Home | History | Annotate | Line # | Download | only in dist
      1 /*
      2  * Copyright 2011      INRIA Saclay
      3  * Copyright 2012-2014 Ecole Normale Superieure
      4  * Copyright 2015-2016 Sven Verdoolaege
      5  * Copyright 2016      INRIA Paris
      6  * Copyright 2017      Sven Verdoolaege
      7  *
      8  * Use of this software is governed by the MIT license
      9  *
     10  * Written by Sven Verdoolaege, INRIA Saclay - Ile-de-France,
     11  * Parc Club Orsay Universite, ZAC des vignes, 4 rue Jacques Monod,
     12  * 91893 Orsay, France
     13  * and Ecole Normale Superieure, 45 rue d'Ulm, 75230 Paris, France
     14  * and Centre de Recherche Inria de Paris, 2 rue Simone Iff - Voie DQ12,
     15  * CS 42112, 75589 Paris Cedex 12, France
     16  */
     17 
     18 #include <isl_ctx_private.h>
     19 #include <isl_map_private.h>
     20 #include <isl_space_private.h>
     21 #include <isl_aff_private.h>
     22 #include <isl/hash.h>
     23 #include <isl/id.h>
     24 #include <isl/constraint.h>
     25 #include <isl/schedule.h>
     26 #include <isl_schedule_constraints.h>
     27 #include <isl/schedule_node.h>
     28 #include <isl_mat_private.h>
     29 #include <isl_vec_private.h>
     30 #include <isl/set.h>
     31 #include <isl_union_set_private.h>
     32 #include <isl_seq.h>
     33 #include <isl_tab.h>
     34 #include <isl_dim_map.h>
     35 #include <isl/map_to_basic_set.h>
     36 #include <isl_sort.h>
     37 #include <isl_options_private.h>
     38 #include <isl_tarjan.h>
     39 #include <isl_morph.h>
     40 #include <isl/ilp.h>
     41 #include <isl_val_private.h>
     42 
     43 #include "isl_scheduler.h"
     44 #include "isl_scheduler_clustering.h"
     45 
     46 /*
     47  * The scheduling algorithm implemented in this file was inspired by
     48  * Bondhugula et al., "Automatic Transformations for Communication-Minimized
     49  * Parallelization and Locality Optimization in the Polyhedral Model".
     50  *
     51  * For a detailed description of the variant implemented in isl,
     52  * see Verdoolaege and Janssens, "Scheduling for PPCG" (2017).
     53  */
     54 
     55 
     56 static isl_bool node_has_tuples(const void *entry, const void *val)
     57 {
     58 	struct isl_sched_node *node = (struct isl_sched_node *)entry;
     59 	isl_space *space = (isl_space *) val;
     60 
     61 	return isl_space_has_equal_tuples(node->space, space);
     62 }
     63 
     64 int isl_sched_node_scc_exactly(struct isl_sched_node *node, int scc)
     65 {
     66 	return node->scc == scc;
     67 }
     68 
     69 static int node_scc_at_most(struct isl_sched_node *node, int scc)
     70 {
     71 	return node->scc <= scc;
     72 }
     73 
     74 static int node_scc_at_least(struct isl_sched_node *node, int scc)
     75 {
     76 	return node->scc >= scc;
     77 }
     78 
     79 /* Is "edge" marked as being of type "type"?
     80  */
     81 int isl_sched_edge_has_type(struct isl_sched_edge *edge,
     82 	enum isl_edge_type type)
     83 {
     84 	return ISL_FL_ISSET(edge->types, 1 << type);
     85 }
     86 
     87 /* Mark "edge" as being of type "type".
     88  */
     89 static void set_type(struct isl_sched_edge *edge, enum isl_edge_type type)
     90 {
     91 	ISL_FL_SET(edge->types, 1 << type);
     92 }
     93 
     94 /* No longer mark "edge" as being of type "type"?
     95  */
     96 static void clear_type(struct isl_sched_edge *edge, enum isl_edge_type type)
     97 {
     98 	ISL_FL_CLR(edge->types, 1 << type);
     99 }
    100 
    101 /* Is "edge" marked as a validity edge?
    102  */
    103 static int is_validity(struct isl_sched_edge *edge)
    104 {
    105 	return isl_sched_edge_has_type(edge, isl_edge_validity);
    106 }
    107 
    108 /* Mark "edge" as a validity edge.
    109  */
    110 static void set_validity(struct isl_sched_edge *edge)
    111 {
    112 	set_type(edge, isl_edge_validity);
    113 }
    114 
    115 /* Is "edge" marked as a proximity edge?
    116  */
    117 int isl_sched_edge_is_proximity(struct isl_sched_edge *edge)
    118 {
    119 	return isl_sched_edge_has_type(edge, isl_edge_proximity);
    120 }
    121 
    122 /* Is "edge" marked as a local edge?
    123  */
    124 static int is_local(struct isl_sched_edge *edge)
    125 {
    126 	return isl_sched_edge_has_type(edge, isl_edge_local);
    127 }
    128 
    129 /* Mark "edge" as a local edge.
    130  */
    131 static void set_local(struct isl_sched_edge *edge)
    132 {
    133 	set_type(edge, isl_edge_local);
    134 }
    135 
    136 /* No longer mark "edge" as a local edge.
    137  */
    138 static void clear_local(struct isl_sched_edge *edge)
    139 {
    140 	clear_type(edge, isl_edge_local);
    141 }
    142 
    143 /* Is "edge" marked as a coincidence edge?
    144  */
    145 static int is_coincidence(struct isl_sched_edge *edge)
    146 {
    147 	return isl_sched_edge_has_type(edge, isl_edge_coincidence);
    148 }
    149 
    150 /* Is "edge" marked as a condition edge?
    151  */
    152 int isl_sched_edge_is_condition(struct isl_sched_edge *edge)
    153 {
    154 	return isl_sched_edge_has_type(edge, isl_edge_condition);
    155 }
    156 
    157 /* Is "edge" marked as a conditional validity edge?
    158  */
    159 int isl_sched_edge_is_conditional_validity(struct isl_sched_edge *edge)
    160 {
    161 	return isl_sched_edge_has_type(edge, isl_edge_conditional_validity);
    162 }
    163 
    164 /* Is "edge" of a type that can appear multiple times between
    165  * the same pair of nodes?
    166  *
    167  * Condition edges and conditional validity edges may have tagged
    168  * dependence relations, in which case an edge is added for each
    169  * pair of tags.
    170  */
    171 static int is_multi_edge_type(struct isl_sched_edge *edge)
    172 {
    173 	return isl_sched_edge_is_condition(edge) ||
    174 		isl_sched_edge_is_conditional_validity(edge);
    175 }
    176 
    177 /* Initialize node_table based on the list of nodes.
    178  */
    179 static int graph_init_table(isl_ctx *ctx, struct isl_sched_graph *graph)
    180 {
    181 	int i;
    182 
    183 	graph->node_table = isl_hash_table_alloc(ctx, graph->n);
    184 	if (!graph->node_table)
    185 		return -1;
    186 
    187 	for (i = 0; i < graph->n; ++i) {
    188 		struct isl_hash_table_entry *entry;
    189 		uint32_t hash;
    190 
    191 		hash = isl_space_get_tuple_hash(graph->node[i].space);
    192 		entry = isl_hash_table_find(ctx, graph->node_table, hash,
    193 					    &node_has_tuples,
    194 					    graph->node[i].space, 1);
    195 		if (!entry)
    196 			return -1;
    197 		entry->data = &graph->node[i];
    198 	}
    199 
    200 	return 0;
    201 }
    202 
    203 /* Return a pointer to the node that lives within the given space,
    204  * an invalid node if there is no such node, or NULL in case of error.
    205  */
    206 struct isl_sched_node *isl_sched_graph_find_node(isl_ctx *ctx,
    207 	struct isl_sched_graph *graph, __isl_keep isl_space *space)
    208 {
    209 	struct isl_hash_table_entry *entry;
    210 	uint32_t hash;
    211 
    212 	if (!space)
    213 		return NULL;
    214 
    215 	hash = isl_space_get_tuple_hash(space);
    216 	entry = isl_hash_table_find(ctx, graph->node_table, hash,
    217 				    &node_has_tuples, space, 0);
    218 	if (!entry)
    219 		return NULL;
    220 	if (entry == isl_hash_table_entry_none)
    221 		return graph->node + graph->n;
    222 
    223 	return entry->data;
    224 }
    225 
    226 /* Is "node" a node in "graph"?
    227  */
    228 int isl_sched_graph_is_node(struct isl_sched_graph *graph,
    229 	struct isl_sched_node *node)
    230 {
    231 	return node && node >= &graph->node[0] && node < &graph->node[graph->n];
    232 }
    233 
    234 static isl_bool edge_has_src_and_dst(const void *entry, const void *val)
    235 {
    236 	const struct isl_sched_edge *edge = entry;
    237 	const struct isl_sched_edge *temp = val;
    238 
    239 	return isl_bool_ok(edge->src == temp->src && edge->dst == temp->dst);
    240 }
    241 
    242 /* Add the given edge to graph->edge_table[type].
    243  */
    244 static isl_stat graph_edge_table_add(isl_ctx *ctx,
    245 	struct isl_sched_graph *graph, enum isl_edge_type type,
    246 	struct isl_sched_edge *edge)
    247 {
    248 	struct isl_hash_table_entry *entry;
    249 	uint32_t hash;
    250 
    251 	hash = isl_hash_init();
    252 	hash = isl_hash_builtin(hash, edge->src);
    253 	hash = isl_hash_builtin(hash, edge->dst);
    254 	entry = isl_hash_table_find(ctx, graph->edge_table[type], hash,
    255 				    &edge_has_src_and_dst, edge, 1);
    256 	if (!entry)
    257 		return isl_stat_error;
    258 	entry->data = edge;
    259 
    260 	return isl_stat_ok;
    261 }
    262 
    263 /* Add "edge" to all relevant edge tables.
    264  * That is, for every type of the edge, add it to the corresponding table.
    265  */
    266 static isl_stat graph_edge_tables_add(isl_ctx *ctx,
    267 	struct isl_sched_graph *graph, struct isl_sched_edge *edge)
    268 {
    269 	enum isl_edge_type t;
    270 
    271 	for (t = isl_edge_first; t <= isl_edge_last; ++t) {
    272 		if (!isl_sched_edge_has_type(edge, t))
    273 			continue;
    274 		if (graph_edge_table_add(ctx, graph, t, edge) < 0)
    275 			return isl_stat_error;
    276 	}
    277 
    278 	return isl_stat_ok;
    279 }
    280 
    281 /* Allocate the edge_tables based on the maximal number of edges of
    282  * each type.
    283  */
    284 static int graph_init_edge_tables(isl_ctx *ctx, struct isl_sched_graph *graph)
    285 {
    286 	int i;
    287 
    288 	for (i = 0; i <= isl_edge_last; ++i) {
    289 		graph->edge_table[i] = isl_hash_table_alloc(ctx,
    290 							    graph->max_edge[i]);
    291 		if (!graph->edge_table[i])
    292 			return -1;
    293 	}
    294 
    295 	return 0;
    296 }
    297 
    298 /* If graph->edge_table[type] contains an edge from the given source
    299  * to the given destination, then return the hash table entry of this edge.
    300  * Otherwise, return NULL.
    301  */
    302 static struct isl_hash_table_entry *graph_find_edge_entry(
    303 	struct isl_sched_graph *graph,
    304 	enum isl_edge_type type,
    305 	struct isl_sched_node *src, struct isl_sched_node *dst)
    306 {
    307 	isl_ctx *ctx = isl_space_get_ctx(src->space);
    308 	uint32_t hash;
    309 	struct isl_sched_edge temp = { .src = src, .dst = dst };
    310 
    311 	hash = isl_hash_init();
    312 	hash = isl_hash_builtin(hash, temp.src);
    313 	hash = isl_hash_builtin(hash, temp.dst);
    314 	return isl_hash_table_find(ctx, graph->edge_table[type], hash,
    315 				    &edge_has_src_and_dst, &temp, 0);
    316 }
    317 
    318 
    319 /* If graph->edge_table[type] contains an edge from the given source
    320  * to the given destination, then return this edge.
    321  * Return "none" if no such edge can be found.
    322  * Return NULL on error.
    323  */
    324 static struct isl_sched_edge *graph_find_edge(struct isl_sched_graph *graph,
    325 	enum isl_edge_type type,
    326 	struct isl_sched_node *src, struct isl_sched_node *dst,
    327 	struct isl_sched_edge *none)
    328 {
    329 	struct isl_hash_table_entry *entry;
    330 
    331 	entry = graph_find_edge_entry(graph, type, src, dst);
    332 	if (!entry)
    333 		return NULL;
    334 	if (entry == isl_hash_table_entry_none)
    335 		return none;
    336 
    337 	return entry->data;
    338 }
    339 
    340 /* Check whether the dependence graph has an edge of the given type
    341  * between the given two nodes.
    342  */
    343 static isl_bool graph_has_edge(struct isl_sched_graph *graph,
    344 	enum isl_edge_type type,
    345 	struct isl_sched_node *src, struct isl_sched_node *dst)
    346 {
    347 	struct isl_sched_edge dummy;
    348 	struct isl_sched_edge *edge;
    349 	isl_bool empty;
    350 
    351 	edge = graph_find_edge(graph, type, src, dst, &dummy);
    352 	if (!edge)
    353 		return isl_bool_error;
    354 	if (edge == &dummy)
    355 		return isl_bool_false;
    356 
    357 	empty = isl_map_plain_is_empty(edge->map);
    358 
    359 	return isl_bool_not(empty);
    360 }
    361 
    362 /* Look for any edge with the same src, dst and map fields as "model".
    363  *
    364  * Return the matching edge if one can be found.
    365  * Return "model" if no matching edge is found.
    366  * Return NULL on error.
    367  */
    368 static struct isl_sched_edge *graph_find_matching_edge(
    369 	struct isl_sched_graph *graph, struct isl_sched_edge *model)
    370 {
    371 	enum isl_edge_type i;
    372 	struct isl_sched_edge *edge;
    373 
    374 	for (i = isl_edge_first; i <= isl_edge_last; ++i) {
    375 		int is_equal;
    376 
    377 		edge = graph_find_edge(graph, i, model->src, model->dst, model);
    378 		if (!edge)
    379 			return NULL;
    380 		if (edge == model)
    381 			continue;
    382 		is_equal = isl_map_plain_is_equal(model->map, edge->map);
    383 		if (is_equal < 0)
    384 			return NULL;
    385 		if (is_equal)
    386 			return edge;
    387 	}
    388 
    389 	return model;
    390 }
    391 
    392 /* Remove the given edge from all the edge_tables that refer to it.
    393  */
    394 static isl_stat graph_remove_edge(struct isl_sched_graph *graph,
    395 	struct isl_sched_edge *edge)
    396 {
    397 	isl_ctx *ctx = isl_map_get_ctx(edge->map);
    398 	enum isl_edge_type i;
    399 
    400 	for (i = isl_edge_first; i <= isl_edge_last; ++i) {
    401 		struct isl_hash_table_entry *entry;
    402 
    403 		entry = graph_find_edge_entry(graph, i, edge->src, edge->dst);
    404 		if (!entry)
    405 			return isl_stat_error;
    406 		if (entry == isl_hash_table_entry_none)
    407 			continue;
    408 		if (entry->data != edge)
    409 			continue;
    410 		isl_hash_table_remove(ctx, graph->edge_table[i], entry);
    411 	}
    412 
    413 	return isl_stat_ok;
    414 }
    415 
    416 /* Check whether the dependence graph has any edge
    417  * between the given two nodes.
    418  */
    419 static isl_bool graph_has_any_edge(struct isl_sched_graph *graph,
    420 	struct isl_sched_node *src, struct isl_sched_node *dst)
    421 {
    422 	enum isl_edge_type i;
    423 	isl_bool r;
    424 
    425 	for (i = isl_edge_first; i <= isl_edge_last; ++i) {
    426 		r = graph_has_edge(graph, i, src, dst);
    427 		if (r < 0 || r)
    428 			return r;
    429 	}
    430 
    431 	return r;
    432 }
    433 
    434 /* Check whether the dependence graph has a validity edge
    435  * between the given two nodes.
    436  *
    437  * Conditional validity edges are essentially validity edges that
    438  * can be ignored if the corresponding condition edges are iteration private.
    439  * Here, we are only checking for the presence of validity
    440  * edges, so we need to consider the conditional validity edges too.
    441  * In particular, this function is used during the detection
    442  * of strongly connected components and we cannot ignore
    443  * conditional validity edges during this detection.
    444  */
    445 isl_bool isl_sched_graph_has_validity_edge(struct isl_sched_graph *graph,
    446 	struct isl_sched_node *src, struct isl_sched_node *dst)
    447 {
    448 	isl_bool r;
    449 
    450 	r = graph_has_edge(graph, isl_edge_validity, src, dst);
    451 	if (r < 0 || r)
    452 		return r;
    453 
    454 	return graph_has_edge(graph, isl_edge_conditional_validity, src, dst);
    455 }
    456 
    457 /* Perform all the required memory allocations for a schedule graph "graph"
    458  * with "n_node" nodes and "n_edge" edge and initialize the corresponding
    459  * fields.
    460  */
    461 static isl_stat graph_alloc(isl_ctx *ctx, struct isl_sched_graph *graph,
    462 	int n_node, int n_edge)
    463 {
    464 	int i;
    465 
    466 	graph->n = n_node;
    467 	graph->n_edge = n_edge;
    468 	graph->node = isl_calloc_array(ctx, struct isl_sched_node, graph->n);
    469 	graph->sorted = isl_calloc_array(ctx, int, graph->n);
    470 	graph->region = isl_alloc_array(ctx,
    471 					struct isl_trivial_region, graph->n);
    472 	graph->edge = isl_calloc_array(ctx,
    473 					struct isl_sched_edge, graph->n_edge);
    474 
    475 	graph->intra_hmap = isl_map_to_basic_set_alloc(ctx, 2 * n_edge);
    476 	graph->intra_hmap_param = isl_map_to_basic_set_alloc(ctx, 2 * n_edge);
    477 	graph->inter_hmap = isl_map_to_basic_set_alloc(ctx, 2 * n_edge);
    478 
    479 	if (!graph->node || !graph->region || (graph->n_edge && !graph->edge) ||
    480 	    !graph->sorted)
    481 		return isl_stat_error;
    482 
    483 	for(i = 0; i < graph->n; ++i)
    484 		graph->sorted[i] = i;
    485 
    486 	return isl_stat_ok;
    487 }
    488 
    489 /* Free the memory associated to node "node" in "graph".
    490  * The "coincident" field is shared by nodes in a graph and its subgraph.
    491  * It therefore only needs to be freed for the original dependence graph,
    492  * i.e., one that is not the result of splitting.
    493  */
    494 static void clear_node(struct isl_sched_graph *graph,
    495 	struct isl_sched_node *node)
    496 {
    497 	isl_space_free(node->space);
    498 	isl_set_free(node->hull);
    499 	isl_multi_aff_free(node->compress);
    500 	isl_pw_multi_aff_free(node->decompress);
    501 	isl_mat_free(node->sched);
    502 	isl_map_free(node->sched_map);
    503 	isl_mat_free(node->indep);
    504 	isl_mat_free(node->vmap);
    505 	if (graph->root == graph)
    506 		free(node->coincident);
    507 	isl_multi_val_free(node->sizes);
    508 	isl_basic_set_free(node->bounds);
    509 	isl_vec_free(node->max);
    510 }
    511 
    512 void isl_sched_graph_free(isl_ctx *ctx, struct isl_sched_graph *graph)
    513 {
    514 	int i;
    515 
    516 	isl_map_to_basic_set_free(graph->intra_hmap);
    517 	isl_map_to_basic_set_free(graph->intra_hmap_param);
    518 	isl_map_to_basic_set_free(graph->inter_hmap);
    519 
    520 	if (graph->node)
    521 		for (i = 0; i < graph->n; ++i)
    522 			clear_node(graph, &graph->node[i]);
    523 	free(graph->node);
    524 	free(graph->sorted);
    525 	if (graph->edge)
    526 		for (i = 0; i < graph->n_edge; ++i) {
    527 			isl_map_free(graph->edge[i].map);
    528 			isl_union_map_free(graph->edge[i].tagged_condition);
    529 			isl_union_map_free(graph->edge[i].tagged_validity);
    530 		}
    531 	free(graph->edge);
    532 	free(graph->region);
    533 	for (i = 0; i <= isl_edge_last; ++i)
    534 		isl_hash_table_free(ctx, graph->edge_table[i]);
    535 	isl_hash_table_free(ctx, graph->node_table);
    536 	isl_basic_set_free(graph->lp);
    537 }
    538 
    539 /* For each "set" on which this function is called, increment
    540  * graph->n by one and update graph->maxvar.
    541  */
    542 static isl_stat init_n_maxvar(__isl_take isl_set *set, void *user)
    543 {
    544 	struct isl_sched_graph *graph = user;
    545 	isl_size nvar = isl_set_dim(set, isl_dim_set);
    546 
    547 	graph->n++;
    548 	if (nvar > graph->maxvar)
    549 		graph->maxvar = nvar;
    550 
    551 	isl_set_free(set);
    552 
    553 	if (nvar < 0)
    554 		return isl_stat_error;
    555 	return isl_stat_ok;
    556 }
    557 
    558 /* Compute the number of rows that should be allocated for the schedule.
    559  * In particular, we need one row for each variable or one row
    560  * for each basic map in the dependences.
    561  * Note that it is practically impossible to exhaust both
    562  * the number of dependences and the number of variables.
    563  */
    564 static isl_stat compute_max_row(struct isl_sched_graph *graph,
    565 	__isl_keep isl_schedule_constraints *sc)
    566 {
    567 	int n_edge;
    568 	isl_stat r;
    569 	isl_union_set *domain;
    570 
    571 	graph->n = 0;
    572 	graph->maxvar = 0;
    573 	domain = isl_schedule_constraints_get_domain(sc);
    574 	r = isl_union_set_foreach_set(domain, &init_n_maxvar, graph);
    575 	isl_union_set_free(domain);
    576 	if (r < 0)
    577 		return isl_stat_error;
    578 	n_edge = isl_schedule_constraints_n_basic_map(sc);
    579 	if (n_edge < 0)
    580 		return isl_stat_error;
    581 	graph->max_row = n_edge + graph->maxvar;
    582 
    583 	return isl_stat_ok;
    584 }
    585 
    586 /* Does "bset" have any defining equalities for its set variables?
    587  */
    588 static isl_bool has_any_defining_equality(__isl_keep isl_basic_set *bset)
    589 {
    590 	int i;
    591 	isl_size n;
    592 
    593 	n = isl_basic_set_dim(bset, isl_dim_set);
    594 	if (n < 0)
    595 		return isl_bool_error;
    596 
    597 	for (i = 0; i < n; ++i) {
    598 		isl_bool has;
    599 
    600 		has = isl_basic_set_has_defining_equality(bset, isl_dim_set, i,
    601 							NULL);
    602 		if (has < 0 || has)
    603 			return has;
    604 	}
    605 
    606 	return isl_bool_false;
    607 }
    608 
    609 /* Set the entries of node->max to the value of the schedule_max_coefficient
    610  * option, if set.
    611  */
    612 static isl_stat set_max_coefficient(isl_ctx *ctx, struct isl_sched_node *node)
    613 {
    614 	int max;
    615 
    616 	max = isl_options_get_schedule_max_coefficient(ctx);
    617 	if (max == -1)
    618 		return isl_stat_ok;
    619 
    620 	node->max = isl_vec_alloc(ctx, node->nvar);
    621 	node->max = isl_vec_set_si(node->max, max);
    622 	if (!node->max)
    623 		return isl_stat_error;
    624 
    625 	return isl_stat_ok;
    626 }
    627 
    628 /* Set the entries of node->max to the minimum of the schedule_max_coefficient
    629  * option (if set) and half of the minimum of the sizes in the other
    630  * dimensions.  Round up when computing the half such that
    631  * if the minimum of the sizes is one, half of the size is taken to be one
    632  * rather than zero.
    633  * If the global minimum is unbounded (i.e., if both
    634  * the schedule_max_coefficient is not set and the sizes in the other
    635  * dimensions are unbounded), then store a negative value.
    636  * If the schedule coefficient is close to the size of the instance set
    637  * in another dimension, then the schedule may represent a loop
    638  * coalescing transformation (especially if the coefficient
    639  * in that other dimension is one).  Forcing the coefficient to be
    640  * smaller than or equal to half the minimal size should avoid this
    641  * situation.
    642  */
    643 static isl_stat compute_max_coefficient(isl_ctx *ctx,
    644 	struct isl_sched_node *node)
    645 {
    646 	int max;
    647 	int i, j;
    648 	isl_vec *v;
    649 
    650 	max = isl_options_get_schedule_max_coefficient(ctx);
    651 	v = isl_vec_alloc(ctx, node->nvar);
    652 	if (!v)
    653 		return isl_stat_error;
    654 
    655 	for (i = 0; i < node->nvar; ++i) {
    656 		isl_int_set_si(v->el[i], max);
    657 		isl_int_mul_si(v->el[i], v->el[i], 2);
    658 	}
    659 
    660 	for (i = 0; i < node->nvar; ++i) {
    661 		isl_val *size;
    662 
    663 		size = isl_multi_val_get_val(node->sizes, i);
    664 		if (!size)
    665 			goto error;
    666 		if (!isl_val_is_int(size)) {
    667 			isl_val_free(size);
    668 			continue;
    669 		}
    670 		for (j = 0; j < node->nvar; ++j) {
    671 			if (j == i)
    672 				continue;
    673 			if (isl_int_is_neg(v->el[j]) ||
    674 			    isl_int_gt(v->el[j], size->n))
    675 				isl_int_set(v->el[j], size->n);
    676 		}
    677 		isl_val_free(size);
    678 	}
    679 
    680 	for (i = 0; i < node->nvar; ++i)
    681 		isl_int_cdiv_q_ui(v->el[i], v->el[i], 2);
    682 
    683 	node->max = v;
    684 	return isl_stat_ok;
    685 error:
    686 	isl_vec_free(v);
    687 	return isl_stat_error;
    688 }
    689 
    690 /* Construct an identifier for node "node", which will represent "set".
    691  * The name of the identifier is either "compressed" or
    692  * "compressed_<name>", with <name> the name of the space of "set".
    693  * The user pointer of the identifier points to "node".
    694  */
    695 static __isl_give isl_id *construct_compressed_id(__isl_keep isl_set *set,
    696 	struct isl_sched_node *node)
    697 {
    698 	isl_bool has_name;
    699 	isl_ctx *ctx;
    700 	isl_id *id;
    701 	isl_printer *p;
    702 	const char *name;
    703 	char *id_name;
    704 
    705 	has_name = isl_set_has_tuple_name(set);
    706 	if (has_name < 0)
    707 		return NULL;
    708 
    709 	ctx = isl_set_get_ctx(set);
    710 	if (!has_name)
    711 		return isl_id_alloc(ctx, "compressed", node);
    712 
    713 	p = isl_printer_to_str(ctx);
    714 	name = isl_set_get_tuple_name(set);
    715 	p = isl_printer_print_str(p, "compressed_");
    716 	p = isl_printer_print_str(p, name);
    717 	id_name = isl_printer_get_str(p);
    718 	isl_printer_free(p);
    719 
    720 	id = isl_id_alloc(ctx, id_name, node);
    721 	free(id_name);
    722 
    723 	return id;
    724 }
    725 
    726 /* Construct a map that isolates the variable in position "pos" in "set".
    727  *
    728  * That is, construct
    729  *
    730  *	[i_0, ..., i_pos-1, i_pos+1, ...] -> [i_pos]
    731  */
    732 static __isl_give isl_map *isolate(__isl_take isl_set *set, int pos)
    733 {
    734 	isl_map *map;
    735 
    736 	map = isl_set_project_onto_map(set, isl_dim_set, pos, 1);
    737 	map = isl_map_project_out(map, isl_dim_in, pos, 1);
    738 	return map;
    739 }
    740 
    741 /* Compute and return the size of "set" in dimension "dim".
    742  * The size is taken to be the difference in values for that variable
    743  * for fixed values of the other variables.
    744  * This assumes that "set" is convex.
    745  * In particular, the variable is first isolated from the other variables
    746  * in the range of a map
    747  *
    748  *	[i_0, ..., i_dim-1, i_dim+1, ...] -> [i_dim]
    749  *
    750  * and then duplicated
    751  *
    752  *	[i_0, ..., i_dim-1, i_dim+1, ...] -> [[i_dim] -> [i_dim']]
    753  *
    754  * The shared variables are then projected out and the maximal value
    755  * of i_dim' - i_dim is computed.
    756  */
    757 static __isl_give isl_val *compute_size(__isl_take isl_set *set, int dim)
    758 {
    759 	isl_map *map;
    760 	isl_local_space *ls;
    761 	isl_aff *obj;
    762 	isl_val *v;
    763 
    764 	map = isolate(set, dim);
    765 	map = isl_map_range_product(map, isl_map_copy(map));
    766 	map = isl_set_unwrap(isl_map_range(map));
    767 	set = isl_map_deltas(map);
    768 	ls = isl_local_space_from_space(isl_set_get_space(set));
    769 	obj = isl_aff_var_on_domain(ls, isl_dim_set, 0);
    770 	v = isl_set_max_val(set, obj);
    771 	isl_aff_free(obj);
    772 	isl_set_free(set);
    773 
    774 	return v;
    775 }
    776 
    777 /* Perform a compression on "node" where "hull" represents the constraints
    778  * that were used to derive the compression, while "compress" and
    779  * "decompress" map the original space to the compressed space and
    780  * vice versa.
    781  *
    782  * If "node" was not compressed already, then simply store
    783  * the compression information.
    784  * Otherwise the "original" space is actually the result
    785  * of a previous compression, which is then combined
    786  * with the present compression.
    787  *
    788  * The dimensionality of the compressed domain is also adjusted.
    789  * Other information, such as the sizes and the maximal coefficient values,
    790  * has not been computed yet and therefore does not need to be adjusted.
    791  */
    792 static isl_stat compress_node(struct isl_sched_node *node,
    793 	__isl_take isl_set *hull, __isl_take isl_multi_aff *compress,
    794 	__isl_take isl_pw_multi_aff *decompress)
    795 {
    796 	node->nvar = isl_multi_aff_dim(compress, isl_dim_out);
    797 	if (!node->compressed) {
    798 		node->compressed = 1;
    799 		node->hull = hull;
    800 		node->compress = compress;
    801 		node->decompress = decompress;
    802 	} else {
    803 		hull = isl_set_preimage_multi_aff(hull,
    804 					    isl_multi_aff_copy(node->compress));
    805 		node->hull = isl_set_intersect(node->hull, hull);
    806 		node->compress = isl_multi_aff_pullback_multi_aff(
    807 						compress, node->compress);
    808 		node->decompress = isl_pw_multi_aff_pullback_pw_multi_aff(
    809 						node->decompress, decompress);
    810 	}
    811 
    812 	if (!node->hull || !node->compress || !node->decompress)
    813 		return isl_stat_error;
    814 
    815 	return isl_stat_ok;
    816 }
    817 
    818 /* Given that dimension "pos" in "set" has a fixed value
    819  * in terms of the other dimensions, (further) compress "node"
    820  * by projecting out this dimension.
    821  * "set" may be the result of a previous compression.
    822  * "uncompressed" is the original domain (without compression).
    823  *
    824  * The compression function simply projects out the dimension.
    825  * The decompression function adds back the dimension
    826  * in the right position as an expression of the other dimensions
    827  * derived from "set".
    828  * As in extract_node, the compressed space has an identifier
    829  * that references "node" such that each compressed space is unique and
    830  * such that the node can be recovered from the compressed space.
    831  *
    832  * The constraint removed through the compression is added to the "hull"
    833  * such that only edges that relate to the original domains
    834  * are taken into account.
    835  * In particular, it is obtained by composing compression and decompression and
    836  * taking the relation among the variables in the range.
    837  */
    838 static isl_stat project_out_fixed(struct isl_sched_node *node,
    839 	__isl_keep isl_set *uncompressed, __isl_take isl_set *set, int pos)
    840 {
    841 	isl_id *id;
    842 	isl_space *space;
    843 	isl_set *domain;
    844 	isl_map *map;
    845 	isl_multi_aff *compress;
    846 	isl_pw_multi_aff *decompress, *pma;
    847 	isl_multi_pw_aff *mpa;
    848 	isl_set *hull;
    849 
    850 	map = isolate(isl_set_copy(set), pos);
    851 	pma = isl_pw_multi_aff_from_map(map);
    852 	domain = isl_pw_multi_aff_domain(isl_pw_multi_aff_copy(pma));
    853 	pma = isl_pw_multi_aff_gist(pma, domain);
    854 	space = isl_pw_multi_aff_get_domain_space(pma);
    855 	mpa = isl_multi_pw_aff_identity(isl_space_map_from_set(space));
    856 	mpa = isl_multi_pw_aff_range_splice(mpa, pos,
    857 				    isl_multi_pw_aff_from_pw_multi_aff(pma));
    858 	decompress = isl_pw_multi_aff_from_multi_pw_aff(mpa);
    859 	space = isl_set_get_space(set);
    860 	compress = isl_multi_aff_project_out_map(space, isl_dim_set, pos, 1);
    861 	id = construct_compressed_id(uncompressed, node);
    862 	compress = isl_multi_aff_set_tuple_id(compress, isl_dim_out, id);
    863 	space = isl_space_reverse(isl_multi_aff_get_space(compress));
    864 	decompress = isl_pw_multi_aff_reset_space(decompress, space);
    865 	pma = isl_pw_multi_aff_pullback_multi_aff(
    866 	    isl_pw_multi_aff_copy(decompress), isl_multi_aff_copy(compress));
    867 	hull = isl_map_range(isl_map_from_pw_multi_aff(pma));
    868 
    869 	isl_set_free(set);
    870 
    871 	return compress_node(node, hull, compress, decompress);
    872 }
    873 
    874 /* Compute the size of the compressed domain in each dimension and
    875  * store the results in node->sizes.
    876  * "uncompressed" is the original domain (without compression).
    877  *
    878  * First compress the domain if needed and then compute the size
    879  * in each direction.
    880  * If the domain is not convex, then the sizes are computed
    881  * on a convex superset in order to avoid picking up sizes
    882  * that are valid for the individual disjuncts, but not for
    883  * the domain as a whole.
    884  *
    885  * If any of the sizes turns out to be zero, then this means
    886  * that this dimension has a fixed value in terms of
    887  * the other dimensions.  Perform an (extra) compression
    888  * to remove this dimension.
    889  */
    890 static isl_stat compute_sizes(struct isl_sched_node *node,
    891 	__isl_keep isl_set *uncompressed)
    892 {
    893 	int j;
    894 	isl_size n;
    895 	isl_multi_val *mv;
    896 	isl_set *set = isl_set_copy(uncompressed);
    897 
    898 	if (node->compressed)
    899 		set = isl_set_preimage_pw_multi_aff(set,
    900 				    isl_pw_multi_aff_copy(node->decompress));
    901 	set = isl_set_from_basic_set(isl_set_simple_hull(set));
    902 	mv = isl_multi_val_zero(isl_set_get_space(set));
    903 	n = isl_set_dim(set, isl_dim_set);
    904 	if (n < 0)
    905 		mv = isl_multi_val_free(mv);
    906 	for (j = 0; j < n; ++j) {
    907 		isl_bool is_zero;
    908 		isl_val *v;
    909 
    910 		v = compute_size(isl_set_copy(set), j);
    911 		is_zero = isl_val_is_zero(v);
    912 		mv = isl_multi_val_set_val(mv, j, v);
    913 		if (is_zero >= 0 && is_zero) {
    914 			isl_multi_val_free(mv);
    915 			if (project_out_fixed(node, uncompressed, set, j) < 0)
    916 				return isl_stat_error;
    917 			return compute_sizes(node, uncompressed);
    918 		}
    919 	}
    920 	node->sizes = mv;
    921 	isl_set_free(set);
    922 	if (!node->sizes)
    923 		return isl_stat_error;
    924 	return isl_stat_ok;
    925 }
    926 
    927 /* Compute the size of the instance set "set" of "node", after compression,
    928  * as well as bounds on the corresponding coefficients, if needed.
    929  *
    930  * The sizes are needed when the schedule_treat_coalescing option is set.
    931  * The bounds are needed when the schedule_treat_coalescing option or
    932  * the schedule_max_coefficient option is set.
    933  *
    934  * If the schedule_treat_coalescing option is not set, then at most
    935  * the bounds need to be set and this is done in set_max_coefficient.
    936  * Otherwise, compute the size of the compressed domain
    937  * in each direction and store the results in node->size.
    938  * Finally, set the bounds on the coefficients based on the sizes
    939  * and the schedule_max_coefficient option in compute_max_coefficient.
    940  */
    941 static isl_stat compute_sizes_and_max(isl_ctx *ctx, struct isl_sched_node *node,
    942 	__isl_take isl_set *set)
    943 {
    944 	isl_stat r;
    945 
    946 	if (!isl_options_get_schedule_treat_coalescing(ctx)) {
    947 		isl_set_free(set);
    948 		return set_max_coefficient(ctx, node);
    949 	}
    950 
    951 	r = compute_sizes(node, set);
    952 	isl_set_free(set);
    953 	if (r < 0)
    954 		return isl_stat_error;
    955 	return compute_max_coefficient(ctx, node);
    956 }
    957 
    958 /* Add a new node to the graph representing the given instance set.
    959  * "nvar" is the (possibly compressed) number of variables and
    960  * may be smaller than then number of set variables in "set"
    961  * if "compressed" is set.
    962  * If "compressed" is set, then "hull" represents the constraints
    963  * that were used to derive the compression, while "compress" and
    964  * "decompress" map the original space to the compressed space and
    965  * vice versa.
    966  * If "compressed" is not set, then "hull", "compress" and "decompress"
    967  * should be NULL.
    968  *
    969  * Compute the size of the instance set and bounds on the coefficients,
    970  * if needed.
    971  */
    972 static isl_stat add_node(struct isl_sched_graph *graph,
    973 	__isl_take isl_set *set, int nvar, int compressed,
    974 	__isl_take isl_set *hull, __isl_take isl_multi_aff *compress,
    975 	__isl_take isl_pw_multi_aff *decompress)
    976 {
    977 	isl_size nparam;
    978 	isl_ctx *ctx;
    979 	isl_mat *sched;
    980 	isl_space *space;
    981 	int *coincident;
    982 	struct isl_sched_node *node;
    983 
    984 	nparam = isl_set_dim(set, isl_dim_param);
    985 	if (nparam < 0)
    986 		goto error;
    987 
    988 	ctx = isl_set_get_ctx(set);
    989 	if (!ctx->opt->schedule_parametric)
    990 		nparam = 0;
    991 	sched = isl_mat_alloc(ctx, 0, 1 + nparam + nvar);
    992 	node = &graph->node[graph->n];
    993 	graph->n++;
    994 	space = isl_set_get_space(set);
    995 	node->space = space;
    996 	node->nvar = nvar;
    997 	node->nparam = nparam;
    998 	node->sched = sched;
    999 	node->sched_map = NULL;
   1000 	coincident = isl_calloc_array(ctx, int, graph->max_row);
   1001 	node->coincident = coincident;
   1002 	node->compressed = compressed;
   1003 	node->hull = hull;
   1004 	node->compress = compress;
   1005 	node->decompress = decompress;
   1006 	if (compute_sizes_and_max(ctx, node, set) < 0)
   1007 		return isl_stat_error;
   1008 
   1009 	if (!space || !sched || (graph->max_row && !coincident))
   1010 		return isl_stat_error;
   1011 	if (compressed && (!hull || !compress || !decompress))
   1012 		return isl_stat_error;
   1013 
   1014 	return isl_stat_ok;
   1015 error:
   1016 	isl_set_free(set);
   1017 	isl_set_free(hull);
   1018 	isl_multi_aff_free(compress);
   1019 	isl_pw_multi_aff_free(decompress);
   1020 	return isl_stat_error;
   1021 }
   1022 
   1023 /* Add a new node to the graph representing the given set.
   1024  *
   1025  * If any of the set variables is defined by an equality, then
   1026  * we perform variable compression such that we can perform
   1027  * the scheduling on the compressed domain.
   1028  * In this case, an identifier is used that references the new node
   1029  * such that each compressed space is unique and
   1030  * such that the node can be recovered from the compressed space.
   1031  */
   1032 static isl_stat extract_node(__isl_take isl_set *set, void *user)
   1033 {
   1034 	isl_size nvar;
   1035 	isl_bool has_equality;
   1036 	isl_id *id;
   1037 	isl_basic_set *hull;
   1038 	isl_set *hull_set;
   1039 	isl_morph *morph;
   1040 	isl_multi_aff *compress, *decompress_ma;
   1041 	isl_pw_multi_aff *decompress;
   1042 	struct isl_sched_graph *graph = user;
   1043 
   1044 	hull = isl_set_affine_hull(isl_set_copy(set));
   1045 	hull = isl_basic_set_remove_divs(hull);
   1046 	nvar = isl_set_dim(set, isl_dim_set);
   1047 	has_equality = has_any_defining_equality(hull);
   1048 
   1049 	if (nvar < 0 || has_equality < 0)
   1050 		goto error;
   1051 	if (!has_equality) {
   1052 		isl_basic_set_free(hull);
   1053 		return add_node(graph, set, nvar, 0, NULL, NULL, NULL);
   1054 	}
   1055 
   1056 	id = construct_compressed_id(set, &graph->node[graph->n]);
   1057 	morph = isl_basic_set_variable_compression_with_id(hull, id);
   1058 	isl_id_free(id);
   1059 	nvar = isl_morph_ran_dim(morph, isl_dim_set);
   1060 	if (nvar < 0)
   1061 		set = isl_set_free(set);
   1062 	compress = isl_morph_get_var_multi_aff(morph);
   1063 	morph = isl_morph_inverse(morph);
   1064 	decompress_ma = isl_morph_get_var_multi_aff(morph);
   1065 	decompress = isl_pw_multi_aff_from_multi_aff(decompress_ma);
   1066 	isl_morph_free(morph);
   1067 
   1068 	hull_set = isl_set_from_basic_set(hull);
   1069 	return add_node(graph, set, nvar, 1, hull_set, compress, decompress);
   1070 error:
   1071 	isl_basic_set_free(hull);
   1072 	isl_set_free(set);
   1073 	return isl_stat_error;
   1074 }
   1075 
   1076 struct isl_extract_edge_data {
   1077 	isl_schedule_constraints *sc;
   1078 	enum isl_edge_type type;
   1079 	struct isl_sched_graph *graph;
   1080 };
   1081 
   1082 /* Merge edge2 into edge1, freeing the contents of edge2.
   1083  * Return 0 on success and -1 on failure.
   1084  *
   1085  * edge1 and edge2 are assumed to have the same value for the map field.
   1086  */
   1087 static int merge_edge(struct isl_sched_edge *edge1,
   1088 	struct isl_sched_edge *edge2)
   1089 {
   1090 	edge1->types |= edge2->types;
   1091 	isl_map_free(edge2->map);
   1092 
   1093 	if (isl_sched_edge_is_condition(edge2)) {
   1094 		if (!edge1->tagged_condition)
   1095 			edge1->tagged_condition = edge2->tagged_condition;
   1096 		else
   1097 			edge1->tagged_condition =
   1098 				isl_union_map_union(edge1->tagged_condition,
   1099 						    edge2->tagged_condition);
   1100 	}
   1101 
   1102 	if (isl_sched_edge_is_conditional_validity(edge2)) {
   1103 		if (!edge1->tagged_validity)
   1104 			edge1->tagged_validity = edge2->tagged_validity;
   1105 		else
   1106 			edge1->tagged_validity =
   1107 				isl_union_map_union(edge1->tagged_validity,
   1108 						    edge2->tagged_validity);
   1109 	}
   1110 
   1111 	if (isl_sched_edge_is_condition(edge2) && !edge1->tagged_condition)
   1112 		return -1;
   1113 	if (isl_sched_edge_is_conditional_validity(edge2) &&
   1114 	    !edge1->tagged_validity)
   1115 		return -1;
   1116 
   1117 	return 0;
   1118 }
   1119 
   1120 /* Insert dummy tags in domain and range of "map".
   1121  *
   1122  * In particular, if "map" is of the form
   1123  *
   1124  *	A -> B
   1125  *
   1126  * then return
   1127  *
   1128  *	[A -> dummy_tag] -> [B -> dummy_tag]
   1129  *
   1130  * where the dummy_tags are identical and equal to any dummy tags
   1131  * introduced by any other call to this function.
   1132  */
   1133 static __isl_give isl_map *insert_dummy_tags(__isl_take isl_map *map)
   1134 {
   1135 	static char dummy;
   1136 	isl_ctx *ctx;
   1137 	isl_id *id;
   1138 	isl_space *space;
   1139 	isl_set *domain, *range;
   1140 
   1141 	ctx = isl_map_get_ctx(map);
   1142 
   1143 	id = isl_id_alloc(ctx, NULL, &dummy);
   1144 	space = isl_space_params(isl_map_get_space(map));
   1145 	space = isl_space_set_from_params(space);
   1146 	space = isl_space_set_tuple_id(space, isl_dim_set, id);
   1147 	space = isl_space_map_from_set(space);
   1148 
   1149 	domain = isl_map_wrap(map);
   1150 	range = isl_map_wrap(isl_map_universe(space));
   1151 	map = isl_map_from_domain_and_range(domain, range);
   1152 	map = isl_map_zip(map);
   1153 
   1154 	return map;
   1155 }
   1156 
   1157 /* Given that at least one of "src" or "dst" is compressed, return
   1158  * a map between the spaces of these nodes restricted to the affine
   1159  * hull that was used in the compression.
   1160  */
   1161 static __isl_give isl_map *extract_hull(struct isl_sched_node *src,
   1162 	struct isl_sched_node *dst)
   1163 {
   1164 	isl_set *dom, *ran;
   1165 
   1166 	if (src->compressed)
   1167 		dom = isl_set_copy(src->hull);
   1168 	else
   1169 		dom = isl_set_universe(isl_space_copy(src->space));
   1170 	if (dst->compressed)
   1171 		ran = isl_set_copy(dst->hull);
   1172 	else
   1173 		ran = isl_set_universe(isl_space_copy(dst->space));
   1174 
   1175 	return isl_map_from_domain_and_range(dom, ran);
   1176 }
   1177 
   1178 /* Intersect the domains of the nested relations in domain and range
   1179  * of "tagged" with "map".
   1180  */
   1181 static __isl_give isl_map *map_intersect_domains(__isl_take isl_map *tagged,
   1182 	__isl_keep isl_map *map)
   1183 {
   1184 	isl_set *set;
   1185 
   1186 	tagged = isl_map_zip(tagged);
   1187 	set = isl_map_wrap(isl_map_copy(map));
   1188 	tagged = isl_map_intersect_domain(tagged, set);
   1189 	tagged = isl_map_zip(tagged);
   1190 	return tagged;
   1191 }
   1192 
   1193 /* Return a pointer to the node that lives in the domain space of "map",
   1194  * an invalid node if there is no such node, or NULL in case of error.
   1195  */
   1196 static struct isl_sched_node *find_domain_node(isl_ctx *ctx,
   1197 	struct isl_sched_graph *graph, __isl_keep isl_map *map)
   1198 {
   1199 	struct isl_sched_node *node;
   1200 	isl_space *space;
   1201 
   1202 	space = isl_space_domain(isl_map_get_space(map));
   1203 	node = isl_sched_graph_find_node(ctx, graph, space);
   1204 	isl_space_free(space);
   1205 
   1206 	return node;
   1207 }
   1208 
   1209 /* Return a pointer to the node that lives in the range space of "map",
   1210  * an invalid node if there is no such node, or NULL in case of error.
   1211  */
   1212 static struct isl_sched_node *find_range_node(isl_ctx *ctx,
   1213 	struct isl_sched_graph *graph, __isl_keep isl_map *map)
   1214 {
   1215 	struct isl_sched_node *node;
   1216 	isl_space *space;
   1217 
   1218 	space = isl_space_range(isl_map_get_space(map));
   1219 	node = isl_sched_graph_find_node(ctx, graph, space);
   1220 	isl_space_free(space);
   1221 
   1222 	return node;
   1223 }
   1224 
   1225 /* Refrain from adding a new edge based on "map".
   1226  * Instead, just free the map.
   1227  * "tagged" is either a copy of "map" with additional tags or NULL.
   1228  */
   1229 static isl_stat skip_edge(__isl_take isl_map *map, __isl_take isl_map *tagged)
   1230 {
   1231 	isl_map_free(map);
   1232 	isl_map_free(tagged);
   1233 
   1234 	return isl_stat_ok;
   1235 }
   1236 
   1237 /* Add a new edge to the graph based on the given map
   1238  * and add it to data->graph->edge_table[data->type].
   1239  * If a dependence relation of a given type happens to be identical
   1240  * to one of the dependence relations of a type that was added before,
   1241  * then we don't create a new edge, but instead mark the original edge
   1242  * as also representing a dependence of the current type.
   1243  *
   1244  * Edges of type isl_edge_condition or isl_edge_conditional_validity
   1245  * may be specified as "tagged" dependence relations.  That is, "map"
   1246  * may contain elements (i -> a) -> (j -> b), where i -> j denotes
   1247  * the dependence on iterations and a and b are tags.
   1248  * edge->map is set to the relation containing the elements i -> j,
   1249  * while edge->tagged_condition and edge->tagged_validity contain
   1250  * the union of all the "map" relations
   1251  * for which extract_edge is called that result in the same edge->map.
   1252  *
   1253  * Compute the gist with respect to the context.
   1254  * This may remove some constraints on the parameters or
   1255  * eliminate some parts of the dependence relation
   1256  * that are not relevant on the context.
   1257  *
   1258  * If the source or the destination node is compressed, then
   1259  * intersect both "map" and "tagged" with the constraints that
   1260  * were used to construct the compression.
   1261  * This ensures that there are no schedule constraints defined
   1262  * outside of these domains, while the scheduler no longer has
   1263  * any control over those outside parts.
   1264  */
   1265 static isl_stat extract_edge(__isl_take isl_map *map, void *user)
   1266 {
   1267 	isl_bool empty;
   1268 	isl_ctx *ctx = isl_map_get_ctx(map);
   1269 	struct isl_extract_edge_data *data = user;
   1270 	struct isl_sched_graph *graph = data->graph;
   1271 	struct isl_sched_node *src, *dst;
   1272 	struct isl_sched_edge *edge;
   1273 	isl_set *context;
   1274 	isl_map *tagged = NULL;
   1275 	isl_schedule_constraints *sc = data->sc;
   1276 
   1277 	if (data->type == isl_edge_condition ||
   1278 	    data->type == isl_edge_conditional_validity) {
   1279 		if (isl_map_can_zip(map)) {
   1280 			tagged = isl_map_copy(map);
   1281 			map = isl_set_unwrap(isl_map_domain(isl_map_zip(map)));
   1282 		} else {
   1283 			tagged = insert_dummy_tags(isl_map_copy(map));
   1284 		}
   1285 	}
   1286 
   1287 	src = find_domain_node(ctx, graph, map);
   1288 	dst = find_range_node(ctx, graph, map);
   1289 
   1290 	if (!src || !dst)
   1291 		goto error;
   1292 	if (!isl_sched_graph_is_node(graph, src) ||
   1293 	    !isl_sched_graph_is_node(graph, dst))
   1294 		return skip_edge(map, tagged);
   1295 
   1296 	context = isl_schedule_constraints_get_context(sc);
   1297 	map = isl_map_gist_params(map, context);
   1298 
   1299 	if (src->compressed || dst->compressed) {
   1300 		isl_map *hull;
   1301 		hull = extract_hull(src, dst);
   1302 		if (tagged)
   1303 			tagged = map_intersect_domains(tagged, hull);
   1304 		map = isl_map_intersect(map, hull);
   1305 	}
   1306 
   1307 	empty = isl_map_plain_is_empty(map);
   1308 	if (empty < 0)
   1309 		goto error;
   1310 	if (empty)
   1311 		return skip_edge(map, tagged);
   1312 
   1313 	graph->edge[graph->n_edge].src = src;
   1314 	graph->edge[graph->n_edge].dst = dst;
   1315 	graph->edge[graph->n_edge].map = map;
   1316 	graph->edge[graph->n_edge].types = 0;
   1317 	graph->edge[graph->n_edge].tagged_condition = NULL;
   1318 	graph->edge[graph->n_edge].tagged_validity = NULL;
   1319 	set_type(&graph->edge[graph->n_edge], data->type);
   1320 	if (data->type == isl_edge_condition)
   1321 		graph->edge[graph->n_edge].tagged_condition =
   1322 					isl_union_map_from_map(tagged);
   1323 	if (data->type == isl_edge_conditional_validity)
   1324 		graph->edge[graph->n_edge].tagged_validity =
   1325 					isl_union_map_from_map(tagged);
   1326 
   1327 	edge = graph_find_matching_edge(graph, &graph->edge[graph->n_edge]);
   1328 	if (!edge) {
   1329 		graph->n_edge++;
   1330 		return isl_stat_error;
   1331 	}
   1332 	if (edge == &graph->edge[graph->n_edge])
   1333 		return graph_edge_table_add(ctx, graph, data->type,
   1334 				    &graph->edge[graph->n_edge++]);
   1335 
   1336 	if (merge_edge(edge, &graph->edge[graph->n_edge]) < 0)
   1337 		return isl_stat_error;
   1338 
   1339 	return graph_edge_table_add(ctx, graph, data->type, edge);
   1340 error:
   1341 	isl_map_free(map);
   1342 	isl_map_free(tagged);
   1343 	return isl_stat_error;
   1344 }
   1345 
   1346 /* Initialize the schedule graph "graph" from the schedule constraints "sc".
   1347  *
   1348  * The context is included in the domain before the nodes of
   1349  * the graphs are extracted in order to be able to exploit
   1350  * any possible additional equalities.
   1351  * Note that this intersection is only performed locally here.
   1352  */
   1353 isl_stat isl_sched_graph_init(struct isl_sched_graph *graph,
   1354 	__isl_keep isl_schedule_constraints *sc)
   1355 {
   1356 	isl_ctx *ctx;
   1357 	isl_union_set *domain;
   1358 	isl_union_map *c;
   1359 	struct isl_extract_edge_data data = { sc };
   1360 	enum isl_edge_type i;
   1361 	isl_stat r;
   1362 	isl_size n;
   1363 
   1364 	if (!sc)
   1365 		return isl_stat_error;
   1366 
   1367 	ctx = isl_schedule_constraints_get_ctx(sc);
   1368 
   1369 	domain = isl_schedule_constraints_get_domain(sc);
   1370 	n = isl_union_set_n_set(domain);
   1371 	graph->n = n;
   1372 	isl_union_set_free(domain);
   1373 	if (n < 0)
   1374 		return isl_stat_error;
   1375 
   1376 	n = isl_schedule_constraints_n_map(sc);
   1377 	if (n < 0 || graph_alloc(ctx, graph, graph->n, n) < 0)
   1378 		return isl_stat_error;
   1379 
   1380 	if (compute_max_row(graph, sc) < 0)
   1381 		return isl_stat_error;
   1382 	graph->root = graph;
   1383 	graph->n = 0;
   1384 	domain = isl_schedule_constraints_get_domain(sc);
   1385 	domain = isl_union_set_intersect_params(domain,
   1386 				    isl_schedule_constraints_get_context(sc));
   1387 	r = isl_union_set_foreach_set(domain, &extract_node, graph);
   1388 	isl_union_set_free(domain);
   1389 	if (r < 0)
   1390 		return isl_stat_error;
   1391 	if (graph_init_table(ctx, graph) < 0)
   1392 		return isl_stat_error;
   1393 	for (i = isl_edge_first; i <= isl_edge_last; ++i) {
   1394 		isl_size n;
   1395 
   1396 		c = isl_schedule_constraints_get(sc, i);
   1397 		n = isl_union_map_n_map(c);
   1398 		graph->max_edge[i] = n;
   1399 		isl_union_map_free(c);
   1400 		if (n < 0)
   1401 			return isl_stat_error;
   1402 	}
   1403 	if (graph_init_edge_tables(ctx, graph) < 0)
   1404 		return isl_stat_error;
   1405 	graph->n_edge = 0;
   1406 	data.graph = graph;
   1407 	for (i = isl_edge_first; i <= isl_edge_last; ++i) {
   1408 		isl_stat r;
   1409 
   1410 		data.type = i;
   1411 		c = isl_schedule_constraints_get(sc, i);
   1412 		r = isl_union_map_foreach_map(c, &extract_edge, &data);
   1413 		isl_union_map_free(c);
   1414 		if (r < 0)
   1415 			return isl_stat_error;
   1416 	}
   1417 
   1418 	return isl_stat_ok;
   1419 }
   1420 
   1421 /* Check whether there is any dependence from node[j] to node[i]
   1422  * or from node[i] to node[j].
   1423  */
   1424 static isl_bool node_follows_weak(int i, int j, void *user)
   1425 {
   1426 	isl_bool f;
   1427 	struct isl_sched_graph *graph = user;
   1428 
   1429 	f = graph_has_any_edge(graph, &graph->node[j], &graph->node[i]);
   1430 	if (f < 0 || f)
   1431 		return f;
   1432 	return graph_has_any_edge(graph, &graph->node[i], &graph->node[j]);
   1433 }
   1434 
   1435 /* Check whether there is a (conditional) validity dependence from node[j]
   1436  * to node[i], forcing node[i] to follow node[j].
   1437  */
   1438 static isl_bool node_follows_strong(int i, int j, void *user)
   1439 {
   1440 	struct isl_sched_graph *graph = user;
   1441 
   1442 	return isl_sched_graph_has_validity_edge(graph, &graph->node[j],
   1443 							&graph->node[i]);
   1444 }
   1445 
   1446 /* Use Tarjan's algorithm for computing the strongly connected components
   1447  * in the dependence graph only considering those edges defined by "follows".
   1448  */
   1449 isl_stat isl_sched_graph_detect_ccs(isl_ctx *ctx,
   1450 	struct isl_sched_graph *graph,
   1451 	isl_bool (*follows)(int i, int j, void *user))
   1452 {
   1453 	int i, n;
   1454 	struct isl_tarjan_graph *g = NULL;
   1455 
   1456 	g = isl_tarjan_graph_init(ctx, graph->n, follows, graph);
   1457 	if (!g)
   1458 		return isl_stat_error;
   1459 
   1460 	graph->scc = 0;
   1461 	i = 0;
   1462 	n = graph->n;
   1463 	while (n) {
   1464 		while (g->order[i] != -1) {
   1465 			graph->node[g->order[i]].scc = graph->scc;
   1466 			--n;
   1467 			++i;
   1468 		}
   1469 		++i;
   1470 		graph->scc++;
   1471 	}
   1472 
   1473 	isl_tarjan_graph_free(g);
   1474 
   1475 	return isl_stat_ok;
   1476 }
   1477 
   1478 /* Apply Tarjan's algorithm to detect the strongly connected components
   1479  * in the dependence graph.
   1480  * Only consider the (conditional) validity dependences and clear "weak".
   1481  */
   1482 static isl_stat detect_sccs(isl_ctx *ctx, struct isl_sched_graph *graph)
   1483 {
   1484 	graph->weak = 0;
   1485 	return isl_sched_graph_detect_ccs(ctx, graph, &node_follows_strong);
   1486 }
   1487 
   1488 /* Apply Tarjan's algorithm to detect the (weakly) connected components
   1489  * in the dependence graph.
   1490  * Consider all dependences and set "weak".
   1491  */
   1492 static isl_stat detect_wccs(isl_ctx *ctx, struct isl_sched_graph *graph)
   1493 {
   1494 	graph->weak = 1;
   1495 	return isl_sched_graph_detect_ccs(ctx, graph, &node_follows_weak);
   1496 }
   1497 
   1498 static int cmp_scc(const void *a, const void *b, void *data)
   1499 {
   1500 	struct isl_sched_graph *graph = data;
   1501 	const int *i1 = a;
   1502 	const int *i2 = b;
   1503 
   1504 	return graph->node[*i1].scc - graph->node[*i2].scc;
   1505 }
   1506 
   1507 /* Sort the elements of graph->sorted according to the corresponding SCCs.
   1508  */
   1509 static int sort_sccs(struct isl_sched_graph *graph)
   1510 {
   1511 	return isl_sort(graph->sorted, graph->n, sizeof(int), &cmp_scc, graph);
   1512 }
   1513 
   1514 /* Return a non-parametric set in the compressed space of "node" that is
   1515  * bounded by the size in each direction
   1516  *
   1517  *	{ [x] : -S_i <= x_i <= S_i }
   1518  *
   1519  * If S_i is infinity in direction i, then there are no constraints
   1520  * in that direction.
   1521  *
   1522  * Cache the result in node->bounds.
   1523  */
   1524 static __isl_give isl_basic_set *get_size_bounds(struct isl_sched_node *node)
   1525 {
   1526 	isl_space *space;
   1527 	isl_basic_set *bounds;
   1528 	int i;
   1529 
   1530 	if (node->bounds)
   1531 		return isl_basic_set_copy(node->bounds);
   1532 
   1533 	if (node->compressed)
   1534 		space = isl_pw_multi_aff_get_domain_space(node->decompress);
   1535 	else
   1536 		space = isl_space_copy(node->space);
   1537 	space = isl_space_drop_all_params(space);
   1538 	bounds = isl_basic_set_universe(space);
   1539 
   1540 	for (i = 0; i < node->nvar; ++i) {
   1541 		isl_val *size;
   1542 
   1543 		size = isl_multi_val_get_val(node->sizes, i);
   1544 		if (!size)
   1545 			return isl_basic_set_free(bounds);
   1546 		if (!isl_val_is_int(size)) {
   1547 			isl_val_free(size);
   1548 			continue;
   1549 		}
   1550 		bounds = isl_basic_set_upper_bound_val(bounds, isl_dim_set, i,
   1551 							isl_val_copy(size));
   1552 		bounds = isl_basic_set_lower_bound_val(bounds, isl_dim_set, i,
   1553 							isl_val_neg(size));
   1554 	}
   1555 
   1556 	node->bounds = isl_basic_set_copy(bounds);
   1557 	return bounds;
   1558 }
   1559 
   1560 /* Compress the dependence relation "map", if needed, i.e.,
   1561  * when the source node "src" and/or the destination node "dst"
   1562  * has been compressed.
   1563  */
   1564 static __isl_give isl_map *compress(__isl_take isl_map *map,
   1565 	struct isl_sched_node *src, struct isl_sched_node *dst)
   1566 {
   1567 	if (src->compressed)
   1568 		map = isl_map_preimage_domain_pw_multi_aff(map,
   1569 					isl_pw_multi_aff_copy(src->decompress));
   1570 	if (dst->compressed)
   1571 		map = isl_map_preimage_range_pw_multi_aff(map,
   1572 					isl_pw_multi_aff_copy(dst->decompress));
   1573 	return map;
   1574 }
   1575 
   1576 /* Drop some constraints from "delta" that could be exploited
   1577  * to construct loop coalescing schedules.
   1578  * In particular, drop those constraint that bound the difference
   1579  * to the size of the domain.
   1580  * First project out the parameters to improve the effectiveness.
   1581  */
   1582 static __isl_give isl_set *drop_coalescing_constraints(
   1583 	__isl_take isl_set *delta, struct isl_sched_node *node)
   1584 {
   1585 	isl_size nparam;
   1586 	isl_basic_set *bounds;
   1587 
   1588 	nparam = isl_set_dim(delta, isl_dim_param);
   1589 	if (nparam < 0)
   1590 		return isl_set_free(delta);
   1591 
   1592 	bounds = get_size_bounds(node);
   1593 
   1594 	delta = isl_set_project_out(delta, isl_dim_param, 0, nparam);
   1595 	delta = isl_set_remove_divs(delta);
   1596 	delta = isl_set_plain_gist_basic_set(delta, bounds);
   1597 	return delta;
   1598 }
   1599 
   1600 /* Given a dependence relation R from "node" to itself,
   1601  * construct the set of coefficients of valid constraints for elements
   1602  * in that dependence relation.
   1603  * In particular, the result contains tuples of coefficients
   1604  * c_0, c_n, c_x such that
   1605  *
   1606  *	c_0 + c_n n + c_x y - c_x x >= 0 for each (x,y) in R
   1607  *
   1608  * or, equivalently,
   1609  *
   1610  *	c_0 + c_n n + c_x d >= 0 for each d in delta R = { y - x | (x,y) in R }
   1611  *
   1612  * We choose here to compute the dual of delta R.
   1613  * Alternatively, we could have computed the dual of R, resulting
   1614  * in a set of tuples c_0, c_n, c_x, c_y, and then
   1615  * plugged in (c_0, c_n, c_x, -c_x).
   1616  *
   1617  * If "need_param" is set, then the resulting coefficients effectively
   1618  * include coefficients for the parameters c_n.  Otherwise, they may
   1619  * have been projected out already.
   1620  * Since the constraints may be different for these two cases,
   1621  * they are stored in separate caches.
   1622  * In particular, if no parameter coefficients are required and
   1623  * the schedule_treat_coalescing option is set, then the parameters
   1624  * are projected out and some constraints that could be exploited
   1625  * to construct coalescing schedules are removed before the dual
   1626  * is computed.
   1627  *
   1628  * If "node" has been compressed, then the dependence relation
   1629  * is also compressed before the set of coefficients is computed.
   1630  */
   1631 static __isl_give isl_basic_set *intra_coefficients(
   1632 	struct isl_sched_graph *graph, struct isl_sched_node *node,
   1633 	__isl_take isl_map *map, int need_param)
   1634 {
   1635 	isl_ctx *ctx;
   1636 	isl_set *delta;
   1637 	isl_map *key;
   1638 	isl_basic_set *coef;
   1639 	isl_maybe_isl_basic_set m;
   1640 	isl_map_to_basic_set **hmap = &graph->intra_hmap;
   1641 	int treat;
   1642 
   1643 	if (!map)
   1644 		return NULL;
   1645 
   1646 	ctx = isl_map_get_ctx(map);
   1647 	treat = !need_param && isl_options_get_schedule_treat_coalescing(ctx);
   1648 	if (!treat)
   1649 		hmap = &graph->intra_hmap_param;
   1650 	m = isl_map_to_basic_set_try_get(*hmap, map);
   1651 	if (m.valid < 0 || m.valid) {
   1652 		isl_map_free(map);
   1653 		return m.value;
   1654 	}
   1655 
   1656 	key = isl_map_copy(map);
   1657 	map = compress(map, node, node);
   1658 	delta = isl_map_deltas(map);
   1659 	if (treat)
   1660 		delta = drop_coalescing_constraints(delta, node);
   1661 	delta = isl_set_remove_divs(delta);
   1662 	coef = isl_set_coefficients(delta);
   1663 	*hmap = isl_map_to_basic_set_set(*hmap, key, isl_basic_set_copy(coef));
   1664 
   1665 	return coef;
   1666 }
   1667 
   1668 /* Given a dependence relation R, construct the set of coefficients
   1669  * of valid constraints for elements in that dependence relation.
   1670  * In particular, the result contains tuples of coefficients
   1671  * c_0, c_n, c_x, c_y such that
   1672  *
   1673  *	c_0 + c_n n + c_x x + c_y y >= 0 for each (x,y) in R
   1674  *
   1675  * If the source or destination nodes of "edge" have been compressed,
   1676  * then the dependence relation is also compressed before
   1677  * the set of coefficients is computed.
   1678  */
   1679 static __isl_give isl_basic_set *inter_coefficients(
   1680 	struct isl_sched_graph *graph, struct isl_sched_edge *edge,
   1681 	__isl_take isl_map *map)
   1682 {
   1683 	isl_set *set;
   1684 	isl_map *key;
   1685 	isl_basic_set *coef;
   1686 	isl_maybe_isl_basic_set m;
   1687 
   1688 	m = isl_map_to_basic_set_try_get(graph->inter_hmap, map);
   1689 	if (m.valid < 0 || m.valid) {
   1690 		isl_map_free(map);
   1691 		return m.value;
   1692 	}
   1693 
   1694 	key = isl_map_copy(map);
   1695 	map = compress(map, edge->src, edge->dst);
   1696 	set = isl_map_wrap(isl_map_remove_divs(map));
   1697 	coef = isl_set_coefficients(set);
   1698 	graph->inter_hmap = isl_map_to_basic_set_set(graph->inter_hmap, key,
   1699 					isl_basic_set_copy(coef));
   1700 
   1701 	return coef;
   1702 }
   1703 
   1704 /* Return the position of the coefficients of the variables in
   1705  * the coefficients constraints "coef".
   1706  *
   1707  * The space of "coef" is of the form
   1708  *
   1709  *	{ coefficients[[cst, params] -> S] }
   1710  *
   1711  * Return the position of S.
   1712  */
   1713 static isl_size coef_var_offset(__isl_keep isl_basic_set *coef)
   1714 {
   1715 	isl_size offset;
   1716 	isl_space *space;
   1717 
   1718 	space = isl_space_unwrap(isl_basic_set_get_space(coef));
   1719 	offset = isl_space_dim(space, isl_dim_in);
   1720 	isl_space_free(space);
   1721 
   1722 	return offset;
   1723 }
   1724 
   1725 /* Return the offset of the coefficient of the constant term of "node"
   1726  * within the (I)LP.
   1727  *
   1728  * Within each node, the coefficients have the following order:
   1729  *	- positive and negative parts of c_i_x
   1730  *	- c_i_n (if parametric)
   1731  *	- c_i_0
   1732  */
   1733 static int node_cst_coef_offset(struct isl_sched_node *node)
   1734 {
   1735 	return node->start + 2 * node->nvar + node->nparam;
   1736 }
   1737 
   1738 /* Return the offset of the coefficients of the parameters of "node"
   1739  * within the (I)LP.
   1740  *
   1741  * Within each node, the coefficients have the following order:
   1742  *	- positive and negative parts of c_i_x
   1743  *	- c_i_n (if parametric)
   1744  *	- c_i_0
   1745  */
   1746 static int node_par_coef_offset(struct isl_sched_node *node)
   1747 {
   1748 	return node->start + 2 * node->nvar;
   1749 }
   1750 
   1751 /* Return the offset of the coefficients of the variables of "node"
   1752  * within the (I)LP.
   1753  *
   1754  * Within each node, the coefficients have the following order:
   1755  *	- positive and negative parts of c_i_x
   1756  *	- c_i_n (if parametric)
   1757  *	- c_i_0
   1758  */
   1759 static int node_var_coef_offset(struct isl_sched_node *node)
   1760 {
   1761 	return node->start;
   1762 }
   1763 
   1764 /* Return the position of the pair of variables encoding
   1765  * coefficient "i" of "node".
   1766  *
   1767  * The order of these variable pairs is the opposite of
   1768  * that of the coefficients, with 2 variables per coefficient.
   1769  */
   1770 static int node_var_coef_pos(struct isl_sched_node *node, int i)
   1771 {
   1772 	return node_var_coef_offset(node) + 2 * (node->nvar - 1 - i);
   1773 }
   1774 
   1775 /* Construct an isl_dim_map for mapping constraints on coefficients
   1776  * for "node" to the corresponding positions in graph->lp.
   1777  * "offset" is the offset of the coefficients for the variables
   1778  * in the input constraints.
   1779  * "s" is the sign of the mapping.
   1780  *
   1781  * The input constraints are given in terms of the coefficients
   1782  * (c_0, c_x) or (c_0, c_n, c_x).
   1783  * The mapping produced by this function essentially plugs in
   1784  * (0, c_i_x^+ - c_i_x^-) if s = 1 and
   1785  * (0, -c_i_x^+ + c_i_x^-) if s = -1 or
   1786  * (0, 0, c_i_x^+ - c_i_x^-) if s = 1 and
   1787  * (0, 0, -c_i_x^+ + c_i_x^-) if s = -1.
   1788  * In graph->lp, the c_i_x^- appear before their c_i_x^+ counterpart.
   1789  * Furthermore, the order of these pairs is the opposite of that
   1790  * of the corresponding coefficients.
   1791  *
   1792  * The caller can extend the mapping to also map the other coefficients
   1793  * (and therefore not plug in 0).
   1794  */
   1795 static __isl_give isl_dim_map *intra_dim_map(isl_ctx *ctx,
   1796 	struct isl_sched_graph *graph, struct isl_sched_node *node,
   1797 	int offset, int s)
   1798 {
   1799 	int pos;
   1800 	isl_size total;
   1801 	isl_dim_map *dim_map;
   1802 
   1803 	total = isl_basic_set_dim(graph->lp, isl_dim_all);
   1804 	if (!node || total < 0)
   1805 		return NULL;
   1806 
   1807 	pos = node_var_coef_pos(node, 0);
   1808 	dim_map = isl_dim_map_alloc(ctx, total);
   1809 	isl_dim_map_range(dim_map, pos, -2, offset, 1, node->nvar, -s);
   1810 	isl_dim_map_range(dim_map, pos + 1, -2, offset, 1, node->nvar, s);
   1811 
   1812 	return dim_map;
   1813 }
   1814 
   1815 /* Construct an isl_dim_map for mapping constraints on coefficients
   1816  * for "src" (node i) and "dst" (node j) to the corresponding positions
   1817  * in graph->lp.
   1818  * "offset" is the offset of the coefficients for the variables of "src"
   1819  * in the input constraints.
   1820  * "s" is the sign of the mapping.
   1821  *
   1822  * The input constraints are given in terms of the coefficients
   1823  * (c_0, c_n, c_x, c_y).
   1824  * The mapping produced by this function essentially plugs in
   1825  * (c_j_0 - c_i_0, c_j_n - c_i_n,
   1826  *  -(c_i_x^+ - c_i_x^-), c_j_x^+ - c_j_x^-) if s = 1 and
   1827  * (-c_j_0 + c_i_0, -c_j_n + c_i_n,
   1828  *  c_i_x^+ - c_i_x^-, -(c_j_x^+ - c_j_x^-)) if s = -1.
   1829  * In graph->lp, the c_*^- appear before their c_*^+ counterpart.
   1830  * Furthermore, the order of these pairs is the opposite of that
   1831  * of the corresponding coefficients.
   1832  *
   1833  * The caller can further extend the mapping.
   1834  */
   1835 static __isl_give isl_dim_map *inter_dim_map(isl_ctx *ctx,
   1836 	struct isl_sched_graph *graph, struct isl_sched_node *src,
   1837 	struct isl_sched_node *dst, int offset, int s)
   1838 {
   1839 	int pos;
   1840 	isl_size total;
   1841 	isl_dim_map *dim_map;
   1842 
   1843 	total = isl_basic_set_dim(graph->lp, isl_dim_all);
   1844 	if (!src || !dst || total < 0)
   1845 		return NULL;
   1846 
   1847 	dim_map = isl_dim_map_alloc(ctx, total);
   1848 
   1849 	pos = node_cst_coef_offset(dst);
   1850 	isl_dim_map_range(dim_map, pos, 0, 0, 0, 1, s);
   1851 	pos = node_par_coef_offset(dst);
   1852 	isl_dim_map_range(dim_map, pos, 1, 1, 1, dst->nparam, s);
   1853 	pos = node_var_coef_pos(dst, 0);
   1854 	isl_dim_map_range(dim_map, pos, -2, offset + src->nvar, 1,
   1855 			  dst->nvar, -s);
   1856 	isl_dim_map_range(dim_map, pos + 1, -2, offset + src->nvar, 1,
   1857 			  dst->nvar, s);
   1858 
   1859 	pos = node_cst_coef_offset(src);
   1860 	isl_dim_map_range(dim_map, pos, 0, 0, 0, 1, -s);
   1861 	pos = node_par_coef_offset(src);
   1862 	isl_dim_map_range(dim_map, pos, 1, 1, 1, src->nparam, -s);
   1863 	pos = node_var_coef_pos(src, 0);
   1864 	isl_dim_map_range(dim_map, pos, -2, offset, 1, src->nvar, s);
   1865 	isl_dim_map_range(dim_map, pos + 1, -2, offset, 1, src->nvar, -s);
   1866 
   1867 	return dim_map;
   1868 }
   1869 
   1870 /* Add the constraints from "src" to "dst" using "dim_map",
   1871  * after making sure there is enough room in "dst" for the extra constraints.
   1872  */
   1873 static __isl_give isl_basic_set *add_constraints_dim_map(
   1874 	__isl_take isl_basic_set *dst, __isl_take isl_basic_set *src,
   1875 	__isl_take isl_dim_map *dim_map)
   1876 {
   1877 	isl_size n_eq, n_ineq;
   1878 
   1879 	n_eq = isl_basic_set_n_equality(src);
   1880 	n_ineq = isl_basic_set_n_inequality(src);
   1881 	if (n_eq < 0 || n_ineq < 0)
   1882 		dst = isl_basic_set_free(dst);
   1883 	dst = isl_basic_set_extend_constraints(dst, n_eq, n_ineq);
   1884 	dst = isl_basic_set_add_constraints_dim_map(dst, src, dim_map);
   1885 	return dst;
   1886 }
   1887 
   1888 /* Add constraints to graph->lp that force validity for the given
   1889  * dependence from a node i to itself.
   1890  * That is, add constraints that enforce
   1891  *
   1892  *	(c_i_0 + c_i_n n + c_i_x y) - (c_i_0 + c_i_n n + c_i_x x)
   1893  *	= c_i_x (y - x) >= 0
   1894  *
   1895  * for each (x,y) in R.
   1896  * We obtain general constraints on coefficients (c_0, c_x)
   1897  * of valid constraints for (y - x) and then plug in (0, c_i_x^+ - c_i_x^-),
   1898  * where c_i_x = c_i_x^+ - c_i_x^-, with c_i_x^+ and c_i_x^- non-negative.
   1899  * In graph->lp, the c_i_x^- appear before their c_i_x^+ counterpart.
   1900  * Note that the result of intra_coefficients may also contain
   1901  * parameter coefficients c_n, in which case 0 is plugged in for them as well.
   1902  */
   1903 static isl_stat add_intra_validity_constraints(struct isl_sched_graph *graph,
   1904 	struct isl_sched_edge *edge)
   1905 {
   1906 	isl_size offset;
   1907 	isl_map *map = isl_map_copy(edge->map);
   1908 	isl_ctx *ctx = isl_map_get_ctx(map);
   1909 	isl_dim_map *dim_map;
   1910 	isl_basic_set *coef;
   1911 	struct isl_sched_node *node = edge->src;
   1912 
   1913 	coef = intra_coefficients(graph, node, map, 0);
   1914 
   1915 	offset = coef_var_offset(coef);
   1916 	if (offset < 0)
   1917 		coef = isl_basic_set_free(coef);
   1918 	if (!coef)
   1919 		return isl_stat_error;
   1920 
   1921 	dim_map = intra_dim_map(ctx, graph, node, offset, 1);
   1922 	graph->lp = add_constraints_dim_map(graph->lp, coef, dim_map);
   1923 
   1924 	return isl_stat_ok;
   1925 }
   1926 
   1927 /* Add constraints to graph->lp that force validity for the given
   1928  * dependence from node i to node j.
   1929  * That is, add constraints that enforce
   1930  *
   1931  *	(c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x) >= 0
   1932  *
   1933  * for each (x,y) in R.
   1934  * We obtain general constraints on coefficients (c_0, c_n, c_x, c_y)
   1935  * of valid constraints for R and then plug in
   1936  * (c_j_0 - c_i_0, c_j_n - c_i_n, -(c_i_x^+ - c_i_x^-), c_j_x^+ - c_j_x^-),
   1937  * where c_* = c_*^+ - c_*^-, with c_*^+ and c_*^- non-negative.
   1938  * In graph->lp, the c_*^- appear before their c_*^+ counterpart.
   1939  */
   1940 static isl_stat add_inter_validity_constraints(struct isl_sched_graph *graph,
   1941 	struct isl_sched_edge *edge)
   1942 {
   1943 	isl_size offset;
   1944 	isl_map *map;
   1945 	isl_ctx *ctx;
   1946 	isl_dim_map *dim_map;
   1947 	isl_basic_set *coef;
   1948 	struct isl_sched_node *src = edge->src;
   1949 	struct isl_sched_node *dst = edge->dst;
   1950 
   1951 	if (!graph->lp)
   1952 		return isl_stat_error;
   1953 
   1954 	map = isl_map_copy(edge->map);
   1955 	ctx = isl_map_get_ctx(map);
   1956 	coef = inter_coefficients(graph, edge, map);
   1957 
   1958 	offset = coef_var_offset(coef);
   1959 	if (offset < 0)
   1960 		coef = isl_basic_set_free(coef);
   1961 	if (!coef)
   1962 		return isl_stat_error;
   1963 
   1964 	dim_map = inter_dim_map(ctx, graph, src, dst, offset, 1);
   1965 
   1966 	edge->start = graph->lp->n_ineq;
   1967 	graph->lp = add_constraints_dim_map(graph->lp, coef, dim_map);
   1968 	if (!graph->lp)
   1969 		return isl_stat_error;
   1970 	edge->end = graph->lp->n_ineq;
   1971 
   1972 	return isl_stat_ok;
   1973 }
   1974 
   1975 /* Add constraints to graph->lp that bound the dependence distance for the given
   1976  * dependence from a node i to itself.
   1977  * If s = 1, we add the constraint
   1978  *
   1979  *	c_i_x (y - x) <= m_0 + m_n n
   1980  *
   1981  * or
   1982  *
   1983  *	-c_i_x (y - x) + m_0 + m_n n >= 0
   1984  *
   1985  * for each (x,y) in R.
   1986  * If s = -1, we add the constraint
   1987  *
   1988  *	-c_i_x (y - x) <= m_0 + m_n n
   1989  *
   1990  * or
   1991  *
   1992  *	c_i_x (y - x) + m_0 + m_n n >= 0
   1993  *
   1994  * for each (x,y) in R.
   1995  * We obtain general constraints on coefficients (c_0, c_n, c_x)
   1996  * of valid constraints for (y - x) and then plug in (m_0, m_n, -s * c_i_x),
   1997  * with each coefficient (except m_0) represented as a pair of non-negative
   1998  * coefficients.
   1999  *
   2000  *
   2001  * If "local" is set, then we add constraints
   2002  *
   2003  *	c_i_x (y - x) <= 0
   2004  *
   2005  * or
   2006  *
   2007  *	-c_i_x (y - x) <= 0
   2008  *
   2009  * instead, forcing the dependence distance to be (less than or) equal to 0.
   2010  * That is, we plug in (0, 0, -s * c_i_x),
   2011  * intra_coefficients is not required to have c_n in its result when
   2012  * "local" is set.  If they are missing, then (0, -s * c_i_x) is plugged in.
   2013  * Note that dependences marked local are treated as validity constraints
   2014  * by add_all_validity_constraints and therefore also have
   2015  * their distances bounded by 0 from below.
   2016  */
   2017 static isl_stat add_intra_proximity_constraints(struct isl_sched_graph *graph,
   2018 	struct isl_sched_edge *edge, int s, int local)
   2019 {
   2020 	isl_size offset;
   2021 	isl_size nparam;
   2022 	isl_map *map = isl_map_copy(edge->map);
   2023 	isl_ctx *ctx = isl_map_get_ctx(map);
   2024 	isl_dim_map *dim_map;
   2025 	isl_basic_set *coef;
   2026 	struct isl_sched_node *node = edge->src;
   2027 
   2028 	coef = intra_coefficients(graph, node, map, !local);
   2029 	nparam = isl_space_dim(node->space, isl_dim_param);
   2030 
   2031 	offset = coef_var_offset(coef);
   2032 	if (nparam < 0 || offset < 0)
   2033 		coef = isl_basic_set_free(coef);
   2034 	if (!coef)
   2035 		return isl_stat_error;
   2036 
   2037 	dim_map = intra_dim_map(ctx, graph, node, offset, -s);
   2038 
   2039 	if (!local) {
   2040 		isl_dim_map_range(dim_map, 1, 0, 0, 0, 1, 1);
   2041 		isl_dim_map_range(dim_map, 4, 2, 1, 1, nparam, -1);
   2042 		isl_dim_map_range(dim_map, 5, 2, 1, 1, nparam, 1);
   2043 	}
   2044 	graph->lp = add_constraints_dim_map(graph->lp, coef, dim_map);
   2045 
   2046 	return isl_stat_ok;
   2047 }
   2048 
   2049 /* Add constraints to graph->lp that bound the dependence distance for the given
   2050  * dependence from node i to node j.
   2051  * If s = 1, we add the constraint
   2052  *
   2053  *	(c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x)
   2054  *		<= m_0 + m_n n
   2055  *
   2056  * or
   2057  *
   2058  *	-(c_j_0 + c_j_n n + c_j_x y) + (c_i_0 + c_i_n n + c_i_x x) +
   2059  *		m_0 + m_n n >= 0
   2060  *
   2061  * for each (x,y) in R.
   2062  * If s = -1, we add the constraint
   2063  *
   2064  *	-((c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x))
   2065  *		<= m_0 + m_n n
   2066  *
   2067  * or
   2068  *
   2069  *	(c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x) +
   2070  *		m_0 + m_n n >= 0
   2071  *
   2072  * for each (x,y) in R.
   2073  * We obtain general constraints on coefficients (c_0, c_n, c_x, c_y)
   2074  * of valid constraints for R and then plug in
   2075  * (m_0 - s*c_j_0 + s*c_i_0, m_n - s*c_j_n + s*c_i_n,
   2076  *  s*c_i_x, -s*c_j_x)
   2077  * with each coefficient (except m_0, c_*_0 and c_*_n)
   2078  * represented as a pair of non-negative coefficients.
   2079  *
   2080  *
   2081  * If "local" is set (and s = 1), then we add constraints
   2082  *
   2083  *	(c_j_0 + c_j_n n + c_j_x y) - (c_i_0 + c_i_n n + c_i_x x) <= 0
   2084  *
   2085  * or
   2086  *
   2087  *	-((c_j_0 + c_j_n n + c_j_x y) + (c_i_0 + c_i_n n + c_i_x x)) >= 0
   2088  *
   2089  * instead, forcing the dependence distance to be (less than or) equal to 0.
   2090  * That is, we plug in
   2091  * (-s*c_j_0 + s*c_i_0, -s*c_j_n + s*c_i_n, s*c_i_x, -s*c_j_x).
   2092  * Note that dependences marked local are treated as validity constraints
   2093  * by add_all_validity_constraints and therefore also have
   2094  * their distances bounded by 0 from below.
   2095  */
   2096 static isl_stat add_inter_proximity_constraints(struct isl_sched_graph *graph,
   2097 	struct isl_sched_edge *edge, int s, int local)
   2098 {
   2099 	isl_size offset;
   2100 	isl_size nparam;
   2101 	isl_map *map = isl_map_copy(edge->map);
   2102 	isl_ctx *ctx = isl_map_get_ctx(map);
   2103 	isl_dim_map *dim_map;
   2104 	isl_basic_set *coef;
   2105 	struct isl_sched_node *src = edge->src;
   2106 	struct isl_sched_node *dst = edge->dst;
   2107 
   2108 	coef = inter_coefficients(graph, edge, map);
   2109 	nparam = isl_space_dim(src->space, isl_dim_param);
   2110 
   2111 	offset = coef_var_offset(coef);
   2112 	if (nparam < 0 || offset < 0)
   2113 		coef = isl_basic_set_free(coef);
   2114 	if (!coef)
   2115 		return isl_stat_error;
   2116 
   2117 	dim_map = inter_dim_map(ctx, graph, src, dst, offset, -s);
   2118 
   2119 	if (!local) {
   2120 		isl_dim_map_range(dim_map, 1, 0, 0, 0, 1, 1);
   2121 		isl_dim_map_range(dim_map, 4, 2, 1, 1, nparam, -1);
   2122 		isl_dim_map_range(dim_map, 5, 2, 1, 1, nparam, 1);
   2123 	}
   2124 
   2125 	graph->lp = add_constraints_dim_map(graph->lp, coef, dim_map);
   2126 
   2127 	return isl_stat_ok;
   2128 }
   2129 
   2130 /* Should the distance over "edge" be forced to zero?
   2131  * That is, is it marked as a local edge?
   2132  * If "use_coincidence" is set, then coincidence edges are treated
   2133  * as local edges.
   2134  */
   2135 static int force_zero(struct isl_sched_edge *edge, int use_coincidence)
   2136 {
   2137 	return is_local(edge) || (use_coincidence && is_coincidence(edge));
   2138 }
   2139 
   2140 /* Add all validity constraints to graph->lp.
   2141  *
   2142  * An edge that is forced to be local needs to have its dependence
   2143  * distances equal to zero.  We take care of bounding them by 0 from below
   2144  * here.  add_all_proximity_constraints takes care of bounding them by 0
   2145  * from above.
   2146  *
   2147  * If "use_coincidence" is set, then we treat coincidence edges as local edges.
   2148  * Otherwise, we ignore them.
   2149  */
   2150 static int add_all_validity_constraints(struct isl_sched_graph *graph,
   2151 	int use_coincidence)
   2152 {
   2153 	int i;
   2154 
   2155 	for (i = 0; i < graph->n_edge; ++i) {
   2156 		struct isl_sched_edge *edge = &graph->edge[i];
   2157 		int zero;
   2158 
   2159 		zero = force_zero(edge, use_coincidence);
   2160 		if (!is_validity(edge) && !zero)
   2161 			continue;
   2162 		if (edge->src != edge->dst)
   2163 			continue;
   2164 		if (add_intra_validity_constraints(graph, edge) < 0)
   2165 			return -1;
   2166 	}
   2167 
   2168 	for (i = 0; i < graph->n_edge; ++i) {
   2169 		struct isl_sched_edge *edge = &graph->edge[i];
   2170 		int zero;
   2171 
   2172 		zero = force_zero(edge, use_coincidence);
   2173 		if (!is_validity(edge) && !zero)
   2174 			continue;
   2175 		if (edge->src == edge->dst)
   2176 			continue;
   2177 		if (add_inter_validity_constraints(graph, edge) < 0)
   2178 			return -1;
   2179 	}
   2180 
   2181 	return 0;
   2182 }
   2183 
   2184 /* Add constraints to graph->lp that bound the dependence distance
   2185  * for all dependence relations.
   2186  * If a given proximity dependence is identical to a validity
   2187  * dependence, then the dependence distance is already bounded
   2188  * from below (by zero), so we only need to bound the distance
   2189  * from above.  (This includes the case of "local" dependences
   2190  * which are treated as validity dependence by add_all_validity_constraints.)
   2191  * Otherwise, we need to bound the distance both from above and from below.
   2192  *
   2193  * If "use_coincidence" is set, then we treat coincidence edges as local edges.
   2194  * Otherwise, we ignore them.
   2195  */
   2196 static int add_all_proximity_constraints(struct isl_sched_graph *graph,
   2197 	int use_coincidence)
   2198 {
   2199 	int i;
   2200 
   2201 	for (i = 0; i < graph->n_edge; ++i) {
   2202 		struct isl_sched_edge *edge = &graph->edge[i];
   2203 		int zero;
   2204 
   2205 		zero = force_zero(edge, use_coincidence);
   2206 		if (!isl_sched_edge_is_proximity(edge) && !zero)
   2207 			continue;
   2208 		if (edge->src == edge->dst &&
   2209 		    add_intra_proximity_constraints(graph, edge, 1, zero) < 0)
   2210 			return -1;
   2211 		if (edge->src != edge->dst &&
   2212 		    add_inter_proximity_constraints(graph, edge, 1, zero) < 0)
   2213 			return -1;
   2214 		if (is_validity(edge) || zero)
   2215 			continue;
   2216 		if (edge->src == edge->dst &&
   2217 		    add_intra_proximity_constraints(graph, edge, -1, 0) < 0)
   2218 			return -1;
   2219 		if (edge->src != edge->dst &&
   2220 		    add_inter_proximity_constraints(graph, edge, -1, 0) < 0)
   2221 			return -1;
   2222 	}
   2223 
   2224 	return 0;
   2225 }
   2226 
   2227 /* Normalize the rows of "indep" such that all rows are lexicographically
   2228  * positive and such that each row contains as many final zeros as possible,
   2229  * given the choice for the previous rows.
   2230  * Do this by performing elementary row operations.
   2231  */
   2232 static __isl_give isl_mat *normalize_independent(__isl_take isl_mat *indep)
   2233 {
   2234 	indep = isl_mat_reverse_gauss(indep);
   2235 	indep = isl_mat_lexnonneg_rows(indep);
   2236 	return indep;
   2237 }
   2238 
   2239 /* Extract the linear part of the current schedule for node "node".
   2240  */
   2241 static __isl_give isl_mat *extract_linear_schedule(struct isl_sched_node *node)
   2242 {
   2243 	isl_size n_row = isl_mat_rows(node->sched);
   2244 
   2245 	if (n_row < 0)
   2246 		return NULL;
   2247 	return isl_mat_sub_alloc(node->sched, 0, n_row,
   2248 			      1 + node->nparam, node->nvar);
   2249 }
   2250 
   2251 /* Compute a basis for the rows in the linear part of the schedule
   2252  * and extend this basis to a full basis.  The remaining rows
   2253  * can then be used to force linear independence from the rows
   2254  * in the schedule.
   2255  *
   2256  * In particular, given the schedule rows S, we compute
   2257  *
   2258  *	S   = H Q
   2259  *	S U = H
   2260  *
   2261  * with H the Hermite normal form of S.  That is, all but the
   2262  * first rank columns of H are zero and so each row in S is
   2263  * a linear combination of the first rank rows of Q.
   2264  * The matrix Q can be used as a variable transformation
   2265  * that isolates the directions of S in the first rank rows.
   2266  * Transposing S U = H yields
   2267  *
   2268  *	U^T S^T = H^T
   2269  *
   2270  * with all but the first rank rows of H^T zero.
   2271  * The last rows of U^T are therefore linear combinations
   2272  * of schedule coefficients that are all zero on schedule
   2273  * coefficients that are linearly dependent on the rows of S.
   2274  * At least one of these combinations is non-zero on
   2275  * linearly independent schedule coefficients.
   2276  * The rows are normalized to involve as few of the last
   2277  * coefficients as possible and to have a positive initial value.
   2278  */
   2279 isl_stat isl_sched_node_update_vmap(struct isl_sched_node *node)
   2280 {
   2281 	isl_mat *H, *U, *Q;
   2282 
   2283 	H = extract_linear_schedule(node);
   2284 
   2285 	H = isl_mat_left_hermite(H, 0, &U, &Q);
   2286 	isl_mat_free(node->indep);
   2287 	isl_mat_free(node->vmap);
   2288 	node->vmap = Q;
   2289 	node->indep = isl_mat_transpose(U);
   2290 	node->rank = isl_mat_initial_non_zero_cols(H);
   2291 	node->indep = isl_mat_drop_rows(node->indep, 0, node->rank);
   2292 	node->indep = normalize_independent(node->indep);
   2293 	isl_mat_free(H);
   2294 
   2295 	if (!node->indep || !node->vmap || node->rank < 0)
   2296 		return isl_stat_error;
   2297 	return isl_stat_ok;
   2298 }
   2299 
   2300 /* Is "edge" marked as a validity or a conditional validity edge?
   2301  */
   2302 static int is_any_validity(struct isl_sched_edge *edge)
   2303 {
   2304 	return is_validity(edge) ||
   2305 		isl_sched_edge_is_conditional_validity(edge);
   2306 }
   2307 
   2308 /* How many times should we count the constraints in "edge"?
   2309  *
   2310  * We count as follows
   2311  * validity		-> 1 (>= 0)
   2312  * validity+proximity	-> 2 (>= 0 and upper bound)
   2313  * proximity		-> 2 (lower and upper bound)
   2314  * local(+any)		-> 2 (>= 0 and <= 0)
   2315  *
   2316  * If an edge is only marked conditional_validity then it counts
   2317  * as zero since it is only checked afterwards.
   2318  *
   2319  * If "use_coincidence" is set, then we treat coincidence edges as local edges.
   2320  * Otherwise, we ignore them.
   2321  */
   2322 static int edge_multiplicity(struct isl_sched_edge *edge, int use_coincidence)
   2323 {
   2324 	if (isl_sched_edge_is_proximity(edge) ||
   2325 	    force_zero(edge, use_coincidence))
   2326 		return 2;
   2327 	if (is_validity(edge))
   2328 		return 1;
   2329 	return 0;
   2330 }
   2331 
   2332 /* How many times should the constraints in "edge" be counted
   2333  * as a parametric intra-node constraint?
   2334  *
   2335  * Only proximity edges that are not forced zero need
   2336  * coefficient constraints that include coefficients for parameters.
   2337  * If the edge is also a validity edge, then only
   2338  * an upper bound is introduced.  Otherwise, both lower and upper bounds
   2339  * are introduced.
   2340  */
   2341 static int parametric_intra_edge_multiplicity(struct isl_sched_edge *edge,
   2342 	int use_coincidence)
   2343 {
   2344 	if (edge->src != edge->dst)
   2345 		return 0;
   2346 	if (!isl_sched_edge_is_proximity(edge))
   2347 		return 0;
   2348 	if (force_zero(edge, use_coincidence))
   2349 		return 0;
   2350 	if (is_validity(edge))
   2351 		return 1;
   2352 	else
   2353 		return 2;
   2354 }
   2355 
   2356 /* Add "f" times the number of equality and inequality constraints of "bset"
   2357  * to "n_eq" and "n_ineq" and free "bset".
   2358  */
   2359 static isl_stat update_count(__isl_take isl_basic_set *bset,
   2360 	int f, int *n_eq, int *n_ineq)
   2361 {
   2362 	isl_size eq, ineq;
   2363 
   2364 	eq = isl_basic_set_n_equality(bset);
   2365 	ineq = isl_basic_set_n_inequality(bset);
   2366 	isl_basic_set_free(bset);
   2367 
   2368 	if (eq < 0 || ineq < 0)
   2369 		return isl_stat_error;
   2370 
   2371 	*n_eq += eq;
   2372 	*n_ineq += ineq;
   2373 
   2374 	return isl_stat_ok;
   2375 }
   2376 
   2377 /* Count the number of equality and inequality constraints
   2378  * that will be added for the given map.
   2379  *
   2380  * The edges that require parameter coefficients are counted separately.
   2381  *
   2382  * "use_coincidence" is set if we should take into account coincidence edges.
   2383  */
   2384 static isl_stat count_map_constraints(struct isl_sched_graph *graph,
   2385 	struct isl_sched_edge *edge, __isl_take isl_map *map,
   2386 	int *n_eq, int *n_ineq, int use_coincidence)
   2387 {
   2388 	isl_map *copy;
   2389 	isl_basic_set *coef;
   2390 	int f = edge_multiplicity(edge, use_coincidence);
   2391 	int fp = parametric_intra_edge_multiplicity(edge, use_coincidence);
   2392 
   2393 	if (f == 0) {
   2394 		isl_map_free(map);
   2395 		return isl_stat_ok;
   2396 	}
   2397 
   2398 	if (edge->src != edge->dst) {
   2399 		coef = inter_coefficients(graph, edge, map);
   2400 		return update_count(coef, f, n_eq, n_ineq);
   2401 	}
   2402 
   2403 	if (fp > 0) {
   2404 		copy = isl_map_copy(map);
   2405 		coef = intra_coefficients(graph, edge->src, copy, 1);
   2406 		if (update_count(coef, fp, n_eq, n_ineq) < 0)
   2407 			goto error;
   2408 	}
   2409 
   2410 	if (f > fp) {
   2411 		copy = isl_map_copy(map);
   2412 		coef = intra_coefficients(graph, edge->src, copy, 0);
   2413 		if (update_count(coef, f - fp, n_eq, n_ineq) < 0)
   2414 			goto error;
   2415 	}
   2416 
   2417 	isl_map_free(map);
   2418 	return isl_stat_ok;
   2419 error:
   2420 	isl_map_free(map);
   2421 	return isl_stat_error;
   2422 }
   2423 
   2424 /* Count the number of equality and inequality constraints
   2425  * that will be added to the main lp problem.
   2426  * We count as follows
   2427  * validity		-> 1 (>= 0)
   2428  * validity+proximity	-> 2 (>= 0 and upper bound)
   2429  * proximity		-> 2 (lower and upper bound)
   2430  * local(+any)		-> 2 (>= 0 and <= 0)
   2431  *
   2432  * If "use_coincidence" is set, then we treat coincidence edges as local edges.
   2433  * Otherwise, we ignore them.
   2434  */
   2435 static int count_constraints(struct isl_sched_graph *graph,
   2436 	int *n_eq, int *n_ineq, int use_coincidence)
   2437 {
   2438 	int i;
   2439 
   2440 	*n_eq = *n_ineq = 0;
   2441 	for (i = 0; i < graph->n_edge; ++i) {
   2442 		struct isl_sched_edge *edge = &graph->edge[i];
   2443 		isl_map *map = isl_map_copy(edge->map);
   2444 
   2445 		if (count_map_constraints(graph, edge, map, n_eq, n_ineq,
   2446 					    use_coincidence) < 0)
   2447 			return -1;
   2448 	}
   2449 
   2450 	return 0;
   2451 }
   2452 
   2453 /* Count the number of constraints that will be added by
   2454  * add_bound_constant_constraints to bound the values of the constant terms
   2455  * and increment *n_eq and *n_ineq accordingly.
   2456  *
   2457  * In practice, add_bound_constant_constraints only adds inequalities.
   2458  */
   2459 static isl_stat count_bound_constant_constraints(isl_ctx *ctx,
   2460 	struct isl_sched_graph *graph, int *n_eq, int *n_ineq)
   2461 {
   2462 	if (isl_options_get_schedule_max_constant_term(ctx) == -1)
   2463 		return isl_stat_ok;
   2464 
   2465 	*n_ineq += graph->n;
   2466 
   2467 	return isl_stat_ok;
   2468 }
   2469 
   2470 /* Add constraints to bound the values of the constant terms in the schedule,
   2471  * if requested by the user.
   2472  *
   2473  * The maximal value of the constant terms is defined by the option
   2474  * "schedule_max_constant_term".
   2475  */
   2476 static isl_stat add_bound_constant_constraints(isl_ctx *ctx,
   2477 	struct isl_sched_graph *graph)
   2478 {
   2479 	int i, k;
   2480 	int max;
   2481 	isl_size total;
   2482 
   2483 	max = isl_options_get_schedule_max_constant_term(ctx);
   2484 	if (max == -1)
   2485 		return isl_stat_ok;
   2486 
   2487 	total = isl_basic_set_dim(graph->lp, isl_dim_set);
   2488 	if (total < 0)
   2489 		return isl_stat_error;
   2490 
   2491 	for (i = 0; i < graph->n; ++i) {
   2492 		struct isl_sched_node *node = &graph->node[i];
   2493 		int pos;
   2494 
   2495 		k = isl_basic_set_alloc_inequality(graph->lp);
   2496 		if (k < 0)
   2497 			return isl_stat_error;
   2498 		isl_seq_clr(graph->lp->ineq[k], 1 + total);
   2499 		pos = node_cst_coef_offset(node);
   2500 		isl_int_set_si(graph->lp->ineq[k][1 + pos], -1);
   2501 		isl_int_set_si(graph->lp->ineq[k][0], max);
   2502 	}
   2503 
   2504 	return isl_stat_ok;
   2505 }
   2506 
   2507 /* Count the number of constraints that will be added by
   2508  * add_bound_coefficient_constraints and increment *n_eq and *n_ineq
   2509  * accordingly.
   2510  *
   2511  * In practice, add_bound_coefficient_constraints only adds inequalities.
   2512  */
   2513 static int count_bound_coefficient_constraints(isl_ctx *ctx,
   2514 	struct isl_sched_graph *graph, int *n_eq, int *n_ineq)
   2515 {
   2516 	int i;
   2517 
   2518 	if (isl_options_get_schedule_max_coefficient(ctx) == -1 &&
   2519 	    !isl_options_get_schedule_treat_coalescing(ctx))
   2520 		return 0;
   2521 
   2522 	for (i = 0; i < graph->n; ++i)
   2523 		*n_ineq += graph->node[i].nparam + 2 * graph->node[i].nvar;
   2524 
   2525 	return 0;
   2526 }
   2527 
   2528 /* Add constraints to graph->lp that bound the values of
   2529  * the parameter schedule coefficients of "node" to "max" and
   2530  * the variable schedule coefficients to the corresponding entry
   2531  * in node->max.
   2532  * In either case, a negative value means that no bound needs to be imposed.
   2533  *
   2534  * For parameter coefficients, this amounts to adding a constraint
   2535  *
   2536  *	c_n <= max
   2537  *
   2538  * i.e.,
   2539  *
   2540  *	-c_n + max >= 0
   2541  *
   2542  * The variables coefficients are, however, not represented directly.
   2543  * Instead, the variable coefficients c_x are written as differences
   2544  * c_x = c_x^+ - c_x^-.
   2545  * That is,
   2546  *
   2547  *	-max_i <= c_x_i <= max_i
   2548  *
   2549  * is encoded as
   2550  *
   2551  *	-max_i <= c_x_i^+ - c_x_i^- <= max_i
   2552  *
   2553  * or
   2554  *
   2555  *	-(c_x_i^+ - c_x_i^-) + max_i >= 0
   2556  *	c_x_i^+ - c_x_i^- + max_i >= 0
   2557  */
   2558 static isl_stat node_add_coefficient_constraints(isl_ctx *ctx,
   2559 	struct isl_sched_graph *graph, struct isl_sched_node *node, int max)
   2560 {
   2561 	int i, j, k;
   2562 	isl_size total;
   2563 	isl_vec *ineq;
   2564 
   2565 	total = isl_basic_set_dim(graph->lp, isl_dim_set);
   2566 	if (total < 0)
   2567 		return isl_stat_error;
   2568 
   2569 	for (j = 0; j < node->nparam; ++j) {
   2570 		int dim;
   2571 
   2572 		if (max < 0)
   2573 			continue;
   2574 
   2575 		k = isl_basic_set_alloc_inequality(graph->lp);
   2576 		if (k < 0)
   2577 			return isl_stat_error;
   2578 		dim = 1 + node_par_coef_offset(node) + j;
   2579 		isl_seq_clr(graph->lp->ineq[k], 1 + total);
   2580 		isl_int_set_si(graph->lp->ineq[k][dim], -1);
   2581 		isl_int_set_si(graph->lp->ineq[k][0], max);
   2582 	}
   2583 
   2584 	ineq = isl_vec_alloc(ctx, 1 + total);
   2585 	ineq = isl_vec_clr(ineq);
   2586 	if (!ineq)
   2587 		return isl_stat_error;
   2588 	for (i = 0; i < node->nvar; ++i) {
   2589 		int pos = 1 + node_var_coef_pos(node, i);
   2590 
   2591 		if (isl_int_is_neg(node->max->el[i]))
   2592 			continue;
   2593 
   2594 		isl_int_set_si(ineq->el[pos], 1);
   2595 		isl_int_set_si(ineq->el[pos + 1], -1);
   2596 		isl_int_set(ineq->el[0], node->max->el[i]);
   2597 
   2598 		k = isl_basic_set_alloc_inequality(graph->lp);
   2599 		if (k < 0)
   2600 			goto error;
   2601 		isl_seq_cpy(graph->lp->ineq[k], ineq->el, 1 + total);
   2602 
   2603 		isl_seq_neg(ineq->el + pos, ineq->el + pos, 2);
   2604 		k = isl_basic_set_alloc_inequality(graph->lp);
   2605 		if (k < 0)
   2606 			goto error;
   2607 		isl_seq_cpy(graph->lp->ineq[k], ineq->el, 1 + total);
   2608 
   2609 		isl_seq_clr(ineq->el + pos, 2);
   2610 	}
   2611 	isl_vec_free(ineq);
   2612 
   2613 	return isl_stat_ok;
   2614 error:
   2615 	isl_vec_free(ineq);
   2616 	return isl_stat_error;
   2617 }
   2618 
   2619 /* Add constraints that bound the values of the variable and parameter
   2620  * coefficients of the schedule.
   2621  *
   2622  * The maximal value of the coefficients is defined by the option
   2623  * 'schedule_max_coefficient' and the entries in node->max.
   2624  * These latter entries are only set if either the schedule_max_coefficient
   2625  * option or the schedule_treat_coalescing option is set.
   2626  */
   2627 static isl_stat add_bound_coefficient_constraints(isl_ctx *ctx,
   2628 	struct isl_sched_graph *graph)
   2629 {
   2630 	int i;
   2631 	int max;
   2632 
   2633 	max = isl_options_get_schedule_max_coefficient(ctx);
   2634 
   2635 	if (max == -1 && !isl_options_get_schedule_treat_coalescing(ctx))
   2636 		return isl_stat_ok;
   2637 
   2638 	for (i = 0; i < graph->n; ++i) {
   2639 		struct isl_sched_node *node = &graph->node[i];
   2640 
   2641 		if (node_add_coefficient_constraints(ctx, graph, node, max) < 0)
   2642 			return isl_stat_error;
   2643 	}
   2644 
   2645 	return isl_stat_ok;
   2646 }
   2647 
   2648 /* Add a constraint to graph->lp that equates the value at position
   2649  * "sum_pos" to the sum of the "n" values starting at "first".
   2650  */
   2651 static isl_stat add_sum_constraint(struct isl_sched_graph *graph,
   2652 	int sum_pos, int first, int n)
   2653 {
   2654 	int i, k;
   2655 	isl_size total;
   2656 
   2657 	total = isl_basic_set_dim(graph->lp, isl_dim_set);
   2658 	if (total < 0)
   2659 		return isl_stat_error;
   2660 
   2661 	k = isl_basic_set_alloc_equality(graph->lp);
   2662 	if (k < 0)
   2663 		return isl_stat_error;
   2664 	isl_seq_clr(graph->lp->eq[k], 1 + total);
   2665 	isl_int_set_si(graph->lp->eq[k][1 + sum_pos], -1);
   2666 	for (i = 0; i < n; ++i)
   2667 		isl_int_set_si(graph->lp->eq[k][1 + first + i], 1);
   2668 
   2669 	return isl_stat_ok;
   2670 }
   2671 
   2672 /* Add a constraint to graph->lp that equates the value at position
   2673  * "sum_pos" to the sum of the parameter coefficients of all nodes.
   2674  */
   2675 static isl_stat add_param_sum_constraint(struct isl_sched_graph *graph,
   2676 	int sum_pos)
   2677 {
   2678 	int i, j, k;
   2679 	isl_size total;
   2680 
   2681 	total = isl_basic_set_dim(graph->lp, isl_dim_set);
   2682 	if (total < 0)
   2683 		return isl_stat_error;
   2684 
   2685 	k = isl_basic_set_alloc_equality(graph->lp);
   2686 	if (k < 0)
   2687 		return isl_stat_error;
   2688 	isl_seq_clr(graph->lp->eq[k], 1 + total);
   2689 	isl_int_set_si(graph->lp->eq[k][1 + sum_pos], -1);
   2690 	for (i = 0; i < graph->n; ++i) {
   2691 		int pos = 1 + node_par_coef_offset(&graph->node[i]);
   2692 
   2693 		for (j = 0; j < graph->node[i].nparam; ++j)
   2694 			isl_int_set_si(graph->lp->eq[k][pos + j], 1);
   2695 	}
   2696 
   2697 	return isl_stat_ok;
   2698 }
   2699 
   2700 /* Add a constraint to graph->lp that equates the value at position
   2701  * "sum_pos" to the sum of the variable coefficients of all nodes.
   2702  */
   2703 static isl_stat add_var_sum_constraint(struct isl_sched_graph *graph,
   2704 	int sum_pos)
   2705 {
   2706 	int i, j, k;
   2707 	isl_size total;
   2708 
   2709 	total = isl_basic_set_dim(graph->lp, isl_dim_set);
   2710 	if (total < 0)
   2711 		return isl_stat_error;
   2712 
   2713 	k = isl_basic_set_alloc_equality(graph->lp);
   2714 	if (k < 0)
   2715 		return isl_stat_error;
   2716 	isl_seq_clr(graph->lp->eq[k], 1 + total);
   2717 	isl_int_set_si(graph->lp->eq[k][1 + sum_pos], -1);
   2718 	for (i = 0; i < graph->n; ++i) {
   2719 		struct isl_sched_node *node = &graph->node[i];
   2720 		int pos = 1 + node_var_coef_offset(node);
   2721 
   2722 		for (j = 0; j < 2 * node->nvar; ++j)
   2723 			isl_int_set_si(graph->lp->eq[k][pos + j], 1);
   2724 	}
   2725 
   2726 	return isl_stat_ok;
   2727 }
   2728 
   2729 /* Construct an ILP problem for finding schedule coefficients
   2730  * that result in non-negative, but small dependence distances
   2731  * over all dependences.
   2732  * In particular, the dependence distances over proximity edges
   2733  * are bounded by m_0 + m_n n and we compute schedule coefficients
   2734  * with small values (preferably zero) of m_n and m_0.
   2735  *
   2736  * All variables of the ILP are non-negative.  The actual coefficients
   2737  * may be negative, so each coefficient is represented as the difference
   2738  * of two non-negative variables.  The negative part always appears
   2739  * immediately before the positive part.
   2740  * Other than that, the variables have the following order
   2741  *
   2742  *	- sum of positive and negative parts of m_n coefficients
   2743  *	- m_0
   2744  *	- sum of all c_n coefficients
   2745  *		(unconstrained when computing non-parametric schedules)
   2746  *	- sum of positive and negative parts of all c_x coefficients
   2747  *	- positive and negative parts of m_n coefficients
   2748  *	- for each node
   2749  *		- positive and negative parts of c_i_x, in opposite order
   2750  *		- c_i_n (if parametric)
   2751  *		- c_i_0
   2752  *
   2753  * The constraints are those from the edges plus two or three equalities
   2754  * to express the sums.
   2755  *
   2756  * If "use_coincidence" is set, then we treat coincidence edges as local edges.
   2757  * Otherwise, we ignore them.
   2758  */
   2759 static isl_stat setup_lp(isl_ctx *ctx, struct isl_sched_graph *graph,
   2760 	int use_coincidence)
   2761 {
   2762 	int i;
   2763 	isl_size nparam;
   2764 	unsigned total;
   2765 	isl_space *space;
   2766 	int parametric;
   2767 	int param_pos;
   2768 	int n_eq, n_ineq;
   2769 
   2770 	parametric = ctx->opt->schedule_parametric;
   2771 	nparam = isl_space_dim(graph->node[0].space, isl_dim_param);
   2772 	if (nparam < 0)
   2773 		return isl_stat_error;
   2774 	param_pos = 4;
   2775 	total = param_pos + 2 * nparam;
   2776 	for (i = 0; i < graph->n; ++i) {
   2777 		struct isl_sched_node *node = &graph->node[graph->sorted[i]];
   2778 		if (isl_sched_node_update_vmap(node) < 0)
   2779 			return isl_stat_error;
   2780 		node->start = total;
   2781 		total += 1 + node->nparam + 2 * node->nvar;
   2782 	}
   2783 
   2784 	if (count_constraints(graph, &n_eq, &n_ineq, use_coincidence) < 0)
   2785 		return isl_stat_error;
   2786 	if (count_bound_constant_constraints(ctx, graph, &n_eq, &n_ineq) < 0)
   2787 		return isl_stat_error;
   2788 	if (count_bound_coefficient_constraints(ctx, graph, &n_eq, &n_ineq) < 0)
   2789 		return isl_stat_error;
   2790 
   2791 	space = isl_space_set_alloc(ctx, 0, total);
   2792 	isl_basic_set_free(graph->lp);
   2793 	n_eq += 2 + parametric;
   2794 
   2795 	graph->lp = isl_basic_set_alloc_space(space, 0, n_eq, n_ineq);
   2796 
   2797 	if (add_sum_constraint(graph, 0, param_pos, 2 * nparam) < 0)
   2798 		return isl_stat_error;
   2799 	if (parametric && add_param_sum_constraint(graph, 2) < 0)
   2800 		return isl_stat_error;
   2801 	if (add_var_sum_constraint(graph, 3) < 0)
   2802 		return isl_stat_error;
   2803 	if (add_bound_constant_constraints(ctx, graph) < 0)
   2804 		return isl_stat_error;
   2805 	if (add_bound_coefficient_constraints(ctx, graph) < 0)
   2806 		return isl_stat_error;
   2807 	if (add_all_validity_constraints(graph, use_coincidence) < 0)
   2808 		return isl_stat_error;
   2809 	if (add_all_proximity_constraints(graph, use_coincidence) < 0)
   2810 		return isl_stat_error;
   2811 
   2812 	return isl_stat_ok;
   2813 }
   2814 
   2815 /* Analyze the conflicting constraint found by
   2816  * isl_tab_basic_set_non_trivial_lexmin.  If it corresponds to the validity
   2817  * constraint of one of the edges between distinct nodes, living, moreover
   2818  * in distinct SCCs, then record the source and sink SCC as this may
   2819  * be a good place to cut between SCCs.
   2820  */
   2821 static int check_conflict(int con, void *user)
   2822 {
   2823 	int i;
   2824 	struct isl_sched_graph *graph = user;
   2825 
   2826 	if (graph->src_scc >= 0)
   2827 		return 0;
   2828 
   2829 	con -= graph->lp->n_eq;
   2830 
   2831 	if (con >= graph->lp->n_ineq)
   2832 		return 0;
   2833 
   2834 	for (i = 0; i < graph->n_edge; ++i) {
   2835 		if (!is_validity(&graph->edge[i]))
   2836 			continue;
   2837 		if (graph->edge[i].src == graph->edge[i].dst)
   2838 			continue;
   2839 		if (graph->edge[i].src->scc == graph->edge[i].dst->scc)
   2840 			continue;
   2841 		if (graph->edge[i].start > con)
   2842 			continue;
   2843 		if (graph->edge[i].end <= con)
   2844 			continue;
   2845 		graph->src_scc = graph->edge[i].src->scc;
   2846 		graph->dst_scc = graph->edge[i].dst->scc;
   2847 	}
   2848 
   2849 	return 0;
   2850 }
   2851 
   2852 /* Check whether the next schedule row of the given node needs to be
   2853  * non-trivial.  Lower-dimensional domains may have some trivial rows,
   2854  * but as soon as the number of remaining required non-trivial rows
   2855  * is as large as the number or remaining rows to be computed,
   2856  * all remaining rows need to be non-trivial.
   2857  */
   2858 static int needs_row(struct isl_sched_graph *graph, struct isl_sched_node *node)
   2859 {
   2860 	return node->nvar - node->rank >= graph->maxvar - graph->n_row;
   2861 }
   2862 
   2863 /* Construct a non-triviality region with triviality directions
   2864  * corresponding to the rows of "indep".
   2865  * The rows of "indep" are expressed in terms of the schedule coefficients c_i,
   2866  * while the triviality directions are expressed in terms of
   2867  * pairs of non-negative variables c^+_i - c^-_i, with c^-_i appearing
   2868  * before c^+_i.  Furthermore,
   2869  * the pairs of non-negative variables representing the coefficients
   2870  * are stored in the opposite order.
   2871  */
   2872 static __isl_give isl_mat *construct_trivial(__isl_keep isl_mat *indep)
   2873 {
   2874 	isl_ctx *ctx;
   2875 	isl_mat *mat;
   2876 	int i, j;
   2877 	isl_size n, n_var;
   2878 
   2879 	n = isl_mat_rows(indep);
   2880 	n_var = isl_mat_cols(indep);
   2881 	if (n < 0 || n_var < 0)
   2882 		return NULL;
   2883 
   2884 	ctx = isl_mat_get_ctx(indep);
   2885 	mat = isl_mat_alloc(ctx, n, 2 * n_var);
   2886 	if (!mat)
   2887 		return NULL;
   2888 	for (i = 0; i < n; ++i) {
   2889 		for (j = 0; j < n_var; ++j) {
   2890 			int nj = n_var - 1 - j;
   2891 			isl_int_neg(mat->row[i][2 * nj], indep->row[i][j]);
   2892 			isl_int_set(mat->row[i][2 * nj + 1], indep->row[i][j]);
   2893 		}
   2894 	}
   2895 
   2896 	return mat;
   2897 }
   2898 
   2899 /* Solve the ILP problem constructed in setup_lp.
   2900  * For each node such that all the remaining rows of its schedule
   2901  * need to be non-trivial, we construct a non-triviality region.
   2902  * This region imposes that the next row is independent of previous rows.
   2903  * In particular, the non-triviality region enforces that at least
   2904  * one of the linear combinations in the rows of node->indep is non-zero.
   2905  */
   2906 static __isl_give isl_vec *solve_lp(isl_ctx *ctx, struct isl_sched_graph *graph)
   2907 {
   2908 	int i;
   2909 	isl_vec *sol;
   2910 	isl_basic_set *lp;
   2911 
   2912 	for (i = 0; i < graph->n; ++i) {
   2913 		struct isl_sched_node *node = &graph->node[i];
   2914 		isl_mat *trivial;
   2915 
   2916 		graph->region[i].pos = node_var_coef_offset(node);
   2917 		if (needs_row(graph, node))
   2918 			trivial = construct_trivial(node->indep);
   2919 		else
   2920 			trivial = isl_mat_zero(ctx, 0, 0);
   2921 		graph->region[i].trivial = trivial;
   2922 	}
   2923 	lp = isl_basic_set_copy(graph->lp);
   2924 	sol = isl_tab_basic_set_non_trivial_lexmin(lp, 2, graph->n,
   2925 				       graph->region, &check_conflict, graph);
   2926 	for (i = 0; i < graph->n; ++i)
   2927 		isl_mat_free(graph->region[i].trivial);
   2928 	return sol;
   2929 }
   2930 
   2931 /* Extract the coefficients for the variables of "node" from "sol".
   2932  *
   2933  * Each schedule coefficient c_i_x is represented as the difference
   2934  * between two non-negative variables c_i_x^+ - c_i_x^-.
   2935  * The c_i_x^- appear before their c_i_x^+ counterpart.
   2936  * Furthermore, the order of these pairs is the opposite of that
   2937  * of the corresponding coefficients.
   2938  *
   2939  * Return c_i_x = c_i_x^+ - c_i_x^-
   2940  */
   2941 static __isl_give isl_vec *extract_var_coef(struct isl_sched_node *node,
   2942 	__isl_keep isl_vec *sol)
   2943 {
   2944 	int i;
   2945 	int pos;
   2946 	isl_vec *csol;
   2947 
   2948 	if (!sol)
   2949 		return NULL;
   2950 	csol = isl_vec_alloc(isl_vec_get_ctx(sol), node->nvar);
   2951 	if (!csol)
   2952 		return NULL;
   2953 
   2954 	pos = 1 + node_var_coef_offset(node);
   2955 	for (i = 0; i < node->nvar; ++i)
   2956 		isl_int_sub(csol->el[node->nvar - 1 - i],
   2957 			    sol->el[pos + 2 * i + 1], sol->el[pos + 2 * i]);
   2958 
   2959 	return csol;
   2960 }
   2961 
   2962 /* Update the schedules of all nodes based on the given solution
   2963  * of the LP problem.
   2964  * The new row is added to the current band.
   2965  * All possibly negative coefficients are encoded as a difference
   2966  * of two non-negative variables, so we need to perform the subtraction
   2967  * here.
   2968  *
   2969  * If coincident is set, then the caller guarantees that the new
   2970  * row satisfies the coincidence constraints.
   2971  */
   2972 static int update_schedule(struct isl_sched_graph *graph,
   2973 	__isl_take isl_vec *sol, int coincident)
   2974 {
   2975 	int i, j;
   2976 	isl_vec *csol = NULL;
   2977 
   2978 	if (!sol)
   2979 		goto error;
   2980 	if (sol->size == 0)
   2981 		isl_die(sol->ctx, isl_error_internal,
   2982 			"no solution found", goto error);
   2983 	if (graph->n_total_row >= graph->max_row)
   2984 		isl_die(sol->ctx, isl_error_internal,
   2985 			"too many schedule rows", goto error);
   2986 
   2987 	for (i = 0; i < graph->n; ++i) {
   2988 		struct isl_sched_node *node = &graph->node[i];
   2989 		int pos;
   2990 		isl_size row = isl_mat_rows(node->sched);
   2991 
   2992 		isl_vec_free(csol);
   2993 		csol = extract_var_coef(node, sol);
   2994 		if (row < 0 || !csol)
   2995 			goto error;
   2996 
   2997 		isl_map_free(node->sched_map);
   2998 		node->sched_map = NULL;
   2999 		node->sched = isl_mat_add_rows(node->sched, 1);
   3000 		if (!node->sched)
   3001 			goto error;
   3002 		pos = node_cst_coef_offset(node);
   3003 		node->sched = isl_mat_set_element(node->sched,
   3004 					row, 0, sol->el[1 + pos]);
   3005 		pos = node_par_coef_offset(node);
   3006 		for (j = 0; j < node->nparam; ++j)
   3007 			node->sched = isl_mat_set_element(node->sched,
   3008 					row, 1 + j, sol->el[1 + pos + j]);
   3009 		for (j = 0; j < node->nvar; ++j)
   3010 			node->sched = isl_mat_set_element(node->sched,
   3011 					row, 1 + node->nparam + j, csol->el[j]);
   3012 		node->coincident[graph->n_total_row] = coincident;
   3013 	}
   3014 	isl_vec_free(sol);
   3015 	isl_vec_free(csol);
   3016 
   3017 	graph->n_row++;
   3018 	graph->n_total_row++;
   3019 
   3020 	return 0;
   3021 error:
   3022 	isl_vec_free(sol);
   3023 	isl_vec_free(csol);
   3024 	return -1;
   3025 }
   3026 
   3027 /* Convert row "row" of node->sched into an isl_aff living in "ls"
   3028  * and return this isl_aff.
   3029  */
   3030 static __isl_give isl_aff *extract_schedule_row(__isl_take isl_local_space *ls,
   3031 	struct isl_sched_node *node, int row)
   3032 {
   3033 	int j;
   3034 	isl_int v;
   3035 	isl_aff *aff;
   3036 
   3037 	isl_int_init(v);
   3038 
   3039 	aff = isl_aff_zero_on_domain(ls);
   3040 	if (isl_mat_get_element(node->sched, row, 0, &v) < 0)
   3041 		goto error;
   3042 	aff = isl_aff_set_constant(aff, v);
   3043 	for (j = 0; j < node->nparam; ++j) {
   3044 		if (isl_mat_get_element(node->sched, row, 1 + j, &v) < 0)
   3045 			goto error;
   3046 		aff = isl_aff_set_coefficient(aff, isl_dim_param, j, v);
   3047 	}
   3048 	for (j = 0; j < node->nvar; ++j) {
   3049 		if (isl_mat_get_element(node->sched, row,
   3050 					1 + node->nparam + j, &v) < 0)
   3051 			goto error;
   3052 		aff = isl_aff_set_coefficient(aff, isl_dim_in, j, v);
   3053 	}
   3054 
   3055 	isl_int_clear(v);
   3056 
   3057 	return aff;
   3058 error:
   3059 	isl_int_clear(v);
   3060 	isl_aff_free(aff);
   3061 	return NULL;
   3062 }
   3063 
   3064 /* Convert the "n" rows starting at "first" of node->sched into a multi_aff
   3065  * and return this multi_aff.
   3066  *
   3067  * The result is defined over the uncompressed node domain.
   3068  */
   3069 __isl_give isl_multi_aff *isl_sched_node_extract_partial_schedule_multi_aff(
   3070 	struct isl_sched_node *node, int first, int n)
   3071 {
   3072 	int i;
   3073 	isl_space *space;
   3074 	isl_local_space *ls;
   3075 	isl_aff *aff;
   3076 	isl_multi_aff *ma;
   3077 	isl_size nrow;
   3078 
   3079 	if (!node)
   3080 		return NULL;
   3081 	nrow = isl_mat_rows(node->sched);
   3082 	if (nrow < 0)
   3083 		return NULL;
   3084 	if (node->compressed)
   3085 		space = isl_pw_multi_aff_get_domain_space(node->decompress);
   3086 	else
   3087 		space = isl_space_copy(node->space);
   3088 	ls = isl_local_space_from_space(isl_space_copy(space));
   3089 	space = isl_space_from_domain(space);
   3090 	space = isl_space_add_dims(space, isl_dim_out, n);
   3091 	ma = isl_multi_aff_zero(space);
   3092 
   3093 	for (i = first; i < first + n; ++i) {
   3094 		aff = extract_schedule_row(isl_local_space_copy(ls), node, i);
   3095 		ma = isl_multi_aff_set_aff(ma, i - first, aff);
   3096 	}
   3097 
   3098 	isl_local_space_free(ls);
   3099 
   3100 	if (node->compressed)
   3101 		ma = isl_multi_aff_pullback_multi_aff(ma,
   3102 					isl_multi_aff_copy(node->compress));
   3103 
   3104 	return ma;
   3105 }
   3106 
   3107 /* Convert node->sched into a multi_aff and return this multi_aff.
   3108  *
   3109  * The result is defined over the uncompressed node domain.
   3110  */
   3111 static __isl_give isl_multi_aff *node_extract_schedule_multi_aff(
   3112 	struct isl_sched_node *node)
   3113 {
   3114 	isl_size nrow;
   3115 
   3116 	nrow = isl_mat_rows(node->sched);
   3117 	if (nrow < 0)
   3118 		return NULL;
   3119 	return isl_sched_node_extract_partial_schedule_multi_aff(node, 0, nrow);
   3120 }
   3121 
   3122 /* Convert node->sched into a map and return this map.
   3123  *
   3124  * The result is cached in node->sched_map, which needs to be released
   3125  * whenever node->sched is updated.
   3126  * It is defined over the uncompressed node domain.
   3127  */
   3128 static __isl_give isl_map *node_extract_schedule(struct isl_sched_node *node)
   3129 {
   3130 	if (!node->sched_map) {
   3131 		isl_multi_aff *ma;
   3132 
   3133 		ma = node_extract_schedule_multi_aff(node);
   3134 		node->sched_map = isl_map_from_multi_aff(ma);
   3135 	}
   3136 
   3137 	return isl_map_copy(node->sched_map);
   3138 }
   3139 
   3140 /* Construct a map that can be used to update a dependence relation
   3141  * based on the current schedule.
   3142  * That is, construct a map expressing that source and sink
   3143  * are executed within the same iteration of the current schedule.
   3144  * This map can then be intersected with the dependence relation.
   3145  * This is not the most efficient way, but this shouldn't be a critical
   3146  * operation.
   3147  */
   3148 static __isl_give isl_map *specializer(struct isl_sched_node *src,
   3149 	struct isl_sched_node *dst)
   3150 {
   3151 	isl_map *src_sched, *dst_sched;
   3152 
   3153 	src_sched = node_extract_schedule(src);
   3154 	dst_sched = node_extract_schedule(dst);
   3155 	return isl_map_apply_range(src_sched, isl_map_reverse(dst_sched));
   3156 }
   3157 
   3158 /* Intersect the domains of the nested relations in domain and range
   3159  * of "umap" with "map".
   3160  */
   3161 static __isl_give isl_union_map *intersect_domains(
   3162 	__isl_take isl_union_map *umap, __isl_keep isl_map *map)
   3163 {
   3164 	isl_union_set *uset;
   3165 
   3166 	umap = isl_union_map_zip(umap);
   3167 	uset = isl_union_set_from_set(isl_map_wrap(isl_map_copy(map)));
   3168 	umap = isl_union_map_intersect_domain(umap, uset);
   3169 	umap = isl_union_map_zip(umap);
   3170 	return umap;
   3171 }
   3172 
   3173 /* Update the dependence relation of the given edge based
   3174  * on the current schedule.
   3175  * If the dependence is carried completely by the current schedule, then
   3176  * it is removed from the edge_tables.  It is kept in the list of edges
   3177  * as otherwise all edge_tables would have to be recomputed.
   3178  *
   3179  * If the edge is of a type that can appear multiple times
   3180  * between the same pair of nodes, then it is added to
   3181  * the edge table (again).  This prevents the situation
   3182  * where none of these edges is referenced from the edge table
   3183  * because the one that was referenced turned out to be empty and
   3184  * was therefore removed from the table.
   3185  */
   3186 static isl_stat update_edge(isl_ctx *ctx, struct isl_sched_graph *graph,
   3187 	struct isl_sched_edge *edge)
   3188 {
   3189 	int empty;
   3190 	isl_map *id;
   3191 
   3192 	id = specializer(edge->src, edge->dst);
   3193 	edge->map = isl_map_intersect(edge->map, isl_map_copy(id));
   3194 	if (!edge->map)
   3195 		goto error;
   3196 
   3197 	if (edge->tagged_condition) {
   3198 		edge->tagged_condition =
   3199 			intersect_domains(edge->tagged_condition, id);
   3200 		if (!edge->tagged_condition)
   3201 			goto error;
   3202 	}
   3203 	if (edge->tagged_validity) {
   3204 		edge->tagged_validity =
   3205 			intersect_domains(edge->tagged_validity, id);
   3206 		if (!edge->tagged_validity)
   3207 			goto error;
   3208 	}
   3209 
   3210 	empty = isl_map_plain_is_empty(edge->map);
   3211 	if (empty < 0)
   3212 		goto error;
   3213 	if (empty) {
   3214 		if (graph_remove_edge(graph, edge) < 0)
   3215 			goto error;
   3216 	} else if (is_multi_edge_type(edge)) {
   3217 		if (graph_edge_tables_add(ctx, graph, edge) < 0)
   3218 			goto error;
   3219 	}
   3220 
   3221 	isl_map_free(id);
   3222 	return isl_stat_ok;
   3223 error:
   3224 	isl_map_free(id);
   3225 	return isl_stat_error;
   3226 }
   3227 
   3228 /* Does the domain of "umap" intersect "uset"?
   3229  */
   3230 static int domain_intersects(__isl_keep isl_union_map *umap,
   3231 	__isl_keep isl_union_set *uset)
   3232 {
   3233 	int empty;
   3234 
   3235 	umap = isl_union_map_copy(umap);
   3236 	umap = isl_union_map_intersect_domain(umap, isl_union_set_copy(uset));
   3237 	empty = isl_union_map_is_empty(umap);
   3238 	isl_union_map_free(umap);
   3239 
   3240 	return empty < 0 ? -1 : !empty;
   3241 }
   3242 
   3243 /* Does the range of "umap" intersect "uset"?
   3244  */
   3245 static int range_intersects(__isl_keep isl_union_map *umap,
   3246 	__isl_keep isl_union_set *uset)
   3247 {
   3248 	int empty;
   3249 
   3250 	umap = isl_union_map_copy(umap);
   3251 	umap = isl_union_map_intersect_range(umap, isl_union_set_copy(uset));
   3252 	empty = isl_union_map_is_empty(umap);
   3253 	isl_union_map_free(umap);
   3254 
   3255 	return empty < 0 ? -1 : !empty;
   3256 }
   3257 
   3258 /* Are the condition dependences of "edge" local with respect to
   3259  * the current schedule?
   3260  *
   3261  * That is, are domain and range of the condition dependences mapped
   3262  * to the same point?
   3263  *
   3264  * In other words, is the condition false?
   3265  */
   3266 static int is_condition_false(struct isl_sched_edge *edge)
   3267 {
   3268 	isl_union_map *umap;
   3269 	isl_map *map, *sched, *test;
   3270 	int empty, local;
   3271 
   3272 	empty = isl_union_map_is_empty(edge->tagged_condition);
   3273 	if (empty < 0 || empty)
   3274 		return empty;
   3275 
   3276 	umap = isl_union_map_copy(edge->tagged_condition);
   3277 	umap = isl_union_map_zip(umap);
   3278 	umap = isl_union_set_unwrap(isl_union_map_domain(umap));
   3279 	map = isl_map_from_union_map(umap);
   3280 
   3281 	sched = node_extract_schedule(edge->src);
   3282 	map = isl_map_apply_domain(map, sched);
   3283 	sched = node_extract_schedule(edge->dst);
   3284 	map = isl_map_apply_range(map, sched);
   3285 
   3286 	test = isl_map_identity(isl_map_get_space(map));
   3287 	local = isl_map_is_subset(map, test);
   3288 	isl_map_free(map);
   3289 	isl_map_free(test);
   3290 
   3291 	return local;
   3292 }
   3293 
   3294 /* For each conditional validity constraint that is adjacent
   3295  * to a condition with domain in condition_source or range in condition_sink,
   3296  * turn it into an unconditional validity constraint.
   3297  */
   3298 static int unconditionalize_adjacent_validity(struct isl_sched_graph *graph,
   3299 	__isl_take isl_union_set *condition_source,
   3300 	__isl_take isl_union_set *condition_sink)
   3301 {
   3302 	int i;
   3303 
   3304 	condition_source = isl_union_set_coalesce(condition_source);
   3305 	condition_sink = isl_union_set_coalesce(condition_sink);
   3306 
   3307 	for (i = 0; i < graph->n_edge; ++i) {
   3308 		int adjacent;
   3309 		isl_union_map *validity;
   3310 
   3311 		if (!isl_sched_edge_is_conditional_validity(&graph->edge[i]))
   3312 			continue;
   3313 		if (is_validity(&graph->edge[i]))
   3314 			continue;
   3315 
   3316 		validity = graph->edge[i].tagged_validity;
   3317 		adjacent = domain_intersects(validity, condition_sink);
   3318 		if (adjacent >= 0 && !adjacent)
   3319 			adjacent = range_intersects(validity, condition_source);
   3320 		if (adjacent < 0)
   3321 			goto error;
   3322 		if (!adjacent)
   3323 			continue;
   3324 
   3325 		set_validity(&graph->edge[i]);
   3326 	}
   3327 
   3328 	isl_union_set_free(condition_source);
   3329 	isl_union_set_free(condition_sink);
   3330 	return 0;
   3331 error:
   3332 	isl_union_set_free(condition_source);
   3333 	isl_union_set_free(condition_sink);
   3334 	return -1;
   3335 }
   3336 
   3337 /* Update the dependence relations of all edges based on the current schedule
   3338  * and enforce conditional validity constraints that are adjacent
   3339  * to satisfied condition constraints.
   3340  *
   3341  * First check if any of the condition constraints are satisfied
   3342  * (i.e., not local to the outer schedule) and keep track of
   3343  * their domain and range.
   3344  * Then update all dependence relations (which removes the non-local
   3345  * constraints).
   3346  * Finally, if any condition constraints turned out to be satisfied,
   3347  * then turn all adjacent conditional validity constraints into
   3348  * unconditional validity constraints.
   3349  */
   3350 static int update_edges(isl_ctx *ctx, struct isl_sched_graph *graph)
   3351 {
   3352 	int i;
   3353 	int any = 0;
   3354 	isl_union_set *source, *sink;
   3355 
   3356 	source = isl_union_set_empty(isl_space_params_alloc(ctx, 0));
   3357 	sink = isl_union_set_empty(isl_space_params_alloc(ctx, 0));
   3358 	for (i = 0; i < graph->n_edge; ++i) {
   3359 		int local;
   3360 		isl_union_set *uset;
   3361 		isl_union_map *umap;
   3362 
   3363 		if (!isl_sched_edge_is_condition(&graph->edge[i]))
   3364 			continue;
   3365 		if (is_local(&graph->edge[i]))
   3366 			continue;
   3367 		local = is_condition_false(&graph->edge[i]);
   3368 		if (local < 0)
   3369 			goto error;
   3370 		if (local)
   3371 			continue;
   3372 
   3373 		any = 1;
   3374 
   3375 		umap = isl_union_map_copy(graph->edge[i].tagged_condition);
   3376 		uset = isl_union_map_domain(umap);
   3377 		source = isl_union_set_union(source, uset);
   3378 
   3379 		umap = isl_union_map_copy(graph->edge[i].tagged_condition);
   3380 		uset = isl_union_map_range(umap);
   3381 		sink = isl_union_set_union(sink, uset);
   3382 	}
   3383 
   3384 	for (i = 0; i < graph->n_edge; ++i) {
   3385 		if (update_edge(ctx, graph, &graph->edge[i]) < 0)
   3386 			goto error;
   3387 	}
   3388 
   3389 	if (any)
   3390 		return unconditionalize_adjacent_validity(graph, source, sink);
   3391 
   3392 	isl_union_set_free(source);
   3393 	isl_union_set_free(sink);
   3394 	return 0;
   3395 error:
   3396 	isl_union_set_free(source);
   3397 	isl_union_set_free(sink);
   3398 	return -1;
   3399 }
   3400 
   3401 static void next_band(struct isl_sched_graph *graph)
   3402 {
   3403 	graph->band_start = graph->n_total_row;
   3404 }
   3405 
   3406 /* Return the union of the universe domains of the nodes in "graph"
   3407  * that satisfy "pred".
   3408  */
   3409 static __isl_give isl_union_set *isl_sched_graph_domain(isl_ctx *ctx,
   3410 	struct isl_sched_graph *graph,
   3411 	int (*pred)(struct isl_sched_node *node, int data), int data)
   3412 {
   3413 	int i;
   3414 	isl_set *set;
   3415 	isl_union_set *dom;
   3416 
   3417 	for (i = 0; i < graph->n; ++i)
   3418 		if (pred(&graph->node[i], data))
   3419 			break;
   3420 
   3421 	if (i >= graph->n)
   3422 		isl_die(ctx, isl_error_internal,
   3423 			"empty component", return NULL);
   3424 
   3425 	set = isl_set_universe(isl_space_copy(graph->node[i].space));
   3426 	dom = isl_union_set_from_set(set);
   3427 
   3428 	for (i = i + 1; i < graph->n; ++i) {
   3429 		if (!pred(&graph->node[i], data))
   3430 			continue;
   3431 		set = isl_set_universe(isl_space_copy(graph->node[i].space));
   3432 		dom = isl_union_set_union(dom, isl_union_set_from_set(set));
   3433 	}
   3434 
   3435 	return dom;
   3436 }
   3437 
   3438 /* Return a union of universe domains corresponding to the nodes
   3439  * in the SCC with index "scc".
   3440  */
   3441 __isl_give isl_union_set *isl_sched_graph_extract_scc(isl_ctx *ctx,
   3442 	struct isl_sched_graph *graph, int scc)
   3443 {
   3444 	return isl_sched_graph_domain(ctx, graph,
   3445 					&isl_sched_node_scc_exactly, scc);
   3446 }
   3447 
   3448 /* Return a list of unions of universe domains, where each element
   3449  * in the list corresponds to an SCC (or WCC) indexed by node->scc.
   3450  */
   3451 __isl_give isl_union_set_list *isl_sched_graph_extract_sccs(isl_ctx *ctx,
   3452 	struct isl_sched_graph *graph)
   3453 {
   3454 	int i;
   3455 	isl_union_set_list *filters;
   3456 
   3457 	filters = isl_union_set_list_alloc(ctx, graph->scc);
   3458 	for (i = 0; i < graph->scc; ++i) {
   3459 		isl_union_set *dom;
   3460 
   3461 		dom = isl_sched_graph_extract_scc(ctx, graph, i);
   3462 		filters = isl_union_set_list_add(filters, dom);
   3463 	}
   3464 
   3465 	return filters;
   3466 }
   3467 
   3468 /* Return a list of two unions of universe domains, one for the SCCs up
   3469  * to and including graph->src_scc and another for the other SCCs.
   3470  */
   3471 static __isl_give isl_union_set_list *extract_split(isl_ctx *ctx,
   3472 	struct isl_sched_graph *graph)
   3473 {
   3474 	isl_union_set *dom;
   3475 	isl_union_set_list *filters;
   3476 
   3477 	filters = isl_union_set_list_alloc(ctx, 2);
   3478 	dom = isl_sched_graph_domain(ctx, graph,
   3479 					&node_scc_at_most, graph->src_scc);
   3480 	filters = isl_union_set_list_add(filters, dom);
   3481 	dom = isl_sched_graph_domain(ctx, graph,
   3482 					&node_scc_at_least, graph->src_scc + 1);
   3483 	filters = isl_union_set_list_add(filters, dom);
   3484 
   3485 	return filters;
   3486 }
   3487 
   3488 /* Copy nodes that satisfy node_pred from the src dependence graph
   3489  * to the dst dependence graph.
   3490  */
   3491 static isl_stat copy_nodes(struct isl_sched_graph *dst,
   3492 	struct isl_sched_graph *src,
   3493 	int (*node_pred)(struct isl_sched_node *node, int data), int data)
   3494 {
   3495 	int i;
   3496 
   3497 	dst->n = 0;
   3498 	for (i = 0; i < src->n; ++i) {
   3499 		int j;
   3500 
   3501 		if (!node_pred(&src->node[i], data))
   3502 			continue;
   3503 
   3504 		j = dst->n;
   3505 		dst->node[j].space = isl_space_copy(src->node[i].space);
   3506 		dst->node[j].compressed = src->node[i].compressed;
   3507 		dst->node[j].hull = isl_set_copy(src->node[i].hull);
   3508 		dst->node[j].compress =
   3509 			isl_multi_aff_copy(src->node[i].compress);
   3510 		dst->node[j].decompress =
   3511 			isl_pw_multi_aff_copy(src->node[i].decompress);
   3512 		dst->node[j].nvar = src->node[i].nvar;
   3513 		dst->node[j].nparam = src->node[i].nparam;
   3514 		dst->node[j].sched = isl_mat_copy(src->node[i].sched);
   3515 		dst->node[j].sched_map = isl_map_copy(src->node[i].sched_map);
   3516 		dst->node[j].coincident = src->node[i].coincident;
   3517 		dst->node[j].sizes = isl_multi_val_copy(src->node[i].sizes);
   3518 		dst->node[j].bounds = isl_basic_set_copy(src->node[i].bounds);
   3519 		dst->node[j].max = isl_vec_copy(src->node[i].max);
   3520 		dst->n++;
   3521 
   3522 		if (!dst->node[j].space || !dst->node[j].sched)
   3523 			return isl_stat_error;
   3524 		if (dst->node[j].compressed &&
   3525 		    (!dst->node[j].hull || !dst->node[j].compress ||
   3526 		     !dst->node[j].decompress))
   3527 			return isl_stat_error;
   3528 	}
   3529 
   3530 	return isl_stat_ok;
   3531 }
   3532 
   3533 /* Copy non-empty edges that satisfy edge_pred from the src dependence graph
   3534  * to the dst dependence graph.
   3535  * If the source or destination node of the edge is not in the destination
   3536  * graph, then it must be a backward proximity edge and it should simply
   3537  * be ignored.
   3538  */
   3539 static isl_stat copy_edges(isl_ctx *ctx, struct isl_sched_graph *dst,
   3540 	struct isl_sched_graph *src,
   3541 	int (*edge_pred)(struct isl_sched_edge *edge, int data), int data)
   3542 {
   3543 	int i;
   3544 
   3545 	dst->n_edge = 0;
   3546 	for (i = 0; i < src->n_edge; ++i) {
   3547 		struct isl_sched_edge *edge = &src->edge[i];
   3548 		isl_map *map;
   3549 		isl_union_map *tagged_condition;
   3550 		isl_union_map *tagged_validity;
   3551 		struct isl_sched_node *dst_src, *dst_dst;
   3552 
   3553 		if (!edge_pred(edge, data))
   3554 			continue;
   3555 
   3556 		if (isl_map_plain_is_empty(edge->map))
   3557 			continue;
   3558 
   3559 		dst_src = isl_sched_graph_find_node(ctx, dst, edge->src->space);
   3560 		dst_dst = isl_sched_graph_find_node(ctx, dst, edge->dst->space);
   3561 		if (!dst_src || !dst_dst)
   3562 			return isl_stat_error;
   3563 		if (!isl_sched_graph_is_node(dst, dst_src) ||
   3564 		    !isl_sched_graph_is_node(dst, dst_dst)) {
   3565 			if (is_validity(edge) ||
   3566 			    isl_sched_edge_is_conditional_validity(edge))
   3567 				isl_die(ctx, isl_error_internal,
   3568 					"backward (conditional) validity edge",
   3569 					return isl_stat_error);
   3570 			continue;
   3571 		}
   3572 
   3573 		map = isl_map_copy(edge->map);
   3574 		tagged_condition = isl_union_map_copy(edge->tagged_condition);
   3575 		tagged_validity = isl_union_map_copy(edge->tagged_validity);
   3576 
   3577 		dst->edge[dst->n_edge].src = dst_src;
   3578 		dst->edge[dst->n_edge].dst = dst_dst;
   3579 		dst->edge[dst->n_edge].map = map;
   3580 		dst->edge[dst->n_edge].tagged_condition = tagged_condition;
   3581 		dst->edge[dst->n_edge].tagged_validity = tagged_validity;
   3582 		dst->edge[dst->n_edge].types = edge->types;
   3583 		dst->n_edge++;
   3584 
   3585 		if (edge->tagged_condition && !tagged_condition)
   3586 			return isl_stat_error;
   3587 		if (edge->tagged_validity && !tagged_validity)
   3588 			return isl_stat_error;
   3589 
   3590 		if (graph_edge_tables_add(ctx, dst,
   3591 					    &dst->edge[dst->n_edge - 1]) < 0)
   3592 			return isl_stat_error;
   3593 	}
   3594 
   3595 	return isl_stat_ok;
   3596 }
   3597 
   3598 /* Compute the maximal number of variables over all nodes.
   3599  * This is the maximal number of linearly independent schedule
   3600  * rows that we need to compute.
   3601  * Just in case we end up in a part of the dependence graph
   3602  * with only lower-dimensional domains, we make sure we will
   3603  * compute the required amount of extra linearly independent rows.
   3604  */
   3605 isl_stat isl_sched_graph_compute_maxvar(struct isl_sched_graph *graph)
   3606 {
   3607 	int i;
   3608 
   3609 	graph->maxvar = 0;
   3610 	for (i = 0; i < graph->n; ++i) {
   3611 		struct isl_sched_node *node = &graph->node[i];
   3612 		int nvar;
   3613 
   3614 		if (isl_sched_node_update_vmap(node) < 0)
   3615 			return isl_stat_error;
   3616 		nvar = node->nvar + graph->n_row - node->rank;
   3617 		if (nvar > graph->maxvar)
   3618 			graph->maxvar = nvar;
   3619 	}
   3620 
   3621 	return isl_stat_ok;
   3622 }
   3623 
   3624 /* Extract the subgraph of "graph" that consists of the nodes satisfying
   3625  * "node_pred" and the edges satisfying "edge_pred" and store
   3626  * the result in "sub".
   3627  */
   3628 isl_stat isl_sched_graph_extract_sub_graph(isl_ctx *ctx,
   3629 	struct isl_sched_graph *graph,
   3630 	int (*node_pred)(struct isl_sched_node *node, int data),
   3631 	int (*edge_pred)(struct isl_sched_edge *edge, int data),
   3632 	int data, struct isl_sched_graph *sub)
   3633 {
   3634 	int i, n = 0, n_edge = 0;
   3635 	int t;
   3636 
   3637 	for (i = 0; i < graph->n; ++i)
   3638 		if (node_pred(&graph->node[i], data))
   3639 			++n;
   3640 	for (i = 0; i < graph->n_edge; ++i)
   3641 		if (edge_pred(&graph->edge[i], data))
   3642 			++n_edge;
   3643 	if (graph_alloc(ctx, sub, n, n_edge) < 0)
   3644 		return isl_stat_error;
   3645 	sub->root = graph->root;
   3646 	if (copy_nodes(sub, graph, node_pred, data) < 0)
   3647 		return isl_stat_error;
   3648 	if (graph_init_table(ctx, sub) < 0)
   3649 		return isl_stat_error;
   3650 	for (t = 0; t <= isl_edge_last; ++t)
   3651 		sub->max_edge[t] = graph->max_edge[t];
   3652 	if (graph_init_edge_tables(ctx, sub) < 0)
   3653 		return isl_stat_error;
   3654 	if (copy_edges(ctx, sub, graph, edge_pred, data) < 0)
   3655 		return isl_stat_error;
   3656 	sub->n_row = graph->n_row;
   3657 	sub->max_row = graph->max_row;
   3658 	sub->n_total_row = graph->n_total_row;
   3659 	sub->band_start = graph->band_start;
   3660 
   3661 	return isl_stat_ok;
   3662 }
   3663 
   3664 static __isl_give isl_schedule_node *compute_schedule(isl_schedule_node *node,
   3665 	struct isl_sched_graph *graph);
   3666 static __isl_give isl_schedule_node *compute_schedule_wcc(
   3667 	isl_schedule_node *node, struct isl_sched_graph *graph);
   3668 
   3669 /* Compute a schedule for a subgraph of "graph".  In particular, for
   3670  * the graph composed of nodes that satisfy node_pred and edges that
   3671  * that satisfy edge_pred.
   3672  * If the subgraph is known to consist of a single component, then wcc should
   3673  * be set and then we call compute_schedule_wcc on the constructed subgraph.
   3674  * Otherwise, we call compute_schedule, which will check whether the subgraph
   3675  * is connected.
   3676  *
   3677  * The schedule is inserted at "node" and the updated schedule node
   3678  * is returned.
   3679  */
   3680 static __isl_give isl_schedule_node *compute_sub_schedule(
   3681 	__isl_take isl_schedule_node *node, isl_ctx *ctx,
   3682 	struct isl_sched_graph *graph,
   3683 	int (*node_pred)(struct isl_sched_node *node, int data),
   3684 	int (*edge_pred)(struct isl_sched_edge *edge, int data),
   3685 	int data, int wcc)
   3686 {
   3687 	struct isl_sched_graph split = { 0 };
   3688 
   3689 	if (isl_sched_graph_extract_sub_graph(ctx, graph, node_pred, edge_pred,
   3690 						data, &split) < 0)
   3691 		goto error;
   3692 
   3693 	if (wcc)
   3694 		node = compute_schedule_wcc(node, &split);
   3695 	else
   3696 		node = compute_schedule(node, &split);
   3697 
   3698 	isl_sched_graph_free(ctx, &split);
   3699 	return node;
   3700 error:
   3701 	isl_sched_graph_free(ctx, &split);
   3702 	return isl_schedule_node_free(node);
   3703 }
   3704 
   3705 int isl_sched_edge_scc_exactly(struct isl_sched_edge *edge, int scc)
   3706 {
   3707 	return edge->src->scc == scc && edge->dst->scc == scc;
   3708 }
   3709 
   3710 static int edge_dst_scc_at_most(struct isl_sched_edge *edge, int scc)
   3711 {
   3712 	return edge->dst->scc <= scc;
   3713 }
   3714 
   3715 static int edge_src_scc_at_least(struct isl_sched_edge *edge, int scc)
   3716 {
   3717 	return edge->src->scc >= scc;
   3718 }
   3719 
   3720 /* Reset the current band by dropping all its schedule rows.
   3721  */
   3722 static isl_stat reset_band(struct isl_sched_graph *graph)
   3723 {
   3724 	int i;
   3725 	int drop;
   3726 
   3727 	drop = graph->n_total_row - graph->band_start;
   3728 	graph->n_total_row -= drop;
   3729 	graph->n_row -= drop;
   3730 
   3731 	for (i = 0; i < graph->n; ++i) {
   3732 		struct isl_sched_node *node = &graph->node[i];
   3733 
   3734 		isl_map_free(node->sched_map);
   3735 		node->sched_map = NULL;
   3736 
   3737 		node->sched = isl_mat_drop_rows(node->sched,
   3738 						graph->band_start, drop);
   3739 
   3740 		if (!node->sched)
   3741 			return isl_stat_error;
   3742 	}
   3743 
   3744 	return isl_stat_ok;
   3745 }
   3746 
   3747 /* Split the current graph into two parts and compute a schedule for each
   3748  * part individually.  In particular, one part consists of all SCCs up
   3749  * to and including graph->src_scc, while the other part contains the other
   3750  * SCCs.  The split is enforced by a sequence node inserted at position "node"
   3751  * in the schedule tree.  Return the updated schedule node.
   3752  * If either of these two parts consists of a sequence, then it is spliced
   3753  * into the sequence containing the two parts.
   3754  *
   3755  * The current band is reset. It would be possible to reuse
   3756  * the previously computed rows as the first rows in the next
   3757  * band, but recomputing them may result in better rows as we are looking
   3758  * at a smaller part of the dependence graph.
   3759  */
   3760 static __isl_give isl_schedule_node *compute_split_schedule(
   3761 	__isl_take isl_schedule_node *node, struct isl_sched_graph *graph)
   3762 {
   3763 	isl_ctx *ctx;
   3764 	isl_union_set_list *filters;
   3765 
   3766 	if (!node)
   3767 		return NULL;
   3768 
   3769 	if (reset_band(graph) < 0)
   3770 		return isl_schedule_node_free(node);
   3771 
   3772 	next_band(graph);
   3773 
   3774 	ctx = isl_schedule_node_get_ctx(node);
   3775 	filters = extract_split(ctx, graph);
   3776 	node = isl_schedule_node_insert_sequence(node, filters);
   3777 	node = isl_schedule_node_grandchild(node, 1, 0);
   3778 
   3779 	node = compute_sub_schedule(node, ctx, graph,
   3780 				&node_scc_at_least, &edge_src_scc_at_least,
   3781 				graph->src_scc + 1, 0);
   3782 	node = isl_schedule_node_grandparent(node);
   3783 	node = isl_schedule_node_grandchild(node, 0, 0);
   3784 	node = compute_sub_schedule(node, ctx, graph,
   3785 				&node_scc_at_most, &edge_dst_scc_at_most,
   3786 				graph->src_scc, 0);
   3787 	node = isl_schedule_node_grandparent(node);
   3788 
   3789 	node = isl_schedule_node_sequence_splice_children(node);
   3790 
   3791 	return node;
   3792 }
   3793 
   3794 /* Insert a band node at position "node" in the schedule tree corresponding
   3795  * to the current band in "graph".  Mark the band node permutable
   3796  * if "permutable" is set.
   3797  * The partial schedules and the coincidence property are extracted
   3798  * from the graph nodes.
   3799  * Return the updated schedule node.
   3800  */
   3801 static __isl_give isl_schedule_node *insert_current_band(
   3802 	__isl_take isl_schedule_node *node, struct isl_sched_graph *graph,
   3803 	int permutable)
   3804 {
   3805 	int i;
   3806 	int start, end, n;
   3807 	isl_multi_aff *ma;
   3808 	isl_multi_pw_aff *mpa;
   3809 	isl_multi_union_pw_aff *mupa;
   3810 
   3811 	if (!node)
   3812 		return NULL;
   3813 
   3814 	if (graph->n < 1)
   3815 		isl_die(isl_schedule_node_get_ctx(node), isl_error_internal,
   3816 			"graph should have at least one node",
   3817 			return isl_schedule_node_free(node));
   3818 
   3819 	start = graph->band_start;
   3820 	end = graph->n_total_row;
   3821 	n = end - start;
   3822 
   3823 	ma = isl_sched_node_extract_partial_schedule_multi_aff(&graph->node[0],
   3824 								start, n);
   3825 	mpa = isl_multi_pw_aff_from_multi_aff(ma);
   3826 	mupa = isl_multi_union_pw_aff_from_multi_pw_aff(mpa);
   3827 
   3828 	for (i = 1; i < graph->n; ++i) {
   3829 		isl_multi_union_pw_aff *mupa_i;
   3830 
   3831 		ma = isl_sched_node_extract_partial_schedule_multi_aff(
   3832 						&graph->node[i], start, n);
   3833 		mpa = isl_multi_pw_aff_from_multi_aff(ma);
   3834 		mupa_i = isl_multi_union_pw_aff_from_multi_pw_aff(mpa);
   3835 		mupa = isl_multi_union_pw_aff_union_add(mupa, mupa_i);
   3836 	}
   3837 	node = isl_schedule_node_insert_partial_schedule(node, mupa);
   3838 
   3839 	for (i = 0; i < n; ++i)
   3840 		node = isl_schedule_node_band_member_set_coincident(node, i,
   3841 					graph->node[0].coincident[start + i]);
   3842 	node = isl_schedule_node_band_set_permutable(node, permutable);
   3843 
   3844 	return node;
   3845 }
   3846 
   3847 /* Update the dependence relations based on the current schedule,
   3848  * add the current band to "node" and then continue with the computation
   3849  * of the next band.
   3850  * Return the updated schedule node.
   3851  */
   3852 static __isl_give isl_schedule_node *compute_next_band(
   3853 	__isl_take isl_schedule_node *node,
   3854 	struct isl_sched_graph *graph, int permutable)
   3855 {
   3856 	isl_ctx *ctx;
   3857 
   3858 	if (!node)
   3859 		return NULL;
   3860 
   3861 	ctx = isl_schedule_node_get_ctx(node);
   3862 	if (update_edges(ctx, graph) < 0)
   3863 		return isl_schedule_node_free(node);
   3864 	node = insert_current_band(node, graph, permutable);
   3865 	next_band(graph);
   3866 
   3867 	node = isl_schedule_node_child(node, 0);
   3868 	node = compute_schedule(node, graph);
   3869 	node = isl_schedule_node_parent(node);
   3870 
   3871 	return node;
   3872 }
   3873 
   3874 /* Add the constraints "coef" derived from an edge from "node" to itself
   3875  * to graph->lp in order to respect the dependences and to try and carry them.
   3876  * "pos" is the sequence number of the edge that needs to be carried.
   3877  * "coef" represents general constraints on coefficients (c_0, c_x)
   3878  * of valid constraints for (y - x) with x and y instances of the node.
   3879  *
   3880  * The constraints added to graph->lp need to enforce
   3881  *
   3882  *	(c_j_0 + c_j_x y) - (c_j_0 + c_j_x x)
   3883  *	= c_j_x (y - x) >= e_i
   3884  *
   3885  * for each (x,y) in the dependence relation of the edge.
   3886  * That is, (-e_i, c_j_x) needs to be plugged in for (c_0, c_x),
   3887  * taking into account that each coefficient in c_j_x is represented
   3888  * as a pair of non-negative coefficients.
   3889  */
   3890 static isl_stat add_intra_constraints(struct isl_sched_graph *graph,
   3891 	struct isl_sched_node *node, __isl_take isl_basic_set *coef, int pos)
   3892 {
   3893 	isl_size offset;
   3894 	isl_ctx *ctx;
   3895 	isl_dim_map *dim_map;
   3896 
   3897 	offset = coef_var_offset(coef);
   3898 	if (offset < 0)
   3899 		coef = isl_basic_set_free(coef);
   3900 	if (!coef)
   3901 		return isl_stat_error;
   3902 
   3903 	ctx = isl_basic_set_get_ctx(coef);
   3904 	dim_map = intra_dim_map(ctx, graph, node, offset, 1);
   3905 	isl_dim_map_range(dim_map, 3 + pos, 0, 0, 0, 1, -1);
   3906 	graph->lp = add_constraints_dim_map(graph->lp, coef, dim_map);
   3907 
   3908 	return isl_stat_ok;
   3909 }
   3910 
   3911 /* Add the constraints "coef" derived from an edge from "src" to "dst"
   3912  * to graph->lp in order to respect the dependences and to try and carry them.
   3913  * "pos" is the sequence number of the edge that needs to be carried or
   3914  * -1 if no attempt should be made to carry the dependences.
   3915  * "coef" represents general constraints on coefficients (c_0, c_n, c_x, c_y)
   3916  * of valid constraints for (x, y) with x and y instances of "src" and "dst".
   3917  *
   3918  * The constraints added to graph->lp need to enforce
   3919  *
   3920  *	(c_k_0 + c_k_n n + c_k_x y) - (c_j_0 + c_j_n n + c_j_x x) >= e_i
   3921  *
   3922  * for each (x,y) in the dependence relation of the edge or
   3923  *
   3924  *	(c_k_0 + c_k_n n + c_k_x y) - (c_j_0 + c_j_n n + c_j_x x) >= 0
   3925  *
   3926  * if pos is -1.
   3927  * That is,
   3928  * (-e_i + c_k_0 - c_j_0, c_k_n - c_j_n, -c_j_x, c_k_x)
   3929  * or
   3930  * (c_k_0 - c_j_0, c_k_n - c_j_n, -c_j_x, c_k_x)
   3931  * needs to be plugged in for (c_0, c_n, c_x, c_y),
   3932  * taking into account that each coefficient in c_j_x and c_k_x is represented
   3933  * as a pair of non-negative coefficients.
   3934  */
   3935 static isl_stat add_inter_constraints(struct isl_sched_graph *graph,
   3936 	struct isl_sched_node *src, struct isl_sched_node *dst,
   3937 	__isl_take isl_basic_set *coef, int pos)
   3938 {
   3939 	isl_size offset;
   3940 	isl_ctx *ctx;
   3941 	isl_dim_map *dim_map;
   3942 
   3943 	offset = coef_var_offset(coef);
   3944 	if (offset < 0)
   3945 		coef = isl_basic_set_free(coef);
   3946 	if (!coef)
   3947 		return isl_stat_error;
   3948 
   3949 	ctx = isl_basic_set_get_ctx(coef);
   3950 	dim_map = inter_dim_map(ctx, graph, src, dst, offset, 1);
   3951 	if (pos >= 0)
   3952 		isl_dim_map_range(dim_map, 3 + pos, 0, 0, 0, 1, -1);
   3953 	graph->lp = add_constraints_dim_map(graph->lp, coef, dim_map);
   3954 
   3955 	return isl_stat_ok;
   3956 }
   3957 
   3958 /* Data structure for keeping track of the data needed
   3959  * to exploit non-trivial lineality spaces.
   3960  *
   3961  * "any_non_trivial" is true if there are any non-trivial lineality spaces.
   3962  * If "any_non_trivial" is not true, then "equivalent" and "mask" may be NULL.
   3963  * "equivalent" connects instances to other instances on the same line(s).
   3964  * "mask" contains the domain spaces of "equivalent".
   3965  * Any instance set not in "mask" does not have a non-trivial lineality space.
   3966  */
   3967 struct isl_exploit_lineality_data {
   3968 	isl_bool any_non_trivial;
   3969 	isl_union_map *equivalent;
   3970 	isl_union_set *mask;
   3971 };
   3972 
   3973 /* Data structure collecting information used during the construction
   3974  * of an LP for carrying dependences.
   3975  *
   3976  * "intra" is a sequence of coefficient constraints for intra-node edges.
   3977  * "inter" is a sequence of coefficient constraints for inter-node edges.
   3978  * "lineality" contains data used to exploit non-trivial lineality spaces.
   3979  */
   3980 struct isl_carry {
   3981 	isl_basic_set_list *intra;
   3982 	isl_basic_set_list *inter;
   3983 	struct isl_exploit_lineality_data lineality;
   3984 };
   3985 
   3986 /* Free all the data stored in "carry".
   3987  */
   3988 static void isl_carry_clear(struct isl_carry *carry)
   3989 {
   3990 	isl_basic_set_list_free(carry->intra);
   3991 	isl_basic_set_list_free(carry->inter);
   3992 	isl_union_map_free(carry->lineality.equivalent);
   3993 	isl_union_set_free(carry->lineality.mask);
   3994 }
   3995 
   3996 /* Return a pointer to the node in "graph" that lives in "space".
   3997  * If the requested node has been compressed, then "space"
   3998  * corresponds to the compressed space.
   3999  * The graph is assumed to have such a node.
   4000  * Return NULL in case of error.
   4001  *
   4002  * First try and see if "space" is the space of an uncompressed node.
   4003  * If so, return that node.
   4004  * Otherwise, "space" was constructed by construct_compressed_id and
   4005  * contains a user pointer pointing to the node in the tuple id.
   4006  * However, this node belongs to the original dependence graph.
   4007  * If "graph" is a subgraph of this original dependence graph,
   4008  * then the node with the same space still needs to be looked up
   4009  * in the current graph.
   4010  */
   4011 static struct isl_sched_node *graph_find_compressed_node(isl_ctx *ctx,
   4012 	struct isl_sched_graph *graph, __isl_keep isl_space *space)
   4013 {
   4014 	isl_id *id;
   4015 	struct isl_sched_node *node;
   4016 
   4017 	if (!space)
   4018 		return NULL;
   4019 
   4020 	node = isl_sched_graph_find_node(ctx, graph, space);
   4021 	if (!node)
   4022 		return NULL;
   4023 	if (isl_sched_graph_is_node(graph, node))
   4024 		return node;
   4025 
   4026 	id = isl_space_get_tuple_id(space, isl_dim_set);
   4027 	node = isl_id_get_user(id);
   4028 	isl_id_free(id);
   4029 
   4030 	if (!node)
   4031 		return NULL;
   4032 
   4033 	if (!isl_sched_graph_is_node(graph->root, node))
   4034 		isl_die(ctx, isl_error_internal,
   4035 			"space points to invalid node", return NULL);
   4036 	if (graph != graph->root)
   4037 		node = isl_sched_graph_find_node(ctx, graph, node->space);
   4038 	if (!isl_sched_graph_is_node(graph, node))
   4039 		isl_die(ctx, isl_error_internal,
   4040 			"unable to find node", return NULL);
   4041 
   4042 	return node;
   4043 }
   4044 
   4045 /* Internal data structure for add_all_constraints.
   4046  *
   4047  * "graph" is the schedule constraint graph for which an LP problem
   4048  * is being constructed.
   4049  * "carry_inter" indicates whether inter-node edges should be carried.
   4050  * "pos" is the position of the next edge that needs to be carried.
   4051  */
   4052 struct isl_add_all_constraints_data {
   4053 	isl_ctx *ctx;
   4054 	struct isl_sched_graph *graph;
   4055 	int carry_inter;
   4056 	int pos;
   4057 };
   4058 
   4059 /* Add the constraints "coef" derived from an edge from a node to itself
   4060  * to data->graph->lp in order to respect the dependences and
   4061  * to try and carry them.
   4062  *
   4063  * The space of "coef" is of the form
   4064  *
   4065  *	coefficients[[c_cst] -> S[c_x]]
   4066  *
   4067  * with S[c_x] the (compressed) space of the node.
   4068  * Extract the node from the space and call add_intra_constraints.
   4069  */
   4070 static isl_stat lp_add_intra(__isl_take isl_basic_set *coef, void *user)
   4071 {
   4072 	struct isl_add_all_constraints_data *data = user;
   4073 	isl_space *space;
   4074 	struct isl_sched_node *node;
   4075 
   4076 	space = isl_basic_set_get_space(coef);
   4077 	space = isl_space_range(isl_space_unwrap(space));
   4078 	node = graph_find_compressed_node(data->ctx, data->graph, space);
   4079 	isl_space_free(space);
   4080 	return add_intra_constraints(data->graph, node, coef, data->pos++);
   4081 }
   4082 
   4083 /* Add the constraints "coef" derived from an edge from a node j
   4084  * to a node k to data->graph->lp in order to respect the dependences and
   4085  * to try and carry them (provided data->carry_inter is set).
   4086  *
   4087  * The space of "coef" is of the form
   4088  *
   4089  *	coefficients[[c_cst, c_n] -> [S_j[c_x] -> S_k[c_y]]]
   4090  *
   4091  * with S_j[c_x] and S_k[c_y] the (compressed) spaces of the nodes.
   4092  * Extract the nodes from the space and call add_inter_constraints.
   4093  */
   4094 static isl_stat lp_add_inter(__isl_take isl_basic_set *coef, void *user)
   4095 {
   4096 	struct isl_add_all_constraints_data *data = user;
   4097 	isl_space *space, *dom;
   4098 	struct isl_sched_node *src, *dst;
   4099 	int pos;
   4100 
   4101 	space = isl_basic_set_get_space(coef);
   4102 	space = isl_space_unwrap(isl_space_range(isl_space_unwrap(space)));
   4103 	dom = isl_space_domain(isl_space_copy(space));
   4104 	src = graph_find_compressed_node(data->ctx, data->graph, dom);
   4105 	isl_space_free(dom);
   4106 	space = isl_space_range(space);
   4107 	dst = graph_find_compressed_node(data->ctx, data->graph, space);
   4108 	isl_space_free(space);
   4109 
   4110 	pos = data->carry_inter ? data->pos++ : -1;
   4111 	return add_inter_constraints(data->graph, src, dst, coef, pos);
   4112 }
   4113 
   4114 /* Add constraints to graph->lp that force all (conditional) validity
   4115  * dependences to be respected and attempt to carry them.
   4116  * "intra" is the sequence of coefficient constraints for intra-node edges.
   4117  * "inter" is the sequence of coefficient constraints for inter-node edges.
   4118  * "carry_inter" indicates whether inter-node edges should be carried or
   4119  * only respected.
   4120  */
   4121 static isl_stat add_all_constraints(isl_ctx *ctx, struct isl_sched_graph *graph,
   4122 	__isl_keep isl_basic_set_list *intra,
   4123 	__isl_keep isl_basic_set_list *inter, int carry_inter)
   4124 {
   4125 	struct isl_add_all_constraints_data data = { ctx, graph, carry_inter };
   4126 
   4127 	data.pos = 0;
   4128 	if (isl_basic_set_list_foreach(intra, &lp_add_intra, &data) < 0)
   4129 		return isl_stat_error;
   4130 	if (isl_basic_set_list_foreach(inter, &lp_add_inter, &data) < 0)
   4131 		return isl_stat_error;
   4132 	return isl_stat_ok;
   4133 }
   4134 
   4135 /* Internal data structure for count_all_constraints
   4136  * for keeping track of the number of equality and inequality constraints.
   4137  */
   4138 struct isl_sched_count {
   4139 	int n_eq;
   4140 	int n_ineq;
   4141 };
   4142 
   4143 /* Add the number of equality and inequality constraints of "bset"
   4144  * to data->n_eq and data->n_ineq.
   4145  */
   4146 static isl_stat bset_update_count(__isl_take isl_basic_set *bset, void *user)
   4147 {
   4148 	struct isl_sched_count *data = user;
   4149 
   4150 	return update_count(bset, 1, &data->n_eq, &data->n_ineq);
   4151 }
   4152 
   4153 /* Count the number of equality and inequality constraints
   4154  * that will be added to the carry_lp problem.
   4155  * We count each edge exactly once.
   4156  * "intra" is the sequence of coefficient constraints for intra-node edges.
   4157  * "inter" is the sequence of coefficient constraints for inter-node edges.
   4158  */
   4159 static isl_stat count_all_constraints(__isl_keep isl_basic_set_list *intra,
   4160 	__isl_keep isl_basic_set_list *inter, int *n_eq, int *n_ineq)
   4161 {
   4162 	struct isl_sched_count data;
   4163 
   4164 	data.n_eq = data.n_ineq = 0;
   4165 	if (isl_basic_set_list_foreach(inter, &bset_update_count, &data) < 0)
   4166 		return isl_stat_error;
   4167 	if (isl_basic_set_list_foreach(intra, &bset_update_count, &data) < 0)
   4168 		return isl_stat_error;
   4169 
   4170 	*n_eq = data.n_eq;
   4171 	*n_ineq = data.n_ineq;
   4172 
   4173 	return isl_stat_ok;
   4174 }
   4175 
   4176 /* Construct an LP problem for finding schedule coefficients
   4177  * such that the schedule carries as many validity dependences as possible.
   4178  * In particular, for each dependence i, we bound the dependence distance
   4179  * from below by e_i, with 0 <= e_i <= 1 and then maximize the sum
   4180  * of all e_i's.  Dependences with e_i = 0 in the solution are simply
   4181  * respected, while those with e_i > 0 (in practice e_i = 1) are carried.
   4182  * "intra" is the sequence of coefficient constraints for intra-node edges.
   4183  * "inter" is the sequence of coefficient constraints for inter-node edges.
   4184  * "n_edge" is the total number of edges.
   4185  * "carry_inter" indicates whether inter-node edges should be carried or
   4186  * only respected.  That is, if "carry_inter" is not set, then
   4187  * no e_i variables are introduced for the inter-node edges.
   4188  *
   4189  * All variables of the LP are non-negative.  The actual coefficients
   4190  * may be negative, so each coefficient is represented as the difference
   4191  * of two non-negative variables.  The negative part always appears
   4192  * immediately before the positive part.
   4193  * Other than that, the variables have the following order
   4194  *
   4195  *	- sum of (1 - e_i) over all edges
   4196  *	- sum of all c_n coefficients
   4197  *		(unconstrained when computing non-parametric schedules)
   4198  *	- sum of positive and negative parts of all c_x coefficients
   4199  *	- for each edge
   4200  *		- e_i
   4201  *	- for each node
   4202  *		- positive and negative parts of c_i_x, in opposite order
   4203  *		- c_i_n (if parametric)
   4204  *		- c_i_0
   4205  *
   4206  * The constraints are those from the (validity) edges plus three equalities
   4207  * to express the sums and n_edge inequalities to express e_i <= 1.
   4208  */
   4209 static isl_stat setup_carry_lp(isl_ctx *ctx, struct isl_sched_graph *graph,
   4210 	int n_edge, __isl_keep isl_basic_set_list *intra,
   4211 	__isl_keep isl_basic_set_list *inter, int carry_inter)
   4212 {
   4213 	int i;
   4214 	int k;
   4215 	isl_space *space;
   4216 	unsigned total;
   4217 	int n_eq, n_ineq;
   4218 
   4219 	total = 3 + n_edge;
   4220 	for (i = 0; i < graph->n; ++i) {
   4221 		struct isl_sched_node *node = &graph->node[graph->sorted[i]];
   4222 		node->start = total;
   4223 		total += 1 + node->nparam + 2 * node->nvar;
   4224 	}
   4225 
   4226 	if (count_all_constraints(intra, inter, &n_eq, &n_ineq) < 0)
   4227 		return isl_stat_error;
   4228 
   4229 	space = isl_space_set_alloc(ctx, 0, total);
   4230 	isl_basic_set_free(graph->lp);
   4231 	n_eq += 3;
   4232 	n_ineq += n_edge;
   4233 	graph->lp = isl_basic_set_alloc_space(space, 0, n_eq, n_ineq);
   4234 	graph->lp = isl_basic_set_set_rational(graph->lp);
   4235 
   4236 	k = isl_basic_set_alloc_equality(graph->lp);
   4237 	if (k < 0)
   4238 		return isl_stat_error;
   4239 	isl_seq_clr(graph->lp->eq[k], 1 + total);
   4240 	isl_int_set_si(graph->lp->eq[k][0], -n_edge);
   4241 	isl_int_set_si(graph->lp->eq[k][1], 1);
   4242 	for (i = 0; i < n_edge; ++i)
   4243 		isl_int_set_si(graph->lp->eq[k][4 + i], 1);
   4244 
   4245 	if (add_param_sum_constraint(graph, 1) < 0)
   4246 		return isl_stat_error;
   4247 	if (add_var_sum_constraint(graph, 2) < 0)
   4248 		return isl_stat_error;
   4249 
   4250 	for (i = 0; i < n_edge; ++i) {
   4251 		k = isl_basic_set_alloc_inequality(graph->lp);
   4252 		if (k < 0)
   4253 			return isl_stat_error;
   4254 		isl_seq_clr(graph->lp->ineq[k], 1 + total);
   4255 		isl_int_set_si(graph->lp->ineq[k][4 + i], -1);
   4256 		isl_int_set_si(graph->lp->ineq[k][0], 1);
   4257 	}
   4258 
   4259 	if (add_all_constraints(ctx, graph, intra, inter, carry_inter) < 0)
   4260 		return isl_stat_error;
   4261 
   4262 	return isl_stat_ok;
   4263 }
   4264 
   4265 static __isl_give isl_schedule_node *compute_component_schedule(
   4266 	__isl_take isl_schedule_node *node, struct isl_sched_graph *graph,
   4267 	int wcc);
   4268 
   4269 /* If the schedule_split_scaled option is set and if the linear
   4270  * parts of the scheduling rows for all nodes in the graphs have
   4271  * a non-trivial common divisor, then remove this
   4272  * common divisor from the linear part.
   4273  * Otherwise, insert a band node directly and continue with
   4274  * the construction of the schedule.
   4275  *
   4276  * If a non-trivial common divisor is found, then
   4277  * the linear part is reduced and the remainder is ignored.
   4278  * The pieces of the graph that are assigned different remainders
   4279  * form (groups of) strongly connected components within
   4280  * the scaled down band.  If needed, they can therefore
   4281  * be ordered along this remainder in a sequence node.
   4282  * However, this ordering is not enforced here in order to allow
   4283  * the scheduler to combine some of the strongly connected components.
   4284  */
   4285 static __isl_give isl_schedule_node *split_scaled(
   4286 	__isl_take isl_schedule_node *node, struct isl_sched_graph *graph)
   4287 {
   4288 	int i;
   4289 	int row;
   4290 	isl_ctx *ctx;
   4291 	isl_int gcd, gcd_i;
   4292 	isl_size n_row;
   4293 
   4294 	if (!node)
   4295 		return NULL;
   4296 
   4297 	ctx = isl_schedule_node_get_ctx(node);
   4298 	if (!ctx->opt->schedule_split_scaled)
   4299 		return compute_next_band(node, graph, 0);
   4300 	if (graph->n <= 1)
   4301 		return compute_next_band(node, graph, 0);
   4302 	n_row = isl_mat_rows(graph->node[0].sched);
   4303 	if (n_row < 0)
   4304 		return isl_schedule_node_free(node);
   4305 
   4306 	isl_int_init(gcd);
   4307 	isl_int_init(gcd_i);
   4308 
   4309 	isl_int_set_si(gcd, 0);
   4310 
   4311 	row = n_row - 1;
   4312 
   4313 	for (i = 0; i < graph->n; ++i) {
   4314 		struct isl_sched_node *node = &graph->node[i];
   4315 		isl_size cols = isl_mat_cols(node->sched);
   4316 
   4317 		if (cols < 0)
   4318 			break;
   4319 		isl_seq_gcd(node->sched->row[row] + 1, cols - 1, &gcd_i);
   4320 		isl_int_gcd(gcd, gcd, gcd_i);
   4321 	}
   4322 
   4323 	isl_int_clear(gcd_i);
   4324 	if (i < graph->n)
   4325 		goto error;
   4326 
   4327 	if (isl_int_cmp_si(gcd, 1) <= 0) {
   4328 		isl_int_clear(gcd);
   4329 		return compute_next_band(node, graph, 0);
   4330 	}
   4331 
   4332 	for (i = 0; i < graph->n; ++i) {
   4333 		struct isl_sched_node *node = &graph->node[i];
   4334 
   4335 		isl_int_fdiv_q(node->sched->row[row][0],
   4336 			       node->sched->row[row][0], gcd);
   4337 		isl_int_mul(node->sched->row[row][0],
   4338 			    node->sched->row[row][0], gcd);
   4339 		node->sched = isl_mat_scale_down_row(node->sched, row, gcd);
   4340 		if (!node->sched)
   4341 			goto error;
   4342 	}
   4343 
   4344 	isl_int_clear(gcd);
   4345 
   4346 	return compute_next_band(node, graph, 0);
   4347 error:
   4348 	isl_int_clear(gcd);
   4349 	return isl_schedule_node_free(node);
   4350 }
   4351 
   4352 /* Is the schedule row "sol" trivial on node "node"?
   4353  * That is, is the solution zero on the dimensions linearly independent of
   4354  * the previously found solutions?
   4355  * Return 1 if the solution is trivial, 0 if it is not and -1 on error.
   4356  *
   4357  * Each coefficient is represented as the difference between
   4358  * two non-negative values in "sol".
   4359  * We construct the schedule row s and check if it is linearly
   4360  * independent of previously computed schedule rows
   4361  * by computing T s, with T the linear combinations that are zero
   4362  * on linearly dependent schedule rows.
   4363  * If the result consists of all zeros, then the solution is trivial.
   4364  */
   4365 static int is_trivial(struct isl_sched_node *node, __isl_keep isl_vec *sol)
   4366 {
   4367 	int trivial;
   4368 	isl_vec *node_sol;
   4369 
   4370 	if (!sol)
   4371 		return -1;
   4372 	if (node->nvar == node->rank)
   4373 		return 0;
   4374 
   4375 	node_sol = extract_var_coef(node, sol);
   4376 	node_sol = isl_mat_vec_product(isl_mat_copy(node->indep), node_sol);
   4377 	if (!node_sol)
   4378 		return -1;
   4379 
   4380 	trivial = isl_seq_first_non_zero(node_sol->el,
   4381 					node->nvar - node->rank) == -1;
   4382 
   4383 	isl_vec_free(node_sol);
   4384 
   4385 	return trivial;
   4386 }
   4387 
   4388 /* Is the schedule row "sol" trivial on any node where it should
   4389  * not be trivial?
   4390  * Return 1 if any solution is trivial, 0 if they are not and -1 on error.
   4391  */
   4392 static int is_any_trivial(struct isl_sched_graph *graph,
   4393 	__isl_keep isl_vec *sol)
   4394 {
   4395 	int i;
   4396 
   4397 	for (i = 0; i < graph->n; ++i) {
   4398 		struct isl_sched_node *node = &graph->node[i];
   4399 		int trivial;
   4400 
   4401 		if (!needs_row(graph, node))
   4402 			continue;
   4403 		trivial = is_trivial(node, sol);
   4404 		if (trivial < 0 || trivial)
   4405 			return trivial;
   4406 	}
   4407 
   4408 	return 0;
   4409 }
   4410 
   4411 /* Does the schedule represented by "sol" perform loop coalescing on "node"?
   4412  * If so, return the position of the coalesced dimension.
   4413  * Otherwise, return node->nvar or -1 on error.
   4414  *
   4415  * In particular, look for pairs of coefficients c_i and c_j such that
   4416  * |c_j/c_i| > ceil(size_i/2), i.e., |c_j| > |c_i * ceil(size_i/2)|.
   4417  * If any such pair is found, then return i.
   4418  * If size_i is infinity, then no check on c_i needs to be performed.
   4419  */
   4420 static int find_node_coalescing(struct isl_sched_node *node,
   4421 	__isl_keep isl_vec *sol)
   4422 {
   4423 	int i, j;
   4424 	isl_int max;
   4425 	isl_vec *csol;
   4426 
   4427 	if (node->nvar <= 1)
   4428 		return node->nvar;
   4429 
   4430 	csol = extract_var_coef(node, sol);
   4431 	if (!csol)
   4432 		return -1;
   4433 	isl_int_init(max);
   4434 	for (i = 0; i < node->nvar; ++i) {
   4435 		isl_val *v;
   4436 
   4437 		if (isl_int_is_zero(csol->el[i]))
   4438 			continue;
   4439 		v = isl_multi_val_get_val(node->sizes, i);
   4440 		if (!v)
   4441 			goto error;
   4442 		if (!isl_val_is_int(v)) {
   4443 			isl_val_free(v);
   4444 			continue;
   4445 		}
   4446 		v = isl_val_div_ui(v, 2);
   4447 		v = isl_val_ceil(v);
   4448 		if (!v)
   4449 			goto error;
   4450 		isl_int_mul(max, v->n, csol->el[i]);
   4451 		isl_val_free(v);
   4452 
   4453 		for (j = 0; j < node->nvar; ++j) {
   4454 			if (j == i)
   4455 				continue;
   4456 			if (isl_int_abs_gt(csol->el[j], max))
   4457 				break;
   4458 		}
   4459 		if (j < node->nvar)
   4460 			break;
   4461 	}
   4462 
   4463 	isl_int_clear(max);
   4464 	isl_vec_free(csol);
   4465 	return i;
   4466 error:
   4467 	isl_int_clear(max);
   4468 	isl_vec_free(csol);
   4469 	return -1;
   4470 }
   4471 
   4472 /* Force the schedule coefficient at position "pos" of "node" to be zero
   4473  * in "tl".
   4474  * The coefficient is encoded as the difference between two non-negative
   4475  * variables.  Force these two variables to have the same value.
   4476  */
   4477 static __isl_give isl_tab_lexmin *zero_out_node_coef(
   4478 	__isl_take isl_tab_lexmin *tl, struct isl_sched_node *node, int pos)
   4479 {
   4480 	int dim;
   4481 	isl_ctx *ctx;
   4482 	isl_vec *eq;
   4483 
   4484 	ctx = isl_space_get_ctx(node->space);
   4485 	dim = isl_tab_lexmin_dim(tl);
   4486 	if (dim < 0)
   4487 		return isl_tab_lexmin_free(tl);
   4488 	eq = isl_vec_alloc(ctx, 1 + dim);
   4489 	eq = isl_vec_clr(eq);
   4490 	if (!eq)
   4491 		return isl_tab_lexmin_free(tl);
   4492 
   4493 	pos = 1 + node_var_coef_pos(node, pos);
   4494 	isl_int_set_si(eq->el[pos], 1);
   4495 	isl_int_set_si(eq->el[pos + 1], -1);
   4496 	tl = isl_tab_lexmin_add_eq(tl, eq->el);
   4497 	isl_vec_free(eq);
   4498 
   4499 	return tl;
   4500 }
   4501 
   4502 /* Return the lexicographically smallest rational point in the basic set
   4503  * from which "tl" was constructed, double checking that this input set
   4504  * was not empty.
   4505  */
   4506 static __isl_give isl_vec *non_empty_solution(__isl_keep isl_tab_lexmin *tl)
   4507 {
   4508 	isl_vec *sol;
   4509 
   4510 	sol = isl_tab_lexmin_get_solution(tl);
   4511 	if (!sol)
   4512 		return NULL;
   4513 	if (sol->size == 0)
   4514 		isl_die(isl_vec_get_ctx(sol), isl_error_internal,
   4515 			"error in schedule construction",
   4516 			return isl_vec_free(sol));
   4517 	return sol;
   4518 }
   4519 
   4520 /* Does the solution "sol" of the LP problem constructed by setup_carry_lp
   4521  * carry any of the "n_edge" groups of dependences?
   4522  * The value in the first position is the sum of (1 - e_i) over all "n_edge"
   4523  * edges, with 0 <= e_i <= 1 equal to 1 when the dependences represented
   4524  * by the edge are carried by the solution.
   4525  * If the sum of the (1 - e_i) is smaller than "n_edge" then at least
   4526  * one of those is carried.
   4527  *
   4528  * Note that despite the fact that the problem is solved using a rational
   4529  * solver, the solution is guaranteed to be integral.
   4530  * Specifically, the dependence distance lower bounds e_i (and therefore
   4531  * also their sum) are integers.  See Lemma 5 of [1].
   4532  *
   4533  * Any potential denominator of the sum is cleared by this function.
   4534  * The denominator is not relevant for any of the other elements
   4535  * in the solution.
   4536  *
   4537  * [1] P. Feautrier, Some Efficient Solutions to the Affine Scheduling
   4538  *     Problem, Part II: Multi-Dimensional Time.
   4539  *     In Intl. Journal of Parallel Programming, 1992.
   4540  */
   4541 static int carries_dependences(__isl_keep isl_vec *sol, int n_edge)
   4542 {
   4543 	isl_int_divexact(sol->el[1], sol->el[1], sol->el[0]);
   4544 	isl_int_set_si(sol->el[0], 1);
   4545 	return isl_int_cmp_si(sol->el[1], n_edge) < 0;
   4546 }
   4547 
   4548 /* Return the lexicographically smallest rational point in "lp",
   4549  * assuming that all variables are non-negative and performing some
   4550  * additional sanity checks.
   4551  * If "want_integral" is set, then compute the lexicographically smallest
   4552  * integer point instead.
   4553  * In particular, "lp" should not be empty by construction.
   4554  * Double check that this is the case.
   4555  * If dependences are not carried for any of the "n_edge" edges,
   4556  * then return an empty vector.
   4557  *
   4558  * If the schedule_treat_coalescing option is set and
   4559  * if the computed schedule performs loop coalescing on a given node,
   4560  * i.e., if it is of the form
   4561  *
   4562  *	c_i i + c_j j + ...
   4563  *
   4564  * with |c_j/c_i| >= size_i, then force the coefficient c_i to be zero
   4565  * to cut out this solution.  Repeat this process until no more loop
   4566  * coalescing occurs or until no more dependences can be carried.
   4567  * In the latter case, revert to the previously computed solution.
   4568  *
   4569  * If the caller requests an integral solution and if coalescing should
   4570  * be treated, then perform the coalescing treatment first as
   4571  * an integral solution computed before coalescing treatment
   4572  * would carry the same number of edges and would therefore probably
   4573  * also be coalescing.
   4574  *
   4575  * To allow the coalescing treatment to be performed first,
   4576  * the initial solution is allowed to be rational and it is only
   4577  * cut out (if needed) in the next iteration, if no coalescing measures
   4578  * were taken.
   4579  */
   4580 static __isl_give isl_vec *non_neg_lexmin(struct isl_sched_graph *graph,
   4581 	__isl_take isl_basic_set *lp, int n_edge, int want_integral)
   4582 {
   4583 	int i, pos, cut;
   4584 	isl_ctx *ctx;
   4585 	isl_tab_lexmin *tl;
   4586 	isl_vec *sol = NULL, *prev;
   4587 	int treat_coalescing;
   4588 	int try_again;
   4589 
   4590 	if (!lp)
   4591 		return NULL;
   4592 	ctx = isl_basic_set_get_ctx(lp);
   4593 	treat_coalescing = isl_options_get_schedule_treat_coalescing(ctx);
   4594 	tl = isl_tab_lexmin_from_basic_set(lp);
   4595 
   4596 	cut = 0;
   4597 	do {
   4598 		int integral;
   4599 
   4600 		try_again = 0;
   4601 		if (cut)
   4602 			tl = isl_tab_lexmin_cut_to_integer(tl);
   4603 		prev = sol;
   4604 		sol = non_empty_solution(tl);
   4605 		if (!sol)
   4606 			goto error;
   4607 
   4608 		integral = isl_int_is_one(sol->el[0]);
   4609 		if (!carries_dependences(sol, n_edge)) {
   4610 			if (!prev)
   4611 				prev = isl_vec_alloc(ctx, 0);
   4612 			isl_vec_free(sol);
   4613 			sol = prev;
   4614 			break;
   4615 		}
   4616 		prev = isl_vec_free(prev);
   4617 		cut = want_integral && !integral;
   4618 		if (cut)
   4619 			try_again = 1;
   4620 		if (!treat_coalescing)
   4621 			continue;
   4622 		for (i = 0; i < graph->n; ++i) {
   4623 			struct isl_sched_node *node = &graph->node[i];
   4624 
   4625 			pos = find_node_coalescing(node, sol);
   4626 			if (pos < 0)
   4627 				goto error;
   4628 			if (pos < node->nvar)
   4629 				break;
   4630 		}
   4631 		if (i < graph->n) {
   4632 			try_again = 1;
   4633 			tl = zero_out_node_coef(tl, &graph->node[i], pos);
   4634 			cut = 0;
   4635 		}
   4636 	} while (try_again);
   4637 
   4638 	isl_tab_lexmin_free(tl);
   4639 
   4640 	return sol;
   4641 error:
   4642 	isl_tab_lexmin_free(tl);
   4643 	isl_vec_free(prev);
   4644 	isl_vec_free(sol);
   4645 	return NULL;
   4646 }
   4647 
   4648 /* If "edge" is an edge from a node to itself, then add the corresponding
   4649  * dependence relation to "umap".
   4650  * If "node" has been compressed, then the dependence relation
   4651  * is also compressed first.
   4652  */
   4653 static __isl_give isl_union_map *add_intra(__isl_take isl_union_map *umap,
   4654 	struct isl_sched_edge *edge)
   4655 {
   4656 	isl_map *map;
   4657 	struct isl_sched_node *node = edge->src;
   4658 
   4659 	if (edge->src != edge->dst)
   4660 		return umap;
   4661 
   4662 	map = isl_map_copy(edge->map);
   4663 	map = compress(map, node, node);
   4664 	umap = isl_union_map_add_map(umap, map);
   4665 	return umap;
   4666 }
   4667 
   4668 /* If "edge" is an edge from a node to another node, then add the corresponding
   4669  * dependence relation to "umap".
   4670  * If the source or destination nodes of "edge" have been compressed,
   4671  * then the dependence relation is also compressed first.
   4672  */
   4673 static __isl_give isl_union_map *add_inter(__isl_take isl_union_map *umap,
   4674 	struct isl_sched_edge *edge)
   4675 {
   4676 	isl_map *map;
   4677 
   4678 	if (edge->src == edge->dst)
   4679 		return umap;
   4680 
   4681 	map = isl_map_copy(edge->map);
   4682 	map = compress(map, edge->src, edge->dst);
   4683 	umap = isl_union_map_add_map(umap, map);
   4684 	return umap;
   4685 }
   4686 
   4687 /* Internal data structure used by union_drop_coalescing_constraints
   4688  * to collect bounds on all relevant statements.
   4689  *
   4690  * "graph" is the schedule constraint graph for which an LP problem
   4691  * is being constructed.
   4692  * "bounds" collects the bounds.
   4693  */
   4694 struct isl_collect_bounds_data {
   4695 	isl_ctx *ctx;
   4696 	struct isl_sched_graph *graph;
   4697 	isl_union_set *bounds;
   4698 };
   4699 
   4700 /* Add the size bounds for the node with instance deltas in "set"
   4701  * to data->bounds.
   4702  */
   4703 static isl_stat collect_bounds(__isl_take isl_set *set, void *user)
   4704 {
   4705 	struct isl_collect_bounds_data *data = user;
   4706 	struct isl_sched_node *node;
   4707 	isl_space *space;
   4708 	isl_set *bounds;
   4709 
   4710 	space = isl_set_get_space(set);
   4711 	isl_set_free(set);
   4712 
   4713 	node = graph_find_compressed_node(data->ctx, data->graph, space);
   4714 	isl_space_free(space);
   4715 
   4716 	bounds = isl_set_from_basic_set(get_size_bounds(node));
   4717 	data->bounds = isl_union_set_add_set(data->bounds, bounds);
   4718 
   4719 	return isl_stat_ok;
   4720 }
   4721 
   4722 /* Drop some constraints from "delta" that could be exploited
   4723  * to construct loop coalescing schedules.
   4724  * In particular, drop those constraint that bound the difference
   4725  * to the size of the domain.
   4726  * Do this for each set/node in "delta" separately.
   4727  * The parameters are assumed to have been projected out by the caller.
   4728  */
   4729 static __isl_give isl_union_set *union_drop_coalescing_constraints(isl_ctx *ctx,
   4730 	struct isl_sched_graph *graph, __isl_take isl_union_set *delta)
   4731 {
   4732 	struct isl_collect_bounds_data data = { ctx, graph };
   4733 
   4734 	data.bounds = isl_union_set_empty(isl_space_params_alloc(ctx, 0));
   4735 	if (isl_union_set_foreach_set(delta, &collect_bounds, &data) < 0)
   4736 		data.bounds = isl_union_set_free(data.bounds);
   4737 	delta = isl_union_set_plain_gist(delta, data.bounds);
   4738 
   4739 	return delta;
   4740 }
   4741 
   4742 /* Given a non-trivial lineality space "lineality", add the corresponding
   4743  * universe set to data->mask and add a map from elements to
   4744  * other elements along the lines in "lineality" to data->equivalent.
   4745  * If this is the first time this function gets called
   4746  * (data->any_non_trivial is still false), then set data->any_non_trivial and
   4747  * initialize data->mask and data->equivalent.
   4748  *
   4749  * In particular, if the lineality space is defined by equality constraints
   4750  *
   4751  *	E x = 0
   4752  *
   4753  * then construct an affine mapping
   4754  *
   4755  *	f : x -> E x
   4756  *
   4757  * and compute the equivalence relation of having the same image under f:
   4758  *
   4759  *	{ x -> x' : E x = E x' }
   4760  */
   4761 static isl_stat add_non_trivial_lineality(__isl_take isl_basic_set *lineality,
   4762 	struct isl_exploit_lineality_data *data)
   4763 {
   4764 	isl_mat *eq;
   4765 	isl_space *space;
   4766 	isl_set *univ;
   4767 	isl_multi_aff *ma;
   4768 	isl_multi_pw_aff *mpa;
   4769 	isl_map *map;
   4770 	isl_size n;
   4771 
   4772 	if (isl_basic_set_check_no_locals(lineality) < 0)
   4773 		goto error;
   4774 
   4775 	space = isl_basic_set_get_space(lineality);
   4776 	if (!data->any_non_trivial) {
   4777 		data->equivalent = isl_union_map_empty(isl_space_copy(space));
   4778 		data->mask = isl_union_set_empty(isl_space_copy(space));
   4779 	}
   4780 	data->any_non_trivial = isl_bool_true;
   4781 
   4782 	univ = isl_set_universe(isl_space_copy(space));
   4783 	data->mask = isl_union_set_add_set(data->mask, univ);
   4784 
   4785 	eq = isl_basic_set_extract_equalities(lineality);
   4786 	n = isl_mat_rows(eq);
   4787 	if (n < 0)
   4788 		space = isl_space_free(space);
   4789 	eq = isl_mat_insert_zero_rows(eq, 0, 1);
   4790 	eq = isl_mat_set_element_si(eq, 0, 0, 1);
   4791 	space = isl_space_from_domain(space);
   4792 	space = isl_space_add_dims(space, isl_dim_out, n);
   4793 	ma = isl_multi_aff_from_aff_mat(space, eq);
   4794 	mpa = isl_multi_pw_aff_from_multi_aff(ma);
   4795 	map = isl_multi_pw_aff_eq_map(mpa, isl_multi_pw_aff_copy(mpa));
   4796 	data->equivalent = isl_union_map_add_map(data->equivalent, map);
   4797 
   4798 	isl_basic_set_free(lineality);
   4799 	return isl_stat_ok;
   4800 error:
   4801 	isl_basic_set_free(lineality);
   4802 	return isl_stat_error;
   4803 }
   4804 
   4805 /* Check if the lineality space "set" is non-trivial (i.e., is not just
   4806  * the origin or, in other words, satisfies a number of equality constraints
   4807  * that is smaller than the dimension of the set).
   4808  * If so, extend data->mask and data->equivalent accordingly.
   4809  *
   4810  * The input should not have any local variables already, but
   4811  * isl_set_remove_divs is called to make sure it does not.
   4812  */
   4813 static isl_stat add_lineality(__isl_take isl_set *set, void *user)
   4814 {
   4815 	struct isl_exploit_lineality_data *data = user;
   4816 	isl_basic_set *hull;
   4817 	isl_size dim;
   4818 	isl_size n_eq;
   4819 
   4820 	set = isl_set_remove_divs(set);
   4821 	hull = isl_set_unshifted_simple_hull(set);
   4822 	dim = isl_basic_set_dim(hull, isl_dim_set);
   4823 	n_eq = isl_basic_set_n_equality(hull);
   4824 	if (dim < 0 || n_eq < 0)
   4825 		goto error;
   4826 	if (dim != n_eq)
   4827 		return add_non_trivial_lineality(hull, data);
   4828 	isl_basic_set_free(hull);
   4829 	return isl_stat_ok;
   4830 error:
   4831 	isl_basic_set_free(hull);
   4832 	return isl_stat_error;
   4833 }
   4834 
   4835 /* Check if the difference set on intra-node schedule constraints "intra"
   4836  * has any non-trivial lineality space.
   4837  * If so, then extend the difference set to a difference set
   4838  * on equivalent elements.  That is, if "intra" is
   4839  *
   4840  *	{ y - x : (x,y) \in V }
   4841  *
   4842  * and elements are equivalent if they have the same image under f,
   4843  * then return
   4844  *
   4845  *	{ y' - x' : (x,y) \in V and f(x) = f(x') and f(y) = f(y') }
   4846  *
   4847  * or, since f is linear,
   4848  *
   4849  *	{ y' - x' : (x,y) \in V and f(y - x) = f(y' - x') }
   4850  *
   4851  * The results of the search for non-trivial lineality spaces is stored
   4852  * in "data".
   4853  */
   4854 static __isl_give isl_union_set *exploit_intra_lineality(
   4855 	__isl_take isl_union_set *intra,
   4856 	struct isl_exploit_lineality_data *data)
   4857 {
   4858 	isl_union_set *lineality;
   4859 	isl_union_set *uset;
   4860 
   4861 	data->any_non_trivial = isl_bool_false;
   4862 	lineality = isl_union_set_copy(intra);
   4863 	lineality = isl_union_set_combined_lineality_space(lineality);
   4864 	if (isl_union_set_foreach_set(lineality, &add_lineality, data) < 0)
   4865 		data->any_non_trivial = isl_bool_error;
   4866 	isl_union_set_free(lineality);
   4867 
   4868 	if (data->any_non_trivial < 0)
   4869 		return isl_union_set_free(intra);
   4870 	if (!data->any_non_trivial)
   4871 		return intra;
   4872 
   4873 	uset = isl_union_set_copy(intra);
   4874 	intra = isl_union_set_subtract(intra, isl_union_set_copy(data->mask));
   4875 	uset = isl_union_set_apply(uset, isl_union_map_copy(data->equivalent));
   4876 	intra = isl_union_set_union(intra, uset);
   4877 
   4878 	intra = isl_union_set_remove_divs(intra);
   4879 
   4880 	return intra;
   4881 }
   4882 
   4883 /* If the difference set on intra-node schedule constraints was found to have
   4884  * any non-trivial lineality space by exploit_intra_lineality,
   4885  * as recorded in "data", then extend the inter-node
   4886  * schedule constraints "inter" to schedule constraints on equivalent elements.
   4887  * That is, if "inter" is V and
   4888  * elements are equivalent if they have the same image under f, then return
   4889  *
   4890  *	{ (x', y') : (x,y) \in V and f(x) = f(x') and f(y) = f(y') }
   4891  */
   4892 static __isl_give isl_union_map *exploit_inter_lineality(
   4893 	__isl_take isl_union_map *inter,
   4894 	struct isl_exploit_lineality_data *data)
   4895 {
   4896 	isl_union_map *umap;
   4897 
   4898 	if (data->any_non_trivial < 0)
   4899 		return isl_union_map_free(inter);
   4900 	if (!data->any_non_trivial)
   4901 		return inter;
   4902 
   4903 	umap = isl_union_map_copy(inter);
   4904 	inter = isl_union_map_subtract_range(inter,
   4905 				isl_union_set_copy(data->mask));
   4906 	umap = isl_union_map_apply_range(umap,
   4907 				isl_union_map_copy(data->equivalent));
   4908 	inter = isl_union_map_union(inter, umap);
   4909 	umap = isl_union_map_copy(inter);
   4910 	inter = isl_union_map_subtract_domain(inter,
   4911 				isl_union_set_copy(data->mask));
   4912 	umap = isl_union_map_apply_range(isl_union_map_copy(data->equivalent),
   4913 				umap);
   4914 	inter = isl_union_map_union(inter, umap);
   4915 
   4916 	inter = isl_union_map_remove_divs(inter);
   4917 
   4918 	return inter;
   4919 }
   4920 
   4921 /* For each (conditional) validity edge in "graph",
   4922  * add the corresponding dependence relation using "add"
   4923  * to a collection of dependence relations and return the result.
   4924  * If "coincidence" is set, then coincidence edges are considered as well.
   4925  */
   4926 static __isl_give isl_union_map *collect_validity(struct isl_sched_graph *graph,
   4927 	__isl_give isl_union_map *(*add)(__isl_take isl_union_map *umap,
   4928 		struct isl_sched_edge *edge), int coincidence)
   4929 {
   4930 	int i;
   4931 	isl_space *space;
   4932 	isl_union_map *umap;
   4933 
   4934 	space = isl_space_copy(graph->node[0].space);
   4935 	umap = isl_union_map_empty(space);
   4936 
   4937 	for (i = 0; i < graph->n_edge; ++i) {
   4938 		struct isl_sched_edge *edge = &graph->edge[i];
   4939 
   4940 		if (!is_any_validity(edge) &&
   4941 		    (!coincidence || !is_coincidence(edge)))
   4942 			continue;
   4943 
   4944 		umap = add(umap, edge);
   4945 	}
   4946 
   4947 	return umap;
   4948 }
   4949 
   4950 /* For each dependence relation on a (conditional) validity edge
   4951  * from a node to itself,
   4952  * construct the set of coefficients of valid constraints for elements
   4953  * in that dependence relation and collect the results.
   4954  * If "coincidence" is set, then coincidence edges are considered as well.
   4955  *
   4956  * In particular, for each dependence relation R, constraints
   4957  * on coefficients (c_0, c_x) are constructed such that
   4958  *
   4959  *	c_0 + c_x d >= 0 for each d in delta R = { y - x | (x,y) in R }
   4960  *
   4961  * If the schedule_treat_coalescing option is set, then some constraints
   4962  * that could be exploited to construct coalescing schedules
   4963  * are removed before the dual is computed, but after the parameters
   4964  * have been projected out.
   4965  * The entire computation is essentially the same as that performed
   4966  * by intra_coefficients, except that it operates on multiple
   4967  * edges together and that the parameters are always projected out.
   4968  *
   4969  * Additionally, exploit any non-trivial lineality space
   4970  * in the difference set after removing coalescing constraints and
   4971  * store the results of the non-trivial lineality space detection in "data".
   4972  * The procedure is currently run unconditionally, but it is unlikely
   4973  * to find any non-trivial lineality spaces if no coalescing constraints
   4974  * have been removed.
   4975  *
   4976  * Note that if a dependence relation is a union of basic maps,
   4977  * then each basic map needs to be treated individually as it may only
   4978  * be possible to carry the dependences expressed by some of those
   4979  * basic maps and not all of them.
   4980  * The collected validity constraints are therefore not coalesced and
   4981  * it is assumed that they are not coalesced automatically.
   4982  * Duplicate basic maps can be removed, however.
   4983  * In particular, if the same basic map appears as a disjunct
   4984  * in multiple edges, then it only needs to be carried once.
   4985  */
   4986 static __isl_give isl_basic_set_list *collect_intra_validity(isl_ctx *ctx,
   4987 	struct isl_sched_graph *graph, int coincidence,
   4988 	struct isl_exploit_lineality_data *data)
   4989 {
   4990 	isl_union_map *intra;
   4991 	isl_union_set *delta;
   4992 	isl_basic_set_list *list;
   4993 
   4994 	intra = collect_validity(graph, &add_intra, coincidence);
   4995 	delta = isl_union_map_deltas(intra);
   4996 	delta = isl_union_set_project_out_all_params(delta);
   4997 	delta = isl_union_set_remove_divs(delta);
   4998 	if (isl_options_get_schedule_treat_coalescing(ctx))
   4999 		delta = union_drop_coalescing_constraints(ctx, graph, delta);
   5000 	delta = exploit_intra_lineality(delta, data);
   5001 	list = isl_union_set_get_basic_set_list(delta);
   5002 	isl_union_set_free(delta);
   5003 
   5004 	return isl_basic_set_list_coefficients(list);
   5005 }
   5006 
   5007 /* For each dependence relation on a (conditional) validity edge
   5008  * from a node to some other node,
   5009  * construct the set of coefficients of valid constraints for elements
   5010  * in that dependence relation and collect the results.
   5011  * If "coincidence" is set, then coincidence edges are considered as well.
   5012  *
   5013  * In particular, for each dependence relation R, constraints
   5014  * on coefficients (c_0, c_n, c_x, c_y) are constructed such that
   5015  *
   5016  *	c_0 + c_n n + c_x x + c_y y >= 0 for each (x,y) in R
   5017  *
   5018  * This computation is essentially the same as that performed
   5019  * by inter_coefficients, except that it operates on multiple
   5020  * edges together.
   5021  *
   5022  * Additionally, exploit any non-trivial lineality space
   5023  * that may have been discovered by collect_intra_validity
   5024  * (as stored in "data").
   5025  *
   5026  * Note that if a dependence relation is a union of basic maps,
   5027  * then each basic map needs to be treated individually as it may only
   5028  * be possible to carry the dependences expressed by some of those
   5029  * basic maps and not all of them.
   5030  * The collected validity constraints are therefore not coalesced and
   5031  * it is assumed that they are not coalesced automatically.
   5032  * Duplicate basic maps can be removed, however.
   5033  * In particular, if the same basic map appears as a disjunct
   5034  * in multiple edges, then it only needs to be carried once.
   5035  */
   5036 static __isl_give isl_basic_set_list *collect_inter_validity(
   5037 	struct isl_sched_graph *graph, int coincidence,
   5038 	struct isl_exploit_lineality_data *data)
   5039 {
   5040 	isl_union_map *inter;
   5041 	isl_union_set *wrap;
   5042 	isl_basic_set_list *list;
   5043 
   5044 	inter = collect_validity(graph, &add_inter, coincidence);
   5045 	inter = exploit_inter_lineality(inter, data);
   5046 	inter = isl_union_map_remove_divs(inter);
   5047 	wrap = isl_union_map_wrap(inter);
   5048 	list = isl_union_set_get_basic_set_list(wrap);
   5049 	isl_union_set_free(wrap);
   5050 	return isl_basic_set_list_coefficients(list);
   5051 }
   5052 
   5053 /* Construct an LP problem for finding schedule coefficients
   5054  * such that the schedule carries as many of the "n_edge" groups of
   5055  * dependences as possible based on the corresponding coefficient
   5056  * constraints and return the lexicographically smallest non-trivial solution.
   5057  * "intra" is the sequence of coefficient constraints for intra-node edges.
   5058  * "inter" is the sequence of coefficient constraints for inter-node edges.
   5059  * If "want_integral" is set, then compute an integral solution
   5060  * for the coefficients rather than using the numerators
   5061  * of a rational solution.
   5062  * "carry_inter" indicates whether inter-node edges should be carried or
   5063  * only respected.
   5064  *
   5065  * If none of the "n_edge" groups can be carried
   5066  * then return an empty vector.
   5067  */
   5068 static __isl_give isl_vec *compute_carrying_sol_coef(isl_ctx *ctx,
   5069 	struct isl_sched_graph *graph, int n_edge,
   5070 	__isl_keep isl_basic_set_list *intra,
   5071 	__isl_keep isl_basic_set_list *inter, int want_integral,
   5072 	int carry_inter)
   5073 {
   5074 	isl_basic_set *lp;
   5075 
   5076 	if (setup_carry_lp(ctx, graph, n_edge, intra, inter, carry_inter) < 0)
   5077 		return NULL;
   5078 
   5079 	lp = isl_basic_set_copy(graph->lp);
   5080 	return non_neg_lexmin(graph, lp, n_edge, want_integral);
   5081 }
   5082 
   5083 /* Construct an LP problem for finding schedule coefficients
   5084  * such that the schedule carries as many of the validity dependences
   5085  * as possible and
   5086  * return the lexicographically smallest non-trivial solution.
   5087  * If "fallback" is set, then the carrying is performed as a fallback
   5088  * for the Pluto-like scheduler.
   5089  * If "coincidence" is set, then try and carry coincidence edges as well.
   5090  *
   5091  * The variable "n_edge" stores the number of groups that should be carried.
   5092  * If none of the "n_edge" groups can be carried
   5093  * then return an empty vector.
   5094  * If, moreover, "n_edge" is zero, then the LP problem does not even
   5095  * need to be constructed.
   5096  *
   5097  * If a fallback solution is being computed, then compute an integral solution
   5098  * for the coefficients rather than using the numerators
   5099  * of a rational solution.
   5100  *
   5101  * If a fallback solution is being computed, if there are any intra-node
   5102  * dependences, and if requested by the user, then first try
   5103  * to only carry those intra-node dependences.
   5104  * If this fails to carry any dependences, then try again
   5105  * with the inter-node dependences included.
   5106  */
   5107 static __isl_give isl_vec *compute_carrying_sol(isl_ctx *ctx,
   5108 	struct isl_sched_graph *graph, int fallback, int coincidence)
   5109 {
   5110 	isl_size n_intra, n_inter;
   5111 	int n_edge;
   5112 	struct isl_carry carry = { 0 };
   5113 	isl_vec *sol;
   5114 
   5115 	carry.intra = collect_intra_validity(ctx, graph, coincidence,
   5116 						&carry.lineality);
   5117 	carry.inter = collect_inter_validity(graph, coincidence,
   5118 						&carry.lineality);
   5119 	n_intra = isl_basic_set_list_n_basic_set(carry.intra);
   5120 	n_inter = isl_basic_set_list_n_basic_set(carry.inter);
   5121 	if (n_intra < 0 || n_inter < 0)
   5122 		goto error;
   5123 
   5124 	if (fallback && n_intra > 0 &&
   5125 	    isl_options_get_schedule_carry_self_first(ctx)) {
   5126 		sol = compute_carrying_sol_coef(ctx, graph, n_intra,
   5127 				carry.intra, carry.inter, fallback, 0);
   5128 		if (!sol || sol->size != 0 || n_inter == 0) {
   5129 			isl_carry_clear(&carry);
   5130 			return sol;
   5131 		}
   5132 		isl_vec_free(sol);
   5133 	}
   5134 
   5135 	n_edge = n_intra + n_inter;
   5136 	if (n_edge == 0) {
   5137 		isl_carry_clear(&carry);
   5138 		return isl_vec_alloc(ctx, 0);
   5139 	}
   5140 
   5141 	sol = compute_carrying_sol_coef(ctx, graph, n_edge,
   5142 				carry.intra, carry.inter, fallback, 1);
   5143 	isl_carry_clear(&carry);
   5144 	return sol;
   5145 error:
   5146 	isl_carry_clear(&carry);
   5147 	return NULL;
   5148 }
   5149 
   5150 /* Construct a schedule row for each node such that as many validity dependences
   5151  * as possible are carried and then continue with the next band.
   5152  * If "fallback" is set, then the carrying is performed as a fallback
   5153  * for the Pluto-like scheduler.
   5154  * If "coincidence" is set, then try and carry coincidence edges as well.
   5155  *
   5156  * If there are no validity dependences, then no dependence can be carried and
   5157  * the procedure is guaranteed to fail.  If there is more than one component,
   5158  * then try computing a schedule on each component separately
   5159  * to prevent or at least postpone this failure.
   5160  *
   5161  * If a schedule row is computed, then check that dependences are carried
   5162  * for at least one of the edges.
   5163  *
   5164  * If the computed schedule row turns out to be trivial on one or
   5165  * more nodes where it should not be trivial, then we throw it away
   5166  * and try again on each component separately.
   5167  *
   5168  * If there is only one component, then we accept the schedule row anyway,
   5169  * but we do not consider it as a complete row and therefore do not
   5170  * increment graph->n_row.  Note that the ranks of the nodes that
   5171  * do get a non-trivial schedule part will get updated regardless and
   5172  * graph->maxvar is computed based on these ranks.  The test for
   5173  * whether more schedule rows are required in compute_schedule_wcc
   5174  * is therefore not affected.
   5175  *
   5176  * Insert a band corresponding to the schedule row at position "node"
   5177  * of the schedule tree and continue with the construction of the schedule.
   5178  * This insertion and the continued construction is performed by split_scaled
   5179  * after optionally checking for non-trivial common divisors.
   5180  */
   5181 static __isl_give isl_schedule_node *carry(__isl_take isl_schedule_node *node,
   5182 	struct isl_sched_graph *graph, int fallback, int coincidence)
   5183 {
   5184 	int trivial;
   5185 	isl_ctx *ctx;
   5186 	isl_vec *sol;
   5187 
   5188 	if (!node)
   5189 		return NULL;
   5190 
   5191 	ctx = isl_schedule_node_get_ctx(node);
   5192 	sol = compute_carrying_sol(ctx, graph, fallback, coincidence);
   5193 	if (!sol)
   5194 		return isl_schedule_node_free(node);
   5195 	if (sol->size == 0) {
   5196 		isl_vec_free(sol);
   5197 		if (graph->scc > 1)
   5198 			return compute_component_schedule(node, graph, 1);
   5199 		isl_die(ctx, isl_error_unknown, "unable to carry dependences",
   5200 			return isl_schedule_node_free(node));
   5201 	}
   5202 
   5203 	trivial = is_any_trivial(graph, sol);
   5204 	if (trivial < 0) {
   5205 		sol = isl_vec_free(sol);
   5206 	} else if (trivial && graph->scc > 1) {
   5207 		isl_vec_free(sol);
   5208 		return compute_component_schedule(node, graph, 1);
   5209 	}
   5210 
   5211 	if (update_schedule(graph, sol, 0) < 0)
   5212 		return isl_schedule_node_free(node);
   5213 	if (trivial)
   5214 		graph->n_row--;
   5215 
   5216 	return split_scaled(node, graph);
   5217 }
   5218 
   5219 /* Construct a schedule row for each node such that as many validity dependences
   5220  * as possible are carried and then continue with the next band.
   5221  * Do so as a fallback for the Pluto-like scheduler.
   5222  * If "coincidence" is set, then try and carry coincidence edges as well.
   5223  */
   5224 static __isl_give isl_schedule_node *carry_fallback(
   5225 	__isl_take isl_schedule_node *node, struct isl_sched_graph *graph,
   5226 	int coincidence)
   5227 {
   5228 	return carry(node, graph, 1, coincidence);
   5229 }
   5230 
   5231 /* Construct a schedule row for each node such that as many validity dependences
   5232  * as possible are carried and then continue with the next band.
   5233  * Do so for the case where the Feautrier scheduler was selected
   5234  * by the user.
   5235  */
   5236 static __isl_give isl_schedule_node *carry_feautrier(
   5237 	__isl_take isl_schedule_node *node, struct isl_sched_graph *graph)
   5238 {
   5239 	return carry(node, graph, 0, 0);
   5240 }
   5241 
   5242 /* Construct a schedule row for each node such that as many validity dependences
   5243  * as possible are carried and then continue with the next band.
   5244  * Do so as a fallback for the Pluto-like scheduler.
   5245  */
   5246 static __isl_give isl_schedule_node *carry_dependences(
   5247 	__isl_take isl_schedule_node *node, struct isl_sched_graph *graph)
   5248 {
   5249 	return carry_fallback(node, graph, 0);
   5250 }
   5251 
   5252 /* Construct a schedule row for each node such that as many validity or
   5253  * coincidence dependences as possible are carried and
   5254  * then continue with the next band.
   5255  * Do so as a fallback for the Pluto-like scheduler.
   5256  */
   5257 static __isl_give isl_schedule_node *carry_coincidence(
   5258 	__isl_take isl_schedule_node *node, struct isl_sched_graph *graph)
   5259 {
   5260 	return carry_fallback(node, graph, 1);
   5261 }
   5262 
   5263 /* Topologically sort statements mapped to the same schedule iteration
   5264  * and add insert a sequence node in front of "node"
   5265  * corresponding to this order.
   5266  * If "initialized" is set, then it may be assumed that
   5267  * isl_sched_graph_compute_maxvar
   5268  * has been called on the current band.  Otherwise, call
   5269  * isl_sched_graph_compute_maxvar if and before carry_dependences gets called.
   5270  *
   5271  * If it turns out to be impossible to sort the statements apart,
   5272  * because different dependences impose different orderings
   5273  * on the statements, then we extend the schedule such that
   5274  * it carries at least one more dependence.
   5275  */
   5276 static __isl_give isl_schedule_node *sort_statements(
   5277 	__isl_take isl_schedule_node *node, struct isl_sched_graph *graph,
   5278 	int initialized)
   5279 {
   5280 	isl_ctx *ctx;
   5281 	isl_union_set_list *filters;
   5282 
   5283 	if (!node)
   5284 		return NULL;
   5285 
   5286 	ctx = isl_schedule_node_get_ctx(node);
   5287 	if (graph->n < 1)
   5288 		isl_die(ctx, isl_error_internal,
   5289 			"graph should have at least one node",
   5290 			return isl_schedule_node_free(node));
   5291 
   5292 	if (graph->n == 1)
   5293 		return node;
   5294 
   5295 	if (update_edges(ctx, graph) < 0)
   5296 		return isl_schedule_node_free(node);
   5297 
   5298 	if (graph->n_edge == 0)
   5299 		return node;
   5300 
   5301 	if (detect_sccs(ctx, graph) < 0)
   5302 		return isl_schedule_node_free(node);
   5303 
   5304 	next_band(graph);
   5305 	if (graph->scc < graph->n) {
   5306 		if (!initialized && isl_sched_graph_compute_maxvar(graph) < 0)
   5307 			return isl_schedule_node_free(node);
   5308 		return carry_dependences(node, graph);
   5309 	}
   5310 
   5311 	filters = isl_sched_graph_extract_sccs(ctx, graph);
   5312 	node = isl_schedule_node_insert_sequence(node, filters);
   5313 
   5314 	return node;
   5315 }
   5316 
   5317 /* Are there any (non-empty) (conditional) validity edges in the graph?
   5318  */
   5319 static int has_validity_edges(struct isl_sched_graph *graph)
   5320 {
   5321 	int i;
   5322 
   5323 	for (i = 0; i < graph->n_edge; ++i) {
   5324 		int empty;
   5325 
   5326 		empty = isl_map_plain_is_empty(graph->edge[i].map);
   5327 		if (empty < 0)
   5328 			return -1;
   5329 		if (empty)
   5330 			continue;
   5331 		if (is_any_validity(&graph->edge[i]))
   5332 			return 1;
   5333 	}
   5334 
   5335 	return 0;
   5336 }
   5337 
   5338 /* Should we apply a Feautrier step?
   5339  * That is, did the user request the Feautrier algorithm and are
   5340  * there any validity dependences (left)?
   5341  */
   5342 static int need_feautrier_step(isl_ctx *ctx, struct isl_sched_graph *graph)
   5343 {
   5344 	if (ctx->opt->schedule_algorithm != ISL_SCHEDULE_ALGORITHM_FEAUTRIER)
   5345 		return 0;
   5346 
   5347 	return has_validity_edges(graph);
   5348 }
   5349 
   5350 /* Compute a schedule for a connected dependence graph using Feautrier's
   5351  * multi-dimensional scheduling algorithm and return the updated schedule node.
   5352  *
   5353  * The original algorithm is described in [1].
   5354  * The main idea is to minimize the number of scheduling dimensions, by
   5355  * trying to satisfy as many dependences as possible per scheduling dimension.
   5356  *
   5357  * [1] P. Feautrier, Some Efficient Solutions to the Affine Scheduling
   5358  *     Problem, Part II: Multi-Dimensional Time.
   5359  *     In Intl. Journal of Parallel Programming, 1992.
   5360  */
   5361 static __isl_give isl_schedule_node *compute_schedule_wcc_feautrier(
   5362 	isl_schedule_node *node, struct isl_sched_graph *graph)
   5363 {
   5364 	return carry_feautrier(node, graph);
   5365 }
   5366 
   5367 /* Turn off the "local" bit on all (condition) edges.
   5368  */
   5369 static void clear_local_edges(struct isl_sched_graph *graph)
   5370 {
   5371 	int i;
   5372 
   5373 	for (i = 0; i < graph->n_edge; ++i)
   5374 		if (isl_sched_edge_is_condition(&graph->edge[i]))
   5375 			clear_local(&graph->edge[i]);
   5376 }
   5377 
   5378 /* Does "graph" have both condition and conditional validity edges?
   5379  */
   5380 static int need_condition_check(struct isl_sched_graph *graph)
   5381 {
   5382 	int i;
   5383 	int any_condition = 0;
   5384 	int any_conditional_validity = 0;
   5385 
   5386 	for (i = 0; i < graph->n_edge; ++i) {
   5387 		if (isl_sched_edge_is_condition(&graph->edge[i]))
   5388 			any_condition = 1;
   5389 		if (isl_sched_edge_is_conditional_validity(&graph->edge[i]))
   5390 			any_conditional_validity = 1;
   5391 	}
   5392 
   5393 	return any_condition && any_conditional_validity;
   5394 }
   5395 
   5396 /* Does "graph" contain any coincidence edge?
   5397  */
   5398 static int has_any_coincidence(struct isl_sched_graph *graph)
   5399 {
   5400 	int i;
   5401 
   5402 	for (i = 0; i < graph->n_edge; ++i)
   5403 		if (is_coincidence(&graph->edge[i]))
   5404 			return 1;
   5405 
   5406 	return 0;
   5407 }
   5408 
   5409 /* Extract the final schedule row as a map with the iteration domain
   5410  * of "node" as domain.
   5411  */
   5412 static __isl_give isl_map *final_row(struct isl_sched_node *node)
   5413 {
   5414 	isl_multi_aff *ma;
   5415 	isl_size n_row;
   5416 
   5417 	n_row = isl_mat_rows(node->sched);
   5418 	if (n_row < 0)
   5419 		return NULL;
   5420 	ma = isl_sched_node_extract_partial_schedule_multi_aff(node,
   5421 								n_row - 1, 1);
   5422 	return isl_map_from_multi_aff(ma);
   5423 }
   5424 
   5425 /* Is the conditional validity dependence in the edge with index "edge_index"
   5426  * violated by the latest (i.e., final) row of the schedule?
   5427  * That is, is i scheduled after j
   5428  * for any conditional validity dependence i -> j?
   5429  */
   5430 static int is_violated(struct isl_sched_graph *graph, int edge_index)
   5431 {
   5432 	isl_map *src_sched, *dst_sched, *map;
   5433 	struct isl_sched_edge *edge = &graph->edge[edge_index];
   5434 	int empty;
   5435 
   5436 	src_sched = final_row(edge->src);
   5437 	dst_sched = final_row(edge->dst);
   5438 	map = isl_map_copy(edge->map);
   5439 	map = isl_map_apply_domain(map, src_sched);
   5440 	map = isl_map_apply_range(map, dst_sched);
   5441 	map = isl_map_order_gt(map, isl_dim_in, 0, isl_dim_out, 0);
   5442 	empty = isl_map_is_empty(map);
   5443 	isl_map_free(map);
   5444 
   5445 	if (empty < 0)
   5446 		return -1;
   5447 
   5448 	return !empty;
   5449 }
   5450 
   5451 /* Does "graph" have any satisfied condition edges that
   5452  * are adjacent to the conditional validity constraint with
   5453  * domain "conditional_source" and range "conditional_sink"?
   5454  *
   5455  * A satisfied condition is one that is not local.
   5456  * If a condition was forced to be local already (i.e., marked as local)
   5457  * then there is no need to check if it is in fact local.
   5458  *
   5459  * Additionally, mark all adjacent condition edges found as local.
   5460  */
   5461 static int has_adjacent_true_conditions(struct isl_sched_graph *graph,
   5462 	__isl_keep isl_union_set *conditional_source,
   5463 	__isl_keep isl_union_set *conditional_sink)
   5464 {
   5465 	int i;
   5466 	int any = 0;
   5467 
   5468 	for (i = 0; i < graph->n_edge; ++i) {
   5469 		int adjacent, local;
   5470 		isl_union_map *condition;
   5471 
   5472 		if (!isl_sched_edge_is_condition(&graph->edge[i]))
   5473 			continue;
   5474 		if (is_local(&graph->edge[i]))
   5475 			continue;
   5476 
   5477 		condition = graph->edge[i].tagged_condition;
   5478 		adjacent = domain_intersects(condition, conditional_sink);
   5479 		if (adjacent >= 0 && !adjacent)
   5480 			adjacent = range_intersects(condition,
   5481 							conditional_source);
   5482 		if (adjacent < 0)
   5483 			return -1;
   5484 		if (!adjacent)
   5485 			continue;
   5486 
   5487 		set_local(&graph->edge[i]);
   5488 
   5489 		local = is_condition_false(&graph->edge[i]);
   5490 		if (local < 0)
   5491 			return -1;
   5492 		if (!local)
   5493 			any = 1;
   5494 	}
   5495 
   5496 	return any;
   5497 }
   5498 
   5499 /* Are there any violated conditional validity dependences with
   5500  * adjacent condition dependences that are not local with respect
   5501  * to the current schedule?
   5502  * That is, is the conditional validity constraint violated?
   5503  *
   5504  * Additionally, mark all those adjacent condition dependences as local.
   5505  * We also mark those adjacent condition dependences that were not marked
   5506  * as local before, but just happened to be local already.  This ensures
   5507  * that they remain local if the schedule is recomputed.
   5508  *
   5509  * We first collect domain and range of all violated conditional validity
   5510  * dependences and then check if there are any adjacent non-local
   5511  * condition dependences.
   5512  */
   5513 static int has_violated_conditional_constraint(isl_ctx *ctx,
   5514 	struct isl_sched_graph *graph)
   5515 {
   5516 	int i;
   5517 	int any = 0;
   5518 	isl_union_set *source, *sink;
   5519 
   5520 	source = isl_union_set_empty(isl_space_params_alloc(ctx, 0));
   5521 	sink = isl_union_set_empty(isl_space_params_alloc(ctx, 0));
   5522 	for (i = 0; i < graph->n_edge; ++i) {
   5523 		isl_union_set *uset;
   5524 		isl_union_map *umap;
   5525 		int violated;
   5526 
   5527 		if (!isl_sched_edge_is_conditional_validity(&graph->edge[i]))
   5528 			continue;
   5529 
   5530 		violated = is_violated(graph, i);
   5531 		if (violated < 0)
   5532 			goto error;
   5533 		if (!violated)
   5534 			continue;
   5535 
   5536 		any = 1;
   5537 
   5538 		umap = isl_union_map_copy(graph->edge[i].tagged_validity);
   5539 		uset = isl_union_map_domain(umap);
   5540 		source = isl_union_set_union(source, uset);
   5541 		source = isl_union_set_coalesce(source);
   5542 
   5543 		umap = isl_union_map_copy(graph->edge[i].tagged_validity);
   5544 		uset = isl_union_map_range(umap);
   5545 		sink = isl_union_set_union(sink, uset);
   5546 		sink = isl_union_set_coalesce(sink);
   5547 	}
   5548 
   5549 	if (any)
   5550 		any = has_adjacent_true_conditions(graph, source, sink);
   5551 
   5552 	isl_union_set_free(source);
   5553 	isl_union_set_free(sink);
   5554 	return any;
   5555 error:
   5556 	isl_union_set_free(source);
   5557 	isl_union_set_free(sink);
   5558 	return -1;
   5559 }
   5560 
   5561 /* Examine the current band (the rows between graph->band_start and
   5562  * graph->n_total_row), deciding whether to drop it or add it to "node"
   5563  * and then continue with the computation of the next band, if any.
   5564  * If "initialized" is set, then it may be assumed that
   5565  * isl_sched_graph_compute_maxvar
   5566  * has been called on the current band.  Otherwise, call
   5567  * isl_sched_graph_compute_maxvar if and before carry_dependences gets called.
   5568  *
   5569  * The caller keeps looking for a new row as long as
   5570  * graph->n_row < graph->maxvar.  If the latest attempt to find
   5571  * such a row failed (i.e., we still have graph->n_row < graph->maxvar),
   5572  * then we either
   5573  * - split between SCCs and start over (assuming we found an interesting
   5574  *	pair of SCCs between which to split)
   5575  * - continue with the next band (assuming the current band has at least
   5576  *	one row)
   5577  * - if there is more than one SCC left, then split along all SCCs
   5578  * - if outer coincidence needs to be enforced, then try to carry as many
   5579  *	validity or coincidence dependences as possible and
   5580  *	continue with the next band
   5581  * - try to carry as many validity dependences as possible and
   5582  *	continue with the next band
   5583  * In each case, we first insert a band node in the schedule tree
   5584  * if any rows have been computed.
   5585  *
   5586  * If the caller managed to complete the schedule and the current band
   5587  * is empty, then finish off by topologically
   5588  * sorting the statements based on the remaining dependences.
   5589  * If, on the other hand, the current band has at least one row,
   5590  * then continue with the next band.  Note that this next band
   5591  * will necessarily be empty, but the graph may still be split up
   5592  * into weakly connected components before arriving back here.
   5593  */
   5594 __isl_give isl_schedule_node *isl_schedule_node_compute_finish_band(
   5595 	__isl_take isl_schedule_node *node, struct isl_sched_graph *graph,
   5596 	int initialized)
   5597 {
   5598 	int empty;
   5599 
   5600 	if (!node)
   5601 		return NULL;
   5602 
   5603 	empty = graph->n_total_row == graph->band_start;
   5604 	if (graph->n_row < graph->maxvar) {
   5605 		isl_ctx *ctx;
   5606 
   5607 		ctx = isl_schedule_node_get_ctx(node);
   5608 		if (!ctx->opt->schedule_maximize_band_depth && !empty)
   5609 			return compute_next_band(node, graph, 1);
   5610 		if (graph->src_scc >= 0)
   5611 			return compute_split_schedule(node, graph);
   5612 		if (!empty)
   5613 			return compute_next_band(node, graph, 1);
   5614 		if (graph->scc > 1)
   5615 			return compute_component_schedule(node, graph, 1);
   5616 		if (!initialized && isl_sched_graph_compute_maxvar(graph) < 0)
   5617 			return isl_schedule_node_free(node);
   5618 		if (isl_options_get_schedule_outer_coincidence(ctx))
   5619 			return carry_coincidence(node, graph);
   5620 		return carry_dependences(node, graph);
   5621 	}
   5622 
   5623 	if (!empty)
   5624 		return compute_next_band(node, graph, 1);
   5625 	return sort_statements(node, graph, initialized);
   5626 }
   5627 
   5628 /* Construct a band of schedule rows for a connected dependence graph.
   5629  * The caller is responsible for determining the strongly connected
   5630  * components and calling isl_sched_graph_compute_maxvar first.
   5631  *
   5632  * We try to find a sequence of as many schedule rows as possible that result
   5633  * in non-negative dependence distances (independent of the previous rows
   5634  * in the sequence, i.e., such that the sequence is tilable), with as
   5635  * many of the initial rows as possible satisfying the coincidence constraints.
   5636  * The computation stops if we can't find any more rows or if we have found
   5637  * all the rows we wanted to find.
   5638  *
   5639  * If ctx->opt->schedule_outer_coincidence is set, then we force the
   5640  * outermost dimension to satisfy the coincidence constraints.  If this
   5641  * turns out to be impossible, we fall back on the general scheme above
   5642  * and try to carry as many dependences as possible.
   5643  *
   5644  * If "graph" contains both condition and conditional validity dependences,
   5645  * then we need to check that that the conditional schedule constraint
   5646  * is satisfied, i.e., there are no violated conditional validity dependences
   5647  * that are adjacent to any non-local condition dependences.
   5648  * If there are, then we mark all those adjacent condition dependences
   5649  * as local and recompute the current band.  Those dependences that
   5650  * are marked local will then be forced to be local.
   5651  * The initial computation is performed with no dependences marked as local.
   5652  * If we are lucky, then there will be no violated conditional validity
   5653  * dependences adjacent to any non-local condition dependences.
   5654  * Otherwise, we mark some additional condition dependences as local and
   5655  * recompute.  We continue this process until there are no violations left or
   5656  * until we are no longer able to compute a schedule.
   5657  * Since there are only a finite number of dependences,
   5658  * there will only be a finite number of iterations.
   5659  */
   5660 isl_stat isl_schedule_node_compute_wcc_band(isl_ctx *ctx,
   5661 	struct isl_sched_graph *graph)
   5662 {
   5663 	int has_coincidence;
   5664 	int use_coincidence;
   5665 	int force_coincidence = 0;
   5666 	int check_conditional;
   5667 
   5668 	if (sort_sccs(graph) < 0)
   5669 		return isl_stat_error;
   5670 
   5671 	clear_local_edges(graph);
   5672 	check_conditional = need_condition_check(graph);
   5673 	has_coincidence = has_any_coincidence(graph);
   5674 
   5675 	if (ctx->opt->schedule_outer_coincidence)
   5676 		force_coincidence = 1;
   5677 
   5678 	use_coincidence = has_coincidence;
   5679 	while (graph->n_row < graph->maxvar) {
   5680 		isl_vec *sol;
   5681 		int violated;
   5682 		int coincident;
   5683 
   5684 		graph->src_scc = -1;
   5685 		graph->dst_scc = -1;
   5686 
   5687 		if (setup_lp(ctx, graph, use_coincidence) < 0)
   5688 			return isl_stat_error;
   5689 		sol = solve_lp(ctx, graph);
   5690 		if (!sol)
   5691 			return isl_stat_error;
   5692 		if (sol->size == 0) {
   5693 			int empty = graph->n_total_row == graph->band_start;
   5694 
   5695 			isl_vec_free(sol);
   5696 			if (use_coincidence && (!force_coincidence || !empty)) {
   5697 				use_coincidence = 0;
   5698 				continue;
   5699 			}
   5700 			return isl_stat_ok;
   5701 		}
   5702 		coincident = !has_coincidence || use_coincidence;
   5703 		if (update_schedule(graph, sol, coincident) < 0)
   5704 			return isl_stat_error;
   5705 
   5706 		if (!check_conditional)
   5707 			continue;
   5708 		violated = has_violated_conditional_constraint(ctx, graph);
   5709 		if (violated < 0)
   5710 			return isl_stat_error;
   5711 		if (!violated)
   5712 			continue;
   5713 		if (reset_band(graph) < 0)
   5714 			return isl_stat_error;
   5715 		use_coincidence = has_coincidence;
   5716 	}
   5717 
   5718 	return isl_stat_ok;
   5719 }
   5720 
   5721 /* Compute a schedule for a connected dependence graph by considering
   5722  * the graph as a whole and return the updated schedule node.
   5723  *
   5724  * The actual schedule rows of the current band are computed by
   5725  * isl_schedule_node_compute_wcc_band.  isl_schedule_node_compute_finish_band
   5726  * takes care of integrating the band into "node" and continuing
   5727  * the computation.
   5728  */
   5729 static __isl_give isl_schedule_node *compute_schedule_wcc_whole(
   5730 	__isl_take isl_schedule_node *node, struct isl_sched_graph *graph)
   5731 {
   5732 	isl_ctx *ctx;
   5733 
   5734 	if (!node)
   5735 		return NULL;
   5736 
   5737 	ctx = isl_schedule_node_get_ctx(node);
   5738 	if (isl_schedule_node_compute_wcc_band(ctx, graph) < 0)
   5739 		return isl_schedule_node_free(node);
   5740 
   5741 	return isl_schedule_node_compute_finish_band(node, graph, 1);
   5742 }
   5743 
   5744 /* Compute a schedule for a connected dependence graph and return
   5745  * the updated schedule node.
   5746  *
   5747  * If Feautrier's algorithm is selected, we first recursively try to satisfy
   5748  * as many validity dependences as possible. When all validity dependences
   5749  * are satisfied we extend the schedule to a full-dimensional schedule.
   5750  *
   5751  * Call compute_schedule_wcc_whole or isl_schedule_node_compute_wcc_clustering
   5752  * depending on whether the user has selected the option to try and
   5753  * compute a schedule for the entire (weakly connected) component first.
   5754  * If there is only a single strongly connected component (SCC), then
   5755  * there is no point in trying to combine SCCs
   5756  * in isl_schedule_node_compute_wcc_clustering, so compute_schedule_wcc_whole
   5757  * is called instead.
   5758  */
   5759 static __isl_give isl_schedule_node *compute_schedule_wcc(
   5760 	__isl_take isl_schedule_node *node, struct isl_sched_graph *graph)
   5761 {
   5762 	isl_ctx *ctx;
   5763 
   5764 	if (!node)
   5765 		return NULL;
   5766 
   5767 	ctx = isl_schedule_node_get_ctx(node);
   5768 	if (detect_sccs(ctx, graph) < 0)
   5769 		return isl_schedule_node_free(node);
   5770 
   5771 	if (isl_sched_graph_compute_maxvar(graph) < 0)
   5772 		return isl_schedule_node_free(node);
   5773 
   5774 	if (need_feautrier_step(ctx, graph))
   5775 		return compute_schedule_wcc_feautrier(node, graph);
   5776 
   5777 	if (graph->scc <= 1 || isl_options_get_schedule_whole_component(ctx))
   5778 		return compute_schedule_wcc_whole(node, graph);
   5779 	else
   5780 		return isl_schedule_node_compute_wcc_clustering(node, graph);
   5781 }
   5782 
   5783 /* Compute a schedule for each group of nodes identified by node->scc
   5784  * separately and then combine them in a sequence node (or as set node
   5785  * if graph->weak is set) inserted at position "node" of the schedule tree.
   5786  * Return the updated schedule node.
   5787  *
   5788  * If "wcc" is set then each of the groups belongs to a single
   5789  * weakly connected component in the dependence graph so that
   5790  * there is no need for compute_sub_schedule to look for weakly
   5791  * connected components.
   5792  *
   5793  * If a set node would be introduced and if the number of components
   5794  * is equal to the number of nodes, then check if the schedule
   5795  * is already complete.  If so, a redundant set node would be introduced
   5796  * (without any further descendants) stating that the statements
   5797  * can be executed in arbitrary order, which is also expressed
   5798  * by the absence of any node.  Refrain from inserting any nodes
   5799  * in this case and simply return.
   5800  */
   5801 static __isl_give isl_schedule_node *compute_component_schedule(
   5802 	__isl_take isl_schedule_node *node, struct isl_sched_graph *graph,
   5803 	int wcc)
   5804 {
   5805 	int component;
   5806 	isl_ctx *ctx;
   5807 	isl_union_set_list *filters;
   5808 
   5809 	if (!node)
   5810 		return NULL;
   5811 
   5812 	if (graph->weak && graph->scc == graph->n) {
   5813 		if (isl_sched_graph_compute_maxvar(graph) < 0)
   5814 			return isl_schedule_node_free(node);
   5815 		if (graph->n_row >= graph->maxvar)
   5816 			return node;
   5817 	}
   5818 
   5819 	ctx = isl_schedule_node_get_ctx(node);
   5820 	filters = isl_sched_graph_extract_sccs(ctx, graph);
   5821 	if (graph->weak)
   5822 		node = isl_schedule_node_insert_set(node, filters);
   5823 	else
   5824 		node = isl_schedule_node_insert_sequence(node, filters);
   5825 
   5826 	for (component = 0; component < graph->scc; ++component) {
   5827 		node = isl_schedule_node_grandchild(node, component, 0);
   5828 		node = compute_sub_schedule(node, ctx, graph,
   5829 				    &isl_sched_node_scc_exactly,
   5830 				    &isl_sched_edge_scc_exactly,
   5831 				    component, wcc);
   5832 		node = isl_schedule_node_grandparent(node);
   5833 	}
   5834 
   5835 	return node;
   5836 }
   5837 
   5838 /* Compute a schedule for the given dependence graph and insert it at "node".
   5839  * Return the updated schedule node.
   5840  *
   5841  * We first check if the graph is connected (through validity and conditional
   5842  * validity dependences) and, if not, compute a schedule
   5843  * for each component separately.
   5844  * If the schedule_serialize_sccs option is set, then we check for strongly
   5845  * connected components instead and compute a separate schedule for
   5846  * each such strongly connected component.
   5847  */
   5848 static __isl_give isl_schedule_node *compute_schedule(isl_schedule_node *node,
   5849 	struct isl_sched_graph *graph)
   5850 {
   5851 	isl_ctx *ctx;
   5852 
   5853 	if (!node)
   5854 		return NULL;
   5855 
   5856 	ctx = isl_schedule_node_get_ctx(node);
   5857 	if (isl_options_get_schedule_serialize_sccs(ctx)) {
   5858 		if (detect_sccs(ctx, graph) < 0)
   5859 			return isl_schedule_node_free(node);
   5860 	} else {
   5861 		if (detect_wccs(ctx, graph) < 0)
   5862 			return isl_schedule_node_free(node);
   5863 	}
   5864 
   5865 	if (graph->scc > 1)
   5866 		return compute_component_schedule(node, graph, 1);
   5867 
   5868 	return compute_schedule_wcc(node, graph);
   5869 }
   5870 
   5871 /* Compute a schedule on sc->domain that respects the given schedule
   5872  * constraints.
   5873  *
   5874  * In particular, the schedule respects all the validity dependences.
   5875  * If the default isl scheduling algorithm is used, it tries to minimize
   5876  * the dependence distances over the proximity dependences.
   5877  * If Feautrier's scheduling algorithm is used, the proximity dependence
   5878  * distances are only minimized during the extension to a full-dimensional
   5879  * schedule.
   5880  *
   5881  * If there are any condition and conditional validity dependences,
   5882  * then the conditional validity dependences may be violated inside
   5883  * a tilable band, provided they have no adjacent non-local
   5884  * condition dependences.
   5885  */
   5886 __isl_give isl_schedule *isl_schedule_constraints_compute_schedule(
   5887 	__isl_take isl_schedule_constraints *sc)
   5888 {
   5889 	isl_ctx *ctx = isl_schedule_constraints_get_ctx(sc);
   5890 	struct isl_sched_graph graph = { 0 };
   5891 	isl_schedule *sched;
   5892 	isl_schedule_node *node;
   5893 	isl_union_set *domain;
   5894 	isl_size n;
   5895 
   5896 	sc = isl_schedule_constraints_align_params(sc);
   5897 
   5898 	domain = isl_schedule_constraints_get_domain(sc);
   5899 	n = isl_union_set_n_set(domain);
   5900 	if (n == 0) {
   5901 		isl_schedule_constraints_free(sc);
   5902 		return isl_schedule_from_domain(domain);
   5903 	}
   5904 
   5905 	if (n < 0 || isl_sched_graph_init(&graph, sc) < 0)
   5906 		domain = isl_union_set_free(domain);
   5907 
   5908 	node = isl_schedule_node_from_domain(domain);
   5909 	node = isl_schedule_node_child(node, 0);
   5910 	if (graph.n > 0)
   5911 		node = compute_schedule(node, &graph);
   5912 	sched = isl_schedule_node_get_schedule(node);
   5913 	isl_schedule_node_free(node);
   5914 
   5915 	isl_sched_graph_free(ctx, &graph);
   5916 	isl_schedule_constraints_free(sc);
   5917 
   5918 	return sched;
   5919 }
   5920 
   5921 /* Compute a schedule for the given union of domains that respects
   5922  * all the validity dependences and minimizes
   5923  * the dependence distances over the proximity dependences.
   5924  *
   5925  * This function is kept for backward compatibility.
   5926  */
   5927 __isl_give isl_schedule *isl_union_set_compute_schedule(
   5928 	__isl_take isl_union_set *domain,
   5929 	__isl_take isl_union_map *validity,
   5930 	__isl_take isl_union_map *proximity)
   5931 {
   5932 	isl_schedule_constraints *sc;
   5933 
   5934 	sc = isl_schedule_constraints_on_domain(domain);
   5935 	sc = isl_schedule_constraints_set_validity(sc, validity);
   5936 	sc = isl_schedule_constraints_set_proximity(sc, proximity);
   5937 
   5938 	return isl_schedule_constraints_compute_schedule(sc);
   5939 }
   5940