (scop
), pbb_nb_scattering_transform (pbb
),
1307 cloog_set_next_scattering (new_scattering
, scattering
);
1308 cloog_set_scattering (new_scattering
, dom
);
1309 scattering
= new_scattering
;
1313 cloog_program_set_loop (prog
, loop_list
);
1314 cloog_program_set_blocklist (prog
, block_list
);
1316 for (i
= 0; i
< nbs
; i
++)
1319 cloog_program_set_scaldims (prog
, scaldims
);
1321 /* Extract scalar dimensions to simplify the code generation problem. */
1322 cloog_program_extract_scalars (prog
, scattering
, options
);
1324 /* Dump a .cloog input file, if requested. This feature is only
1325 enabled in the Graphite branch. */
1328 static size_t file_scop_number
= 0;
1329 FILE *cloog_file
= init_cloog_input_file (file_scop_number
);
1331 cloog_program_dump_cloog (cloog_file
, prog
, scattering
);
1335 /* Apply scattering. */
1336 cloog_program_scatter (prog
, scattering
, options
);
1337 free_scattering (scattering
);
1339 /* Iterators corresponding to scalar dimensions have to be extracted. */
1340 cloog_names_scalarize (cloog_program_names (prog
), nbs
,
1341 cloog_program_scaldims (prog
));
1343 /* Free blocklist. */
1345 CloogBlockList
*next
= cloog_program_blocklist (prog
);
1349 CloogBlockList
*toDelete
= next
;
1350 next
= cloog_block_list_next (next
);
1351 cloog_block_list_set_next (toDelete
, NULL
);
1352 cloog_block_list_set_block (toDelete
, NULL
);
1353 cloog_block_list_free (toDelete
);
1355 cloog_program_set_blocklist (prog
, NULL
);
1359 /* Return the options that will be used in GLOOG. */
1361 static CloogOptions
*
1362 set_cloog_options (void)
1364 CloogOptions
*options
= cloog_options_malloc (cloog_state
);
1366 /* Change cloog output language to C. If we do use FORTRAN instead, cloog
1367 will stop e.g. with "ERROR: unbounded loops not allowed in FORTRAN.", if
1368 we pass an incomplete program to cloog. */
1369 options
->language
= LANGUAGE_C
;
1371 /* Enable complex equality spreading: removes dummy statements
1372 (assignments) in the generated code which repeats the
1373 substitution equations for statements. This is useless for
1378 /* Silence CLooG to avoid failing tests due to debug output to stderr. */
1381 /* Enable C pretty-printing mode: normalizes the substitution
1382 equations for statements. */
1386 /* Allow cloog to build strides with a stride width different to one.
1387 This example has stride = 4:
1389 for (i = 0; i < 20; i += 4)
1391 options
->strides
= 1;
1393 /* Disable optimizations and make cloog generate source code closer to the
1394 input. This is useful for debugging, but later we want the optimized
1397 XXX: We can not disable optimizations, as loop blocking is not working
1402 options
->l
= INT_MAX
;
1408 /* Prints STMT to STDERR. */
1411 print_clast_stmt (FILE *file
, struct clast_stmt
*stmt
)
1413 CloogOptions
*options
= set_cloog_options ();
1415 clast_pprint (file
, stmt
, 0, options
);
1416 cloog_options_free (options
);
1419 /* Prints STMT to STDERR. */
1422 debug_clast_stmt (struct clast_stmt
*stmt
)
1424 print_clast_stmt (stderr
, stmt
);
1427 /* Translate SCOP to a CLooG program and clast. These two
1428 representations should be freed together: a clast cannot be used
1429 without a program. */
1432 scop_to_clast (scop_p scop
)
1434 CloogOptions
*options
= set_cloog_options ();
1435 cloog_prog_clast pc
;
1437 /* Connect new cloog prog generation to graphite. */
1438 pc
.prog
= cloog_program_malloc ();
1439 build_cloog_prog (scop
, pc
.prog
, options
);
1440 pc
.prog
= cloog_program_generate (pc
.prog
, options
);
1441 pc
.stmt
= cloog_clast_create (pc
.prog
, options
);
1443 cloog_options_free (options
);
1447 /* Prints to FILE the code generated by CLooG for SCOP. */
1450 print_generated_program (FILE *file
, scop_p scop
)
1452 CloogOptions
*options
= set_cloog_options ();
1454 cloog_prog_clast pc
= scop_to_clast (scop
);
1456 fprintf (file
, " (prog: \n");
1457 cloog_program_print (file
, pc
.prog
);
1458 fprintf (file
, " )\n");
1460 fprintf (file
, " (clast: \n");
1461 clast_pprint (file
, pc
.stmt
, 0, options
);
1462 fprintf (file
, " )\n");
1464 cloog_options_free (options
);
1465 cloog_clast_free (pc
.stmt
);
1466 cloog_program_free (pc
.prog
);
1469 /* Prints to STDERR the code generated by CLooG for SCOP. */
1472 debug_generated_program (scop_p scop
)
1474 print_generated_program (stderr
, scop
);
1477 /* Add CLooG names to parameter index. The index is used to translate
1478 back from CLooG names to GCC trees. */
1481 create_params_index (htab_t index_table
, CloogProgram
*prog
) {
1482 CloogNames
* names
= cloog_program_names (prog
);
1483 int nb_parameters
= cloog_names_nb_parameters (names
);
1484 char **parameters
= cloog_names_parameters (names
);
1487 for (i
= 0; i
< nb_parameters
; i
++)
1488 save_clast_name_index (index_table
, parameters
[i
], i
);
1491 /* GIMPLE Loop Generator: generates loops from STMT in GIMPLE form for
1492 the given SCOP. Return true if code generation succeeded.
1493 BB_PBB_MAPPING is a basic_block and it's related poly_bb_p mapping.
1497 gloog (scop_p scop
, htab_t bb_pbb_mapping
)
1499 VEC (tree
, heap
) *newivs
= VEC_alloc (tree
, heap
, 10);
1500 loop_p context_loop
;
1501 sese region
= SCOP_REGION (scop
);
1502 ifsese if_region
= NULL
;
1503 htab_t newivs_index
, params_index
;
1504 cloog_prog_clast pc
;
1506 timevar_push (TV_GRAPHITE_CODE_GEN
);
1507 gloog_error
= false;
1509 pc
= scop_to_clast (scop
);
1511 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1513 fprintf (dump_file
, "\nCLAST generated by CLooG: \n");
1514 print_clast_stmt (dump_file
, pc
.stmt
);
1515 fprintf (dump_file
, "\n");
1518 recompute_all_dominators ();
1521 if_region
= move_sese_in_condition (region
);
1522 sese_insert_phis_for_liveouts (region
,
1523 if_region
->region
->exit
->src
,
1524 if_region
->false_region
->exit
,
1525 if_region
->true_region
->exit
);
1526 recompute_all_dominators ();
1529 context_loop
= SESE_ENTRY (region
)->src
->loop_father
;
1530 newivs_index
= htab_create (10, clast_name_index_elt_info
,
1531 eq_clast_name_indexes
, free
);
1532 params_index
= htab_create (10, clast_name_index_elt_info
,
1533 eq_clast_name_indexes
, free
);
1535 create_params_index (params_index
, pc
.prog
);
1537 translate_clast (region
, context_loop
, pc
.stmt
,
1538 if_region
->true_region
->entry
,
1539 &newivs
, newivs_index
,
1540 bb_pbb_mapping
, 1, params_index
);
1543 recompute_all_dominators ();
1547 set_ifsese_condition (if_region
, integer_zero_node
);
1549 free (if_region
->true_region
);
1550 free (if_region
->region
);
1553 htab_delete (newivs_index
);
1554 htab_delete (params_index
);
1555 VEC_free (tree
, heap
, newivs
);
1556 cloog_clast_free (pc
.stmt
);
1557 cloog_program_free (pc
.prog
);
1558 timevar_pop (TV_GRAPHITE_CODE_GEN
);
1560 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1564 int num_no_dependency
= 0;
1566 FOR_EACH_LOOP (li
, loop
, 0)
1567 if (loop
->can_be_parallel
)
1568 num_no_dependency
++;
1570 fprintf (dump_file
, "\n%d loops carried no dependency.\n",
1574 return !gloog_error
;