2 * Copyright 2005-2007 Universiteit Leiden
3 * Copyright 2008-2009 Katholieke Universiteit Leuven
4 * Copyright 2010 INRIA Saclay
6 * Use of this software is governed by the GNU LGPLv2.1 license
8 * Written by Sven Verdoolaege, Leiden Institute of Advanced Computer Science,
9 * Universiteit Leiden, Niels Bohrweg 1, 2333 CA Leiden, The Netherlands
10 * and K.U.Leuven, Departement Computerwetenschappen, Celestijnenlaan 200A,
11 * B-3001 Leuven, Belgium
12 * and INRIA Saclay - Ile-de-France, Parc Club Orsay Universite,
13 * ZAC des vignes, 4 rue Jacques Monod, 91893 Orsay, France
19 #include <isl_qsort.h>
21 /* A private structure to keep track of a mapping together with
22 * a user-specified identifier and a boolean indicating whether
23 * the map represents a must or may access/dependence.
25 struct isl_labeled_map {
31 /* A structure containing the input for dependence analysis:
33 * - n_must + n_may (<= max_source) sources
34 * - a function for determining the relative order of sources and sink
35 * The must sources are placed before the may sources.
37 * domain_map is an auxiliary map that maps the sink access relation
38 * to the domain of this access relation.
40 * restrict_sources is a callback that (if not NULL) will be called
41 * right before any lexicographical maximization.
43 struct isl_access_info {
45 struct isl_labeled_map sink;
46 isl_access_level_before level_before;
47 isl_access_restrict_sources restrict_sources;
51 struct isl_labeled_map source[1];
54 /* A structure containing the output of dependence analysis:
55 * - n_source dependences
56 * - a wrapped subset of the sink for which definitely no source could be found
57 * - a wrapped subset of the sink for which possibly no source could be found
60 isl_set *must_no_source;
61 isl_set *may_no_source;
63 struct isl_labeled_map *dep;
66 /* Construct an isl_access_info structure and fill it up with
67 * the given data. The number of sources is set to 0.
69 __isl_give isl_access_info *isl_access_info_alloc(__isl_take isl_map *sink,
70 void *sink_user, isl_access_level_before fn, int max_source)
73 struct isl_access_info *acc;
78 ctx = isl_map_get_ctx(sink);
79 isl_assert(ctx, max_source >= 0, goto error);
81 acc = isl_calloc(ctx, struct isl_access_info,
82 sizeof(struct isl_access_info) +
83 (max_source - 1) * sizeof(struct isl_labeled_map));
88 acc->sink.data = sink_user;
89 acc->level_before = fn;
90 acc->max_source = max_source;
100 /* Free the given isl_access_info structure.
102 void isl_access_info_free(__isl_take isl_access_info *acc)
108 isl_map_free(acc->domain_map);
109 isl_map_free(acc->sink.map);
110 for (i = 0; i < acc->n_must + acc->n_may; ++i)
111 isl_map_free(acc->source[i].map);
115 isl_ctx *isl_access_info_get_ctx(__isl_keep isl_access_info *acc)
117 return acc ? isl_map_get_ctx(acc->sink.map) : NULL;
120 __isl_give isl_access_info *isl_access_info_set_restrict_sources(
121 __isl_take isl_access_info *acc, isl_access_restrict_sources fn)
125 acc->restrict_sources = fn;
129 /* Add another source to an isl_access_info structure, making
130 * sure the "must" sources are placed before the "may" sources.
131 * This function may be called at most max_source times on a
132 * given isl_access_info structure, with max_source as specified
133 * in the call to isl_access_info_alloc that constructed the structure.
135 __isl_give isl_access_info *isl_access_info_add_source(
136 __isl_take isl_access_info *acc, __isl_take isl_map *source,
137 int must, void *source_user)
143 ctx = isl_map_get_ctx(acc->sink.map);
144 isl_assert(ctx, acc->n_must + acc->n_may < acc->max_source, goto error);
148 acc->source[acc->n_must + acc->n_may] =
149 acc->source[acc->n_must];
150 acc->source[acc->n_must].map = source;
151 acc->source[acc->n_must].data = source_user;
152 acc->source[acc->n_must].must = 1;
155 acc->source[acc->n_must + acc->n_may].map = source;
156 acc->source[acc->n_must + acc->n_may].data = source_user;
157 acc->source[acc->n_must + acc->n_may].must = 0;
163 isl_map_free(source);
164 isl_access_info_free(acc);
168 /* Return -n, 0 or n (with n a positive value), depending on whether
169 * the source access identified by p1 should be sorted before, together
170 * or after that identified by p2.
172 * If p1 and p2 share a different number of levels with the sink,
173 * then the one with the lowest number of shared levels should be
175 * If they both share no levels, then the order is irrelevant.
176 * Otherwise, if p1 appears before p2, then it should be sorted first.
177 * For more generic initial schedules, it is possible that neither
178 * p1 nor p2 appears before the other, or at least not in any obvious way.
179 * We therefore also check if p2 appears before p1, in which case p2
180 * should be sorted first.
181 * If not, we try to order the two statements based on the description
182 * of the iteration domains. This results in an arbitrary, but fairly
185 static int access_sort_cmp(const void *p1, const void *p2, void *user)
187 isl_access_info *acc = user;
188 const struct isl_labeled_map *i1, *i2;
191 i1 = (const struct isl_labeled_map *) p1;
192 i2 = (const struct isl_labeled_map *) p2;
194 level1 = acc->level_before(i1->data, acc->sink.data);
195 level2 = acc->level_before(i2->data, acc->sink.data);
197 if (level1 != level2 || !level1)
198 return level1 - level2;
200 level1 = acc->level_before(i1->data, i2->data);
204 level2 = acc->level_before(i2->data, i1->data);
208 h1 = isl_map_get_hash(i1->map);
209 h2 = isl_map_get_hash(i2->map);
210 return h1 > h2 ? 1 : h1 < h2 ? -1 : 0;
213 /* Sort the must source accesses in order of increasing number of shared
214 * levels with the sink access.
215 * Source accesses with the same number of shared levels are sorted
216 * in their textual order.
218 static __isl_give isl_access_info *isl_access_info_sort_sources(
219 __isl_take isl_access_info *acc)
223 if (acc->n_must <= 1)
226 isl_quicksort(acc->source, acc->n_must, sizeof(struct isl_labeled_map),
227 access_sort_cmp, acc);
232 /* Align the parameters of the two spaces if needed and then call
235 static __isl_give isl_space *space_align_and_join(__isl_take isl_space *left,
236 __isl_take isl_space *right)
238 if (isl_space_match(left, isl_dim_param, right, isl_dim_param))
239 return isl_space_join(left, right);
241 left = isl_space_align_params(left, isl_space_copy(right));
242 right = isl_space_align_params(right, isl_space_copy(left));
243 return isl_space_join(left, right);
246 /* Initialize an empty isl_flow structure corresponding to a given
247 * isl_access_info structure.
248 * For each must access, two dependences are created (initialized
249 * to the empty relation), one for the resulting must dependences
250 * and one for the resulting may dependences. May accesses can
251 * only lead to may dependences, so only one dependence is created
253 * This function is private as isl_flow structures are only supposed
254 * to be created by isl_access_info_compute_flow.
256 static __isl_give isl_flow *isl_flow_alloc(__isl_keep isl_access_info *acc)
260 struct isl_flow *dep;
265 ctx = isl_map_get_ctx(acc->sink.map);
266 dep = isl_calloc_type(ctx, struct isl_flow);
270 dep->dep = isl_calloc_array(ctx, struct isl_labeled_map,
271 2 * acc->n_must + acc->n_may);
275 dep->n_source = 2 * acc->n_must + acc->n_may;
276 for (i = 0; i < acc->n_must; ++i) {
278 dim = space_align_and_join(
279 isl_map_get_space(acc->source[i].map),
280 isl_space_reverse(isl_map_get_space(acc->sink.map)));
281 dep->dep[2 * i].map = isl_map_empty(dim);
282 dep->dep[2 * i + 1].map = isl_map_copy(dep->dep[2 * i].map);
283 dep->dep[2 * i].data = acc->source[i].data;
284 dep->dep[2 * i + 1].data = acc->source[i].data;
285 dep->dep[2 * i].must = 1;
286 dep->dep[2 * i + 1].must = 0;
287 if (!dep->dep[2 * i].map || !dep->dep[2 * i + 1].map)
290 for (i = acc->n_must; i < acc->n_must + acc->n_may; ++i) {
292 dim = space_align_and_join(
293 isl_map_get_space(acc->source[i].map),
294 isl_space_reverse(isl_map_get_space(acc->sink.map)));
295 dep->dep[acc->n_must + i].map = isl_map_empty(dim);
296 dep->dep[acc->n_must + i].data = acc->source[i].data;
297 dep->dep[acc->n_must + i].must = 0;
298 if (!dep->dep[acc->n_must + i].map)
308 /* Iterate over all sources and for each resulting flow dependence
309 * that is not empty, call the user specfied function.
310 * The second argument in this function call identifies the source,
311 * while the third argument correspond to the final argument of
312 * the isl_flow_foreach call.
314 int isl_flow_foreach(__isl_keep isl_flow *deps,
315 int (*fn)(__isl_take isl_map *dep, int must, void *dep_user, void *user),
323 for (i = 0; i < deps->n_source; ++i) {
324 if (isl_map_plain_is_empty(deps->dep[i].map))
326 if (fn(isl_map_copy(deps->dep[i].map), deps->dep[i].must,
327 deps->dep[i].data, user) < 0)
334 /* Return a copy of the subset of the sink for which no source could be found.
336 __isl_give isl_map *isl_flow_get_no_source(__isl_keep isl_flow *deps, int must)
342 return isl_set_unwrap(isl_set_copy(deps->must_no_source));
344 return isl_set_unwrap(isl_set_copy(deps->may_no_source));
347 void isl_flow_free(__isl_take isl_flow *deps)
353 isl_set_free(deps->must_no_source);
354 isl_set_free(deps->may_no_source);
356 for (i = 0; i < deps->n_source; ++i)
357 isl_map_free(deps->dep[i].map);
363 isl_ctx *isl_flow_get_ctx(__isl_keep isl_flow *deps)
365 return deps ? isl_set_get_ctx(deps->must_no_source) : NULL;
368 /* Return a map that enforces that the domain iteration occurs after
369 * the range iteration at the given level.
370 * If level is odd, then the domain iteration should occur after
371 * the target iteration in their shared level/2 outermost loops.
372 * In this case we simply need to enforce that these outermost
373 * loop iterations are the same.
374 * If level is even, then the loop iterator of the domain should
375 * be greater than the loop iterator of the range at the last
376 * of the level/2 shared loops, i.e., loop level/2 - 1.
378 static __isl_give isl_map *after_at_level(__isl_take isl_space *dim, int level)
380 struct isl_basic_map *bmap;
383 bmap = isl_basic_map_equal(dim, level/2);
385 bmap = isl_basic_map_more_at(dim, level/2 - 1);
387 return isl_map_from_basic_map(bmap);
390 /* Check if the user has set acc->restrict_sources and if so
391 * intersect the range of "dep" with the result of a call to this function.
393 * Since the user expects a mapping from sink iterations to source iterations,
394 * whereas the domain of "dep" is a wrapped map, mapping sink iterations
395 * to accessed array elements, we first need to project out the accessed
396 * sink array elements by applying acc->domain_map.
398 static __isl_give isl_map *restrict_sources(__isl_take isl_map *dep,
399 struct isl_access_info *acc, int source)
404 if (!acc->restrict_sources)
407 source_map = isl_map_copy(dep);
408 source_map = isl_map_apply_domain(source_map,
409 isl_map_copy(acc->domain_map));
410 param = acc->restrict_sources(source_map, acc->sink.data,
411 acc->source[source].data);
412 dep = isl_map_intersect_range(dep, param);
416 /* Compute the last iteration of must source j that precedes the sink
417 * at the given level for sink iterations in set_C.
418 * The subset of set_C for which no such iteration can be found is returned
421 static struct isl_map *last_source(struct isl_access_info *acc,
422 struct isl_set *set_C,
423 int j, int level, struct isl_set **empty)
425 struct isl_map *read_map;
426 struct isl_map *write_map;
427 struct isl_map *dep_map;
428 struct isl_map *after;
429 struct isl_map *result;
431 read_map = isl_map_copy(acc->sink.map);
432 write_map = isl_map_copy(acc->source[j].map);
433 write_map = isl_map_reverse(write_map);
434 dep_map = isl_map_apply_range(read_map, write_map);
435 after = after_at_level(isl_map_get_space(dep_map), level);
436 dep_map = isl_map_intersect(dep_map, after);
437 dep_map = restrict_sources(dep_map, acc, j);
438 result = isl_map_partial_lexmax(dep_map, set_C, empty);
439 result = isl_map_reverse(result);
444 /* For a given mapping between iterations of must source j and iterations
445 * of the sink, compute the last iteration of must source k preceding
446 * the sink at level before_level for any of the sink iterations,
447 * but following the corresponding iteration of must source j at level
450 static struct isl_map *last_later_source(struct isl_access_info *acc,
451 struct isl_map *old_map,
452 int j, int before_level,
453 int k, int after_level,
454 struct isl_set **empty)
457 struct isl_set *set_C;
458 struct isl_map *read_map;
459 struct isl_map *write_map;
460 struct isl_map *dep_map;
461 struct isl_map *after_write;
462 struct isl_map *before_read;
463 struct isl_map *result;
465 set_C = isl_map_range(isl_map_copy(old_map));
466 read_map = isl_map_copy(acc->sink.map);
467 write_map = isl_map_copy(acc->source[k].map);
469 write_map = isl_map_reverse(write_map);
470 dep_map = isl_map_apply_range(read_map, write_map);
471 dim = space_align_and_join(isl_map_get_space(acc->source[k].map),
472 isl_space_reverse(isl_map_get_space(acc->source[j].map)));
473 after_write = after_at_level(dim, after_level);
474 after_write = isl_map_apply_range(after_write, old_map);
475 after_write = isl_map_reverse(after_write);
476 dep_map = isl_map_intersect(dep_map, after_write);
477 before_read = after_at_level(isl_map_get_space(dep_map), before_level);
478 dep_map = isl_map_intersect(dep_map, before_read);
479 dep_map = restrict_sources(dep_map, acc, k);
480 result = isl_map_partial_lexmax(dep_map, set_C, empty);
481 result = isl_map_reverse(result);
486 /* Given a shared_level between two accesses, return 1 if the
487 * the first can precede the second at the requested target_level.
488 * If the target level is odd, i.e., refers to a statement level
489 * dimension, then first needs to precede second at the requested
490 * level, i.e., shared_level must be equal to target_level.
491 * If the target level is odd, then the two loops should share
492 * at least the requested number of outer loops.
494 static int can_precede_at_level(int shared_level, int target_level)
496 if (shared_level < target_level)
498 if ((target_level % 2) && shared_level > target_level)
503 /* Given a possible flow dependence temp_rel[j] between source j and the sink
504 * at level sink_level, remove those elements for which
505 * there is an iteration of another source k < j that is closer to the sink.
506 * The flow dependences temp_rel[k] are updated with the improved sources.
507 * Any improved source needs to precede the sink at the same level
508 * and needs to follow source j at the same or a deeper level.
509 * The lower this level, the later the execution date of source k.
510 * We therefore consider lower levels first.
512 * If temp_rel[j] is empty, then there can be no improvement and
513 * we return immediately.
515 static int intermediate_sources(__isl_keep isl_access_info *acc,
516 struct isl_map **temp_rel, int j, int sink_level)
519 int depth = 2 * isl_map_dim(acc->source[j].map, isl_dim_in) + 1;
521 if (isl_map_plain_is_empty(temp_rel[j]))
524 for (k = j - 1; k >= 0; --k) {
526 plevel = acc->level_before(acc->source[k].data, acc->sink.data);
527 if (!can_precede_at_level(plevel, sink_level))
530 plevel2 = acc->level_before(acc->source[j].data,
531 acc->source[k].data);
533 for (level = sink_level; level <= depth; ++level) {
535 struct isl_set *trest;
536 struct isl_map *copy;
538 if (!can_precede_at_level(plevel2, level))
541 copy = isl_map_copy(temp_rel[j]);
542 T = last_later_source(acc, copy, j, sink_level, k,
544 if (isl_map_plain_is_empty(T)) {
549 temp_rel[j] = isl_map_intersect_range(temp_rel[j], trest);
550 temp_rel[k] = isl_map_union_disjoint(temp_rel[k], T);
557 /* Compute all iterations of may source j that precedes the sink at the given
558 * level for sink iterations in set_C.
560 static __isl_give isl_map *all_sources(__isl_keep isl_access_info *acc,
561 __isl_take isl_set *set_C, int j, int level)
568 read_map = isl_map_copy(acc->sink.map);
569 read_map = isl_map_intersect_domain(read_map, set_C);
570 write_map = isl_map_copy(acc->source[acc->n_must + j].map);
571 write_map = isl_map_reverse(write_map);
572 dep_map = isl_map_apply_range(read_map, write_map);
573 after = after_at_level(isl_map_get_space(dep_map), level);
574 dep_map = isl_map_intersect(dep_map, after);
576 return isl_map_reverse(dep_map);
579 /* For a given mapping between iterations of must source k and iterations
580 * of the sink, compute the all iteration of may source j preceding
581 * the sink at level before_level for any of the sink iterations,
582 * but following the corresponding iteration of must source k at level
585 static __isl_give isl_map *all_later_sources(__isl_keep isl_access_info *acc,
586 __isl_keep isl_map *old_map,
587 int j, int before_level, int k, int after_level)
594 isl_map *after_write;
595 isl_map *before_read;
597 set_C = isl_map_range(isl_map_copy(old_map));
598 read_map = isl_map_copy(acc->sink.map);
599 read_map = isl_map_intersect_domain(read_map, set_C);
600 write_map = isl_map_copy(acc->source[acc->n_must + j].map);
602 write_map = isl_map_reverse(write_map);
603 dep_map = isl_map_apply_range(read_map, write_map);
604 dim = isl_space_join(isl_map_get_space(acc->source[acc->n_must + j].map),
605 isl_space_reverse(isl_map_get_space(acc->source[k].map)));
606 after_write = after_at_level(dim, after_level);
607 after_write = isl_map_apply_range(after_write, old_map);
608 after_write = isl_map_reverse(after_write);
609 dep_map = isl_map_intersect(dep_map, after_write);
610 before_read = after_at_level(isl_map_get_space(dep_map), before_level);
611 dep_map = isl_map_intersect(dep_map, before_read);
612 return isl_map_reverse(dep_map);
615 /* Given the must and may dependence relations for the must accesses
616 * for level sink_level, check if there are any accesses of may access j
617 * that occur in between and return their union.
618 * If some of these accesses are intermediate with respect to
619 * (previously thought to be) must dependences, then these
620 * must dependences are turned into may dependences.
622 static __isl_give isl_map *all_intermediate_sources(
623 __isl_keep isl_access_info *acc, __isl_take isl_map *map,
624 struct isl_map **must_rel, struct isl_map **may_rel,
625 int j, int sink_level)
628 int depth = 2 * isl_map_dim(acc->source[acc->n_must + j].map,
631 for (k = 0; k < acc->n_must; ++k) {
634 if (isl_map_plain_is_empty(may_rel[k]) &&
635 isl_map_plain_is_empty(must_rel[k]))
638 plevel = acc->level_before(acc->source[k].data,
639 acc->source[acc->n_must + j].data);
641 for (level = sink_level; level <= depth; ++level) {
646 if (!can_precede_at_level(plevel, level))
649 copy = isl_map_copy(may_rel[k]);
650 T = all_later_sources(acc, copy, j, sink_level, k, level);
651 map = isl_map_union(map, T);
653 copy = isl_map_copy(must_rel[k]);
654 T = all_later_sources(acc, copy, j, sink_level, k, level);
655 ran = isl_map_range(isl_map_copy(T));
656 map = isl_map_union(map, T);
657 may_rel[k] = isl_map_union_disjoint(may_rel[k],
658 isl_map_intersect_range(isl_map_copy(must_rel[k]),
660 T = isl_map_from_domain_and_range(
662 isl_space_domain(isl_map_get_space(must_rel[k]))),
664 must_rel[k] = isl_map_subtract(must_rel[k], T);
671 /* Compute dependences for the case where all accesses are "may"
672 * accesses, which boils down to computing memory based dependences.
673 * The generic algorithm would also work in this case, but it would
674 * be overkill to use it.
676 static __isl_give isl_flow *compute_mem_based_dependences(
677 __isl_keep isl_access_info *acc)
684 res = isl_flow_alloc(acc);
688 mustdo = isl_map_domain(isl_map_copy(acc->sink.map));
689 maydo = isl_set_copy(mustdo);
691 for (i = 0; i < acc->n_may; ++i) {
698 plevel = acc->level_before(acc->source[i].data, acc->sink.data);
699 is_before = plevel & 1;
702 dim = isl_map_get_space(res->dep[i].map);
704 before = isl_map_lex_le_first(dim, plevel);
706 before = isl_map_lex_lt_first(dim, plevel);
707 dep = isl_map_apply_range(isl_map_copy(acc->source[i].map),
708 isl_map_reverse(isl_map_copy(acc->sink.map)));
709 dep = isl_map_intersect(dep, before);
710 mustdo = isl_set_subtract(mustdo,
711 isl_map_range(isl_map_copy(dep)));
712 res->dep[i].map = isl_map_union(res->dep[i].map, dep);
715 res->may_no_source = isl_set_subtract(maydo, isl_set_copy(mustdo));
716 res->must_no_source = mustdo;
721 /* Compute dependences for the case where there is at least one
724 * The core algorithm considers all levels in which a source may precede
725 * the sink, where a level may either be a statement level or a loop level.
726 * The outermost statement level is 1, the first loop level is 2, etc...
727 * The algorithm basically does the following:
728 * for all levels l of the read access from innermost to outermost
729 * for all sources w that may precede the sink access at that level
730 * compute the last iteration of the source that precedes the sink access
732 * add result to possible last accesses at level l of source w
733 * for all sources w2 that we haven't considered yet at this level that may
734 * also precede the sink access
735 * for all levels l2 of w from l to innermost
736 * for all possible last accesses dep of w at l
737 * compute last iteration of w2 between the source and sink
739 * add result to possible last accesses at level l of write w2
740 * and replace possible last accesses dep by the remainder
743 * The above algorithm is applied to the must access. During the course
744 * of the algorithm, we keep track of sink iterations that still
745 * need to be considered. These iterations are split into those that
746 * haven't been matched to any source access (mustdo) and those that have only
747 * been matched to may accesses (maydo).
748 * At the end of each level, we also consider the may accesses.
749 * In particular, we consider may accesses that precede the remaining
750 * sink iterations, moving elements from mustdo to maydo when appropriate,
751 * and may accesses that occur between a must source and a sink of any
752 * dependences found at the current level, turning must dependences into
753 * may dependences when appropriate.
756 static __isl_give isl_flow *compute_val_based_dependences(
757 __isl_keep isl_access_info *acc)
761 isl_set *mustdo = NULL;
762 isl_set *maydo = NULL;
765 isl_map **must_rel = NULL;
766 isl_map **may_rel = NULL;
771 res = isl_flow_alloc(acc);
774 ctx = isl_map_get_ctx(acc->sink.map);
776 depth = 2 * isl_map_dim(acc->sink.map, isl_dim_in) + 1;
777 mustdo = isl_map_domain(isl_map_copy(acc->sink.map));
778 maydo = isl_set_empty_like(mustdo);
779 if (!mustdo || !maydo)
781 if (isl_set_plain_is_empty(mustdo))
784 must_rel = isl_alloc_array(ctx, struct isl_map *, acc->n_must);
785 may_rel = isl_alloc_array(ctx, struct isl_map *, acc->n_must);
786 if (!must_rel || !may_rel)
789 for (level = depth; level >= 1; --level) {
790 for (j = acc->n_must-1; j >=0; --j) {
791 must_rel[j] = isl_map_empty_like(res->dep[j].map);
792 may_rel[j] = isl_map_copy(must_rel[j]);
795 for (j = acc->n_must - 1; j >= 0; --j) {
797 struct isl_set *rest;
800 plevel = acc->level_before(acc->source[j].data,
802 if (!can_precede_at_level(plevel, level))
805 T = last_source(acc, mustdo, j, level, &rest);
806 must_rel[j] = isl_map_union_disjoint(must_rel[j], T);
809 intermediate_sources(acc, must_rel, j, level);
811 T = last_source(acc, maydo, j, level, &rest);
812 may_rel[j] = isl_map_union_disjoint(may_rel[j], T);
815 intermediate_sources(acc, may_rel, j, level);
817 if (isl_set_plain_is_empty(mustdo) &&
818 isl_set_plain_is_empty(maydo))
821 for (j = j - 1; j >= 0; --j) {
824 plevel = acc->level_before(acc->source[j].data,
826 if (!can_precede_at_level(plevel, level))
829 intermediate_sources(acc, must_rel, j, level);
830 intermediate_sources(acc, may_rel, j, level);
833 for (j = 0; j < acc->n_may; ++j) {
838 plevel = acc->level_before(acc->source[acc->n_must + j].data,
840 if (!can_precede_at_level(plevel, level))
843 T = all_sources(acc, isl_set_copy(maydo), j, level);
844 res->dep[2 * acc->n_must + j].map =
845 isl_map_union(res->dep[2 * acc->n_must + j].map, T);
846 T = all_sources(acc, isl_set_copy(mustdo), j, level);
847 ran = isl_map_range(isl_map_copy(T));
848 res->dep[2 * acc->n_must + j].map =
849 isl_map_union(res->dep[2 * acc->n_must + j].map, T);
850 mustdo = isl_set_subtract(mustdo, isl_set_copy(ran));
851 maydo = isl_set_union_disjoint(maydo, ran);
853 T = res->dep[2 * acc->n_must + j].map;
854 T = all_intermediate_sources(acc, T, must_rel, may_rel,
856 res->dep[2 * acc->n_must + j].map = T;
859 for (j = acc->n_must - 1; j >= 0; --j) {
860 res->dep[2 * j].map =
861 isl_map_union_disjoint(res->dep[2 * j].map,
863 res->dep[2 * j + 1].map =
864 isl_map_union_disjoint(res->dep[2 * j + 1].map,
868 if (isl_set_plain_is_empty(mustdo) &&
869 isl_set_plain_is_empty(maydo))
876 res->must_no_source = mustdo;
877 res->may_no_source = maydo;
881 isl_set_free(mustdo);
888 /* Given a "sink" access, a list of n "source" accesses,
889 * compute for each iteration of the sink access
890 * and for each element accessed by that iteration,
891 * the source access in the list that last accessed the
892 * element accessed by the sink access before this sink access.
893 * Each access is given as a map from the loop iterators
894 * to the array indices.
895 * The result is a list of n relations between source and sink
896 * iterations and a subset of the domain of the sink access,
897 * corresponding to those iterations that access an element
898 * not previously accessed.
900 * To deal with multi-valued sink access relations, the sink iteration
901 * domain is first extended with dimensions that correspond to the data
902 * space. After the computation is finished, these extra dimensions are
903 * projected out again.
905 __isl_give isl_flow *isl_access_info_compute_flow(__isl_take isl_access_info *acc)
908 struct isl_flow *res = NULL;
913 acc->domain_map = isl_map_domain_map(isl_map_copy(acc->sink.map));
914 acc->sink.map = isl_map_range_map(acc->sink.map);
918 if (acc->n_must == 0)
919 res = compute_mem_based_dependences(acc);
921 acc = isl_access_info_sort_sources(acc);
922 res = compute_val_based_dependences(acc);
927 for (j = 0; j < res->n_source; ++j) {
928 res->dep[j].map = isl_map_apply_range(res->dep[j].map,
929 isl_map_copy(acc->domain_map));
930 if (!res->dep[j].map)
933 if (!res->must_no_source || !res->may_no_source)
936 isl_access_info_free(acc);
939 isl_access_info_free(acc);
945 /* Keep track of some information about a schedule for a given
946 * access. In particular, keep track of which dimensions
947 * have a constant value and of the actual constant values.
949 struct isl_sched_info {
954 static void sched_info_free(__isl_take struct isl_sched_info *info)
958 isl_vec_free(info->cst);
963 /* Extract information on the constant dimensions of the schedule
964 * for a given access. The "map" is of the form
968 * with S the schedule domain, D the iteration domain and A the data domain.
970 static __isl_give struct isl_sched_info *sched_info_alloc(
971 __isl_keep isl_map *map)
975 struct isl_sched_info *info;
981 dim = isl_space_unwrap(isl_space_domain(isl_map_get_space(map)));
984 n = isl_space_dim(dim, isl_dim_in);
987 ctx = isl_map_get_ctx(map);
988 info = isl_alloc_type(ctx, struct isl_sched_info);
991 info->is_cst = isl_alloc_array(ctx, int, n);
992 info->cst = isl_vec_alloc(ctx, n);
993 if (!info->is_cst || !info->cst)
996 for (i = 0; i < n; ++i)
997 info->is_cst[i] = isl_map_plain_is_fixed(map, isl_dim_in, i,
1002 sched_info_free(info);
1006 struct isl_compute_flow_data {
1007 isl_union_map *must_source;
1008 isl_union_map *may_source;
1009 isl_union_map *must_dep;
1010 isl_union_map *may_dep;
1011 isl_union_map *must_no_source;
1012 isl_union_map *may_no_source;
1017 struct isl_sched_info *sink_info;
1018 struct isl_sched_info **source_info;
1019 isl_access_info *accesses;
1022 static int count_matching_array(__isl_take isl_map *map, void *user)
1026 struct isl_compute_flow_data *data;
1028 data = (struct isl_compute_flow_data *)user;
1030 dim = isl_space_range(isl_map_get_space(map));
1032 eq = isl_space_is_equal(dim, data->dim);
1034 isl_space_free(dim);
1045 static int collect_matching_array(__isl_take isl_map *map, void *user)
1049 struct isl_sched_info *info;
1050 struct isl_compute_flow_data *data;
1052 data = (struct isl_compute_flow_data *)user;
1054 dim = isl_space_range(isl_map_get_space(map));
1056 eq = isl_space_is_equal(dim, data->dim);
1058 isl_space_free(dim);
1067 info = sched_info_alloc(map);
1068 data->source_info[data->count] = info;
1070 data->accesses = isl_access_info_add_source(data->accesses,
1071 map, data->must, info);
1081 /* Determine the shared nesting level and the "textual order" of
1082 * the given accesses.
1084 * We first determine the minimal schedule dimension for both accesses.
1086 * If among those dimensions, we can find one where both have a fixed
1087 * value and if moreover those values are different, then the previous
1088 * dimension is the last shared nesting level and the textual order
1089 * is determined based on the order of the fixed values.
1090 * If no such fixed values can be found, then we set the shared
1091 * nesting level to the minimal schedule dimension, with no textual ordering.
1093 static int before(void *first, void *second)
1095 struct isl_sched_info *info1 = first;
1096 struct isl_sched_info *info2 = second;
1100 n1 = info1->cst->size;
1101 n2 = info2->cst->size;
1106 for (i = 0; i < n1; ++i) {
1107 if (!info1->is_cst[i])
1109 if (!info2->is_cst[i])
1111 if (isl_int_eq(info1->cst->el[i], info2->cst->el[i]))
1113 return 2 * i + isl_int_lt(info1->cst->el[i], info2->cst->el[i]);
1119 /* Given a sink access, look for all the source accesses that access
1120 * the same array and perform dataflow analysis on them using
1121 * isl_access_info_compute_flow.
1123 static int compute_flow(__isl_take isl_map *map, void *user)
1127 struct isl_compute_flow_data *data;
1130 data = (struct isl_compute_flow_data *)user;
1132 ctx = isl_map_get_ctx(map);
1134 data->accesses = NULL;
1135 data->sink_info = NULL;
1136 data->source_info = NULL;
1138 data->dim = isl_space_range(isl_map_get_space(map));
1140 if (isl_union_map_foreach_map(data->must_source,
1141 &count_matching_array, data) < 0)
1143 if (isl_union_map_foreach_map(data->may_source,
1144 &count_matching_array, data) < 0)
1147 data->sink_info = sched_info_alloc(map);
1148 data->source_info = isl_calloc_array(ctx, struct isl_sched_info *,
1151 data->accesses = isl_access_info_alloc(isl_map_copy(map),
1152 data->sink_info, &before, data->count);
1153 if (!data->sink_info || !data->source_info || !data->accesses)
1157 if (isl_union_map_foreach_map(data->must_source,
1158 &collect_matching_array, data) < 0)
1161 if (isl_union_map_foreach_map(data->may_source,
1162 &collect_matching_array, data) < 0)
1165 flow = isl_access_info_compute_flow(data->accesses);
1166 data->accesses = NULL;
1171 data->must_no_source = isl_union_map_union(data->must_no_source,
1172 isl_union_map_from_map(isl_flow_get_no_source(flow, 1)));
1173 data->may_no_source = isl_union_map_union(data->may_no_source,
1174 isl_union_map_from_map(isl_flow_get_no_source(flow, 0)));
1176 for (i = 0; i < flow->n_source; ++i) {
1178 dep = isl_union_map_from_map(isl_map_copy(flow->dep[i].map));
1179 if (flow->dep[i].must)
1180 data->must_dep = isl_union_map_union(data->must_dep, dep);
1182 data->may_dep = isl_union_map_union(data->may_dep, dep);
1185 isl_flow_free(flow);
1187 sched_info_free(data->sink_info);
1188 if (data->source_info) {
1189 for (i = 0; i < data->count; ++i)
1190 sched_info_free(data->source_info[i]);
1191 free(data->source_info);
1193 isl_space_free(data->dim);
1198 isl_access_info_free(data->accesses);
1199 sched_info_free(data->sink_info);
1200 if (data->source_info) {
1201 for (i = 0; i < data->count; ++i)
1202 sched_info_free(data->source_info[i]);
1203 free(data->source_info);
1205 isl_space_free(data->dim);
1211 /* Given a collection of "sink" and "source" accesses,
1212 * compute for each iteration of a sink access
1213 * and for each element accessed by that iteration,
1214 * the source access in the list that last accessed the
1215 * element accessed by the sink access before this sink access.
1216 * Each access is given as a map from the loop iterators
1217 * to the array indices.
1218 * The result is a relations between source and sink
1219 * iterations and a subset of the domain of the sink accesses,
1220 * corresponding to those iterations that access an element
1221 * not previously accessed.
1223 * We first prepend the schedule dimensions to the domain
1224 * of the accesses so that we can easily compare their relative order.
1225 * Then we consider each sink access individually in compute_flow.
1227 int isl_union_map_compute_flow(__isl_take isl_union_map *sink,
1228 __isl_take isl_union_map *must_source,
1229 __isl_take isl_union_map *may_source,
1230 __isl_take isl_union_map *schedule,
1231 __isl_give isl_union_map **must_dep, __isl_give isl_union_map **may_dep,
1232 __isl_give isl_union_map **must_no_source,
1233 __isl_give isl_union_map **may_no_source)
1236 isl_union_map *range_map = NULL;
1237 struct isl_compute_flow_data data;
1239 sink = isl_union_map_align_params(sink,
1240 isl_union_map_get_space(must_source));
1241 sink = isl_union_map_align_params(sink,
1242 isl_union_map_get_space(may_source));
1243 sink = isl_union_map_align_params(sink,
1244 isl_union_map_get_space(schedule));
1245 dim = isl_union_map_get_space(sink);
1246 must_source = isl_union_map_align_params(must_source, isl_space_copy(dim));
1247 may_source = isl_union_map_align_params(may_source, isl_space_copy(dim));
1248 schedule = isl_union_map_align_params(schedule, isl_space_copy(dim));
1250 schedule = isl_union_map_reverse(schedule);
1251 range_map = isl_union_map_range_map(schedule);
1252 schedule = isl_union_map_reverse(isl_union_map_copy(range_map));
1253 sink = isl_union_map_apply_domain(sink, isl_union_map_copy(schedule));
1254 must_source = isl_union_map_apply_domain(must_source,
1255 isl_union_map_copy(schedule));
1256 may_source = isl_union_map_apply_domain(may_source, schedule);
1258 data.must_source = must_source;
1259 data.may_source = may_source;
1260 data.must_dep = must_dep ?
1261 isl_union_map_empty(isl_space_copy(dim)) : NULL;
1262 data.may_dep = may_dep ? isl_union_map_empty(isl_space_copy(dim)) : NULL;
1263 data.must_no_source = must_no_source ?
1264 isl_union_map_empty(isl_space_copy(dim)) : NULL;
1265 data.may_no_source = may_no_source ?
1266 isl_union_map_empty(isl_space_copy(dim)) : NULL;
1268 isl_space_free(dim);
1270 if (isl_union_map_foreach_map(sink, &compute_flow, &data) < 0)
1273 isl_union_map_free(sink);
1274 isl_union_map_free(must_source);
1275 isl_union_map_free(may_source);
1278 data.must_dep = isl_union_map_apply_domain(data.must_dep,
1279 isl_union_map_copy(range_map));
1280 data.must_dep = isl_union_map_apply_range(data.must_dep,
1281 isl_union_map_copy(range_map));
1282 *must_dep = data.must_dep;
1285 data.may_dep = isl_union_map_apply_domain(data.may_dep,
1286 isl_union_map_copy(range_map));
1287 data.may_dep = isl_union_map_apply_range(data.may_dep,
1288 isl_union_map_copy(range_map));
1289 *may_dep = data.may_dep;
1291 if (must_no_source) {
1292 data.must_no_source = isl_union_map_apply_domain(
1293 data.must_no_source, isl_union_map_copy(range_map));
1294 *must_no_source = data.must_no_source;
1296 if (may_no_source) {
1297 data.may_no_source = isl_union_map_apply_domain(
1298 data.may_no_source, isl_union_map_copy(range_map));
1299 *may_no_source = data.may_no_source;
1302 isl_union_map_free(range_map);
1306 isl_union_map_free(range_map);
1307 isl_union_map_free(sink);
1308 isl_union_map_free(must_source);
1309 isl_union_map_free(may_source);
1310 isl_union_map_free(data.must_dep);
1311 isl_union_map_free(data.may_dep);
1312 isl_union_map_free(data.must_no_source);
1313 isl_union_map_free(data.may_no_source);
1320 *must_no_source = NULL;
1322 *may_no_source = NULL;