Remove a layer of indirection from hash_table
[gcc.git] / gcc / graphite-sese-to-poly.c
1 /* Conversion of SESE regions to Polyhedra.
2 Copyright (C) 2009-2014 Free Software Foundation, Inc.
3 Contributed by Sebastian Pop <sebastian.pop@amd.com>.
4
5 This file is part of GCC.
6
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.
11
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.
16
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/>. */
20
21 #include "config.h"
22
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
33
34 #include "system.h"
35 #include "coretypes.h"
36 #include "tree.h"
37 #include "basic-block.h"
38 #include "tree-ssa-alias.h"
39 #include "internal-fn.h"
40 #include "gimple-expr.h"
41 #include "is-a.h"
42 #include "gimple.h"
43 #include "gimple-iterator.h"
44 #include "gimplify.h"
45 #include "gimplify-me.h"
46 #include "gimple-ssa.h"
47 #include "tree-cfg.h"
48 #include "tree-phinodes.h"
49 #include "ssa-iterators.h"
50 #include "stringpool.h"
51 #include "tree-ssanames.h"
52 #include "tree-ssa-loop-manip.h"
53 #include "tree-ssa-loop-niter.h"
54 #include "tree-ssa-loop.h"
55 #include "tree-into-ssa.h"
56 #include "tree-pass.h"
57 #include "cfgloop.h"
58 #include "tree-chrec.h"
59 #include "tree-data-ref.h"
60 #include "tree-scalar-evolution.h"
61 #include "domwalk.h"
62 #include "sese.h"
63 #include "tree-ssa-propagate.h"
64
65 #ifdef HAVE_cloog
66 #include "expr.h"
67 #include "graphite-poly.h"
68 #include "graphite-sese-to-poly.h"
69
70
71 /* Assigns to RES the value of the INTEGER_CST T. */
72
73 static inline void
74 tree_int_to_gmp (tree t, mpz_t res)
75 {
76 wi::to_mpz (t, res, TYPE_SIGN (TREE_TYPE (t)));
77 }
78
79 /* Returns the index of the PHI argument defined in the outermost
80 loop. */
81
82 static size_t
83 phi_arg_in_outermost_loop (gimple phi)
84 {
85 loop_p loop = gimple_bb (phi)->loop_father;
86 size_t i, res = 0;
87
88 for (i = 0; i < gimple_phi_num_args (phi); i++)
89 if (!flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, i)->src))
90 {
91 loop = gimple_phi_arg_edge (phi, i)->src->loop_father;
92 res = i;
93 }
94
95 return res;
96 }
97
98 /* Removes a simple copy phi node "RES = phi (INIT, RES)" at position
99 PSI by inserting on the loop ENTRY edge assignment "RES = INIT". */
100
101 static void
102 remove_simple_copy_phi (gimple_stmt_iterator *psi)
103 {
104 gimple phi = gsi_stmt (*psi);
105 tree res = gimple_phi_result (phi);
106 size_t entry = phi_arg_in_outermost_loop (phi);
107 tree init = gimple_phi_arg_def (phi, entry);
108 gimple stmt = gimple_build_assign (res, init);
109 edge e = gimple_phi_arg_edge (phi, entry);
110
111 remove_phi_node (psi, false);
112 gsi_insert_on_edge_immediate (e, stmt);
113 }
114
115 /* Removes an invariant phi node at position PSI by inserting on the
116 loop ENTRY edge the assignment RES = INIT. */
117
118 static void
119 remove_invariant_phi (sese region, gimple_stmt_iterator *psi)
120 {
121 gimple phi = gsi_stmt (*psi);
122 loop_p loop = loop_containing_stmt (phi);
123 tree res = gimple_phi_result (phi);
124 tree scev = scalar_evolution_in_region (region, loop, res);
125 size_t entry = phi_arg_in_outermost_loop (phi);
126 edge e = gimple_phi_arg_edge (phi, entry);
127 tree var;
128 gimple stmt;
129 gimple_seq stmts = NULL;
130
131 if (tree_contains_chrecs (scev, NULL))
132 scev = gimple_phi_arg_def (phi, entry);
133
134 var = force_gimple_operand (scev, &stmts, true, NULL_TREE);
135 stmt = gimple_build_assign (res, var);
136 remove_phi_node (psi, false);
137
138 gimple_seq_add_stmt (&stmts, stmt);
139 gsi_insert_seq_on_edge (e, stmts);
140 gsi_commit_edge_inserts ();
141 SSA_NAME_DEF_STMT (res) = stmt;
142 }
143
144 /* Returns true when the phi node at PSI is of the form "a = phi (a, x)". */
145
146 static inline bool
147 simple_copy_phi_p (gimple phi)
148 {
149 tree res;
150
151 if (gimple_phi_num_args (phi) != 2)
152 return false;
153
154 res = gimple_phi_result (phi);
155 return (res == gimple_phi_arg_def (phi, 0)
156 || res == gimple_phi_arg_def (phi, 1));
157 }
158
159 /* Returns true when the phi node at position PSI is a reduction phi
160 node in REGION. Otherwise moves the pointer PSI to the next phi to
161 be considered. */
162
163 static bool
164 reduction_phi_p (sese region, gimple_stmt_iterator *psi)
165 {
166 loop_p loop;
167 gimple phi = gsi_stmt (*psi);
168 tree res = gimple_phi_result (phi);
169
170 loop = loop_containing_stmt (phi);
171
172 if (simple_copy_phi_p (phi))
173 {
174 /* PRE introduces phi nodes like these, for an example,
175 see id-5.f in the fortran graphite testsuite:
176
177 # prephitmp.85_265 = PHI <prephitmp.85_258(33), prephitmp.85_265(18)>
178 */
179 remove_simple_copy_phi (psi);
180 return false;
181 }
182
183 if (scev_analyzable_p (res, region))
184 {
185 tree scev = scalar_evolution_in_region (region, loop, res);
186
187 if (evolution_function_is_invariant_p (scev, loop->num))
188 remove_invariant_phi (region, psi);
189 else
190 gsi_next (psi);
191
192 return false;
193 }
194
195 /* All the other cases are considered reductions. */
196 return true;
197 }
198
199 /* Store the GRAPHITE representation of BB. */
200
201 static gimple_bb_p
202 new_gimple_bb (basic_block bb, vec<data_reference_p> drs)
203 {
204 struct gimple_bb *gbb;
205
206 gbb = XNEW (struct gimple_bb);
207 bb->aux = gbb;
208 GBB_BB (gbb) = bb;
209 GBB_DATA_REFS (gbb) = drs;
210 GBB_CONDITIONS (gbb).create (0);
211 GBB_CONDITION_CASES (gbb).create (0);
212
213 return gbb;
214 }
215
216 static void
217 free_data_refs_aux (vec<data_reference_p> datarefs)
218 {
219 unsigned int i;
220 struct data_reference *dr;
221
222 FOR_EACH_VEC_ELT (datarefs, i, dr)
223 if (dr->aux)
224 {
225 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
226
227 free (bap->alias_set);
228
229 free (bap);
230 dr->aux = NULL;
231 }
232 }
233 /* Frees GBB. */
234
235 static void
236 free_gimple_bb (struct gimple_bb *gbb)
237 {
238 free_data_refs_aux (GBB_DATA_REFS (gbb));
239 free_data_refs (GBB_DATA_REFS (gbb));
240
241 GBB_CONDITIONS (gbb).release ();
242 GBB_CONDITION_CASES (gbb).release ();
243 GBB_BB (gbb)->aux = 0;
244 XDELETE (gbb);
245 }
246
247 /* Deletes all gimple bbs in SCOP. */
248
249 static void
250 remove_gbbs_in_scop (scop_p scop)
251 {
252 int i;
253 poly_bb_p pbb;
254
255 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
256 free_gimple_bb (PBB_BLACK_BOX (pbb));
257 }
258
259 /* Deletes all scops in SCOPS. */
260
261 void
262 free_scops (vec<scop_p> scops)
263 {
264 int i;
265 scop_p scop;
266
267 FOR_EACH_VEC_ELT (scops, i, scop)
268 {
269 remove_gbbs_in_scop (scop);
270 free_sese (SCOP_REGION (scop));
271 free_scop (scop);
272 }
273
274 scops.release ();
275 }
276
277 /* Same as outermost_loop_in_sese, returns the outermost loop
278 containing BB in REGION, but makes sure that the returned loop
279 belongs to the REGION, and so this returns the first loop in the
280 REGION when the loop containing BB does not belong to REGION. */
281
282 static loop_p
283 outermost_loop_in_sese_1 (sese region, basic_block bb)
284 {
285 loop_p nest = outermost_loop_in_sese (region, bb);
286
287 if (loop_in_sese_p (nest, region))
288 return nest;
289
290 /* When the basic block BB does not belong to a loop in the region,
291 return the first loop in the region. */
292 nest = nest->inner;
293 while (nest)
294 if (loop_in_sese_p (nest, region))
295 break;
296 else
297 nest = nest->next;
298
299 gcc_assert (nest);
300 return nest;
301 }
302
303 /* Generates a polyhedral black box only if the bb contains interesting
304 information. */
305
306 static gimple_bb_p
307 try_generate_gimple_bb (scop_p scop, basic_block bb)
308 {
309 vec<data_reference_p> drs;
310 drs.create (5);
311 sese region = SCOP_REGION (scop);
312 loop_p nest = outermost_loop_in_sese_1 (region, bb);
313 gimple_stmt_iterator gsi;
314
315 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
316 {
317 gimple stmt = gsi_stmt (gsi);
318 loop_p loop;
319
320 if (is_gimple_debug (stmt))
321 continue;
322
323 loop = loop_containing_stmt (stmt);
324 if (!loop_in_sese_p (loop, region))
325 loop = nest;
326
327 graphite_find_data_references_in_stmt (nest, loop, stmt, &drs);
328 }
329
330 return new_gimple_bb (bb, drs);
331 }
332
333 /* Returns true if all predecessors of BB, that are not dominated by BB, are
334 marked in MAP. The predecessors dominated by BB are loop latches and will
335 be handled after BB. */
336
337 static bool
338 all_non_dominated_preds_marked_p (basic_block bb, sbitmap map)
339 {
340 edge e;
341 edge_iterator ei;
342
343 FOR_EACH_EDGE (e, ei, bb->preds)
344 if (!bitmap_bit_p (map, e->src->index)
345 && !dominated_by_p (CDI_DOMINATORS, e->src, bb))
346 return false;
347
348 return true;
349 }
350
351 /* Compare the depth of two basic_block's P1 and P2. */
352
353 static int
354 compare_bb_depths (const void *p1, const void *p2)
355 {
356 const_basic_block const bb1 = *(const_basic_block const*)p1;
357 const_basic_block const bb2 = *(const_basic_block const*)p2;
358 int d1 = loop_depth (bb1->loop_father);
359 int d2 = loop_depth (bb2->loop_father);
360
361 if (d1 < d2)
362 return 1;
363
364 if (d1 > d2)
365 return -1;
366
367 return 0;
368 }
369
370 /* Sort the basic blocks from DOM such that the first are the ones at
371 a deepest loop level. */
372
373 static void
374 graphite_sort_dominated_info (vec<basic_block> dom)
375 {
376 dom.qsort (compare_bb_depths);
377 }
378
379 /* Recursive helper function for build_scops_bbs. */
380
381 static void
382 build_scop_bbs_1 (scop_p scop, sbitmap visited, basic_block bb)
383 {
384 sese region = SCOP_REGION (scop);
385 vec<basic_block> dom;
386 poly_bb_p pbb;
387
388 if (bitmap_bit_p (visited, bb->index)
389 || !bb_in_sese_p (bb, region))
390 return;
391
392 pbb = new_poly_bb (scop, try_generate_gimple_bb (scop, bb));
393 SCOP_BBS (scop).safe_push (pbb);
394 bitmap_set_bit (visited, bb->index);
395
396 dom = get_dominated_by (CDI_DOMINATORS, bb);
397
398 if (!dom.exists ())
399 return;
400
401 graphite_sort_dominated_info (dom);
402
403 while (!dom.is_empty ())
404 {
405 int i;
406 basic_block dom_bb;
407
408 FOR_EACH_VEC_ELT (dom, i, dom_bb)
409 if (all_non_dominated_preds_marked_p (dom_bb, visited))
410 {
411 build_scop_bbs_1 (scop, visited, dom_bb);
412 dom.unordered_remove (i);
413 break;
414 }
415 }
416
417 dom.release ();
418 }
419
420 /* Gather the basic blocks belonging to the SCOP. */
421
422 static void
423 build_scop_bbs (scop_p scop)
424 {
425 sbitmap visited = sbitmap_alloc (last_basic_block_for_fn (cfun));
426 sese region = SCOP_REGION (scop);
427
428 bitmap_clear (visited);
429 build_scop_bbs_1 (scop, visited, SESE_ENTRY_BB (region));
430 sbitmap_free (visited);
431 }
432
433 /* Return an ISL identifier for the polyhedral basic block PBB. */
434
435 static isl_id *
436 isl_id_for_pbb (scop_p s, poly_bb_p pbb)
437 {
438 char name[50];
439 snprintf (name, sizeof (name), "S_%d", pbb_index (pbb));
440 return isl_id_alloc (s->ctx, name, pbb);
441 }
442
443 /* Converts the STATIC_SCHEDULE of PBB into a scattering polyhedron.
444 We generate SCATTERING_DIMENSIONS scattering dimensions.
445
446 CLooG 0.15.0 and previous versions require, that all
447 scattering functions of one CloogProgram have the same number of
448 scattering dimensions, therefore we allow to specify it. This
449 should be removed in future versions of CLooG.
450
451 The scattering polyhedron consists of these dimensions: scattering,
452 loop_iterators, parameters.
453
454 Example:
455
456 | scattering_dimensions = 5
457 | used_scattering_dimensions = 3
458 | nb_iterators = 1
459 | scop_nb_params = 2
460 |
461 | Schedule:
462 | i
463 | 4 5
464 |
465 | Scattering polyhedron:
466 |
467 | scattering: {s1, s2, s3, s4, s5}
468 | loop_iterators: {i}
469 | parameters: {p1, p2}
470 |
471 | s1 s2 s3 s4 s5 i p1 p2 1
472 | 1 0 0 0 0 0 0 0 -4 = 0
473 | 0 1 0 0 0 -1 0 0 0 = 0
474 | 0 0 1 0 0 0 0 0 -5 = 0 */
475
476 static void
477 build_pbb_scattering_polyhedrons (isl_aff *static_sched,
478 poly_bb_p pbb, int scattering_dimensions)
479 {
480 int i;
481 int nb_iterators = pbb_dim_iter_domain (pbb);
482 int used_scattering_dimensions = nb_iterators * 2 + 1;
483 isl_int val;
484 isl_space *dc, *dm;
485
486 gcc_assert (scattering_dimensions >= used_scattering_dimensions);
487
488 isl_int_init (val);
489
490 dc = isl_set_get_space (pbb->domain);
491 dm = isl_space_add_dims (isl_space_from_domain (dc),
492 isl_dim_out, scattering_dimensions);
493 pbb->schedule = isl_map_universe (dm);
494
495 for (i = 0; i < scattering_dimensions; i++)
496 {
497 /* Textual order inside this loop. */
498 if ((i % 2) == 0)
499 {
500 isl_constraint *c = isl_equality_alloc
501 (isl_local_space_from_space (isl_map_get_space (pbb->schedule)));
502
503 if (0 != isl_aff_get_coefficient (static_sched, isl_dim_in,
504 i / 2, &val))
505 gcc_unreachable ();
506
507 isl_int_neg (val, val);
508 c = isl_constraint_set_constant (c, val);
509 c = isl_constraint_set_coefficient_si (c, isl_dim_out, i, 1);
510 pbb->schedule = isl_map_add_constraint (pbb->schedule, c);
511 }
512
513 /* Iterations of this loop. */
514 else /* if ((i % 2) == 1) */
515 {
516 int loop = (i - 1) / 2;
517 pbb->schedule = isl_map_equate (pbb->schedule, isl_dim_in, loop,
518 isl_dim_out, i);
519 }
520 }
521
522 isl_int_clear (val);
523
524 pbb->transformed = isl_map_copy (pbb->schedule);
525 }
526
527 /* Build for BB the static schedule.
528
529 The static schedule is a Dewey numbering of the abstract syntax
530 tree: http://en.wikipedia.org/wiki/Dewey_Decimal_Classification
531
532 The following example informally defines the static schedule:
533
534 A
535 for (i: ...)
536 {
537 for (j: ...)
538 {
539 B
540 C
541 }
542
543 for (k: ...)
544 {
545 D
546 E
547 }
548 }
549 F
550
551 Static schedules for A to F:
552
553 DEPTH
554 0 1 2
555 A 0
556 B 1 0 0
557 C 1 0 1
558 D 1 1 0
559 E 1 1 1
560 F 2
561 */
562
563 static void
564 build_scop_scattering (scop_p scop)
565 {
566 int i;
567 poly_bb_p pbb;
568 gimple_bb_p previous_gbb = NULL;
569 isl_space *dc = isl_set_get_space (scop->context);
570 isl_aff *static_sched;
571
572 dc = isl_space_add_dims (dc, isl_dim_set, number_of_loops (cfun));
573 static_sched = isl_aff_zero_on_domain (isl_local_space_from_space (dc));
574
575 /* We have to start schedules at 0 on the first component and
576 because we cannot compare_prefix_loops against a previous loop,
577 prefix will be equal to zero, and that index will be
578 incremented before copying. */
579 static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in, 0, -1);
580
581 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
582 {
583 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
584 int prefix;
585 int nb_scat_dims = pbb_dim_iter_domain (pbb) * 2 + 1;
586
587 if (previous_gbb)
588 prefix = nb_common_loops (SCOP_REGION (scop), previous_gbb, gbb);
589 else
590 prefix = 0;
591
592 previous_gbb = gbb;
593
594 static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in,
595 prefix, 1);
596 build_pbb_scattering_polyhedrons (static_sched, pbb, nb_scat_dims);
597 }
598
599 isl_aff_free (static_sched);
600 }
601
602 static isl_pw_aff *extract_affine (scop_p, tree, __isl_take isl_space *space);
603
604 /* Extract an affine expression from the chain of recurrence E. */
605
606 static isl_pw_aff *
607 extract_affine_chrec (scop_p s, tree e, __isl_take isl_space *space)
608 {
609 isl_pw_aff *lhs = extract_affine (s, CHREC_LEFT (e), isl_space_copy (space));
610 isl_pw_aff *rhs = extract_affine (s, CHREC_RIGHT (e), isl_space_copy (space));
611 isl_local_space *ls = isl_local_space_from_space (space);
612 unsigned pos = sese_loop_depth ((sese) s->region, get_chrec_loop (e)) - 1;
613 isl_aff *loop = isl_aff_set_coefficient_si
614 (isl_aff_zero_on_domain (ls), isl_dim_in, pos, 1);
615 isl_pw_aff *l = isl_pw_aff_from_aff (loop);
616
617 /* Before multiplying, make sure that the result is affine. */
618 gcc_assert (isl_pw_aff_is_cst (rhs)
619 || isl_pw_aff_is_cst (l));
620
621 return isl_pw_aff_add (lhs, isl_pw_aff_mul (rhs, l));
622 }
623
624 /* Extract an affine expression from the mult_expr E. */
625
626 static isl_pw_aff *
627 extract_affine_mul (scop_p s, tree e, __isl_take isl_space *space)
628 {
629 isl_pw_aff *lhs = extract_affine (s, TREE_OPERAND (e, 0),
630 isl_space_copy (space));
631 isl_pw_aff *rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
632
633 if (!isl_pw_aff_is_cst (lhs)
634 && !isl_pw_aff_is_cst (rhs))
635 {
636 isl_pw_aff_free (lhs);
637 isl_pw_aff_free (rhs);
638 return NULL;
639 }
640
641 return isl_pw_aff_mul (lhs, rhs);
642 }
643
644 /* Return an ISL identifier from the name of the ssa_name E. */
645
646 static isl_id *
647 isl_id_for_ssa_name (scop_p s, tree e)
648 {
649 const char *name = get_name (e);
650 isl_id *id;
651
652 if (name)
653 id = isl_id_alloc (s->ctx, name, e);
654 else
655 {
656 char name1[50];
657 snprintf (name1, sizeof (name1), "P_%d", SSA_NAME_VERSION (e));
658 id = isl_id_alloc (s->ctx, name1, e);
659 }
660
661 return id;
662 }
663
664 /* Return an ISL identifier for the data reference DR. */
665
666 static isl_id *
667 isl_id_for_dr (scop_p s, data_reference_p dr ATTRIBUTE_UNUSED)
668 {
669 /* Data references all get the same isl_id. They need to be comparable
670 and are distinguished through the first dimension, which contains the
671 alias set number. */
672 return isl_id_alloc (s->ctx, "", 0);
673 }
674
675 /* Extract an affine expression from the ssa_name E. */
676
677 static isl_pw_aff *
678 extract_affine_name (scop_p s, tree e, __isl_take isl_space *space)
679 {
680 isl_aff *aff;
681 isl_set *dom;
682 isl_id *id;
683 int dimension;
684
685 id = isl_id_for_ssa_name (s, e);
686 dimension = isl_space_find_dim_by_id (space, isl_dim_param, id);
687 isl_id_free (id);
688 dom = isl_set_universe (isl_space_copy (space));
689 aff = isl_aff_zero_on_domain (isl_local_space_from_space (space));
690 aff = isl_aff_add_coefficient_si (aff, isl_dim_param, dimension, 1);
691 return isl_pw_aff_alloc (dom, aff);
692 }
693
694 /* Extract an affine expression from the gmp constant G. */
695
696 static isl_pw_aff *
697 extract_affine_gmp (mpz_t g, __isl_take isl_space *space)
698 {
699 isl_local_space *ls = isl_local_space_from_space (isl_space_copy (space));
700 isl_aff *aff = isl_aff_zero_on_domain (ls);
701 isl_set *dom = isl_set_universe (space);
702 isl_int v;
703
704 isl_int_init (v);
705 isl_int_set_gmp (v, g);
706 aff = isl_aff_add_constant (aff, v);
707 isl_int_clear (v);
708
709 return isl_pw_aff_alloc (dom, aff);
710 }
711
712 /* Extract an affine expression from the integer_cst E. */
713
714 static isl_pw_aff *
715 extract_affine_int (tree e, __isl_take isl_space *space)
716 {
717 isl_pw_aff *res;
718 mpz_t g;
719
720 mpz_init (g);
721 tree_int_to_gmp (e, g);
722 res = extract_affine_gmp (g, space);
723 mpz_clear (g);
724
725 return res;
726 }
727
728 /* Compute pwaff mod 2^width. */
729
730 static isl_pw_aff *
731 wrap (isl_pw_aff *pwaff, unsigned width)
732 {
733 isl_int mod;
734
735 isl_int_init (mod);
736 isl_int_set_si (mod, 1);
737 isl_int_mul_2exp (mod, mod, width);
738
739 pwaff = isl_pw_aff_mod (pwaff, mod);
740
741 isl_int_clear (mod);
742
743 return pwaff;
744 }
745
746 /* When parameter NAME is in REGION, returns its index in SESE_PARAMS.
747 Otherwise returns -1. */
748
749 static inline int
750 parameter_index_in_region_1 (tree name, sese region)
751 {
752 int i;
753 tree p;
754
755 gcc_assert (TREE_CODE (name) == SSA_NAME);
756
757 FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, p)
758 if (p == name)
759 return i;
760
761 return -1;
762 }
763
764 /* When the parameter NAME is in REGION, returns its index in
765 SESE_PARAMS. Otherwise this function inserts NAME in SESE_PARAMS
766 and returns the index of NAME. */
767
768 static int
769 parameter_index_in_region (tree name, sese region)
770 {
771 int i;
772
773 gcc_assert (TREE_CODE (name) == SSA_NAME);
774
775 i = parameter_index_in_region_1 (name, region);
776 if (i != -1)
777 return i;
778
779 gcc_assert (SESE_ADD_PARAMS (region));
780
781 i = SESE_PARAMS (region).length ();
782 SESE_PARAMS (region).safe_push (name);
783 return i;
784 }
785
786 /* Extract an affine expression from the tree E in the scop S. */
787
788 static isl_pw_aff *
789 extract_affine (scop_p s, tree e, __isl_take isl_space *space)
790 {
791 isl_pw_aff *lhs, *rhs, *res;
792 tree type;
793
794 if (e == chrec_dont_know) {
795 isl_space_free (space);
796 return NULL;
797 }
798
799 switch (TREE_CODE (e))
800 {
801 case POLYNOMIAL_CHREC:
802 res = extract_affine_chrec (s, e, space);
803 break;
804
805 case MULT_EXPR:
806 res = extract_affine_mul (s, e, space);
807 break;
808
809 case PLUS_EXPR:
810 case POINTER_PLUS_EXPR:
811 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
812 rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
813 res = isl_pw_aff_add (lhs, rhs);
814 break;
815
816 case MINUS_EXPR:
817 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
818 rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
819 res = isl_pw_aff_sub (lhs, rhs);
820 break;
821
822 case NEGATE_EXPR:
823 case BIT_NOT_EXPR:
824 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
825 rhs = extract_affine (s, integer_minus_one_node, space);
826 res = isl_pw_aff_mul (lhs, rhs);
827 break;
828
829 case SSA_NAME:
830 gcc_assert (-1 != parameter_index_in_region_1 (e, SCOP_REGION (s)));
831 res = extract_affine_name (s, e, space);
832 break;
833
834 case INTEGER_CST:
835 res = extract_affine_int (e, space);
836 /* No need to wrap a single integer. */
837 return res;
838
839 CASE_CONVERT:
840 case NON_LVALUE_EXPR:
841 res = extract_affine (s, TREE_OPERAND (e, 0), space);
842 break;
843
844 default:
845 gcc_unreachable ();
846 break;
847 }
848
849 type = TREE_TYPE (e);
850 if (TYPE_UNSIGNED (type))
851 res = wrap (res, TYPE_PRECISION (type));
852
853 return res;
854 }
855
856 /* In the context of sese S, scan the expression E and translate it to
857 a linear expression C. When parsing a symbolic multiplication, K
858 represents the constant multiplier of an expression containing
859 parameters. */
860
861 static void
862 scan_tree_for_params (sese s, tree e)
863 {
864 if (e == chrec_dont_know)
865 return;
866
867 switch (TREE_CODE (e))
868 {
869 case POLYNOMIAL_CHREC:
870 scan_tree_for_params (s, CHREC_LEFT (e));
871 break;
872
873 case MULT_EXPR:
874 if (chrec_contains_symbols (TREE_OPERAND (e, 0)))
875 scan_tree_for_params (s, TREE_OPERAND (e, 0));
876 else
877 scan_tree_for_params (s, TREE_OPERAND (e, 1));
878 break;
879
880 case PLUS_EXPR:
881 case POINTER_PLUS_EXPR:
882 case MINUS_EXPR:
883 scan_tree_for_params (s, TREE_OPERAND (e, 0));
884 scan_tree_for_params (s, TREE_OPERAND (e, 1));
885 break;
886
887 case NEGATE_EXPR:
888 case BIT_NOT_EXPR:
889 CASE_CONVERT:
890 case NON_LVALUE_EXPR:
891 scan_tree_for_params (s, TREE_OPERAND (e, 0));
892 break;
893
894 case SSA_NAME:
895 parameter_index_in_region (e, s);
896 break;
897
898 case INTEGER_CST:
899 case ADDR_EXPR:
900 break;
901
902 default:
903 gcc_unreachable ();
904 break;
905 }
906 }
907
908 /* Find parameters with respect to REGION in BB. We are looking in memory
909 access functions, conditions and loop bounds. */
910
911 static void
912 find_params_in_bb (sese region, gimple_bb_p gbb)
913 {
914 int i;
915 unsigned j;
916 data_reference_p dr;
917 gimple stmt;
918 loop_p loop = GBB_BB (gbb)->loop_father;
919
920 /* Find parameters in the access functions of data references. */
921 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr)
922 for (j = 0; j < DR_NUM_DIMENSIONS (dr); j++)
923 scan_tree_for_params (region, DR_ACCESS_FN (dr, j));
924
925 /* Find parameters in conditional statements. */
926 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt)
927 {
928 tree lhs = scalar_evolution_in_region (region, loop,
929 gimple_cond_lhs (stmt));
930 tree rhs = scalar_evolution_in_region (region, loop,
931 gimple_cond_rhs (stmt));
932
933 scan_tree_for_params (region, lhs);
934 scan_tree_for_params (region, rhs);
935 }
936 }
937
938 /* Record the parameters used in the SCOP. A variable is a parameter
939 in a scop if it does not vary during the execution of that scop. */
940
941 static void
942 find_scop_parameters (scop_p scop)
943 {
944 poly_bb_p pbb;
945 unsigned i;
946 sese region = SCOP_REGION (scop);
947 struct loop *loop;
948 int nbp;
949
950 /* Find the parameters used in the loop bounds. */
951 FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop)
952 {
953 tree nb_iters = number_of_latch_executions (loop);
954
955 if (!chrec_contains_symbols (nb_iters))
956 continue;
957
958 nb_iters = scalar_evolution_in_region (region, loop, nb_iters);
959 scan_tree_for_params (region, nb_iters);
960 }
961
962 /* Find the parameters used in data accesses. */
963 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
964 find_params_in_bb (region, PBB_BLACK_BOX (pbb));
965
966 nbp = sese_nb_params (region);
967 scop_set_nb_params (scop, nbp);
968 SESE_ADD_PARAMS (region) = false;
969
970 {
971 tree e;
972 isl_space *space = isl_space_set_alloc (scop->ctx, nbp, 0);
973
974 FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, e)
975 space = isl_space_set_dim_id (space, isl_dim_param, i,
976 isl_id_for_ssa_name (scop, e));
977
978 scop->context = isl_set_universe (space);
979 }
980 }
981
982 /* Builds the constraint polyhedra for LOOP in SCOP. OUTER_PH gives
983 the constraints for the surrounding loops. */
984
985 static void
986 build_loop_iteration_domains (scop_p scop, struct loop *loop,
987 int nb,
988 isl_set *outer, isl_set **doms)
989 {
990 tree nb_iters = number_of_latch_executions (loop);
991 sese region = SCOP_REGION (scop);
992
993 isl_set *inner = isl_set_copy (outer);
994 isl_space *space;
995 isl_constraint *c;
996 int pos = isl_set_dim (outer, isl_dim_set);
997 isl_int v;
998 mpz_t g;
999
1000 mpz_init (g);
1001 isl_int_init (v);
1002
1003 inner = isl_set_add_dims (inner, isl_dim_set, 1);
1004 space = isl_set_get_space (inner);
1005
1006 /* 0 <= loop_i */
1007 c = isl_inequality_alloc
1008 (isl_local_space_from_space (isl_space_copy (space)));
1009 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, 1);
1010 inner = isl_set_add_constraint (inner, c);
1011
1012 /* loop_i <= cst_nb_iters */
1013 if (TREE_CODE (nb_iters) == INTEGER_CST)
1014 {
1015 c = isl_inequality_alloc
1016 (isl_local_space_from_space (isl_space_copy (space)));
1017 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1);
1018 tree_int_to_gmp (nb_iters, g);
1019 isl_int_set_gmp (v, g);
1020 c = isl_constraint_set_constant (c, v);
1021 inner = isl_set_add_constraint (inner, c);
1022 }
1023
1024 /* loop_i <= expr_nb_iters */
1025 else if (!chrec_contains_undetermined (nb_iters))
1026 {
1027 widest_int nit;
1028 isl_pw_aff *aff;
1029 isl_set *valid;
1030 isl_local_space *ls;
1031 isl_aff *al;
1032 isl_set *le;
1033
1034 nb_iters = scalar_evolution_in_region (region, loop, nb_iters);
1035
1036 aff = extract_affine (scop, nb_iters, isl_set_get_space (inner));
1037 valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (aff));
1038 valid = isl_set_project_out (valid, isl_dim_set, 0,
1039 isl_set_dim (valid, isl_dim_set));
1040 scop->context = isl_set_intersect (scop->context, valid);
1041
1042 ls = isl_local_space_from_space (isl_space_copy (space));
1043 al = isl_aff_set_coefficient_si (isl_aff_zero_on_domain (ls),
1044 isl_dim_in, pos, 1);
1045 le = isl_pw_aff_le_set (isl_pw_aff_from_aff (al),
1046 isl_pw_aff_copy (aff));
1047 inner = isl_set_intersect (inner, le);
1048
1049 if (max_stmt_executions (loop, &nit))
1050 {
1051 /* Insert in the context the constraints from the
1052 estimation of the number of iterations NIT and the
1053 symbolic number of iterations (involving parameter
1054 names) NB_ITERS. First, build the affine expression
1055 "NIT - NB_ITERS" and then say that it is positive,
1056 i.e., NIT approximates NB_ITERS: "NIT >= NB_ITERS". */
1057 isl_pw_aff *approx;
1058 mpz_t g;
1059 isl_set *x;
1060 isl_constraint *c;
1061
1062 mpz_init (g);
1063 wi::to_mpz (nit, g, SIGNED);
1064 mpz_sub_ui (g, g, 1);
1065 approx = extract_affine_gmp (g, isl_set_get_space (inner));
1066 x = isl_pw_aff_ge_set (approx, aff);
1067 x = isl_set_project_out (x, isl_dim_set, 0,
1068 isl_set_dim (x, isl_dim_set));
1069 scop->context = isl_set_intersect (scop->context, x);
1070
1071 c = isl_inequality_alloc
1072 (isl_local_space_from_space (isl_space_copy (space)));
1073 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1);
1074 isl_int_set_gmp (v, g);
1075 mpz_clear (g);
1076 c = isl_constraint_set_constant (c, v);
1077 inner = isl_set_add_constraint (inner, c);
1078 }
1079 else
1080 isl_pw_aff_free (aff);
1081 }
1082 else
1083 gcc_unreachable ();
1084
1085 if (loop->inner && loop_in_sese_p (loop->inner, region))
1086 build_loop_iteration_domains (scop, loop->inner, nb + 1,
1087 isl_set_copy (inner), doms);
1088
1089 if (nb != 0
1090 && loop->next
1091 && loop_in_sese_p (loop->next, region))
1092 build_loop_iteration_domains (scop, loop->next, nb,
1093 isl_set_copy (outer), doms);
1094
1095 doms[loop->num] = inner;
1096
1097 isl_set_free (outer);
1098 isl_space_free (space);
1099 isl_int_clear (v);
1100 mpz_clear (g);
1101 }
1102
1103 /* Returns a linear expression for tree T evaluated in PBB. */
1104
1105 static isl_pw_aff *
1106 create_pw_aff_from_tree (poly_bb_p pbb, tree t)
1107 {
1108 scop_p scop = PBB_SCOP (pbb);
1109
1110 t = scalar_evolution_in_region (SCOP_REGION (scop), pbb_loop (pbb), t);
1111 gcc_assert (!automatically_generated_chrec_p (t));
1112
1113 return extract_affine (scop, t, isl_set_get_space (pbb->domain));
1114 }
1115
1116 /* Add conditional statement STMT to pbb. CODE is used as the comparison
1117 operator. This allows us to invert the condition or to handle
1118 inequalities. */
1119
1120 static void
1121 add_condition_to_pbb (poly_bb_p pbb, gimple stmt, enum tree_code code)
1122 {
1123 isl_pw_aff *lhs = create_pw_aff_from_tree (pbb, gimple_cond_lhs (stmt));
1124 isl_pw_aff *rhs = create_pw_aff_from_tree (pbb, gimple_cond_rhs (stmt));
1125 isl_set *cond;
1126
1127 switch (code)
1128 {
1129 case LT_EXPR:
1130 cond = isl_pw_aff_lt_set (lhs, rhs);
1131 break;
1132
1133 case GT_EXPR:
1134 cond = isl_pw_aff_gt_set (lhs, rhs);
1135 break;
1136
1137 case LE_EXPR:
1138 cond = isl_pw_aff_le_set (lhs, rhs);
1139 break;
1140
1141 case GE_EXPR:
1142 cond = isl_pw_aff_ge_set (lhs, rhs);
1143 break;
1144
1145 case EQ_EXPR:
1146 cond = isl_pw_aff_eq_set (lhs, rhs);
1147 break;
1148
1149 case NE_EXPR:
1150 cond = isl_pw_aff_ne_set (lhs, rhs);
1151 break;
1152
1153 default:
1154 isl_pw_aff_free (lhs);
1155 isl_pw_aff_free (rhs);
1156 return;
1157 }
1158
1159 cond = isl_set_coalesce (cond);
1160 cond = isl_set_set_tuple_id (cond, isl_set_get_tuple_id (pbb->domain));
1161 pbb->domain = isl_set_intersect (pbb->domain, cond);
1162 }
1163
1164 /* Add conditions to the domain of PBB. */
1165
1166 static void
1167 add_conditions_to_domain (poly_bb_p pbb)
1168 {
1169 unsigned int i;
1170 gimple stmt;
1171 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
1172
1173 if (GBB_CONDITIONS (gbb).is_empty ())
1174 return;
1175
1176 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt)
1177 switch (gimple_code (stmt))
1178 {
1179 case GIMPLE_COND:
1180 {
1181 enum tree_code code = gimple_cond_code (stmt);
1182
1183 /* The conditions for ELSE-branches are inverted. */
1184 if (!GBB_CONDITION_CASES (gbb)[i])
1185 code = invert_tree_comparison (code, false);
1186
1187 add_condition_to_pbb (pbb, stmt, code);
1188 break;
1189 }
1190
1191 case GIMPLE_SWITCH:
1192 /* Switch statements are not supported right now - fall through. */
1193
1194 default:
1195 gcc_unreachable ();
1196 break;
1197 }
1198 }
1199
1200 /* Traverses all the GBBs of the SCOP and add their constraints to the
1201 iteration domains. */
1202
1203 static void
1204 add_conditions_to_constraints (scop_p scop)
1205 {
1206 int i;
1207 poly_bb_p pbb;
1208
1209 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1210 add_conditions_to_domain (pbb);
1211 }
1212
1213 /* Returns a COND_EXPR statement when BB has a single predecessor, the
1214 edge between BB and its predecessor is not a loop exit edge, and
1215 the last statement of the single predecessor is a COND_EXPR. */
1216
1217 static gimple
1218 single_pred_cond_non_loop_exit (basic_block bb)
1219 {
1220 if (single_pred_p (bb))
1221 {
1222 edge e = single_pred_edge (bb);
1223 basic_block pred = e->src;
1224 gimple stmt;
1225
1226 if (loop_depth (pred->loop_father) > loop_depth (bb->loop_father))
1227 return NULL;
1228
1229 stmt = last_stmt (pred);
1230
1231 if (stmt && gimple_code (stmt) == GIMPLE_COND)
1232 return stmt;
1233 }
1234
1235 return NULL;
1236 }
1237
1238 class sese_dom_walker : public dom_walker
1239 {
1240 public:
1241 sese_dom_walker (cdi_direction, sese);
1242
1243 virtual void before_dom_children (basic_block);
1244 virtual void after_dom_children (basic_block);
1245
1246 private:
1247 auto_vec<gimple, 3> m_conditions, m_cases;
1248 sese m_region;
1249 };
1250
1251 sese_dom_walker::sese_dom_walker (cdi_direction direction, sese region)
1252 : dom_walker (direction), m_region (region)
1253 {
1254 }
1255
1256 /* Call-back for dom_walk executed before visiting the dominated
1257 blocks. */
1258
1259 void
1260 sese_dom_walker::before_dom_children (basic_block bb)
1261 {
1262 gimple_bb_p gbb;
1263 gimple stmt;
1264
1265 if (!bb_in_sese_p (bb, m_region))
1266 return;
1267
1268 stmt = single_pred_cond_non_loop_exit (bb);
1269
1270 if (stmt)
1271 {
1272 edge e = single_pred_edge (bb);
1273
1274 m_conditions.safe_push (stmt);
1275
1276 if (e->flags & EDGE_TRUE_VALUE)
1277 m_cases.safe_push (stmt);
1278 else
1279 m_cases.safe_push (NULL);
1280 }
1281
1282 gbb = gbb_from_bb (bb);
1283
1284 if (gbb)
1285 {
1286 GBB_CONDITIONS (gbb) = m_conditions.copy ();
1287 GBB_CONDITION_CASES (gbb) = m_cases.copy ();
1288 }
1289 }
1290
1291 /* Call-back for dom_walk executed after visiting the dominated
1292 blocks. */
1293
1294 void
1295 sese_dom_walker::after_dom_children (basic_block bb)
1296 {
1297 if (!bb_in_sese_p (bb, m_region))
1298 return;
1299
1300 if (single_pred_cond_non_loop_exit (bb))
1301 {
1302 m_conditions.pop ();
1303 m_cases.pop ();
1304 }
1305 }
1306
1307 /* Add constraints on the possible values of parameter P from the type
1308 of P. */
1309
1310 static void
1311 add_param_constraints (scop_p scop, graphite_dim_t p)
1312 {
1313 tree parameter = SESE_PARAMS (SCOP_REGION (scop))[p];
1314 tree type = TREE_TYPE (parameter);
1315 tree lb = NULL_TREE;
1316 tree ub = NULL_TREE;
1317
1318 if (POINTER_TYPE_P (type) || !TYPE_MIN_VALUE (type))
1319 lb = lower_bound_in_type (type, type);
1320 else
1321 lb = TYPE_MIN_VALUE (type);
1322
1323 if (POINTER_TYPE_P (type) || !TYPE_MAX_VALUE (type))
1324 ub = upper_bound_in_type (type, type);
1325 else
1326 ub = TYPE_MAX_VALUE (type);
1327
1328 if (lb)
1329 {
1330 isl_space *space = isl_set_get_space (scop->context);
1331 isl_constraint *c;
1332 mpz_t g;
1333 isl_int v;
1334
1335 c = isl_inequality_alloc (isl_local_space_from_space (space));
1336 mpz_init (g);
1337 isl_int_init (v);
1338 tree_int_to_gmp (lb, g);
1339 isl_int_set_gmp (v, g);
1340 isl_int_neg (v, v);
1341 mpz_clear (g);
1342 c = isl_constraint_set_constant (c, v);
1343 isl_int_clear (v);
1344 c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, 1);
1345
1346 scop->context = isl_set_add_constraint (scop->context, c);
1347 }
1348
1349 if (ub)
1350 {
1351 isl_space *space = isl_set_get_space (scop->context);
1352 isl_constraint *c;
1353 mpz_t g;
1354 isl_int v;
1355
1356 c = isl_inequality_alloc (isl_local_space_from_space (space));
1357
1358 mpz_init (g);
1359 isl_int_init (v);
1360 tree_int_to_gmp (ub, g);
1361 isl_int_set_gmp (v, g);
1362 mpz_clear (g);
1363 c = isl_constraint_set_constant (c, v);
1364 isl_int_clear (v);
1365 c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, -1);
1366
1367 scop->context = isl_set_add_constraint (scop->context, c);
1368 }
1369 }
1370
1371 /* Build the context of the SCOP. The context usually contains extra
1372 constraints that are added to the iteration domains that constrain
1373 some parameters. */
1374
1375 static void
1376 build_scop_context (scop_p scop)
1377 {
1378 graphite_dim_t p, n = scop_nb_params (scop);
1379
1380 for (p = 0; p < n; p++)
1381 add_param_constraints (scop, p);
1382 }
1383
1384 /* Build the iteration domains: the loops belonging to the current
1385 SCOP, and that vary for the execution of the current basic block.
1386 Returns false if there is no loop in SCOP. */
1387
1388 static void
1389 build_scop_iteration_domain (scop_p scop)
1390 {
1391 struct loop *loop;
1392 sese region = SCOP_REGION (scop);
1393 int i;
1394 poly_bb_p pbb;
1395 int nb_loops = number_of_loops (cfun);
1396 isl_set **doms = XCNEWVEC (isl_set *, nb_loops);
1397
1398 FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop)
1399 if (!loop_in_sese_p (loop_outer (loop), region))
1400 build_loop_iteration_domains (scop, loop, 0,
1401 isl_set_copy (scop->context), doms);
1402
1403 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1404 {
1405 loop = pbb_loop (pbb);
1406
1407 if (doms[loop->num])
1408 pbb->domain = isl_set_copy (doms[loop->num]);
1409 else
1410 pbb->domain = isl_set_copy (scop->context);
1411
1412 pbb->domain = isl_set_set_tuple_id (pbb->domain,
1413 isl_id_for_pbb (scop, pbb));
1414 }
1415
1416 for (i = 0; i < nb_loops; i++)
1417 if (doms[i])
1418 isl_set_free (doms[i]);
1419
1420 free (doms);
1421 }
1422
1423 /* Add a constrain to the ACCESSES polyhedron for the alias set of
1424 data reference DR. ACCESSP_NB_DIMS is the dimension of the
1425 ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration
1426 domain. */
1427
1428 static isl_map *
1429 pdr_add_alias_set (isl_map *acc, data_reference_p dr)
1430 {
1431 isl_constraint *c;
1432 int alias_set_num = 0;
1433 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
1434
1435 if (bap && bap->alias_set)
1436 alias_set_num = *(bap->alias_set);
1437
1438 c = isl_equality_alloc
1439 (isl_local_space_from_space (isl_map_get_space (acc)));
1440 c = isl_constraint_set_constant_si (c, -alias_set_num);
1441 c = isl_constraint_set_coefficient_si (c, isl_dim_out, 0, 1);
1442
1443 return isl_map_add_constraint (acc, c);
1444 }
1445
1446 /* Assign the affine expression INDEX to the output dimension POS of
1447 MAP and return the result. */
1448
1449 static isl_map *
1450 set_index (isl_map *map, int pos, isl_pw_aff *index)
1451 {
1452 isl_map *index_map;
1453 int len = isl_map_dim (map, isl_dim_out);
1454 isl_id *id;
1455
1456 index_map = isl_map_from_pw_aff (index);
1457 index_map = isl_map_insert_dims (index_map, isl_dim_out, 0, pos);
1458 index_map = isl_map_add_dims (index_map, isl_dim_out, len - pos - 1);
1459
1460 id = isl_map_get_tuple_id (map, isl_dim_out);
1461 index_map = isl_map_set_tuple_id (index_map, isl_dim_out, id);
1462 id = isl_map_get_tuple_id (map, isl_dim_in);
1463 index_map = isl_map_set_tuple_id (index_map, isl_dim_in, id);
1464
1465 return isl_map_intersect (map, index_map);
1466 }
1467
1468 /* Add to ACCESSES polyhedron equalities defining the access functions
1469 to the memory. ACCESSP_NB_DIMS is the dimension of the ACCESSES
1470 polyhedron, DOM_NB_DIMS is the dimension of the iteration domain.
1471 PBB is the poly_bb_p that contains the data reference DR. */
1472
1473 static isl_map *
1474 pdr_add_memory_accesses (isl_map *acc, data_reference_p dr, poly_bb_p pbb)
1475 {
1476 int i, nb_subscripts = DR_NUM_DIMENSIONS (dr);
1477 scop_p scop = PBB_SCOP (pbb);
1478
1479 for (i = 0; i < nb_subscripts; i++)
1480 {
1481 isl_pw_aff *aff;
1482 tree afn = DR_ACCESS_FN (dr, nb_subscripts - 1 - i);
1483
1484 aff = extract_affine (scop, afn,
1485 isl_space_domain (isl_map_get_space (acc)));
1486 acc = set_index (acc, i + 1, aff);
1487 }
1488
1489 return acc;
1490 }
1491
1492 /* Add constrains representing the size of the accessed data to the
1493 ACCESSES polyhedron. ACCESSP_NB_DIMS is the dimension of the
1494 ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration
1495 domain. */
1496
1497 static isl_set *
1498 pdr_add_data_dimensions (isl_set *extent, scop_p scop, data_reference_p dr)
1499 {
1500 tree ref = DR_REF (dr);
1501 int i, nb_subscripts = DR_NUM_DIMENSIONS (dr);
1502
1503 for (i = nb_subscripts - 1; i >= 0; i--, ref = TREE_OPERAND (ref, 0))
1504 {
1505 tree low, high;
1506
1507 if (TREE_CODE (ref) != ARRAY_REF)
1508 break;
1509
1510 low = array_ref_low_bound (ref);
1511 high = array_ref_up_bound (ref);
1512
1513 /* XXX The PPL code dealt separately with
1514 subscript - low >= 0 and high - subscript >= 0 in case one of
1515 the two bounds isn't known. Do the same here? */
1516
1517 if (tree_fits_shwi_p (low)
1518 && high
1519 && tree_fits_shwi_p (high)
1520 /* 1-element arrays at end of structures may extend over
1521 their declared size. */
1522 && !(array_at_struct_end_p (ref)
1523 && operand_equal_p (low, high, 0)))
1524 {
1525 isl_id *id;
1526 isl_aff *aff;
1527 isl_set *univ, *lbs, *ubs;
1528 isl_pw_aff *index;
1529 isl_space *space;
1530 isl_set *valid;
1531 isl_pw_aff *lb = extract_affine_int (low, isl_set_get_space (extent));
1532 isl_pw_aff *ub = extract_affine_int (high, isl_set_get_space (extent));
1533
1534 /* high >= 0 */
1535 valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (ub));
1536 valid = isl_set_project_out (valid, isl_dim_set, 0,
1537 isl_set_dim (valid, isl_dim_set));
1538 scop->context = isl_set_intersect (scop->context, valid);
1539
1540 space = isl_set_get_space (extent);
1541 aff = isl_aff_zero_on_domain (isl_local_space_from_space (space));
1542 aff = isl_aff_add_coefficient_si (aff, isl_dim_in, i + 1, 1);
1543 univ = isl_set_universe (isl_space_domain (isl_aff_get_space (aff)));
1544 index = isl_pw_aff_alloc (univ, aff);
1545
1546 id = isl_set_get_tuple_id (extent);
1547 lb = isl_pw_aff_set_tuple_id (lb, isl_dim_in, isl_id_copy (id));
1548 ub = isl_pw_aff_set_tuple_id (ub, isl_dim_in, id);
1549
1550 /* low <= sub_i <= high */
1551 lbs = isl_pw_aff_ge_set (isl_pw_aff_copy (index), lb);
1552 ubs = isl_pw_aff_le_set (index, ub);
1553 extent = isl_set_intersect (extent, lbs);
1554 extent = isl_set_intersect (extent, ubs);
1555 }
1556 }
1557
1558 return extent;
1559 }
1560
1561 /* Build data accesses for DR in PBB. */
1562
1563 static void
1564 build_poly_dr (data_reference_p dr, poly_bb_p pbb)
1565 {
1566 int dr_base_object_set;
1567 isl_map *acc;
1568 isl_set *extent;
1569 scop_p scop = PBB_SCOP (pbb);
1570
1571 {
1572 isl_space *dc = isl_set_get_space (pbb->domain);
1573 int nb_out = 1 + DR_NUM_DIMENSIONS (dr);
1574 isl_space *space = isl_space_add_dims (isl_space_from_domain (dc),
1575 isl_dim_out, nb_out);
1576
1577 acc = isl_map_universe (space);
1578 acc = isl_map_set_tuple_id (acc, isl_dim_out, isl_id_for_dr (scop, dr));
1579 }
1580
1581 acc = pdr_add_alias_set (acc, dr);
1582 acc = pdr_add_memory_accesses (acc, dr, pbb);
1583
1584 {
1585 isl_id *id = isl_id_for_dr (scop, dr);
1586 int nb = 1 + DR_NUM_DIMENSIONS (dr);
1587 isl_space *space = isl_space_set_alloc (scop->ctx, 0, nb);
1588 int alias_set_num = 0;
1589 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
1590
1591 if (bap && bap->alias_set)
1592 alias_set_num = *(bap->alias_set);
1593
1594 space = isl_space_set_tuple_id (space, isl_dim_set, id);
1595 extent = isl_set_nat_universe (space);
1596 extent = isl_set_fix_si (extent, isl_dim_set, 0, alias_set_num);
1597 extent = pdr_add_data_dimensions (extent, scop, dr);
1598 }
1599
1600 gcc_assert (dr->aux);
1601 dr_base_object_set = ((base_alias_pair *)(dr->aux))->base_obj_set;
1602
1603 new_poly_dr (pbb, dr_base_object_set,
1604 DR_IS_READ (dr) ? PDR_READ : PDR_WRITE,
1605 dr, DR_NUM_DIMENSIONS (dr), acc, extent);
1606 }
1607
1608 /* Write to FILE the alias graph of data references in DIMACS format. */
1609
1610 static inline bool
1611 write_alias_graph_to_ascii_dimacs (FILE *file, char *comment,
1612 vec<data_reference_p> drs)
1613 {
1614 int num_vertex = drs.length ();
1615 int edge_num = 0;
1616 data_reference_p dr1, dr2;
1617 int i, j;
1618
1619 if (num_vertex == 0)
1620 return true;
1621
1622 FOR_EACH_VEC_ELT (drs, i, dr1)
1623 for (j = i + 1; drs.iterate (j, &dr2); j++)
1624 if (dr_may_alias_p (dr1, dr2, true))
1625 edge_num++;
1626
1627 fprintf (file, "$\n");
1628
1629 if (comment)
1630 fprintf (file, "c %s\n", comment);
1631
1632 fprintf (file, "p edge %d %d\n", num_vertex, edge_num);
1633
1634 FOR_EACH_VEC_ELT (drs, i, dr1)
1635 for (j = i + 1; drs.iterate (j, &dr2); j++)
1636 if (dr_may_alias_p (dr1, dr2, true))
1637 fprintf (file, "e %d %d\n", i + 1, j + 1);
1638
1639 return true;
1640 }
1641
1642 /* Write to FILE the alias graph of data references in DOT format. */
1643
1644 static inline bool
1645 write_alias_graph_to_ascii_dot (FILE *file, char *comment,
1646 vec<data_reference_p> drs)
1647 {
1648 int num_vertex = drs.length ();
1649 data_reference_p dr1, dr2;
1650 int i, j;
1651
1652 if (num_vertex == 0)
1653 return true;
1654
1655 fprintf (file, "$\n");
1656
1657 if (comment)
1658 fprintf (file, "c %s\n", comment);
1659
1660 /* First print all the vertices. */
1661 FOR_EACH_VEC_ELT (drs, i, dr1)
1662 fprintf (file, "n%d;\n", i);
1663
1664 FOR_EACH_VEC_ELT (drs, i, dr1)
1665 for (j = i + 1; drs.iterate (j, &dr2); j++)
1666 if (dr_may_alias_p (dr1, dr2, true))
1667 fprintf (file, "n%d n%d\n", i, j);
1668
1669 return true;
1670 }
1671
1672 /* Write to FILE the alias graph of data references in ECC format. */
1673
1674 static inline bool
1675 write_alias_graph_to_ascii_ecc (FILE *file, char *comment,
1676 vec<data_reference_p> drs)
1677 {
1678 int num_vertex = drs.length ();
1679 data_reference_p dr1, dr2;
1680 int i, j;
1681
1682 if (num_vertex == 0)
1683 return true;
1684
1685 fprintf (file, "$\n");
1686
1687 if (comment)
1688 fprintf (file, "c %s\n", comment);
1689
1690 FOR_EACH_VEC_ELT (drs, i, dr1)
1691 for (j = i + 1; drs.iterate (j, &dr2); j++)
1692 if (dr_may_alias_p (dr1, dr2, true))
1693 fprintf (file, "%d %d\n", i, j);
1694
1695 return true;
1696 }
1697
1698 /* Check if DR1 and DR2 are in the same object set. */
1699
1700 static bool
1701 dr_same_base_object_p (const struct data_reference *dr1,
1702 const struct data_reference *dr2)
1703 {
1704 return operand_equal_p (DR_BASE_OBJECT (dr1), DR_BASE_OBJECT (dr2), 0);
1705 }
1706
1707 /* Uses DFS component number as representative of alias-sets. Also tests for
1708 optimality by verifying if every connected component is a clique. Returns
1709 true (1) if the above test is true, and false (0) otherwise. */
1710
1711 static int
1712 build_alias_set_optimal_p (vec<data_reference_p> drs)
1713 {
1714 int num_vertices = drs.length ();
1715 struct graph *g = new_graph (num_vertices);
1716 data_reference_p dr1, dr2;
1717 int i, j;
1718 int num_connected_components;
1719 int v_indx1, v_indx2, num_vertices_in_component;
1720 int *all_vertices;
1721 int *vertices;
1722 struct graph_edge *e;
1723 int this_component_is_clique;
1724 int all_components_are_cliques = 1;
1725
1726 FOR_EACH_VEC_ELT (drs, i, dr1)
1727 for (j = i+1; drs.iterate (j, &dr2); j++)
1728 if (dr_may_alias_p (dr1, dr2, true))
1729 {
1730 add_edge (g, i, j);
1731 add_edge (g, j, i);
1732 }
1733
1734 all_vertices = XNEWVEC (int, num_vertices);
1735 vertices = XNEWVEC (int, num_vertices);
1736 for (i = 0; i < num_vertices; i++)
1737 all_vertices[i] = i;
1738
1739 num_connected_components = graphds_dfs (g, all_vertices, num_vertices,
1740 NULL, true, NULL);
1741 for (i = 0; i < g->n_vertices; i++)
1742 {
1743 data_reference_p dr = drs[i];
1744 base_alias_pair *bap;
1745
1746 gcc_assert (dr->aux);
1747 bap = (base_alias_pair *)(dr->aux);
1748
1749 bap->alias_set = XNEW (int);
1750 *(bap->alias_set) = g->vertices[i].component + 1;
1751 }
1752
1753 /* Verify if the DFS numbering results in optimal solution. */
1754 for (i = 0; i < num_connected_components; i++)
1755 {
1756 num_vertices_in_component = 0;
1757 /* Get all vertices whose DFS component number is the same as i. */
1758 for (j = 0; j < num_vertices; j++)
1759 if (g->vertices[j].component == i)
1760 vertices[num_vertices_in_component++] = j;
1761
1762 /* Now test if the vertices in 'vertices' form a clique, by testing
1763 for edges among each pair. */
1764 this_component_is_clique = 1;
1765 for (v_indx1 = 0; v_indx1 < num_vertices_in_component; v_indx1++)
1766 {
1767 for (v_indx2 = v_indx1+1; v_indx2 < num_vertices_in_component; v_indx2++)
1768 {
1769 /* Check if the two vertices are connected by iterating
1770 through all the edges which have one of these are source. */
1771 e = g->vertices[vertices[v_indx2]].pred;
1772 while (e)
1773 {
1774 if (e->src == vertices[v_indx1])
1775 break;
1776 e = e->pred_next;
1777 }
1778 if (!e)
1779 {
1780 this_component_is_clique = 0;
1781 break;
1782 }
1783 }
1784 if (!this_component_is_clique)
1785 all_components_are_cliques = 0;
1786 }
1787 }
1788
1789 free (all_vertices);
1790 free (vertices);
1791 free_graph (g);
1792 return all_components_are_cliques;
1793 }
1794
1795 /* Group each data reference in DRS with its base object set num. */
1796
1797 static void
1798 build_base_obj_set_for_drs (vec<data_reference_p> drs)
1799 {
1800 int num_vertex = drs.length ();
1801 struct graph *g = new_graph (num_vertex);
1802 data_reference_p dr1, dr2;
1803 int i, j;
1804 int *queue;
1805
1806 FOR_EACH_VEC_ELT (drs, i, dr1)
1807 for (j = i + 1; drs.iterate (j, &dr2); j++)
1808 if (dr_same_base_object_p (dr1, dr2))
1809 {
1810 add_edge (g, i, j);
1811 add_edge (g, j, i);
1812 }
1813
1814 queue = XNEWVEC (int, num_vertex);
1815 for (i = 0; i < num_vertex; i++)
1816 queue[i] = i;
1817
1818 graphds_dfs (g, queue, num_vertex, NULL, true, NULL);
1819
1820 for (i = 0; i < g->n_vertices; i++)
1821 {
1822 data_reference_p dr = drs[i];
1823 base_alias_pair *bap;
1824
1825 gcc_assert (dr->aux);
1826 bap = (base_alias_pair *)(dr->aux);
1827
1828 bap->base_obj_set = g->vertices[i].component + 1;
1829 }
1830
1831 free (queue);
1832 free_graph (g);
1833 }
1834
1835 /* Build the data references for PBB. */
1836
1837 static void
1838 build_pbb_drs (poly_bb_p pbb)
1839 {
1840 int j;
1841 data_reference_p dr;
1842 vec<data_reference_p> gbb_drs = GBB_DATA_REFS (PBB_BLACK_BOX (pbb));
1843
1844 FOR_EACH_VEC_ELT (gbb_drs, j, dr)
1845 build_poly_dr (dr, pbb);
1846 }
1847
1848 /* Dump to file the alias graphs for the data references in DRS. */
1849
1850 static void
1851 dump_alias_graphs (vec<data_reference_p> drs)
1852 {
1853 char comment[100];
1854 FILE *file_dimacs, *file_ecc, *file_dot;
1855
1856 file_dimacs = fopen ("/tmp/dr_alias_graph_dimacs", "ab");
1857 if (file_dimacs)
1858 {
1859 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1860 current_function_name ());
1861 write_alias_graph_to_ascii_dimacs (file_dimacs, comment, drs);
1862 fclose (file_dimacs);
1863 }
1864
1865 file_ecc = fopen ("/tmp/dr_alias_graph_ecc", "ab");
1866 if (file_ecc)
1867 {
1868 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1869 current_function_name ());
1870 write_alias_graph_to_ascii_ecc (file_ecc, comment, drs);
1871 fclose (file_ecc);
1872 }
1873
1874 file_dot = fopen ("/tmp/dr_alias_graph_dot", "ab");
1875 if (file_dot)
1876 {
1877 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1878 current_function_name ());
1879 write_alias_graph_to_ascii_dot (file_dot, comment, drs);
1880 fclose (file_dot);
1881 }
1882 }
1883
1884 /* Build data references in SCOP. */
1885
1886 static void
1887 build_scop_drs (scop_p scop)
1888 {
1889 int i, j;
1890 poly_bb_p pbb;
1891 data_reference_p dr;
1892 auto_vec<data_reference_p, 3> drs;
1893
1894 /* Remove all the PBBs that do not have data references: these basic
1895 blocks are not handled in the polyhedral representation. */
1896 for (i = 0; SCOP_BBS (scop).iterate (i, &pbb); i++)
1897 if (GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).is_empty ())
1898 {
1899 free_gimple_bb (PBB_BLACK_BOX (pbb));
1900 free_poly_bb (pbb);
1901 SCOP_BBS (scop).ordered_remove (i);
1902 i--;
1903 }
1904
1905 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1906 for (j = 0; GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).iterate (j, &dr); j++)
1907 drs.safe_push (dr);
1908
1909 FOR_EACH_VEC_ELT (drs, i, dr)
1910 dr->aux = XNEW (base_alias_pair);
1911
1912 if (!build_alias_set_optimal_p (drs))
1913 {
1914 /* TODO: Add support when building alias set is not optimal. */
1915 ;
1916 }
1917
1918 build_base_obj_set_for_drs (drs);
1919
1920 /* When debugging, enable the following code. This cannot be used
1921 in production compilers. */
1922 if (0)
1923 dump_alias_graphs (drs);
1924
1925 drs.release ();
1926
1927 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1928 build_pbb_drs (pbb);
1929 }
1930
1931 /* Return a gsi at the position of the phi node STMT. */
1932
1933 static gimple_stmt_iterator
1934 gsi_for_phi_node (gimple stmt)
1935 {
1936 gimple_stmt_iterator psi;
1937 basic_block bb = gimple_bb (stmt);
1938
1939 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
1940 if (stmt == gsi_stmt (psi))
1941 return psi;
1942
1943 gcc_unreachable ();
1944 return psi;
1945 }
1946
1947 /* Analyze all the data references of STMTS and add them to the
1948 GBB_DATA_REFS vector of BB. */
1949
1950 static void
1951 analyze_drs_in_stmts (scop_p scop, basic_block bb, vec<gimple> stmts)
1952 {
1953 loop_p nest;
1954 gimple_bb_p gbb;
1955 gimple stmt;
1956 int i;
1957 sese region = SCOP_REGION (scop);
1958
1959 if (!bb_in_sese_p (bb, region))
1960 return;
1961
1962 nest = outermost_loop_in_sese_1 (region, bb);
1963 gbb = gbb_from_bb (bb);
1964
1965 FOR_EACH_VEC_ELT (stmts, i, stmt)
1966 {
1967 loop_p loop;
1968
1969 if (is_gimple_debug (stmt))
1970 continue;
1971
1972 loop = loop_containing_stmt (stmt);
1973 if (!loop_in_sese_p (loop, region))
1974 loop = nest;
1975
1976 graphite_find_data_references_in_stmt (nest, loop, stmt,
1977 &GBB_DATA_REFS (gbb));
1978 }
1979 }
1980
1981 /* Insert STMT at the end of the STMTS sequence and then insert the
1982 statements from STMTS at INSERT_GSI and call analyze_drs_in_stmts
1983 on STMTS. */
1984
1985 static void
1986 insert_stmts (scop_p scop, gimple stmt, gimple_seq stmts,
1987 gimple_stmt_iterator insert_gsi)
1988 {
1989 gimple_stmt_iterator gsi;
1990 auto_vec<gimple, 3> x;
1991
1992 gimple_seq_add_stmt (&stmts, stmt);
1993 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
1994 x.safe_push (gsi_stmt (gsi));
1995
1996 gsi_insert_seq_before (&insert_gsi, stmts, GSI_SAME_STMT);
1997 analyze_drs_in_stmts (scop, gsi_bb (insert_gsi), x);
1998 }
1999
2000 /* Insert the assignment "RES := EXPR" just after AFTER_STMT. */
2001
2002 static void
2003 insert_out_of_ssa_copy (scop_p scop, tree res, tree expr, gimple after_stmt)
2004 {
2005 gimple_seq stmts;
2006 gimple_stmt_iterator gsi;
2007 tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE);
2008 gimple stmt = gimple_build_assign (unshare_expr (res), var);
2009 auto_vec<gimple, 3> x;
2010
2011 gimple_seq_add_stmt (&stmts, stmt);
2012 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2013 x.safe_push (gsi_stmt (gsi));
2014
2015 if (gimple_code (after_stmt) == GIMPLE_PHI)
2016 {
2017 gsi = gsi_after_labels (gimple_bb (after_stmt));
2018 gsi_insert_seq_before (&gsi, stmts, GSI_NEW_STMT);
2019 }
2020 else
2021 {
2022 gsi = gsi_for_stmt (after_stmt);
2023 gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT);
2024 }
2025
2026 analyze_drs_in_stmts (scop, gimple_bb (after_stmt), x);
2027 }
2028
2029 /* Creates a poly_bb_p for basic_block BB from the existing PBB. */
2030
2031 static void
2032 new_pbb_from_pbb (scop_p scop, poly_bb_p pbb, basic_block bb)
2033 {
2034 vec<data_reference_p> drs;
2035 drs.create (3);
2036 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
2037 gimple_bb_p gbb1 = new_gimple_bb (bb, drs);
2038 poly_bb_p pbb1 = new_poly_bb (scop, gbb1);
2039 int index, n = SCOP_BBS (scop).length ();
2040
2041 /* The INDEX of PBB in SCOP_BBS. */
2042 for (index = 0; index < n; index++)
2043 if (SCOP_BBS (scop)[index] == pbb)
2044 break;
2045
2046 pbb1->domain = isl_set_copy (pbb->domain);
2047
2048 GBB_PBB (gbb1) = pbb1;
2049 GBB_CONDITIONS (gbb1) = GBB_CONDITIONS (gbb).copy ();
2050 GBB_CONDITION_CASES (gbb1) = GBB_CONDITION_CASES (gbb).copy ();
2051 SCOP_BBS (scop).safe_insert (index + 1, pbb1);
2052 }
2053
2054 /* Insert on edge E the assignment "RES := EXPR". */
2055
2056 static void
2057 insert_out_of_ssa_copy_on_edge (scop_p scop, edge e, tree res, tree expr)
2058 {
2059 gimple_stmt_iterator gsi;
2060 gimple_seq stmts = NULL;
2061 tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE);
2062 gimple stmt = gimple_build_assign (unshare_expr (res), var);
2063 basic_block bb;
2064 auto_vec<gimple, 3> x;
2065
2066 gimple_seq_add_stmt (&stmts, stmt);
2067 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2068 x.safe_push (gsi_stmt (gsi));
2069
2070 gsi_insert_seq_on_edge (e, stmts);
2071 gsi_commit_edge_inserts ();
2072 bb = gimple_bb (stmt);
2073
2074 if (!bb_in_sese_p (bb, SCOP_REGION (scop)))
2075 return;
2076
2077 if (!gbb_from_bb (bb))
2078 new_pbb_from_pbb (scop, pbb_from_bb (e->src), bb);
2079
2080 analyze_drs_in_stmts (scop, bb, x);
2081 }
2082
2083 /* Creates a zero dimension array of the same type as VAR. */
2084
2085 static tree
2086 create_zero_dim_array (tree var, const char *base_name)
2087 {
2088 tree index_type = build_index_type (integer_zero_node);
2089 tree elt_type = TREE_TYPE (var);
2090 tree array_type = build_array_type (elt_type, index_type);
2091 tree base = create_tmp_var (array_type, base_name);
2092
2093 return build4 (ARRAY_REF, elt_type, base, integer_zero_node, NULL_TREE,
2094 NULL_TREE);
2095 }
2096
2097 /* Returns true when PHI is a loop close phi node. */
2098
2099 static bool
2100 scalar_close_phi_node_p (gimple phi)
2101 {
2102 if (gimple_code (phi) != GIMPLE_PHI
2103 || virtual_operand_p (gimple_phi_result (phi)))
2104 return false;
2105
2106 /* Note that loop close phi nodes should have a single argument
2107 because we translated the representation into a canonical form
2108 before Graphite: see canonicalize_loop_closed_ssa_form. */
2109 return (gimple_phi_num_args (phi) == 1);
2110 }
2111
2112 /* For a definition DEF in REGION, propagates the expression EXPR in
2113 all the uses of DEF outside REGION. */
2114
2115 static void
2116 propagate_expr_outside_region (tree def, tree expr, sese region)
2117 {
2118 imm_use_iterator imm_iter;
2119 gimple use_stmt;
2120 gimple_seq stmts;
2121 bool replaced_once = false;
2122
2123 gcc_assert (TREE_CODE (def) == SSA_NAME);
2124
2125 expr = force_gimple_operand (unshare_expr (expr), &stmts, true,
2126 NULL_TREE);
2127
2128 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2129 if (!is_gimple_debug (use_stmt)
2130 && !bb_in_sese_p (gimple_bb (use_stmt), region))
2131 {
2132 ssa_op_iter iter;
2133 use_operand_p use_p;
2134
2135 FOR_EACH_PHI_OR_STMT_USE (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2136 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0)
2137 && (replaced_once = true))
2138 replace_exp (use_p, expr);
2139
2140 update_stmt (use_stmt);
2141 }
2142
2143 if (replaced_once)
2144 {
2145 gsi_insert_seq_on_edge (SESE_ENTRY (region), stmts);
2146 gsi_commit_edge_inserts ();
2147 }
2148 }
2149
2150 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2151 dimension array for it. */
2152
2153 static void
2154 rewrite_close_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi)
2155 {
2156 sese region = SCOP_REGION (scop);
2157 gimple phi = gsi_stmt (*psi);
2158 tree res = gimple_phi_result (phi);
2159 basic_block bb = gimple_bb (phi);
2160 gimple_stmt_iterator gsi = gsi_after_labels (bb);
2161 tree arg = gimple_phi_arg_def (phi, 0);
2162 gimple stmt;
2163
2164 /* Note that loop close phi nodes should have a single argument
2165 because we translated the representation into a canonical form
2166 before Graphite: see canonicalize_loop_closed_ssa_form. */
2167 gcc_assert (gimple_phi_num_args (phi) == 1);
2168
2169 /* The phi node can be a non close phi node, when its argument is
2170 invariant, or a default definition. */
2171 if (is_gimple_min_invariant (arg)
2172 || SSA_NAME_IS_DEFAULT_DEF (arg))
2173 {
2174 propagate_expr_outside_region (res, arg, region);
2175 gsi_next (psi);
2176 return;
2177 }
2178
2179 else if (gimple_bb (SSA_NAME_DEF_STMT (arg))->loop_father == bb->loop_father)
2180 {
2181 propagate_expr_outside_region (res, arg, region);
2182 stmt = gimple_build_assign (res, arg);
2183 remove_phi_node (psi, false);
2184 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2185 return;
2186 }
2187
2188 /* If res is scev analyzable and is not a scalar value, it is safe
2189 to ignore the close phi node: it will be code generated in the
2190 out of Graphite pass. */
2191 else if (scev_analyzable_p (res, region))
2192 {
2193 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (res));
2194 tree scev;
2195
2196 if (!loop_in_sese_p (loop, region))
2197 {
2198 loop = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2199 scev = scalar_evolution_in_region (region, loop, arg);
2200 scev = compute_overall_effect_of_inner_loop (loop, scev);
2201 }
2202 else
2203 scev = scalar_evolution_in_region (region, loop, res);
2204
2205 if (tree_does_not_contain_chrecs (scev))
2206 propagate_expr_outside_region (res, scev, region);
2207
2208 gsi_next (psi);
2209 return;
2210 }
2211 else
2212 {
2213 tree zero_dim_array = create_zero_dim_array (res, "Close_Phi");
2214
2215 stmt = gimple_build_assign (res, unshare_expr (zero_dim_array));
2216
2217 if (TREE_CODE (arg) == SSA_NAME)
2218 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2219 SSA_NAME_DEF_STMT (arg));
2220 else
2221 insert_out_of_ssa_copy_on_edge (scop, single_pred_edge (bb),
2222 zero_dim_array, arg);
2223 }
2224
2225 remove_phi_node (psi, false);
2226 SSA_NAME_DEF_STMT (res) = stmt;
2227
2228 insert_stmts (scop, stmt, NULL, gsi_after_labels (bb));
2229 }
2230
2231 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2232 dimension array for it. */
2233
2234 static void
2235 rewrite_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi)
2236 {
2237 size_t i;
2238 gimple phi = gsi_stmt (*psi);
2239 basic_block bb = gimple_bb (phi);
2240 tree res = gimple_phi_result (phi);
2241 tree zero_dim_array = create_zero_dim_array (res, "phi_out_of_ssa");
2242 gimple stmt;
2243
2244 for (i = 0; i < gimple_phi_num_args (phi); i++)
2245 {
2246 tree arg = gimple_phi_arg_def (phi, i);
2247 edge e = gimple_phi_arg_edge (phi, i);
2248
2249 /* Avoid the insertion of code in the loop latch to please the
2250 pattern matching of the vectorizer. */
2251 if (TREE_CODE (arg) == SSA_NAME
2252 && !SSA_NAME_IS_DEFAULT_DEF (arg)
2253 && e->src == bb->loop_father->latch)
2254 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2255 SSA_NAME_DEF_STMT (arg));
2256 else
2257 insert_out_of_ssa_copy_on_edge (scop, e, zero_dim_array, arg);
2258 }
2259
2260 stmt = gimple_build_assign (res, unshare_expr (zero_dim_array));
2261 remove_phi_node (psi, false);
2262 insert_stmts (scop, stmt, NULL, gsi_after_labels (bb));
2263 }
2264
2265 /* Rewrite the degenerate phi node at position PSI from the degenerate
2266 form "x = phi (y, y, ..., y)" to "x = y". */
2267
2268 static void
2269 rewrite_degenerate_phi (gimple_stmt_iterator *psi)
2270 {
2271 tree rhs;
2272 gimple stmt;
2273 gimple_stmt_iterator gsi;
2274 gimple phi = gsi_stmt (*psi);
2275 tree res = gimple_phi_result (phi);
2276 basic_block bb;
2277
2278 bb = gimple_bb (phi);
2279 rhs = degenerate_phi_result (phi);
2280 gcc_assert (rhs);
2281
2282 stmt = gimple_build_assign (res, rhs);
2283 remove_phi_node (psi, false);
2284
2285 gsi = gsi_after_labels (bb);
2286 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2287 }
2288
2289 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2290
2291 static void
2292 rewrite_reductions_out_of_ssa (scop_p scop)
2293 {
2294 basic_block bb;
2295 gimple_stmt_iterator psi;
2296 sese region = SCOP_REGION (scop);
2297
2298 FOR_EACH_BB_FN (bb, cfun)
2299 if (bb_in_sese_p (bb, region))
2300 for (psi = gsi_start_phis (bb); !gsi_end_p (psi);)
2301 {
2302 gimple phi = gsi_stmt (psi);
2303
2304 if (virtual_operand_p (gimple_phi_result (phi)))
2305 {
2306 gsi_next (&psi);
2307 continue;
2308 }
2309
2310 if (gimple_phi_num_args (phi) > 1
2311 && degenerate_phi_result (phi))
2312 rewrite_degenerate_phi (&psi);
2313
2314 else if (scalar_close_phi_node_p (phi))
2315 rewrite_close_phi_out_of_ssa (scop, &psi);
2316
2317 else if (reduction_phi_p (region, &psi))
2318 rewrite_phi_out_of_ssa (scop, &psi);
2319 }
2320
2321 update_ssa (TODO_update_ssa);
2322 #ifdef ENABLE_CHECKING
2323 verify_loop_closed_ssa (true);
2324 #endif
2325 }
2326
2327 /* Rewrite the scalar dependence of DEF used in USE_STMT with a memory
2328 read from ZERO_DIM_ARRAY. */
2329
2330 static void
2331 rewrite_cross_bb_scalar_dependence (scop_p scop, tree zero_dim_array,
2332 tree def, gimple use_stmt)
2333 {
2334 gimple name_stmt;
2335 tree name;
2336 ssa_op_iter iter;
2337 use_operand_p use_p;
2338
2339 gcc_assert (gimple_code (use_stmt) != GIMPLE_PHI);
2340
2341 name = copy_ssa_name (def, NULL);
2342 name_stmt = gimple_build_assign (name, zero_dim_array);
2343
2344 gimple_assign_set_lhs (name_stmt, name);
2345 insert_stmts (scop, name_stmt, NULL, gsi_for_stmt (use_stmt));
2346
2347 FOR_EACH_SSA_USE_OPERAND (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2348 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0))
2349 replace_exp (use_p, name);
2350
2351 update_stmt (use_stmt);
2352 }
2353
2354 /* For every definition DEF in the SCOP that is used outside the scop,
2355 insert a closing-scop definition in the basic block just after this
2356 SCOP. */
2357
2358 static void
2359 handle_scalar_deps_crossing_scop_limits (scop_p scop, tree def, gimple stmt)
2360 {
2361 tree var = create_tmp_reg (TREE_TYPE (def), NULL);
2362 tree new_name = make_ssa_name (var, stmt);
2363 bool needs_copy = false;
2364 use_operand_p use_p;
2365 imm_use_iterator imm_iter;
2366 gimple use_stmt;
2367 sese region = SCOP_REGION (scop);
2368
2369 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2370 {
2371 if (!bb_in_sese_p (gimple_bb (use_stmt), region))
2372 {
2373 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
2374 {
2375 SET_USE (use_p, new_name);
2376 }
2377 update_stmt (use_stmt);
2378 needs_copy = true;
2379 }
2380 }
2381
2382 /* Insert in the empty BB just after the scop a use of DEF such
2383 that the rewrite of cross_bb_scalar_dependences won't insert
2384 arrays everywhere else. */
2385 if (needs_copy)
2386 {
2387 gimple assign = gimple_build_assign (new_name, def);
2388 gimple_stmt_iterator psi = gsi_after_labels (SESE_EXIT (region)->dest);
2389
2390 update_stmt (assign);
2391 gsi_insert_before (&psi, assign, GSI_SAME_STMT);
2392 }
2393 }
2394
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. */
2398
2399 static bool
2400 rewrite_cross_bb_scalar_deps (scop_p scop, gimple_stmt_iterator *gsi)
2401 {
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;
2410
2411 switch (gimple_code (stmt))
2412 {
2413 case GIMPLE_ASSIGN:
2414 def = gimple_assign_lhs (stmt);
2415 break;
2416
2417 case GIMPLE_CALL:
2418 def = gimple_call_lhs (stmt);
2419 break;
2420
2421 default:
2422 return false;
2423 }
2424
2425 if (!def
2426 || !is_gimple_reg (def))
2427 return false;
2428
2429 if (scev_analyzable_p (def, region))
2430 {
2431 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (def));
2432 tree scev = scalar_evolution_in_region (region, loop, def);
2433
2434 if (tree_contains_chrecs (scev, NULL))
2435 return false;
2436
2437 propagate_expr_outside_region (def, scev, region);
2438 return true;
2439 }
2440
2441 def_bb = gimple_bb (stmt);
2442
2443 handle_scalar_deps_crossing_scop_limits (scop, def, stmt);
2444
2445 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2446 if (gimple_code (use_stmt) == GIMPLE_PHI
2447 && (res = true))
2448 {
2449 gimple_stmt_iterator psi = gsi_for_stmt (use_stmt);
2450
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);
2455 }
2456
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))
2462 {
2463 if (!zero_dim_array)
2464 {
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);
2470 }
2471
2472 rewrite_cross_bb_scalar_dependence (scop, unshare_expr (zero_dim_array),
2473 def, use_stmt);
2474 }
2475
2476 return res;
2477 }
2478
2479 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2480
2481 static void
2482 rewrite_cross_bb_scalar_deps_out_of_ssa (scop_p scop)
2483 {
2484 basic_block bb;
2485 gimple_stmt_iterator psi;
2486 sese region = SCOP_REGION (scop);
2487 bool changed = false;
2488
2489 /* Create an extra empty BB after the scop. */
2490 split_edge (SESE_EXIT (region));
2491
2492 FOR_EACH_BB_FN (bb, cfun)
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);
2496
2497 if (changed)
2498 {
2499 scev_reset_htab ();
2500 update_ssa (TODO_update_ssa);
2501 #ifdef ENABLE_CHECKING
2502 verify_loop_closed_ssa (true);
2503 #endif
2504 }
2505 }
2506
2507 /* Returns the number of pbbs that are in loops contained in SCOP. */
2508
2509 static int
2510 nb_pbbs_in_loops (scop_p scop)
2511 {
2512 int i;
2513 poly_bb_p pbb;
2514 int res = 0;
2515
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++;
2519
2520 return res;
2521 }
2522
2523 /* Return the number of data references in BB that write in
2524 memory. */
2525
2526 static int
2527 nb_data_writes_in_bb (basic_block bb)
2528 {
2529 int res = 0;
2530 gimple_stmt_iterator gsi;
2531
2532 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
2533 if (gimple_vdef (gsi_stmt (gsi)))
2534 res++;
2535
2536 return res;
2537 }
2538
2539 /* Splits at STMT the basic block BB represented as PBB in the
2540 polyhedral form. */
2541
2542 static edge
2543 split_pbb (scop_p scop, poly_bb_p pbb, basic_block bb, gimple stmt)
2544 {
2545 edge e1 = split_block (bb, stmt);
2546 new_pbb_from_pbb (scop, pbb, e1->dest);
2547 return e1;
2548 }
2549
2550 /* Splits STMT out of its current BB. This is done for reduction
2551 statements for which we want to ignore data dependences. */
2552
2553 static basic_block
2554 split_reduction_stmt (scop_p scop, gimple stmt)
2555 {
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;
2562
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;
2569
2570 e1 = split_pbb (scop, pbb, bb, stmt);
2571
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)))
2575 {
2576 gimple_stmt_iterator gsi = gsi_last_bb (bb);
2577 gsi_prev (&gsi);
2578 e1 = split_pbb (scop, pbb, bb, gsi_stmt (gsi));
2579 }
2580
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)
2585 {
2586 basic_block bb1 = gimple_bb (DR_STMT (dr));
2587
2588 if (bb1 != bb)
2589 {
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--;
2594 }
2595 }
2596
2597 return e1->dest;
2598 }
2599
2600 /* Return true when stmt is a reduction operation. */
2601
2602 static inline bool
2603 is_reduction_operation_p (gimple stmt)
2604 {
2605 enum tree_code code;
2606
2607 gcc_assert (is_gimple_assign (stmt));
2608 code = gimple_assign_rhs_code (stmt);
2609
2610 return flag_associative_math
2611 && commutative_tree_code (code)
2612 && associative_tree_code (code);
2613 }
2614
2615 /* Returns true when PHI contains an argument ARG. */
2616
2617 static bool
2618 phi_contains_arg (gimple phi, tree arg)
2619 {
2620 size_t i;
2621
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;
2625
2626 return false;
2627 }
2628
2629 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2630
2631 static gimple
2632 follow_ssa_with_commutative_ops (tree arg, tree lhs)
2633 {
2634 gimple stmt;
2635
2636 if (TREE_CODE (arg) != SSA_NAME)
2637 return NULL;
2638
2639 stmt = SSA_NAME_DEF_STMT (arg);
2640
2641 if (gimple_code (stmt) == GIMPLE_NOP
2642 || gimple_code (stmt) == GIMPLE_CALL)
2643 return NULL;
2644
2645 if (gimple_code (stmt) == GIMPLE_PHI)
2646 {
2647 if (phi_contains_arg (stmt, lhs))
2648 return stmt;
2649 return NULL;
2650 }
2651
2652 if (!is_gimple_assign (stmt))
2653 return NULL;
2654
2655 if (gimple_num_ops (stmt) == 2)
2656 return follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2657
2658 if (is_reduction_operation_p (stmt))
2659 {
2660 gimple res = follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2661
2662 return res ? res :
2663 follow_ssa_with_commutative_ops (gimple_assign_rhs2 (stmt), lhs);
2664 }
2665
2666 return NULL;
2667 }
2668
2669 /* Detect commutative and associative scalar reductions starting at
2670 the STMT. Return the phi node of the reduction cycle, or NULL. */
2671
2672 static gimple
2673 detect_commutative_reduction_arg (tree lhs, gimple stmt, tree arg,
2674 vec<gimple> *in,
2675 vec<gimple> *out)
2676 {
2677 gimple phi = follow_ssa_with_commutative_ops (arg, lhs);
2678
2679 if (!phi)
2680 return NULL;
2681
2682 in->safe_push (stmt);
2683 out->safe_push (stmt);
2684 return phi;
2685 }
2686
2687 /* Detect commutative and associative scalar reductions starting at
2688 STMT. Return the phi node of the reduction cycle, or NULL. */
2689
2690 static gimple
2691 detect_commutative_reduction_assign (gimple stmt, vec<gimple> *in,
2692 vec<gimple> *out)
2693 {
2694 tree lhs = gimple_assign_lhs (stmt);
2695
2696 if (gimple_num_ops (stmt) == 2)
2697 return detect_commutative_reduction_arg (lhs, stmt,
2698 gimple_assign_rhs1 (stmt),
2699 in, out);
2700
2701 if (is_reduction_operation_p (stmt))
2702 {
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);
2710 }
2711
2712 return NULL;
2713 }
2714
2715 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2716
2717 static gimple
2718 follow_inital_value_to_phi (tree arg, tree lhs)
2719 {
2720 gimple stmt;
2721
2722 if (!arg || TREE_CODE (arg) != SSA_NAME)
2723 return NULL;
2724
2725 stmt = SSA_NAME_DEF_STMT (arg);
2726
2727 if (gimple_code (stmt) == GIMPLE_PHI
2728 && phi_contains_arg (stmt, lhs))
2729 return stmt;
2730
2731 return NULL;
2732 }
2733
2734
2735 /* Return the argument of the loop PHI that is the initial value coming
2736 from outside the loop. */
2737
2738 static edge
2739 edge_initial_value_for_loop_phi (gimple phi)
2740 {
2741 size_t i;
2742
2743 for (i = 0; i < gimple_phi_num_args (phi); i++)
2744 {
2745 edge e = gimple_phi_arg_edge (phi, i);
2746
2747 if (loop_depth (e->src->loop_father)
2748 < loop_depth (e->dest->loop_father))
2749 return e;
2750 }
2751
2752 return NULL;
2753 }
2754
2755 /* Return the argument of the loop PHI that is the initial value coming
2756 from outside the loop. */
2757
2758 static tree
2759 initial_value_for_loop_phi (gimple phi)
2760 {
2761 size_t i;
2762
2763 for (i = 0; i < gimple_phi_num_args (phi); i++)
2764 {
2765 edge e = gimple_phi_arg_edge (phi, i);
2766
2767 if (loop_depth (e->src->loop_father)
2768 < loop_depth (e->dest->loop_father))
2769 return gimple_phi_arg_def (phi, i);
2770 }
2771
2772 return NULL_TREE;
2773 }
2774
2775 /* Returns true when DEF is used outside the reduction cycle of
2776 LOOP_PHI. */
2777
2778 static bool
2779 used_outside_reduction (tree def, gimple loop_phi)
2780 {
2781 use_operand_p use_p;
2782 imm_use_iterator imm_iter;
2783 loop_p loop = loop_containing_stmt (loop_phi);
2784
2785 /* In LOOP, DEF should be used only in LOOP_PHI. */
2786 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2787 {
2788 gimple stmt = USE_STMT (use_p);
2789
2790 if (stmt != loop_phi
2791 && !is_gimple_debug (stmt)
2792 && flow_bb_inside_loop_p (loop, gimple_bb (stmt)))
2793 return true;
2794 }
2795
2796 return false;
2797 }
2798
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. */
2802
2803 static gimple
2804 detect_commutative_reduction (scop_p scop, gimple stmt, vec<gimple> *in,
2805 vec<gimple> *out)
2806 {
2807 if (scalar_close_phi_node_p (stmt))
2808 {
2809 gimple def, loop_phi, phi, close_phi = stmt;
2810 tree init, lhs, arg = gimple_phi_arg_def (close_phi, 0);
2811
2812 if (TREE_CODE (arg) != SSA_NAME)
2813 return NULL;
2814
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);
2819
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;
2824
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);
2828
2829 if (phi && (used_outside_reduction (lhs, phi)
2830 || !has_single_use (gimple_phi_result (phi))))
2831 return NULL;
2832
2833 in->safe_push (loop_phi);
2834 out->safe_push (close_phi);
2835 return phi;
2836 }
2837
2838 if (gimple_code (stmt) == GIMPLE_ASSIGN)
2839 return detect_commutative_reduction_assign (stmt, in, out);
2840
2841 return NULL;
2842 }
2843
2844 /* Translate the scalar reduction statement STMT to an array RED
2845 knowing that its recursive phi node is LOOP_PHI. */
2846
2847 static void
2848 translate_scalar_reduction_to_array_for_stmt (scop_p scop, tree red,
2849 gimple stmt, gimple loop_phi)
2850 {
2851 tree res = gimple_phi_result (loop_phi);
2852 gimple assign = gimple_build_assign (res, unshare_expr (red));
2853 gimple_stmt_iterator gsi;
2854
2855 insert_stmts (scop, assign, NULL, gsi_after_labels (gimple_bb (loop_phi)));
2856
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);
2861 }
2862
2863 /* Removes the PHI node and resets all the debug stmts that are using
2864 the PHI_RESULT. */
2865
2866 static void
2867 remove_phi (gimple phi)
2868 {
2869 imm_use_iterator imm_iter;
2870 tree def;
2871 use_operand_p use_p;
2872 gimple_stmt_iterator gsi;
2873 auto_vec<gimple, 3> update;
2874 unsigned int i;
2875 gimple stmt;
2876
2877 def = PHI_RESULT (phi);
2878 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2879 {
2880 stmt = USE_STMT (use_p);
2881
2882 if (is_gimple_debug (stmt))
2883 {
2884 gimple_debug_bind_reset_value (stmt);
2885 update.safe_push (stmt);
2886 }
2887 }
2888
2889 FOR_EACH_VEC_ELT (update, i, stmt)
2890 update_stmt (stmt);
2891
2892 gsi = gsi_for_phi_node (phi);
2893 remove_phi_node (&gsi, false);
2894 }
2895
2896 /* Helper function for for_each_index. For each INDEX of the data
2897 reference REF, returns true when its indices are valid in the loop
2898 nest LOOP passed in as DATA. */
2899
2900 static bool
2901 dr_indices_valid_in_loop (tree ref ATTRIBUTE_UNUSED, tree *index, void *data)
2902 {
2903 loop_p loop;
2904 basic_block header, def_bb;
2905 gimple stmt;
2906
2907 if (TREE_CODE (*index) != SSA_NAME)
2908 return true;
2909
2910 loop = *((loop_p *) data);
2911 header = loop->header;
2912 stmt = SSA_NAME_DEF_STMT (*index);
2913
2914 if (!stmt)
2915 return true;
2916
2917 def_bb = gimple_bb (stmt);
2918
2919 if (!def_bb)
2920 return true;
2921
2922 return dominated_by_p (CDI_DOMINATORS, header, def_bb);
2923 }
2924
2925 /* When the result of a CLOSE_PHI is written to a memory location,
2926 return a pointer to that memory reference, otherwise return
2927 NULL_TREE. */
2928
2929 static tree
2930 close_phi_written_to_memory (gimple close_phi)
2931 {
2932 imm_use_iterator imm_iter;
2933 use_operand_p use_p;
2934 gimple stmt;
2935 tree res, def = gimple_phi_result (close_phi);
2936
2937 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2938 if ((stmt = USE_STMT (use_p))
2939 && gimple_code (stmt) == GIMPLE_ASSIGN
2940 && (res = gimple_assign_lhs (stmt)))
2941 {
2942 switch (TREE_CODE (res))
2943 {
2944 case VAR_DECL:
2945 case PARM_DECL:
2946 case RESULT_DECL:
2947 return res;
2948
2949 case ARRAY_REF:
2950 case MEM_REF:
2951 {
2952 tree arg = gimple_phi_arg_def (close_phi, 0);
2953 loop_p nest = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2954
2955 /* FIXME: this restriction is for id-{24,25}.f and
2956 could be handled by duplicating the computation of
2957 array indices before the loop of the close_phi. */
2958 if (for_each_index (&res, dr_indices_valid_in_loop, &nest))
2959 return res;
2960 }
2961 /* Fallthru. */
2962
2963 default:
2964 continue;
2965 }
2966 }
2967 return NULL_TREE;
2968 }
2969
2970 /* Rewrite out of SSA the reduction described by the loop phi nodes
2971 IN, and the close phi nodes OUT. IN and OUT are structured by loop
2972 levels like this:
2973
2974 IN: stmt, loop_n, ..., loop_0
2975 OUT: stmt, close_n, ..., close_0
2976
2977 the first element is the reduction statement, and the next elements
2978 are the loop and close phi nodes of each of the outer loops. */
2979
2980 static void
2981 translate_scalar_reduction_to_array (scop_p scop,
2982 vec<gimple> in,
2983 vec<gimple> out)
2984 {
2985 gimple loop_phi;
2986 unsigned int i = out.length () - 1;
2987 tree red = close_phi_written_to_memory (out[i]);
2988
2989 FOR_EACH_VEC_ELT (in, i, loop_phi)
2990 {
2991 gimple close_phi = out[i];
2992
2993 if (i == 0)
2994 {
2995 gimple stmt = loop_phi;
2996 basic_block bb = split_reduction_stmt (scop, stmt);
2997 poly_bb_p pbb = pbb_from_bb (bb);
2998 PBB_IS_REDUCTION (pbb) = true;
2999 gcc_assert (close_phi == loop_phi);
3000
3001 if (!red)
3002 red = create_zero_dim_array
3003 (gimple_assign_lhs (stmt), "Commutative_Associative_Reduction");
3004
3005 translate_scalar_reduction_to_array_for_stmt (scop, red, stmt, in[1]);
3006 continue;
3007 }
3008
3009 if (i == in.length () - 1)
3010 {
3011 insert_out_of_ssa_copy (scop, gimple_phi_result (close_phi),
3012 unshare_expr (red), close_phi);
3013 insert_out_of_ssa_copy_on_edge
3014 (scop, edge_initial_value_for_loop_phi (loop_phi),
3015 unshare_expr (red), initial_value_for_loop_phi (loop_phi));
3016 }
3017
3018 remove_phi (loop_phi);
3019 remove_phi (close_phi);
3020 }
3021 }
3022
3023 /* Rewrites out of SSA a commutative reduction at CLOSE_PHI. Returns
3024 true when something has been changed. */
3025
3026 static bool
3027 rewrite_commutative_reductions_out_of_ssa_close_phi (scop_p scop,
3028 gimple close_phi)
3029 {
3030 bool res;
3031 auto_vec<gimple, 10> in;
3032 auto_vec<gimple, 10> out;
3033
3034 detect_commutative_reduction (scop, close_phi, &in, &out);
3035 res = in.length () > 1;
3036 if (res)
3037 translate_scalar_reduction_to_array (scop, in, out);
3038
3039 return res;
3040 }
3041
3042 /* Rewrites all the commutative reductions from LOOP out of SSA.
3043 Returns true when something has been changed. */
3044
3045 static bool
3046 rewrite_commutative_reductions_out_of_ssa_loop (scop_p scop,
3047 loop_p loop)
3048 {
3049 gimple_stmt_iterator gsi;
3050 edge exit = single_exit (loop);
3051 tree res;
3052 bool changed = false;
3053
3054 if (!exit)
3055 return false;
3056
3057 for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi))
3058 if ((res = gimple_phi_result (gsi_stmt (gsi)))
3059 && !virtual_operand_p (res)
3060 && !scev_analyzable_p (res, SCOP_REGION (scop)))
3061 changed |= rewrite_commutative_reductions_out_of_ssa_close_phi
3062 (scop, gsi_stmt (gsi));
3063
3064 return changed;
3065 }
3066
3067 /* Rewrites all the commutative reductions from SCOP out of SSA. */
3068
3069 static void
3070 rewrite_commutative_reductions_out_of_ssa (scop_p scop)
3071 {
3072 loop_p loop;
3073 bool changed = false;
3074 sese region = SCOP_REGION (scop);
3075
3076 FOR_EACH_LOOP (loop, 0)
3077 if (loop_in_sese_p (loop, region))
3078 changed |= rewrite_commutative_reductions_out_of_ssa_loop (scop, loop);
3079
3080 if (changed)
3081 {
3082 scev_reset_htab ();
3083 gsi_commit_edge_inserts ();
3084 update_ssa (TODO_update_ssa);
3085 #ifdef ENABLE_CHECKING
3086 verify_loop_closed_ssa (true);
3087 #endif
3088 }
3089 }
3090
3091 /* Can all ivs be represented by a signed integer?
3092 As CLooG might generate negative values in its expressions, signed loop ivs
3093 are required in the backend. */
3094
3095 static bool
3096 scop_ivs_can_be_represented (scop_p scop)
3097 {
3098 loop_p loop;
3099 gimple_stmt_iterator psi;
3100 bool result = true;
3101
3102 FOR_EACH_LOOP (loop, 0)
3103 {
3104 if (!loop_in_sese_p (loop, SCOP_REGION (scop)))
3105 continue;
3106
3107 for (psi = gsi_start_phis (loop->header);
3108 !gsi_end_p (psi); gsi_next (&psi))
3109 {
3110 gimple phi = gsi_stmt (psi);
3111 tree res = PHI_RESULT (phi);
3112 tree type = TREE_TYPE (res);
3113
3114 if (TYPE_UNSIGNED (type)
3115 && TYPE_PRECISION (type) >= TYPE_PRECISION (long_long_integer_type_node))
3116 {
3117 result = false;
3118 break;
3119 }
3120 }
3121 if (!result)
3122 break;
3123 }
3124
3125 return result;
3126 }
3127
3128 /* Builds the polyhedral representation for a SESE region. */
3129
3130 void
3131 build_poly_scop (scop_p scop)
3132 {
3133 sese region = SCOP_REGION (scop);
3134 graphite_dim_t max_dim;
3135
3136 build_scop_bbs (scop);
3137
3138 /* FIXME: This restriction is needed to avoid a problem in CLooG.
3139 Once CLooG is fixed, remove this guard. Anyways, it makes no
3140 sense to optimize a scop containing only PBBs that do not belong
3141 to any loops. */
3142 if (nb_pbbs_in_loops (scop) == 0)
3143 return;
3144
3145 if (!scop_ivs_can_be_represented (scop))
3146 return;
3147
3148 if (flag_associative_math)
3149 rewrite_commutative_reductions_out_of_ssa (scop);
3150
3151 build_sese_loop_nests (region);
3152 /* Record all conditions in REGION. */
3153 sese_dom_walker (CDI_DOMINATORS, region).walk (cfun->cfg->x_entry_block_ptr);
3154 find_scop_parameters (scop);
3155
3156 max_dim = PARAM_VALUE (PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS);
3157 if (scop_nb_params (scop) > max_dim)
3158 return;
3159
3160 build_scop_iteration_domain (scop);
3161 build_scop_context (scop);
3162 add_conditions_to_constraints (scop);
3163
3164 /* Rewrite out of SSA only after having translated the
3165 representation to the polyhedral representation to avoid scev
3166 analysis failures. That means that these functions will insert
3167 new data references that they create in the right place. */
3168 rewrite_reductions_out_of_ssa (scop);
3169 rewrite_cross_bb_scalar_deps_out_of_ssa (scop);
3170
3171 build_scop_drs (scop);
3172 scop_to_lst (scop);
3173 build_scop_scattering (scop);
3174
3175 /* This SCoP has been translated to the polyhedral
3176 representation. */
3177 POLY_SCOP_P (scop) = true;
3178 }
3179 #endif