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-cloog-util.h"
50 #include "graphite-ppl.h"
52 #include "graphite-poly.h"
53 #include "graphite-scop-detection.h"
54 #include "graphite-clast-to-gimple.h"
55 #include "graphite-dependences.h"
56 #include "graphite-cloog-compat.h"
58 /* This flag is set when an error occurred during the translation of
60 static bool gloog_error
;
62 /* Verifies properties that GRAPHITE should maintain during translation. */
65 graphite_verify (void)
67 #ifdef ENABLE_CHECKING
68 verify_loop_structure ();
69 verify_dominators (CDI_DOMINATORS
);
70 verify_dominators (CDI_POST_DOMINATORS
);
71 verify_loop_closed_ssa (true);
75 /* Stores the INDEX in a vector for a given clast NAME. */
77 typedef struct clast_name_index
{
80 } *clast_name_index_p
;
82 /* Returns a pointer to a new element of type clast_name_index_p built
83 from NAME and INDEX. */
85 static inline clast_name_index_p
86 new_clast_name_index (const char *name
, int index
)
88 clast_name_index_p res
= XNEW (struct clast_name_index
);
95 /* For a given clast NAME, returns -1 if it does not correspond to any
96 parameter, or otherwise, returns the index in the PARAMS or
97 SCATTERING_DIMENSIONS vector. */
100 clast_name_to_index (clast_name_p name
, htab_t index_table
)
102 struct clast_name_index tmp
;
106 gcc_assert (name
->type
== clast_expr_name
);
107 tmp
.name
= ((const struct clast_name
*) name
)->name
;
112 slot
= htab_find_slot (index_table
, &tmp
, NO_INSERT
);
115 return ((struct clast_name_index
*) *slot
)->index
;
120 /* Records in INDEX_TABLE the INDEX for NAME. */
123 save_clast_name_index (htab_t index_table
, const char *name
, int index
)
125 struct clast_name_index tmp
;
129 slot
= htab_find_slot (index_table
, &tmp
, INSERT
);
136 *slot
= new_clast_name_index (name
, index
);
140 /* Computes a hash function for database element ELT. */
142 static inline hashval_t
143 clast_name_index_elt_info (const void *elt
)
145 return htab_hash_pointer (((const struct clast_name_index
*) elt
)->name
);
148 /* Compares database elements E1 and E2. */
151 eq_clast_name_indexes (const void *e1
, const void *e2
)
153 const struct clast_name_index
*elt1
= (const struct clast_name_index
*) e1
;
154 const struct clast_name_index
*elt2
= (const struct clast_name_index
*) e2
;
156 return (elt1
->name
== elt2
->name
);
159 /* For a given scattering dimension, return the new induction variable
163 newivs_to_depth_to_newiv (VEC (tree
, heap
) *newivs
, int depth
)
165 return VEC_index (tree
, newivs
, depth
);
170 /* Returns the tree variable from the name NAME that was given in
171 Cloog representation. */
174 clast_name_to_gcc (clast_name_p name
, sese region
, VEC (tree
, heap
) *newivs
,
175 htab_t newivs_index
, htab_t params_index
)
178 VEC (tree
, heap
) *params
= SESE_PARAMS (region
);
180 if (params
&& params_index
)
182 index
= clast_name_to_index (name
, params_index
);
185 return VEC_index (tree
, params
, index
);
188 gcc_assert (newivs
&& newivs_index
);
189 index
= clast_name_to_index (name
, newivs_index
);
190 gcc_assert (index
>= 0);
192 return newivs_to_depth_to_newiv (newivs
, index
);
195 /* Returns the signed maximal precision type for expressions TYPE1 and TYPE2. */
198 max_signed_precision_type (tree type1
, tree type2
)
200 int p1
= TYPE_PRECISION (type1
);
201 int p2
= TYPE_PRECISION (type2
);
204 enum machine_mode mode
;
207 precision
= TYPE_UNSIGNED (type1
) ? p1
* 2 : p1
;
209 precision
= TYPE_UNSIGNED (type2
) ? p2
* 2 : p2
;
211 if (precision
> BITS_PER_WORD
)
214 return integer_type_node
;
217 mode
= smallest_mode_for_size (precision
, MODE_INT
);
218 precision
= GET_MODE_PRECISION (mode
);
219 type
= build_nonstandard_integer_type (precision
, false);
224 return integer_type_node
;
230 /* Returns the maximal precision type for expressions TYPE1 and TYPE2. */
233 max_precision_type (tree type1
, tree type2
)
235 if (POINTER_TYPE_P (type1
))
238 if (POINTER_TYPE_P (type2
))
241 if (!TYPE_UNSIGNED (type1
)
242 || !TYPE_UNSIGNED (type2
))
243 return max_signed_precision_type (type1
, type2
);
245 return TYPE_PRECISION (type1
) > TYPE_PRECISION (type2
) ? type1
: type2
;
249 clast_to_gcc_expression (tree
, struct clast_expr
*, sese
, VEC (tree
, heap
) *,
252 /* Converts a Cloog reduction expression R with reduction operation OP
253 to a GCC expression tree of type TYPE. */
256 clast_to_gcc_expression_red (tree type
, enum tree_code op
,
257 struct clast_reduction
*r
,
258 sese region
, VEC (tree
, heap
) *newivs
,
259 htab_t newivs_index
, htab_t params_index
)
262 tree res
= clast_to_gcc_expression (type
, r
->elts
[0], region
, newivs
,
263 newivs_index
, params_index
);
264 tree operand_type
= (op
== POINTER_PLUS_EXPR
) ? sizetype
: type
;
266 for (i
= 1; i
< r
->n
; i
++)
268 tree t
= clast_to_gcc_expression (operand_type
, r
->elts
[i
], region
,
269 newivs
, newivs_index
, params_index
);
270 res
= fold_build2 (op
, type
, res
, t
);
276 /* Converts a Cloog AST expression E back to a GCC expression tree of
280 clast_to_gcc_expression (tree type
, struct clast_expr
*e
,
281 sese region
, VEC (tree
, heap
) *newivs
,
282 htab_t newivs_index
, htab_t params_index
)
286 case clast_expr_term
:
288 struct clast_term
*t
= (struct clast_term
*) e
;
292 if (mpz_cmp_si (t
->val
, 1) == 0)
294 tree name
= clast_name_to_gcc (t
->var
, region
, newivs
,
295 newivs_index
, params_index
);
297 if (POINTER_TYPE_P (TREE_TYPE (name
)) != POINTER_TYPE_P (type
))
298 name
= fold_convert (sizetype
, name
);
300 name
= fold_convert (type
, name
);
304 else if (mpz_cmp_si (t
->val
, -1) == 0)
306 tree name
= clast_name_to_gcc (t
->var
, region
, newivs
,
307 newivs_index
, params_index
);
309 if (POINTER_TYPE_P (TREE_TYPE (name
)) != POINTER_TYPE_P (type
))
310 name
= fold_convert (sizetype
, name
);
312 name
= fold_convert (type
, name
);
314 return fold_build1 (NEGATE_EXPR
, type
, name
);
318 tree name
= clast_name_to_gcc (t
->var
, region
, newivs
,
319 newivs_index
, params_index
);
320 tree cst
= gmp_cst_to_tree (type
, t
->val
);
322 if (POINTER_TYPE_P (TREE_TYPE (name
)) != POINTER_TYPE_P (type
))
323 name
= fold_convert (sizetype
, name
);
325 name
= fold_convert (type
, name
);
327 if (!POINTER_TYPE_P (type
))
328 return fold_build2 (MULT_EXPR
, type
, cst
, name
);
335 return gmp_cst_to_tree (type
, t
->val
);
340 struct clast_reduction
*r
= (struct clast_reduction
*) e
;
345 return clast_to_gcc_expression_red
346 (type
, POINTER_TYPE_P (type
) ? POINTER_PLUS_EXPR
: PLUS_EXPR
,
347 r
, region
, newivs
, newivs_index
, params_index
);
350 return clast_to_gcc_expression_red (type
, MIN_EXPR
, r
, region
,
351 newivs
, newivs_index
,
355 return clast_to_gcc_expression_red (type
, MAX_EXPR
, r
, region
,
356 newivs
, newivs_index
,
367 struct clast_binary
*b
= (struct clast_binary
*) e
;
368 struct clast_expr
*lhs
= (struct clast_expr
*) b
->LHS
;
369 tree tl
= clast_to_gcc_expression (type
, lhs
, region
, newivs
,
370 newivs_index
, params_index
);
371 tree tr
= gmp_cst_to_tree (type
, b
->RHS
);
376 return fold_build2 (FLOOR_DIV_EXPR
, type
, tl
, tr
);
379 return fold_build2 (CEIL_DIV_EXPR
, type
, tl
, tr
);
382 return fold_build2 (EXACT_DIV_EXPR
, type
, tl
, tr
);
385 return fold_build2 (TRUNC_MOD_EXPR
, type
, tl
, tr
);
399 /* Return the precision needed to represent the value VAL. */
402 precision_for_value (mpz_t val
)
418 while (mpz_cmp (y
, x
) >= 0)
431 /* Return the precision needed to represent the values between LOW and
435 precision_for_interval (mpz_t low
, mpz_t up
)
440 gcc_assert (mpz_cmp (low
, up
) <= 0);
443 mpz_sub (diff
, up
, low
);
444 precision
= precision_for_value (diff
);
450 /* Return a type that could represent the integer value VAL. */
453 gcc_type_for_interval (mpz_t low
, mpz_t up
)
455 bool unsigned_p
= true;
456 int precision
, prec_up
, prec_int
;
458 enum machine_mode mode
;
460 gcc_assert (mpz_cmp (low
, up
) <= 0);
462 prec_up
= precision_for_value (up
);
463 prec_int
= precision_for_interval (low
, up
);
464 precision
= MAX (prec_up
, prec_int
);
466 if (precision
> BITS_PER_WORD
)
469 return integer_type_node
;
472 if (mpz_sgn (low
) <= 0)
475 else if (precision
< BITS_PER_WORD
)
481 mode
= smallest_mode_for_size (precision
, MODE_INT
);
482 precision
= GET_MODE_PRECISION (mode
);
483 type
= build_nonstandard_integer_type (precision
, unsigned_p
);
488 return integer_type_node
;
494 /* Return a type that could represent the integer value VAL, or
495 otherwise return NULL_TREE. */
498 gcc_type_for_value (mpz_t val
)
500 return gcc_type_for_interval (val
, val
);
503 /* Return the type for the clast_term T used in STMT. */
506 gcc_type_for_clast_term (struct clast_term
*t
,
507 sese region
, VEC (tree
, heap
) *newivs
,
508 htab_t newivs_index
, htab_t params_index
)
510 gcc_assert (t
->expr
.type
== clast_expr_term
);
513 return gcc_type_for_value (t
->val
);
515 return TREE_TYPE (clast_name_to_gcc (t
->var
, region
, newivs
,
516 newivs_index
, params_index
));
520 gcc_type_for_clast_expr (struct clast_expr
*, sese
,
521 VEC (tree
, heap
) *, htab_t
, htab_t
);
523 /* Return the type for the clast_reduction R used in STMT. */
526 gcc_type_for_clast_red (struct clast_reduction
*r
, sese region
,
527 VEC (tree
, heap
) *newivs
,
528 htab_t newivs_index
, htab_t params_index
)
531 tree type
= NULL_TREE
;
534 return gcc_type_for_clast_expr (r
->elts
[0], region
, newivs
,
535 newivs_index
, params_index
);
542 type
= gcc_type_for_clast_expr (r
->elts
[0], region
, newivs
,
543 newivs_index
, params_index
);
544 for (i
= 1; i
< r
->n
; i
++)
545 type
= max_precision_type (type
, gcc_type_for_clast_expr
546 (r
->elts
[i
], region
, newivs
,
547 newivs_index
, params_index
));
559 /* Return the type for the clast_binary B used in STMT. */
562 gcc_type_for_clast_bin (struct clast_binary
*b
,
563 sese region
, VEC (tree
, heap
) *newivs
,
564 htab_t newivs_index
, htab_t params_index
)
566 tree l
= gcc_type_for_clast_expr ((struct clast_expr
*) b
->LHS
, region
,
567 newivs
, newivs_index
, params_index
);
568 tree r
= gcc_type_for_value (b
->RHS
);
569 return max_signed_precision_type (l
, r
);
572 /* Returns the type for the CLAST expression E when used in statement
576 gcc_type_for_clast_expr (struct clast_expr
*e
,
577 sese region
, VEC (tree
, heap
) *newivs
,
578 htab_t newivs_index
, htab_t params_index
)
582 case clast_expr_term
:
583 return gcc_type_for_clast_term ((struct clast_term
*) e
, region
,
584 newivs
, newivs_index
, params_index
);
587 return gcc_type_for_clast_red ((struct clast_reduction
*) e
, region
,
588 newivs
, newivs_index
, params_index
);
591 return gcc_type_for_clast_bin ((struct clast_binary
*) e
, region
,
592 newivs
, newivs_index
, params_index
);
601 /* Returns the type for the equation CLEQ. */
604 gcc_type_for_clast_eq (struct clast_equation
*cleq
,
605 sese region
, VEC (tree
, heap
) *newivs
,
606 htab_t newivs_index
, htab_t params_index
)
608 tree l
= gcc_type_for_clast_expr (cleq
->LHS
, region
, newivs
,
609 newivs_index
, params_index
);
610 tree r
= gcc_type_for_clast_expr (cleq
->RHS
, region
, newivs
,
611 newivs_index
, params_index
);
612 return max_precision_type (l
, r
);
615 /* Translates a clast equation CLEQ to a tree. */
618 graphite_translate_clast_equation (sese region
,
619 struct clast_equation
*cleq
,
620 VEC (tree
, heap
) *newivs
,
621 htab_t newivs_index
, htab_t params_index
)
624 tree type
= gcc_type_for_clast_eq (cleq
, region
, newivs
, newivs_index
,
626 tree lhs
= clast_to_gcc_expression (type
, cleq
->LHS
, region
, newivs
,
627 newivs_index
, params_index
);
628 tree rhs
= clast_to_gcc_expression (type
, cleq
->RHS
, region
, newivs
,
629 newivs_index
, params_index
);
634 else if (cleq
->sign
> 0)
640 return fold_build2 (comp
, boolean_type_node
, lhs
, rhs
);
643 /* Creates the test for the condition in STMT. */
646 graphite_create_guard_cond_expr (sese region
, struct clast_guard
*stmt
,
647 VEC (tree
, heap
) *newivs
,
648 htab_t newivs_index
, htab_t params_index
)
653 for (i
= 0; i
< stmt
->n
; i
++)
655 tree eq
= graphite_translate_clast_equation (region
, &stmt
->eq
[i
],
656 newivs
, newivs_index
,
660 cond
= fold_build2 (TRUTH_AND_EXPR
, TREE_TYPE (eq
), cond
, eq
);
668 /* Creates a new if region corresponding to Cloog's guard. */
671 graphite_create_new_guard (sese region
, edge entry_edge
,
672 struct clast_guard
*stmt
,
673 VEC (tree
, heap
) *newivs
,
674 htab_t newivs_index
, htab_t params_index
)
676 tree cond_expr
= graphite_create_guard_cond_expr (region
, stmt
, newivs
,
677 newivs_index
, params_index
);
678 edge exit_edge
= create_empty_if_region_on_edge (entry_edge
, cond_expr
);
682 /* Compute the lower bound LOW and upper bound UP for the induction
683 variable at LEVEL for the statement PBB, based on the transformed
684 scattering of PBB: T|I|G|Cst, with T the scattering transform, I
685 the iteration domain, and G the context parameters. */
688 compute_bounds_for_level (poly_bb_p pbb
, int level
, mpz_t low
, mpz_t up
)
690 ppl_Pointset_Powerset_C_Polyhedron_t ps
;
691 ppl_Linear_Expression_t le
;
693 combine_context_id_scat (&ps
, pbb
, false);
695 /* Prepare the linear expression corresponding to the level that we
696 want to maximize/minimize. */
698 ppl_dimension_type dim
= pbb_nb_scattering_transform (pbb
)
699 + pbb_dim_iter_domain (pbb
) + pbb_nb_params (pbb
);
701 ppl_new_Linear_Expression_with_dimension (&le
, dim
);
702 ppl_set_coef (le
, 2 * level
+ 1, 1);
705 ppl_max_for_le_pointset (ps
, le
, up
);
706 ppl_min_for_le_pointset (ps
, le
, low
);
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 /* Build cloog program for SCoP. */
1223 build_cloog_prog (scop_p scop
, CloogProgram
*prog
,
1224 CloogOptions
*options
, CloogState
*state ATTRIBUTE_UNUSED
)
1227 int max_nb_loops
= scop_max_loop_depth (scop
);
1229 CloogLoop
*loop_list
= NULL
;
1230 CloogBlockList
*block_list
= NULL
;
1231 CloogScatteringList
*scattering
= NULL
;
1232 int nbs
= 2 * max_nb_loops
+ 1;
1235 cloog_program_set_context
1236 (prog
, new_Cloog_Domain_from_ppl_Pointset_Powerset (SCOP_CONTEXT (scop
),
1237 scop_nb_params (scop
), state
));
1238 nbs
= unify_scattering_dimensions (scop
);
1239 scaldims
= (int *) xmalloc (nbs
* (sizeof (int)));
1240 cloog_program_set_nb_scattdims (prog
, nbs
);
1241 initialize_cloog_names (scop
, prog
);
1243 FOR_EACH_VEC_ELT (poly_bb_p
, SCOP_BBS (scop
), i
, pbb
)
1245 CloogStatement
*stmt
;
1249 /* Dead code elimination: when the domain of a PBB is empty,
1250 don't generate code for the PBB. */
1251 if (ppl_Pointset_Powerset_C_Polyhedron_is_empty (PBB_DOMAIN (pbb
)))
1254 /* Build the new statement and its block. */
1255 stmt
= cloog_statement_alloc (state
, pbb_index (pbb
));
1256 dom
= new_Cloog_Domain_from_ppl_Pointset_Powerset (PBB_DOMAIN (pbb
),
1257 scop_nb_params (scop
),
1259 block
= cloog_block_alloc (stmt
, 0, NULL
, pbb_dim_iter_domain (pbb
));
1260 cloog_statement_set_usr (stmt
, pbb
);
1262 /* Build loop list. */
1264 CloogLoop
*new_loop_list
= cloog_loop_malloc (state
);
1265 cloog_loop_set_next (new_loop_list
, loop_list
);
1266 cloog_loop_set_domain (new_loop_list
, dom
);
1267 cloog_loop_set_block (new_loop_list
, block
);
1268 loop_list
= new_loop_list
;
1271 /* Build block list. */
1273 CloogBlockList
*new_block_list
= cloog_block_list_malloc ();
1275 cloog_block_list_set_next (new_block_list
, block_list
);
1276 cloog_block_list_set_block (new_block_list
, block
);
1277 block_list
= new_block_list
;
1280 /* Build scattering list. */
1282 /* XXX: Replace with cloog_domain_list_alloc(), when available. */
1283 CloogScatteringList
*new_scattering
1284 = (CloogScatteringList
*) xmalloc (sizeof (CloogScatteringList
));
1285 ppl_Polyhedron_t scat
;
1286 CloogScattering
*dom
;
1288 scat
= PBB_TRANSFORMED_SCATTERING (pbb
);
1289 dom
= new_Cloog_Scattering_from_ppl_Polyhedron
1290 (scat
, scop_nb_params (scop
), pbb_nb_scattering_transform (pbb
),
1293 cloog_set_next_scattering (new_scattering
, scattering
);
1294 cloog_set_scattering (new_scattering
, dom
);
1295 scattering
= new_scattering
;
1299 cloog_program_set_loop (prog
, loop_list
);
1300 cloog_program_set_blocklist (prog
, block_list
);
1302 for (i
= 0; i
< nbs
; i
++)
1305 cloog_program_set_scaldims (prog
, scaldims
);
1307 /* Extract scalar dimensions to simplify the code generation problem. */
1308 cloog_program_extract_scalars (prog
, scattering
, options
);
1310 /* Apply scattering. */
1311 cloog_program_scatter (prog
, scattering
, options
);
1312 free_scattering (scattering
);
1314 /* Iterators corresponding to scalar dimensions have to be extracted. */
1315 cloog_names_scalarize (cloog_program_names (prog
), nbs
,
1316 cloog_program_scaldims (prog
));
1318 /* Free blocklist. */
1320 CloogBlockList
*next
= cloog_program_blocklist (prog
);
1324 CloogBlockList
*toDelete
= next
;
1325 next
= cloog_block_list_next (next
);
1326 cloog_block_list_set_next (toDelete
, NULL
);
1327 cloog_block_list_set_block (toDelete
, NULL
);
1328 cloog_block_list_free (toDelete
);
1330 cloog_program_set_blocklist (prog
, NULL
);
1334 /* Return the options that will be used in GLOOG. */
1336 static CloogOptions
*
1337 set_cloog_options (CloogState
*state ATTRIBUTE_UNUSED
)
1339 CloogOptions
*options
= cloog_options_malloc (state
);
1341 /* Change cloog output language to C. If we do use FORTRAN instead, cloog
1342 will stop e.g. with "ERROR: unbounded loops not allowed in FORTRAN.", if
1343 we pass an incomplete program to cloog. */
1344 options
->language
= LANGUAGE_C
;
1346 /* Enable complex equality spreading: removes dummy statements
1347 (assignments) in the generated code which repeats the
1348 substitution equations for statements. This is useless for
1353 /* Silence CLooG to avoid failing tests due to debug output to stderr. */
1356 /* Enable C pretty-printing mode: normalizes the substitution
1357 equations for statements. */
1361 /* Allow cloog to build strides with a stride width different to one.
1362 This example has stride = 4:
1364 for (i = 0; i < 20; i += 4)
1366 options
->strides
= 1;
1368 /* Disable optimizations and make cloog generate source code closer to the
1369 input. This is useful for debugging, but later we want the optimized
1372 XXX: We can not disable optimizations, as loop blocking is not working
1377 options
->l
= INT_MAX
;
1383 /* Prints STMT to STDERR. */
1386 print_clast_stmt (FILE *file
, struct clast_stmt
*stmt
)
1388 CloogState
*state
= cloog_state_malloc ();
1389 CloogOptions
*options
= set_cloog_options (state
);
1391 clast_pprint (file
, stmt
, 0, options
);
1392 cloog_options_free (options
);
1393 cloog_state_free (state
);
1396 /* Prints STMT to STDERR. */
1399 debug_clast_stmt (struct clast_stmt
*stmt
)
1401 print_clast_stmt (stderr
, stmt
);
1404 /* Translate SCOP to a CLooG program and clast. These two
1405 representations should be freed together: a clast cannot be used
1406 without a program. */
1409 scop_to_clast (scop_p scop
, CloogState
*state
)
1411 CloogOptions
*options
= set_cloog_options (state
);
1412 cloog_prog_clast pc
;
1414 /* Connect new cloog prog generation to graphite. */
1415 pc
.prog
= cloog_program_malloc ();
1416 build_cloog_prog (scop
, pc
.prog
, options
, state
);
1417 pc
.prog
= cloog_program_generate (pc
.prog
, options
);
1418 pc
.stmt
= cloog_clast_create (pc
.prog
, options
);
1420 cloog_options_free (options
);
1424 /* Prints to FILE the code generated by CLooG for SCOP. */
1427 print_generated_program (FILE *file
, scop_p scop
)
1429 CloogState
*state
= cloog_state_malloc ();
1430 CloogOptions
*options
= set_cloog_options (state
);
1432 cloog_prog_clast pc
= scop_to_clast (scop
, state
);
1434 fprintf (file
, " (prog: \n");
1435 cloog_program_print (file
, pc
.prog
);
1436 fprintf (file
, " )\n");
1438 fprintf (file
, " (clast: \n");
1439 clast_pprint (file
, pc
.stmt
, 0, options
);
1440 fprintf (file
, " )\n");
1442 cloog_options_free (options
);
1443 cloog_clast_free (pc
.stmt
);
1444 cloog_program_free (pc
.prog
);
1447 /* Prints to STDERR the code generated by CLooG for SCOP. */
1450 debug_generated_program (scop_p scop
)
1452 print_generated_program (stderr
, scop
);
1455 /* Add CLooG names to parameter index. The index is used to translate
1456 back from CLooG names to GCC trees. */
1459 create_params_index (htab_t index_table
, CloogProgram
*prog
) {
1460 CloogNames
* names
= cloog_program_names (prog
);
1461 int nb_parameters
= cloog_names_nb_parameters (names
);
1462 char **parameters
= cloog_names_parameters (names
);
1465 for (i
= 0; i
< nb_parameters
; i
++)
1466 save_clast_name_index (index_table
, parameters
[i
], i
);
1469 /* GIMPLE Loop Generator: generates loops from STMT in GIMPLE form for
1470 the given SCOP. Return true if code generation succeeded.
1471 BB_PBB_MAPPING is a basic_block and it's related poly_bb_p mapping.
1475 gloog (scop_p scop
, htab_t bb_pbb_mapping
)
1477 VEC (tree
, heap
) *newivs
= VEC_alloc (tree
, heap
, 10);
1478 loop_p context_loop
;
1479 sese region
= SCOP_REGION (scop
);
1480 ifsese if_region
= NULL
;
1481 htab_t newivs_index
, params_index
;
1482 cloog_prog_clast pc
;
1485 state
= cloog_state_malloc ();
1486 timevar_push (TV_GRAPHITE_CODE_GEN
);
1487 gloog_error
= false;
1489 pc
= scop_to_clast (scop
, state
);
1491 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1493 fprintf (dump_file
, "\nCLAST generated by CLooG: \n");
1494 print_clast_stmt (dump_file
, pc
.stmt
);
1495 fprintf (dump_file
, "\n");
1498 recompute_all_dominators ();
1501 if_region
= move_sese_in_condition (region
);
1502 sese_insert_phis_for_liveouts (region
,
1503 if_region
->region
->exit
->src
,
1504 if_region
->false_region
->exit
,
1505 if_region
->true_region
->exit
);
1506 recompute_all_dominators ();
1509 context_loop
= SESE_ENTRY (region
)->src
->loop_father
;
1510 newivs_index
= htab_create (10, clast_name_index_elt_info
,
1511 eq_clast_name_indexes
, free
);
1512 params_index
= htab_create (10, clast_name_index_elt_info
,
1513 eq_clast_name_indexes
, free
);
1515 create_params_index (params_index
, pc
.prog
);
1517 translate_clast (region
, context_loop
, pc
.stmt
,
1518 if_region
->true_region
->entry
,
1519 &newivs
, newivs_index
,
1520 bb_pbb_mapping
, 1, params_index
);
1523 recompute_all_dominators ();
1527 set_ifsese_condition (if_region
, integer_zero_node
);
1529 free (if_region
->true_region
);
1530 free (if_region
->region
);
1533 htab_delete (newivs_index
);
1534 htab_delete (params_index
);
1535 VEC_free (tree
, heap
, newivs
);
1536 cloog_clast_free (pc
.stmt
);
1537 cloog_program_free (pc
.prog
);
1538 timevar_pop (TV_GRAPHITE_CODE_GEN
);
1540 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1544 int num_no_dependency
= 0;
1546 FOR_EACH_LOOP (li
, loop
, 0)
1547 if (loop
->can_be_parallel
)
1548 num_no_dependency
++;
1550 fprintf (dump_file
, "\n%d loops carried no dependency.\n",
1554 cloog_state_free (state
);
1556 return !gloog_error
;