c1686221187725928f22679d6ed4d86f0a38ba12
[gcc.git] / gcc / graphite-sese-to-poly.c
1 /* Conversion of SESE regions to Polyhedra.
2 Copyright (C) 2009, 2010, 2011 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-flow.h"
37 #include "tree-pass.h"
38 #include "cfgloop.h"
39 #include "tree-chrec.h"
40 #include "tree-data-ref.h"
41 #include "tree-scalar-evolution.h"
42 #include "domwalk.h"
43 #include "sese.h"
44
45 #ifdef HAVE_cloog
46 #include "graphite-poly.h"
47 #include "graphite-sese-to-poly.h"
48
49
50 /* Assigns to RES the value of the INTEGER_CST T. */
51
52 static inline void
53 tree_int_to_gmp (tree t, mpz_t res)
54 {
55 double_int di = tree_to_double_int (t);
56 mpz_set_double_int (res, di, TYPE_UNSIGNED (TREE_TYPE (t)));
57 }
58
59 /* Returns the index of the PHI argument defined in the outermost
60 loop. */
61
62 static size_t
63 phi_arg_in_outermost_loop (gimple phi)
64 {
65 loop_p loop = gimple_bb (phi)->loop_father;
66 size_t i, res = 0;
67
68 for (i = 0; i < gimple_phi_num_args (phi); i++)
69 if (!flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, i)->src))
70 {
71 loop = gimple_phi_arg_edge (phi, i)->src->loop_father;
72 res = i;
73 }
74
75 return res;
76 }
77
78 /* Removes a simple copy phi node "RES = phi (INIT, RES)" at position
79 PSI by inserting on the loop ENTRY edge assignment "RES = INIT". */
80
81 static void
82 remove_simple_copy_phi (gimple_stmt_iterator *psi)
83 {
84 gimple phi = gsi_stmt (*psi);
85 tree res = gimple_phi_result (phi);
86 size_t entry = phi_arg_in_outermost_loop (phi);
87 tree init = gimple_phi_arg_def (phi, entry);
88 gimple stmt = gimple_build_assign (res, init);
89 edge e = gimple_phi_arg_edge (phi, entry);
90
91 remove_phi_node (psi, false);
92 gsi_insert_on_edge_immediate (e, stmt);
93 SSA_NAME_DEF_STMT (res) = stmt;
94 }
95
96 /* Removes an invariant phi node at position PSI by inserting on the
97 loop ENTRY edge the assignment RES = INIT. */
98
99 static void
100 remove_invariant_phi (sese region, gimple_stmt_iterator *psi)
101 {
102 gimple phi = gsi_stmt (*psi);
103 loop_p loop = loop_containing_stmt (phi);
104 tree res = gimple_phi_result (phi);
105 tree scev = scalar_evolution_in_region (region, loop, res);
106 size_t entry = phi_arg_in_outermost_loop (phi);
107 edge e = gimple_phi_arg_edge (phi, entry);
108 tree var;
109 gimple stmt;
110 gimple_seq stmts = NULL;
111
112 if (tree_contains_chrecs (scev, NULL))
113 scev = gimple_phi_arg_def (phi, entry);
114
115 var = force_gimple_operand (scev, &stmts, true, NULL_TREE);
116 stmt = gimple_build_assign (res, var);
117 remove_phi_node (psi, false);
118
119 gimple_seq_add_stmt (&stmts, stmt);
120 gsi_insert_seq_on_edge (e, stmts);
121 gsi_commit_edge_inserts ();
122 SSA_NAME_DEF_STMT (res) = stmt;
123 }
124
125 /* Returns true when the phi node at PSI is of the form "a = phi (a, x)". */
126
127 static inline bool
128 simple_copy_phi_p (gimple phi)
129 {
130 tree res;
131
132 if (gimple_phi_num_args (phi) != 2)
133 return false;
134
135 res = gimple_phi_result (phi);
136 return (res == gimple_phi_arg_def (phi, 0)
137 || res == gimple_phi_arg_def (phi, 1));
138 }
139
140 /* Returns true when the phi node at position PSI is a reduction phi
141 node in REGION. Otherwise moves the pointer PSI to the next phi to
142 be considered. */
143
144 static bool
145 reduction_phi_p (sese region, gimple_stmt_iterator *psi)
146 {
147 loop_p loop;
148 gimple phi = gsi_stmt (*psi);
149 tree res = gimple_phi_result (phi);
150
151 loop = loop_containing_stmt (phi);
152
153 if (simple_copy_phi_p (phi))
154 {
155 /* PRE introduces phi nodes like these, for an example,
156 see id-5.f in the fortran graphite testsuite:
157
158 # prephitmp.85_265 = PHI <prephitmp.85_258(33), prephitmp.85_265(18)>
159 */
160 remove_simple_copy_phi (psi);
161 return false;
162 }
163
164 if (scev_analyzable_p (res, region))
165 {
166 tree scev = scalar_evolution_in_region (region, loop, res);
167
168 if (evolution_function_is_invariant_p (scev, loop->num))
169 remove_invariant_phi (region, psi);
170 else
171 gsi_next (psi);
172
173 return false;
174 }
175
176 /* All the other cases are considered reductions. */
177 return true;
178 }
179
180 /* Store the GRAPHITE representation of BB. */
181
182 static gimple_bb_p
183 new_gimple_bb (basic_block bb, VEC (data_reference_p, heap) *drs)
184 {
185 struct gimple_bb *gbb;
186
187 gbb = XNEW (struct gimple_bb);
188 bb->aux = gbb;
189 GBB_BB (gbb) = bb;
190 GBB_DATA_REFS (gbb) = drs;
191 GBB_CONDITIONS (gbb) = NULL;
192 GBB_CONDITION_CASES (gbb) = NULL;
193
194 return gbb;
195 }
196
197 static void
198 free_data_refs_aux (VEC (data_reference_p, heap) *datarefs)
199 {
200 unsigned int i;
201 struct data_reference *dr;
202
203 FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr)
204 if (dr->aux)
205 {
206 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
207
208 free (bap->alias_set);
209
210 free (bap);
211 dr->aux = NULL;
212 }
213 }
214 /* Frees GBB. */
215
216 static void
217 free_gimple_bb (struct gimple_bb *gbb)
218 {
219 free_data_refs_aux (GBB_DATA_REFS (gbb));
220 free_data_refs (GBB_DATA_REFS (gbb));
221
222 VEC_free (gimple, heap, GBB_CONDITIONS (gbb));
223 VEC_free (gimple, heap, GBB_CONDITION_CASES (gbb));
224 GBB_BB (gbb)->aux = 0;
225 XDELETE (gbb);
226 }
227
228 /* Deletes all gimple bbs in SCOP. */
229
230 static void
231 remove_gbbs_in_scop (scop_p scop)
232 {
233 int i;
234 poly_bb_p pbb;
235
236 FOR_EACH_VEC_ELT (poly_bb_p, SCOP_BBS (scop), i, pbb)
237 free_gimple_bb (PBB_BLACK_BOX (pbb));
238 }
239
240 /* Deletes all scops in SCOPS. */
241
242 void
243 free_scops (VEC (scop_p, heap) *scops)
244 {
245 int i;
246 scop_p scop;
247
248 FOR_EACH_VEC_ELT (scop_p, scops, i, scop)
249 {
250 remove_gbbs_in_scop (scop);
251 free_sese (SCOP_REGION (scop));
252 free_scop (scop);
253 }
254
255 VEC_free (scop_p, heap, scops);
256 }
257
258 /* Same as outermost_loop_in_sese, returns the outermost loop
259 containing BB in REGION, but makes sure that the returned loop
260 belongs to the REGION, and so this returns the first loop in the
261 REGION when the loop containing BB does not belong to REGION. */
262
263 static loop_p
264 outermost_loop_in_sese_1 (sese region, basic_block bb)
265 {
266 loop_p nest = outermost_loop_in_sese (region, bb);
267
268 if (loop_in_sese_p (nest, region))
269 return nest;
270
271 /* When the basic block BB does not belong to a loop in the region,
272 return the first loop in the region. */
273 nest = nest->inner;
274 while (nest)
275 if (loop_in_sese_p (nest, region))
276 break;
277 else
278 nest = nest->next;
279
280 gcc_assert (nest);
281 return nest;
282 }
283
284 /* Generates a polyhedral black box only if the bb contains interesting
285 information. */
286
287 static gimple_bb_p
288 try_generate_gimple_bb (scop_p scop, basic_block bb)
289 {
290 VEC (data_reference_p, heap) *drs = VEC_alloc (data_reference_p, heap, 5);
291 sese region = SCOP_REGION (scop);
292 loop_p nest = outermost_loop_in_sese_1 (region, bb);
293 gimple_stmt_iterator gsi;
294
295 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
296 {
297 gimple stmt = gsi_stmt (gsi);
298 loop_p loop;
299
300 if (is_gimple_debug (stmt))
301 continue;
302
303 loop = loop_containing_stmt (stmt);
304 if (!loop_in_sese_p (loop, region))
305 loop = nest;
306
307 graphite_find_data_references_in_stmt (nest, loop, stmt, &drs);
308 }
309
310 return new_gimple_bb (bb, drs);
311 }
312
313 /* Returns true if all predecessors of BB, that are not dominated by BB, are
314 marked in MAP. The predecessors dominated by BB are loop latches and will
315 be handled after BB. */
316
317 static bool
318 all_non_dominated_preds_marked_p (basic_block bb, sbitmap map)
319 {
320 edge e;
321 edge_iterator ei;
322
323 FOR_EACH_EDGE (e, ei, bb->preds)
324 if (!TEST_BIT (map, e->src->index)
325 && !dominated_by_p (CDI_DOMINATORS, e->src, bb))
326 return false;
327
328 return true;
329 }
330
331 /* Compare the depth of two basic_block's P1 and P2. */
332
333 static int
334 compare_bb_depths (const void *p1, const void *p2)
335 {
336 const_basic_block const bb1 = *(const_basic_block const*)p1;
337 const_basic_block const bb2 = *(const_basic_block const*)p2;
338 int d1 = loop_depth (bb1->loop_father);
339 int d2 = loop_depth (bb2->loop_father);
340
341 if (d1 < d2)
342 return 1;
343
344 if (d1 > d2)
345 return -1;
346
347 return 0;
348 }
349
350 /* Sort the basic blocks from DOM such that the first are the ones at
351 a deepest loop level. */
352
353 static void
354 graphite_sort_dominated_info (VEC (basic_block, heap) *dom)
355 {
356 VEC_qsort (basic_block, dom, compare_bb_depths);
357 }
358
359 /* Recursive helper function for build_scops_bbs. */
360
361 static void
362 build_scop_bbs_1 (scop_p scop, sbitmap visited, basic_block bb)
363 {
364 sese region = SCOP_REGION (scop);
365 VEC (basic_block, heap) *dom;
366 poly_bb_p pbb;
367
368 if (TEST_BIT (visited, bb->index)
369 || !bb_in_sese_p (bb, region))
370 return;
371
372 pbb = new_poly_bb (scop, try_generate_gimple_bb (scop, bb));
373 VEC_safe_push (poly_bb_p, heap, SCOP_BBS (scop), pbb);
374 SET_BIT (visited, bb->index);
375
376 dom = get_dominated_by (CDI_DOMINATORS, bb);
377
378 if (dom == NULL)
379 return;
380
381 graphite_sort_dominated_info (dom);
382
383 while (!VEC_empty (basic_block, dom))
384 {
385 int i;
386 basic_block dom_bb;
387
388 FOR_EACH_VEC_ELT (basic_block, dom, i, dom_bb)
389 if (all_non_dominated_preds_marked_p (dom_bb, visited))
390 {
391 build_scop_bbs_1 (scop, visited, dom_bb);
392 VEC_unordered_remove (basic_block, dom, i);
393 break;
394 }
395 }
396
397 VEC_free (basic_block, heap, dom);
398 }
399
400 /* Gather the basic blocks belonging to the SCOP. */
401
402 static void
403 build_scop_bbs (scop_p scop)
404 {
405 sbitmap visited = sbitmap_alloc (last_basic_block);
406 sese region = SCOP_REGION (scop);
407
408 sbitmap_zero (visited);
409 build_scop_bbs_1 (scop, visited, SESE_ENTRY_BB (region));
410 sbitmap_free (visited);
411 }
412
413 /* Return an ISL identifier for the polyhedral basic block PBB. */
414
415 static isl_id *
416 isl_id_for_pbb (scop_p s, poly_bb_p pbb)
417 {
418 char name[50];
419 snprintf (name, sizeof (name), "S_%d", pbb_index (pbb));
420 return isl_id_alloc (s->ctx, name, pbb);
421 }
422
423 /* Converts the STATIC_SCHEDULE of PBB into a scattering polyhedron.
424 We generate SCATTERING_DIMENSIONS scattering dimensions.
425
426 CLooG 0.15.0 and previous versions require, that all
427 scattering functions of one CloogProgram have the same number of
428 scattering dimensions, therefore we allow to specify it. This
429 should be removed in future versions of CLooG.
430
431 The scattering polyhedron consists of these dimensions: scattering,
432 loop_iterators, parameters.
433
434 Example:
435
436 | scattering_dimensions = 5
437 | used_scattering_dimensions = 3
438 | nb_iterators = 1
439 | scop_nb_params = 2
440 |
441 | Schedule:
442 | i
443 | 4 5
444 |
445 | Scattering polyhedron:
446 |
447 | scattering: {s1, s2, s3, s4, s5}
448 | loop_iterators: {i}
449 | parameters: {p1, p2}
450 |
451 | s1 s2 s3 s4 s5 i p1 p2 1
452 | 1 0 0 0 0 0 0 0 -4 = 0
453 | 0 1 0 0 0 -1 0 0 0 = 0
454 | 0 0 1 0 0 0 0 0 -5 = 0 */
455
456 static void
457 build_pbb_scattering_polyhedrons (isl_aff *static_sched,
458 poly_bb_p pbb, int scattering_dimensions)
459 {
460 int i;
461 int nb_iterators = pbb_dim_iter_domain (pbb);
462 int used_scattering_dimensions = nb_iterators * 2 + 1;
463 isl_int val;
464 isl_space *dc, *dm;
465
466 gcc_assert (scattering_dimensions >= used_scattering_dimensions);
467
468 isl_int_init (val);
469
470 dc = isl_set_get_space (pbb->domain);
471 dm = isl_space_add_dims (isl_space_from_domain (dc),
472 isl_dim_out, scattering_dimensions);
473 pbb->schedule = isl_map_universe (dm);
474
475 for (i = 0; i < scattering_dimensions; i++)
476 {
477 /* Textual order inside this loop. */
478 if ((i % 2) == 0)
479 {
480 isl_constraint *c = isl_equality_alloc
481 (isl_local_space_from_space (isl_map_get_space (pbb->schedule)));
482
483 if (0 != isl_aff_get_coefficient (static_sched, isl_dim_in,
484 i / 2, &val))
485 gcc_unreachable ();
486
487 isl_int_neg (val, val);
488 c = isl_constraint_set_constant (c, val);
489 c = isl_constraint_set_coefficient_si (c, isl_dim_out, i, 1);
490 pbb->schedule = isl_map_add_constraint (pbb->schedule, c);
491 }
492
493 /* Iterations of this loop. */
494 else /* if ((i % 2) == 1) */
495 {
496 int loop = (i - 1) / 2;
497 pbb->schedule = isl_map_equate (pbb->schedule, isl_dim_in, loop,
498 isl_dim_out, i);
499 }
500 }
501
502 isl_int_clear (val);
503
504 pbb->transformed = isl_map_copy (pbb->schedule);
505 }
506
507 /* Build for BB the static schedule.
508
509 The static schedule is a Dewey numbering of the abstract syntax
510 tree: http://en.wikipedia.org/wiki/Dewey_Decimal_Classification
511
512 The following example informally defines the static schedule:
513
514 A
515 for (i: ...)
516 {
517 for (j: ...)
518 {
519 B
520 C
521 }
522
523 for (k: ...)
524 {
525 D
526 E
527 }
528 }
529 F
530
531 Static schedules for A to F:
532
533 DEPTH
534 0 1 2
535 A 0
536 B 1 0 0
537 C 1 0 1
538 D 1 1 0
539 E 1 1 1
540 F 2
541 */
542
543 static void
544 build_scop_scattering (scop_p scop)
545 {
546 int i;
547 poly_bb_p pbb;
548 gimple_bb_p previous_gbb = NULL;
549 isl_space *dc = isl_set_get_space (scop->context);
550 isl_aff *static_sched;
551
552 dc = isl_space_add_dims (dc, isl_dim_set, number_of_loops());
553 static_sched = isl_aff_zero_on_domain (isl_local_space_from_space (dc));
554
555 /* We have to start schedules at 0 on the first component and
556 because we cannot compare_prefix_loops against a previous loop,
557 prefix will be equal to zero, and that index will be
558 incremented before copying. */
559 static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in, 0, -1);
560
561 FOR_EACH_VEC_ELT (poly_bb_p, SCOP_BBS (scop), i, pbb)
562 {
563 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
564 int prefix;
565 int nb_scat_dims = pbb_dim_iter_domain (pbb) * 2 + 1;
566
567 if (previous_gbb)
568 prefix = nb_common_loops (SCOP_REGION (scop), previous_gbb, gbb);
569 else
570 prefix = 0;
571
572 previous_gbb = gbb;
573
574 static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in,
575 prefix, 1);
576 build_pbb_scattering_polyhedrons (static_sched, pbb, nb_scat_dims);
577 }
578
579 isl_aff_free (static_sched);
580 }
581
582 static isl_pw_aff *extract_affine (scop_p, tree, __isl_take isl_space *space);
583
584 /* Extract an affine expression from the chain of recurrence E. */
585
586 static isl_pw_aff *
587 extract_affine_chrec (scop_p s, tree e, __isl_take isl_space *space)
588 {
589 isl_pw_aff *lhs = extract_affine (s, CHREC_LEFT (e), isl_space_copy (space));
590 isl_pw_aff *rhs = extract_affine (s, CHREC_RIGHT (e), isl_space_copy (space));
591 isl_local_space *ls = isl_local_space_from_space (space);
592 unsigned pos = sese_loop_depth ((sese) s->region,
593 get_loop (CHREC_VARIABLE (e))) - 1;
594 isl_aff *loop = isl_aff_set_coefficient_si
595 (isl_aff_zero_on_domain (ls), isl_dim_in, pos, 1);
596 isl_pw_aff *l = isl_pw_aff_from_aff (loop);
597
598 /* Before multiplying, make sure that the result is affine. */
599 gcc_assert (isl_pw_aff_is_cst (rhs)
600 || isl_pw_aff_is_cst (l));
601
602 return isl_pw_aff_add (lhs, isl_pw_aff_mul (rhs, l));
603 }
604
605 /* Extract an affine expression from the mult_expr E. */
606
607 static isl_pw_aff *
608 extract_affine_mul (scop_p s, tree e, __isl_take isl_space *space)
609 {
610 isl_pw_aff *lhs = extract_affine (s, TREE_OPERAND (e, 0),
611 isl_space_copy (space));
612 isl_pw_aff *rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
613
614 if (!isl_pw_aff_is_cst (lhs)
615 && !isl_pw_aff_is_cst (rhs))
616 {
617 isl_pw_aff_free (lhs);
618 isl_pw_aff_free (rhs);
619 return NULL;
620 }
621
622 return isl_pw_aff_mul (lhs, rhs);
623 }
624
625 /* Return an ISL identifier from the name of the ssa_name E. */
626
627 static isl_id *
628 isl_id_for_ssa_name (scop_p s, tree e)
629 {
630 const char *name = get_name (e);
631 isl_id *id;
632
633 if (name)
634 id = isl_id_alloc (s->ctx, name, e);
635 else
636 {
637 char name1[50];
638 snprintf (name1, sizeof (name1), "P_%d", SSA_NAME_VERSION (e));
639 id = isl_id_alloc (s->ctx, name1, e);
640 }
641
642 return id;
643 }
644
645 /* Return an ISL identifier for the data reference DR. */
646
647 static isl_id *
648 isl_id_for_dr (scop_p s, data_reference_p dr ATTRIBUTE_UNUSED)
649 {
650 /* Data references all get the same isl_id. They need to be comparable
651 and are distinguished through the first dimension, which contains the
652 alias set number. */
653 return isl_id_alloc (s->ctx, "", 0);
654 }
655
656 /* Extract an affine expression from the ssa_name E. */
657
658 static isl_pw_aff *
659 extract_affine_name (scop_p s, tree e, __isl_take isl_space *space)
660 {
661 isl_aff *aff;
662 isl_set *dom;
663 isl_id *id;
664 int dimension;
665
666 id = isl_id_for_ssa_name (s, e);
667 dimension = isl_space_find_dim_by_id (space, isl_dim_param, id);
668 isl_id_free(id);
669 dom = isl_set_universe (isl_space_copy (space));
670 aff = isl_aff_zero_on_domain (isl_local_space_from_space (space));
671 aff = isl_aff_add_coefficient_si (aff, isl_dim_param, dimension, 1);
672 return isl_pw_aff_alloc (dom, aff);
673 }
674
675 /* Extract an affine expression from the gmp constant G. */
676
677 static isl_pw_aff *
678 extract_affine_gmp (mpz_t g, __isl_take isl_space *space)
679 {
680 isl_local_space *ls = isl_local_space_from_space (isl_space_copy (space));
681 isl_aff *aff = isl_aff_zero_on_domain (ls);
682 isl_set *dom = isl_set_universe (space);
683 isl_int v;
684
685 isl_int_init (v);
686 isl_int_set_gmp (v, g);
687 aff = isl_aff_add_constant (aff, v);
688 isl_int_clear (v);
689
690 return isl_pw_aff_alloc (dom, aff);
691 }
692
693 /* Extract an affine expression from the integer_cst E. */
694
695 static isl_pw_aff *
696 extract_affine_int (tree e, __isl_take isl_space *space)
697 {
698 isl_pw_aff *res;
699 mpz_t g;
700
701 mpz_init (g);
702 tree_int_to_gmp (e, g);
703 res = extract_affine_gmp (g, space);
704 mpz_clear (g);
705
706 return res;
707 }
708
709 /* Compute pwaff mod 2^width. */
710
711 static isl_pw_aff *
712 wrap (isl_pw_aff *pwaff, unsigned width)
713 {
714 isl_int mod;
715
716 isl_int_init (mod);
717 isl_int_set_si (mod, 1);
718 isl_int_mul_2exp (mod, mod, width);
719
720 pwaff = isl_pw_aff_mod (pwaff, mod);
721
722 isl_int_clear (mod);
723
724 return pwaff;
725 }
726
727 /* When parameter NAME is in REGION, returns its index in SESE_PARAMS.
728 Otherwise returns -1. */
729
730 static inline int
731 parameter_index_in_region_1 (tree name, sese region)
732 {
733 int i;
734 tree p;
735
736 gcc_assert (TREE_CODE (name) == SSA_NAME);
737
738 FOR_EACH_VEC_ELT (tree, SESE_PARAMS (region), i, p)
739 if (p == name)
740 return i;
741
742 return -1;
743 }
744
745 /* When the parameter NAME is in REGION, returns its index in
746 SESE_PARAMS. Otherwise this function inserts NAME in SESE_PARAMS
747 and returns the index of NAME. */
748
749 static int
750 parameter_index_in_region (tree name, sese region)
751 {
752 int i;
753
754 gcc_assert (TREE_CODE (name) == SSA_NAME);
755
756 i = parameter_index_in_region_1 (name, region);
757 if (i != -1)
758 return i;
759
760 gcc_assert (SESE_ADD_PARAMS (region));
761
762 i = VEC_length (tree, SESE_PARAMS (region));
763 VEC_safe_push (tree, heap, SESE_PARAMS (region), name);
764 return i;
765 }
766
767 /* Extract an affine expression from the tree E in the scop S. */
768
769 static isl_pw_aff *
770 extract_affine (scop_p s, tree e, __isl_take isl_space *space)
771 {
772 isl_pw_aff *lhs, *rhs, *res;
773 tree type;
774
775 if (e == chrec_dont_know) {
776 isl_space_free (space);
777 return NULL;
778 }
779
780 switch (TREE_CODE (e))
781 {
782 case POLYNOMIAL_CHREC:
783 res = extract_affine_chrec (s, e, space);
784 break;
785
786 case MULT_EXPR:
787 res = extract_affine_mul (s, e, space);
788 break;
789
790 case PLUS_EXPR:
791 case POINTER_PLUS_EXPR:
792 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
793 rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
794 res = isl_pw_aff_add (lhs, rhs);
795 break;
796
797 case MINUS_EXPR:
798 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
799 rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
800 res = isl_pw_aff_sub (lhs, rhs);
801 break;
802
803 case NEGATE_EXPR:
804 case BIT_NOT_EXPR:
805 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
806 rhs = extract_affine (s, integer_minus_one_node, space);
807 res = isl_pw_aff_mul (lhs, rhs);
808 break;
809
810 case SSA_NAME:
811 gcc_assert (-1 != parameter_index_in_region_1 (e, SCOP_REGION (s)));
812 res = extract_affine_name (s, e, space);
813 break;
814
815 case INTEGER_CST:
816 res = extract_affine_int (e, space);
817 /* No need to wrap a single integer. */
818 return res;
819
820 CASE_CONVERT:
821 case NON_LVALUE_EXPR:
822 res = extract_affine (s, TREE_OPERAND (e, 0), space);
823 break;
824
825 default:
826 gcc_unreachable ();
827 break;
828 }
829
830 type = TREE_TYPE (e);
831 if (TYPE_UNSIGNED (type))
832 res = wrap (res, TYPE_PRECISION (type));
833
834 return res;
835 }
836
837 /* In the context of sese S, scan the expression E and translate it to
838 a linear expression C. When parsing a symbolic multiplication, K
839 represents the constant multiplier of an expression containing
840 parameters. */
841
842 static void
843 scan_tree_for_params (sese s, tree e)
844 {
845 if (e == chrec_dont_know)
846 return;
847
848 switch (TREE_CODE (e))
849 {
850 case POLYNOMIAL_CHREC:
851 scan_tree_for_params (s, CHREC_LEFT (e));
852 break;
853
854 case MULT_EXPR:
855 if (chrec_contains_symbols (TREE_OPERAND (e, 0)))
856 scan_tree_for_params (s, TREE_OPERAND (e, 0));
857 else
858 scan_tree_for_params (s, TREE_OPERAND (e, 1));
859 break;
860
861 case PLUS_EXPR:
862 case POINTER_PLUS_EXPR:
863 case MINUS_EXPR:
864 scan_tree_for_params (s, TREE_OPERAND (e, 0));
865 scan_tree_for_params (s, TREE_OPERAND (e, 1));
866 break;
867
868 case NEGATE_EXPR:
869 case BIT_NOT_EXPR:
870 CASE_CONVERT:
871 case NON_LVALUE_EXPR:
872 scan_tree_for_params (s, TREE_OPERAND (e, 0));
873 break;
874
875 case SSA_NAME:
876 parameter_index_in_region (e, s);
877 break;
878
879 case INTEGER_CST:
880 case ADDR_EXPR:
881 break;
882
883 default:
884 gcc_unreachable ();
885 break;
886 }
887 }
888
889 /* Find parameters with respect to REGION in BB. We are looking in memory
890 access functions, conditions and loop bounds. */
891
892 static void
893 find_params_in_bb (sese region, gimple_bb_p gbb)
894 {
895 int i;
896 unsigned j;
897 data_reference_p dr;
898 gimple stmt;
899 loop_p loop = GBB_BB (gbb)->loop_father;
900
901 /* Find parameters in the access functions of data references. */
902 FOR_EACH_VEC_ELT (data_reference_p, GBB_DATA_REFS (gbb), i, dr)
903 for (j = 0; j < DR_NUM_DIMENSIONS (dr); j++)
904 scan_tree_for_params (region, DR_ACCESS_FN (dr, j));
905
906 /* Find parameters in conditional statements. */
907 FOR_EACH_VEC_ELT (gimple, GBB_CONDITIONS (gbb), i, stmt)
908 {
909 tree lhs = scalar_evolution_in_region (region, loop,
910 gimple_cond_lhs (stmt));
911 tree rhs = scalar_evolution_in_region (region, loop,
912 gimple_cond_rhs (stmt));
913
914 scan_tree_for_params (region, lhs);
915 scan_tree_for_params (region, rhs);
916 }
917 }
918
919 /* Record the parameters used in the SCOP. A variable is a parameter
920 in a scop if it does not vary during the execution of that scop. */
921
922 static void
923 find_scop_parameters (scop_p scop)
924 {
925 poly_bb_p pbb;
926 unsigned i;
927 sese region = SCOP_REGION (scop);
928 struct loop *loop;
929 int nbp;
930
931 /* Find the parameters used in the loop bounds. */
932 FOR_EACH_VEC_ELT (loop_p, SESE_LOOP_NEST (region), i, loop)
933 {
934 tree nb_iters = number_of_latch_executions (loop);
935
936 if (!chrec_contains_symbols (nb_iters))
937 continue;
938
939 nb_iters = scalar_evolution_in_region (region, loop, nb_iters);
940 scan_tree_for_params (region, nb_iters);
941 }
942
943 /* Find the parameters used in data accesses. */
944 FOR_EACH_VEC_ELT (poly_bb_p, SCOP_BBS (scop), i, pbb)
945 find_params_in_bb (region, PBB_BLACK_BOX (pbb));
946
947 nbp = sese_nb_params (region);
948 scop_set_nb_params (scop, nbp);
949 SESE_ADD_PARAMS (region) = false;
950
951 {
952 tree e;
953 isl_space *space = isl_space_set_alloc (scop->ctx, nbp, 0);
954
955 FOR_EACH_VEC_ELT (tree, SESE_PARAMS (region), i, e)
956 space = isl_space_set_dim_id (space, isl_dim_param, i,
957 isl_id_for_ssa_name (scop, e));
958
959 scop->context = isl_set_universe (space);
960 }
961 }
962
963 /* Builds the constraint polyhedra for LOOP in SCOP. OUTER_PH gives
964 the constraints for the surrounding loops. */
965
966 static void
967 build_loop_iteration_domains (scop_p scop, struct loop *loop,
968 int nb,
969 isl_set *outer, isl_set **doms)
970 {
971 tree nb_iters = number_of_latch_executions (loop);
972 sese region = SCOP_REGION (scop);
973
974 isl_set *inner = isl_set_copy (outer);
975 isl_space *space;
976 isl_constraint *c;
977 int pos = isl_set_dim (outer, isl_dim_set);
978 isl_int v;
979 mpz_t g;
980
981 mpz_init (g);
982 isl_int_init (v);
983
984 inner = isl_set_add_dims (inner, isl_dim_set, 1);
985 space = isl_set_get_space (inner);
986
987 /* 0 <= loop_i */
988 c = isl_inequality_alloc
989 (isl_local_space_from_space (isl_space_copy (space)));
990 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, 1);
991 inner = isl_set_add_constraint (inner, c);
992
993 /* loop_i <= cst_nb_iters */
994 if (TREE_CODE (nb_iters) == INTEGER_CST)
995 {
996 c = isl_inequality_alloc
997 (isl_local_space_from_space(isl_space_copy (space)));
998 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1);
999 tree_int_to_gmp (nb_iters, g);
1000 isl_int_set_gmp (v, g);
1001 c = isl_constraint_set_constant (c, v);
1002 inner = isl_set_add_constraint (inner, c);
1003 }
1004
1005 /* loop_i <= expr_nb_iters */
1006 else if (!chrec_contains_undetermined (nb_iters))
1007 {
1008 double_int nit;
1009 isl_pw_aff *aff;
1010 isl_set *valid;
1011 isl_local_space *ls;
1012 isl_aff *al;
1013 isl_set *le;
1014
1015 nb_iters = scalar_evolution_in_region (region, loop, nb_iters);
1016
1017 aff = extract_affine (scop, nb_iters, isl_set_get_space (inner));
1018 valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (aff));
1019 valid = isl_set_project_out (valid, isl_dim_set, 0,
1020 isl_set_dim (valid, isl_dim_set));
1021 scop->context = isl_set_intersect (scop->context, valid);
1022
1023 ls = isl_local_space_from_space (isl_space_copy (space));
1024 al = isl_aff_set_coefficient_si (isl_aff_zero_on_domain (ls),
1025 isl_dim_in, pos, 1);
1026 le = isl_pw_aff_le_set (isl_pw_aff_from_aff (al),
1027 isl_pw_aff_copy (aff));
1028 inner = isl_set_intersect (inner, le);
1029
1030 if (max_stmt_executions (loop, &nit))
1031 {
1032 /* Insert in the context the constraints from the
1033 estimation of the number of iterations NIT and the
1034 symbolic number of iterations (involving parameter
1035 names) NB_ITERS. First, build the affine expression
1036 "NIT - NB_ITERS" and then say that it is positive,
1037 i.e., NIT approximates NB_ITERS: "NIT >= NB_ITERS". */
1038 isl_pw_aff *approx;
1039 mpz_t g;
1040 isl_set *x;
1041 isl_constraint *c;
1042
1043 mpz_init (g);
1044 mpz_set_double_int (g, nit, false);
1045 mpz_sub_ui (g, g, 1);
1046 approx = extract_affine_gmp (g, isl_set_get_space (inner));
1047 x = isl_pw_aff_ge_set (approx, aff);
1048 x = isl_set_project_out (x, isl_dim_set, 0,
1049 isl_set_dim (x, isl_dim_set));
1050 scop->context = isl_set_intersect (scop->context, x);
1051
1052 c = isl_inequality_alloc
1053 (isl_local_space_from_space (isl_space_copy (space)));
1054 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1);
1055 isl_int_set_gmp (v, g);
1056 mpz_clear (g);
1057 c = isl_constraint_set_constant (c, v);
1058 inner = isl_set_add_constraint (inner, c);
1059 }
1060 }
1061 else
1062 gcc_unreachable ();
1063
1064 if (loop->inner && loop_in_sese_p (loop->inner, region))
1065 build_loop_iteration_domains (scop, loop->inner, nb + 1,
1066 isl_set_copy (inner), doms);
1067
1068 if (nb != 0
1069 && loop->next
1070 && loop_in_sese_p (loop->next, region))
1071 build_loop_iteration_domains (scop, loop->next, nb,
1072 isl_set_copy (outer), doms);
1073
1074 doms[loop->num] = inner;
1075
1076 isl_set_free (outer);
1077 isl_space_free (space);
1078 isl_int_clear (v);
1079 mpz_clear (g);
1080 }
1081
1082 /* Returns a linear expression for tree T evaluated in PBB. */
1083
1084 static isl_pw_aff *
1085 create_pw_aff_from_tree (poly_bb_p pbb, tree t)
1086 {
1087 scop_p scop = PBB_SCOP (pbb);
1088
1089 t = scalar_evolution_in_region (SCOP_REGION (scop), pbb_loop (pbb), t);
1090 gcc_assert (!automatically_generated_chrec_p (t));
1091
1092 return extract_affine (scop, t, isl_set_get_space (pbb->domain));
1093 }
1094
1095 /* Add conditional statement STMT to pbb. CODE is used as the comparison
1096 operator. This allows us to invert the condition or to handle
1097 inequalities. */
1098
1099 static void
1100 add_condition_to_pbb (poly_bb_p pbb, gimple stmt, enum tree_code code)
1101 {
1102 isl_pw_aff *lhs = create_pw_aff_from_tree (pbb, gimple_cond_lhs (stmt));
1103 isl_pw_aff *rhs = create_pw_aff_from_tree (pbb, gimple_cond_rhs (stmt));
1104 isl_set *cond;
1105
1106 switch (code)
1107 {
1108 case LT_EXPR:
1109 cond = isl_pw_aff_lt_set (lhs, rhs);
1110 break;
1111
1112 case GT_EXPR:
1113 cond = isl_pw_aff_gt_set (lhs, rhs);
1114 break;
1115
1116 case LE_EXPR:
1117 cond = isl_pw_aff_le_set (lhs, rhs);
1118 break;
1119
1120 case GE_EXPR:
1121 cond = isl_pw_aff_ge_set (lhs, rhs);
1122 break;
1123
1124 case EQ_EXPR:
1125 cond = isl_pw_aff_eq_set (lhs, rhs);
1126 break;
1127
1128 case NE_EXPR:
1129 cond = isl_pw_aff_ne_set (lhs, rhs);
1130 break;
1131
1132 default:
1133 isl_pw_aff_free(lhs);
1134 isl_pw_aff_free(rhs);
1135 return;
1136 }
1137
1138 cond = isl_set_coalesce (cond);
1139 cond = isl_set_set_tuple_id (cond, isl_set_get_tuple_id (pbb->domain));
1140 pbb->domain = isl_set_intersect (pbb->domain, cond);
1141 }
1142
1143 /* Add conditions to the domain of PBB. */
1144
1145 static void
1146 add_conditions_to_domain (poly_bb_p pbb)
1147 {
1148 unsigned int i;
1149 gimple stmt;
1150 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
1151
1152 if (VEC_empty (gimple, GBB_CONDITIONS (gbb)))
1153 return;
1154
1155 FOR_EACH_VEC_ELT (gimple, GBB_CONDITIONS (gbb), i, stmt)
1156 switch (gimple_code (stmt))
1157 {
1158 case GIMPLE_COND:
1159 {
1160 enum tree_code code = gimple_cond_code (stmt);
1161
1162 /* The conditions for ELSE-branches are inverted. */
1163 if (!VEC_index (gimple, GBB_CONDITION_CASES (gbb), i))
1164 code = invert_tree_comparison (code, false);
1165
1166 add_condition_to_pbb (pbb, stmt, code);
1167 break;
1168 }
1169
1170 case GIMPLE_SWITCH:
1171 /* Switch statements are not supported right now - fall through. */
1172
1173 default:
1174 gcc_unreachable ();
1175 break;
1176 }
1177 }
1178
1179 /* Traverses all the GBBs of the SCOP and add their constraints to the
1180 iteration domains. */
1181
1182 static void
1183 add_conditions_to_constraints (scop_p scop)
1184 {
1185 int i;
1186 poly_bb_p pbb;
1187
1188 FOR_EACH_VEC_ELT (poly_bb_p, SCOP_BBS (scop), i, pbb)
1189 add_conditions_to_domain (pbb);
1190 }
1191
1192 /* Structure used to pass data to dom_walk. */
1193
1194 struct bsc
1195 {
1196 VEC (gimple, heap) **conditions, **cases;
1197 sese region;
1198 };
1199
1200 /* Returns a COND_EXPR statement when BB has a single predecessor, the
1201 edge between BB and its predecessor is not a loop exit edge, and
1202 the last statement of the single predecessor is a COND_EXPR. */
1203
1204 static gimple
1205 single_pred_cond_non_loop_exit (basic_block bb)
1206 {
1207 if (single_pred_p (bb))
1208 {
1209 edge e = single_pred_edge (bb);
1210 basic_block pred = e->src;
1211 gimple stmt;
1212
1213 if (loop_depth (pred->loop_father) > loop_depth (bb->loop_father))
1214 return NULL;
1215
1216 stmt = last_stmt (pred);
1217
1218 if (stmt && gimple_code (stmt) == GIMPLE_COND)
1219 return stmt;
1220 }
1221
1222 return NULL;
1223 }
1224
1225 /* Call-back for dom_walk executed before visiting the dominated
1226 blocks. */
1227
1228 static void
1229 build_sese_conditions_before (struct dom_walk_data *dw_data,
1230 basic_block bb)
1231 {
1232 struct bsc *data = (struct bsc *) dw_data->global_data;
1233 VEC (gimple, heap) **conditions = data->conditions;
1234 VEC (gimple, heap) **cases = data->cases;
1235 gimple_bb_p gbb;
1236 gimple stmt;
1237
1238 if (!bb_in_sese_p (bb, data->region))
1239 return;
1240
1241 stmt = single_pred_cond_non_loop_exit (bb);
1242
1243 if (stmt)
1244 {
1245 edge e = single_pred_edge (bb);
1246
1247 VEC_safe_push (gimple, heap, *conditions, stmt);
1248
1249 if (e->flags & EDGE_TRUE_VALUE)
1250 VEC_safe_push (gimple, heap, *cases, stmt);
1251 else
1252 VEC_safe_push (gimple, heap, *cases, NULL);
1253 }
1254
1255 gbb = gbb_from_bb (bb);
1256
1257 if (gbb)
1258 {
1259 GBB_CONDITIONS (gbb) = VEC_copy (gimple, heap, *conditions);
1260 GBB_CONDITION_CASES (gbb) = VEC_copy (gimple, heap, *cases);
1261 }
1262 }
1263
1264 /* Call-back for dom_walk executed after visiting the dominated
1265 blocks. */
1266
1267 static void
1268 build_sese_conditions_after (struct dom_walk_data *dw_data,
1269 basic_block bb)
1270 {
1271 struct bsc *data = (struct bsc *) dw_data->global_data;
1272 VEC (gimple, heap) **conditions = data->conditions;
1273 VEC (gimple, heap) **cases = data->cases;
1274
1275 if (!bb_in_sese_p (bb, data->region))
1276 return;
1277
1278 if (single_pred_cond_non_loop_exit (bb))
1279 {
1280 VEC_pop (gimple, *conditions);
1281 VEC_pop (gimple, *cases);
1282 }
1283 }
1284
1285 /* Record all conditions in REGION. */
1286
1287 static void
1288 build_sese_conditions (sese region)
1289 {
1290 struct dom_walk_data walk_data;
1291 VEC (gimple, heap) *conditions = VEC_alloc (gimple, heap, 3);
1292 VEC (gimple, heap) *cases = VEC_alloc (gimple, heap, 3);
1293 struct bsc data;
1294
1295 data.conditions = &conditions;
1296 data.cases = &cases;
1297 data.region = region;
1298
1299 walk_data.dom_direction = CDI_DOMINATORS;
1300 walk_data.initialize_block_local_data = NULL;
1301 walk_data.before_dom_children = build_sese_conditions_before;
1302 walk_data.after_dom_children = build_sese_conditions_after;
1303 walk_data.global_data = &data;
1304 walk_data.block_local_data_size = 0;
1305
1306 init_walk_dominator_tree (&walk_data);
1307 walk_dominator_tree (&walk_data, SESE_ENTRY_BB (region));
1308 fini_walk_dominator_tree (&walk_data);
1309
1310 VEC_free (gimple, heap, conditions);
1311 VEC_free (gimple, heap, cases);
1312 }
1313
1314 /* Add constraints on the possible values of parameter P from the type
1315 of P. */
1316
1317 static void
1318 add_param_constraints (scop_p scop, graphite_dim_t p)
1319 {
1320 tree parameter = VEC_index (tree, SESE_PARAMS (SCOP_REGION (scop)), p);
1321 tree type = TREE_TYPE (parameter);
1322 tree lb = NULL_TREE;
1323 tree ub = NULL_TREE;
1324
1325 if (POINTER_TYPE_P (type) || !TYPE_MIN_VALUE (type))
1326 lb = lower_bound_in_type (type, type);
1327 else
1328 lb = TYPE_MIN_VALUE (type);
1329
1330 if (POINTER_TYPE_P (type) || !TYPE_MAX_VALUE (type))
1331 ub = upper_bound_in_type (type, type);
1332 else
1333 ub = TYPE_MAX_VALUE (type);
1334
1335 if (lb)
1336 {
1337 isl_space *space = isl_set_get_space (scop->context);
1338 isl_constraint *c;
1339 mpz_t g;
1340 isl_int v;
1341
1342 c = isl_inequality_alloc (isl_local_space_from_space (space));
1343 mpz_init (g);
1344 isl_int_init (v);
1345 tree_int_to_gmp (lb, g);
1346 isl_int_set_gmp (v, g);
1347 isl_int_neg (v, v);
1348 mpz_clear (g);
1349 c = isl_constraint_set_constant (c, v);
1350 isl_int_clear (v);
1351 c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, 1);
1352
1353 scop->context = isl_set_add_constraint (scop->context, c);
1354 }
1355
1356 if (ub)
1357 {
1358 isl_space *space = isl_set_get_space (scop->context);
1359 isl_constraint *c;
1360 mpz_t g;
1361 isl_int v;
1362
1363 c = isl_inequality_alloc (isl_local_space_from_space (space));
1364
1365 mpz_init (g);
1366 isl_int_init (v);
1367 tree_int_to_gmp (ub, g);
1368 isl_int_set_gmp (v, g);
1369 mpz_clear (g);
1370 c = isl_constraint_set_constant (c, v);
1371 isl_int_clear (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 ();
1403 isl_set **doms = XCNEWVEC (isl_set *, nb_loops);
1404
1405 FOR_EACH_VEC_ELT (loop_p, 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 (poly_bb_p, 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 (host_integerp (low, 0)
1525 && high
1526 && host_integerp (high, 0)
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, heap) *drs)
1620 {
1621 int num_vertex = VEC_length (data_reference_p, drs);
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 (data_reference_p, drs, i, dr1)
1630 for (j = i + 1; VEC_iterate (data_reference_p, drs, j, dr2); j++)
1631 if (dr_may_alias_p (dr1, dr2, true))
1632 edge_num++;
1633
1634 fprintf (file, "$\n");
1635
1636 if (comment)
1637 fprintf (file, "c %s\n", comment);
1638
1639 fprintf (file, "p edge %d %d\n", num_vertex, edge_num);
1640
1641 FOR_EACH_VEC_ELT (data_reference_p, drs, i, dr1)
1642 for (j = i + 1; VEC_iterate (data_reference_p, drs, j, dr2); j++)
1643 if (dr_may_alias_p (dr1, dr2, true))
1644 fprintf (file, "e %d %d\n", i + 1, j + 1);
1645
1646 return true;
1647 }
1648
1649 /* Write to FILE the alias graph of data references in DOT format. */
1650
1651 static inline bool
1652 write_alias_graph_to_ascii_dot (FILE *file, char *comment,
1653 VEC (data_reference_p, heap) *drs)
1654 {
1655 int num_vertex = VEC_length (data_reference_p, drs);
1656 data_reference_p dr1, dr2;
1657 int i, j;
1658
1659 if (num_vertex == 0)
1660 return true;
1661
1662 fprintf (file, "$\n");
1663
1664 if (comment)
1665 fprintf (file, "c %s\n", comment);
1666
1667 /* First print all the vertices. */
1668 FOR_EACH_VEC_ELT (data_reference_p, drs, i, dr1)
1669 fprintf (file, "n%d;\n", i);
1670
1671 FOR_EACH_VEC_ELT (data_reference_p, drs, i, dr1)
1672 for (j = i + 1; VEC_iterate (data_reference_p, drs, 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, heap) *drs)
1684 {
1685 int num_vertex = VEC_length (data_reference_p, drs);
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 (data_reference_p, drs, i, dr1)
1698 for (j = i + 1; VEC_iterate (data_reference_p, drs, 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, heap) *drs)
1720 {
1721 int num_vertices = VEC_length (data_reference_p, drs);
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 (data_reference_p, drs, i, dr1)
1734 for (j = i+1; VEC_iterate (data_reference_p, drs, 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 = VEC_index (data_reference_p, 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, heap) *drs)
1806 {
1807 int num_vertex = VEC_length (data_reference_p, drs);
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 (data_reference_p, drs, i, dr1)
1814 for (j = i + 1; VEC_iterate (data_reference_p, drs, 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 = VEC_index (data_reference_p, 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, heap) *gbb_drs = GBB_DATA_REFS (PBB_BLACK_BOX (pbb));
1850
1851 FOR_EACH_VEC_ELT (data_reference_p, 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, heap) *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 VEC (data_reference_p, heap) *drs = VEC_alloc (data_reference_p, heap, 3);
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; VEC_iterate (poly_bb_p, SCOP_BBS (scop), i, pbb); i++)
1904 if (VEC_empty (data_reference_p, GBB_DATA_REFS (PBB_BLACK_BOX (pbb))))
1905 {
1906 free_gimple_bb (PBB_BLACK_BOX (pbb));
1907 free_poly_bb (pbb);
1908 VEC_ordered_remove (poly_bb_p, SCOP_BBS (scop), i);
1909 i--;
1910 }
1911
1912 FOR_EACH_VEC_ELT (poly_bb_p, SCOP_BBS (scop), i, pbb)
1913 for (j = 0; VEC_iterate (data_reference_p,
1914 GBB_DATA_REFS (PBB_BLACK_BOX (pbb)), j, dr); j++)
1915 VEC_safe_push (data_reference_p, heap, drs, dr);
1916
1917 FOR_EACH_VEC_ELT (data_reference_p, drs, i, dr)
1918 dr->aux = XNEW (base_alias_pair);
1919
1920 if (!build_alias_set_optimal_p (drs))
1921 {
1922 /* TODO: Add support when building alias set is not optimal. */
1923 ;
1924 }
1925
1926 build_base_obj_set_for_drs (drs);
1927
1928 /* When debugging, enable the following code. This cannot be used
1929 in production compilers. */
1930 if (0)
1931 dump_alias_graphs (drs);
1932
1933 VEC_free (data_reference_p, heap, drs);
1934
1935 FOR_EACH_VEC_ELT (poly_bb_p, SCOP_BBS (scop), i, pbb)
1936 build_pbb_drs (pbb);
1937 }
1938
1939 /* Return a gsi at the position of the phi node STMT. */
1940
1941 static gimple_stmt_iterator
1942 gsi_for_phi_node (gimple stmt)
1943 {
1944 gimple_stmt_iterator psi;
1945 basic_block bb = gimple_bb (stmt);
1946
1947 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
1948 if (stmt == gsi_stmt (psi))
1949 return psi;
1950
1951 gcc_unreachable ();
1952 return psi;
1953 }
1954
1955 /* Analyze all the data references of STMTS and add them to the
1956 GBB_DATA_REFS vector of BB. */
1957
1958 static void
1959 analyze_drs_in_stmts (scop_p scop, basic_block bb, VEC (gimple, heap) *stmts)
1960 {
1961 loop_p nest;
1962 gimple_bb_p gbb;
1963 gimple stmt;
1964 int i;
1965 sese region = SCOP_REGION (scop);
1966
1967 if (!bb_in_sese_p (bb, region))
1968 return;
1969
1970 nest = outermost_loop_in_sese_1 (region, bb);
1971 gbb = gbb_from_bb (bb);
1972
1973 FOR_EACH_VEC_ELT (gimple, stmts, i, stmt)
1974 {
1975 loop_p loop;
1976
1977 if (is_gimple_debug (stmt))
1978 continue;
1979
1980 loop = loop_containing_stmt (stmt);
1981 if (!loop_in_sese_p (loop, region))
1982 loop = nest;
1983
1984 graphite_find_data_references_in_stmt (nest, loop, stmt,
1985 &GBB_DATA_REFS (gbb));
1986 }
1987 }
1988
1989 /* Insert STMT at the end of the STMTS sequence and then insert the
1990 statements from STMTS at INSERT_GSI and call analyze_drs_in_stmts
1991 on STMTS. */
1992
1993 static void
1994 insert_stmts (scop_p scop, gimple stmt, gimple_seq stmts,
1995 gimple_stmt_iterator insert_gsi)
1996 {
1997 gimple_stmt_iterator gsi;
1998 VEC (gimple, heap) *x = VEC_alloc (gimple, heap, 3);
1999
2000 gimple_seq_add_stmt (&stmts, stmt);
2001 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2002 VEC_safe_push (gimple, heap, x, gsi_stmt (gsi));
2003
2004 gsi_insert_seq_before (&insert_gsi, stmts, GSI_SAME_STMT);
2005 analyze_drs_in_stmts (scop, gsi_bb (insert_gsi), x);
2006 VEC_free (gimple, heap, x);
2007 }
2008
2009 /* Insert the assignment "RES := EXPR" just after AFTER_STMT. */
2010
2011 static void
2012 insert_out_of_ssa_copy (scop_p scop, tree res, tree expr, gimple after_stmt)
2013 {
2014 gimple_seq stmts;
2015 gimple_stmt_iterator gsi;
2016 tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE);
2017 gimple stmt = gimple_build_assign (res, var);
2018 VEC (gimple, heap) *x = VEC_alloc (gimple, heap, 3);
2019
2020 gimple_seq_add_stmt (&stmts, stmt);
2021 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2022 VEC_safe_push (gimple, heap, x, gsi_stmt (gsi));
2023
2024 if (gimple_code (after_stmt) == GIMPLE_PHI)
2025 {
2026 gsi = gsi_after_labels (gimple_bb (after_stmt));
2027 gsi_insert_seq_before (&gsi, stmts, GSI_NEW_STMT);
2028 }
2029 else
2030 {
2031 gsi = gsi_for_stmt (after_stmt);
2032 gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT);
2033 }
2034
2035 analyze_drs_in_stmts (scop, gimple_bb (after_stmt), x);
2036 VEC_free (gimple, heap, x);
2037 }
2038
2039 /* Creates a poly_bb_p for basic_block BB from the existing PBB. */
2040
2041 static void
2042 new_pbb_from_pbb (scop_p scop, poly_bb_p pbb, basic_block bb)
2043 {
2044 VEC (data_reference_p, heap) *drs = VEC_alloc (data_reference_p, heap, 3);
2045 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
2046 gimple_bb_p gbb1 = new_gimple_bb (bb, drs);
2047 poly_bb_p pbb1 = new_poly_bb (scop, gbb1);
2048 int index, n = VEC_length (poly_bb_p, SCOP_BBS (scop));
2049
2050 /* The INDEX of PBB in SCOP_BBS. */
2051 for (index = 0; index < n; index++)
2052 if (VEC_index (poly_bb_p, SCOP_BBS (scop), index) == pbb)
2053 break;
2054
2055 pbb1->domain = isl_set_copy (pbb->domain);
2056
2057 GBB_PBB (gbb1) = pbb1;
2058 GBB_CONDITIONS (gbb1) = VEC_copy (gimple, heap, GBB_CONDITIONS (gbb));
2059 GBB_CONDITION_CASES (gbb1) = VEC_copy (gimple, heap, GBB_CONDITION_CASES (gbb));
2060 VEC_safe_insert (poly_bb_p, heap, SCOP_BBS (scop), 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 (res, var);
2072 basic_block bb;
2073 VEC (gimple, heap) *x = VEC_alloc (gimple, heap, 3);
2074
2075 gimple_seq_add_stmt (&stmts, stmt);
2076 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2077 VEC_safe_push (gimple, heap, x, 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 VEC_free (gimple, heap, x);
2091 }
2092
2093 /* Creates a zero dimension array of the same type as VAR. */
2094
2095 static tree
2096 create_zero_dim_array (tree var, const char *base_name)
2097 {
2098 tree index_type = build_index_type (integer_zero_node);
2099 tree elt_type = TREE_TYPE (var);
2100 tree array_type = build_array_type (elt_type, index_type);
2101 tree base = create_tmp_var (array_type, base_name);
2102
2103 return build4 (ARRAY_REF, elt_type, base, integer_zero_node, NULL_TREE,
2104 NULL_TREE);
2105 }
2106
2107 /* Returns true when PHI is a loop close phi node. */
2108
2109 static bool
2110 scalar_close_phi_node_p (gimple phi)
2111 {
2112 if (gimple_code (phi) != GIMPLE_PHI
2113 || !is_gimple_reg (gimple_phi_result (phi)))
2114 return false;
2115
2116 /* Note that loop close phi nodes should have a single argument
2117 because we translated the representation into a canonical form
2118 before Graphite: see canonicalize_loop_closed_ssa_form. */
2119 return (gimple_phi_num_args (phi) == 1);
2120 }
2121
2122 /* For a definition DEF in REGION, propagates the expression EXPR in
2123 all the uses of DEF outside REGION. */
2124
2125 static void
2126 propagate_expr_outside_region (tree def, tree expr, sese region)
2127 {
2128 imm_use_iterator imm_iter;
2129 gimple use_stmt;
2130 gimple_seq stmts;
2131 bool replaced_once = false;
2132
2133 gcc_assert (TREE_CODE (def) == SSA_NAME);
2134
2135 expr = force_gimple_operand (unshare_expr (expr), &stmts, true,
2136 NULL_TREE);
2137
2138 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2139 if (!is_gimple_debug (use_stmt)
2140 && !bb_in_sese_p (gimple_bb (use_stmt), region))
2141 {
2142 ssa_op_iter iter;
2143 use_operand_p use_p;
2144
2145 FOR_EACH_PHI_OR_STMT_USE (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2146 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0)
2147 && (replaced_once = true))
2148 replace_exp (use_p, expr);
2149
2150 update_stmt (use_stmt);
2151 }
2152
2153 if (replaced_once)
2154 {
2155 gsi_insert_seq_on_edge (SESE_ENTRY (region), stmts);
2156 gsi_commit_edge_inserts ();
2157 }
2158 }
2159
2160 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2161 dimension array for it. */
2162
2163 static void
2164 rewrite_close_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi)
2165 {
2166 sese region = SCOP_REGION (scop);
2167 gimple phi = gsi_stmt (*psi);
2168 tree res = gimple_phi_result (phi);
2169 basic_block bb = gimple_bb (phi);
2170 gimple_stmt_iterator gsi = gsi_after_labels (bb);
2171 tree arg = gimple_phi_arg_def (phi, 0);
2172 gimple stmt;
2173
2174 /* Note that loop close phi nodes should have a single argument
2175 because we translated the representation into a canonical form
2176 before Graphite: see canonicalize_loop_closed_ssa_form. */
2177 gcc_assert (gimple_phi_num_args (phi) == 1);
2178
2179 /* The phi node can be a non close phi node, when its argument is
2180 invariant, or a default definition. */
2181 if (is_gimple_min_invariant (arg)
2182 || SSA_NAME_IS_DEFAULT_DEF (arg))
2183 {
2184 propagate_expr_outside_region (res, arg, region);
2185 gsi_next (psi);
2186 return;
2187 }
2188
2189 else if (gimple_bb (SSA_NAME_DEF_STMT (arg))->loop_father == bb->loop_father)
2190 {
2191 propagate_expr_outside_region (res, arg, region);
2192 stmt = gimple_build_assign (res, arg);
2193 remove_phi_node (psi, false);
2194 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2195 SSA_NAME_DEF_STMT (res) = stmt;
2196 return;
2197 }
2198
2199 /* If res is scev analyzable and is not a scalar value, it is safe
2200 to ignore the close phi node: it will be code generated in the
2201 out of Graphite pass. */
2202 else if (scev_analyzable_p (res, region))
2203 {
2204 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (res));
2205 tree scev;
2206
2207 if (!loop_in_sese_p (loop, region))
2208 {
2209 loop = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2210 scev = scalar_evolution_in_region (region, loop, arg);
2211 scev = compute_overall_effect_of_inner_loop (loop, scev);
2212 }
2213 else
2214 scev = scalar_evolution_in_region (region, loop, res);
2215
2216 if (tree_does_not_contain_chrecs (scev))
2217 propagate_expr_outside_region (res, scev, region);
2218
2219 gsi_next (psi);
2220 return;
2221 }
2222 else
2223 {
2224 tree zero_dim_array = create_zero_dim_array (res, "Close_Phi");
2225
2226 stmt = gimple_build_assign (res, zero_dim_array);
2227
2228 if (TREE_CODE (arg) == SSA_NAME)
2229 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2230 SSA_NAME_DEF_STMT (arg));
2231 else
2232 insert_out_of_ssa_copy_on_edge (scop, single_pred_edge (bb),
2233 zero_dim_array, arg);
2234 }
2235
2236 remove_phi_node (psi, false);
2237 SSA_NAME_DEF_STMT (res) = stmt;
2238
2239 insert_stmts (scop, stmt, NULL, gsi_after_labels (bb));
2240 }
2241
2242 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2243 dimension array for it. */
2244
2245 static void
2246 rewrite_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi)
2247 {
2248 size_t i;
2249 gimple phi = gsi_stmt (*psi);
2250 basic_block bb = gimple_bb (phi);
2251 tree res = gimple_phi_result (phi);
2252 tree var;
2253 tree zero_dim_array = create_zero_dim_array (res, "phi_out_of_ssa");
2254 gimple stmt;
2255 gimple_seq stmts;
2256
2257 for (i = 0; i < gimple_phi_num_args (phi); i++)
2258 {
2259 tree arg = gimple_phi_arg_def (phi, i);
2260 edge e = gimple_phi_arg_edge (phi, i);
2261
2262 /* Avoid the insertion of code in the loop latch to please the
2263 pattern matching of the vectorizer. */
2264 if (TREE_CODE (arg) == SSA_NAME
2265 && e->src == bb->loop_father->latch)
2266 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2267 SSA_NAME_DEF_STMT (arg));
2268 else
2269 insert_out_of_ssa_copy_on_edge (scop, e, zero_dim_array, arg);
2270 }
2271
2272 var = force_gimple_operand (zero_dim_array, &stmts, true, NULL_TREE);
2273
2274 stmt = gimple_build_assign (res, var);
2275 remove_phi_node (psi, false);
2276 SSA_NAME_DEF_STMT (res) = stmt;
2277
2278 insert_stmts (scop, stmt, stmts, gsi_after_labels (bb));
2279 }
2280
2281 /* Rewrite the degenerate phi node at position PSI from the degenerate
2282 form "x = phi (y, y, ..., y)" to "x = y". */
2283
2284 static void
2285 rewrite_degenerate_phi (gimple_stmt_iterator *psi)
2286 {
2287 tree rhs;
2288 gimple stmt;
2289 gimple_stmt_iterator gsi;
2290 gimple phi = gsi_stmt (*psi);
2291 tree res = gimple_phi_result (phi);
2292 basic_block bb;
2293
2294 bb = gimple_bb (phi);
2295 rhs = degenerate_phi_result (phi);
2296 gcc_assert (rhs);
2297
2298 stmt = gimple_build_assign (res, rhs);
2299 remove_phi_node (psi, false);
2300 SSA_NAME_DEF_STMT (res) = stmt;
2301
2302 gsi = gsi_after_labels (bb);
2303 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2304 }
2305
2306 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2307
2308 static void
2309 rewrite_reductions_out_of_ssa (scop_p scop)
2310 {
2311 basic_block bb;
2312 gimple_stmt_iterator psi;
2313 sese region = SCOP_REGION (scop);
2314
2315 FOR_EACH_BB (bb)
2316 if (bb_in_sese_p (bb, region))
2317 for (psi = gsi_start_phis (bb); !gsi_end_p (psi);)
2318 {
2319 gimple phi = gsi_stmt (psi);
2320
2321 if (!is_gimple_reg (gimple_phi_result (phi)))
2322 {
2323 gsi_next (&psi);
2324 continue;
2325 }
2326
2327 if (gimple_phi_num_args (phi) > 1
2328 && degenerate_phi_result (phi))
2329 rewrite_degenerate_phi (&psi);
2330
2331 else if (scalar_close_phi_node_p (phi))
2332 rewrite_close_phi_out_of_ssa (scop, &psi);
2333
2334 else if (reduction_phi_p (region, &psi))
2335 rewrite_phi_out_of_ssa (scop, &psi);
2336 }
2337
2338 update_ssa (TODO_update_ssa);
2339 #ifdef ENABLE_CHECKING
2340 verify_loop_closed_ssa (true);
2341 #endif
2342 }
2343
2344 /* Rewrite the scalar dependence of DEF used in USE_STMT with a memory
2345 read from ZERO_DIM_ARRAY. */
2346
2347 static void
2348 rewrite_cross_bb_scalar_dependence (scop_p scop, tree zero_dim_array,
2349 tree def, gimple use_stmt)
2350 {
2351 gimple name_stmt;
2352 tree name;
2353 ssa_op_iter iter;
2354 use_operand_p use_p;
2355
2356 gcc_assert (gimple_code (use_stmt) != GIMPLE_PHI);
2357
2358 name = copy_ssa_name (def, NULL);
2359 name_stmt = gimple_build_assign (name, zero_dim_array);
2360
2361 gimple_assign_set_lhs (name_stmt, name);
2362 insert_stmts (scop, name_stmt, NULL, gsi_for_stmt (use_stmt));
2363
2364 FOR_EACH_SSA_USE_OPERAND (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2365 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0))
2366 replace_exp (use_p, name);
2367
2368 update_stmt (use_stmt);
2369 }
2370
2371 /* For every definition DEF in the SCOP that is used outside the scop,
2372 insert a closing-scop definition in the basic block just after this
2373 SCOP. */
2374
2375 static void
2376 handle_scalar_deps_crossing_scop_limits (scop_p scop, tree def, gimple stmt)
2377 {
2378 tree var = create_tmp_reg (TREE_TYPE (def), NULL);
2379 tree new_name = make_ssa_name (var, stmt);
2380 bool needs_copy = false;
2381 use_operand_p use_p;
2382 imm_use_iterator imm_iter;
2383 gimple use_stmt;
2384 sese region = SCOP_REGION (scop);
2385
2386 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2387 {
2388 if (!bb_in_sese_p (gimple_bb (use_stmt), region))
2389 {
2390 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
2391 {
2392 SET_USE (use_p, new_name);
2393 }
2394 update_stmt (use_stmt);
2395 needs_copy = true;
2396 }
2397 }
2398
2399 /* Insert in the empty BB just after the scop a use of DEF such
2400 that the rewrite of cross_bb_scalar_dependences won't insert
2401 arrays everywhere else. */
2402 if (needs_copy)
2403 {
2404 gimple assign = gimple_build_assign (new_name, def);
2405 gimple_stmt_iterator psi = gsi_after_labels (SESE_EXIT (region)->dest);
2406
2407 SSA_NAME_DEF_STMT (new_name) = assign;
2408 update_stmt (assign);
2409 gsi_insert_before (&psi, assign, GSI_SAME_STMT);
2410 }
2411 }
2412
2413 /* Rewrite the scalar dependences crossing the boundary of the BB
2414 containing STMT with an array. Return true when something has been
2415 changed. */
2416
2417 static bool
2418 rewrite_cross_bb_scalar_deps (scop_p scop, gimple_stmt_iterator *gsi)
2419 {
2420 sese region = SCOP_REGION (scop);
2421 gimple stmt = gsi_stmt (*gsi);
2422 imm_use_iterator imm_iter;
2423 tree def;
2424 basic_block def_bb;
2425 tree zero_dim_array = NULL_TREE;
2426 gimple use_stmt;
2427 bool res = false;
2428
2429 switch (gimple_code (stmt))
2430 {
2431 case GIMPLE_ASSIGN:
2432 def = gimple_assign_lhs (stmt);
2433 break;
2434
2435 case GIMPLE_CALL:
2436 def = gimple_call_lhs (stmt);
2437 break;
2438
2439 default:
2440 return false;
2441 }
2442
2443 if (!def
2444 || !is_gimple_reg (def))
2445 return false;
2446
2447 if (scev_analyzable_p (def, region))
2448 {
2449 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (def));
2450 tree scev = scalar_evolution_in_region (region, loop, def);
2451
2452 if (tree_contains_chrecs (scev, NULL))
2453 return false;
2454
2455 propagate_expr_outside_region (def, scev, region);
2456 return true;
2457 }
2458
2459 def_bb = gimple_bb (stmt);
2460
2461 handle_scalar_deps_crossing_scop_limits (scop, def, stmt);
2462
2463 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2464 if (gimple_code (use_stmt) == GIMPLE_PHI
2465 && (res = true))
2466 {
2467 gimple_stmt_iterator psi = gsi_for_stmt (use_stmt);
2468
2469 if (scalar_close_phi_node_p (gsi_stmt (psi)))
2470 rewrite_close_phi_out_of_ssa (scop, &psi);
2471 else
2472 rewrite_phi_out_of_ssa (scop, &psi);
2473 }
2474
2475 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2476 if (gimple_code (use_stmt) != GIMPLE_PHI
2477 && def_bb != gimple_bb (use_stmt)
2478 && !is_gimple_debug (use_stmt)
2479 && (res = true))
2480 {
2481 if (!zero_dim_array)
2482 {
2483 zero_dim_array = create_zero_dim_array
2484 (def, "Cross_BB_scalar_dependence");
2485 insert_out_of_ssa_copy (scop, zero_dim_array, def,
2486 SSA_NAME_DEF_STMT (def));
2487 gsi_next (gsi);
2488 }
2489
2490 rewrite_cross_bb_scalar_dependence (scop, zero_dim_array,
2491 def, use_stmt);
2492 }
2493
2494 return res;
2495 }
2496
2497 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2498
2499 static void
2500 rewrite_cross_bb_scalar_deps_out_of_ssa (scop_p scop)
2501 {
2502 basic_block bb;
2503 gimple_stmt_iterator psi;
2504 sese region = SCOP_REGION (scop);
2505 bool changed = false;
2506
2507 /* Create an extra empty BB after the scop. */
2508 split_edge (SESE_EXIT (region));
2509
2510 FOR_EACH_BB (bb)
2511 if (bb_in_sese_p (bb, region))
2512 for (psi = gsi_start_bb (bb); !gsi_end_p (psi); gsi_next (&psi))
2513 changed |= rewrite_cross_bb_scalar_deps (scop, &psi);
2514
2515 if (changed)
2516 {
2517 scev_reset_htab ();
2518 update_ssa (TODO_update_ssa);
2519 #ifdef ENABLE_CHECKING
2520 verify_loop_closed_ssa (true);
2521 #endif
2522 }
2523 }
2524
2525 /* Returns the number of pbbs that are in loops contained in SCOP. */
2526
2527 static int
2528 nb_pbbs_in_loops (scop_p scop)
2529 {
2530 int i;
2531 poly_bb_p pbb;
2532 int res = 0;
2533
2534 FOR_EACH_VEC_ELT (poly_bb_p, SCOP_BBS (scop), i, pbb)
2535 if (loop_in_sese_p (gbb_loop (PBB_BLACK_BOX (pbb)), SCOP_REGION (scop)))
2536 res++;
2537
2538 return res;
2539 }
2540
2541 /* Return the number of data references in BB that write in
2542 memory. */
2543
2544 static int
2545 nb_data_writes_in_bb (basic_block bb)
2546 {
2547 int res = 0;
2548 gimple_stmt_iterator gsi;
2549
2550 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
2551 if (gimple_vdef (gsi_stmt (gsi)))
2552 res++;
2553
2554 return res;
2555 }
2556
2557 /* Splits at STMT the basic block BB represented as PBB in the
2558 polyhedral form. */
2559
2560 static edge
2561 split_pbb (scop_p scop, poly_bb_p pbb, basic_block bb, gimple stmt)
2562 {
2563 edge e1 = split_block (bb, stmt);
2564 new_pbb_from_pbb (scop, pbb, e1->dest);
2565 return e1;
2566 }
2567
2568 /* Splits STMT out of its current BB. This is done for reduction
2569 statements for which we want to ignore data dependences. */
2570
2571 static basic_block
2572 split_reduction_stmt (scop_p scop, gimple stmt)
2573 {
2574 basic_block bb = gimple_bb (stmt);
2575 poly_bb_p pbb = pbb_from_bb (bb);
2576 gimple_bb_p gbb = gbb_from_bb (bb);
2577 edge e1;
2578 int i;
2579 data_reference_p dr;
2580
2581 /* Do not split basic blocks with no writes to memory: the reduction
2582 will be the only write to memory. */
2583 if (nb_data_writes_in_bb (bb) == 0
2584 /* Or if we have already marked BB as a reduction. */
2585 || PBB_IS_REDUCTION (pbb_from_bb (bb)))
2586 return bb;
2587
2588 e1 = split_pbb (scop, pbb, bb, stmt);
2589
2590 /* Split once more only when the reduction stmt is not the only one
2591 left in the original BB. */
2592 if (!gsi_one_before_end_p (gsi_start_nondebug_bb (bb)))
2593 {
2594 gimple_stmt_iterator gsi = gsi_last_bb (bb);
2595 gsi_prev (&gsi);
2596 e1 = split_pbb (scop, pbb, bb, gsi_stmt (gsi));
2597 }
2598
2599 /* A part of the data references will end in a different basic block
2600 after the split: move the DRs from the original GBB to the newly
2601 created GBB1. */
2602 FOR_EACH_VEC_ELT (data_reference_p, GBB_DATA_REFS (gbb), i, dr)
2603 {
2604 basic_block bb1 = gimple_bb (DR_STMT (dr));
2605
2606 if (bb1 != bb)
2607 {
2608 gimple_bb_p gbb1 = gbb_from_bb (bb1);
2609 VEC_safe_push (data_reference_p, heap, GBB_DATA_REFS (gbb1), dr);
2610 VEC_ordered_remove (data_reference_p, GBB_DATA_REFS (gbb), i);
2611 i--;
2612 }
2613 }
2614
2615 return e1->dest;
2616 }
2617
2618 /* Return true when stmt is a reduction operation. */
2619
2620 static inline bool
2621 is_reduction_operation_p (gimple stmt)
2622 {
2623 enum tree_code code;
2624
2625 gcc_assert (is_gimple_assign (stmt));
2626 code = gimple_assign_rhs_code (stmt);
2627
2628 return flag_associative_math
2629 && commutative_tree_code (code)
2630 && associative_tree_code (code);
2631 }
2632
2633 /* Returns true when PHI contains an argument ARG. */
2634
2635 static bool
2636 phi_contains_arg (gimple phi, tree arg)
2637 {
2638 size_t i;
2639
2640 for (i = 0; i < gimple_phi_num_args (phi); i++)
2641 if (operand_equal_p (arg, gimple_phi_arg_def (phi, i), 0))
2642 return true;
2643
2644 return false;
2645 }
2646
2647 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2648
2649 static gimple
2650 follow_ssa_with_commutative_ops (tree arg, tree lhs)
2651 {
2652 gimple stmt;
2653
2654 if (TREE_CODE (arg) != SSA_NAME)
2655 return NULL;
2656
2657 stmt = SSA_NAME_DEF_STMT (arg);
2658
2659 if (gimple_code (stmt) == GIMPLE_NOP
2660 || gimple_code (stmt) == GIMPLE_CALL)
2661 return NULL;
2662
2663 if (gimple_code (stmt) == GIMPLE_PHI)
2664 {
2665 if (phi_contains_arg (stmt, lhs))
2666 return stmt;
2667 return NULL;
2668 }
2669
2670 if (!is_gimple_assign (stmt))
2671 return NULL;
2672
2673 if (gimple_num_ops (stmt) == 2)
2674 return follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2675
2676 if (is_reduction_operation_p (stmt))
2677 {
2678 gimple res = follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2679
2680 return res ? res :
2681 follow_ssa_with_commutative_ops (gimple_assign_rhs2 (stmt), lhs);
2682 }
2683
2684 return NULL;
2685 }
2686
2687 /* Detect commutative and associative scalar reductions starting at
2688 the STMT. Return the phi node of the reduction cycle, or NULL. */
2689
2690 static gimple
2691 detect_commutative_reduction_arg (tree lhs, gimple stmt, tree arg,
2692 VEC (gimple, heap) **in,
2693 VEC (gimple, heap) **out)
2694 {
2695 gimple phi = follow_ssa_with_commutative_ops (arg, lhs);
2696
2697 if (!phi)
2698 return NULL;
2699
2700 VEC_safe_push (gimple, heap, *in, stmt);
2701 VEC_safe_push (gimple, heap, *out, stmt);
2702 return phi;
2703 }
2704
2705 /* Detect commutative and associative scalar reductions starting at
2706 STMT. Return the phi node of the reduction cycle, or NULL. */
2707
2708 static gimple
2709 detect_commutative_reduction_assign (gimple stmt, VEC (gimple, heap) **in,
2710 VEC (gimple, heap) **out)
2711 {
2712 tree lhs = gimple_assign_lhs (stmt);
2713
2714 if (gimple_num_ops (stmt) == 2)
2715 return detect_commutative_reduction_arg (lhs, stmt,
2716 gimple_assign_rhs1 (stmt),
2717 in, out);
2718
2719 if (is_reduction_operation_p (stmt))
2720 {
2721 gimple res = detect_commutative_reduction_arg (lhs, stmt,
2722 gimple_assign_rhs1 (stmt),
2723 in, out);
2724 return res ? res
2725 : detect_commutative_reduction_arg (lhs, stmt,
2726 gimple_assign_rhs2 (stmt),
2727 in, out);
2728 }
2729
2730 return NULL;
2731 }
2732
2733 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2734
2735 static gimple
2736 follow_inital_value_to_phi (tree arg, tree lhs)
2737 {
2738 gimple stmt;
2739
2740 if (!arg || TREE_CODE (arg) != SSA_NAME)
2741 return NULL;
2742
2743 stmt = SSA_NAME_DEF_STMT (arg);
2744
2745 if (gimple_code (stmt) == GIMPLE_PHI
2746 && phi_contains_arg (stmt, lhs))
2747 return stmt;
2748
2749 return NULL;
2750 }
2751
2752
2753 /* Return the argument of the loop PHI that is the initial value coming
2754 from outside the loop. */
2755
2756 static edge
2757 edge_initial_value_for_loop_phi (gimple phi)
2758 {
2759 size_t i;
2760
2761 for (i = 0; i < gimple_phi_num_args (phi); i++)
2762 {
2763 edge e = gimple_phi_arg_edge (phi, i);
2764
2765 if (loop_depth (e->src->loop_father)
2766 < loop_depth (e->dest->loop_father))
2767 return e;
2768 }
2769
2770 return NULL;
2771 }
2772
2773 /* Return the argument of the loop PHI that is the initial value coming
2774 from outside the loop. */
2775
2776 static tree
2777 initial_value_for_loop_phi (gimple phi)
2778 {
2779 size_t i;
2780
2781 for (i = 0; i < gimple_phi_num_args (phi); i++)
2782 {
2783 edge e = gimple_phi_arg_edge (phi, i);
2784
2785 if (loop_depth (e->src->loop_father)
2786 < loop_depth (e->dest->loop_father))
2787 return gimple_phi_arg_def (phi, i);
2788 }
2789
2790 return NULL_TREE;
2791 }
2792
2793 /* Returns true when DEF is used outside the reduction cycle of
2794 LOOP_PHI. */
2795
2796 static bool
2797 used_outside_reduction (tree def, gimple loop_phi)
2798 {
2799 use_operand_p use_p;
2800 imm_use_iterator imm_iter;
2801 loop_p loop = loop_containing_stmt (loop_phi);
2802
2803 /* In LOOP, DEF should be used only in LOOP_PHI. */
2804 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2805 {
2806 gimple stmt = USE_STMT (use_p);
2807
2808 if (stmt != loop_phi
2809 && !is_gimple_debug (stmt)
2810 && flow_bb_inside_loop_p (loop, gimple_bb (stmt)))
2811 return true;
2812 }
2813
2814 return false;
2815 }
2816
2817 /* Detect commutative and associative scalar reductions belonging to
2818 the SCOP starting at the loop closed phi node STMT. Return the phi
2819 node of the reduction cycle, or NULL. */
2820
2821 static gimple
2822 detect_commutative_reduction (scop_p scop, gimple stmt, VEC (gimple, heap) **in,
2823 VEC (gimple, heap) **out)
2824 {
2825 if (scalar_close_phi_node_p (stmt))
2826 {
2827 gimple def, loop_phi, phi, close_phi = stmt;
2828 tree init, lhs, arg = gimple_phi_arg_def (close_phi, 0);
2829
2830 if (TREE_CODE (arg) != SSA_NAME)
2831 return NULL;
2832
2833 /* Note that loop close phi nodes should have a single argument
2834 because we translated the representation into a canonical form
2835 before Graphite: see canonicalize_loop_closed_ssa_form. */
2836 gcc_assert (gimple_phi_num_args (close_phi) == 1);
2837
2838 def = SSA_NAME_DEF_STMT (arg);
2839 if (!stmt_in_sese_p (def, SCOP_REGION (scop))
2840 || !(loop_phi = detect_commutative_reduction (scop, def, in, out)))
2841 return NULL;
2842
2843 lhs = gimple_phi_result (close_phi);
2844 init = initial_value_for_loop_phi (loop_phi);
2845 phi = follow_inital_value_to_phi (init, lhs);
2846
2847 if (phi && (used_outside_reduction (lhs, phi)
2848 || !has_single_use (gimple_phi_result (phi))))
2849 return NULL;
2850
2851 VEC_safe_push (gimple, heap, *in, loop_phi);
2852 VEC_safe_push (gimple, heap, *out, close_phi);
2853 return phi;
2854 }
2855
2856 if (gimple_code (stmt) == GIMPLE_ASSIGN)
2857 return detect_commutative_reduction_assign (stmt, in, out);
2858
2859 return NULL;
2860 }
2861
2862 /* Translate the scalar reduction statement STMT to an array RED
2863 knowing that its recursive phi node is LOOP_PHI. */
2864
2865 static void
2866 translate_scalar_reduction_to_array_for_stmt (scop_p scop, tree red,
2867 gimple stmt, gimple loop_phi)
2868 {
2869 tree res = gimple_phi_result (loop_phi);
2870 gimple assign = gimple_build_assign (res, unshare_expr (red));
2871 gimple_stmt_iterator gsi;
2872
2873 insert_stmts (scop, assign, NULL, gsi_after_labels (gimple_bb (loop_phi)));
2874
2875 assign = gimple_build_assign (unshare_expr (red), gimple_assign_lhs (stmt));
2876 gsi = gsi_for_stmt (stmt);
2877 gsi_next (&gsi);
2878 insert_stmts (scop, assign, NULL, gsi);
2879 }
2880
2881 /* Removes the PHI node and resets all the debug stmts that are using
2882 the PHI_RESULT. */
2883
2884 static void
2885 remove_phi (gimple phi)
2886 {
2887 imm_use_iterator imm_iter;
2888 tree def;
2889 use_operand_p use_p;
2890 gimple_stmt_iterator gsi;
2891 VEC (gimple, heap) *update = VEC_alloc (gimple, heap, 3);
2892 unsigned int i;
2893 gimple stmt;
2894
2895 def = PHI_RESULT (phi);
2896 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2897 {
2898 stmt = USE_STMT (use_p);
2899
2900 if (is_gimple_debug (stmt))
2901 {
2902 gimple_debug_bind_reset_value (stmt);
2903 VEC_safe_push (gimple, heap, update, stmt);
2904 }
2905 }
2906
2907 FOR_EACH_VEC_ELT (gimple, update, i, stmt)
2908 update_stmt (stmt);
2909
2910 VEC_free (gimple, heap, update);
2911
2912 gsi = gsi_for_phi_node (phi);
2913 remove_phi_node (&gsi, false);
2914 }
2915
2916 /* Helper function for for_each_index. For each INDEX of the data
2917 reference REF, returns true when its indices are valid in the loop
2918 nest LOOP passed in as DATA. */
2919
2920 static bool
2921 dr_indices_valid_in_loop (tree ref ATTRIBUTE_UNUSED, tree *index, void *data)
2922 {
2923 loop_p loop;
2924 basic_block header, def_bb;
2925 gimple stmt;
2926
2927 if (TREE_CODE (*index) != SSA_NAME)
2928 return true;
2929
2930 loop = *((loop_p *) data);
2931 header = loop->header;
2932 stmt = SSA_NAME_DEF_STMT (*index);
2933
2934 if (!stmt)
2935 return true;
2936
2937 def_bb = gimple_bb (stmt);
2938
2939 if (!def_bb)
2940 return true;
2941
2942 return dominated_by_p (CDI_DOMINATORS, header, def_bb);
2943 }
2944
2945 /* When the result of a CLOSE_PHI is written to a memory location,
2946 return a pointer to that memory reference, otherwise return
2947 NULL_TREE. */
2948
2949 static tree
2950 close_phi_written_to_memory (gimple close_phi)
2951 {
2952 imm_use_iterator imm_iter;
2953 use_operand_p use_p;
2954 gimple stmt;
2955 tree res, def = gimple_phi_result (close_phi);
2956
2957 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2958 if ((stmt = USE_STMT (use_p))
2959 && gimple_code (stmt) == GIMPLE_ASSIGN
2960 && (res = gimple_assign_lhs (stmt)))
2961 {
2962 switch (TREE_CODE (res))
2963 {
2964 case VAR_DECL:
2965 case PARM_DECL:
2966 case RESULT_DECL:
2967 return res;
2968
2969 case ARRAY_REF:
2970 case MEM_REF:
2971 {
2972 tree arg = gimple_phi_arg_def (close_phi, 0);
2973 loop_p nest = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2974
2975 /* FIXME: this restriction is for id-{24,25}.f and
2976 could be handled by duplicating the computation of
2977 array indices before the loop of the close_phi. */
2978 if (for_each_index (&res, dr_indices_valid_in_loop, &nest))
2979 return res;
2980 }
2981 /* Fallthru. */
2982
2983 default:
2984 continue;
2985 }
2986 }
2987 return NULL_TREE;
2988 }
2989
2990 /* Rewrite out of SSA the reduction described by the loop phi nodes
2991 IN, and the close phi nodes OUT. IN and OUT are structured by loop
2992 levels like this:
2993
2994 IN: stmt, loop_n, ..., loop_0
2995 OUT: stmt, close_n, ..., close_0
2996
2997 the first element is the reduction statement, and the next elements
2998 are the loop and close phi nodes of each of the outer loops. */
2999
3000 static void
3001 translate_scalar_reduction_to_array (scop_p scop,
3002 VEC (gimple, heap) *in,
3003 VEC (gimple, heap) *out)
3004 {
3005 gimple loop_phi;
3006 unsigned int i = VEC_length (gimple, out) - 1;
3007 tree red = close_phi_written_to_memory (VEC_index (gimple, out, i));
3008
3009 FOR_EACH_VEC_ELT (gimple, in, i, loop_phi)
3010 {
3011 gimple close_phi = VEC_index (gimple, out, i);
3012
3013 if (i == 0)
3014 {
3015 gimple stmt = loop_phi;
3016 basic_block bb = split_reduction_stmt (scop, stmt);
3017 poly_bb_p pbb = pbb_from_bb (bb);
3018 PBB_IS_REDUCTION (pbb) = true;
3019 gcc_assert (close_phi == loop_phi);
3020
3021 if (!red)
3022 red = create_zero_dim_array
3023 (gimple_assign_lhs (stmt), "Commutative_Associative_Reduction");
3024
3025 translate_scalar_reduction_to_array_for_stmt
3026 (scop, red, stmt, VEC_index (gimple, in, 1));
3027 continue;
3028 }
3029
3030 if (i == VEC_length (gimple, in) - 1)
3031 {
3032 insert_out_of_ssa_copy (scop, gimple_phi_result (close_phi),
3033 unshare_expr (red), close_phi);
3034 insert_out_of_ssa_copy_on_edge
3035 (scop, edge_initial_value_for_loop_phi (loop_phi),
3036 unshare_expr (red), initial_value_for_loop_phi (loop_phi));
3037 }
3038
3039 remove_phi (loop_phi);
3040 remove_phi (close_phi);
3041 }
3042 }
3043
3044 /* Rewrites out of SSA a commutative reduction at CLOSE_PHI. Returns
3045 true when something has been changed. */
3046
3047 static bool
3048 rewrite_commutative_reductions_out_of_ssa_close_phi (scop_p scop,
3049 gimple close_phi)
3050 {
3051 bool res;
3052 VEC (gimple, heap) *in = VEC_alloc (gimple, heap, 10);
3053 VEC (gimple, heap) *out = VEC_alloc (gimple, heap, 10);
3054
3055 detect_commutative_reduction (scop, close_phi, &in, &out);
3056 res = VEC_length (gimple, in) > 1;
3057 if (res)
3058 translate_scalar_reduction_to_array (scop, in, out);
3059
3060 VEC_free (gimple, heap, in);
3061 VEC_free (gimple, heap, out);
3062 return res;
3063 }
3064
3065 /* Rewrites all the commutative reductions from LOOP out of SSA.
3066 Returns true when something has been changed. */
3067
3068 static bool
3069 rewrite_commutative_reductions_out_of_ssa_loop (scop_p scop,
3070 loop_p loop)
3071 {
3072 gimple_stmt_iterator gsi;
3073 edge exit = single_exit (loop);
3074 tree res;
3075 bool changed = false;
3076
3077 if (!exit)
3078 return false;
3079
3080 for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi))
3081 if ((res = gimple_phi_result (gsi_stmt (gsi)))
3082 && is_gimple_reg (res)
3083 && !scev_analyzable_p (res, SCOP_REGION (scop)))
3084 changed |= rewrite_commutative_reductions_out_of_ssa_close_phi
3085 (scop, gsi_stmt (gsi));
3086
3087 return changed;
3088 }
3089
3090 /* Rewrites all the commutative reductions from SCOP out of SSA. */
3091
3092 static void
3093 rewrite_commutative_reductions_out_of_ssa (scop_p scop)
3094 {
3095 loop_iterator li;
3096 loop_p loop;
3097 bool changed = false;
3098 sese region = SCOP_REGION (scop);
3099
3100 FOR_EACH_LOOP (li, loop, 0)
3101 if (loop_in_sese_p (loop, region))
3102 changed |= rewrite_commutative_reductions_out_of_ssa_loop (scop, loop);
3103
3104 if (changed)
3105 {
3106 scev_reset_htab ();
3107 gsi_commit_edge_inserts ();
3108 update_ssa (TODO_update_ssa);
3109 #ifdef ENABLE_CHECKING
3110 verify_loop_closed_ssa (true);
3111 #endif
3112 }
3113 }
3114
3115 /* Can all ivs be represented by a signed integer?
3116 As CLooG might generate negative values in its expressions, signed loop ivs
3117 are required in the backend. */
3118
3119 static bool
3120 scop_ivs_can_be_represented (scop_p scop)
3121 {
3122 loop_iterator li;
3123 loop_p loop;
3124 gimple_stmt_iterator psi;
3125
3126 FOR_EACH_LOOP (li, loop, 0)
3127 {
3128 if (!loop_in_sese_p (loop, SCOP_REGION (scop)))
3129 continue;
3130
3131 for (psi = gsi_start_phis (loop->header);
3132 !gsi_end_p (psi); gsi_next (&psi))
3133 {
3134 gimple phi = gsi_stmt (psi);
3135 tree res = PHI_RESULT (phi);
3136 tree type = TREE_TYPE (res);
3137
3138 if (TYPE_UNSIGNED (type)
3139 && TYPE_PRECISION (type) >= TYPE_PRECISION (long_long_integer_type_node))
3140 return false;
3141 }
3142 }
3143
3144 return true;
3145 }
3146
3147 /* Builds the polyhedral representation for a SESE region. */
3148
3149 void
3150 build_poly_scop (scop_p scop)
3151 {
3152 sese region = SCOP_REGION (scop);
3153 graphite_dim_t max_dim;
3154
3155 build_scop_bbs (scop);
3156
3157 /* FIXME: This restriction is needed to avoid a problem in CLooG.
3158 Once CLooG is fixed, remove this guard. Anyways, it makes no
3159 sense to optimize a scop containing only PBBs that do not belong
3160 to any loops. */
3161 if (nb_pbbs_in_loops (scop) == 0)
3162 return;
3163
3164 if (!scop_ivs_can_be_represented (scop))
3165 return;
3166
3167 if (flag_associative_math)
3168 rewrite_commutative_reductions_out_of_ssa (scop);
3169
3170 build_sese_loop_nests (region);
3171 build_sese_conditions (region);
3172 find_scop_parameters (scop);
3173
3174 max_dim = PARAM_VALUE (PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS);
3175 if (scop_nb_params (scop) > max_dim)
3176 return;
3177
3178 build_scop_iteration_domain (scop);
3179 build_scop_context (scop);
3180 add_conditions_to_constraints (scop);
3181
3182 /* Rewrite out of SSA only after having translated the
3183 representation to the polyhedral representation to avoid scev
3184 analysis failures. That means that these functions will insert
3185 new data references that they create in the right place. */
3186 rewrite_reductions_out_of_ssa (scop);
3187 rewrite_cross_bb_scalar_deps_out_of_ssa (scop);
3188
3189 build_scop_drs (scop);
3190 scop_to_lst (scop);
3191 build_scop_scattering (scop);
3192
3193 /* This SCoP has been translated to the polyhedral
3194 representation. */
3195 POLY_SCOP_P (scop) = true;
3196 }
3197 #endif