isl_basic_map_set_tuple_name: finalize result
[isl.git] / isl_flow.c
blob3d5aabb2005c175dcdf14995e76d0fb9b06165dd
1 /*
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
16 #include <isl/flow.h>
18 /* A private structure to keep track of a mapping together with
19 * a user-specified identifier and a boolean indicating whether
20 * the map represents a must or may access/dependence.
22 struct isl_labeled_map {
23 struct isl_map *map;
24 void *data;
25 int must;
28 /* A structure containing the input for dependence analysis:
29 * - a sink
30 * - n_must + n_may (<= max_source) sources
31 * - a function for determining the relative order of sources and sink
32 * The must sources are placed before the may sources.
34 struct isl_access_info {
35 struct isl_labeled_map sink;
36 isl_access_level_before level_before;
37 int max_source;
38 int n_must;
39 int n_may;
40 struct isl_labeled_map source[1];
43 /* A structure containing the output of dependence analysis:
44 * - n_source dependences
45 * - a subset of the sink for which definitely no source could be found
46 * - a subset of the sink for which possibly no source could be found
48 struct isl_flow {
49 isl_set *must_no_source;
50 isl_set *may_no_source;
51 int n_source;
52 struct isl_labeled_map *dep;
55 /* Construct an isl_access_info structure and fill it up with
56 * the given data. The number of sources is set to 0.
58 __isl_give isl_access_info *isl_access_info_alloc(__isl_take isl_map *sink,
59 void *sink_user, isl_access_level_before fn, int max_source)
61 isl_ctx *ctx;
62 struct isl_access_info *acc;
64 if (!sink)
65 return NULL;
67 ctx = isl_map_get_ctx(sink);
68 isl_assert(ctx, max_source >= 0, goto error);
70 acc = isl_alloc(ctx, struct isl_access_info,
71 sizeof(struct isl_access_info) +
72 (max_source - 1) * sizeof(struct isl_labeled_map));
73 if (!acc)
74 goto error;
76 acc->sink.map = sink;
77 acc->sink.data = sink_user;
78 acc->level_before = fn;
79 acc->max_source = max_source;
80 acc->n_must = 0;
81 acc->n_may = 0;
83 return acc;
84 error:
85 isl_map_free(sink);
86 return NULL;
89 /* Free the given isl_access_info structure.
91 void isl_access_info_free(__isl_take isl_access_info *acc)
93 int i;
95 if (!acc)
96 return;
97 isl_map_free(acc->sink.map);
98 for (i = 0; i < acc->n_must + acc->n_may; ++i)
99 isl_map_free(acc->source[i].map);
100 free(acc);
103 /* Add another source to an isl_access_info structure, making
104 * sure the "must" sources are placed before the "may" sources.
105 * This function may be called at most max_source times on a
106 * given isl_access_info structure, with max_source as specified
107 * in the call to isl_access_info_alloc that constructed the structure.
109 __isl_give isl_access_info *isl_access_info_add_source(
110 __isl_take isl_access_info *acc, __isl_take isl_map *source,
111 int must, void *source_user)
113 isl_ctx *ctx;
115 if (!acc)
116 return NULL;
117 ctx = isl_map_get_ctx(acc->sink.map);
118 isl_assert(ctx, acc->n_must + acc->n_may < acc->max_source, goto error);
120 if (must) {
121 if (acc->n_may)
122 acc->source[acc->n_must + acc->n_may] =
123 acc->source[acc->n_must];
124 acc->source[acc->n_must].map = source;
125 acc->source[acc->n_must].data = source_user;
126 acc->source[acc->n_must].must = 1;
127 acc->n_must++;
128 } else {
129 acc->source[acc->n_must + acc->n_may].map = source;
130 acc->source[acc->n_must + acc->n_may].data = source_user;
131 acc->source[acc->n_must + acc->n_may].must = 0;
132 acc->n_may++;
135 return acc;
136 error:
137 isl_map_free(source);
138 isl_access_info_free(acc);
139 return NULL;
142 /* A temporary structure used while sorting the accesses in an isl_access_info.
144 struct isl_access_sort_info {
145 struct isl_map *source_map;
146 void *source_data;
147 struct isl_access_info *acc;
150 /* Return -n, 0 or n (with n a positive value), depending on whether
151 * the source access identified by p1 should be sorted before, together
152 * or after that identified by p2.
154 * If p1 and p2 share a different number of levels with the sink,
155 * then the one with the lowest number of shared levels should be
156 * sorted first.
157 * If they both share no levels, then the order is irrelevant.
158 * Otherwise, if p1 appears before p2, then it should be sorted first.
159 * For more generic initial schedules, it is possible that neither
160 * p1 nor p2 appears before the other, or at least not in any obvious way.
161 * We therefore also check if p2 appears before p1, in which case p2
162 * should be sorted first.
163 * If not, we try to order the two statements based on the description
164 * of the iteration domains. This results in an arbitrary, but fairly
165 * stable ordering.
167 static int access_sort_cmp(const void *p1, const void *p2)
169 const struct isl_access_sort_info *i1, *i2;
170 int level1, level2;
171 uint32_t h1, h2;
172 i1 = (const struct isl_access_sort_info *) p1;
173 i2 = (const struct isl_access_sort_info *) p2;
175 level1 = i1->acc->level_before(i1->source_data, i1->acc->sink.data);
176 level2 = i2->acc->level_before(i2->source_data, i2->acc->sink.data);
178 if (level1 != level2 || !level1)
179 return level1 - level2;
181 level1 = i1->acc->level_before(i1->source_data, i2->source_data);
182 if (level1 % 2)
183 return -1;
185 level2 = i1->acc->level_before(i2->source_data, i1->source_data);
186 if (level2 % 2)
187 return 1;
189 h1 = isl_map_get_hash(i1->source_map);
190 h2 = isl_map_get_hash(i2->source_map);
191 return h1 - h2;
194 /* Sort the must source accesses in order of increasing number of shared
195 * levels with the sink access.
196 * Source accesses with the same number of shared levels are sorted
197 * in their textual order.
199 static __isl_give isl_access_info *isl_access_info_sort_sources(
200 __isl_take isl_access_info *acc)
202 int i;
203 isl_ctx *ctx;
204 struct isl_access_sort_info *array;
206 if (!acc)
207 return NULL;
208 if (acc->n_must <= 1)
209 return acc;
211 ctx = isl_map_get_ctx(acc->sink.map);
212 array = isl_alloc_array(ctx, struct isl_access_sort_info, acc->n_must);
213 if (!array)
214 goto error;
216 for (i = 0; i < acc->n_must; ++i) {
217 array[i].source_map = acc->source[i].map;
218 array[i].source_data = acc->source[i].data;
219 array[i].acc = acc;
222 qsort(array, acc->n_must, sizeof(struct isl_access_sort_info),
223 access_sort_cmp);
225 for (i = 0; i < acc->n_must; ++i) {
226 acc->source[i].map = array[i].source_map;
227 acc->source[i].data = array[i].source_data;
230 free(array);
232 return acc;
233 error:
234 isl_access_info_free(acc);
235 return NULL;
238 /* Initialize an empty isl_flow structure corresponding to a given
239 * isl_access_info structure.
240 * For each must access, two dependences are created (initialized
241 * to the empty relation), one for the resulting must dependences
242 * and one for the resulting may dependences. May accesses can
243 * only lead to may dependences, so only one dependence is created
244 * for each of them.
245 * This function is private as isl_flow structures are only supposed
246 * to be created by isl_access_info_compute_flow.
248 static __isl_give isl_flow *isl_flow_alloc(__isl_keep isl_access_info *acc)
250 int i;
251 struct isl_ctx *ctx;
252 struct isl_flow *dep;
254 if (!acc)
255 return NULL;
257 ctx = isl_map_get_ctx(acc->sink.map);
258 dep = isl_calloc_type(ctx, struct isl_flow);
259 if (!dep)
260 return NULL;
262 dep->dep = isl_calloc_array(ctx, struct isl_labeled_map,
263 2 * acc->n_must + acc->n_may);
264 if (!dep->dep)
265 goto error;
267 dep->n_source = 2 * acc->n_must + acc->n_may;
268 for (i = 0; i < acc->n_must; ++i) {
269 struct isl_dim *dim;
270 dim = isl_dim_join(isl_map_get_dim(acc->source[i].map),
271 isl_dim_reverse(isl_map_get_dim(acc->sink.map)));
272 dep->dep[2 * i].map = isl_map_empty(dim);
273 dep->dep[2 * i + 1].map = isl_map_copy(dep->dep[2 * i].map);
274 dep->dep[2 * i].data = acc->source[i].data;
275 dep->dep[2 * i + 1].data = acc->source[i].data;
276 dep->dep[2 * i].must = 1;
277 dep->dep[2 * i + 1].must = 0;
278 if (!dep->dep[2 * i].map || !dep->dep[2 * i + 1].map)
279 goto error;
281 for (i = acc->n_must; i < acc->n_must + acc->n_may; ++i) {
282 struct isl_dim *dim;
283 dim = isl_dim_join(isl_map_get_dim(acc->source[i].map),
284 isl_dim_reverse(isl_map_get_dim(acc->sink.map)));
285 dep->dep[acc->n_must + i].map = isl_map_empty(dim);
286 dep->dep[acc->n_must + i].data = acc->source[i].data;
287 dep->dep[acc->n_must + i].must = 0;
288 if (!dep->dep[acc->n_must + i].map)
289 goto error;
292 return dep;
293 error:
294 isl_flow_free(dep);
295 return NULL;
298 /* Iterate over all sources and for each resulting flow dependence
299 * that is not empty, call the user specfied function.
300 * The second argument in this function call identifies the source,
301 * while the third argument correspond to the final argument of
302 * the isl_flow_foreach call.
304 int isl_flow_foreach(__isl_keep isl_flow *deps,
305 int (*fn)(__isl_take isl_map *dep, int must, void *dep_user, void *user),
306 void *user)
308 int i;
310 if (!deps)
311 return -1;
313 for (i = 0; i < deps->n_source; ++i) {
314 if (isl_map_fast_is_empty(deps->dep[i].map))
315 continue;
316 if (fn(isl_map_copy(deps->dep[i].map), deps->dep[i].must,
317 deps->dep[i].data, user) < 0)
318 return -1;
321 return 0;
324 /* Return a copy of the subset of the sink for which no source could be found.
326 __isl_give isl_set *isl_flow_get_no_source(__isl_keep isl_flow *deps, int must)
328 if (!deps)
329 return NULL;
331 if (must)
332 return isl_set_copy(deps->must_no_source);
333 else
334 return isl_set_copy(deps->may_no_source);
337 void isl_flow_free(__isl_take isl_flow *deps)
339 int i;
341 if (!deps)
342 return;
343 isl_set_free(deps->must_no_source);
344 isl_set_free(deps->may_no_source);
345 if (deps->dep) {
346 for (i = 0; i < deps->n_source; ++i)
347 isl_map_free(deps->dep[i].map);
348 free(deps->dep);
350 free(deps);
353 /* Return a map that enforces that the domain iteration occurs after
354 * the range iteration at the given level.
355 * If level is odd, then the domain iteration should occur after
356 * the target iteration in their shared level/2 outermost loops.
357 * In this case we simply need to enforce that these outermost
358 * loop iterations are the same.
359 * If level is even, then the loop iterator of the domain should
360 * be greater than the loop iterator of the range at the last
361 * of the level/2 shared loops, i.e., loop level/2 - 1.
363 static __isl_give isl_map *after_at_level(struct isl_dim *dim, int level)
365 struct isl_basic_map *bmap;
367 if (level % 2)
368 bmap = isl_basic_map_equal(dim, level/2);
369 else
370 bmap = isl_basic_map_more_at(dim, level/2 - 1);
372 return isl_map_from_basic_map(bmap);
375 /* Compute the last iteration of must source j that precedes the sink
376 * at the given level for sink iterations in set_C.
377 * The subset of set_C for which no such iteration can be found is returned
378 * in *empty.
380 static struct isl_map *last_source(struct isl_access_info *acc,
381 struct isl_set *set_C,
382 int j, int level, struct isl_set **empty)
384 struct isl_map *read_map;
385 struct isl_map *write_map;
386 struct isl_map *dep_map;
387 struct isl_map *after;
388 struct isl_map *result;
390 read_map = isl_map_copy(acc->sink.map);
391 write_map = isl_map_copy(acc->source[j].map);
392 write_map = isl_map_reverse(write_map);
393 dep_map = isl_map_apply_range(read_map, write_map);
394 after = after_at_level(isl_map_get_dim(dep_map), level);
395 dep_map = isl_map_intersect(dep_map, after);
396 result = isl_map_partial_lexmax(dep_map, set_C, empty);
397 result = isl_map_reverse(result);
399 return result;
402 /* For a given mapping between iterations of must source j and iterations
403 * of the sink, compute the last iteration of must source k preceding
404 * the sink at level before_level for any of the sink iterations,
405 * but following the corresponding iteration of must source j at level
406 * after_level.
408 static struct isl_map *last_later_source(struct isl_access_info *acc,
409 struct isl_map *old_map,
410 int j, int before_level,
411 int k, int after_level,
412 struct isl_set **empty)
414 struct isl_dim *dim;
415 struct isl_set *set_C;
416 struct isl_map *read_map;
417 struct isl_map *write_map;
418 struct isl_map *dep_map;
419 struct isl_map *after_write;
420 struct isl_map *before_read;
421 struct isl_map *result;
423 set_C = isl_map_range(isl_map_copy(old_map));
424 read_map = isl_map_copy(acc->sink.map);
425 write_map = isl_map_copy(acc->source[k].map);
427 write_map = isl_map_reverse(write_map);
428 dep_map = isl_map_apply_range(read_map, write_map);
429 dim = isl_dim_join(isl_map_get_dim(acc->source[k].map),
430 isl_dim_reverse(isl_map_get_dim(acc->source[j].map)));
431 after_write = after_at_level(dim, after_level);
432 after_write = isl_map_apply_range(after_write, old_map);
433 after_write = isl_map_reverse(after_write);
434 dep_map = isl_map_intersect(dep_map, after_write);
435 before_read = after_at_level(isl_map_get_dim(dep_map), before_level);
436 dep_map = isl_map_intersect(dep_map, before_read);
437 result = isl_map_partial_lexmax(dep_map, set_C, empty);
438 result = isl_map_reverse(result);
440 return result;
443 /* Given a shared_level between two accesses, return 1 if the
444 * the first can precede the second at the requested target_level.
445 * If the target level is odd, i.e., refers to a statement level
446 * dimension, then first needs to precede second at the requested
447 * level, i.e., shared_level must be equal to target_level.
448 * If the target level is odd, then the two loops should share
449 * at least the requested number of outer loops.
451 static int can_precede_at_level(int shared_level, int target_level)
453 if (shared_level < target_level)
454 return 0;
455 if ((target_level % 2) && shared_level > target_level)
456 return 0;
457 return 1;
460 /* Given a possible flow dependence temp_rel[j] between source j and the sink
461 * at level sink_level, remove those elements for which
462 * there is an iteration of another source k < j that is closer to the sink.
463 * The flow dependences temp_rel[k] are updated with the improved sources.
464 * Any improved source needs to precede the sink at the same level
465 * and needs to follow source j at the same or a deeper level.
466 * The lower this level, the later the execution date of source k.
467 * We therefore consider lower levels first.
469 * If temp_rel[j] is empty, then there can be no improvement and
470 * we return immediately.
472 static int intermediate_sources(__isl_keep isl_access_info *acc,
473 struct isl_map **temp_rel, int j, int sink_level)
475 int k, level;
476 int depth = 2 * isl_map_dim(acc->source[j].map, isl_dim_in) + 1;
478 if (isl_map_fast_is_empty(temp_rel[j]))
479 return 0;
481 for (k = j - 1; k >= 0; --k) {
482 int plevel, plevel2;
483 plevel = acc->level_before(acc->source[k].data, acc->sink.data);
484 if (!can_precede_at_level(plevel, sink_level))
485 continue;
487 plevel2 = acc->level_before(acc->source[j].data,
488 acc->source[k].data);
490 for (level = sink_level; level <= depth; ++level) {
491 struct isl_map *T;
492 struct isl_set *trest;
493 struct isl_map *copy;
495 if (!can_precede_at_level(plevel2, level))
496 continue;
498 copy = isl_map_copy(temp_rel[j]);
499 T = last_later_source(acc, copy, j, sink_level, k,
500 level, &trest);
501 if (isl_map_fast_is_empty(T)) {
502 isl_set_free(trest);
503 isl_map_free(T);
504 continue;
506 temp_rel[j] = isl_map_intersect_range(temp_rel[j], trest);
507 temp_rel[k] = isl_map_union_disjoint(temp_rel[k], T);
511 return 0;
514 /* Compute all iterations of may source j that precedes the sink at the given
515 * level for sink iterations in set_C.
517 static __isl_give isl_map *all_sources(__isl_keep isl_access_info *acc,
518 __isl_take isl_set *set_C, int j, int level)
520 isl_map *read_map;
521 isl_map *write_map;
522 isl_map *dep_map;
523 isl_map *after;
525 read_map = isl_map_copy(acc->sink.map);
526 read_map = isl_map_intersect_domain(read_map, set_C);
527 write_map = isl_map_copy(acc->source[acc->n_must + j].map);
528 write_map = isl_map_reverse(write_map);
529 dep_map = isl_map_apply_range(read_map, write_map);
530 after = after_at_level(isl_map_get_dim(dep_map), level);
531 dep_map = isl_map_intersect(dep_map, after);
533 return isl_map_reverse(dep_map);
536 /* For a given mapping between iterations of must source k and iterations
537 * of the sink, compute the all iteration of may source j preceding
538 * the sink at level before_level for any of the sink iterations,
539 * but following the corresponding iteration of must source k at level
540 * after_level.
542 static __isl_give isl_map *all_later_sources(__isl_keep isl_access_info *acc,
543 __isl_keep isl_map *old_map,
544 int j, int before_level, int k, int after_level)
546 isl_dim *dim;
547 isl_set *set_C;
548 isl_map *read_map;
549 isl_map *write_map;
550 isl_map *dep_map;
551 isl_map *after_write;
552 isl_map *before_read;
554 set_C = isl_map_range(isl_map_copy(old_map));
555 read_map = isl_map_copy(acc->sink.map);
556 read_map = isl_map_intersect_domain(read_map, set_C);
557 write_map = isl_map_copy(acc->source[acc->n_must + j].map);
559 write_map = isl_map_reverse(write_map);
560 dep_map = isl_map_apply_range(read_map, write_map);
561 dim = isl_dim_join(isl_map_get_dim(acc->source[acc->n_must + j].map),
562 isl_dim_reverse(isl_map_get_dim(acc->source[k].map)));
563 after_write = after_at_level(dim, after_level);
564 after_write = isl_map_apply_range(after_write, old_map);
565 after_write = isl_map_reverse(after_write);
566 dep_map = isl_map_intersect(dep_map, after_write);
567 before_read = after_at_level(isl_map_get_dim(dep_map), before_level);
568 dep_map = isl_map_intersect(dep_map, before_read);
569 return isl_map_reverse(dep_map);
572 /* Given the must and may dependence relations for the must accesses
573 * for level sink_level, check if there are any accesses of may access j
574 * that occur in between and return their union.
575 * If some of these accesses are intermediate with respect to
576 * (previously thought to be) must dependences, then these
577 * must dependences are turned into may dependences.
579 static __isl_give isl_map *all_intermediate_sources(
580 __isl_keep isl_access_info *acc, __isl_take isl_map *map,
581 struct isl_map **must_rel, struct isl_map **may_rel,
582 int j, int sink_level)
584 int k, level;
585 int depth = 2 * isl_map_dim(acc->source[acc->n_must + j].map,
586 isl_dim_in) + 1;
588 for (k = 0; k < acc->n_must; ++k) {
589 int plevel;
591 if (isl_map_fast_is_empty(may_rel[k]) &&
592 isl_map_fast_is_empty(must_rel[k]))
593 continue;
595 plevel = acc->level_before(acc->source[k].data,
596 acc->source[acc->n_must + j].data);
598 for (level = sink_level; level <= depth; ++level) {
599 isl_map *T;
600 isl_map *copy;
601 isl_set *ran;
603 if (!can_precede_at_level(plevel, level))
604 continue;
606 copy = isl_map_copy(may_rel[k]);
607 T = all_later_sources(acc, copy, j, sink_level, k, level);
608 map = isl_map_union(map, T);
610 copy = isl_map_copy(must_rel[k]);
611 T = all_later_sources(acc, copy, j, sink_level, k, level);
612 ran = isl_map_range(isl_map_copy(T));
613 map = isl_map_union(map, T);
614 may_rel[k] = isl_map_union_disjoint(may_rel[k],
615 isl_map_intersect_range(isl_map_copy(must_rel[k]),
616 isl_set_copy(ran)));
617 T = isl_map_from_domain_and_range(
618 isl_set_universe(
619 isl_dim_domain(isl_map_get_dim(must_rel[k]))),
620 ran);
621 must_rel[k] = isl_map_subtract(must_rel[k], T);
625 return map;
628 /* Compute dependences for the case where all accesses are "may"
629 * accesses, which boils down to computing memory based dependences.
630 * The generic algorithm would also work in this case, but it would
631 * be overkill to use it.
633 static __isl_give isl_flow *compute_mem_based_dependences(
634 __isl_take isl_access_info *acc)
636 int i;
637 isl_set *mustdo;
638 isl_set *maydo;
639 isl_flow *res;
641 res = isl_flow_alloc(acc);
642 if (!res)
643 goto error;
645 mustdo = isl_map_domain(isl_map_copy(acc->sink.map));
646 maydo = isl_set_copy(mustdo);
648 for (i = 0; i < acc->n_may; ++i) {
649 int plevel;
650 int is_before;
651 isl_dim *dim;
652 isl_map *before;
653 isl_map *dep;
655 plevel = acc->level_before(acc->source[i].data, acc->sink.data);
656 is_before = plevel & 1;
657 plevel >>= 1;
659 dim = isl_map_get_dim(res->dep[i].map);
660 if (is_before)
661 before = isl_map_lex_le_first(dim, plevel);
662 else
663 before = isl_map_lex_lt_first(dim, plevel);
664 dep = isl_map_apply_range(isl_map_copy(acc->source[i].map),
665 isl_map_reverse(isl_map_copy(acc->sink.map)));
666 dep = isl_map_intersect(dep, before);
667 mustdo = isl_set_subtract(mustdo,
668 isl_map_range(isl_map_copy(dep)));
669 res->dep[i].map = isl_map_union(res->dep[i].map, dep);
672 res->may_no_source = isl_set_subtract(maydo, isl_set_copy(mustdo));
673 res->must_no_source = mustdo;
675 isl_access_info_free(acc);
677 return res;
678 error:
679 isl_access_info_free(acc);
680 return NULL;
683 /* Compute dependences for the case where there is at least one
684 * "must" access.
686 * The core algorithm considers all levels in which a source may precede
687 * the sink, where a level may either be a statement level or a loop level.
688 * The outermost statement level is 1, the first loop level is 2, etc...
689 * The algorithm basically does the following:
690 * for all levels l of the read access from innermost to outermost
691 * for all sources w that may precede the sink access at that level
692 * compute the last iteration of the source that precedes the sink access
693 * at that level
694 * add result to possible last accesses at level l of source w
695 * for all sources w2 that we haven't considered yet at this level that may
696 * also precede the sink access
697 * for all levels l2 of w from l to innermost
698 * for all possible last accesses dep of w at l
699 * compute last iteration of w2 between the source and sink
700 * of dep
701 * add result to possible last accesses at level l of write w2
702 * and replace possible last accesses dep by the remainder
705 * The above algorithm is applied to the must access. During the course
706 * of the algorithm, we keep track of sink iterations that still
707 * need to be considered. These iterations are split into those that
708 * haven't been matched to any source access (mustdo) and those that have only
709 * been matched to may accesses (maydo).
710 * At the end of each level, we also consider the may accesses.
711 * In particular, we consider may accesses that precede the remaining
712 * sink iterations, moving elements from mustdo to maydo when appropriate,
713 * and may accesses that occur between a must source and a sink of any
714 * dependences found at the current level, turning must dependences into
715 * may dependences when appropriate.
718 static __isl_give isl_flow *compute_val_based_dependences(
719 __isl_take isl_access_info *acc)
721 isl_ctx *ctx;
722 isl_flow *res;
723 isl_set *mustdo = NULL;
724 isl_set *maydo = NULL;
725 int level, j;
726 int depth;
727 isl_map **must_rel = NULL;
728 isl_map **may_rel = NULL;
730 acc = isl_access_info_sort_sources(acc);
731 if (!acc)
732 return NULL;
734 res = isl_flow_alloc(acc);
735 if (!res)
736 goto error;
737 ctx = isl_map_get_ctx(acc->sink.map);
739 depth = 2 * isl_map_dim(acc->sink.map, isl_dim_in) + 1;
740 mustdo = isl_map_domain(isl_map_copy(acc->sink.map));
741 maydo = isl_set_empty_like(mustdo);
742 if (!mustdo || !maydo)
743 goto error;
744 if (isl_set_fast_is_empty(mustdo))
745 goto done;
747 must_rel = isl_alloc_array(ctx, struct isl_map *, acc->n_must);
748 may_rel = isl_alloc_array(ctx, struct isl_map *, acc->n_must);
749 if (!must_rel || !may_rel)
750 goto error;
752 for (level = depth; level >= 1; --level) {
753 for (j = acc->n_must-1; j >=0; --j) {
754 must_rel[j] = isl_map_empty_like(res->dep[j].map);
755 may_rel[j] = isl_map_copy(must_rel[j]);
758 for (j = acc->n_must - 1; j >= 0; --j) {
759 struct isl_map *T;
760 struct isl_set *rest;
761 int plevel;
763 plevel = acc->level_before(acc->source[j].data,
764 acc->sink.data);
765 if (!can_precede_at_level(plevel, level))
766 continue;
768 T = last_source(acc, mustdo, j, level, &rest);
769 must_rel[j] = isl_map_union_disjoint(must_rel[j], T);
770 mustdo = rest;
772 intermediate_sources(acc, must_rel, j, level);
774 T = last_source(acc, maydo, j, level, &rest);
775 may_rel[j] = isl_map_union_disjoint(may_rel[j], T);
776 maydo = rest;
778 intermediate_sources(acc, may_rel, j, level);
780 if (isl_set_fast_is_empty(mustdo) &&
781 isl_set_fast_is_empty(maydo))
782 break;
784 for (j = j - 1; j >= 0; --j) {
785 int plevel;
787 plevel = acc->level_before(acc->source[j].data,
788 acc->sink.data);
789 if (!can_precede_at_level(plevel, level))
790 continue;
792 intermediate_sources(acc, must_rel, j, level);
793 intermediate_sources(acc, may_rel, j, level);
796 for (j = 0; j < acc->n_may; ++j) {
797 int plevel;
798 isl_map *T;
799 isl_set *ran;
801 plevel = acc->level_before(acc->source[acc->n_must + j].data,
802 acc->sink.data);
803 if (!can_precede_at_level(plevel, level))
804 continue;
806 T = all_sources(acc, isl_set_copy(maydo), j, level);
807 res->dep[2 * acc->n_must + j].map =
808 isl_map_union(res->dep[2 * acc->n_must + j].map, T);
809 T = all_sources(acc, isl_set_copy(mustdo), j, level);
810 ran = isl_map_range(isl_map_copy(T));
811 res->dep[2 * acc->n_must + j].map =
812 isl_map_union(res->dep[2 * acc->n_must + j].map, T);
813 mustdo = isl_set_subtract(mustdo, isl_set_copy(ran));
814 maydo = isl_set_union_disjoint(maydo, ran);
816 T = res->dep[2 * acc->n_must + j].map;
817 T = all_intermediate_sources(acc, T, must_rel, may_rel,
818 j, level);
819 res->dep[2 * acc->n_must + j].map = T;
822 for (j = acc->n_must - 1; j >= 0; --j) {
823 res->dep[2 * j].map =
824 isl_map_union_disjoint(res->dep[2 * j].map,
825 must_rel[j]);
826 res->dep[2 * j + 1].map =
827 isl_map_union_disjoint(res->dep[2 * j + 1].map,
828 may_rel[j]);
831 if (isl_set_fast_is_empty(mustdo) &&
832 isl_set_fast_is_empty(maydo))
833 break;
836 free(must_rel);
837 free(may_rel);
838 done:
839 res->must_no_source = mustdo;
840 res->may_no_source = maydo;
841 isl_access_info_free(acc);
842 return res;
843 error:
844 isl_access_info_free(acc);
845 isl_flow_free(res);
846 isl_set_free(mustdo);
847 isl_set_free(maydo);
848 free(must_rel);
849 free(may_rel);
850 return NULL;
853 /* Given a "sink" access, a list of n "source" accesses,
854 * compute for each iteration of the sink access
855 * and for each element accessed by that iteration,
856 * the source access in the list that last accessed the
857 * element accessed by the sink access before this sink access.
858 * Each access is given as a map from the loop iterators
859 * to the array indices.
860 * The result is a list of n relations between source and sink
861 * iterations and a subset of the domain of the sink access,
862 * corresponding to those iterations that access an element
863 * not previously accessed.
865 * To deal with multi-valued sink access relations, the sink iteration
866 * domain is first extended with dimensions that correspond to the data
867 * space. After the computation is finished, these extra dimensions are
868 * projected out again.
870 __isl_give isl_flow *isl_access_info_compute_flow(__isl_take isl_access_info *acc)
872 int j;
873 struct isl_flow *res;
874 isl_map *domain_map = NULL;
876 if (!acc)
877 return NULL;
879 domain_map = isl_map_domain_map(isl_map_copy(acc->sink.map));
880 acc->sink.map = isl_map_range_map(acc->sink.map);
881 if (!acc->sink.map)
882 goto error;
884 if (acc->n_must == 0)
885 res = compute_mem_based_dependences(acc);
886 else
887 res = compute_val_based_dependences(acc);
888 if (!res)
889 return NULL;
891 for (j = 0; j < res->n_source; ++j) {
892 res->dep[j].map = isl_map_apply_range(res->dep[j].map,
893 isl_map_copy(domain_map));
894 if (!res->dep[j].map)
895 goto error2;
897 res->must_no_source = isl_set_apply(res->must_no_source,
898 isl_map_copy(domain_map));
899 res->may_no_source = isl_set_apply(res->may_no_source,
900 isl_map_copy(domain_map));
901 if (!res->must_no_source || !res->may_no_source)
902 goto error2;
904 isl_map_free(domain_map);
905 return res;
906 error:
907 isl_map_free(domain_map);
908 isl_access_info_free(acc);
909 return NULL;
910 error2:
911 isl_map_free(domain_map);
912 isl_flow_free(res);
913 return NULL;
917 /* Keep track of some information about a schedule for a given
918 * access. In particular, keep track of which dimensions
919 * have a constant value and of the actual constant values.
921 struct isl_sched_info {
922 int *is_cst;
923 isl_vec *cst;
926 static void sched_info_free(__isl_take struct isl_sched_info *info)
928 if (!info)
929 return;
930 isl_vec_free(info->cst);
931 free(info->is_cst);
932 free(info);
935 /* Extract information on the constant dimensions of the schedule
936 * for a given access. The "map" is of the form
938 * [S -> D] -> A
940 * with S the schedule domain, D the iteration domain and A the data domain.
942 static __isl_give struct isl_sched_info *sched_info_alloc(
943 __isl_keep isl_map *map)
945 isl_ctx *ctx;
946 isl_dim *dim;
947 struct isl_sched_info *info;
948 int i, n;
950 if (!map)
951 return NULL;
953 dim = isl_dim_unwrap(isl_dim_domain(isl_map_get_dim(map)));
954 if (!dim)
955 return NULL;
956 n = isl_dim_size(dim, isl_dim_in);
957 isl_dim_free(dim);
959 ctx = isl_map_get_ctx(map);
960 info = isl_alloc_type(ctx, struct isl_sched_info);
961 if (!info)
962 return NULL;
963 info->is_cst = isl_alloc_array(ctx, int, n);
964 info->cst = isl_vec_alloc(ctx, n);
965 if (!info->is_cst || !info->cst)
966 goto error;
968 for (i = 0; i < n; ++i)
969 info->is_cst[i] = isl_map_fast_is_fixed(map, isl_dim_in, i,
970 &info->cst->el[i]);
972 return info;
973 error:
974 sched_info_free(info);
975 return NULL;
978 struct isl_compute_flow_data {
979 isl_union_map *must_source;
980 isl_union_map *may_source;
981 isl_union_map *must_dep;
982 isl_union_map *may_dep;
983 isl_union_set *must_no_source;
984 isl_union_set *may_no_source;
986 int count;
987 int must;
988 isl_dim *dim;
989 struct isl_sched_info *sink_info;
990 struct isl_sched_info **source_info;
991 isl_access_info *accesses;
994 static int count_matching_array(__isl_take isl_map *map, void *user)
996 int eq;
997 isl_dim *dim;
998 struct isl_compute_flow_data *data;
1000 data = (struct isl_compute_flow_data *)user;
1002 dim = isl_dim_range(isl_map_get_dim(map));
1004 eq = isl_dim_equal(dim, data->dim);
1006 isl_dim_free(dim);
1007 isl_map_free(map);
1009 if (eq < 0)
1010 return -1;
1011 if (eq)
1012 data->count++;
1014 return 0;
1017 static int collect_matching_array(__isl_take isl_map *map, void *user)
1019 int eq;
1020 isl_dim *dim;
1021 struct isl_sched_info *info;
1022 struct isl_compute_flow_data *data;
1024 data = (struct isl_compute_flow_data *)user;
1026 dim = isl_dim_range(isl_map_get_dim(map));
1028 eq = isl_dim_equal(dim, data->dim);
1030 isl_dim_free(dim);
1032 if (eq < 0)
1033 goto error;
1034 if (!eq) {
1035 isl_map_free(map);
1036 return 0;
1039 info = sched_info_alloc(map);
1040 data->source_info[data->count] = info;
1042 data->accesses = isl_access_info_add_source(data->accesses,
1043 map, data->must, info);
1045 data->count++;
1047 return 0;
1048 error:
1049 isl_map_free(map);
1050 return -1;
1053 /* Determine the shared nesting level and the "textual order" of
1054 * the given accesses.
1056 * We first determine the minimal schedule dimension for both accesses.
1058 * If among those dimensions, we can find one where both have a fixed
1059 * value and if moreover those values are different, then the previous
1060 * dimension is the last shared nesting level and the textual order
1061 * is determined based on the order of the fixed values.
1062 * If no such fixed values can be found, then we set the shared
1063 * nesting level to the minimal schedule dimension, with no textual ordering.
1065 static int before(void *first, void *second)
1067 struct isl_sched_info *info1 = first;
1068 struct isl_sched_info *info2 = second;
1069 int n1, n2;
1070 int i;
1072 n1 = info1->cst->size;
1073 n2 = info2->cst->size;
1075 if (n2 < n1)
1076 n1 = n2;
1078 for (i = 0; i < n1; ++i) {
1079 if (!info1->is_cst[i])
1080 continue;
1081 if (!info2->is_cst[i])
1082 continue;
1083 if (isl_int_eq(info1->cst->el[i], info2->cst->el[i]))
1084 continue;
1085 return 2 * i + isl_int_lt(info1->cst->el[i], info2->cst->el[i]);
1088 return 2 * n1;
1091 /* Given a sink access, look for all the source accesses that access
1092 * the same array and perform dataflow analysis on them using
1093 * isl_access_info_compute_flow.
1095 static int compute_flow(__isl_take isl_map *map, void *user)
1097 int i;
1098 isl_ctx *ctx;
1099 struct isl_compute_flow_data *data;
1100 isl_flow *flow;
1102 data = (struct isl_compute_flow_data *)user;
1104 ctx = isl_map_get_ctx(map);
1106 data->accesses = NULL;
1107 data->sink_info = NULL;
1108 data->source_info = NULL;
1109 data->count = 0;
1110 data->dim = isl_dim_range(isl_map_get_dim(map));
1112 if (isl_union_map_foreach_map(data->must_source,
1113 &count_matching_array, data) < 0)
1114 goto error;
1115 if (isl_union_map_foreach_map(data->may_source,
1116 &count_matching_array, data) < 0)
1117 goto error;
1119 data->sink_info = sched_info_alloc(map);
1120 data->source_info = isl_calloc_array(ctx, struct isl_sched_info *,
1121 data->count);
1123 data->accesses = isl_access_info_alloc(isl_map_copy(map),
1124 data->sink_info, &before, data->count);
1125 if (!data->sink_info || !data->source_info || !data->accesses)
1126 goto error;
1127 data->count = 0;
1128 data->must = 1;
1129 if (isl_union_map_foreach_map(data->must_source,
1130 &collect_matching_array, data) < 0)
1131 goto error;
1132 data->must = 0;
1133 if (isl_union_map_foreach_map(data->may_source,
1134 &collect_matching_array, data) < 0)
1135 goto error;
1137 flow = isl_access_info_compute_flow(data->accesses);
1138 data->accesses = NULL;
1140 if (!flow)
1141 goto error;
1143 data->must_no_source = isl_union_set_union(data->must_no_source,
1144 isl_union_set_from_set(isl_set_copy(flow->must_no_source)));
1145 data->may_no_source = isl_union_set_union(data->may_no_source,
1146 isl_union_set_from_set(isl_set_copy(flow->may_no_source)));
1148 for (i = 0; i < flow->n_source; ++i) {
1149 isl_union_map *dep;
1150 dep = isl_union_map_from_map(isl_map_copy(flow->dep[i].map));
1151 if (flow->dep[i].must)
1152 data->must_dep = isl_union_map_union(data->must_dep, dep);
1153 else
1154 data->may_dep = isl_union_map_union(data->may_dep, dep);
1157 isl_flow_free(flow);
1159 sched_info_free(data->sink_info);
1160 if (data->source_info) {
1161 for (i = 0; i < data->count; ++i)
1162 sched_info_free(data->source_info[i]);
1163 free(data->source_info);
1165 isl_dim_free(data->dim);
1166 isl_map_free(map);
1168 return 0;
1169 error:
1170 isl_access_info_free(data->accesses);
1171 sched_info_free(data->sink_info);
1172 if (data->source_info) {
1173 for (i = 0; i < data->count; ++i)
1174 sched_info_free(data->source_info[i]);
1175 free(data->source_info);
1177 isl_dim_free(data->dim);
1178 isl_map_free(map);
1180 return -1;
1183 /* Given a collection of "sink" and "source" accesses,
1184 * compute for each iteration of a sink access
1185 * and for each element accessed by that iteration,
1186 * the source access in the list that last accessed the
1187 * element accessed by the sink access before this sink access.
1188 * Each access is given as a map from the loop iterators
1189 * to the array indices.
1190 * The result is a relations between source and sink
1191 * iterations and a subset of the domain of the sink accesses,
1192 * corresponding to those iterations that access an element
1193 * not previously accessed.
1195 * We first prepend the schedule dimensions to the domain
1196 * of the accesses so that we can easily compare their relative order.
1197 * Then we consider each sink access individually in compute_flow.
1199 int isl_union_map_compute_flow(__isl_take isl_union_map *sink,
1200 __isl_take isl_union_map *must_source,
1201 __isl_take isl_union_map *may_source,
1202 __isl_take isl_union_map *schedule,
1203 __isl_give isl_union_map **must_dep, __isl_give isl_union_map **may_dep,
1204 __isl_give isl_union_set **must_no_source,
1205 __isl_give isl_union_set **may_no_source)
1207 isl_dim *dim;
1208 isl_union_map *range_map = NULL;
1209 struct isl_compute_flow_data data;
1211 sink = isl_union_map_align_params(sink,
1212 isl_union_map_get_dim(must_source));
1213 sink = isl_union_map_align_params(sink,
1214 isl_union_map_get_dim(may_source));
1215 sink = isl_union_map_align_params(sink,
1216 isl_union_map_get_dim(schedule));
1217 dim = isl_union_map_get_dim(sink);
1218 must_source = isl_union_map_align_params(must_source, isl_dim_copy(dim));
1219 may_source = isl_union_map_align_params(may_source, isl_dim_copy(dim));
1220 schedule = isl_union_map_align_params(schedule, isl_dim_copy(dim));
1222 schedule = isl_union_map_reverse(schedule);
1223 range_map = isl_union_map_range_map(schedule);
1224 schedule = isl_union_map_reverse(isl_union_map_copy(range_map));
1225 sink = isl_union_map_apply_domain(sink, isl_union_map_copy(schedule));
1226 must_source = isl_union_map_apply_domain(must_source,
1227 isl_union_map_copy(schedule));
1228 may_source = isl_union_map_apply_domain(may_source, schedule);
1230 data.must_source = must_source;
1231 data.may_source = may_source;
1232 data.must_dep = must_dep ?
1233 isl_union_map_empty(isl_dim_copy(dim)) : NULL;
1234 data.may_dep = may_dep ? isl_union_map_empty(isl_dim_copy(dim)) : NULL;
1235 data.must_no_source = must_no_source ?
1236 isl_union_set_empty(isl_dim_copy(dim)) : NULL;
1237 data.may_no_source = may_no_source ?
1238 isl_union_set_empty(isl_dim_copy(dim)) : NULL;
1240 isl_dim_free(dim);
1242 if (isl_union_map_foreach_map(sink, &compute_flow, &data) < 0)
1243 goto error;
1245 isl_union_map_free(sink);
1246 isl_union_map_free(must_source);
1247 isl_union_map_free(may_source);
1249 if (must_dep) {
1250 data.must_dep = isl_union_map_apply_domain(data.must_dep,
1251 isl_union_map_copy(range_map));
1252 data.must_dep = isl_union_map_apply_range(data.must_dep,
1253 isl_union_map_copy(range_map));
1254 *must_dep = data.must_dep;
1256 if (may_dep) {
1257 data.may_dep = isl_union_map_apply_domain(data.may_dep,
1258 isl_union_map_copy(range_map));
1259 data.may_dep = isl_union_map_apply_range(data.may_dep,
1260 isl_union_map_copy(range_map));
1261 *may_dep = data.may_dep;
1263 if (must_no_source) {
1264 data.must_no_source = isl_union_set_apply(data.must_no_source,
1265 isl_union_map_copy(range_map));
1266 *must_no_source = data.must_no_source;
1268 if (may_no_source) {
1269 data.may_no_source = isl_union_set_apply(data.may_no_source,
1270 isl_union_map_copy(range_map));
1271 *may_no_source = data.may_no_source;
1274 isl_union_map_free(range_map);
1276 return 0;
1277 error:
1278 isl_union_map_free(range_map);
1279 isl_union_map_free(sink);
1280 isl_union_map_free(must_source);
1281 isl_union_map_free(may_source);
1282 isl_union_map_free(data.must_dep);
1283 isl_union_map_free(data.may_dep);
1284 isl_union_set_free(data.must_no_source);
1285 isl_union_set_free(data.may_no_source);
1287 if (must_dep)
1288 *must_dep = NULL;
1289 if (may_dep)
1290 *may_dep = NULL;
1291 if (must_no_source)
1292 *must_no_source = NULL;
1293 if (may_no_source)
1294 *may_no_source = NULL;
1295 return -1;