gcc/
[official-gcc.git] / gcc / graphite-sese-to-poly.c
blob49856914782262a4b340209d4aa0be197b835f7f
1 /* Conversion of SESE regions to Polyhedra.
2 Copyright (C) 2009-2013 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)
10 any later version.
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/>. */
21 #include "config.h"
23 #ifdef HAVE_cloog
24 #include <isl/set.h>
25 #include <isl/map.h>
26 #include <isl/union_map.h>
27 #include <isl/constraint.h>
28 #include <isl/aff.h>
29 #include <cloog/cloog.h>
30 #include <cloog/cloog.h>
31 #include <cloog/isl/domain.h>
32 #endif
34 #include "system.h"
35 #include "coretypes.h"
36 #include "tree-ssa.h"
37 #include "tree-pass.h"
38 #include "cfgloop.h"
39 #include "tree-chrec.h"
40 #include "tree-data-ref.h"
41 #include "tree-scalar-evolution.h"
42 #include "domwalk.h"
43 #include "sese.h"
45 #ifdef HAVE_cloog
46 #include "graphite-poly.h"
47 #include "graphite-sese-to-poly.h"
50 /* Assigns to RES the value of the INTEGER_CST T. */
52 static inline void
53 tree_int_to_gmp (tree t, mpz_t res)
55 double_int di = tree_to_double_int (t);
56 mpz_set_double_int (res, di, TYPE_UNSIGNED (TREE_TYPE (t)));
59 /* Returns the index of the PHI argument defined in the outermost
60 loop. */
62 static size_t
63 phi_arg_in_outermost_loop (gimple phi)
65 loop_p loop = gimple_bb (phi)->loop_father;
66 size_t i, res = 0;
68 for (i = 0; i < gimple_phi_num_args (phi); i++)
69 if (!flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, i)->src))
71 loop = gimple_phi_arg_edge (phi, i)->src->loop_father;
72 res = i;
75 return res;
78 /* Removes a simple copy phi node "RES = phi (INIT, RES)" at position
79 PSI by inserting on the loop ENTRY edge assignment "RES = INIT". */
81 static void
82 remove_simple_copy_phi (gimple_stmt_iterator *psi)
84 gimple phi = gsi_stmt (*psi);
85 tree res = gimple_phi_result (phi);
86 size_t entry = phi_arg_in_outermost_loop (phi);
87 tree init = gimple_phi_arg_def (phi, entry);
88 gimple stmt = gimple_build_assign (res, init);
89 edge e = gimple_phi_arg_edge (phi, entry);
91 remove_phi_node (psi, false);
92 gsi_insert_on_edge_immediate (e, stmt);
93 SSA_NAME_DEF_STMT (res) = stmt;
96 /* Removes an invariant phi node at position PSI by inserting on the
97 loop ENTRY edge the assignment RES = INIT. */
99 static void
100 remove_invariant_phi (sese region, gimple_stmt_iterator *psi)
102 gimple phi = gsi_stmt (*psi);
103 loop_p loop = loop_containing_stmt (phi);
104 tree res = gimple_phi_result (phi);
105 tree scev = scalar_evolution_in_region (region, loop, res);
106 size_t entry = phi_arg_in_outermost_loop (phi);
107 edge e = gimple_phi_arg_edge (phi, entry);
108 tree var;
109 gimple stmt;
110 gimple_seq stmts = NULL;
112 if (tree_contains_chrecs (scev, NULL))
113 scev = gimple_phi_arg_def (phi, entry);
115 var = force_gimple_operand (scev, &stmts, true, NULL_TREE);
116 stmt = gimple_build_assign (res, var);
117 remove_phi_node (psi, false);
119 gimple_seq_add_stmt (&stmts, stmt);
120 gsi_insert_seq_on_edge (e, stmts);
121 gsi_commit_edge_inserts ();
122 SSA_NAME_DEF_STMT (res) = stmt;
125 /* Returns true when the phi node at PSI is of the form "a = phi (a, x)". */
127 static inline bool
128 simple_copy_phi_p (gimple phi)
130 tree res;
132 if (gimple_phi_num_args (phi) != 2)
133 return false;
135 res = gimple_phi_result (phi);
136 return (res == gimple_phi_arg_def (phi, 0)
137 || res == gimple_phi_arg_def (phi, 1));
140 /* Returns true when the phi node at position PSI is a reduction phi
141 node in REGION. Otherwise moves the pointer PSI to the next phi to
142 be considered. */
144 static bool
145 reduction_phi_p (sese region, gimple_stmt_iterator *psi)
147 loop_p loop;
148 gimple phi = gsi_stmt (*psi);
149 tree res = gimple_phi_result (phi);
151 loop = loop_containing_stmt (phi);
153 if (simple_copy_phi_p (phi))
155 /* PRE introduces phi nodes like these, for an example,
156 see id-5.f in the fortran graphite testsuite:
158 # prephitmp.85_265 = PHI <prephitmp.85_258(33), prephitmp.85_265(18)>
160 remove_simple_copy_phi (psi);
161 return false;
164 if (scev_analyzable_p (res, region))
166 tree scev = scalar_evolution_in_region (region, loop, res);
168 if (evolution_function_is_invariant_p (scev, loop->num))
169 remove_invariant_phi (region, psi);
170 else
171 gsi_next (psi);
173 return false;
176 /* All the other cases are considered reductions. */
177 return true;
180 /* Store the GRAPHITE representation of BB. */
182 static gimple_bb_p
183 new_gimple_bb (basic_block bb, vec<data_reference_p> drs)
185 struct gimple_bb *gbb;
187 gbb = XNEW (struct gimple_bb);
188 bb->aux = gbb;
189 GBB_BB (gbb) = bb;
190 GBB_DATA_REFS (gbb) = drs;
191 GBB_CONDITIONS (gbb).create (0);
192 GBB_CONDITION_CASES (gbb).create (0);
194 return gbb;
197 static void
198 free_data_refs_aux (vec<data_reference_p> datarefs)
200 unsigned int i;
201 struct data_reference *dr;
203 FOR_EACH_VEC_ELT (datarefs, i, dr)
204 if (dr->aux)
206 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
208 free (bap->alias_set);
210 free (bap);
211 dr->aux = NULL;
214 /* Frees GBB. */
216 static void
217 free_gimple_bb (struct gimple_bb *gbb)
219 free_data_refs_aux (GBB_DATA_REFS (gbb));
220 free_data_refs (GBB_DATA_REFS (gbb));
222 GBB_CONDITIONS (gbb).release ();
223 GBB_CONDITION_CASES (gbb).release ();
224 GBB_BB (gbb)->aux = 0;
225 XDELETE (gbb);
228 /* Deletes all gimple bbs in SCOP. */
230 static void
231 remove_gbbs_in_scop (scop_p scop)
233 int i;
234 poly_bb_p pbb;
236 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
237 free_gimple_bb (PBB_BLACK_BOX (pbb));
240 /* Deletes all scops in SCOPS. */
242 void
243 free_scops (vec<scop_p> scops)
245 int i;
246 scop_p scop;
248 FOR_EACH_VEC_ELT (scops, i, scop)
250 remove_gbbs_in_scop (scop);
251 free_sese (SCOP_REGION (scop));
252 free_scop (scop);
255 scops.release ();
258 /* Same as outermost_loop_in_sese, returns the outermost loop
259 containing BB in REGION, but makes sure that the returned loop
260 belongs to the REGION, and so this returns the first loop in the
261 REGION when the loop containing BB does not belong to REGION. */
263 static loop_p
264 outermost_loop_in_sese_1 (sese region, basic_block bb)
266 loop_p nest = outermost_loop_in_sese (region, bb);
268 if (loop_in_sese_p (nest, region))
269 return nest;
271 /* When the basic block BB does not belong to a loop in the region,
272 return the first loop in the region. */
273 nest = nest->inner;
274 while (nest)
275 if (loop_in_sese_p (nest, region))
276 break;
277 else
278 nest = nest->next;
280 gcc_assert (nest);
281 return nest;
284 /* Generates a polyhedral black box only if the bb contains interesting
285 information. */
287 static gimple_bb_p
288 try_generate_gimple_bb (scop_p scop, basic_block bb)
290 vec<data_reference_p> drs;
291 drs.create (5);
292 sese region = SCOP_REGION (scop);
293 loop_p nest = outermost_loop_in_sese_1 (region, bb);
294 gimple_stmt_iterator gsi;
296 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
298 gimple stmt = gsi_stmt (gsi);
299 loop_p loop;
301 if (is_gimple_debug (stmt))
302 continue;
304 loop = loop_containing_stmt (stmt);
305 if (!loop_in_sese_p (loop, region))
306 loop = nest;
308 graphite_find_data_references_in_stmt (nest, loop, stmt, &drs);
311 return new_gimple_bb (bb, drs);
314 /* Returns true if all predecessors of BB, that are not dominated by BB, are
315 marked in MAP. The predecessors dominated by BB are loop latches and will
316 be handled after BB. */
318 static bool
319 all_non_dominated_preds_marked_p (basic_block bb, sbitmap map)
321 edge e;
322 edge_iterator ei;
324 FOR_EACH_EDGE (e, ei, bb->preds)
325 if (!bitmap_bit_p (map, e->src->index)
326 && !dominated_by_p (CDI_DOMINATORS, e->src, bb))
327 return false;
329 return true;
332 /* Compare the depth of two basic_block's P1 and P2. */
334 static int
335 compare_bb_depths (const void *p1, const void *p2)
337 const_basic_block const bb1 = *(const_basic_block const*)p1;
338 const_basic_block const bb2 = *(const_basic_block const*)p2;
339 int d1 = loop_depth (bb1->loop_father);
340 int d2 = loop_depth (bb2->loop_father);
342 if (d1 < d2)
343 return 1;
345 if (d1 > d2)
346 return -1;
348 return 0;
351 /* Sort the basic blocks from DOM such that the first are the ones at
352 a deepest loop level. */
354 static void
355 graphite_sort_dominated_info (vec<basic_block> dom)
357 dom.qsort (compare_bb_depths);
360 /* Recursive helper function for build_scops_bbs. */
362 static void
363 build_scop_bbs_1 (scop_p scop, sbitmap visited, basic_block bb)
365 sese region = SCOP_REGION (scop);
366 vec<basic_block> dom;
367 poly_bb_p pbb;
369 if (bitmap_bit_p (visited, bb->index)
370 || !bb_in_sese_p (bb, region))
371 return;
373 pbb = new_poly_bb (scop, try_generate_gimple_bb (scop, bb));
374 SCOP_BBS (scop).safe_push (pbb);
375 bitmap_set_bit (visited, bb->index);
377 dom = get_dominated_by (CDI_DOMINATORS, bb);
379 if (!dom.exists ())
380 return;
382 graphite_sort_dominated_info (dom);
384 while (!dom.is_empty ())
386 int i;
387 basic_block dom_bb;
389 FOR_EACH_VEC_ELT (dom, i, dom_bb)
390 if (all_non_dominated_preds_marked_p (dom_bb, visited))
392 build_scop_bbs_1 (scop, visited, dom_bb);
393 dom.unordered_remove (i);
394 break;
398 dom.release ();
401 /* Gather the basic blocks belonging to the SCOP. */
403 static void
404 build_scop_bbs (scop_p scop)
406 sbitmap visited = sbitmap_alloc (last_basic_block);
407 sese region = SCOP_REGION (scop);
409 bitmap_clear (visited);
410 build_scop_bbs_1 (scop, visited, SESE_ENTRY_BB (region));
411 sbitmap_free (visited);
414 /* Return an ISL identifier for the polyhedral basic block PBB. */
416 static isl_id *
417 isl_id_for_pbb (scop_p s, poly_bb_p pbb)
419 char name[50];
420 snprintf (name, sizeof (name), "S_%d", pbb_index (pbb));
421 return isl_id_alloc (s->ctx, name, pbb);
424 /* Converts the STATIC_SCHEDULE of PBB into a scattering polyhedron.
425 We generate SCATTERING_DIMENSIONS scattering dimensions.
427 CLooG 0.15.0 and previous versions require, that all
428 scattering functions of one CloogProgram have the same number of
429 scattering dimensions, therefore we allow to specify it. This
430 should be removed in future versions of CLooG.
432 The scattering polyhedron consists of these dimensions: scattering,
433 loop_iterators, parameters.
435 Example:
437 | scattering_dimensions = 5
438 | used_scattering_dimensions = 3
439 | nb_iterators = 1
440 | scop_nb_params = 2
442 | Schedule:
444 | 4 5
446 | Scattering polyhedron:
448 | scattering: {s1, s2, s3, s4, s5}
449 | loop_iterators: {i}
450 | parameters: {p1, p2}
452 | s1 s2 s3 s4 s5 i p1 p2 1
453 | 1 0 0 0 0 0 0 0 -4 = 0
454 | 0 1 0 0 0 -1 0 0 0 = 0
455 | 0 0 1 0 0 0 0 0 -5 = 0 */
457 static void
458 build_pbb_scattering_polyhedrons (isl_aff *static_sched,
459 poly_bb_p pbb, int scattering_dimensions)
461 int i;
462 int nb_iterators = pbb_dim_iter_domain (pbb);
463 int used_scattering_dimensions = nb_iterators * 2 + 1;
464 isl_int val;
465 isl_space *dc, *dm;
467 gcc_assert (scattering_dimensions >= used_scattering_dimensions);
469 isl_int_init (val);
471 dc = isl_set_get_space (pbb->domain);
472 dm = isl_space_add_dims (isl_space_from_domain (dc),
473 isl_dim_out, scattering_dimensions);
474 pbb->schedule = isl_map_universe (dm);
476 for (i = 0; i < scattering_dimensions; i++)
478 /* Textual order inside this loop. */
479 if ((i % 2) == 0)
481 isl_constraint *c = isl_equality_alloc
482 (isl_local_space_from_space (isl_map_get_space (pbb->schedule)));
484 if (0 != isl_aff_get_coefficient (static_sched, isl_dim_in,
485 i / 2, &val))
486 gcc_unreachable ();
488 isl_int_neg (val, val);
489 c = isl_constraint_set_constant (c, val);
490 c = isl_constraint_set_coefficient_si (c, isl_dim_out, i, 1);
491 pbb->schedule = isl_map_add_constraint (pbb->schedule, c);
494 /* Iterations of this loop. */
495 else /* if ((i % 2) == 1) */
497 int loop = (i - 1) / 2;
498 pbb->schedule = isl_map_equate (pbb->schedule, isl_dim_in, loop,
499 isl_dim_out, i);
503 isl_int_clear (val);
505 pbb->transformed = isl_map_copy (pbb->schedule);
508 /* Build for BB the static schedule.
510 The static schedule is a Dewey numbering of the abstract syntax
511 tree: http://en.wikipedia.org/wiki/Dewey_Decimal_Classification
513 The following example informally defines the static schedule:
516 for (i: ...)
518 for (j: ...)
524 for (k: ...)
532 Static schedules for A to F:
534 DEPTH
535 0 1 2
537 B 1 0 0
538 C 1 0 1
539 D 1 1 0
540 E 1 1 1
544 static void
545 build_scop_scattering (scop_p scop)
547 int i;
548 poly_bb_p pbb;
549 gimple_bb_p previous_gbb = NULL;
550 isl_space *dc = isl_set_get_space (scop->context);
551 isl_aff *static_sched;
553 dc = isl_space_add_dims (dc, isl_dim_set, number_of_loops (cfun));
554 static_sched = isl_aff_zero_on_domain (isl_local_space_from_space (dc));
556 /* We have to start schedules at 0 on the first component and
557 because we cannot compare_prefix_loops against a previous loop,
558 prefix will be equal to zero, and that index will be
559 incremented before copying. */
560 static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in, 0, -1);
562 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
564 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
565 int prefix;
566 int nb_scat_dims = pbb_dim_iter_domain (pbb) * 2 + 1;
568 if (previous_gbb)
569 prefix = nb_common_loops (SCOP_REGION (scop), previous_gbb, gbb);
570 else
571 prefix = 0;
573 previous_gbb = gbb;
575 static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in,
576 prefix, 1);
577 build_pbb_scattering_polyhedrons (static_sched, pbb, nb_scat_dims);
580 isl_aff_free (static_sched);
583 static isl_pw_aff *extract_affine (scop_p, tree, __isl_take isl_space *space);
585 /* Extract an affine expression from the chain of recurrence E. */
587 static isl_pw_aff *
588 extract_affine_chrec (scop_p s, tree e, __isl_take isl_space *space)
590 isl_pw_aff *lhs = extract_affine (s, CHREC_LEFT (e), isl_space_copy (space));
591 isl_pw_aff *rhs = extract_affine (s, CHREC_RIGHT (e), isl_space_copy (space));
592 isl_local_space *ls = isl_local_space_from_space (space);
593 unsigned pos = sese_loop_depth ((sese) s->region, get_chrec_loop (e)) - 1;
594 isl_aff *loop = isl_aff_set_coefficient_si
595 (isl_aff_zero_on_domain (ls), isl_dim_in, pos, 1);
596 isl_pw_aff *l = isl_pw_aff_from_aff (loop);
598 /* Before multiplying, make sure that the result is affine. */
599 gcc_assert (isl_pw_aff_is_cst (rhs)
600 || isl_pw_aff_is_cst (l));
602 return isl_pw_aff_add (lhs, isl_pw_aff_mul (rhs, l));
605 /* Extract an affine expression from the mult_expr E. */
607 static isl_pw_aff *
608 extract_affine_mul (scop_p s, tree e, __isl_take isl_space *space)
610 isl_pw_aff *lhs = extract_affine (s, TREE_OPERAND (e, 0),
611 isl_space_copy (space));
612 isl_pw_aff *rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
614 if (!isl_pw_aff_is_cst (lhs)
615 && !isl_pw_aff_is_cst (rhs))
617 isl_pw_aff_free (lhs);
618 isl_pw_aff_free (rhs);
619 return NULL;
622 return isl_pw_aff_mul (lhs, rhs);
625 /* Return an ISL identifier from the name of the ssa_name E. */
627 static isl_id *
628 isl_id_for_ssa_name (scop_p s, tree e)
630 const char *name = get_name (e);
631 isl_id *id;
633 if (name)
634 id = isl_id_alloc (s->ctx, name, e);
635 else
637 char name1[50];
638 snprintf (name1, sizeof (name1), "P_%d", SSA_NAME_VERSION (e));
639 id = isl_id_alloc (s->ctx, name1, e);
642 return id;
645 /* Return an ISL identifier for the data reference DR. */
647 static isl_id *
648 isl_id_for_dr (scop_p s, data_reference_p dr ATTRIBUTE_UNUSED)
650 /* Data references all get the same isl_id. They need to be comparable
651 and are distinguished through the first dimension, which contains the
652 alias set number. */
653 return isl_id_alloc (s->ctx, "", 0);
656 /* Extract an affine expression from the ssa_name E. */
658 static isl_pw_aff *
659 extract_affine_name (scop_p s, tree e, __isl_take isl_space *space)
661 isl_aff *aff;
662 isl_set *dom;
663 isl_id *id;
664 int dimension;
666 id = isl_id_for_ssa_name (s, e);
667 dimension = isl_space_find_dim_by_id (space, isl_dim_param, id);
668 isl_id_free(id);
669 dom = isl_set_universe (isl_space_copy (space));
670 aff = isl_aff_zero_on_domain (isl_local_space_from_space (space));
671 aff = isl_aff_add_coefficient_si (aff, isl_dim_param, dimension, 1);
672 return isl_pw_aff_alloc (dom, aff);
675 /* Extract an affine expression from the gmp constant G. */
677 static isl_pw_aff *
678 extract_affine_gmp (mpz_t g, __isl_take isl_space *space)
680 isl_local_space *ls = isl_local_space_from_space (isl_space_copy (space));
681 isl_aff *aff = isl_aff_zero_on_domain (ls);
682 isl_set *dom = isl_set_universe (space);
683 isl_int v;
685 isl_int_init (v);
686 isl_int_set_gmp (v, g);
687 aff = isl_aff_add_constant (aff, v);
688 isl_int_clear (v);
690 return isl_pw_aff_alloc (dom, aff);
693 /* Extract an affine expression from the integer_cst E. */
695 static isl_pw_aff *
696 extract_affine_int (tree e, __isl_take isl_space *space)
698 isl_pw_aff *res;
699 mpz_t g;
701 mpz_init (g);
702 tree_int_to_gmp (e, g);
703 res = extract_affine_gmp (g, space);
704 mpz_clear (g);
706 return res;
709 /* Compute pwaff mod 2^width. */
711 static isl_pw_aff *
712 wrap (isl_pw_aff *pwaff, unsigned width)
714 isl_int mod;
716 isl_int_init (mod);
717 isl_int_set_si (mod, 1);
718 isl_int_mul_2exp (mod, mod, width);
720 pwaff = isl_pw_aff_mod (pwaff, mod);
722 isl_int_clear (mod);
724 return pwaff;
727 /* When parameter NAME is in REGION, returns its index in SESE_PARAMS.
728 Otherwise returns -1. */
730 static inline int
731 parameter_index_in_region_1 (tree name, sese region)
733 int i;
734 tree p;
736 gcc_assert (TREE_CODE (name) == SSA_NAME);
738 FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, p)
739 if (p == name)
740 return i;
742 return -1;
745 /* When the parameter NAME is in REGION, returns its index in
746 SESE_PARAMS. Otherwise this function inserts NAME in SESE_PARAMS
747 and returns the index of NAME. */
749 static int
750 parameter_index_in_region (tree name, sese region)
752 int i;
754 gcc_assert (TREE_CODE (name) == SSA_NAME);
756 i = parameter_index_in_region_1 (name, region);
757 if (i != -1)
758 return i;
760 gcc_assert (SESE_ADD_PARAMS (region));
762 i = SESE_PARAMS (region).length ();
763 SESE_PARAMS (region).safe_push (name);
764 return i;
767 /* Extract an affine expression from the tree E in the scop S. */
769 static isl_pw_aff *
770 extract_affine (scop_p s, tree e, __isl_take isl_space *space)
772 isl_pw_aff *lhs, *rhs, *res;
773 tree type;
775 if (e == chrec_dont_know) {
776 isl_space_free (space);
777 return NULL;
780 switch (TREE_CODE (e))
782 case POLYNOMIAL_CHREC:
783 res = extract_affine_chrec (s, e, space);
784 break;
786 case MULT_EXPR:
787 res = extract_affine_mul (s, e, space);
788 break;
790 case PLUS_EXPR:
791 case POINTER_PLUS_EXPR:
792 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
793 rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
794 res = isl_pw_aff_add (lhs, rhs);
795 break;
797 case MINUS_EXPR:
798 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
799 rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
800 res = isl_pw_aff_sub (lhs, rhs);
801 break;
803 case NEGATE_EXPR:
804 case BIT_NOT_EXPR:
805 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
806 rhs = extract_affine (s, integer_minus_one_node, space);
807 res = isl_pw_aff_mul (lhs, rhs);
808 break;
810 case SSA_NAME:
811 gcc_assert (-1 != parameter_index_in_region_1 (e, SCOP_REGION (s)));
812 res = extract_affine_name (s, e, space);
813 break;
815 case INTEGER_CST:
816 res = extract_affine_int (e, space);
817 /* No need to wrap a single integer. */
818 return res;
820 CASE_CONVERT:
821 case NON_LVALUE_EXPR:
822 res = extract_affine (s, TREE_OPERAND (e, 0), space);
823 break;
825 default:
826 gcc_unreachable ();
827 break;
830 type = TREE_TYPE (e);
831 if (TYPE_UNSIGNED (type))
832 res = wrap (res, TYPE_PRECISION (type));
834 return res;
837 /* In the context of sese S, scan the expression E and translate it to
838 a linear expression C. When parsing a symbolic multiplication, K
839 represents the constant multiplier of an expression containing
840 parameters. */
842 static void
843 scan_tree_for_params (sese s, tree e)
845 if (e == chrec_dont_know)
846 return;
848 switch (TREE_CODE (e))
850 case POLYNOMIAL_CHREC:
851 scan_tree_for_params (s, CHREC_LEFT (e));
852 break;
854 case MULT_EXPR:
855 if (chrec_contains_symbols (TREE_OPERAND (e, 0)))
856 scan_tree_for_params (s, TREE_OPERAND (e, 0));
857 else
858 scan_tree_for_params (s, TREE_OPERAND (e, 1));
859 break;
861 case PLUS_EXPR:
862 case POINTER_PLUS_EXPR:
863 case MINUS_EXPR:
864 scan_tree_for_params (s, TREE_OPERAND (e, 0));
865 scan_tree_for_params (s, TREE_OPERAND (e, 1));
866 break;
868 case NEGATE_EXPR:
869 case BIT_NOT_EXPR:
870 CASE_CONVERT:
871 case NON_LVALUE_EXPR:
872 scan_tree_for_params (s, TREE_OPERAND (e, 0));
873 break;
875 case SSA_NAME:
876 parameter_index_in_region (e, s);
877 break;
879 case INTEGER_CST:
880 case ADDR_EXPR:
881 break;
883 default:
884 gcc_unreachable ();
885 break;
889 /* Find parameters with respect to REGION in BB. We are looking in memory
890 access functions, conditions and loop bounds. */
892 static void
893 find_params_in_bb (sese region, gimple_bb_p gbb)
895 int i;
896 unsigned j;
897 data_reference_p dr;
898 gimple stmt;
899 loop_p loop = GBB_BB (gbb)->loop_father;
901 /* Find parameters in the access functions of data references. */
902 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr)
903 for (j = 0; j < DR_NUM_DIMENSIONS (dr); j++)
904 scan_tree_for_params (region, DR_ACCESS_FN (dr, j));
906 /* Find parameters in conditional statements. */
907 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt)
909 tree lhs = scalar_evolution_in_region (region, loop,
910 gimple_cond_lhs (stmt));
911 tree rhs = scalar_evolution_in_region (region, loop,
912 gimple_cond_rhs (stmt));
914 scan_tree_for_params (region, lhs);
915 scan_tree_for_params (region, rhs);
919 /* Record the parameters used in the SCOP. A variable is a parameter
920 in a scop if it does not vary during the execution of that scop. */
922 static void
923 find_scop_parameters (scop_p scop)
925 poly_bb_p pbb;
926 unsigned i;
927 sese region = SCOP_REGION (scop);
928 struct loop *loop;
929 int nbp;
931 /* Find the parameters used in the loop bounds. */
932 FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop)
934 tree nb_iters = number_of_latch_executions (loop);
936 if (!chrec_contains_symbols (nb_iters))
937 continue;
939 nb_iters = scalar_evolution_in_region (region, loop, nb_iters);
940 scan_tree_for_params (region, nb_iters);
943 /* Find the parameters used in data accesses. */
944 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
945 find_params_in_bb (region, PBB_BLACK_BOX (pbb));
947 nbp = sese_nb_params (region);
948 scop_set_nb_params (scop, nbp);
949 SESE_ADD_PARAMS (region) = false;
952 tree e;
953 isl_space *space = isl_space_set_alloc (scop->ctx, nbp, 0);
955 FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, e)
956 space = isl_space_set_dim_id (space, isl_dim_param, i,
957 isl_id_for_ssa_name (scop, e));
959 scop->context = isl_set_universe (space);
963 /* Builds the constraint polyhedra for LOOP in SCOP. OUTER_PH gives
964 the constraints for the surrounding loops. */
966 static void
967 build_loop_iteration_domains (scop_p scop, struct loop *loop,
968 int nb,
969 isl_set *outer, isl_set **doms)
971 tree nb_iters = number_of_latch_executions (loop);
972 sese region = SCOP_REGION (scop);
974 isl_set *inner = isl_set_copy (outer);
975 isl_space *space;
976 isl_constraint *c;
977 int pos = isl_set_dim (outer, isl_dim_set);
978 isl_int v;
979 mpz_t g;
981 mpz_init (g);
982 isl_int_init (v);
984 inner = isl_set_add_dims (inner, isl_dim_set, 1);
985 space = isl_set_get_space (inner);
987 /* 0 <= loop_i */
988 c = isl_inequality_alloc
989 (isl_local_space_from_space (isl_space_copy (space)));
990 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, 1);
991 inner = isl_set_add_constraint (inner, c);
993 /* loop_i <= cst_nb_iters */
994 if (TREE_CODE (nb_iters) == INTEGER_CST)
996 c = isl_inequality_alloc
997 (isl_local_space_from_space(isl_space_copy (space)));
998 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1);
999 tree_int_to_gmp (nb_iters, g);
1000 isl_int_set_gmp (v, g);
1001 c = isl_constraint_set_constant (c, v);
1002 inner = isl_set_add_constraint (inner, c);
1005 /* loop_i <= expr_nb_iters */
1006 else if (!chrec_contains_undetermined (nb_iters))
1008 double_int nit;
1009 isl_pw_aff *aff;
1010 isl_set *valid;
1011 isl_local_space *ls;
1012 isl_aff *al;
1013 isl_set *le;
1015 nb_iters = scalar_evolution_in_region (region, loop, nb_iters);
1017 aff = extract_affine (scop, nb_iters, isl_set_get_space (inner));
1018 valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (aff));
1019 valid = isl_set_project_out (valid, isl_dim_set, 0,
1020 isl_set_dim (valid, isl_dim_set));
1021 scop->context = isl_set_intersect (scop->context, valid);
1023 ls = isl_local_space_from_space (isl_space_copy (space));
1024 al = isl_aff_set_coefficient_si (isl_aff_zero_on_domain (ls),
1025 isl_dim_in, pos, 1);
1026 le = isl_pw_aff_le_set (isl_pw_aff_from_aff (al),
1027 isl_pw_aff_copy (aff));
1028 inner = isl_set_intersect (inner, le);
1030 if (max_stmt_executions (loop, &nit))
1032 /* Insert in the context the constraints from the
1033 estimation of the number of iterations NIT and the
1034 symbolic number of iterations (involving parameter
1035 names) NB_ITERS. First, build the affine expression
1036 "NIT - NB_ITERS" and then say that it is positive,
1037 i.e., NIT approximates NB_ITERS: "NIT >= NB_ITERS". */
1038 isl_pw_aff *approx;
1039 mpz_t g;
1040 isl_set *x;
1041 isl_constraint *c;
1043 mpz_init (g);
1044 mpz_set_double_int (g, nit, false);
1045 mpz_sub_ui (g, g, 1);
1046 approx = extract_affine_gmp (g, isl_set_get_space (inner));
1047 x = isl_pw_aff_ge_set (approx, aff);
1048 x = isl_set_project_out (x, isl_dim_set, 0,
1049 isl_set_dim (x, isl_dim_set));
1050 scop->context = isl_set_intersect (scop->context, x);
1052 c = isl_inequality_alloc
1053 (isl_local_space_from_space (isl_space_copy (space)));
1054 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1);
1055 isl_int_set_gmp (v, g);
1056 mpz_clear (g);
1057 c = isl_constraint_set_constant (c, v);
1058 inner = isl_set_add_constraint (inner, c);
1060 else
1061 isl_pw_aff_free (aff);
1063 else
1064 gcc_unreachable ();
1066 if (loop->inner && loop_in_sese_p (loop->inner, region))
1067 build_loop_iteration_domains (scop, loop->inner, nb + 1,
1068 isl_set_copy (inner), doms);
1070 if (nb != 0
1071 && loop->next
1072 && loop_in_sese_p (loop->next, region))
1073 build_loop_iteration_domains (scop, loop->next, nb,
1074 isl_set_copy (outer), doms);
1076 doms[loop->num] = inner;
1078 isl_set_free (outer);
1079 isl_space_free (space);
1080 isl_int_clear (v);
1081 mpz_clear (g);
1084 /* Returns a linear expression for tree T evaluated in PBB. */
1086 static isl_pw_aff *
1087 create_pw_aff_from_tree (poly_bb_p pbb, tree t)
1089 scop_p scop = PBB_SCOP (pbb);
1091 t = scalar_evolution_in_region (SCOP_REGION (scop), pbb_loop (pbb), t);
1092 gcc_assert (!automatically_generated_chrec_p (t));
1094 return extract_affine (scop, t, isl_set_get_space (pbb->domain));
1097 /* Add conditional statement STMT to pbb. CODE is used as the comparison
1098 operator. This allows us to invert the condition or to handle
1099 inequalities. */
1101 static void
1102 add_condition_to_pbb (poly_bb_p pbb, gimple stmt, enum tree_code code)
1104 isl_pw_aff *lhs = create_pw_aff_from_tree (pbb, gimple_cond_lhs (stmt));
1105 isl_pw_aff *rhs = create_pw_aff_from_tree (pbb, gimple_cond_rhs (stmt));
1106 isl_set *cond;
1108 switch (code)
1110 case LT_EXPR:
1111 cond = isl_pw_aff_lt_set (lhs, rhs);
1112 break;
1114 case GT_EXPR:
1115 cond = isl_pw_aff_gt_set (lhs, rhs);
1116 break;
1118 case LE_EXPR:
1119 cond = isl_pw_aff_le_set (lhs, rhs);
1120 break;
1122 case GE_EXPR:
1123 cond = isl_pw_aff_ge_set (lhs, rhs);
1124 break;
1126 case EQ_EXPR:
1127 cond = isl_pw_aff_eq_set (lhs, rhs);
1128 break;
1130 case NE_EXPR:
1131 cond = isl_pw_aff_ne_set (lhs, rhs);
1132 break;
1134 default:
1135 isl_pw_aff_free(lhs);
1136 isl_pw_aff_free(rhs);
1137 return;
1140 cond = isl_set_coalesce (cond);
1141 cond = isl_set_set_tuple_id (cond, isl_set_get_tuple_id (pbb->domain));
1142 pbb->domain = isl_set_intersect (pbb->domain, cond);
1145 /* Add conditions to the domain of PBB. */
1147 static void
1148 add_conditions_to_domain (poly_bb_p pbb)
1150 unsigned int i;
1151 gimple stmt;
1152 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
1154 if (GBB_CONDITIONS (gbb).is_empty ())
1155 return;
1157 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt)
1158 switch (gimple_code (stmt))
1160 case GIMPLE_COND:
1162 enum tree_code code = gimple_cond_code (stmt);
1164 /* The conditions for ELSE-branches are inverted. */
1165 if (!GBB_CONDITION_CASES (gbb)[i])
1166 code = invert_tree_comparison (code, false);
1168 add_condition_to_pbb (pbb, stmt, code);
1169 break;
1172 case GIMPLE_SWITCH:
1173 /* Switch statements are not supported right now - fall through. */
1175 default:
1176 gcc_unreachable ();
1177 break;
1181 /* Traverses all the GBBs of the SCOP and add their constraints to the
1182 iteration domains. */
1184 static void
1185 add_conditions_to_constraints (scop_p scop)
1187 int i;
1188 poly_bb_p pbb;
1190 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1191 add_conditions_to_domain (pbb);
1194 /* Returns a COND_EXPR statement when BB has a single predecessor, the
1195 edge between BB and its predecessor is not a loop exit edge, and
1196 the last statement of the single predecessor is a COND_EXPR. */
1198 static gimple
1199 single_pred_cond_non_loop_exit (basic_block bb)
1201 if (single_pred_p (bb))
1203 edge e = single_pred_edge (bb);
1204 basic_block pred = e->src;
1205 gimple stmt;
1207 if (loop_depth (pred->loop_father) > loop_depth (bb->loop_father))
1208 return NULL;
1210 stmt = last_stmt (pred);
1212 if (stmt && gimple_code (stmt) == GIMPLE_COND)
1213 return stmt;
1216 return NULL;
1219 class sese_dom_walker : public dom_walker
1221 public:
1222 sese_dom_walker (cdi_direction, sese);
1223 ~sese_dom_walker ();
1225 virtual void before_dom_children (basic_block);
1226 virtual void after_dom_children (basic_block);
1228 private:
1229 vec<gimple> conditions_, cases_;
1230 sese region_;
1233 sese_dom_walker::sese_dom_walker (cdi_direction direction, sese region)
1234 : dom_walker (direction), region_ (region)
1236 conditions_.create (3);
1237 cases_.create (3);
1240 sese_dom_walker::~sese_dom_walker ()
1242 conditions_.release ();
1243 cases_.release ();
1246 /* Call-back for dom_walk executed before visiting the dominated
1247 blocks. */
1249 void
1250 sese_dom_walker::before_dom_children (basic_block bb)
1252 gimple_bb_p gbb;
1253 gimple stmt;
1255 if (!bb_in_sese_p (bb, region_))
1256 return;
1258 stmt = single_pred_cond_non_loop_exit (bb);
1260 if (stmt)
1262 edge e = single_pred_edge (bb);
1264 conditions_.safe_push (stmt);
1266 if (e->flags & EDGE_TRUE_VALUE)
1267 cases_.safe_push (stmt);
1268 else
1269 cases_.safe_push (NULL);
1272 gbb = gbb_from_bb (bb);
1274 if (gbb)
1276 GBB_CONDITIONS (gbb) = conditions_.copy ();
1277 GBB_CONDITION_CASES (gbb) = cases_.copy ();
1281 /* Call-back for dom_walk executed after visiting the dominated
1282 blocks. */
1284 void
1285 sese_dom_walker::after_dom_children (basic_block bb)
1287 if (!bb_in_sese_p (bb, region_))
1288 return;
1290 if (single_pred_cond_non_loop_exit (bb))
1292 conditions_.pop ();
1293 cases_.pop ();
1297 /* Add constraints on the possible values of parameter P from the type
1298 of P. */
1300 static void
1301 add_param_constraints (scop_p scop, graphite_dim_t p)
1303 tree parameter = SESE_PARAMS (SCOP_REGION (scop))[p];
1304 tree type = TREE_TYPE (parameter);
1305 tree lb = NULL_TREE;
1306 tree ub = NULL_TREE;
1308 if (POINTER_TYPE_P (type) || !TYPE_MIN_VALUE (type))
1309 lb = lower_bound_in_type (type, type);
1310 else
1311 lb = TYPE_MIN_VALUE (type);
1313 if (POINTER_TYPE_P (type) || !TYPE_MAX_VALUE (type))
1314 ub = upper_bound_in_type (type, type);
1315 else
1316 ub = TYPE_MAX_VALUE (type);
1318 if (lb)
1320 isl_space *space = isl_set_get_space (scop->context);
1321 isl_constraint *c;
1322 mpz_t g;
1323 isl_int v;
1325 c = isl_inequality_alloc (isl_local_space_from_space (space));
1326 mpz_init (g);
1327 isl_int_init (v);
1328 tree_int_to_gmp (lb, g);
1329 isl_int_set_gmp (v, g);
1330 isl_int_neg (v, v);
1331 mpz_clear (g);
1332 c = isl_constraint_set_constant (c, v);
1333 isl_int_clear (v);
1334 c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, 1);
1336 scop->context = isl_set_add_constraint (scop->context, c);
1339 if (ub)
1341 isl_space *space = isl_set_get_space (scop->context);
1342 isl_constraint *c;
1343 mpz_t g;
1344 isl_int v;
1346 c = isl_inequality_alloc (isl_local_space_from_space (space));
1348 mpz_init (g);
1349 isl_int_init (v);
1350 tree_int_to_gmp (ub, g);
1351 isl_int_set_gmp (v, g);
1352 mpz_clear (g);
1353 c = isl_constraint_set_constant (c, v);
1354 isl_int_clear (v);
1355 c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, -1);
1357 scop->context = isl_set_add_constraint (scop->context, c);
1361 /* Build the context of the SCOP. The context usually contains extra
1362 constraints that are added to the iteration domains that constrain
1363 some parameters. */
1365 static void
1366 build_scop_context (scop_p scop)
1368 graphite_dim_t p, n = scop_nb_params (scop);
1370 for (p = 0; p < n; p++)
1371 add_param_constraints (scop, p);
1374 /* Build the iteration domains: the loops belonging to the current
1375 SCOP, and that vary for the execution of the current basic block.
1376 Returns false if there is no loop in SCOP. */
1378 static void
1379 build_scop_iteration_domain (scop_p scop)
1381 struct loop *loop;
1382 sese region = SCOP_REGION (scop);
1383 int i;
1384 poly_bb_p pbb;
1385 int nb_loops = number_of_loops (cfun);
1386 isl_set **doms = XCNEWVEC (isl_set *, nb_loops);
1388 FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop)
1389 if (!loop_in_sese_p (loop_outer (loop), region))
1390 build_loop_iteration_domains (scop, loop, 0,
1391 isl_set_copy (scop->context), doms);
1393 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1395 loop = pbb_loop (pbb);
1397 if (doms[loop->num])
1398 pbb->domain = isl_set_copy (doms[loop->num]);
1399 else
1400 pbb->domain = isl_set_copy (scop->context);
1402 pbb->domain = isl_set_set_tuple_id (pbb->domain,
1403 isl_id_for_pbb (scop, pbb));
1406 for (i = 0; i < nb_loops; i++)
1407 if (doms[i])
1408 isl_set_free (doms[i]);
1410 free (doms);
1413 /* Add a constrain to the ACCESSES polyhedron for the alias set of
1414 data reference DR. ACCESSP_NB_DIMS is the dimension of the
1415 ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration
1416 domain. */
1418 static isl_map *
1419 pdr_add_alias_set (isl_map *acc, data_reference_p dr)
1421 isl_constraint *c;
1422 int alias_set_num = 0;
1423 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
1425 if (bap && bap->alias_set)
1426 alias_set_num = *(bap->alias_set);
1428 c = isl_equality_alloc
1429 (isl_local_space_from_space (isl_map_get_space (acc)));
1430 c = isl_constraint_set_constant_si (c, -alias_set_num);
1431 c = isl_constraint_set_coefficient_si (c, isl_dim_out, 0, 1);
1433 return isl_map_add_constraint (acc, c);
1436 /* Assign the affine expression INDEX to the output dimension POS of
1437 MAP and return the result. */
1439 static isl_map *
1440 set_index (isl_map *map, int pos, isl_pw_aff *index)
1442 isl_map *index_map;
1443 int len = isl_map_dim (map, isl_dim_out);
1444 isl_id *id;
1446 index_map = isl_map_from_pw_aff (index);
1447 index_map = isl_map_insert_dims (index_map, isl_dim_out, 0, pos);
1448 index_map = isl_map_add_dims (index_map, isl_dim_out, len - pos - 1);
1450 id = isl_map_get_tuple_id (map, isl_dim_out);
1451 index_map = isl_map_set_tuple_id (index_map, isl_dim_out, id);
1452 id = isl_map_get_tuple_id (map, isl_dim_in);
1453 index_map = isl_map_set_tuple_id (index_map, isl_dim_in, id);
1455 return isl_map_intersect (map, index_map);
1458 /* Add to ACCESSES polyhedron equalities defining the access functions
1459 to the memory. ACCESSP_NB_DIMS is the dimension of the ACCESSES
1460 polyhedron, DOM_NB_DIMS is the dimension of the iteration domain.
1461 PBB is the poly_bb_p that contains the data reference DR. */
1463 static isl_map *
1464 pdr_add_memory_accesses (isl_map *acc, data_reference_p dr, poly_bb_p pbb)
1466 int i, nb_subscripts = DR_NUM_DIMENSIONS (dr);
1467 scop_p scop = PBB_SCOP (pbb);
1469 for (i = 0; i < nb_subscripts; i++)
1471 isl_pw_aff *aff;
1472 tree afn = DR_ACCESS_FN (dr, nb_subscripts - 1 - i);
1474 aff = extract_affine (scop, afn,
1475 isl_space_domain (isl_map_get_space (acc)));
1476 acc = set_index (acc, i + 1, aff);
1479 return acc;
1482 /* Add constrains representing the size of the accessed data to the
1483 ACCESSES polyhedron. ACCESSP_NB_DIMS is the dimension of the
1484 ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration
1485 domain. */
1487 static isl_set *
1488 pdr_add_data_dimensions (isl_set *extent, scop_p scop, data_reference_p dr)
1490 tree ref = DR_REF (dr);
1491 int i, nb_subscripts = DR_NUM_DIMENSIONS (dr);
1493 for (i = nb_subscripts - 1; i >= 0; i--, ref = TREE_OPERAND (ref, 0))
1495 tree low, high;
1497 if (TREE_CODE (ref) != ARRAY_REF)
1498 break;
1500 low = array_ref_low_bound (ref);
1501 high = array_ref_up_bound (ref);
1503 /* XXX The PPL code dealt separately with
1504 subscript - low >= 0 and high - subscript >= 0 in case one of
1505 the two bounds isn't known. Do the same here? */
1507 if (host_integerp (low, 0)
1508 && high
1509 && host_integerp (high, 0)
1510 /* 1-element arrays at end of structures may extend over
1511 their declared size. */
1512 && !(array_at_struct_end_p (ref)
1513 && operand_equal_p (low, high, 0)))
1515 isl_id *id;
1516 isl_aff *aff;
1517 isl_set *univ, *lbs, *ubs;
1518 isl_pw_aff *index;
1519 isl_space *space;
1520 isl_set *valid;
1521 isl_pw_aff *lb = extract_affine_int (low, isl_set_get_space (extent));
1522 isl_pw_aff *ub = extract_affine_int (high, isl_set_get_space (extent));
1524 /* high >= 0 */
1525 valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (ub));
1526 valid = isl_set_project_out (valid, isl_dim_set, 0,
1527 isl_set_dim (valid, isl_dim_set));
1528 scop->context = isl_set_intersect (scop->context, valid);
1530 space = isl_set_get_space (extent);
1531 aff = isl_aff_zero_on_domain (isl_local_space_from_space (space));
1532 aff = isl_aff_add_coefficient_si (aff, isl_dim_in, i + 1, 1);
1533 univ = isl_set_universe (isl_space_domain (isl_aff_get_space (aff)));
1534 index = isl_pw_aff_alloc (univ, aff);
1536 id = isl_set_get_tuple_id (extent);
1537 lb = isl_pw_aff_set_tuple_id (lb, isl_dim_in, isl_id_copy (id));
1538 ub = isl_pw_aff_set_tuple_id (ub, isl_dim_in, id);
1540 /* low <= sub_i <= high */
1541 lbs = isl_pw_aff_ge_set (isl_pw_aff_copy (index), lb);
1542 ubs = isl_pw_aff_le_set (index, ub);
1543 extent = isl_set_intersect (extent, lbs);
1544 extent = isl_set_intersect (extent, ubs);
1548 return extent;
1551 /* Build data accesses for DR in PBB. */
1553 static void
1554 build_poly_dr (data_reference_p dr, poly_bb_p pbb)
1556 int dr_base_object_set;
1557 isl_map *acc;
1558 isl_set *extent;
1559 scop_p scop = PBB_SCOP (pbb);
1562 isl_space *dc = isl_set_get_space (pbb->domain);
1563 int nb_out = 1 + DR_NUM_DIMENSIONS (dr);
1564 isl_space *space = isl_space_add_dims (isl_space_from_domain (dc),
1565 isl_dim_out, nb_out);
1567 acc = isl_map_universe (space);
1568 acc = isl_map_set_tuple_id (acc, isl_dim_out, isl_id_for_dr (scop, dr));
1571 acc = pdr_add_alias_set (acc, dr);
1572 acc = pdr_add_memory_accesses (acc, dr, pbb);
1575 isl_id *id = isl_id_for_dr (scop, dr);
1576 int nb = 1 + DR_NUM_DIMENSIONS (dr);
1577 isl_space *space = isl_space_set_alloc (scop->ctx, 0, nb);
1578 int alias_set_num = 0;
1579 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
1581 if (bap && bap->alias_set)
1582 alias_set_num = *(bap->alias_set);
1584 space = isl_space_set_tuple_id (space, isl_dim_set, id);
1585 extent = isl_set_nat_universe (space);
1586 extent = isl_set_fix_si (extent, isl_dim_set, 0, alias_set_num);
1587 extent = pdr_add_data_dimensions (extent, scop, dr);
1590 gcc_assert (dr->aux);
1591 dr_base_object_set = ((base_alias_pair *)(dr->aux))->base_obj_set;
1593 new_poly_dr (pbb, dr_base_object_set,
1594 DR_IS_READ (dr) ? PDR_READ : PDR_WRITE,
1595 dr, DR_NUM_DIMENSIONS (dr), acc, extent);
1598 /* Write to FILE the alias graph of data references in DIMACS format. */
1600 static inline bool
1601 write_alias_graph_to_ascii_dimacs (FILE *file, char *comment,
1602 vec<data_reference_p> drs)
1604 int num_vertex = drs.length ();
1605 int edge_num = 0;
1606 data_reference_p dr1, dr2;
1607 int i, j;
1609 if (num_vertex == 0)
1610 return true;
1612 FOR_EACH_VEC_ELT (drs, i, dr1)
1613 for (j = i + 1; drs.iterate (j, &dr2); j++)
1614 if (dr_may_alias_p (dr1, dr2, true))
1615 edge_num++;
1617 fprintf (file, "$\n");
1619 if (comment)
1620 fprintf (file, "c %s\n", comment);
1622 fprintf (file, "p edge %d %d\n", num_vertex, edge_num);
1624 FOR_EACH_VEC_ELT (drs, i, dr1)
1625 for (j = i + 1; drs.iterate (j, &dr2); j++)
1626 if (dr_may_alias_p (dr1, dr2, true))
1627 fprintf (file, "e %d %d\n", i + 1, j + 1);
1629 return true;
1632 /* Write to FILE the alias graph of data references in DOT format. */
1634 static inline bool
1635 write_alias_graph_to_ascii_dot (FILE *file, char *comment,
1636 vec<data_reference_p> drs)
1638 int num_vertex = drs.length ();
1639 data_reference_p dr1, dr2;
1640 int i, j;
1642 if (num_vertex == 0)
1643 return true;
1645 fprintf (file, "$\n");
1647 if (comment)
1648 fprintf (file, "c %s\n", comment);
1650 /* First print all the vertices. */
1651 FOR_EACH_VEC_ELT (drs, i, dr1)
1652 fprintf (file, "n%d;\n", i);
1654 FOR_EACH_VEC_ELT (drs, i, dr1)
1655 for (j = i + 1; drs.iterate (j, &dr2); j++)
1656 if (dr_may_alias_p (dr1, dr2, true))
1657 fprintf (file, "n%d n%d\n", i, j);
1659 return true;
1662 /* Write to FILE the alias graph of data references in ECC format. */
1664 static inline bool
1665 write_alias_graph_to_ascii_ecc (FILE *file, char *comment,
1666 vec<data_reference_p> drs)
1668 int num_vertex = drs.length ();
1669 data_reference_p dr1, dr2;
1670 int i, j;
1672 if (num_vertex == 0)
1673 return true;
1675 fprintf (file, "$\n");
1677 if (comment)
1678 fprintf (file, "c %s\n", comment);
1680 FOR_EACH_VEC_ELT (drs, i, dr1)
1681 for (j = i + 1; drs.iterate (j, &dr2); j++)
1682 if (dr_may_alias_p (dr1, dr2, true))
1683 fprintf (file, "%d %d\n", i, j);
1685 return true;
1688 /* Check if DR1 and DR2 are in the same object set. */
1690 static bool
1691 dr_same_base_object_p (const struct data_reference *dr1,
1692 const struct data_reference *dr2)
1694 return operand_equal_p (DR_BASE_OBJECT (dr1), DR_BASE_OBJECT (dr2), 0);
1697 /* Uses DFS component number as representative of alias-sets. Also tests for
1698 optimality by verifying if every connected component is a clique. Returns
1699 true (1) if the above test is true, and false (0) otherwise. */
1701 static int
1702 build_alias_set_optimal_p (vec<data_reference_p> drs)
1704 int num_vertices = drs.length ();
1705 struct graph *g = new_graph (num_vertices);
1706 data_reference_p dr1, dr2;
1707 int i, j;
1708 int num_connected_components;
1709 int v_indx1, v_indx2, num_vertices_in_component;
1710 int *all_vertices;
1711 int *vertices;
1712 struct graph_edge *e;
1713 int this_component_is_clique;
1714 int all_components_are_cliques = 1;
1716 FOR_EACH_VEC_ELT (drs, i, dr1)
1717 for (j = i+1; drs.iterate (j, &dr2); j++)
1718 if (dr_may_alias_p (dr1, dr2, true))
1720 add_edge (g, i, j);
1721 add_edge (g, j, i);
1724 all_vertices = XNEWVEC (int, num_vertices);
1725 vertices = XNEWVEC (int, num_vertices);
1726 for (i = 0; i < num_vertices; i++)
1727 all_vertices[i] = i;
1729 num_connected_components = graphds_dfs (g, all_vertices, num_vertices,
1730 NULL, true, NULL);
1731 for (i = 0; i < g->n_vertices; i++)
1733 data_reference_p dr = drs[i];
1734 base_alias_pair *bap;
1736 gcc_assert (dr->aux);
1737 bap = (base_alias_pair *)(dr->aux);
1739 bap->alias_set = XNEW (int);
1740 *(bap->alias_set) = g->vertices[i].component + 1;
1743 /* Verify if the DFS numbering results in optimal solution. */
1744 for (i = 0; i < num_connected_components; i++)
1746 num_vertices_in_component = 0;
1747 /* Get all vertices whose DFS component number is the same as i. */
1748 for (j = 0; j < num_vertices; j++)
1749 if (g->vertices[j].component == i)
1750 vertices[num_vertices_in_component++] = j;
1752 /* Now test if the vertices in 'vertices' form a clique, by testing
1753 for edges among each pair. */
1754 this_component_is_clique = 1;
1755 for (v_indx1 = 0; v_indx1 < num_vertices_in_component; v_indx1++)
1757 for (v_indx2 = v_indx1+1; v_indx2 < num_vertices_in_component; v_indx2++)
1759 /* Check if the two vertices are connected by iterating
1760 through all the edges which have one of these are source. */
1761 e = g->vertices[vertices[v_indx2]].pred;
1762 while (e)
1764 if (e->src == vertices[v_indx1])
1765 break;
1766 e = e->pred_next;
1768 if (!e)
1770 this_component_is_clique = 0;
1771 break;
1774 if (!this_component_is_clique)
1775 all_components_are_cliques = 0;
1779 free (all_vertices);
1780 free (vertices);
1781 free_graph (g);
1782 return all_components_are_cliques;
1785 /* Group each data reference in DRS with its base object set num. */
1787 static void
1788 build_base_obj_set_for_drs (vec<data_reference_p> drs)
1790 int num_vertex = drs.length ();
1791 struct graph *g = new_graph (num_vertex);
1792 data_reference_p dr1, dr2;
1793 int i, j;
1794 int *queue;
1796 FOR_EACH_VEC_ELT (drs, i, dr1)
1797 for (j = i + 1; drs.iterate (j, &dr2); j++)
1798 if (dr_same_base_object_p (dr1, dr2))
1800 add_edge (g, i, j);
1801 add_edge (g, j, i);
1804 queue = XNEWVEC (int, num_vertex);
1805 for (i = 0; i < num_vertex; i++)
1806 queue[i] = i;
1808 graphds_dfs (g, queue, num_vertex, NULL, true, NULL);
1810 for (i = 0; i < g->n_vertices; i++)
1812 data_reference_p dr = drs[i];
1813 base_alias_pair *bap;
1815 gcc_assert (dr->aux);
1816 bap = (base_alias_pair *)(dr->aux);
1818 bap->base_obj_set = g->vertices[i].component + 1;
1821 free (queue);
1822 free_graph (g);
1825 /* Build the data references for PBB. */
1827 static void
1828 build_pbb_drs (poly_bb_p pbb)
1830 int j;
1831 data_reference_p dr;
1832 vec<data_reference_p> gbb_drs = GBB_DATA_REFS (PBB_BLACK_BOX (pbb));
1834 FOR_EACH_VEC_ELT (gbb_drs, j, dr)
1835 build_poly_dr (dr, pbb);
1838 /* Dump to file the alias graphs for the data references in DRS. */
1840 static void
1841 dump_alias_graphs (vec<data_reference_p> drs)
1843 char comment[100];
1844 FILE *file_dimacs, *file_ecc, *file_dot;
1846 file_dimacs = fopen ("/tmp/dr_alias_graph_dimacs", "ab");
1847 if (file_dimacs)
1849 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1850 current_function_name ());
1851 write_alias_graph_to_ascii_dimacs (file_dimacs, comment, drs);
1852 fclose (file_dimacs);
1855 file_ecc = fopen ("/tmp/dr_alias_graph_ecc", "ab");
1856 if (file_ecc)
1858 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1859 current_function_name ());
1860 write_alias_graph_to_ascii_ecc (file_ecc, comment, drs);
1861 fclose (file_ecc);
1864 file_dot = fopen ("/tmp/dr_alias_graph_dot", "ab");
1865 if (file_dot)
1867 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1868 current_function_name ());
1869 write_alias_graph_to_ascii_dot (file_dot, comment, drs);
1870 fclose (file_dot);
1874 /* Build data references in SCOP. */
1876 static void
1877 build_scop_drs (scop_p scop)
1879 int i, j;
1880 poly_bb_p pbb;
1881 data_reference_p dr;
1882 vec<data_reference_p> drs;
1883 drs.create (3);
1885 /* Remove all the PBBs that do not have data references: these basic
1886 blocks are not handled in the polyhedral representation. */
1887 for (i = 0; SCOP_BBS (scop).iterate (i, &pbb); i++)
1888 if (GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).is_empty ())
1890 free_gimple_bb (PBB_BLACK_BOX (pbb));
1891 free_poly_bb (pbb);
1892 SCOP_BBS (scop).ordered_remove (i);
1893 i--;
1896 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1897 for (j = 0; GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).iterate (j, &dr); j++)
1898 drs.safe_push (dr);
1900 FOR_EACH_VEC_ELT (drs, i, dr)
1901 dr->aux = XNEW (base_alias_pair);
1903 if (!build_alias_set_optimal_p (drs))
1905 /* TODO: Add support when building alias set is not optimal. */
1909 build_base_obj_set_for_drs (drs);
1911 /* When debugging, enable the following code. This cannot be used
1912 in production compilers. */
1913 if (0)
1914 dump_alias_graphs (drs);
1916 drs.release ();
1918 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1919 build_pbb_drs (pbb);
1922 /* Return a gsi at the position of the phi node STMT. */
1924 static gimple_stmt_iterator
1925 gsi_for_phi_node (gimple stmt)
1927 gimple_stmt_iterator psi;
1928 basic_block bb = gimple_bb (stmt);
1930 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
1931 if (stmt == gsi_stmt (psi))
1932 return psi;
1934 gcc_unreachable ();
1935 return psi;
1938 /* Analyze all the data references of STMTS and add them to the
1939 GBB_DATA_REFS vector of BB. */
1941 static void
1942 analyze_drs_in_stmts (scop_p scop, basic_block bb, vec<gimple> stmts)
1944 loop_p nest;
1945 gimple_bb_p gbb;
1946 gimple stmt;
1947 int i;
1948 sese region = SCOP_REGION (scop);
1950 if (!bb_in_sese_p (bb, region))
1951 return;
1953 nest = outermost_loop_in_sese_1 (region, bb);
1954 gbb = gbb_from_bb (bb);
1956 FOR_EACH_VEC_ELT (stmts, i, stmt)
1958 loop_p loop;
1960 if (is_gimple_debug (stmt))
1961 continue;
1963 loop = loop_containing_stmt (stmt);
1964 if (!loop_in_sese_p (loop, region))
1965 loop = nest;
1967 graphite_find_data_references_in_stmt (nest, loop, stmt,
1968 &GBB_DATA_REFS (gbb));
1972 /* Insert STMT at the end of the STMTS sequence and then insert the
1973 statements from STMTS at INSERT_GSI and call analyze_drs_in_stmts
1974 on STMTS. */
1976 static void
1977 insert_stmts (scop_p scop, gimple stmt, gimple_seq stmts,
1978 gimple_stmt_iterator insert_gsi)
1980 gimple_stmt_iterator gsi;
1981 vec<gimple> x;
1982 x.create (3);
1984 gimple_seq_add_stmt (&stmts, stmt);
1985 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
1986 x.safe_push (gsi_stmt (gsi));
1988 gsi_insert_seq_before (&insert_gsi, stmts, GSI_SAME_STMT);
1989 analyze_drs_in_stmts (scop, gsi_bb (insert_gsi), x);
1990 x.release ();
1993 /* Insert the assignment "RES := EXPR" just after AFTER_STMT. */
1995 static void
1996 insert_out_of_ssa_copy (scop_p scop, tree res, tree expr, gimple after_stmt)
1998 gimple_seq stmts;
1999 gimple_stmt_iterator gsi;
2000 tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE);
2001 gimple stmt = gimple_build_assign (unshare_expr (res), var);
2002 vec<gimple> x;
2003 x.create (3);
2005 gimple_seq_add_stmt (&stmts, stmt);
2006 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2007 x.safe_push (gsi_stmt (gsi));
2009 if (gimple_code (after_stmt) == GIMPLE_PHI)
2011 gsi = gsi_after_labels (gimple_bb (after_stmt));
2012 gsi_insert_seq_before (&gsi, stmts, GSI_NEW_STMT);
2014 else
2016 gsi = gsi_for_stmt (after_stmt);
2017 gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT);
2020 analyze_drs_in_stmts (scop, gimple_bb (after_stmt), x);
2021 x.release ();
2024 /* Creates a poly_bb_p for basic_block BB from the existing PBB. */
2026 static void
2027 new_pbb_from_pbb (scop_p scop, poly_bb_p pbb, basic_block bb)
2029 vec<data_reference_p> drs;
2030 drs.create (3);
2031 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
2032 gimple_bb_p gbb1 = new_gimple_bb (bb, drs);
2033 poly_bb_p pbb1 = new_poly_bb (scop, gbb1);
2034 int index, n = SCOP_BBS (scop).length ();
2036 /* The INDEX of PBB in SCOP_BBS. */
2037 for (index = 0; index < n; index++)
2038 if (SCOP_BBS (scop)[index] == pbb)
2039 break;
2041 pbb1->domain = isl_set_copy (pbb->domain);
2043 GBB_PBB (gbb1) = pbb1;
2044 GBB_CONDITIONS (gbb1) = GBB_CONDITIONS (gbb).copy ();
2045 GBB_CONDITION_CASES (gbb1) = GBB_CONDITION_CASES (gbb).copy ();
2046 SCOP_BBS (scop).safe_insert (index + 1, pbb1);
2049 /* Insert on edge E the assignment "RES := EXPR". */
2051 static void
2052 insert_out_of_ssa_copy_on_edge (scop_p scop, edge e, tree res, tree expr)
2054 gimple_stmt_iterator gsi;
2055 gimple_seq stmts = NULL;
2056 tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE);
2057 gimple stmt = gimple_build_assign (unshare_expr (res), var);
2058 basic_block bb;
2059 vec<gimple> x;
2060 x.create (3);
2062 gimple_seq_add_stmt (&stmts, stmt);
2063 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2064 x.safe_push (gsi_stmt (gsi));
2066 gsi_insert_seq_on_edge (e, stmts);
2067 gsi_commit_edge_inserts ();
2068 bb = gimple_bb (stmt);
2070 if (!bb_in_sese_p (bb, SCOP_REGION (scop)))
2071 return;
2073 if (!gbb_from_bb (bb))
2074 new_pbb_from_pbb (scop, pbb_from_bb (e->src), bb);
2076 analyze_drs_in_stmts (scop, bb, x);
2077 x.release ();
2080 /* Creates a zero dimension array of the same type as VAR. */
2082 static tree
2083 create_zero_dim_array (tree var, const char *base_name)
2085 tree index_type = build_index_type (integer_zero_node);
2086 tree elt_type = TREE_TYPE (var);
2087 tree array_type = build_array_type (elt_type, index_type);
2088 tree base = create_tmp_var (array_type, base_name);
2090 return build4 (ARRAY_REF, elt_type, base, integer_zero_node, NULL_TREE,
2091 NULL_TREE);
2094 /* Returns true when PHI is a loop close phi node. */
2096 static bool
2097 scalar_close_phi_node_p (gimple phi)
2099 if (gimple_code (phi) != GIMPLE_PHI
2100 || virtual_operand_p (gimple_phi_result (phi)))
2101 return false;
2103 /* Note that loop close phi nodes should have a single argument
2104 because we translated the representation into a canonical form
2105 before Graphite: see canonicalize_loop_closed_ssa_form. */
2106 return (gimple_phi_num_args (phi) == 1);
2109 /* For a definition DEF in REGION, propagates the expression EXPR in
2110 all the uses of DEF outside REGION. */
2112 static void
2113 propagate_expr_outside_region (tree def, tree expr, sese region)
2115 imm_use_iterator imm_iter;
2116 gimple use_stmt;
2117 gimple_seq stmts;
2118 bool replaced_once = false;
2120 gcc_assert (TREE_CODE (def) == SSA_NAME);
2122 expr = force_gimple_operand (unshare_expr (expr), &stmts, true,
2123 NULL_TREE);
2125 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2126 if (!is_gimple_debug (use_stmt)
2127 && !bb_in_sese_p (gimple_bb (use_stmt), region))
2129 ssa_op_iter iter;
2130 use_operand_p use_p;
2132 FOR_EACH_PHI_OR_STMT_USE (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2133 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0)
2134 && (replaced_once = true))
2135 replace_exp (use_p, expr);
2137 update_stmt (use_stmt);
2140 if (replaced_once)
2142 gsi_insert_seq_on_edge (SESE_ENTRY (region), stmts);
2143 gsi_commit_edge_inserts ();
2147 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2148 dimension array for it. */
2150 static void
2151 rewrite_close_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi)
2153 sese region = SCOP_REGION (scop);
2154 gimple phi = gsi_stmt (*psi);
2155 tree res = gimple_phi_result (phi);
2156 basic_block bb = gimple_bb (phi);
2157 gimple_stmt_iterator gsi = gsi_after_labels (bb);
2158 tree arg = gimple_phi_arg_def (phi, 0);
2159 gimple stmt;
2161 /* Note that loop close phi nodes should have a single argument
2162 because we translated the representation into a canonical form
2163 before Graphite: see canonicalize_loop_closed_ssa_form. */
2164 gcc_assert (gimple_phi_num_args (phi) == 1);
2166 /* The phi node can be a non close phi node, when its argument is
2167 invariant, or a default definition. */
2168 if (is_gimple_min_invariant (arg)
2169 || SSA_NAME_IS_DEFAULT_DEF (arg))
2171 propagate_expr_outside_region (res, arg, region);
2172 gsi_next (psi);
2173 return;
2176 else if (gimple_bb (SSA_NAME_DEF_STMT (arg))->loop_father == bb->loop_father)
2178 propagate_expr_outside_region (res, arg, region);
2179 stmt = gimple_build_assign (res, arg);
2180 remove_phi_node (psi, false);
2181 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2182 SSA_NAME_DEF_STMT (res) = stmt;
2183 return;
2186 /* If res is scev analyzable and is not a scalar value, it is safe
2187 to ignore the close phi node: it will be code generated in the
2188 out of Graphite pass. */
2189 else if (scev_analyzable_p (res, region))
2191 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (res));
2192 tree scev;
2194 if (!loop_in_sese_p (loop, region))
2196 loop = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2197 scev = scalar_evolution_in_region (region, loop, arg);
2198 scev = compute_overall_effect_of_inner_loop (loop, scev);
2200 else
2201 scev = scalar_evolution_in_region (region, loop, res);
2203 if (tree_does_not_contain_chrecs (scev))
2204 propagate_expr_outside_region (res, scev, region);
2206 gsi_next (psi);
2207 return;
2209 else
2211 tree zero_dim_array = create_zero_dim_array (res, "Close_Phi");
2213 stmt = gimple_build_assign (res, unshare_expr (zero_dim_array));
2215 if (TREE_CODE (arg) == SSA_NAME)
2216 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2217 SSA_NAME_DEF_STMT (arg));
2218 else
2219 insert_out_of_ssa_copy_on_edge (scop, single_pred_edge (bb),
2220 zero_dim_array, arg);
2223 remove_phi_node (psi, false);
2224 SSA_NAME_DEF_STMT (res) = stmt;
2226 insert_stmts (scop, stmt, NULL, gsi_after_labels (bb));
2229 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2230 dimension array for it. */
2232 static void
2233 rewrite_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi)
2235 size_t i;
2236 gimple phi = gsi_stmt (*psi);
2237 basic_block bb = gimple_bb (phi);
2238 tree res = gimple_phi_result (phi);
2239 tree zero_dim_array = create_zero_dim_array (res, "phi_out_of_ssa");
2240 gimple stmt;
2242 for (i = 0; i < gimple_phi_num_args (phi); i++)
2244 tree arg = gimple_phi_arg_def (phi, i);
2245 edge e = gimple_phi_arg_edge (phi, i);
2247 /* Avoid the insertion of code in the loop latch to please the
2248 pattern matching of the vectorizer. */
2249 if (TREE_CODE (arg) == SSA_NAME
2250 && e->src == bb->loop_father->latch)
2251 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2252 SSA_NAME_DEF_STMT (arg));
2253 else
2254 insert_out_of_ssa_copy_on_edge (scop, e, zero_dim_array, arg);
2257 stmt = gimple_build_assign (res, unshare_expr (zero_dim_array));
2258 remove_phi_node (psi, false);
2259 SSA_NAME_DEF_STMT (res) = stmt;
2260 insert_stmts (scop, stmt, NULL, gsi_after_labels (bb));
2263 /* Rewrite the degenerate phi node at position PSI from the degenerate
2264 form "x = phi (y, y, ..., y)" to "x = y". */
2266 static void
2267 rewrite_degenerate_phi (gimple_stmt_iterator *psi)
2269 tree rhs;
2270 gimple stmt;
2271 gimple_stmt_iterator gsi;
2272 gimple phi = gsi_stmt (*psi);
2273 tree res = gimple_phi_result (phi);
2274 basic_block bb;
2276 bb = gimple_bb (phi);
2277 rhs = degenerate_phi_result (phi);
2278 gcc_assert (rhs);
2280 stmt = gimple_build_assign (res, rhs);
2281 remove_phi_node (psi, false);
2282 SSA_NAME_DEF_STMT (res) = stmt;
2284 gsi = gsi_after_labels (bb);
2285 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2288 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2290 static void
2291 rewrite_reductions_out_of_ssa (scop_p scop)
2293 basic_block bb;
2294 gimple_stmt_iterator psi;
2295 sese region = SCOP_REGION (scop);
2297 FOR_EACH_BB (bb)
2298 if (bb_in_sese_p (bb, region))
2299 for (psi = gsi_start_phis (bb); !gsi_end_p (psi);)
2301 gimple phi = gsi_stmt (psi);
2303 if (virtual_operand_p (gimple_phi_result (phi)))
2305 gsi_next (&psi);
2306 continue;
2309 if (gimple_phi_num_args (phi) > 1
2310 && degenerate_phi_result (phi))
2311 rewrite_degenerate_phi (&psi);
2313 else if (scalar_close_phi_node_p (phi))
2314 rewrite_close_phi_out_of_ssa (scop, &psi);
2316 else if (reduction_phi_p (region, &psi))
2317 rewrite_phi_out_of_ssa (scop, &psi);
2320 update_ssa (TODO_update_ssa);
2321 #ifdef ENABLE_CHECKING
2322 verify_loop_closed_ssa (true);
2323 #endif
2326 /* Rewrite the scalar dependence of DEF used in USE_STMT with a memory
2327 read from ZERO_DIM_ARRAY. */
2329 static void
2330 rewrite_cross_bb_scalar_dependence (scop_p scop, tree zero_dim_array,
2331 tree def, gimple use_stmt)
2333 gimple name_stmt;
2334 tree name;
2335 ssa_op_iter iter;
2336 use_operand_p use_p;
2338 gcc_assert (gimple_code (use_stmt) != GIMPLE_PHI);
2340 name = copy_ssa_name (def, NULL);
2341 name_stmt = gimple_build_assign (name, zero_dim_array);
2343 gimple_assign_set_lhs (name_stmt, name);
2344 insert_stmts (scop, name_stmt, NULL, gsi_for_stmt (use_stmt));
2346 FOR_EACH_SSA_USE_OPERAND (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2347 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0))
2348 replace_exp (use_p, name);
2350 update_stmt (use_stmt);
2353 /* For every definition DEF in the SCOP that is used outside the scop,
2354 insert a closing-scop definition in the basic block just after this
2355 SCOP. */
2357 static void
2358 handle_scalar_deps_crossing_scop_limits (scop_p scop, tree def, gimple stmt)
2360 tree var = create_tmp_reg (TREE_TYPE (def), NULL);
2361 tree new_name = make_ssa_name (var, stmt);
2362 bool needs_copy = false;
2363 use_operand_p use_p;
2364 imm_use_iterator imm_iter;
2365 gimple use_stmt;
2366 sese region = SCOP_REGION (scop);
2368 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2370 if (!bb_in_sese_p (gimple_bb (use_stmt), region))
2372 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
2374 SET_USE (use_p, new_name);
2376 update_stmt (use_stmt);
2377 needs_copy = true;
2381 /* Insert in the empty BB just after the scop a use of DEF such
2382 that the rewrite of cross_bb_scalar_dependences won't insert
2383 arrays everywhere else. */
2384 if (needs_copy)
2386 gimple assign = gimple_build_assign (new_name, def);
2387 gimple_stmt_iterator psi = gsi_after_labels (SESE_EXIT (region)->dest);
2389 SSA_NAME_DEF_STMT (new_name) = assign;
2390 update_stmt (assign);
2391 gsi_insert_before (&psi, assign, GSI_SAME_STMT);
2395 /* Rewrite the scalar dependences crossing the boundary of the BB
2396 containing STMT with an array. Return true when something has been
2397 changed. */
2399 static bool
2400 rewrite_cross_bb_scalar_deps (scop_p scop, gimple_stmt_iterator *gsi)
2402 sese region = SCOP_REGION (scop);
2403 gimple stmt = gsi_stmt (*gsi);
2404 imm_use_iterator imm_iter;
2405 tree def;
2406 basic_block def_bb;
2407 tree zero_dim_array = NULL_TREE;
2408 gimple use_stmt;
2409 bool res = false;
2411 switch (gimple_code (stmt))
2413 case GIMPLE_ASSIGN:
2414 def = gimple_assign_lhs (stmt);
2415 break;
2417 case GIMPLE_CALL:
2418 def = gimple_call_lhs (stmt);
2419 break;
2421 default:
2422 return false;
2425 if (!def
2426 || !is_gimple_reg (def))
2427 return false;
2429 if (scev_analyzable_p (def, region))
2431 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (def));
2432 tree scev = scalar_evolution_in_region (region, loop, def);
2434 if (tree_contains_chrecs (scev, NULL))
2435 return false;
2437 propagate_expr_outside_region (def, scev, region);
2438 return true;
2441 def_bb = gimple_bb (stmt);
2443 handle_scalar_deps_crossing_scop_limits (scop, def, stmt);
2445 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2446 if (gimple_code (use_stmt) == GIMPLE_PHI
2447 && (res = true))
2449 gimple_stmt_iterator psi = gsi_for_stmt (use_stmt);
2451 if (scalar_close_phi_node_p (gsi_stmt (psi)))
2452 rewrite_close_phi_out_of_ssa (scop, &psi);
2453 else
2454 rewrite_phi_out_of_ssa (scop, &psi);
2457 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2458 if (gimple_code (use_stmt) != GIMPLE_PHI
2459 && def_bb != gimple_bb (use_stmt)
2460 && !is_gimple_debug (use_stmt)
2461 && (res = true))
2463 if (!zero_dim_array)
2465 zero_dim_array = create_zero_dim_array
2466 (def, "Cross_BB_scalar_dependence");
2467 insert_out_of_ssa_copy (scop, zero_dim_array, def,
2468 SSA_NAME_DEF_STMT (def));
2469 gsi_next (gsi);
2472 rewrite_cross_bb_scalar_dependence (scop, zero_dim_array,
2473 def, use_stmt);
2476 return res;
2479 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2481 static void
2482 rewrite_cross_bb_scalar_deps_out_of_ssa (scop_p scop)
2484 basic_block bb;
2485 gimple_stmt_iterator psi;
2486 sese region = SCOP_REGION (scop);
2487 bool changed = false;
2489 /* Create an extra empty BB after the scop. */
2490 split_edge (SESE_EXIT (region));
2492 FOR_EACH_BB (bb)
2493 if (bb_in_sese_p (bb, region))
2494 for (psi = gsi_start_bb (bb); !gsi_end_p (psi); gsi_next (&psi))
2495 changed |= rewrite_cross_bb_scalar_deps (scop, &psi);
2497 if (changed)
2499 scev_reset_htab ();
2500 update_ssa (TODO_update_ssa);
2501 #ifdef ENABLE_CHECKING
2502 verify_loop_closed_ssa (true);
2503 #endif
2507 /* Returns the number of pbbs that are in loops contained in SCOP. */
2509 static int
2510 nb_pbbs_in_loops (scop_p scop)
2512 int i;
2513 poly_bb_p pbb;
2514 int res = 0;
2516 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
2517 if (loop_in_sese_p (gbb_loop (PBB_BLACK_BOX (pbb)), SCOP_REGION (scop)))
2518 res++;
2520 return res;
2523 /* Return the number of data references in BB that write in
2524 memory. */
2526 static int
2527 nb_data_writes_in_bb (basic_block bb)
2529 int res = 0;
2530 gimple_stmt_iterator gsi;
2532 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
2533 if (gimple_vdef (gsi_stmt (gsi)))
2534 res++;
2536 return res;
2539 /* Splits at STMT the basic block BB represented as PBB in the
2540 polyhedral form. */
2542 static edge
2543 split_pbb (scop_p scop, poly_bb_p pbb, basic_block bb, gimple stmt)
2545 edge e1 = split_block (bb, stmt);
2546 new_pbb_from_pbb (scop, pbb, e1->dest);
2547 return e1;
2550 /* Splits STMT out of its current BB. This is done for reduction
2551 statements for which we want to ignore data dependences. */
2553 static basic_block
2554 split_reduction_stmt (scop_p scop, gimple stmt)
2556 basic_block bb = gimple_bb (stmt);
2557 poly_bb_p pbb = pbb_from_bb (bb);
2558 gimple_bb_p gbb = gbb_from_bb (bb);
2559 edge e1;
2560 int i;
2561 data_reference_p dr;
2563 /* Do not split basic blocks with no writes to memory: the reduction
2564 will be the only write to memory. */
2565 if (nb_data_writes_in_bb (bb) == 0
2566 /* Or if we have already marked BB as a reduction. */
2567 || PBB_IS_REDUCTION (pbb_from_bb (bb)))
2568 return bb;
2570 e1 = split_pbb (scop, pbb, bb, stmt);
2572 /* Split once more only when the reduction stmt is not the only one
2573 left in the original BB. */
2574 if (!gsi_one_before_end_p (gsi_start_nondebug_bb (bb)))
2576 gimple_stmt_iterator gsi = gsi_last_bb (bb);
2577 gsi_prev (&gsi);
2578 e1 = split_pbb (scop, pbb, bb, gsi_stmt (gsi));
2581 /* A part of the data references will end in a different basic block
2582 after the split: move the DRs from the original GBB to the newly
2583 created GBB1. */
2584 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr)
2586 basic_block bb1 = gimple_bb (DR_STMT (dr));
2588 if (bb1 != bb)
2590 gimple_bb_p gbb1 = gbb_from_bb (bb1);
2591 GBB_DATA_REFS (gbb1).safe_push (dr);
2592 GBB_DATA_REFS (gbb).ordered_remove (i);
2593 i--;
2597 return e1->dest;
2600 /* Return true when stmt is a reduction operation. */
2602 static inline bool
2603 is_reduction_operation_p (gimple stmt)
2605 enum tree_code code;
2607 gcc_assert (is_gimple_assign (stmt));
2608 code = gimple_assign_rhs_code (stmt);
2610 return flag_associative_math
2611 && commutative_tree_code (code)
2612 && associative_tree_code (code);
2615 /* Returns true when PHI contains an argument ARG. */
2617 static bool
2618 phi_contains_arg (gimple phi, tree arg)
2620 size_t i;
2622 for (i = 0; i < gimple_phi_num_args (phi); i++)
2623 if (operand_equal_p (arg, gimple_phi_arg_def (phi, i), 0))
2624 return true;
2626 return false;
2629 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2631 static gimple
2632 follow_ssa_with_commutative_ops (tree arg, tree lhs)
2634 gimple stmt;
2636 if (TREE_CODE (arg) != SSA_NAME)
2637 return NULL;
2639 stmt = SSA_NAME_DEF_STMT (arg);
2641 if (gimple_code (stmt) == GIMPLE_NOP
2642 || gimple_code (stmt) == GIMPLE_CALL)
2643 return NULL;
2645 if (gimple_code (stmt) == GIMPLE_PHI)
2647 if (phi_contains_arg (stmt, lhs))
2648 return stmt;
2649 return NULL;
2652 if (!is_gimple_assign (stmt))
2653 return NULL;
2655 if (gimple_num_ops (stmt) == 2)
2656 return follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2658 if (is_reduction_operation_p (stmt))
2660 gimple res = follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2662 return res ? res :
2663 follow_ssa_with_commutative_ops (gimple_assign_rhs2 (stmt), lhs);
2666 return NULL;
2669 /* Detect commutative and associative scalar reductions starting at
2670 the STMT. Return the phi node of the reduction cycle, or NULL. */
2672 static gimple
2673 detect_commutative_reduction_arg (tree lhs, gimple stmt, tree arg,
2674 vec<gimple> *in,
2675 vec<gimple> *out)
2677 gimple phi = follow_ssa_with_commutative_ops (arg, lhs);
2679 if (!phi)
2680 return NULL;
2682 in->safe_push (stmt);
2683 out->safe_push (stmt);
2684 return phi;
2687 /* Detect commutative and associative scalar reductions starting at
2688 STMT. Return the phi node of the reduction cycle, or NULL. */
2690 static gimple
2691 detect_commutative_reduction_assign (gimple stmt, vec<gimple> *in,
2692 vec<gimple> *out)
2694 tree lhs = gimple_assign_lhs (stmt);
2696 if (gimple_num_ops (stmt) == 2)
2697 return detect_commutative_reduction_arg (lhs, stmt,
2698 gimple_assign_rhs1 (stmt),
2699 in, out);
2701 if (is_reduction_operation_p (stmt))
2703 gimple res = detect_commutative_reduction_arg (lhs, stmt,
2704 gimple_assign_rhs1 (stmt),
2705 in, out);
2706 return res ? res
2707 : detect_commutative_reduction_arg (lhs, stmt,
2708 gimple_assign_rhs2 (stmt),
2709 in, out);
2712 return NULL;
2715 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2717 static gimple
2718 follow_inital_value_to_phi (tree arg, tree lhs)
2720 gimple stmt;
2722 if (!arg || TREE_CODE (arg) != SSA_NAME)
2723 return NULL;
2725 stmt = SSA_NAME_DEF_STMT (arg);
2727 if (gimple_code (stmt) == GIMPLE_PHI
2728 && phi_contains_arg (stmt, lhs))
2729 return stmt;
2731 return NULL;
2735 /* Return the argument of the loop PHI that is the initial value coming
2736 from outside the loop. */
2738 static edge
2739 edge_initial_value_for_loop_phi (gimple phi)
2741 size_t i;
2743 for (i = 0; i < gimple_phi_num_args (phi); i++)
2745 edge e = gimple_phi_arg_edge (phi, i);
2747 if (loop_depth (e->src->loop_father)
2748 < loop_depth (e->dest->loop_father))
2749 return e;
2752 return NULL;
2755 /* Return the argument of the loop PHI that is the initial value coming
2756 from outside the loop. */
2758 static tree
2759 initial_value_for_loop_phi (gimple phi)
2761 size_t i;
2763 for (i = 0; i < gimple_phi_num_args (phi); i++)
2765 edge e = gimple_phi_arg_edge (phi, i);
2767 if (loop_depth (e->src->loop_father)
2768 < loop_depth (e->dest->loop_father))
2769 return gimple_phi_arg_def (phi, i);
2772 return NULL_TREE;
2775 /* Returns true when DEF is used outside the reduction cycle of
2776 LOOP_PHI. */
2778 static bool
2779 used_outside_reduction (tree def, gimple loop_phi)
2781 use_operand_p use_p;
2782 imm_use_iterator imm_iter;
2783 loop_p loop = loop_containing_stmt (loop_phi);
2785 /* In LOOP, DEF should be used only in LOOP_PHI. */
2786 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2788 gimple stmt = USE_STMT (use_p);
2790 if (stmt != loop_phi
2791 && !is_gimple_debug (stmt)
2792 && flow_bb_inside_loop_p (loop, gimple_bb (stmt)))
2793 return true;
2796 return false;
2799 /* Detect commutative and associative scalar reductions belonging to
2800 the SCOP starting at the loop closed phi node STMT. Return the phi
2801 node of the reduction cycle, or NULL. */
2803 static gimple
2804 detect_commutative_reduction (scop_p scop, gimple stmt, vec<gimple> *in,
2805 vec<gimple> *out)
2807 if (scalar_close_phi_node_p (stmt))
2809 gimple def, loop_phi, phi, close_phi = stmt;
2810 tree init, lhs, arg = gimple_phi_arg_def (close_phi, 0);
2812 if (TREE_CODE (arg) != SSA_NAME)
2813 return NULL;
2815 /* Note that loop close phi nodes should have a single argument
2816 because we translated the representation into a canonical form
2817 before Graphite: see canonicalize_loop_closed_ssa_form. */
2818 gcc_assert (gimple_phi_num_args (close_phi) == 1);
2820 def = SSA_NAME_DEF_STMT (arg);
2821 if (!stmt_in_sese_p (def, SCOP_REGION (scop))
2822 || !(loop_phi = detect_commutative_reduction (scop, def, in, out)))
2823 return NULL;
2825 lhs = gimple_phi_result (close_phi);
2826 init = initial_value_for_loop_phi (loop_phi);
2827 phi = follow_inital_value_to_phi (init, lhs);
2829 if (phi && (used_outside_reduction (lhs, phi)
2830 || !has_single_use (gimple_phi_result (phi))))
2831 return NULL;
2833 in->safe_push (loop_phi);
2834 out->safe_push (close_phi);
2835 return phi;
2838 if (gimple_code (stmt) == GIMPLE_ASSIGN)
2839 return detect_commutative_reduction_assign (stmt, in, out);
2841 return NULL;
2844 /* Translate the scalar reduction statement STMT to an array RED
2845 knowing that its recursive phi node is LOOP_PHI. */
2847 static void
2848 translate_scalar_reduction_to_array_for_stmt (scop_p scop, tree red,
2849 gimple stmt, gimple loop_phi)
2851 tree res = gimple_phi_result (loop_phi);
2852 gimple assign = gimple_build_assign (res, unshare_expr (red));
2853 gimple_stmt_iterator gsi;
2855 insert_stmts (scop, assign, NULL, gsi_after_labels (gimple_bb (loop_phi)));
2857 assign = gimple_build_assign (unshare_expr (red), gimple_assign_lhs (stmt));
2858 gsi = gsi_for_stmt (stmt);
2859 gsi_next (&gsi);
2860 insert_stmts (scop, assign, NULL, gsi);
2863 /* Removes the PHI node and resets all the debug stmts that are using
2864 the PHI_RESULT. */
2866 static void
2867 remove_phi (gimple phi)
2869 imm_use_iterator imm_iter;
2870 tree def;
2871 use_operand_p use_p;
2872 gimple_stmt_iterator gsi;
2873 vec<gimple> update;
2874 update.create (3);
2875 unsigned int i;
2876 gimple stmt;
2878 def = PHI_RESULT (phi);
2879 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2881 stmt = USE_STMT (use_p);
2883 if (is_gimple_debug (stmt))
2885 gimple_debug_bind_reset_value (stmt);
2886 update.safe_push (stmt);
2890 FOR_EACH_VEC_ELT (update, i, stmt)
2891 update_stmt (stmt);
2893 update.release ();
2895 gsi = gsi_for_phi_node (phi);
2896 remove_phi_node (&gsi, false);
2899 /* Helper function for for_each_index. For each INDEX of the data
2900 reference REF, returns true when its indices are valid in the loop
2901 nest LOOP passed in as DATA. */
2903 static bool
2904 dr_indices_valid_in_loop (tree ref ATTRIBUTE_UNUSED, tree *index, void *data)
2906 loop_p loop;
2907 basic_block header, def_bb;
2908 gimple stmt;
2910 if (TREE_CODE (*index) != SSA_NAME)
2911 return true;
2913 loop = *((loop_p *) data);
2914 header = loop->header;
2915 stmt = SSA_NAME_DEF_STMT (*index);
2917 if (!stmt)
2918 return true;
2920 def_bb = gimple_bb (stmt);
2922 if (!def_bb)
2923 return true;
2925 return dominated_by_p (CDI_DOMINATORS, header, def_bb);
2928 /* When the result of a CLOSE_PHI is written to a memory location,
2929 return a pointer to that memory reference, otherwise return
2930 NULL_TREE. */
2932 static tree
2933 close_phi_written_to_memory (gimple close_phi)
2935 imm_use_iterator imm_iter;
2936 use_operand_p use_p;
2937 gimple stmt;
2938 tree res, def = gimple_phi_result (close_phi);
2940 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2941 if ((stmt = USE_STMT (use_p))
2942 && gimple_code (stmt) == GIMPLE_ASSIGN
2943 && (res = gimple_assign_lhs (stmt)))
2945 switch (TREE_CODE (res))
2947 case VAR_DECL:
2948 case PARM_DECL:
2949 case RESULT_DECL:
2950 return res;
2952 case ARRAY_REF:
2953 case MEM_REF:
2955 tree arg = gimple_phi_arg_def (close_phi, 0);
2956 loop_p nest = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2958 /* FIXME: this restriction is for id-{24,25}.f and
2959 could be handled by duplicating the computation of
2960 array indices before the loop of the close_phi. */
2961 if (for_each_index (&res, dr_indices_valid_in_loop, &nest))
2962 return res;
2964 /* Fallthru. */
2966 default:
2967 continue;
2970 return NULL_TREE;
2973 /* Rewrite out of SSA the reduction described by the loop phi nodes
2974 IN, and the close phi nodes OUT. IN and OUT are structured by loop
2975 levels like this:
2977 IN: stmt, loop_n, ..., loop_0
2978 OUT: stmt, close_n, ..., close_0
2980 the first element is the reduction statement, and the next elements
2981 are the loop and close phi nodes of each of the outer loops. */
2983 static void
2984 translate_scalar_reduction_to_array (scop_p scop,
2985 vec<gimple> in,
2986 vec<gimple> out)
2988 gimple loop_phi;
2989 unsigned int i = out.length () - 1;
2990 tree red = close_phi_written_to_memory (out[i]);
2992 FOR_EACH_VEC_ELT (in, i, loop_phi)
2994 gimple close_phi = out[i];
2996 if (i == 0)
2998 gimple stmt = loop_phi;
2999 basic_block bb = split_reduction_stmt (scop, stmt);
3000 poly_bb_p pbb = pbb_from_bb (bb);
3001 PBB_IS_REDUCTION (pbb) = true;
3002 gcc_assert (close_phi == loop_phi);
3004 if (!red)
3005 red = create_zero_dim_array
3006 (gimple_assign_lhs (stmt), "Commutative_Associative_Reduction");
3008 translate_scalar_reduction_to_array_for_stmt (scop, red, stmt, in[1]);
3009 continue;
3012 if (i == in.length () - 1)
3014 insert_out_of_ssa_copy (scop, gimple_phi_result (close_phi),
3015 unshare_expr (red), close_phi);
3016 insert_out_of_ssa_copy_on_edge
3017 (scop, edge_initial_value_for_loop_phi (loop_phi),
3018 unshare_expr (red), initial_value_for_loop_phi (loop_phi));
3021 remove_phi (loop_phi);
3022 remove_phi (close_phi);
3026 /* Rewrites out of SSA a commutative reduction at CLOSE_PHI. Returns
3027 true when something has been changed. */
3029 static bool
3030 rewrite_commutative_reductions_out_of_ssa_close_phi (scop_p scop,
3031 gimple close_phi)
3033 bool res;
3034 vec<gimple> in;
3035 in.create (10);
3036 vec<gimple> out;
3037 out.create (10);
3039 detect_commutative_reduction (scop, close_phi, &in, &out);
3040 res = in.length () > 1;
3041 if (res)
3042 translate_scalar_reduction_to_array (scop, in, out);
3044 in.release ();
3045 out.release ();
3046 return res;
3049 /* Rewrites all the commutative reductions from LOOP out of SSA.
3050 Returns true when something has been changed. */
3052 static bool
3053 rewrite_commutative_reductions_out_of_ssa_loop (scop_p scop,
3054 loop_p loop)
3056 gimple_stmt_iterator gsi;
3057 edge exit = single_exit (loop);
3058 tree res;
3059 bool changed = false;
3061 if (!exit)
3062 return false;
3064 for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi))
3065 if ((res = gimple_phi_result (gsi_stmt (gsi)))
3066 && !virtual_operand_p (res)
3067 && !scev_analyzable_p (res, SCOP_REGION (scop)))
3068 changed |= rewrite_commutative_reductions_out_of_ssa_close_phi
3069 (scop, gsi_stmt (gsi));
3071 return changed;
3074 /* Rewrites all the commutative reductions from SCOP out of SSA. */
3076 static void
3077 rewrite_commutative_reductions_out_of_ssa (scop_p scop)
3079 loop_iterator li;
3080 loop_p loop;
3081 bool changed = false;
3082 sese region = SCOP_REGION (scop);
3084 FOR_EACH_LOOP (li, loop, 0)
3085 if (loop_in_sese_p (loop, region))
3086 changed |= rewrite_commutative_reductions_out_of_ssa_loop (scop, loop);
3088 if (changed)
3090 scev_reset_htab ();
3091 gsi_commit_edge_inserts ();
3092 update_ssa (TODO_update_ssa);
3093 #ifdef ENABLE_CHECKING
3094 verify_loop_closed_ssa (true);
3095 #endif
3099 /* Can all ivs be represented by a signed integer?
3100 As CLooG might generate negative values in its expressions, signed loop ivs
3101 are required in the backend. */
3103 static bool
3104 scop_ivs_can_be_represented (scop_p scop)
3106 loop_iterator li;
3107 loop_p loop;
3108 gimple_stmt_iterator psi;
3109 bool result = true;
3111 FOR_EACH_LOOP (li, loop, 0)
3113 if (!loop_in_sese_p (loop, SCOP_REGION (scop)))
3114 continue;
3116 for (psi = gsi_start_phis (loop->header);
3117 !gsi_end_p (psi); gsi_next (&psi))
3119 gimple phi = gsi_stmt (psi);
3120 tree res = PHI_RESULT (phi);
3121 tree type = TREE_TYPE (res);
3123 if (TYPE_UNSIGNED (type)
3124 && TYPE_PRECISION (type) >= TYPE_PRECISION (long_long_integer_type_node))
3126 result = false;
3127 break;
3130 if (!result)
3131 FOR_EACH_LOOP_BREAK (li);
3134 return result;
3137 /* Builds the polyhedral representation for a SESE region. */
3139 void
3140 build_poly_scop (scop_p scop)
3142 sese region = SCOP_REGION (scop);
3143 graphite_dim_t max_dim;
3145 build_scop_bbs (scop);
3147 /* FIXME: This restriction is needed to avoid a problem in CLooG.
3148 Once CLooG is fixed, remove this guard. Anyways, it makes no
3149 sense to optimize a scop containing only PBBs that do not belong
3150 to any loops. */
3151 if (nb_pbbs_in_loops (scop) == 0)
3152 return;
3154 if (!scop_ivs_can_be_represented (scop))
3155 return;
3157 if (flag_associative_math)
3158 rewrite_commutative_reductions_out_of_ssa (scop);
3160 build_sese_loop_nests (region);
3161 /* Record all conditions in REGION. */
3162 sese_dom_walker (CDI_DOMINATORS, region).walk (cfun->cfg->x_entry_block_ptr);
3163 find_scop_parameters (scop);
3165 max_dim = PARAM_VALUE (PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS);
3166 if (scop_nb_params (scop) > max_dim)
3167 return;
3169 build_scop_iteration_domain (scop);
3170 build_scop_context (scop);
3171 add_conditions_to_constraints (scop);
3173 /* Rewrite out of SSA only after having translated the
3174 representation to the polyhedral representation to avoid scev
3175 analysis failures. That means that these functions will insert
3176 new data references that they create in the right place. */
3177 rewrite_reductions_out_of_ssa (scop);
3178 rewrite_cross_bb_scalar_deps_out_of_ssa (scop);
3180 build_scop_drs (scop);
3181 scop_to_lst (scop);
3182 build_scop_scattering (scop);
3184 /* This SCoP has been translated to the polyhedral
3185 representation. */
3186 POLY_SCOP_P (scop) = true;
3188 #endif