Merger of git branch "gimple-classes-v2-option-3"
[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_isl
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 <isl/val.h>
30 /* For C++ linkage of C functions.
31 Missing from isl/val_gmp.h in isl 0.12 versions.
32 Appearing in isl/val_gmp.h in isl 0.13.
33 To be removed when passing to isl 0.13. */
34 #if defined(__cplusplus)
35 extern "C" {
36 #endif
37 #include <isl/val_gmp.h>
38 #if defined(__cplusplus)
39 }
40 #endif
41 #endif
42
43 #include "system.h"
44 #include "coretypes.h"
45 #include "tree.h"
46 #include "predict.h"
47 #include "vec.h"
48 #include "hashtab.h"
49 #include "hash-set.h"
50 #include "machmode.h"
51 #include "tm.h"
52 #include "hard-reg-set.h"
53 #include "input.h"
54 #include "function.h"
55 #include "dominance.h"
56 #include "cfg.h"
57 #include "basic-block.h"
58 #include "tree-ssa-alias.h"
59 #include "internal-fn.h"
60 #include "gimple-expr.h"
61 #include "is-a.h"
62 #include "gimple.h"
63 #include "gimple-iterator.h"
64 #include "gimplify.h"
65 #include "gimplify-me.h"
66 #include "gimple-ssa.h"
67 #include "tree-cfg.h"
68 #include "tree-phinodes.h"
69 #include "ssa-iterators.h"
70 #include "stringpool.h"
71 #include "tree-ssanames.h"
72 #include "tree-ssa-loop-manip.h"
73 #include "tree-ssa-loop-niter.h"
74 #include "tree-ssa-loop.h"
75 #include "tree-into-ssa.h"
76 #include "tree-pass.h"
77 #include "cfgloop.h"
78 #include "tree-chrec.h"
79 #include "tree-data-ref.h"
80 #include "tree-scalar-evolution.h"
81 #include "domwalk.h"
82 #include "sese.h"
83 #include "tree-ssa-propagate.h"
84
85 #ifdef HAVE_isl
86 #include "expr.h"
87 #include "graphite-poly.h"
88 #include "graphite-sese-to-poly.h"
89
90
91 /* Assigns to RES the value of the INTEGER_CST T. */
92
93 static inline void
94 tree_int_to_gmp (tree t, mpz_t res)
95 {
96 wi::to_mpz (t, res, TYPE_SIGN (TREE_TYPE (t)));
97 }
98
99 /* Returns the index of the PHI argument defined in the outermost
100 loop. */
101
102 static size_t
103 phi_arg_in_outermost_loop (gphi *phi)
104 {
105 loop_p loop = gimple_bb (phi)->loop_father;
106 size_t i, res = 0;
107
108 for (i = 0; i < gimple_phi_num_args (phi); i++)
109 if (!flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, i)->src))
110 {
111 loop = gimple_phi_arg_edge (phi, i)->src->loop_father;
112 res = i;
113 }
114
115 return res;
116 }
117
118 /* Removes a simple copy phi node "RES = phi (INIT, RES)" at position
119 PSI by inserting on the loop ENTRY edge assignment "RES = INIT". */
120
121 static void
122 remove_simple_copy_phi (gphi_iterator *psi)
123 {
124 gphi *phi = psi->phi ();
125 tree res = gimple_phi_result (phi);
126 size_t entry = phi_arg_in_outermost_loop (phi);
127 tree init = gimple_phi_arg_def (phi, entry);
128 gassign *stmt = gimple_build_assign (res, init);
129 edge e = gimple_phi_arg_edge (phi, entry);
130
131 remove_phi_node (psi, false);
132 gsi_insert_on_edge_immediate (e, stmt);
133 }
134
135 /* Removes an invariant phi node at position PSI by inserting on the
136 loop ENTRY edge the assignment RES = INIT. */
137
138 static void
139 remove_invariant_phi (sese region, gphi_iterator *psi)
140 {
141 gphi *phi = psi->phi ();
142 loop_p loop = loop_containing_stmt (phi);
143 tree res = gimple_phi_result (phi);
144 tree scev = scalar_evolution_in_region (region, loop, res);
145 size_t entry = phi_arg_in_outermost_loop (phi);
146 edge e = gimple_phi_arg_edge (phi, entry);
147 tree var;
148 gassign *stmt;
149 gimple_seq stmts = NULL;
150
151 if (tree_contains_chrecs (scev, NULL))
152 scev = gimple_phi_arg_def (phi, entry);
153
154 var = force_gimple_operand (scev, &stmts, true, NULL_TREE);
155 stmt = gimple_build_assign (res, var);
156 remove_phi_node (psi, false);
157
158 gimple_seq_add_stmt (&stmts, stmt);
159 gsi_insert_seq_on_edge (e, stmts);
160 gsi_commit_edge_inserts ();
161 SSA_NAME_DEF_STMT (res) = stmt;
162 }
163
164 /* Returns true when the phi node at PSI is of the form "a = phi (a, x)". */
165
166 static inline bool
167 simple_copy_phi_p (gphi *phi)
168 {
169 tree res;
170
171 if (gimple_phi_num_args (phi) != 2)
172 return false;
173
174 res = gimple_phi_result (phi);
175 return (res == gimple_phi_arg_def (phi, 0)
176 || res == gimple_phi_arg_def (phi, 1));
177 }
178
179 /* Returns true when the phi node at position PSI is a reduction phi
180 node in REGION. Otherwise moves the pointer PSI to the next phi to
181 be considered. */
182
183 static bool
184 reduction_phi_p (sese region, gphi_iterator *psi)
185 {
186 loop_p loop;
187 gphi *phi = psi->phi ();
188 tree res = gimple_phi_result (phi);
189
190 loop = loop_containing_stmt (phi);
191
192 if (simple_copy_phi_p (phi))
193 {
194 /* PRE introduces phi nodes like these, for an example,
195 see id-5.f in the fortran graphite testsuite:
196
197 # prephitmp.85_265 = PHI <prephitmp.85_258(33), prephitmp.85_265(18)>
198 */
199 remove_simple_copy_phi (psi);
200 return false;
201 }
202
203 if (scev_analyzable_p (res, region))
204 {
205 tree scev = scalar_evolution_in_region (region, loop, res);
206
207 if (evolution_function_is_invariant_p (scev, loop->num))
208 remove_invariant_phi (region, psi);
209 else
210 gsi_next (psi);
211
212 return false;
213 }
214
215 /* All the other cases are considered reductions. */
216 return true;
217 }
218
219 /* Store the GRAPHITE representation of BB. */
220
221 static gimple_bb_p
222 new_gimple_bb (basic_block bb, vec<data_reference_p> drs)
223 {
224 struct gimple_bb *gbb;
225
226 gbb = XNEW (struct gimple_bb);
227 bb->aux = gbb;
228 GBB_BB (gbb) = bb;
229 GBB_DATA_REFS (gbb) = drs;
230 GBB_CONDITIONS (gbb).create (0);
231 GBB_CONDITION_CASES (gbb).create (0);
232
233 return gbb;
234 }
235
236 static void
237 free_data_refs_aux (vec<data_reference_p> datarefs)
238 {
239 unsigned int i;
240 struct data_reference *dr;
241
242 FOR_EACH_VEC_ELT (datarefs, i, dr)
243 if (dr->aux)
244 {
245 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
246
247 free (bap->alias_set);
248
249 free (bap);
250 dr->aux = NULL;
251 }
252 }
253 /* Frees GBB. */
254
255 static void
256 free_gimple_bb (struct gimple_bb *gbb)
257 {
258 free_data_refs_aux (GBB_DATA_REFS (gbb));
259 free_data_refs (GBB_DATA_REFS (gbb));
260
261 GBB_CONDITIONS (gbb).release ();
262 GBB_CONDITION_CASES (gbb).release ();
263 GBB_BB (gbb)->aux = 0;
264 XDELETE (gbb);
265 }
266
267 /* Deletes all gimple bbs in SCOP. */
268
269 static void
270 remove_gbbs_in_scop (scop_p scop)
271 {
272 int i;
273 poly_bb_p pbb;
274
275 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
276 free_gimple_bb (PBB_BLACK_BOX (pbb));
277 }
278
279 /* Deletes all scops in SCOPS. */
280
281 void
282 free_scops (vec<scop_p> scops)
283 {
284 int i;
285 scop_p scop;
286
287 FOR_EACH_VEC_ELT (scops, i, scop)
288 {
289 remove_gbbs_in_scop (scop);
290 free_sese (SCOP_REGION (scop));
291 free_scop (scop);
292 }
293
294 scops.release ();
295 }
296
297 /* Same as outermost_loop_in_sese, returns the outermost loop
298 containing BB in REGION, but makes sure that the returned loop
299 belongs to the REGION, and so this returns the first loop in the
300 REGION when the loop containing BB does not belong to REGION. */
301
302 static loop_p
303 outermost_loop_in_sese_1 (sese region, basic_block bb)
304 {
305 loop_p nest = outermost_loop_in_sese (region, bb);
306
307 if (loop_in_sese_p (nest, region))
308 return nest;
309
310 /* When the basic block BB does not belong to a loop in the region,
311 return the first loop in the region. */
312 nest = nest->inner;
313 while (nest)
314 if (loop_in_sese_p (nest, region))
315 break;
316 else
317 nest = nest->next;
318
319 gcc_assert (nest);
320 return nest;
321 }
322
323 /* Generates a polyhedral black box only if the bb contains interesting
324 information. */
325
326 static gimple_bb_p
327 try_generate_gimple_bb (scop_p scop, basic_block bb)
328 {
329 vec<data_reference_p> drs;
330 drs.create (5);
331 sese region = SCOP_REGION (scop);
332 loop_p nest = outermost_loop_in_sese_1 (region, bb);
333 gimple_stmt_iterator gsi;
334
335 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
336 {
337 gimple stmt = gsi_stmt (gsi);
338 loop_p loop;
339
340 if (is_gimple_debug (stmt))
341 continue;
342
343 loop = loop_containing_stmt (stmt);
344 if (!loop_in_sese_p (loop, region))
345 loop = nest;
346
347 graphite_find_data_references_in_stmt (nest, loop, stmt, &drs);
348 }
349
350 return new_gimple_bb (bb, drs);
351 }
352
353 /* Returns true if all predecessors of BB, that are not dominated by BB, are
354 marked in MAP. The predecessors dominated by BB are loop latches and will
355 be handled after BB. */
356
357 static bool
358 all_non_dominated_preds_marked_p (basic_block bb, sbitmap map)
359 {
360 edge e;
361 edge_iterator ei;
362
363 FOR_EACH_EDGE (e, ei, bb->preds)
364 if (!bitmap_bit_p (map, e->src->index)
365 && !dominated_by_p (CDI_DOMINATORS, e->src, bb))
366 return false;
367
368 return true;
369 }
370
371 /* Compare the depth of two basic_block's P1 and P2. */
372
373 static int
374 compare_bb_depths (const void *p1, const void *p2)
375 {
376 const_basic_block const bb1 = *(const_basic_block const*)p1;
377 const_basic_block const bb2 = *(const_basic_block const*)p2;
378 int d1 = loop_depth (bb1->loop_father);
379 int d2 = loop_depth (bb2->loop_father);
380
381 if (d1 < d2)
382 return 1;
383
384 if (d1 > d2)
385 return -1;
386
387 return 0;
388 }
389
390 /* Sort the basic blocks from DOM such that the first are the ones at
391 a deepest loop level. */
392
393 static void
394 graphite_sort_dominated_info (vec<basic_block> dom)
395 {
396 dom.qsort (compare_bb_depths);
397 }
398
399 /* Recursive helper function for build_scops_bbs. */
400
401 static void
402 build_scop_bbs_1 (scop_p scop, sbitmap visited, basic_block bb)
403 {
404 sese region = SCOP_REGION (scop);
405 vec<basic_block> dom;
406 poly_bb_p pbb;
407
408 if (bitmap_bit_p (visited, bb->index)
409 || !bb_in_sese_p (bb, region))
410 return;
411
412 pbb = new_poly_bb (scop, try_generate_gimple_bb (scop, bb));
413 SCOP_BBS (scop).safe_push (pbb);
414 bitmap_set_bit (visited, bb->index);
415
416 dom = get_dominated_by (CDI_DOMINATORS, bb);
417
418 if (!dom.exists ())
419 return;
420
421 graphite_sort_dominated_info (dom);
422
423 while (!dom.is_empty ())
424 {
425 int i;
426 basic_block dom_bb;
427
428 FOR_EACH_VEC_ELT (dom, i, dom_bb)
429 if (all_non_dominated_preds_marked_p (dom_bb, visited))
430 {
431 build_scop_bbs_1 (scop, visited, dom_bb);
432 dom.unordered_remove (i);
433 break;
434 }
435 }
436
437 dom.release ();
438 }
439
440 /* Gather the basic blocks belonging to the SCOP. */
441
442 static void
443 build_scop_bbs (scop_p scop)
444 {
445 sbitmap visited = sbitmap_alloc (last_basic_block_for_fn (cfun));
446 sese region = SCOP_REGION (scop);
447
448 bitmap_clear (visited);
449 build_scop_bbs_1 (scop, visited, SESE_ENTRY_BB (region));
450 sbitmap_free (visited);
451 }
452
453 /* Return an ISL identifier for the polyhedral basic block PBB. */
454
455 static isl_id *
456 isl_id_for_pbb (scop_p s, poly_bb_p pbb)
457 {
458 char name[50];
459 snprintf (name, sizeof (name), "S_%d", pbb_index (pbb));
460 return isl_id_alloc (s->ctx, name, pbb);
461 }
462
463 /* Converts the STATIC_SCHEDULE of PBB into a scattering polyhedron.
464 We generate SCATTERING_DIMENSIONS scattering dimensions.
465
466 CLooG 0.15.0 and previous versions require, that all
467 scattering functions of one CloogProgram have the same number of
468 scattering dimensions, therefore we allow to specify it. This
469 should be removed in future versions of CLooG.
470
471 The scattering polyhedron consists of these dimensions: scattering,
472 loop_iterators, parameters.
473
474 Example:
475
476 | scattering_dimensions = 5
477 | used_scattering_dimensions = 3
478 | nb_iterators = 1
479 | scop_nb_params = 2
480 |
481 | Schedule:
482 | i
483 | 4 5
484 |
485 | Scattering polyhedron:
486 |
487 | scattering: {s1, s2, s3, s4, s5}
488 | loop_iterators: {i}
489 | parameters: {p1, p2}
490 |
491 | s1 s2 s3 s4 s5 i p1 p2 1
492 | 1 0 0 0 0 0 0 0 -4 = 0
493 | 0 1 0 0 0 -1 0 0 0 = 0
494 | 0 0 1 0 0 0 0 0 -5 = 0 */
495
496 static void
497 build_pbb_scattering_polyhedrons (isl_aff *static_sched,
498 poly_bb_p pbb, int scattering_dimensions)
499 {
500 int i;
501 int nb_iterators = pbb_dim_iter_domain (pbb);
502 int used_scattering_dimensions = nb_iterators * 2 + 1;
503 isl_val *val;
504 isl_space *dc, *dm;
505
506 gcc_assert (scattering_dimensions >= used_scattering_dimensions);
507
508 dc = isl_set_get_space (pbb->domain);
509 dm = isl_space_add_dims (isl_space_from_domain (dc),
510 isl_dim_out, scattering_dimensions);
511 pbb->schedule = isl_map_universe (dm);
512
513 for (i = 0; i < scattering_dimensions; i++)
514 {
515 /* Textual order inside this loop. */
516 if ((i % 2) == 0)
517 {
518 isl_constraint *c = isl_equality_alloc
519 (isl_local_space_from_space (isl_map_get_space (pbb->schedule)));
520
521 val = isl_aff_get_coefficient_val (static_sched, isl_dim_in, i / 2);
522
523 val = isl_val_neg (val);
524 c = isl_constraint_set_constant_val (c, val);
525 c = isl_constraint_set_coefficient_si (c, isl_dim_out, i, 1);
526 pbb->schedule = isl_map_add_constraint (pbb->schedule, c);
527 }
528
529 /* Iterations of this loop. */
530 else /* if ((i % 2) == 1) */
531 {
532 int loop = (i - 1) / 2;
533 pbb->schedule = isl_map_equate (pbb->schedule, isl_dim_in, loop,
534 isl_dim_out, i);
535 }
536 }
537
538 pbb->transformed = isl_map_copy (pbb->schedule);
539 }
540
541 /* Build for BB the static schedule.
542
543 The static schedule is a Dewey numbering of the abstract syntax
544 tree: http://en.wikipedia.org/wiki/Dewey_Decimal_Classification
545
546 The following example informally defines the static schedule:
547
548 A
549 for (i: ...)
550 {
551 for (j: ...)
552 {
553 B
554 C
555 }
556
557 for (k: ...)
558 {
559 D
560 E
561 }
562 }
563 F
564
565 Static schedules for A to F:
566
567 DEPTH
568 0 1 2
569 A 0
570 B 1 0 0
571 C 1 0 1
572 D 1 1 0
573 E 1 1 1
574 F 2
575 */
576
577 static void
578 build_scop_scattering (scop_p scop)
579 {
580 int i;
581 poly_bb_p pbb;
582 gimple_bb_p previous_gbb = NULL;
583 isl_space *dc = isl_set_get_space (scop->context);
584 isl_aff *static_sched;
585
586 dc = isl_space_add_dims (dc, isl_dim_set, number_of_loops (cfun));
587 static_sched = isl_aff_zero_on_domain (isl_local_space_from_space (dc));
588
589 /* We have to start schedules at 0 on the first component and
590 because we cannot compare_prefix_loops against a previous loop,
591 prefix will be equal to zero, and that index will be
592 incremented before copying. */
593 static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in, 0, -1);
594
595 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
596 {
597 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
598 int prefix;
599 int nb_scat_dims = pbb_dim_iter_domain (pbb) * 2 + 1;
600
601 if (previous_gbb)
602 prefix = nb_common_loops (SCOP_REGION (scop), previous_gbb, gbb);
603 else
604 prefix = 0;
605
606 previous_gbb = gbb;
607
608 static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in,
609 prefix, 1);
610 build_pbb_scattering_polyhedrons (static_sched, pbb, nb_scat_dims);
611 }
612
613 isl_aff_free (static_sched);
614 }
615
616 static isl_pw_aff *extract_affine (scop_p, tree, __isl_take isl_space *space);
617
618 /* Extract an affine expression from the chain of recurrence E. */
619
620 static isl_pw_aff *
621 extract_affine_chrec (scop_p s, tree e, __isl_take isl_space *space)
622 {
623 isl_pw_aff *lhs = extract_affine (s, CHREC_LEFT (e), isl_space_copy (space));
624 isl_pw_aff *rhs = extract_affine (s, CHREC_RIGHT (e), isl_space_copy (space));
625 isl_local_space *ls = isl_local_space_from_space (space);
626 unsigned pos = sese_loop_depth ((sese) s->region, get_chrec_loop (e)) - 1;
627 isl_aff *loop = isl_aff_set_coefficient_si
628 (isl_aff_zero_on_domain (ls), isl_dim_in, pos, 1);
629 isl_pw_aff *l = isl_pw_aff_from_aff (loop);
630
631 /* Before multiplying, make sure that the result is affine. */
632 gcc_assert (isl_pw_aff_is_cst (rhs)
633 || isl_pw_aff_is_cst (l));
634
635 return isl_pw_aff_add (lhs, isl_pw_aff_mul (rhs, l));
636 }
637
638 /* Extract an affine expression from the mult_expr E. */
639
640 static isl_pw_aff *
641 extract_affine_mul (scop_p s, tree e, __isl_take isl_space *space)
642 {
643 isl_pw_aff *lhs = extract_affine (s, TREE_OPERAND (e, 0),
644 isl_space_copy (space));
645 isl_pw_aff *rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
646
647 if (!isl_pw_aff_is_cst (lhs)
648 && !isl_pw_aff_is_cst (rhs))
649 {
650 isl_pw_aff_free (lhs);
651 isl_pw_aff_free (rhs);
652 return NULL;
653 }
654
655 return isl_pw_aff_mul (lhs, rhs);
656 }
657
658 /* Return an ISL identifier from the name of the ssa_name E. */
659
660 static isl_id *
661 isl_id_for_ssa_name (scop_p s, tree e)
662 {
663 const char *name = get_name (e);
664 isl_id *id;
665
666 if (name)
667 id = isl_id_alloc (s->ctx, name, e);
668 else
669 {
670 char name1[50];
671 snprintf (name1, sizeof (name1), "P_%d", SSA_NAME_VERSION (e));
672 id = isl_id_alloc (s->ctx, name1, e);
673 }
674
675 return id;
676 }
677
678 /* Return an ISL identifier for the data reference DR. */
679
680 static isl_id *
681 isl_id_for_dr (scop_p s, data_reference_p dr ATTRIBUTE_UNUSED)
682 {
683 /* Data references all get the same isl_id. They need to be comparable
684 and are distinguished through the first dimension, which contains the
685 alias set number. */
686 return isl_id_alloc (s->ctx, "", 0);
687 }
688
689 /* Extract an affine expression from the ssa_name E. */
690
691 static isl_pw_aff *
692 extract_affine_name (scop_p s, tree e, __isl_take isl_space *space)
693 {
694 isl_aff *aff;
695 isl_set *dom;
696 isl_id *id;
697 int dimension;
698
699 id = isl_id_for_ssa_name (s, e);
700 dimension = isl_space_find_dim_by_id (space, isl_dim_param, id);
701 isl_id_free (id);
702 dom = isl_set_universe (isl_space_copy (space));
703 aff = isl_aff_zero_on_domain (isl_local_space_from_space (space));
704 aff = isl_aff_add_coefficient_si (aff, isl_dim_param, dimension, 1);
705 return isl_pw_aff_alloc (dom, aff);
706 }
707
708 /* Extract an affine expression from the gmp constant G. */
709
710 static isl_pw_aff *
711 extract_affine_gmp (mpz_t g, __isl_take isl_space *space)
712 {
713 isl_local_space *ls = isl_local_space_from_space (isl_space_copy (space));
714 isl_aff *aff = isl_aff_zero_on_domain (ls);
715 isl_set *dom = isl_set_universe (space);
716 isl_val *v;
717 isl_ctx *ct;
718
719 ct = isl_aff_get_ctx (aff);
720 v = isl_val_int_from_gmp (ct, g);
721 aff = isl_aff_add_constant_val (aff, v);
722
723 return isl_pw_aff_alloc (dom, aff);
724 }
725
726 /* Extract an affine expression from the integer_cst E. */
727
728 static isl_pw_aff *
729 extract_affine_int (tree e, __isl_take isl_space *space)
730 {
731 isl_pw_aff *res;
732 mpz_t g;
733
734 mpz_init (g);
735 tree_int_to_gmp (e, g);
736 res = extract_affine_gmp (g, space);
737 mpz_clear (g);
738
739 return res;
740 }
741
742 /* Compute pwaff mod 2^width. */
743
744 extern isl_ctx *the_isl_ctx;
745
746 static isl_pw_aff *
747 wrap (isl_pw_aff *pwaff, unsigned width)
748 {
749 isl_val *mod;
750
751 mod = isl_val_int_from_ui(the_isl_ctx, width);
752 mod = isl_val_2exp (mod);
753 pwaff = isl_pw_aff_mod_val (pwaff, mod);
754
755 return pwaff;
756 }
757
758 /* When parameter NAME is in REGION, returns its index in SESE_PARAMS.
759 Otherwise returns -1. */
760
761 static inline int
762 parameter_index_in_region_1 (tree name, sese region)
763 {
764 int i;
765 tree p;
766
767 gcc_assert (TREE_CODE (name) == SSA_NAME);
768
769 FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, p)
770 if (p == name)
771 return i;
772
773 return -1;
774 }
775
776 /* When the parameter NAME is in REGION, returns its index in
777 SESE_PARAMS. Otherwise this function inserts NAME in SESE_PARAMS
778 and returns the index of NAME. */
779
780 static int
781 parameter_index_in_region (tree name, sese region)
782 {
783 int i;
784
785 gcc_assert (TREE_CODE (name) == SSA_NAME);
786
787 i = parameter_index_in_region_1 (name, region);
788 if (i != -1)
789 return i;
790
791 gcc_assert (SESE_ADD_PARAMS (region));
792
793 i = SESE_PARAMS (region).length ();
794 SESE_PARAMS (region).safe_push (name);
795 return i;
796 }
797
798 /* Extract an affine expression from the tree E in the scop S. */
799
800 static isl_pw_aff *
801 extract_affine (scop_p s, tree e, __isl_take isl_space *space)
802 {
803 isl_pw_aff *lhs, *rhs, *res;
804 tree type;
805
806 if (e == chrec_dont_know) {
807 isl_space_free (space);
808 return NULL;
809 }
810
811 switch (TREE_CODE (e))
812 {
813 case POLYNOMIAL_CHREC:
814 res = extract_affine_chrec (s, e, space);
815 break;
816
817 case MULT_EXPR:
818 res = extract_affine_mul (s, e, space);
819 break;
820
821 case PLUS_EXPR:
822 case POINTER_PLUS_EXPR:
823 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
824 rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
825 res = isl_pw_aff_add (lhs, rhs);
826 break;
827
828 case MINUS_EXPR:
829 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
830 rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
831 res = isl_pw_aff_sub (lhs, rhs);
832 break;
833
834 case NEGATE_EXPR:
835 case BIT_NOT_EXPR:
836 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
837 rhs = extract_affine (s, integer_minus_one_node, space);
838 res = isl_pw_aff_mul (lhs, rhs);
839 break;
840
841 case SSA_NAME:
842 gcc_assert (-1 != parameter_index_in_region_1 (e, SCOP_REGION (s)));
843 res = extract_affine_name (s, e, space);
844 break;
845
846 case INTEGER_CST:
847 res = extract_affine_int (e, space);
848 /* No need to wrap a single integer. */
849 return res;
850
851 CASE_CONVERT:
852 case NON_LVALUE_EXPR:
853 res = extract_affine (s, TREE_OPERAND (e, 0), space);
854 break;
855
856 default:
857 gcc_unreachable ();
858 break;
859 }
860
861 type = TREE_TYPE (e);
862 if (TYPE_UNSIGNED (type))
863 res = wrap (res, TYPE_PRECISION (type));
864
865 return res;
866 }
867
868 /* In the context of sese S, scan the expression E and translate it to
869 a linear expression C. When parsing a symbolic multiplication, K
870 represents the constant multiplier of an expression containing
871 parameters. */
872
873 static void
874 scan_tree_for_params (sese s, tree e)
875 {
876 if (e == chrec_dont_know)
877 return;
878
879 switch (TREE_CODE (e))
880 {
881 case POLYNOMIAL_CHREC:
882 scan_tree_for_params (s, CHREC_LEFT (e));
883 break;
884
885 case MULT_EXPR:
886 if (chrec_contains_symbols (TREE_OPERAND (e, 0)))
887 scan_tree_for_params (s, TREE_OPERAND (e, 0));
888 else
889 scan_tree_for_params (s, TREE_OPERAND (e, 1));
890 break;
891
892 case PLUS_EXPR:
893 case POINTER_PLUS_EXPR:
894 case MINUS_EXPR:
895 scan_tree_for_params (s, TREE_OPERAND (e, 0));
896 scan_tree_for_params (s, TREE_OPERAND (e, 1));
897 break;
898
899 case NEGATE_EXPR:
900 case BIT_NOT_EXPR:
901 CASE_CONVERT:
902 case NON_LVALUE_EXPR:
903 scan_tree_for_params (s, TREE_OPERAND (e, 0));
904 break;
905
906 case SSA_NAME:
907 parameter_index_in_region (e, s);
908 break;
909
910 case INTEGER_CST:
911 case ADDR_EXPR:
912 break;
913
914 default:
915 gcc_unreachable ();
916 break;
917 }
918 }
919
920 /* Find parameters with respect to REGION in BB. We are looking in memory
921 access functions, conditions and loop bounds. */
922
923 static void
924 find_params_in_bb (sese region, gimple_bb_p gbb)
925 {
926 int i;
927 unsigned j;
928 data_reference_p dr;
929 gimple stmt;
930 loop_p loop = GBB_BB (gbb)->loop_father;
931
932 /* Find parameters in the access functions of data references. */
933 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr)
934 for (j = 0; j < DR_NUM_DIMENSIONS (dr); j++)
935 scan_tree_for_params (region, DR_ACCESS_FN (dr, j));
936
937 /* Find parameters in conditional statements. */
938 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt)
939 {
940 tree lhs = scalar_evolution_in_region (region, loop,
941 gimple_cond_lhs (stmt));
942 tree rhs = scalar_evolution_in_region (region, loop,
943 gimple_cond_rhs (stmt));
944
945 scan_tree_for_params (region, lhs);
946 scan_tree_for_params (region, rhs);
947 }
948 }
949
950 /* Record the parameters used in the SCOP. A variable is a parameter
951 in a scop if it does not vary during the execution of that scop. */
952
953 static void
954 find_scop_parameters (scop_p scop)
955 {
956 poly_bb_p pbb;
957 unsigned i;
958 sese region = SCOP_REGION (scop);
959 struct loop *loop;
960 int nbp;
961
962 /* Find the parameters used in the loop bounds. */
963 FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop)
964 {
965 tree nb_iters = number_of_latch_executions (loop);
966
967 if (!chrec_contains_symbols (nb_iters))
968 continue;
969
970 nb_iters = scalar_evolution_in_region (region, loop, nb_iters);
971 scan_tree_for_params (region, nb_iters);
972 }
973
974 /* Find the parameters used in data accesses. */
975 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
976 find_params_in_bb (region, PBB_BLACK_BOX (pbb));
977
978 nbp = sese_nb_params (region);
979 scop_set_nb_params (scop, nbp);
980 SESE_ADD_PARAMS (region) = false;
981
982 {
983 tree e;
984 isl_space *space = isl_space_set_alloc (scop->ctx, nbp, 0);
985
986 FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, e)
987 space = isl_space_set_dim_id (space, isl_dim_param, i,
988 isl_id_for_ssa_name (scop, e));
989
990 scop->context = isl_set_universe (space);
991 }
992 }
993
994 /* Builds the constraint polyhedra for LOOP in SCOP. OUTER_PH gives
995 the constraints for the surrounding loops. */
996
997 static void
998 build_loop_iteration_domains (scop_p scop, struct loop *loop,
999 int nb,
1000 isl_set *outer, isl_set **doms)
1001 {
1002 tree nb_iters = number_of_latch_executions (loop);
1003 sese region = SCOP_REGION (scop);
1004
1005 isl_set *inner = isl_set_copy (outer);
1006 isl_space *space;
1007 isl_constraint *c;
1008 int pos = isl_set_dim (outer, isl_dim_set);
1009 isl_val *v;
1010 mpz_t g;
1011
1012 mpz_init (g);
1013
1014 inner = isl_set_add_dims (inner, isl_dim_set, 1);
1015 space = isl_set_get_space (inner);
1016
1017 /* 0 <= loop_i */
1018 c = isl_inequality_alloc
1019 (isl_local_space_from_space (isl_space_copy (space)));
1020 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, 1);
1021 inner = isl_set_add_constraint (inner, c);
1022
1023 /* loop_i <= cst_nb_iters */
1024 if (TREE_CODE (nb_iters) == INTEGER_CST)
1025 {
1026 c = isl_inequality_alloc
1027 (isl_local_space_from_space (isl_space_copy (space)));
1028 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1);
1029 tree_int_to_gmp (nb_iters, g);
1030 v = isl_val_int_from_gmp (the_isl_ctx, g);
1031 c = isl_constraint_set_constant_val (c, v);
1032 inner = isl_set_add_constraint (inner, c);
1033 }
1034
1035 /* loop_i <= expr_nb_iters */
1036 else if (!chrec_contains_undetermined (nb_iters))
1037 {
1038 widest_int nit;
1039 isl_pw_aff *aff;
1040 isl_set *valid;
1041 isl_local_space *ls;
1042 isl_aff *al;
1043 isl_set *le;
1044
1045 nb_iters = scalar_evolution_in_region (region, loop, nb_iters);
1046
1047 aff = extract_affine (scop, nb_iters, isl_set_get_space (inner));
1048 valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (aff));
1049 valid = isl_set_project_out (valid, isl_dim_set, 0,
1050 isl_set_dim (valid, isl_dim_set));
1051 scop->context = isl_set_intersect (scop->context, valid);
1052
1053 ls = isl_local_space_from_space (isl_space_copy (space));
1054 al = isl_aff_set_coefficient_si (isl_aff_zero_on_domain (ls),
1055 isl_dim_in, pos, 1);
1056 le = isl_pw_aff_le_set (isl_pw_aff_from_aff (al),
1057 isl_pw_aff_copy (aff));
1058 inner = isl_set_intersect (inner, le);
1059
1060 if (max_stmt_executions (loop, &nit))
1061 {
1062 /* Insert in the context the constraints from the
1063 estimation of the number of iterations NIT and the
1064 symbolic number of iterations (involving parameter
1065 names) NB_ITERS. First, build the affine expression
1066 "NIT - NB_ITERS" and then say that it is positive,
1067 i.e., NIT approximates NB_ITERS: "NIT >= NB_ITERS". */
1068 isl_pw_aff *approx;
1069 mpz_t g;
1070 isl_set *x;
1071 isl_constraint *c;
1072
1073 mpz_init (g);
1074 wi::to_mpz (nit, g, SIGNED);
1075 mpz_sub_ui (g, g, 1);
1076 approx = extract_affine_gmp (g, isl_set_get_space (inner));
1077 x = isl_pw_aff_ge_set (approx, aff);
1078 x = isl_set_project_out (x, isl_dim_set, 0,
1079 isl_set_dim (x, isl_dim_set));
1080 scop->context = isl_set_intersect (scop->context, x);
1081
1082 c = isl_inequality_alloc
1083 (isl_local_space_from_space (isl_space_copy (space)));
1084 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1);
1085 v = isl_val_int_from_gmp (the_isl_ctx, g);
1086 mpz_clear (g);
1087 c = isl_constraint_set_constant_val (c, v);
1088 inner = isl_set_add_constraint (inner, c);
1089 }
1090 else
1091 isl_pw_aff_free (aff);
1092 }
1093 else
1094 gcc_unreachable ();
1095
1096 if (loop->inner && loop_in_sese_p (loop->inner, region))
1097 build_loop_iteration_domains (scop, loop->inner, nb + 1,
1098 isl_set_copy (inner), doms);
1099
1100 if (nb != 0
1101 && loop->next
1102 && loop_in_sese_p (loop->next, region))
1103 build_loop_iteration_domains (scop, loop->next, nb,
1104 isl_set_copy (outer), doms);
1105
1106 doms[loop->num] = inner;
1107
1108 isl_set_free (outer);
1109 isl_space_free (space);
1110 mpz_clear (g);
1111 }
1112
1113 /* Returns a linear expression for tree T evaluated in PBB. */
1114
1115 static isl_pw_aff *
1116 create_pw_aff_from_tree (poly_bb_p pbb, tree t)
1117 {
1118 scop_p scop = PBB_SCOP (pbb);
1119
1120 t = scalar_evolution_in_region (SCOP_REGION (scop), pbb_loop (pbb), t);
1121 gcc_assert (!automatically_generated_chrec_p (t));
1122
1123 return extract_affine (scop, t, isl_set_get_space (pbb->domain));
1124 }
1125
1126 /* Add conditional statement STMT to pbb. CODE is used as the comparison
1127 operator. This allows us to invert the condition or to handle
1128 inequalities. */
1129
1130 static void
1131 add_condition_to_pbb (poly_bb_p pbb, gcond *stmt, enum tree_code code)
1132 {
1133 isl_pw_aff *lhs = create_pw_aff_from_tree (pbb, gimple_cond_lhs (stmt));
1134 isl_pw_aff *rhs = create_pw_aff_from_tree (pbb, gimple_cond_rhs (stmt));
1135 isl_set *cond;
1136
1137 switch (code)
1138 {
1139 case LT_EXPR:
1140 cond = isl_pw_aff_lt_set (lhs, rhs);
1141 break;
1142
1143 case GT_EXPR:
1144 cond = isl_pw_aff_gt_set (lhs, rhs);
1145 break;
1146
1147 case LE_EXPR:
1148 cond = isl_pw_aff_le_set (lhs, rhs);
1149 break;
1150
1151 case GE_EXPR:
1152 cond = isl_pw_aff_ge_set (lhs, rhs);
1153 break;
1154
1155 case EQ_EXPR:
1156 cond = isl_pw_aff_eq_set (lhs, rhs);
1157 break;
1158
1159 case NE_EXPR:
1160 cond = isl_pw_aff_ne_set (lhs, rhs);
1161 break;
1162
1163 default:
1164 isl_pw_aff_free (lhs);
1165 isl_pw_aff_free (rhs);
1166 return;
1167 }
1168
1169 cond = isl_set_coalesce (cond);
1170 cond = isl_set_set_tuple_id (cond, isl_set_get_tuple_id (pbb->domain));
1171 pbb->domain = isl_set_intersect (pbb->domain, cond);
1172 }
1173
1174 /* Add conditions to the domain of PBB. */
1175
1176 static void
1177 add_conditions_to_domain (poly_bb_p pbb)
1178 {
1179 unsigned int i;
1180 gimple stmt;
1181 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
1182
1183 if (GBB_CONDITIONS (gbb).is_empty ())
1184 return;
1185
1186 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt)
1187 switch (gimple_code (stmt))
1188 {
1189 case GIMPLE_COND:
1190 {
1191 gcond *cond_stmt = as_a <gcond *> (stmt);
1192 enum tree_code code = gimple_cond_code (cond_stmt);
1193
1194 /* The conditions for ELSE-branches are inverted. */
1195 if (!GBB_CONDITION_CASES (gbb)[i])
1196 code = invert_tree_comparison (code, false);
1197
1198 add_condition_to_pbb (pbb, cond_stmt, code);
1199 break;
1200 }
1201
1202 case GIMPLE_SWITCH:
1203 /* Switch statements are not supported right now - fall through. */
1204
1205 default:
1206 gcc_unreachable ();
1207 break;
1208 }
1209 }
1210
1211 /* Traverses all the GBBs of the SCOP and add their constraints to the
1212 iteration domains. */
1213
1214 static void
1215 add_conditions_to_constraints (scop_p scop)
1216 {
1217 int i;
1218 poly_bb_p pbb;
1219
1220 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1221 add_conditions_to_domain (pbb);
1222 }
1223
1224 /* Returns a COND_EXPR statement when BB has a single predecessor, the
1225 edge between BB and its predecessor is not a loop exit edge, and
1226 the last statement of the single predecessor is a COND_EXPR. */
1227
1228 static gcond *
1229 single_pred_cond_non_loop_exit (basic_block bb)
1230 {
1231 if (single_pred_p (bb))
1232 {
1233 edge e = single_pred_edge (bb);
1234 basic_block pred = e->src;
1235 gimple stmt;
1236
1237 if (loop_depth (pred->loop_father) > loop_depth (bb->loop_father))
1238 return NULL;
1239
1240 stmt = last_stmt (pred);
1241
1242 if (stmt && gimple_code (stmt) == GIMPLE_COND)
1243 return as_a <gcond *> (stmt);
1244 }
1245
1246 return NULL;
1247 }
1248
1249 class sese_dom_walker : public dom_walker
1250 {
1251 public:
1252 sese_dom_walker (cdi_direction, sese);
1253
1254 virtual void before_dom_children (basic_block);
1255 virtual void after_dom_children (basic_block);
1256
1257 private:
1258 auto_vec<gimple, 3> m_conditions, m_cases;
1259 sese m_region;
1260 };
1261
1262 sese_dom_walker::sese_dom_walker (cdi_direction direction, sese region)
1263 : dom_walker (direction), m_region (region)
1264 {
1265 }
1266
1267 /* Call-back for dom_walk executed before visiting the dominated
1268 blocks. */
1269
1270 void
1271 sese_dom_walker::before_dom_children (basic_block bb)
1272 {
1273 gimple_bb_p gbb;
1274 gcond *stmt;
1275
1276 if (!bb_in_sese_p (bb, m_region))
1277 return;
1278
1279 stmt = single_pred_cond_non_loop_exit (bb);
1280
1281 if (stmt)
1282 {
1283 edge e = single_pred_edge (bb);
1284
1285 m_conditions.safe_push (stmt);
1286
1287 if (e->flags & EDGE_TRUE_VALUE)
1288 m_cases.safe_push (stmt);
1289 else
1290 m_cases.safe_push (NULL);
1291 }
1292
1293 gbb = gbb_from_bb (bb);
1294
1295 if (gbb)
1296 {
1297 GBB_CONDITIONS (gbb) = m_conditions.copy ();
1298 GBB_CONDITION_CASES (gbb) = m_cases.copy ();
1299 }
1300 }
1301
1302 /* Call-back for dom_walk executed after visiting the dominated
1303 blocks. */
1304
1305 void
1306 sese_dom_walker::after_dom_children (basic_block bb)
1307 {
1308 if (!bb_in_sese_p (bb, m_region))
1309 return;
1310
1311 if (single_pred_cond_non_loop_exit (bb))
1312 {
1313 m_conditions.pop ();
1314 m_cases.pop ();
1315 }
1316 }
1317
1318 /* Add constraints on the possible values of parameter P from the type
1319 of P. */
1320
1321 static void
1322 add_param_constraints (scop_p scop, graphite_dim_t p)
1323 {
1324 tree parameter = SESE_PARAMS (SCOP_REGION (scop))[p];
1325 tree type = TREE_TYPE (parameter);
1326 tree lb = NULL_TREE;
1327 tree ub = NULL_TREE;
1328
1329 if (POINTER_TYPE_P (type) || !TYPE_MIN_VALUE (type))
1330 lb = lower_bound_in_type (type, type);
1331 else
1332 lb = TYPE_MIN_VALUE (type);
1333
1334 if (POINTER_TYPE_P (type) || !TYPE_MAX_VALUE (type))
1335 ub = upper_bound_in_type (type, type);
1336 else
1337 ub = TYPE_MAX_VALUE (type);
1338
1339 if (lb)
1340 {
1341 isl_space *space = isl_set_get_space (scop->context);
1342 isl_constraint *c;
1343 mpz_t g;
1344 isl_val *v;
1345
1346 c = isl_inequality_alloc (isl_local_space_from_space (space));
1347 mpz_init (g);
1348 tree_int_to_gmp (lb, g);
1349 v = isl_val_int_from_gmp (the_isl_ctx, g);
1350 v = isl_val_neg (v);
1351 mpz_clear (g);
1352 c = isl_constraint_set_constant_val (c, v);
1353 c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, 1);
1354
1355 scop->context = isl_set_add_constraint (scop->context, c);
1356 }
1357
1358 if (ub)
1359 {
1360 isl_space *space = isl_set_get_space (scop->context);
1361 isl_constraint *c;
1362 mpz_t g;
1363 isl_val *v;
1364
1365 c = isl_inequality_alloc (isl_local_space_from_space (space));
1366
1367 mpz_init (g);
1368 tree_int_to_gmp (ub, g);
1369 v = isl_val_int_from_gmp (the_isl_ctx, g);
1370 mpz_clear (g);
1371 c = isl_constraint_set_constant_val (c, v);
1372 c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, -1);
1373
1374 scop->context = isl_set_add_constraint (scop->context, c);
1375 }
1376 }
1377
1378 /* Build the context of the SCOP. The context usually contains extra
1379 constraints that are added to the iteration domains that constrain
1380 some parameters. */
1381
1382 static void
1383 build_scop_context (scop_p scop)
1384 {
1385 graphite_dim_t p, n = scop_nb_params (scop);
1386
1387 for (p = 0; p < n; p++)
1388 add_param_constraints (scop, p);
1389 }
1390
1391 /* Build the iteration domains: the loops belonging to the current
1392 SCOP, and that vary for the execution of the current basic block.
1393 Returns false if there is no loop in SCOP. */
1394
1395 static void
1396 build_scop_iteration_domain (scop_p scop)
1397 {
1398 struct loop *loop;
1399 sese region = SCOP_REGION (scop);
1400 int i;
1401 poly_bb_p pbb;
1402 int nb_loops = number_of_loops (cfun);
1403 isl_set **doms = XCNEWVEC (isl_set *, nb_loops);
1404
1405 FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop)
1406 if (!loop_in_sese_p (loop_outer (loop), region))
1407 build_loop_iteration_domains (scop, loop, 0,
1408 isl_set_copy (scop->context), doms);
1409
1410 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1411 {
1412 loop = pbb_loop (pbb);
1413
1414 if (doms[loop->num])
1415 pbb->domain = isl_set_copy (doms[loop->num]);
1416 else
1417 pbb->domain = isl_set_copy (scop->context);
1418
1419 pbb->domain = isl_set_set_tuple_id (pbb->domain,
1420 isl_id_for_pbb (scop, pbb));
1421 }
1422
1423 for (i = 0; i < nb_loops; i++)
1424 if (doms[i])
1425 isl_set_free (doms[i]);
1426
1427 free (doms);
1428 }
1429
1430 /* Add a constrain to the ACCESSES polyhedron for the alias set of
1431 data reference DR. ACCESSP_NB_DIMS is the dimension of the
1432 ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration
1433 domain. */
1434
1435 static isl_map *
1436 pdr_add_alias_set (isl_map *acc, data_reference_p dr)
1437 {
1438 isl_constraint *c;
1439 int alias_set_num = 0;
1440 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
1441
1442 if (bap && bap->alias_set)
1443 alias_set_num = *(bap->alias_set);
1444
1445 c = isl_equality_alloc
1446 (isl_local_space_from_space (isl_map_get_space (acc)));
1447 c = isl_constraint_set_constant_si (c, -alias_set_num);
1448 c = isl_constraint_set_coefficient_si (c, isl_dim_out, 0, 1);
1449
1450 return isl_map_add_constraint (acc, c);
1451 }
1452
1453 /* Assign the affine expression INDEX to the output dimension POS of
1454 MAP and return the result. */
1455
1456 static isl_map *
1457 set_index (isl_map *map, int pos, isl_pw_aff *index)
1458 {
1459 isl_map *index_map;
1460 int len = isl_map_dim (map, isl_dim_out);
1461 isl_id *id;
1462
1463 index_map = isl_map_from_pw_aff (index);
1464 index_map = isl_map_insert_dims (index_map, isl_dim_out, 0, pos);
1465 index_map = isl_map_add_dims (index_map, isl_dim_out, len - pos - 1);
1466
1467 id = isl_map_get_tuple_id (map, isl_dim_out);
1468 index_map = isl_map_set_tuple_id (index_map, isl_dim_out, id);
1469 id = isl_map_get_tuple_id (map, isl_dim_in);
1470 index_map = isl_map_set_tuple_id (index_map, isl_dim_in, id);
1471
1472 return isl_map_intersect (map, index_map);
1473 }
1474
1475 /* Add to ACCESSES polyhedron equalities defining the access functions
1476 to the memory. ACCESSP_NB_DIMS is the dimension of the ACCESSES
1477 polyhedron, DOM_NB_DIMS is the dimension of the iteration domain.
1478 PBB is the poly_bb_p that contains the data reference DR. */
1479
1480 static isl_map *
1481 pdr_add_memory_accesses (isl_map *acc, data_reference_p dr, poly_bb_p pbb)
1482 {
1483 int i, nb_subscripts = DR_NUM_DIMENSIONS (dr);
1484 scop_p scop = PBB_SCOP (pbb);
1485
1486 for (i = 0; i < nb_subscripts; i++)
1487 {
1488 isl_pw_aff *aff;
1489 tree afn = DR_ACCESS_FN (dr, nb_subscripts - 1 - i);
1490
1491 aff = extract_affine (scop, afn,
1492 isl_space_domain (isl_map_get_space (acc)));
1493 acc = set_index (acc, i + 1, aff);
1494 }
1495
1496 return acc;
1497 }
1498
1499 /* Add constrains representing the size of the accessed data to the
1500 ACCESSES polyhedron. ACCESSP_NB_DIMS is the dimension of the
1501 ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration
1502 domain. */
1503
1504 static isl_set *
1505 pdr_add_data_dimensions (isl_set *extent, scop_p scop, data_reference_p dr)
1506 {
1507 tree ref = DR_REF (dr);
1508 int i, nb_subscripts = DR_NUM_DIMENSIONS (dr);
1509
1510 for (i = nb_subscripts - 1; i >= 0; i--, ref = TREE_OPERAND (ref, 0))
1511 {
1512 tree low, high;
1513
1514 if (TREE_CODE (ref) != ARRAY_REF)
1515 break;
1516
1517 low = array_ref_low_bound (ref);
1518 high = array_ref_up_bound (ref);
1519
1520 /* XXX The PPL code dealt separately with
1521 subscript - low >= 0 and high - subscript >= 0 in case one of
1522 the two bounds isn't known. Do the same here? */
1523
1524 if (tree_fits_shwi_p (low)
1525 && high
1526 && tree_fits_shwi_p (high)
1527 /* 1-element arrays at end of structures may extend over
1528 their declared size. */
1529 && !(array_at_struct_end_p (ref)
1530 && operand_equal_p (low, high, 0)))
1531 {
1532 isl_id *id;
1533 isl_aff *aff;
1534 isl_set *univ, *lbs, *ubs;
1535 isl_pw_aff *index;
1536 isl_space *space;
1537 isl_set *valid;
1538 isl_pw_aff *lb = extract_affine_int (low, isl_set_get_space (extent));
1539 isl_pw_aff *ub = extract_affine_int (high, isl_set_get_space (extent));
1540
1541 /* high >= 0 */
1542 valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (ub));
1543 valid = isl_set_project_out (valid, isl_dim_set, 0,
1544 isl_set_dim (valid, isl_dim_set));
1545 scop->context = isl_set_intersect (scop->context, valid);
1546
1547 space = isl_set_get_space (extent);
1548 aff = isl_aff_zero_on_domain (isl_local_space_from_space (space));
1549 aff = isl_aff_add_coefficient_si (aff, isl_dim_in, i + 1, 1);
1550 univ = isl_set_universe (isl_space_domain (isl_aff_get_space (aff)));
1551 index = isl_pw_aff_alloc (univ, aff);
1552
1553 id = isl_set_get_tuple_id (extent);
1554 lb = isl_pw_aff_set_tuple_id (lb, isl_dim_in, isl_id_copy (id));
1555 ub = isl_pw_aff_set_tuple_id (ub, isl_dim_in, id);
1556
1557 /* low <= sub_i <= high */
1558 lbs = isl_pw_aff_ge_set (isl_pw_aff_copy (index), lb);
1559 ubs = isl_pw_aff_le_set (index, ub);
1560 extent = isl_set_intersect (extent, lbs);
1561 extent = isl_set_intersect (extent, ubs);
1562 }
1563 }
1564
1565 return extent;
1566 }
1567
1568 /* Build data accesses for DR in PBB. */
1569
1570 static void
1571 build_poly_dr (data_reference_p dr, poly_bb_p pbb)
1572 {
1573 int dr_base_object_set;
1574 isl_map *acc;
1575 isl_set *extent;
1576 scop_p scop = PBB_SCOP (pbb);
1577
1578 {
1579 isl_space *dc = isl_set_get_space (pbb->domain);
1580 int nb_out = 1 + DR_NUM_DIMENSIONS (dr);
1581 isl_space *space = isl_space_add_dims (isl_space_from_domain (dc),
1582 isl_dim_out, nb_out);
1583
1584 acc = isl_map_universe (space);
1585 acc = isl_map_set_tuple_id (acc, isl_dim_out, isl_id_for_dr (scop, dr));
1586 }
1587
1588 acc = pdr_add_alias_set (acc, dr);
1589 acc = pdr_add_memory_accesses (acc, dr, pbb);
1590
1591 {
1592 isl_id *id = isl_id_for_dr (scop, dr);
1593 int nb = 1 + DR_NUM_DIMENSIONS (dr);
1594 isl_space *space = isl_space_set_alloc (scop->ctx, 0, nb);
1595 int alias_set_num = 0;
1596 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
1597
1598 if (bap && bap->alias_set)
1599 alias_set_num = *(bap->alias_set);
1600
1601 space = isl_space_set_tuple_id (space, isl_dim_set, id);
1602 extent = isl_set_nat_universe (space);
1603 extent = isl_set_fix_si (extent, isl_dim_set, 0, alias_set_num);
1604 extent = pdr_add_data_dimensions (extent, scop, dr);
1605 }
1606
1607 gcc_assert (dr->aux);
1608 dr_base_object_set = ((base_alias_pair *)(dr->aux))->base_obj_set;
1609
1610 new_poly_dr (pbb, dr_base_object_set,
1611 DR_IS_READ (dr) ? PDR_READ : PDR_WRITE,
1612 dr, DR_NUM_DIMENSIONS (dr), acc, extent);
1613 }
1614
1615 /* Write to FILE the alias graph of data references in DIMACS format. */
1616
1617 static inline bool
1618 write_alias_graph_to_ascii_dimacs (FILE *file, char *comment,
1619 vec<data_reference_p> drs)
1620 {
1621 int num_vertex = drs.length ();
1622 int edge_num = 0;
1623 data_reference_p dr1, dr2;
1624 int i, j;
1625
1626 if (num_vertex == 0)
1627 return true;
1628
1629 FOR_EACH_VEC_ELT (drs, i, dr1)
1630 for (j = i + 1; drs.iterate (j, &dr2); j++)
1631 if (dr_may_alias_p (dr1, dr2, true))
1632 edge_num++;
1633
1634 fprintf (file, "$\n");
1635
1636 if (comment)
1637 fprintf (file, "c %s\n", comment);
1638
1639 fprintf (file, "p edge %d %d\n", num_vertex, edge_num);
1640
1641 FOR_EACH_VEC_ELT (drs, i, dr1)
1642 for (j = i + 1; drs.iterate (j, &dr2); j++)
1643 if (dr_may_alias_p (dr1, dr2, true))
1644 fprintf (file, "e %d %d\n", i + 1, j + 1);
1645
1646 return true;
1647 }
1648
1649 /* Write to FILE the alias graph of data references in DOT format. */
1650
1651 static inline bool
1652 write_alias_graph_to_ascii_dot (FILE *file, char *comment,
1653 vec<data_reference_p> drs)
1654 {
1655 int num_vertex = drs.length ();
1656 data_reference_p dr1, dr2;
1657 int i, j;
1658
1659 if (num_vertex == 0)
1660 return true;
1661
1662 fprintf (file, "$\n");
1663
1664 if (comment)
1665 fprintf (file, "c %s\n", comment);
1666
1667 /* First print all the vertices. */
1668 FOR_EACH_VEC_ELT (drs, i, dr1)
1669 fprintf (file, "n%d;\n", i);
1670
1671 FOR_EACH_VEC_ELT (drs, i, dr1)
1672 for (j = i + 1; drs.iterate (j, &dr2); j++)
1673 if (dr_may_alias_p (dr1, dr2, true))
1674 fprintf (file, "n%d n%d\n", i, j);
1675
1676 return true;
1677 }
1678
1679 /* Write to FILE the alias graph of data references in ECC format. */
1680
1681 static inline bool
1682 write_alias_graph_to_ascii_ecc (FILE *file, char *comment,
1683 vec<data_reference_p> drs)
1684 {
1685 int num_vertex = drs.length ();
1686 data_reference_p dr1, dr2;
1687 int i, j;
1688
1689 if (num_vertex == 0)
1690 return true;
1691
1692 fprintf (file, "$\n");
1693
1694 if (comment)
1695 fprintf (file, "c %s\n", comment);
1696
1697 FOR_EACH_VEC_ELT (drs, i, dr1)
1698 for (j = i + 1; drs.iterate (j, &dr2); j++)
1699 if (dr_may_alias_p (dr1, dr2, true))
1700 fprintf (file, "%d %d\n", i, j);
1701
1702 return true;
1703 }
1704
1705 /* Check if DR1 and DR2 are in the same object set. */
1706
1707 static bool
1708 dr_same_base_object_p (const struct data_reference *dr1,
1709 const struct data_reference *dr2)
1710 {
1711 return operand_equal_p (DR_BASE_OBJECT (dr1), DR_BASE_OBJECT (dr2), 0);
1712 }
1713
1714 /* Uses DFS component number as representative of alias-sets. Also tests for
1715 optimality by verifying if every connected component is a clique. Returns
1716 true (1) if the above test is true, and false (0) otherwise. */
1717
1718 static int
1719 build_alias_set_optimal_p (vec<data_reference_p> drs)
1720 {
1721 int num_vertices = drs.length ();
1722 struct graph *g = new_graph (num_vertices);
1723 data_reference_p dr1, dr2;
1724 int i, j;
1725 int num_connected_components;
1726 int v_indx1, v_indx2, num_vertices_in_component;
1727 int *all_vertices;
1728 int *vertices;
1729 struct graph_edge *e;
1730 int this_component_is_clique;
1731 int all_components_are_cliques = 1;
1732
1733 FOR_EACH_VEC_ELT (drs, i, dr1)
1734 for (j = i+1; drs.iterate (j, &dr2); j++)
1735 if (dr_may_alias_p (dr1, dr2, true))
1736 {
1737 add_edge (g, i, j);
1738 add_edge (g, j, i);
1739 }
1740
1741 all_vertices = XNEWVEC (int, num_vertices);
1742 vertices = XNEWVEC (int, num_vertices);
1743 for (i = 0; i < num_vertices; i++)
1744 all_vertices[i] = i;
1745
1746 num_connected_components = graphds_dfs (g, all_vertices, num_vertices,
1747 NULL, true, NULL);
1748 for (i = 0; i < g->n_vertices; i++)
1749 {
1750 data_reference_p dr = drs[i];
1751 base_alias_pair *bap;
1752
1753 gcc_assert (dr->aux);
1754 bap = (base_alias_pair *)(dr->aux);
1755
1756 bap->alias_set = XNEW (int);
1757 *(bap->alias_set) = g->vertices[i].component + 1;
1758 }
1759
1760 /* Verify if the DFS numbering results in optimal solution. */
1761 for (i = 0; i < num_connected_components; i++)
1762 {
1763 num_vertices_in_component = 0;
1764 /* Get all vertices whose DFS component number is the same as i. */
1765 for (j = 0; j < num_vertices; j++)
1766 if (g->vertices[j].component == i)
1767 vertices[num_vertices_in_component++] = j;
1768
1769 /* Now test if the vertices in 'vertices' form a clique, by testing
1770 for edges among each pair. */
1771 this_component_is_clique = 1;
1772 for (v_indx1 = 0; v_indx1 < num_vertices_in_component; v_indx1++)
1773 {
1774 for (v_indx2 = v_indx1+1; v_indx2 < num_vertices_in_component; v_indx2++)
1775 {
1776 /* Check if the two vertices are connected by iterating
1777 through all the edges which have one of these are source. */
1778 e = g->vertices[vertices[v_indx2]].pred;
1779 while (e)
1780 {
1781 if (e->src == vertices[v_indx1])
1782 break;
1783 e = e->pred_next;
1784 }
1785 if (!e)
1786 {
1787 this_component_is_clique = 0;
1788 break;
1789 }
1790 }
1791 if (!this_component_is_clique)
1792 all_components_are_cliques = 0;
1793 }
1794 }
1795
1796 free (all_vertices);
1797 free (vertices);
1798 free_graph (g);
1799 return all_components_are_cliques;
1800 }
1801
1802 /* Group each data reference in DRS with its base object set num. */
1803
1804 static void
1805 build_base_obj_set_for_drs (vec<data_reference_p> drs)
1806 {
1807 int num_vertex = drs.length ();
1808 struct graph *g = new_graph (num_vertex);
1809 data_reference_p dr1, dr2;
1810 int i, j;
1811 int *queue;
1812
1813 FOR_EACH_VEC_ELT (drs, i, dr1)
1814 for (j = i + 1; drs.iterate (j, &dr2); j++)
1815 if (dr_same_base_object_p (dr1, dr2))
1816 {
1817 add_edge (g, i, j);
1818 add_edge (g, j, i);
1819 }
1820
1821 queue = XNEWVEC (int, num_vertex);
1822 for (i = 0; i < num_vertex; i++)
1823 queue[i] = i;
1824
1825 graphds_dfs (g, queue, num_vertex, NULL, true, NULL);
1826
1827 for (i = 0; i < g->n_vertices; i++)
1828 {
1829 data_reference_p dr = drs[i];
1830 base_alias_pair *bap;
1831
1832 gcc_assert (dr->aux);
1833 bap = (base_alias_pair *)(dr->aux);
1834
1835 bap->base_obj_set = g->vertices[i].component + 1;
1836 }
1837
1838 free (queue);
1839 free_graph (g);
1840 }
1841
1842 /* Build the data references for PBB. */
1843
1844 static void
1845 build_pbb_drs (poly_bb_p pbb)
1846 {
1847 int j;
1848 data_reference_p dr;
1849 vec<data_reference_p> gbb_drs = GBB_DATA_REFS (PBB_BLACK_BOX (pbb));
1850
1851 FOR_EACH_VEC_ELT (gbb_drs, j, dr)
1852 build_poly_dr (dr, pbb);
1853 }
1854
1855 /* Dump to file the alias graphs for the data references in DRS. */
1856
1857 static void
1858 dump_alias_graphs (vec<data_reference_p> drs)
1859 {
1860 char comment[100];
1861 FILE *file_dimacs, *file_ecc, *file_dot;
1862
1863 file_dimacs = fopen ("/tmp/dr_alias_graph_dimacs", "ab");
1864 if (file_dimacs)
1865 {
1866 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1867 current_function_name ());
1868 write_alias_graph_to_ascii_dimacs (file_dimacs, comment, drs);
1869 fclose (file_dimacs);
1870 }
1871
1872 file_ecc = fopen ("/tmp/dr_alias_graph_ecc", "ab");
1873 if (file_ecc)
1874 {
1875 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1876 current_function_name ());
1877 write_alias_graph_to_ascii_ecc (file_ecc, comment, drs);
1878 fclose (file_ecc);
1879 }
1880
1881 file_dot = fopen ("/tmp/dr_alias_graph_dot", "ab");
1882 if (file_dot)
1883 {
1884 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1885 current_function_name ());
1886 write_alias_graph_to_ascii_dot (file_dot, comment, drs);
1887 fclose (file_dot);
1888 }
1889 }
1890
1891 /* Build data references in SCOP. */
1892
1893 static void
1894 build_scop_drs (scop_p scop)
1895 {
1896 int i, j;
1897 poly_bb_p pbb;
1898 data_reference_p dr;
1899 auto_vec<data_reference_p, 3> drs;
1900
1901 /* Remove all the PBBs that do not have data references: these basic
1902 blocks are not handled in the polyhedral representation. */
1903 for (i = 0; SCOP_BBS (scop).iterate (i, &pbb); i++)
1904 if (GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).is_empty ())
1905 {
1906 free_gimple_bb (PBB_BLACK_BOX (pbb));
1907 free_poly_bb (pbb);
1908 SCOP_BBS (scop).ordered_remove (i);
1909 i--;
1910 }
1911
1912 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1913 for (j = 0; GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).iterate (j, &dr); j++)
1914 drs.safe_push (dr);
1915
1916 FOR_EACH_VEC_ELT (drs, i, dr)
1917 dr->aux = XNEW (base_alias_pair);
1918
1919 if (!build_alias_set_optimal_p (drs))
1920 {
1921 /* TODO: Add support when building alias set is not optimal. */
1922 ;
1923 }
1924
1925 build_base_obj_set_for_drs (drs);
1926
1927 /* When debugging, enable the following code. This cannot be used
1928 in production compilers. */
1929 if (0)
1930 dump_alias_graphs (drs);
1931
1932 drs.release ();
1933
1934 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1935 build_pbb_drs (pbb);
1936 }
1937
1938 /* Return a gsi at the position of the phi node STMT. */
1939
1940 static gphi_iterator
1941 gsi_for_phi_node (gphi *stmt)
1942 {
1943 gphi_iterator psi;
1944 basic_block bb = gimple_bb (stmt);
1945
1946 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
1947 if (stmt == psi.phi ())
1948 return psi;
1949
1950 gcc_unreachable ();
1951 return psi;
1952 }
1953
1954 /* Analyze all the data references of STMTS and add them to the
1955 GBB_DATA_REFS vector of BB. */
1956
1957 static void
1958 analyze_drs_in_stmts (scop_p scop, basic_block bb, vec<gimple> stmts)
1959 {
1960 loop_p nest;
1961 gimple_bb_p gbb;
1962 gimple stmt;
1963 int i;
1964 sese region = SCOP_REGION (scop);
1965
1966 if (!bb_in_sese_p (bb, region))
1967 return;
1968
1969 nest = outermost_loop_in_sese_1 (region, bb);
1970 gbb = gbb_from_bb (bb);
1971
1972 FOR_EACH_VEC_ELT (stmts, i, stmt)
1973 {
1974 loop_p loop;
1975
1976 if (is_gimple_debug (stmt))
1977 continue;
1978
1979 loop = loop_containing_stmt (stmt);
1980 if (!loop_in_sese_p (loop, region))
1981 loop = nest;
1982
1983 graphite_find_data_references_in_stmt (nest, loop, stmt,
1984 &GBB_DATA_REFS (gbb));
1985 }
1986 }
1987
1988 /* Insert STMT at the end of the STMTS sequence and then insert the
1989 statements from STMTS at INSERT_GSI and call analyze_drs_in_stmts
1990 on STMTS. */
1991
1992 static void
1993 insert_stmts (scop_p scop, gimple stmt, gimple_seq stmts,
1994 gimple_stmt_iterator insert_gsi)
1995 {
1996 gimple_stmt_iterator gsi;
1997 auto_vec<gimple, 3> x;
1998
1999 gimple_seq_add_stmt (&stmts, stmt);
2000 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2001 x.safe_push (gsi_stmt (gsi));
2002
2003 gsi_insert_seq_before (&insert_gsi, stmts, GSI_SAME_STMT);
2004 analyze_drs_in_stmts (scop, gsi_bb (insert_gsi), x);
2005 }
2006
2007 /* Insert the assignment "RES := EXPR" just after AFTER_STMT. */
2008
2009 static void
2010 insert_out_of_ssa_copy (scop_p scop, tree res, tree expr, gimple after_stmt)
2011 {
2012 gimple_seq stmts;
2013 gimple_stmt_iterator gsi;
2014 tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE);
2015 gassign *stmt = gimple_build_assign (unshare_expr (res), var);
2016 auto_vec<gimple, 3> x;
2017
2018 gimple_seq_add_stmt (&stmts, stmt);
2019 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2020 x.safe_push (gsi_stmt (gsi));
2021
2022 if (gimple_code (after_stmt) == GIMPLE_PHI)
2023 {
2024 gsi = gsi_after_labels (gimple_bb (after_stmt));
2025 gsi_insert_seq_before (&gsi, stmts, GSI_NEW_STMT);
2026 }
2027 else
2028 {
2029 gsi = gsi_for_stmt (after_stmt);
2030 gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT);
2031 }
2032
2033 analyze_drs_in_stmts (scop, gimple_bb (after_stmt), x);
2034 }
2035
2036 /* Creates a poly_bb_p for basic_block BB from the existing PBB. */
2037
2038 static void
2039 new_pbb_from_pbb (scop_p scop, poly_bb_p pbb, basic_block bb)
2040 {
2041 vec<data_reference_p> drs;
2042 drs.create (3);
2043 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
2044 gimple_bb_p gbb1 = new_gimple_bb (bb, drs);
2045 poly_bb_p pbb1 = new_poly_bb (scop, gbb1);
2046 int index, n = SCOP_BBS (scop).length ();
2047
2048 /* The INDEX of PBB in SCOP_BBS. */
2049 for (index = 0; index < n; index++)
2050 if (SCOP_BBS (scop)[index] == pbb)
2051 break;
2052
2053 pbb1->domain = isl_set_copy (pbb->domain);
2054 pbb1->domain = isl_set_set_tuple_id (pbb1->domain,
2055 isl_id_for_pbb (scop, pbb1));
2056
2057 GBB_PBB (gbb1) = pbb1;
2058 GBB_CONDITIONS (gbb1) = GBB_CONDITIONS (gbb).copy ();
2059 GBB_CONDITION_CASES (gbb1) = GBB_CONDITION_CASES (gbb).copy ();
2060 SCOP_BBS (scop).safe_insert (index + 1, pbb1);
2061 }
2062
2063 /* Insert on edge E the assignment "RES := EXPR". */
2064
2065 static void
2066 insert_out_of_ssa_copy_on_edge (scop_p scop, edge e, tree res, tree expr)
2067 {
2068 gimple_stmt_iterator gsi;
2069 gimple_seq stmts = NULL;
2070 tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE);
2071 gimple stmt = gimple_build_assign (unshare_expr (res), var);
2072 basic_block bb;
2073 auto_vec<gimple, 3> x;
2074
2075 gimple_seq_add_stmt (&stmts, stmt);
2076 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2077 x.safe_push (gsi_stmt (gsi));
2078
2079 gsi_insert_seq_on_edge (e, stmts);
2080 gsi_commit_edge_inserts ();
2081 bb = gimple_bb (stmt);
2082
2083 if (!bb_in_sese_p (bb, SCOP_REGION (scop)))
2084 return;
2085
2086 if (!gbb_from_bb (bb))
2087 new_pbb_from_pbb (scop, pbb_from_bb (e->src), bb);
2088
2089 analyze_drs_in_stmts (scop, bb, x);
2090 }
2091
2092 /* Creates a zero dimension array of the same type as VAR. */
2093
2094 static tree
2095 create_zero_dim_array (tree var, const char *base_name)
2096 {
2097 tree index_type = build_index_type (integer_zero_node);
2098 tree elt_type = TREE_TYPE (var);
2099 tree array_type = build_array_type (elt_type, index_type);
2100 tree base = create_tmp_var (array_type, base_name);
2101
2102 return build4 (ARRAY_REF, elt_type, base, integer_zero_node, NULL_TREE,
2103 NULL_TREE);
2104 }
2105
2106 /* Returns true when PHI is a loop close phi node. */
2107
2108 static bool
2109 scalar_close_phi_node_p (gimple phi)
2110 {
2111 if (gimple_code (phi) != GIMPLE_PHI
2112 || virtual_operand_p (gimple_phi_result (phi)))
2113 return false;
2114
2115 /* Note that loop close phi nodes should have a single argument
2116 because we translated the representation into a canonical form
2117 before Graphite: see canonicalize_loop_closed_ssa_form. */
2118 return (gimple_phi_num_args (phi) == 1);
2119 }
2120
2121 /* For a definition DEF in REGION, propagates the expression EXPR in
2122 all the uses of DEF outside REGION. */
2123
2124 static void
2125 propagate_expr_outside_region (tree def, tree expr, sese region)
2126 {
2127 imm_use_iterator imm_iter;
2128 gimple use_stmt;
2129 gimple_seq stmts;
2130 bool replaced_once = false;
2131
2132 gcc_assert (TREE_CODE (def) == SSA_NAME);
2133
2134 expr = force_gimple_operand (unshare_expr (expr), &stmts, true,
2135 NULL_TREE);
2136
2137 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2138 if (!is_gimple_debug (use_stmt)
2139 && !bb_in_sese_p (gimple_bb (use_stmt), region))
2140 {
2141 ssa_op_iter iter;
2142 use_operand_p use_p;
2143
2144 FOR_EACH_PHI_OR_STMT_USE (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2145 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0)
2146 && (replaced_once = true))
2147 replace_exp (use_p, expr);
2148
2149 update_stmt (use_stmt);
2150 }
2151
2152 if (replaced_once)
2153 {
2154 gsi_insert_seq_on_edge (SESE_ENTRY (region), stmts);
2155 gsi_commit_edge_inserts ();
2156 }
2157 }
2158
2159 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2160 dimension array for it. */
2161
2162 static void
2163 rewrite_close_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi)
2164 {
2165 sese region = SCOP_REGION (scop);
2166 gimple phi = gsi_stmt (*psi);
2167 tree res = gimple_phi_result (phi);
2168 basic_block bb = gimple_bb (phi);
2169 gimple_stmt_iterator gsi = gsi_after_labels (bb);
2170 tree arg = gimple_phi_arg_def (phi, 0);
2171 gimple stmt;
2172
2173 /* Note that loop close phi nodes should have a single argument
2174 because we translated the representation into a canonical form
2175 before Graphite: see canonicalize_loop_closed_ssa_form. */
2176 gcc_assert (gimple_phi_num_args (phi) == 1);
2177
2178 /* The phi node can be a non close phi node, when its argument is
2179 invariant, or a default definition. */
2180 if (is_gimple_min_invariant (arg)
2181 || SSA_NAME_IS_DEFAULT_DEF (arg))
2182 {
2183 propagate_expr_outside_region (res, arg, region);
2184 gsi_next (psi);
2185 return;
2186 }
2187
2188 else if (gimple_bb (SSA_NAME_DEF_STMT (arg))->loop_father == bb->loop_father)
2189 {
2190 propagate_expr_outside_region (res, arg, region);
2191 stmt = gimple_build_assign (res, arg);
2192 remove_phi_node (psi, false);
2193 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2194 return;
2195 }
2196
2197 /* If res is scev analyzable and is not a scalar value, it is safe
2198 to ignore the close phi node: it will be code generated in the
2199 out of Graphite pass. */
2200 else if (scev_analyzable_p (res, region))
2201 {
2202 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (res));
2203 tree scev;
2204
2205 if (!loop_in_sese_p (loop, region))
2206 {
2207 loop = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2208 scev = scalar_evolution_in_region (region, loop, arg);
2209 scev = compute_overall_effect_of_inner_loop (loop, scev);
2210 }
2211 else
2212 scev = scalar_evolution_in_region (region, loop, res);
2213
2214 if (tree_does_not_contain_chrecs (scev))
2215 propagate_expr_outside_region (res, scev, region);
2216
2217 gsi_next (psi);
2218 return;
2219 }
2220 else
2221 {
2222 tree zero_dim_array = create_zero_dim_array (res, "Close_Phi");
2223
2224 stmt = gimple_build_assign (res, unshare_expr (zero_dim_array));
2225
2226 if (TREE_CODE (arg) == SSA_NAME)
2227 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2228 SSA_NAME_DEF_STMT (arg));
2229 else
2230 insert_out_of_ssa_copy_on_edge (scop, single_pred_edge (bb),
2231 zero_dim_array, arg);
2232 }
2233
2234 remove_phi_node (psi, false);
2235 SSA_NAME_DEF_STMT (res) = stmt;
2236
2237 insert_stmts (scop, stmt, NULL, gsi_after_labels (bb));
2238 }
2239
2240 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2241 dimension array for it. */
2242
2243 static void
2244 rewrite_phi_out_of_ssa (scop_p scop, gphi_iterator *psi)
2245 {
2246 size_t i;
2247 gphi *phi = psi->phi ();
2248 basic_block bb = gimple_bb (phi);
2249 tree res = gimple_phi_result (phi);
2250 tree zero_dim_array = create_zero_dim_array (res, "phi_out_of_ssa");
2251 gimple stmt;
2252
2253 for (i = 0; i < gimple_phi_num_args (phi); i++)
2254 {
2255 tree arg = gimple_phi_arg_def (phi, i);
2256 edge e = gimple_phi_arg_edge (phi, i);
2257
2258 /* Avoid the insertion of code in the loop latch to please the
2259 pattern matching of the vectorizer. */
2260 if (TREE_CODE (arg) == SSA_NAME
2261 && !SSA_NAME_IS_DEFAULT_DEF (arg)
2262 && e->src == bb->loop_father->latch)
2263 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2264 SSA_NAME_DEF_STMT (arg));
2265 else
2266 insert_out_of_ssa_copy_on_edge (scop, e, zero_dim_array, arg);
2267 }
2268
2269 stmt = gimple_build_assign (res, unshare_expr (zero_dim_array));
2270 remove_phi_node (psi, false);
2271 insert_stmts (scop, stmt, NULL, gsi_after_labels (bb));
2272 }
2273
2274 /* Rewrite the degenerate phi node at position PSI from the degenerate
2275 form "x = phi (y, y, ..., y)" to "x = y". */
2276
2277 static void
2278 rewrite_degenerate_phi (gphi_iterator *psi)
2279 {
2280 tree rhs;
2281 gimple stmt;
2282 gimple_stmt_iterator gsi;
2283 gphi *phi = psi->phi ();
2284 tree res = gimple_phi_result (phi);
2285 basic_block bb;
2286
2287 bb = gimple_bb (phi);
2288 rhs = degenerate_phi_result (phi);
2289 gcc_assert (rhs);
2290
2291 stmt = gimple_build_assign (res, rhs);
2292 remove_phi_node (psi, false);
2293
2294 gsi = gsi_after_labels (bb);
2295 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2296 }
2297
2298 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2299
2300 static void
2301 rewrite_reductions_out_of_ssa (scop_p scop)
2302 {
2303 basic_block bb;
2304 gphi_iterator psi;
2305 sese region = SCOP_REGION (scop);
2306
2307 FOR_EACH_BB_FN (bb, cfun)
2308 if (bb_in_sese_p (bb, region))
2309 for (psi = gsi_start_phis (bb); !gsi_end_p (psi);)
2310 {
2311 gphi *phi = psi.phi ();
2312
2313 if (virtual_operand_p (gimple_phi_result (phi)))
2314 {
2315 gsi_next (&psi);
2316 continue;
2317 }
2318
2319 if (gimple_phi_num_args (phi) > 1
2320 && degenerate_phi_result (phi))
2321 rewrite_degenerate_phi (&psi);
2322
2323 else if (scalar_close_phi_node_p (phi))
2324 rewrite_close_phi_out_of_ssa (scop, &psi);
2325
2326 else if (reduction_phi_p (region, &psi))
2327 rewrite_phi_out_of_ssa (scop, &psi);
2328 }
2329
2330 update_ssa (TODO_update_ssa);
2331 #ifdef ENABLE_CHECKING
2332 verify_loop_closed_ssa (true);
2333 #endif
2334 }
2335
2336 /* Rewrite the scalar dependence of DEF used in USE_STMT with a memory
2337 read from ZERO_DIM_ARRAY. */
2338
2339 static void
2340 rewrite_cross_bb_scalar_dependence (scop_p scop, tree zero_dim_array,
2341 tree def, gimple use_stmt)
2342 {
2343 gimple name_stmt;
2344 tree name;
2345 ssa_op_iter iter;
2346 use_operand_p use_p;
2347
2348 gcc_assert (gimple_code (use_stmt) != GIMPLE_PHI);
2349
2350 name = copy_ssa_name (def, NULL);
2351 name_stmt = gimple_build_assign (name, zero_dim_array);
2352
2353 gimple_assign_set_lhs (name_stmt, name);
2354 insert_stmts (scop, name_stmt, NULL, gsi_for_stmt (use_stmt));
2355
2356 FOR_EACH_SSA_USE_OPERAND (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2357 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0))
2358 replace_exp (use_p, name);
2359
2360 update_stmt (use_stmt);
2361 }
2362
2363 /* For every definition DEF in the SCOP that is used outside the scop,
2364 insert a closing-scop definition in the basic block just after this
2365 SCOP. */
2366
2367 static void
2368 handle_scalar_deps_crossing_scop_limits (scop_p scop, tree def, gimple stmt)
2369 {
2370 tree var = create_tmp_reg (TREE_TYPE (def), NULL);
2371 tree new_name = make_ssa_name (var, stmt);
2372 bool needs_copy = false;
2373 use_operand_p use_p;
2374 imm_use_iterator imm_iter;
2375 gimple use_stmt;
2376 sese region = SCOP_REGION (scop);
2377
2378 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2379 {
2380 if (!bb_in_sese_p (gimple_bb (use_stmt), region))
2381 {
2382 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
2383 {
2384 SET_USE (use_p, new_name);
2385 }
2386 update_stmt (use_stmt);
2387 needs_copy = true;
2388 }
2389 }
2390
2391 /* Insert in the empty BB just after the scop a use of DEF such
2392 that the rewrite of cross_bb_scalar_dependences won't insert
2393 arrays everywhere else. */
2394 if (needs_copy)
2395 {
2396 gimple assign = gimple_build_assign (new_name, def);
2397 gimple_stmt_iterator psi = gsi_after_labels (SESE_EXIT (region)->dest);
2398
2399 update_stmt (assign);
2400 gsi_insert_before (&psi, assign, GSI_SAME_STMT);
2401 }
2402 }
2403
2404 /* Rewrite the scalar dependences crossing the boundary of the BB
2405 containing STMT with an array. Return true when something has been
2406 changed. */
2407
2408 static bool
2409 rewrite_cross_bb_scalar_deps (scop_p scop, gimple_stmt_iterator *gsi)
2410 {
2411 sese region = SCOP_REGION (scop);
2412 gimple stmt = gsi_stmt (*gsi);
2413 imm_use_iterator imm_iter;
2414 tree def;
2415 basic_block def_bb;
2416 tree zero_dim_array = NULL_TREE;
2417 gimple use_stmt;
2418 bool res = false;
2419
2420 switch (gimple_code (stmt))
2421 {
2422 case GIMPLE_ASSIGN:
2423 def = gimple_assign_lhs (stmt);
2424 break;
2425
2426 case GIMPLE_CALL:
2427 def = gimple_call_lhs (stmt);
2428 break;
2429
2430 default:
2431 return false;
2432 }
2433
2434 if (!def
2435 || !is_gimple_reg (def))
2436 return false;
2437
2438 if (scev_analyzable_p (def, region))
2439 {
2440 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (def));
2441 tree scev = scalar_evolution_in_region (region, loop, def);
2442
2443 if (tree_contains_chrecs (scev, NULL))
2444 return false;
2445
2446 propagate_expr_outside_region (def, scev, region);
2447 return true;
2448 }
2449
2450 def_bb = gimple_bb (stmt);
2451
2452 handle_scalar_deps_crossing_scop_limits (scop, def, stmt);
2453
2454 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2455 if (gimple_code (use_stmt) == GIMPLE_PHI
2456 && (res = true))
2457 {
2458 gphi_iterator psi = gsi_start_phis (gimple_bb (use_stmt));
2459
2460 if (scalar_close_phi_node_p (gsi_stmt (psi)))
2461 rewrite_close_phi_out_of_ssa (scop, &psi);
2462 else
2463 rewrite_phi_out_of_ssa (scop, &psi);
2464 }
2465
2466 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2467 if (gimple_code (use_stmt) != GIMPLE_PHI
2468 && def_bb != gimple_bb (use_stmt)
2469 && !is_gimple_debug (use_stmt)
2470 && (res = true))
2471 {
2472 if (!zero_dim_array)
2473 {
2474 zero_dim_array = create_zero_dim_array
2475 (def, "Cross_BB_scalar_dependence");
2476 insert_out_of_ssa_copy (scop, zero_dim_array, def,
2477 SSA_NAME_DEF_STMT (def));
2478 gsi_next (gsi);
2479 }
2480
2481 rewrite_cross_bb_scalar_dependence (scop, unshare_expr (zero_dim_array),
2482 def, use_stmt);
2483 }
2484
2485 return res;
2486 }
2487
2488 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2489
2490 static void
2491 rewrite_cross_bb_scalar_deps_out_of_ssa (scop_p scop)
2492 {
2493 basic_block bb;
2494 gimple_stmt_iterator psi;
2495 sese region = SCOP_REGION (scop);
2496 bool changed = false;
2497
2498 /* Create an extra empty BB after the scop. */
2499 split_edge (SESE_EXIT (region));
2500
2501 FOR_EACH_BB_FN (bb, cfun)
2502 if (bb_in_sese_p (bb, region))
2503 for (psi = gsi_start_bb (bb); !gsi_end_p (psi); gsi_next (&psi))
2504 changed |= rewrite_cross_bb_scalar_deps (scop, &psi);
2505
2506 if (changed)
2507 {
2508 scev_reset_htab ();
2509 update_ssa (TODO_update_ssa);
2510 #ifdef ENABLE_CHECKING
2511 verify_loop_closed_ssa (true);
2512 #endif
2513 }
2514 }
2515
2516 /* Returns the number of pbbs that are in loops contained in SCOP. */
2517
2518 static int
2519 nb_pbbs_in_loops (scop_p scop)
2520 {
2521 int i;
2522 poly_bb_p pbb;
2523 int res = 0;
2524
2525 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
2526 if (loop_in_sese_p (gbb_loop (PBB_BLACK_BOX (pbb)), SCOP_REGION (scop)))
2527 res++;
2528
2529 return res;
2530 }
2531
2532 /* Return the number of data references in BB that write in
2533 memory. */
2534
2535 static int
2536 nb_data_writes_in_bb (basic_block bb)
2537 {
2538 int res = 0;
2539 gimple_stmt_iterator gsi;
2540
2541 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
2542 if (gimple_vdef (gsi_stmt (gsi)))
2543 res++;
2544
2545 return res;
2546 }
2547
2548 /* Splits at STMT the basic block BB represented as PBB in the
2549 polyhedral form. */
2550
2551 static edge
2552 split_pbb (scop_p scop, poly_bb_p pbb, basic_block bb, gimple stmt)
2553 {
2554 edge e1 = split_block (bb, stmt);
2555 new_pbb_from_pbb (scop, pbb, e1->dest);
2556 return e1;
2557 }
2558
2559 /* Splits STMT out of its current BB. This is done for reduction
2560 statements for which we want to ignore data dependences. */
2561
2562 static basic_block
2563 split_reduction_stmt (scop_p scop, gimple stmt)
2564 {
2565 basic_block bb = gimple_bb (stmt);
2566 poly_bb_p pbb = pbb_from_bb (bb);
2567 gimple_bb_p gbb = gbb_from_bb (bb);
2568 edge e1;
2569 int i;
2570 data_reference_p dr;
2571
2572 /* Do not split basic blocks with no writes to memory: the reduction
2573 will be the only write to memory. */
2574 if (nb_data_writes_in_bb (bb) == 0
2575 /* Or if we have already marked BB as a reduction. */
2576 || PBB_IS_REDUCTION (pbb_from_bb (bb)))
2577 return bb;
2578
2579 e1 = split_pbb (scop, pbb, bb, stmt);
2580
2581 /* Split once more only when the reduction stmt is not the only one
2582 left in the original BB. */
2583 if (!gsi_one_before_end_p (gsi_start_nondebug_bb (bb)))
2584 {
2585 gimple_stmt_iterator gsi = gsi_last_bb (bb);
2586 gsi_prev (&gsi);
2587 e1 = split_pbb (scop, pbb, bb, gsi_stmt (gsi));
2588 }
2589
2590 /* A part of the data references will end in a different basic block
2591 after the split: move the DRs from the original GBB to the newly
2592 created GBB1. */
2593 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr)
2594 {
2595 basic_block bb1 = gimple_bb (DR_STMT (dr));
2596
2597 if (bb1 != bb)
2598 {
2599 gimple_bb_p gbb1 = gbb_from_bb (bb1);
2600 GBB_DATA_REFS (gbb1).safe_push (dr);
2601 GBB_DATA_REFS (gbb).ordered_remove (i);
2602 i--;
2603 }
2604 }
2605
2606 return e1->dest;
2607 }
2608
2609 /* Return true when stmt is a reduction operation. */
2610
2611 static inline bool
2612 is_reduction_operation_p (gimple stmt)
2613 {
2614 enum tree_code code;
2615
2616 gcc_assert (is_gimple_assign (stmt));
2617 code = gimple_assign_rhs_code (stmt);
2618
2619 return flag_associative_math
2620 && commutative_tree_code (code)
2621 && associative_tree_code (code);
2622 }
2623
2624 /* Returns true when PHI contains an argument ARG. */
2625
2626 static bool
2627 phi_contains_arg (gphi *phi, tree arg)
2628 {
2629 size_t i;
2630
2631 for (i = 0; i < gimple_phi_num_args (phi); i++)
2632 if (operand_equal_p (arg, gimple_phi_arg_def (phi, i), 0))
2633 return true;
2634
2635 return false;
2636 }
2637
2638 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2639
2640 static gphi *
2641 follow_ssa_with_commutative_ops (tree arg, tree lhs)
2642 {
2643 gimple stmt;
2644
2645 if (TREE_CODE (arg) != SSA_NAME)
2646 return NULL;
2647
2648 stmt = SSA_NAME_DEF_STMT (arg);
2649
2650 if (gimple_code (stmt) == GIMPLE_NOP
2651 || gimple_code (stmt) == GIMPLE_CALL)
2652 return NULL;
2653
2654 if (gphi *phi = dyn_cast <gphi *> (stmt))
2655 {
2656 if (phi_contains_arg (phi, lhs))
2657 return phi;
2658 return NULL;
2659 }
2660
2661 if (!is_gimple_assign (stmt))
2662 return NULL;
2663
2664 if (gimple_num_ops (stmt) == 2)
2665 return follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2666
2667 if (is_reduction_operation_p (stmt))
2668 {
2669 gphi *res
2670 = follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2671
2672 return res ? res :
2673 follow_ssa_with_commutative_ops (gimple_assign_rhs2 (stmt), lhs);
2674 }
2675
2676 return NULL;
2677 }
2678
2679 /* Detect commutative and associative scalar reductions starting at
2680 the STMT. Return the phi node of the reduction cycle, or NULL. */
2681
2682 static gphi *
2683 detect_commutative_reduction_arg (tree lhs, gimple stmt, tree arg,
2684 vec<gimple> *in,
2685 vec<gimple> *out)
2686 {
2687 gphi *phi = follow_ssa_with_commutative_ops (arg, lhs);
2688
2689 if (!phi)
2690 return NULL;
2691
2692 in->safe_push (stmt);
2693 out->safe_push (stmt);
2694 return phi;
2695 }
2696
2697 /* Detect commutative and associative scalar reductions starting at
2698 STMT. Return the phi node of the reduction cycle, or NULL. */
2699
2700 static gphi *
2701 detect_commutative_reduction_assign (gimple stmt, vec<gimple> *in,
2702 vec<gimple> *out)
2703 {
2704 tree lhs = gimple_assign_lhs (stmt);
2705
2706 if (gimple_num_ops (stmt) == 2)
2707 return detect_commutative_reduction_arg (lhs, stmt,
2708 gimple_assign_rhs1 (stmt),
2709 in, out);
2710
2711 if (is_reduction_operation_p (stmt))
2712 {
2713 gphi *res = detect_commutative_reduction_arg (lhs, stmt,
2714 gimple_assign_rhs1 (stmt),
2715 in, out);
2716 return res ? res
2717 : detect_commutative_reduction_arg (lhs, stmt,
2718 gimple_assign_rhs2 (stmt),
2719 in, out);
2720 }
2721
2722 return NULL;
2723 }
2724
2725 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2726
2727 static gphi *
2728 follow_inital_value_to_phi (tree arg, tree lhs)
2729 {
2730 gimple stmt;
2731
2732 if (!arg || TREE_CODE (arg) != SSA_NAME)
2733 return NULL;
2734
2735 stmt = SSA_NAME_DEF_STMT (arg);
2736
2737 if (gphi *phi = dyn_cast <gphi *> (stmt))
2738 if (phi_contains_arg (phi, lhs))
2739 return phi;
2740
2741 return NULL;
2742 }
2743
2744
2745 /* Return the argument of the loop PHI that is the initial value coming
2746 from outside the loop. */
2747
2748 static edge
2749 edge_initial_value_for_loop_phi (gphi *phi)
2750 {
2751 size_t i;
2752
2753 for (i = 0; i < gimple_phi_num_args (phi); i++)
2754 {
2755 edge e = gimple_phi_arg_edge (phi, i);
2756
2757 if (loop_depth (e->src->loop_father)
2758 < loop_depth (e->dest->loop_father))
2759 return e;
2760 }
2761
2762 return NULL;
2763 }
2764
2765 /* Return the argument of the loop PHI that is the initial value coming
2766 from outside the loop. */
2767
2768 static tree
2769 initial_value_for_loop_phi (gphi *phi)
2770 {
2771 size_t i;
2772
2773 for (i = 0; i < gimple_phi_num_args (phi); i++)
2774 {
2775 edge e = gimple_phi_arg_edge (phi, i);
2776
2777 if (loop_depth (e->src->loop_father)
2778 < loop_depth (e->dest->loop_father))
2779 return gimple_phi_arg_def (phi, i);
2780 }
2781
2782 return NULL_TREE;
2783 }
2784
2785 /* Returns true when DEF is used outside the reduction cycle of
2786 LOOP_PHI. */
2787
2788 static bool
2789 used_outside_reduction (tree def, gimple loop_phi)
2790 {
2791 use_operand_p use_p;
2792 imm_use_iterator imm_iter;
2793 loop_p loop = loop_containing_stmt (loop_phi);
2794
2795 /* In LOOP, DEF should be used only in LOOP_PHI. */
2796 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2797 {
2798 gimple stmt = USE_STMT (use_p);
2799
2800 if (stmt != loop_phi
2801 && !is_gimple_debug (stmt)
2802 && flow_bb_inside_loop_p (loop, gimple_bb (stmt)))
2803 return true;
2804 }
2805
2806 return false;
2807 }
2808
2809 /* Detect commutative and associative scalar reductions belonging to
2810 the SCOP starting at the loop closed phi node STMT. Return the phi
2811 node of the reduction cycle, or NULL. */
2812
2813 static gphi *
2814 detect_commutative_reduction (scop_p scop, gimple stmt, vec<gimple> *in,
2815 vec<gimple> *out)
2816 {
2817 if (scalar_close_phi_node_p (stmt))
2818 {
2819 gimple def;
2820 gphi *loop_phi, *phi, *close_phi = as_a <gphi *> (stmt);
2821 tree init, lhs, arg = gimple_phi_arg_def (close_phi, 0);
2822
2823 if (TREE_CODE (arg) != SSA_NAME)
2824 return NULL;
2825
2826 /* Note that loop close phi nodes should have a single argument
2827 because we translated the representation into a canonical form
2828 before Graphite: see canonicalize_loop_closed_ssa_form. */
2829 gcc_assert (gimple_phi_num_args (close_phi) == 1);
2830
2831 def = SSA_NAME_DEF_STMT (arg);
2832 if (!stmt_in_sese_p (def, SCOP_REGION (scop))
2833 || !(loop_phi = detect_commutative_reduction (scop, def, in, out)))
2834 return NULL;
2835
2836 lhs = gimple_phi_result (close_phi);
2837 init = initial_value_for_loop_phi (loop_phi);
2838 phi = follow_inital_value_to_phi (init, lhs);
2839
2840 if (phi && (used_outside_reduction (lhs, phi)
2841 || !has_single_use (gimple_phi_result (phi))))
2842 return NULL;
2843
2844 in->safe_push (loop_phi);
2845 out->safe_push (close_phi);
2846 return phi;
2847 }
2848
2849 if (gimple_code (stmt) == GIMPLE_ASSIGN)
2850 return detect_commutative_reduction_assign (stmt, in, out);
2851
2852 return NULL;
2853 }
2854
2855 /* Translate the scalar reduction statement STMT to an array RED
2856 knowing that its recursive phi node is LOOP_PHI. */
2857
2858 static void
2859 translate_scalar_reduction_to_array_for_stmt (scop_p scop, tree red,
2860 gimple stmt, gphi *loop_phi)
2861 {
2862 tree res = gimple_phi_result (loop_phi);
2863 gassign *assign = gimple_build_assign (res, unshare_expr (red));
2864 gimple_stmt_iterator gsi;
2865
2866 insert_stmts (scop, assign, NULL, gsi_after_labels (gimple_bb (loop_phi)));
2867
2868 assign = gimple_build_assign (unshare_expr (red), gimple_assign_lhs (stmt));
2869 gsi = gsi_for_stmt (stmt);
2870 gsi_next (&gsi);
2871 insert_stmts (scop, assign, NULL, gsi);
2872 }
2873
2874 /* Removes the PHI node and resets all the debug stmts that are using
2875 the PHI_RESULT. */
2876
2877 static void
2878 remove_phi (gphi *phi)
2879 {
2880 imm_use_iterator imm_iter;
2881 tree def;
2882 use_operand_p use_p;
2883 gimple_stmt_iterator gsi;
2884 auto_vec<gimple, 3> update;
2885 unsigned int i;
2886 gimple stmt;
2887
2888 def = PHI_RESULT (phi);
2889 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2890 {
2891 stmt = USE_STMT (use_p);
2892
2893 if (is_gimple_debug (stmt))
2894 {
2895 gimple_debug_bind_reset_value (stmt);
2896 update.safe_push (stmt);
2897 }
2898 }
2899
2900 FOR_EACH_VEC_ELT (update, i, stmt)
2901 update_stmt (stmt);
2902
2903 gsi = gsi_for_phi_node (phi);
2904 remove_phi_node (&gsi, false);
2905 }
2906
2907 /* Helper function for for_each_index. For each INDEX of the data
2908 reference REF, returns true when its indices are valid in the loop
2909 nest LOOP passed in as DATA. */
2910
2911 static bool
2912 dr_indices_valid_in_loop (tree ref ATTRIBUTE_UNUSED, tree *index, void *data)
2913 {
2914 loop_p loop;
2915 basic_block header, def_bb;
2916 gimple stmt;
2917
2918 if (TREE_CODE (*index) != SSA_NAME)
2919 return true;
2920
2921 loop = *((loop_p *) data);
2922 header = loop->header;
2923 stmt = SSA_NAME_DEF_STMT (*index);
2924
2925 if (!stmt)
2926 return true;
2927
2928 def_bb = gimple_bb (stmt);
2929
2930 if (!def_bb)
2931 return true;
2932
2933 return dominated_by_p (CDI_DOMINATORS, header, def_bb);
2934 }
2935
2936 /* When the result of a CLOSE_PHI is written to a memory location,
2937 return a pointer to that memory reference, otherwise return
2938 NULL_TREE. */
2939
2940 static tree
2941 close_phi_written_to_memory (gphi *close_phi)
2942 {
2943 imm_use_iterator imm_iter;
2944 use_operand_p use_p;
2945 gimple stmt;
2946 tree res, def = gimple_phi_result (close_phi);
2947
2948 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2949 if ((stmt = USE_STMT (use_p))
2950 && gimple_code (stmt) == GIMPLE_ASSIGN
2951 && (res = gimple_assign_lhs (stmt)))
2952 {
2953 switch (TREE_CODE (res))
2954 {
2955 case VAR_DECL:
2956 case PARM_DECL:
2957 case RESULT_DECL:
2958 return res;
2959
2960 case ARRAY_REF:
2961 case MEM_REF:
2962 {
2963 tree arg = gimple_phi_arg_def (close_phi, 0);
2964 loop_p nest = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2965
2966 /* FIXME: this restriction is for id-{24,25}.f and
2967 could be handled by duplicating the computation of
2968 array indices before the loop of the close_phi. */
2969 if (for_each_index (&res, dr_indices_valid_in_loop, &nest))
2970 return res;
2971 }
2972 /* Fallthru. */
2973
2974 default:
2975 continue;
2976 }
2977 }
2978 return NULL_TREE;
2979 }
2980
2981 /* Rewrite out of SSA the reduction described by the loop phi nodes
2982 IN, and the close phi nodes OUT. IN and OUT are structured by loop
2983 levels like this:
2984
2985 IN: stmt, loop_n, ..., loop_0
2986 OUT: stmt, close_n, ..., close_0
2987
2988 the first element is the reduction statement, and the next elements
2989 are the loop and close phi nodes of each of the outer loops. */
2990
2991 static void
2992 translate_scalar_reduction_to_array (scop_p scop,
2993 vec<gimple> in,
2994 vec<gimple> out)
2995 {
2996 gimple loop_stmt;
2997 unsigned int i = out.length () - 1;
2998 tree red = close_phi_written_to_memory (as_a <gphi *> (out[i]));
2999
3000 FOR_EACH_VEC_ELT (in, i, loop_stmt)
3001 {
3002 gimple close_stmt = out[i];
3003
3004 if (i == 0)
3005 {
3006 basic_block bb = split_reduction_stmt (scop, loop_stmt);
3007 poly_bb_p pbb = pbb_from_bb (bb);
3008 PBB_IS_REDUCTION (pbb) = true;
3009 gcc_assert (close_stmt == loop_stmt);
3010
3011 if (!red)
3012 red = create_zero_dim_array
3013 (gimple_assign_lhs (loop_stmt), "Commutative_Associative_Reduction");
3014
3015 translate_scalar_reduction_to_array_for_stmt (scop, red, loop_stmt,
3016 as_a <gphi *> (in[1]));
3017 continue;
3018 }
3019
3020 gphi *loop_phi = as_a <gphi *> (loop_stmt);
3021 gphi *close_phi = as_a <gphi *> (close_stmt);
3022
3023 if (i == in.length () - 1)
3024 {
3025 insert_out_of_ssa_copy (scop, gimple_phi_result (close_phi),
3026 unshare_expr (red), close_phi);
3027 insert_out_of_ssa_copy_on_edge
3028 (scop, edge_initial_value_for_loop_phi (loop_phi),
3029 unshare_expr (red), initial_value_for_loop_phi (loop_phi));
3030 }
3031
3032 remove_phi (loop_phi);
3033 remove_phi (close_phi);
3034 }
3035 }
3036
3037 /* Rewrites out of SSA a commutative reduction at CLOSE_PHI. Returns
3038 true when something has been changed. */
3039
3040 static bool
3041 rewrite_commutative_reductions_out_of_ssa_close_phi (scop_p scop,
3042 gphi *close_phi)
3043 {
3044 bool res;
3045 auto_vec<gimple, 10> in;
3046 auto_vec<gimple, 10> out;
3047
3048 detect_commutative_reduction (scop, close_phi, &in, &out);
3049 res = in.length () > 1;
3050 if (res)
3051 translate_scalar_reduction_to_array (scop, in, out);
3052
3053 return res;
3054 }
3055
3056 /* Rewrites all the commutative reductions from LOOP out of SSA.
3057 Returns true when something has been changed. */
3058
3059 static bool
3060 rewrite_commutative_reductions_out_of_ssa_loop (scop_p scop,
3061 loop_p loop)
3062 {
3063 gphi_iterator gsi;
3064 edge exit = single_exit (loop);
3065 tree res;
3066 bool changed = false;
3067
3068 if (!exit)
3069 return false;
3070
3071 for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi))
3072 if ((res = gimple_phi_result (gsi.phi ()))
3073 && !virtual_operand_p (res)
3074 && !scev_analyzable_p (res, SCOP_REGION (scop)))
3075 changed |= rewrite_commutative_reductions_out_of_ssa_close_phi
3076 (scop, gsi.phi ());
3077
3078 return changed;
3079 }
3080
3081 /* Rewrites all the commutative reductions from SCOP out of SSA. */
3082
3083 static void
3084 rewrite_commutative_reductions_out_of_ssa (scop_p scop)
3085 {
3086 loop_p loop;
3087 bool changed = false;
3088 sese region = SCOP_REGION (scop);
3089
3090 FOR_EACH_LOOP (loop, 0)
3091 if (loop_in_sese_p (loop, region))
3092 changed |= rewrite_commutative_reductions_out_of_ssa_loop (scop, loop);
3093
3094 if (changed)
3095 {
3096 scev_reset_htab ();
3097 gsi_commit_edge_inserts ();
3098 update_ssa (TODO_update_ssa);
3099 #ifdef ENABLE_CHECKING
3100 verify_loop_closed_ssa (true);
3101 #endif
3102 }
3103 }
3104
3105 /* Can all ivs be represented by a signed integer?
3106 As CLooG might generate negative values in its expressions, signed loop ivs
3107 are required in the backend. */
3108
3109 static bool
3110 scop_ivs_can_be_represented (scop_p scop)
3111 {
3112 loop_p loop;
3113 gphi_iterator psi;
3114 bool result = true;
3115
3116 FOR_EACH_LOOP (loop, 0)
3117 {
3118 if (!loop_in_sese_p (loop, SCOP_REGION (scop)))
3119 continue;
3120
3121 for (psi = gsi_start_phis (loop->header);
3122 !gsi_end_p (psi); gsi_next (&psi))
3123 {
3124 gphi *phi = psi.phi ();
3125 tree res = PHI_RESULT (phi);
3126 tree type = TREE_TYPE (res);
3127
3128 if (TYPE_UNSIGNED (type)
3129 && TYPE_PRECISION (type) >= TYPE_PRECISION (long_long_integer_type_node))
3130 {
3131 result = false;
3132 break;
3133 }
3134 }
3135 if (!result)
3136 break;
3137 }
3138
3139 return result;
3140 }
3141
3142 /* Builds the polyhedral representation for a SESE region. */
3143
3144 void
3145 build_poly_scop (scop_p scop)
3146 {
3147 sese region = SCOP_REGION (scop);
3148 graphite_dim_t max_dim;
3149
3150 build_scop_bbs (scop);
3151
3152 /* FIXME: This restriction is needed to avoid a problem in CLooG.
3153 Once CLooG is fixed, remove this guard. Anyways, it makes no
3154 sense to optimize a scop containing only PBBs that do not belong
3155 to any loops. */
3156 if (nb_pbbs_in_loops (scop) == 0)
3157 return;
3158
3159 if (!scop_ivs_can_be_represented (scop))
3160 return;
3161
3162 if (flag_associative_math)
3163 rewrite_commutative_reductions_out_of_ssa (scop);
3164
3165 build_sese_loop_nests (region);
3166 /* Record all conditions in REGION. */
3167 sese_dom_walker (CDI_DOMINATORS, region).walk (cfun->cfg->x_entry_block_ptr);
3168 find_scop_parameters (scop);
3169
3170 max_dim = PARAM_VALUE (PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS);
3171 if (scop_nb_params (scop) > max_dim)
3172 return;
3173
3174 build_scop_iteration_domain (scop);
3175 build_scop_context (scop);
3176 add_conditions_to_constraints (scop);
3177
3178 /* Rewrite out of SSA only after having translated the
3179 representation to the polyhedral representation to avoid scev
3180 analysis failures. That means that these functions will insert
3181 new data references that they create in the right place. */
3182 rewrite_reductions_out_of_ssa (scop);
3183 rewrite_cross_bb_scalar_deps_out_of_ssa (scop);
3184
3185 build_scop_drs (scop);
3186 scop_to_lst (scop);
3187 build_scop_scattering (scop);
3188
3189 /* This SCoP has been translated to the polyhedral
3190 representation. */
3191 POLY_SCOP_P (scop) = true;
3192 }
3193 #endif