Home | History | Annotate | Line # | Download | only in dist
isl_ast_codegen.c revision 1.1
      1 /*
      2  * Copyright 2012-2014 Ecole Normale Superieure
      3  * Copyright 2014      INRIA Rocquencourt
      4  *
      5  * Use of this software is governed by the MIT license
      6  *
      7  * Written by Sven Verdoolaege,
      8  * Ecole Normale Superieure, 45 rue dUlm, 75230 Paris, France
      9  * and Inria Paris - Rocquencourt, Domaine de Voluceau - Rocquencourt,
     10  * B.P. 105 - 78153 Le Chesnay, France
     11  */
     12 
     13 #include <limits.h>
     14 #include <isl/id.h>
     15 #include <isl/val.h>
     16 #include <isl/space.h>
     17 #include <isl/aff.h>
     18 #include <isl/constraint.h>
     19 #include <isl/set.h>
     20 #include <isl/ilp.h>
     21 #include <isl/union_set.h>
     22 #include <isl/union_map.h>
     23 #include <isl/schedule_node.h>
     24 #include <isl/options.h>
     25 #include <isl_sort.h>
     26 #include <isl_tarjan.h>
     27 #include <isl_ast_private.h>
     28 #include <isl_ast_build_expr.h>
     29 #include <isl_ast_build_private.h>
     30 #include <isl_ast_graft_private.h>
     31 
     32 /* Try and reduce the number of disjuncts in the representation of "set",
     33  * without dropping explicit representations of local variables.
     34  */
     35 static __isl_give isl_set *isl_set_coalesce_preserve(__isl_take isl_set *set)
     36 {
     37 	isl_ctx *ctx;
     38 	int save_preserve;
     39 
     40 	if (!set)
     41 		return NULL;
     42 
     43 	ctx = isl_set_get_ctx(set);
     44 	save_preserve = isl_options_get_coalesce_preserve_locals(ctx);
     45 	isl_options_set_coalesce_preserve_locals(ctx, 1);
     46 	set = isl_set_coalesce(set);
     47 	isl_options_set_coalesce_preserve_locals(ctx, save_preserve);
     48 	return set;
     49 }
     50 
     51 /* Data used in generate_domain.
     52  *
     53  * "build" is the input build.
     54  * "list" collects the results.
     55  */
     56 struct isl_generate_domain_data {
     57 	isl_ast_build *build;
     58 
     59 	isl_ast_graft_list *list;
     60 };
     61 
     62 static __isl_give isl_ast_graft_list *generate_next_level(
     63 	__isl_take isl_union_map *executed,
     64 	__isl_take isl_ast_build *build);
     65 static __isl_give isl_ast_graft_list *generate_code(
     66 	__isl_take isl_union_map *executed, __isl_take isl_ast_build *build,
     67 	int internal);
     68 
     69 /* Generate an AST for a single domain based on
     70  * the (non single valued) inverse schedule "executed".
     71  *
     72  * We extend the schedule with the iteration domain
     73  * and continue generating through a call to generate_code.
     74  *
     75  * In particular, if executed has the form
     76  *
     77  *	S -> D
     78  *
     79  * then we continue generating code on
     80  *
     81  *	[S -> D] -> D
     82  *
     83  * The extended inverse schedule is clearly single valued
     84  * ensuring that the nested generate_code will not reach this function,
     85  * but will instead create calls to all elements of D that need
     86  * to be executed from the current schedule domain.
     87  */
     88 static isl_stat generate_non_single_valued(__isl_take isl_map *executed,
     89 	struct isl_generate_domain_data *data)
     90 {
     91 	isl_map *identity;
     92 	isl_ast_build *build;
     93 	isl_ast_graft_list *list;
     94 
     95 	build = isl_ast_build_copy(data->build);
     96 
     97 	identity = isl_set_identity(isl_map_range(isl_map_copy(executed)));
     98 	executed = isl_map_domain_product(executed, identity);
     99 	build = isl_ast_build_set_single_valued(build, 1);
    100 
    101 	list = generate_code(isl_union_map_from_map(executed), build, 1);
    102 
    103 	data->list = isl_ast_graft_list_concat(data->list, list);
    104 
    105 	return isl_stat_ok;
    106 }
    107 
    108 /* Call the at_each_domain callback, if requested by the user,
    109  * after recording the current inverse schedule in the build.
    110  */
    111 static __isl_give isl_ast_graft *at_each_domain(__isl_take isl_ast_graft *graft,
    112 	__isl_keep isl_map *executed, __isl_keep isl_ast_build *build)
    113 {
    114 	if (!graft || !build)
    115 		return isl_ast_graft_free(graft);
    116 	if (!build->at_each_domain)
    117 		return graft;
    118 
    119 	build = isl_ast_build_copy(build);
    120 	build = isl_ast_build_set_executed(build,
    121 			isl_union_map_from_map(isl_map_copy(executed)));
    122 	if (!build)
    123 		return isl_ast_graft_free(graft);
    124 
    125 	graft->node = build->at_each_domain(graft->node,
    126 					build, build->at_each_domain_user);
    127 	isl_ast_build_free(build);
    128 
    129 	if (!graft->node)
    130 		graft = isl_ast_graft_free(graft);
    131 
    132 	return graft;
    133 }
    134 
    135 /* Generate a call expression for the single executed
    136  * domain element "map" and put a guard around it based its (simplified)
    137  * domain.  "executed" is the original inverse schedule from which "map"
    138  * has been derived.  In particular, "map" is either identical to "executed"
    139  * or it is the result of gisting "executed" with respect to the build domain.
    140  * "executed" is only used if there is an at_each_domain callback.
    141  *
    142  * At this stage, any pending constraints in the build can no longer
    143  * be simplified with respect to any enforced constraints since
    144  * the call node does not have any enforced constraints.
    145  * Since all pending constraints not covered by any enforced constraints
    146  * will be added as a guard to the graft in create_node_scaled,
    147  * even in the eliminated case, the pending constraints
    148  * can be considered to have been generated by outer constructs.
    149  *
    150  * If the user has set an at_each_domain callback, it is called
    151  * on the constructed call expression node.
    152  */
    153 static isl_stat add_domain(__isl_take isl_map *executed,
    154 	__isl_take isl_map *map, struct isl_generate_domain_data *data)
    155 {
    156 	isl_ast_build *build;
    157 	isl_ast_graft *graft;
    158 	isl_ast_graft_list *list;
    159 	isl_set *guard, *pending;
    160 
    161 	build = isl_ast_build_copy(data->build);
    162 	pending = isl_ast_build_get_pending(build);
    163 	build = isl_ast_build_replace_pending_by_guard(build, pending);
    164 
    165 	guard = isl_map_domain(isl_map_copy(map));
    166 	guard = isl_set_compute_divs(guard);
    167 	guard = isl_set_coalesce_preserve(guard);
    168 	guard = isl_set_gist(guard, isl_ast_build_get_generated(build));
    169 	guard = isl_ast_build_specialize(build, guard);
    170 
    171 	graft = isl_ast_graft_alloc_domain(map, build);
    172 	graft = at_each_domain(graft, executed, build);
    173 	isl_ast_build_free(build);
    174 	isl_map_free(executed);
    175 	graft = isl_ast_graft_add_guard(graft, guard, data->build);
    176 
    177 	list = isl_ast_graft_list_from_ast_graft(graft);
    178 	data->list = isl_ast_graft_list_concat(data->list, list);
    179 
    180 	return isl_stat_ok;
    181 }
    182 
    183 /* Generate an AST for a single domain based on
    184  * the inverse schedule "executed" and add it to data->list.
    185  *
    186  * If there is more than one domain element associated to the current
    187  * schedule "time", then we need to continue the generation process
    188  * in generate_non_single_valued.
    189  * Note that the inverse schedule being single-valued may depend
    190  * on constraints that are only available in the original context
    191  * domain specified by the user.  We therefore first introduce
    192  * some of the constraints of data->build->domain.  In particular,
    193  * we intersect with a single-disjunct approximation of this set.
    194  * We perform this approximation to avoid further splitting up
    195  * the executed relation, possibly introducing a disjunctive guard
    196  * on the statement.
    197  *
    198  * On the other hand, we only perform the test after having taken the gist
    199  * of the domain as the resulting map is the one from which the call
    200  * expression is constructed.  Using this map to construct the call
    201  * expression usually yields simpler results in cases where the original
    202  * map is not obviously single-valued.
    203  * If the original map is obviously single-valued, then the gist
    204  * operation is skipped.
    205  *
    206  * Because we perform the single-valuedness test on the gisted map,
    207  * we may in rare cases fail to recognize that the inverse schedule
    208  * is single-valued.  This becomes problematic if this happens
    209  * from the recursive call through generate_non_single_valued
    210  * as we would then end up in an infinite recursion.
    211  * We therefore check if we are inside a call to generate_non_single_valued
    212  * and revert to the ungisted map if the gisted map turns out not to be
    213  * single-valued.
    214  *
    215  * Otherwise, call add_domain to generate a call expression (with guard) and
    216  * to call the at_each_domain callback, if any.
    217  */
    218 static isl_stat generate_domain(__isl_take isl_map *executed, void *user)
    219 {
    220 	struct isl_generate_domain_data *data = user;
    221 	isl_set *domain;
    222 	isl_map *map = NULL;
    223 	int empty, sv;
    224 
    225 	domain = isl_ast_build_get_domain(data->build);
    226 	domain = isl_set_from_basic_set(isl_set_simple_hull(domain));
    227 	executed = isl_map_intersect_domain(executed, domain);
    228 	empty = isl_map_is_empty(executed);
    229 	if (empty < 0)
    230 		goto error;
    231 	if (empty) {
    232 		isl_map_free(executed);
    233 		return isl_stat_ok;
    234 	}
    235 
    236 	sv = isl_map_plain_is_single_valued(executed);
    237 	if (sv < 0)
    238 		goto error;
    239 	if (sv)
    240 		return add_domain(executed, isl_map_copy(executed), data);
    241 
    242 	executed = isl_map_coalesce(executed);
    243 	map = isl_map_copy(executed);
    244 	map = isl_ast_build_compute_gist_map_domain(data->build, map);
    245 	sv = isl_map_is_single_valued(map);
    246 	if (sv < 0)
    247 		goto error;
    248 	if (!sv) {
    249 		isl_map_free(map);
    250 		if (data->build->single_valued)
    251 			map = isl_map_copy(executed);
    252 		else
    253 			return generate_non_single_valued(executed, data);
    254 	}
    255 
    256 	return add_domain(executed, map, data);
    257 error:
    258 	isl_map_free(map);
    259 	isl_map_free(executed);
    260 	return isl_stat_error;
    261 }
    262 
    263 /* Call build->create_leaf to a create "leaf" node in the AST,
    264  * encapsulate the result in an isl_ast_graft and return the result
    265  * as a 1-element list.
    266  *
    267  * Note that the node returned by the user may be an entire tree.
    268  *
    269  * Since the node itself cannot enforce any constraints, we turn
    270  * all pending constraints into guards and add them to the resulting
    271  * graft to ensure that they will be generated.
    272  *
    273  * Before we pass control to the user, we first clear some information
    274  * from the build that is (presumbably) only meaningful
    275  * for the current code generation.
    276  * This includes the create_leaf callback itself, so we make a copy
    277  * of the build first.
    278  */
    279 static __isl_give isl_ast_graft_list *call_create_leaf(
    280 	__isl_take isl_union_map *executed, __isl_take isl_ast_build *build)
    281 {
    282 	isl_set *guard;
    283 	isl_ast_node *node;
    284 	isl_ast_graft *graft;
    285 	isl_ast_build *user_build;
    286 
    287 	guard = isl_ast_build_get_pending(build);
    288 	user_build = isl_ast_build_copy(build);
    289 	user_build = isl_ast_build_replace_pending_by_guard(user_build,
    290 							isl_set_copy(guard));
    291 	user_build = isl_ast_build_set_executed(user_build, executed);
    292 	user_build = isl_ast_build_clear_local_info(user_build);
    293 	if (!user_build)
    294 		node = NULL;
    295 	else
    296 		node = build->create_leaf(user_build, build->create_leaf_user);
    297 	graft = isl_ast_graft_alloc(node, build);
    298 	graft = isl_ast_graft_add_guard(graft, guard, build);
    299 	isl_ast_build_free(build);
    300 	return isl_ast_graft_list_from_ast_graft(graft);
    301 }
    302 
    303 static __isl_give isl_ast_graft_list *build_ast_from_child(
    304 	__isl_take isl_ast_build *build, __isl_take isl_schedule_node *node,
    305 	__isl_take isl_union_map *executed);
    306 
    307 /* Generate an AST after having handled the complete schedule
    308  * of this call to the code generator or the complete band
    309  * if we are generating an AST from a schedule tree.
    310  *
    311  * If we are inside a band node, then move on to the child of the band.
    312  *
    313  * If the user has specified a create_leaf callback, control
    314  * is passed to the user in call_create_leaf.
    315  *
    316  * Otherwise, we generate one or more calls for each individual
    317  * domain in generate_domain.
    318  */
    319 static __isl_give isl_ast_graft_list *generate_inner_level(
    320 	__isl_take isl_union_map *executed, __isl_take isl_ast_build *build)
    321 {
    322 	isl_ctx *ctx;
    323 	struct isl_generate_domain_data data = { build };
    324 
    325 	if (!build || !executed)
    326 		goto error;
    327 
    328 	if (isl_ast_build_has_schedule_node(build)) {
    329 		isl_schedule_node *node;
    330 		node = isl_ast_build_get_schedule_node(build);
    331 		build = isl_ast_build_reset_schedule_node(build);
    332 		return build_ast_from_child(build, node, executed);
    333 	}
    334 
    335 	if (build->create_leaf)
    336 		return call_create_leaf(executed, build);
    337 
    338 	ctx = isl_union_map_get_ctx(executed);
    339 	data.list = isl_ast_graft_list_alloc(ctx, 0);
    340 	if (isl_union_map_foreach_map(executed, &generate_domain, &data) < 0)
    341 		data.list = isl_ast_graft_list_free(data.list);
    342 
    343 	if (0)
    344 error:		data.list = NULL;
    345 	isl_ast_build_free(build);
    346 	isl_union_map_free(executed);
    347 	return data.list;
    348 }
    349 
    350 /* Call the before_each_for callback, if requested by the user.
    351  */
    352 static __isl_give isl_ast_node *before_each_for(__isl_take isl_ast_node *node,
    353 	__isl_keep isl_ast_build *build)
    354 {
    355 	isl_id *id;
    356 
    357 	if (!node || !build)
    358 		return isl_ast_node_free(node);
    359 	if (!build->before_each_for)
    360 		return node;
    361 	id = build->before_each_for(build, build->before_each_for_user);
    362 	node = isl_ast_node_set_annotation(node, id);
    363 	return node;
    364 }
    365 
    366 /* Call the after_each_for callback, if requested by the user.
    367  */
    368 static __isl_give isl_ast_graft *after_each_for(__isl_take isl_ast_graft *graft,
    369 	__isl_keep isl_ast_build *build)
    370 {
    371 	if (!graft || !build)
    372 		return isl_ast_graft_free(graft);
    373 	if (!build->after_each_for)
    374 		return graft;
    375 	graft->node = build->after_each_for(graft->node, build,
    376 						build->after_each_for_user);
    377 	if (!graft->node)
    378 		return isl_ast_graft_free(graft);
    379 	return graft;
    380 }
    381 
    382 /* Plug in all the know values of the current and outer dimensions
    383  * in the domain of "executed".  In principle, we only need to plug
    384  * in the known value of the current dimension since the values of
    385  * outer dimensions have been plugged in already.
    386  * However, it turns out to be easier to just plug in all known values.
    387  */
    388 static __isl_give isl_union_map *plug_in_values(
    389 	__isl_take isl_union_map *executed, __isl_keep isl_ast_build *build)
    390 {
    391 	return isl_ast_build_substitute_values_union_map_domain(build,
    392 								    executed);
    393 }
    394 
    395 /* Check if the constraint "c" is a lower bound on dimension "pos",
    396  * an upper bound, or independent of dimension "pos".
    397  */
    398 static int constraint_type(isl_constraint *c, int pos)
    399 {
    400 	if (isl_constraint_is_lower_bound(c, isl_dim_set, pos))
    401 		return 1;
    402 	if (isl_constraint_is_upper_bound(c, isl_dim_set, pos))
    403 		return 2;
    404 	return 0;
    405 }
    406 
    407 /* Compare the types of the constraints "a" and "b",
    408  * resulting in constraints that are independent of "depth"
    409  * to be sorted before the lower bounds on "depth", which in
    410  * turn are sorted before the upper bounds on "depth".
    411  */
    412 static int cmp_constraint(__isl_keep isl_constraint *a,
    413 	__isl_keep isl_constraint *b, void *user)
    414 {
    415 	int *depth = user;
    416 	int t1 = constraint_type(a, *depth);
    417 	int t2 = constraint_type(b, *depth);
    418 
    419 	return t1 - t2;
    420 }
    421 
    422 /* Extract a lower bound on dimension "pos" from constraint "c".
    423  *
    424  * If the constraint is of the form
    425  *
    426  *	a x + f(...) >= 0
    427  *
    428  * then we essentially return
    429  *
    430  *	l = ceil(-f(...)/a)
    431  *
    432  * However, if the current dimension is strided, then we need to make
    433  * sure that the lower bound we construct is of the form
    434  *
    435  *	f + s a
    436  *
    437  * with f the offset and s the stride.
    438  * We therefore compute
    439  *
    440  *	f + s * ceil((l - f)/s)
    441  */
    442 static __isl_give isl_aff *lower_bound(__isl_keep isl_constraint *c,
    443 	int pos, __isl_keep isl_ast_build *build)
    444 {
    445 	isl_aff *aff;
    446 
    447 	aff = isl_constraint_get_bound(c, isl_dim_set, pos);
    448 	aff = isl_aff_ceil(aff);
    449 
    450 	if (isl_ast_build_has_stride(build, pos)) {
    451 		isl_aff *offset;
    452 		isl_val *stride;
    453 
    454 		offset = isl_ast_build_get_offset(build, pos);
    455 		stride = isl_ast_build_get_stride(build, pos);
    456 
    457 		aff = isl_aff_sub(aff, isl_aff_copy(offset));
    458 		aff = isl_aff_scale_down_val(aff, isl_val_copy(stride));
    459 		aff = isl_aff_ceil(aff);
    460 		aff = isl_aff_scale_val(aff, stride);
    461 		aff = isl_aff_add(aff, offset);
    462 	}
    463 
    464 	aff = isl_ast_build_compute_gist_aff(build, aff);
    465 
    466 	return aff;
    467 }
    468 
    469 /* Return the exact lower bound (or upper bound if "upper" is set)
    470  * of "domain" as a piecewise affine expression.
    471  *
    472  * If we are computing a lower bound (of a strided dimension), then
    473  * we need to make sure it is of the form
    474  *
    475  *	f + s a
    476  *
    477  * where f is the offset and s is the stride.
    478  * We therefore need to include the stride constraint before computing
    479  * the minimum.
    480  */
    481 static __isl_give isl_pw_aff *exact_bound(__isl_keep isl_set *domain,
    482 	__isl_keep isl_ast_build *build, int upper)
    483 {
    484 	isl_set *stride;
    485 	isl_map *it_map;
    486 	isl_pw_aff *pa;
    487 	isl_pw_multi_aff *pma;
    488 
    489 	domain = isl_set_copy(domain);
    490 	if (!upper) {
    491 		stride = isl_ast_build_get_stride_constraint(build);
    492 		domain = isl_set_intersect(domain, stride);
    493 	}
    494 	it_map = isl_ast_build_map_to_iterator(build, domain);
    495 	if (upper)
    496 		pma = isl_map_lexmax_pw_multi_aff(it_map);
    497 	else
    498 		pma = isl_map_lexmin_pw_multi_aff(it_map);
    499 	pa = isl_pw_multi_aff_get_pw_aff(pma, 0);
    500 	isl_pw_multi_aff_free(pma);
    501 	pa = isl_ast_build_compute_gist_pw_aff(build, pa);
    502 	pa = isl_pw_aff_coalesce(pa);
    503 
    504 	return pa;
    505 }
    506 
    507 /* Callback for sorting the isl_pw_aff_list passed to reduce_list and
    508  * remove_redundant_lower_bounds.
    509  */
    510 static int reduce_list_cmp(__isl_keep isl_pw_aff *a, __isl_keep isl_pw_aff *b,
    511 	void *user)
    512 {
    513 	return isl_pw_aff_plain_cmp(a, b);
    514 }
    515 
    516 /* Given a list of lower bounds "list", remove those that are redundant
    517  * with respect to the other bounds in "list" and the domain of "build".
    518  *
    519  * We first sort the bounds in the same way as they would be sorted
    520  * by set_for_node_expressions so that we can try and remove the last
    521  * bounds first.
    522  *
    523  * For a lower bound to be effective, there needs to be at least
    524  * one domain element for which it is larger than all other lower bounds.
    525  * For each lower bound we therefore intersect the domain with
    526  * the conditions that it is larger than all other bounds and
    527  * check whether the result is empty.  If so, the bound can be removed.
    528  */
    529 static __isl_give isl_pw_aff_list *remove_redundant_lower_bounds(
    530 	__isl_take isl_pw_aff_list *list, __isl_keep isl_ast_build *build)
    531 {
    532 	int i, j;
    533 	isl_size n;
    534 	isl_set *domain;
    535 
    536 	list = isl_pw_aff_list_sort(list, &reduce_list_cmp, NULL);
    537 
    538 	n = isl_pw_aff_list_n_pw_aff(list);
    539 	if (n < 0)
    540 		return isl_pw_aff_list_free(list);
    541 	if (n <= 1)
    542 		return list;
    543 
    544 	domain = isl_ast_build_get_domain(build);
    545 
    546 	for (i = n - 1; i >= 0; --i) {
    547 		isl_pw_aff *pa_i;
    548 		isl_set *domain_i;
    549 		int empty;
    550 
    551 		domain_i = isl_set_copy(domain);
    552 		pa_i = isl_pw_aff_list_get_pw_aff(list, i);
    553 
    554 		for (j = 0; j < n; ++j) {
    555 			isl_pw_aff *pa_j;
    556 			isl_set *better;
    557 
    558 			if (j == i)
    559 				continue;
    560 
    561 			pa_j = isl_pw_aff_list_get_pw_aff(list, j);
    562 			better = isl_pw_aff_gt_set(isl_pw_aff_copy(pa_i), pa_j);
    563 			domain_i = isl_set_intersect(domain_i, better);
    564 		}
    565 
    566 		empty = isl_set_is_empty(domain_i);
    567 
    568 		isl_set_free(domain_i);
    569 		isl_pw_aff_free(pa_i);
    570 
    571 		if (empty < 0)
    572 			goto error;
    573 		if (!empty)
    574 			continue;
    575 		list = isl_pw_aff_list_drop(list, i, 1);
    576 		n--;
    577 	}
    578 
    579 	isl_set_free(domain);
    580 
    581 	return list;
    582 error:
    583 	isl_set_free(domain);
    584 	return isl_pw_aff_list_free(list);
    585 }
    586 
    587 /* Extract a lower bound on dimension "pos" from each constraint
    588  * in "constraints" and return the list of lower bounds.
    589  * If "constraints" has zero elements, then we extract a lower bound
    590  * from "domain" instead.
    591  *
    592  * If the current dimension is strided, then the lower bound
    593  * is adjusted by lower_bound to match the stride information.
    594  * This modification may make one or more lower bounds redundant
    595  * with respect to the other lower bounds.  We therefore check
    596  * for this condition and remove the redundant lower bounds.
    597  */
    598 static __isl_give isl_pw_aff_list *lower_bounds(
    599 	__isl_keep isl_constraint_list *constraints, int pos,
    600 	__isl_keep isl_set *domain, __isl_keep isl_ast_build *build)
    601 {
    602 	isl_ctx *ctx;
    603 	isl_pw_aff_list *list;
    604 	int i;
    605 	isl_size n;
    606 
    607 	if (!build)
    608 		return NULL;
    609 
    610 	n = isl_constraint_list_n_constraint(constraints);
    611 	if (n < 0)
    612 		return NULL;
    613 	if (n == 0) {
    614 		isl_pw_aff *pa;
    615 		pa = exact_bound(domain, build, 0);
    616 		return isl_pw_aff_list_from_pw_aff(pa);
    617 	}
    618 
    619 	ctx = isl_ast_build_get_ctx(build);
    620 	list = isl_pw_aff_list_alloc(ctx,n);
    621 
    622 	for (i = 0; i < n; ++i) {
    623 		isl_aff *aff;
    624 		isl_constraint *c;
    625 
    626 		c = isl_constraint_list_get_constraint(constraints, i);
    627 		aff = lower_bound(c, pos, build);
    628 		isl_constraint_free(c);
    629 		list = isl_pw_aff_list_add(list, isl_pw_aff_from_aff(aff));
    630 	}
    631 
    632 	if (isl_ast_build_has_stride(build, pos))
    633 		list = remove_redundant_lower_bounds(list, build);
    634 
    635 	return list;
    636 }
    637 
    638 /* Extract an upper bound on dimension "pos" from each constraint
    639  * in "constraints" and return the list of upper bounds.
    640  * If "constraints" has zero elements, then we extract an upper bound
    641  * from "domain" instead.
    642  */
    643 static __isl_give isl_pw_aff_list *upper_bounds(
    644 	__isl_keep isl_constraint_list *constraints, int pos,
    645 	__isl_keep isl_set *domain, __isl_keep isl_ast_build *build)
    646 {
    647 	isl_ctx *ctx;
    648 	isl_pw_aff_list *list;
    649 	int i;
    650 	isl_size n;
    651 
    652 	n = isl_constraint_list_n_constraint(constraints);
    653 	if (n < 0)
    654 		return NULL;
    655 	if (n == 0) {
    656 		isl_pw_aff *pa;
    657 		pa = exact_bound(domain, build, 1);
    658 		return isl_pw_aff_list_from_pw_aff(pa);
    659 	}
    660 
    661 	ctx = isl_ast_build_get_ctx(build);
    662 	list = isl_pw_aff_list_alloc(ctx,n);
    663 
    664 	for (i = 0; i < n; ++i) {
    665 		isl_aff *aff;
    666 		isl_constraint *c;
    667 
    668 		c = isl_constraint_list_get_constraint(constraints, i);
    669 		aff = isl_constraint_get_bound(c, isl_dim_set, pos);
    670 		isl_constraint_free(c);
    671 		aff = isl_aff_floor(aff);
    672 		list = isl_pw_aff_list_add(list, isl_pw_aff_from_aff(aff));
    673 	}
    674 
    675 	return list;
    676 }
    677 
    678 /* Return an isl_ast_expr that performs the reduction of type "type"
    679  * on AST expressions corresponding to the elements in "list".
    680  *
    681  * The list is assumed to contain at least one element.
    682  * If the list contains exactly one element, then the returned isl_ast_expr
    683  * simply computes that affine expression.
    684  * If the list contains more than one element, then we sort it
    685  * using a fairly arbitrary but hopefully reasonably stable order.
    686  */
    687 static __isl_give isl_ast_expr *reduce_list(enum isl_ast_expr_op_type type,
    688 	__isl_keep isl_pw_aff_list *list, __isl_keep isl_ast_build *build)
    689 {
    690 	int i;
    691 	isl_size n;
    692 	isl_ctx *ctx;
    693 	isl_ast_expr *expr;
    694 
    695 	n = isl_pw_aff_list_n_pw_aff(list);
    696 	if (n < 0)
    697 		return NULL;
    698 
    699 	if (n == 1)
    700 		return isl_ast_build_expr_from_pw_aff_internal(build,
    701 				isl_pw_aff_list_get_pw_aff(list, 0));
    702 
    703 	ctx = isl_pw_aff_list_get_ctx(list);
    704 	expr = isl_ast_expr_alloc_op(ctx, type, n);
    705 
    706 	list = isl_pw_aff_list_copy(list);
    707 	list = isl_pw_aff_list_sort(list, &reduce_list_cmp, NULL);
    708 	if (!list)
    709 		return isl_ast_expr_free(expr);
    710 
    711 	for (i = 0; i < n; ++i) {
    712 		isl_ast_expr *expr_i;
    713 
    714 		expr_i = isl_ast_build_expr_from_pw_aff_internal(build,
    715 				isl_pw_aff_list_get_pw_aff(list, i));
    716 		expr = isl_ast_expr_op_add_arg(expr, expr_i);
    717 	}
    718 
    719 	isl_pw_aff_list_free(list);
    720 	return expr;
    721 }
    722 
    723 /* Add guards implied by the "generated constraints",
    724  * but not (necessarily) enforced by the generated AST to "guard".
    725  * In particular, if there is any stride constraints,
    726  * then add the guard implied by those constraints.
    727  * If we have generated a degenerate loop, then add the guard
    728  * implied by "bounds" on the outer dimensions, i.e., the guard
    729  * that ensures that the single value actually exists.
    730  * Since there may also be guards implied by a combination
    731  * of these constraints, we first combine them before
    732  * deriving the implied constraints.
    733  */
    734 static __isl_give isl_set *add_implied_guards(__isl_take isl_set *guard,
    735 	int degenerate, __isl_keep isl_basic_set *bounds,
    736 	__isl_keep isl_ast_build *build)
    737 {
    738 	isl_size depth;
    739 	isl_bool has_stride;
    740 	isl_space *space;
    741 	isl_set *dom, *set;
    742 
    743 	depth = isl_ast_build_get_depth(build);
    744 	has_stride = isl_ast_build_has_stride(build, depth);
    745 	if (depth < 0 || has_stride < 0)
    746 		return isl_set_free(guard);
    747 	if (!has_stride && !degenerate)
    748 		return guard;
    749 
    750 	space = isl_basic_set_get_space(bounds);
    751 	dom = isl_set_universe(space);
    752 
    753 	if (degenerate) {
    754 		bounds = isl_basic_set_copy(bounds);
    755 		bounds = isl_basic_set_drop_constraints_not_involving_dims(
    756 					bounds, isl_dim_set, depth, 1);
    757 		set = isl_set_from_basic_set(bounds);
    758 		dom = isl_set_intersect(dom, set);
    759 	}
    760 
    761 	if (has_stride) {
    762 		set = isl_ast_build_get_stride_constraint(build);
    763 		dom = isl_set_intersect(dom, set);
    764 	}
    765 
    766 	dom = isl_set_eliminate(dom, isl_dim_set, depth, 1);
    767 	dom = isl_ast_build_compute_gist(build, dom);
    768 	guard = isl_set_intersect(guard, dom);
    769 
    770 	return guard;
    771 }
    772 
    773 /* Update "graft" based on "sub_build" for the degenerate case.
    774  *
    775  * "build" is the build in which graft->node was created
    776  * "sub_build" contains information about the current level itself,
    777  * including the single value attained.
    778  *
    779  * We set the initialization part of the for loop to the single
    780  * value attained by the current dimension.
    781  * The increment and condition are not strictly needed as they are known
    782  * to be "1" and "iterator <= value" respectively.
    783  */
    784 static __isl_give isl_ast_graft *refine_degenerate(
    785 	__isl_take isl_ast_graft *graft, __isl_keep isl_ast_build *build,
    786 	__isl_keep isl_ast_build *sub_build)
    787 {
    788 	isl_pw_aff *value;
    789 	isl_ast_expr *init;
    790 
    791 	if (!graft || !sub_build)
    792 		return isl_ast_graft_free(graft);
    793 
    794 	value = isl_pw_aff_copy(sub_build->value);
    795 
    796 	init = isl_ast_build_expr_from_pw_aff_internal(build, value);
    797 	graft->node = isl_ast_node_for_set_init(graft->node, init);
    798 	if (!graft->node)
    799 		return isl_ast_graft_free(graft);
    800 
    801 	return graft;
    802 }
    803 
    804 /* Return the intersection of constraints in "list" as a set.
    805  */
    806 static __isl_give isl_set *intersect_constraints(
    807 	__isl_keep isl_constraint_list *list)
    808 {
    809 	int i;
    810 	isl_size n;
    811 	isl_basic_set *bset;
    812 
    813 	n = isl_constraint_list_n_constraint(list);
    814 	if (n < 0)
    815 		return NULL;
    816 	if (n < 1)
    817 		isl_die(isl_constraint_list_get_ctx(list), isl_error_internal,
    818 			"expecting at least one constraint", return NULL);
    819 
    820 	bset = isl_basic_set_from_constraint(
    821 				isl_constraint_list_get_constraint(list, 0));
    822 	for (i = 1; i < n; ++i) {
    823 		isl_basic_set *bset_i;
    824 
    825 		bset_i = isl_basic_set_from_constraint(
    826 				isl_constraint_list_get_constraint(list, i));
    827 		bset = isl_basic_set_intersect(bset, bset_i);
    828 	}
    829 
    830 	return isl_set_from_basic_set(bset);
    831 }
    832 
    833 /* Compute the constraints on the outer dimensions enforced by
    834  * graft->node and add those constraints to graft->enforced,
    835  * in case the upper bound is expressed as a set "upper".
    836  *
    837  * In particular, if l(...) is a lower bound in "lower", and
    838  *
    839  *	-a i + f(...) >= 0		or	a i <= f(...)
    840  *
    841  * is an upper bound ocnstraint on the current dimension i,
    842  * then the for loop enforces the constraint
    843  *
    844  *	-a l(...) + f(...) >= 0		or	a l(...) <= f(...)
    845  *
    846  * We therefore simply take each lower bound in turn, plug it into
    847  * the upper bounds and compute the intersection over all lower bounds.
    848  *
    849  * If a lower bound is a rational expression, then
    850  * isl_basic_set_preimage_multi_aff will force this rational
    851  * expression to have only integer values.  However, the loop
    852  * itself does not enforce this integrality constraint.  We therefore
    853  * use the ceil of the lower bounds instead of the lower bounds themselves.
    854  * Other constraints will make sure that the for loop is only executed
    855  * when each of the lower bounds attains an integral value.
    856  * In particular, potentially rational values only occur in
    857  * lower_bound if the offset is a (seemingly) rational expression,
    858  * but then outer conditions will make sure that this rational expression
    859  * only attains integer values.
    860  */
    861 static __isl_give isl_ast_graft *set_enforced_from_set(
    862 	__isl_take isl_ast_graft *graft,
    863 	__isl_keep isl_pw_aff_list *lower, int pos, __isl_keep isl_set *upper)
    864 {
    865 	isl_space *space;
    866 	isl_basic_set *enforced;
    867 	isl_pw_multi_aff *pma;
    868 	int i;
    869 	isl_size n;
    870 
    871 	n = isl_pw_aff_list_n_pw_aff(lower);
    872 	if (!graft || n < 0)
    873 		return isl_ast_graft_free(graft);
    874 
    875 	space = isl_set_get_space(upper);
    876 	enforced = isl_basic_set_universe(isl_space_copy(space));
    877 
    878 	space = isl_space_map_from_set(space);
    879 	pma = isl_pw_multi_aff_identity(space);
    880 
    881 	for (i = 0; i < n; ++i) {
    882 		isl_pw_aff *pa;
    883 		isl_set *enforced_i;
    884 		isl_basic_set *hull;
    885 		isl_pw_multi_aff *pma_i;
    886 
    887 		pa = isl_pw_aff_list_get_pw_aff(lower, i);
    888 		pa = isl_pw_aff_ceil(pa);
    889 		pma_i = isl_pw_multi_aff_copy(pma);
    890 		pma_i = isl_pw_multi_aff_set_pw_aff(pma_i, pos, pa);
    891 		enforced_i = isl_set_copy(upper);
    892 		enforced_i = isl_set_preimage_pw_multi_aff(enforced_i, pma_i);
    893 		hull = isl_set_simple_hull(enforced_i);
    894 		enforced = isl_basic_set_intersect(enforced, hull);
    895 	}
    896 
    897 	isl_pw_multi_aff_free(pma);
    898 
    899 	graft = isl_ast_graft_enforce(graft, enforced);
    900 
    901 	return graft;
    902 }
    903 
    904 /* Compute the constraints on the outer dimensions enforced by
    905  * graft->node and add those constraints to graft->enforced,
    906  * in case the upper bound is expressed as
    907  * a list of affine expressions "upper".
    908  *
    909  * The enforced condition is that each lower bound expression is less
    910  * than or equal to each upper bound expression.
    911  */
    912 static __isl_give isl_ast_graft *set_enforced_from_list(
    913 	__isl_take isl_ast_graft *graft,
    914 	__isl_keep isl_pw_aff_list *lower, __isl_keep isl_pw_aff_list *upper)
    915 {
    916 	isl_set *cond;
    917 	isl_basic_set *enforced;
    918 
    919 	lower = isl_pw_aff_list_copy(lower);
    920 	upper = isl_pw_aff_list_copy(upper);
    921 	cond = isl_pw_aff_list_le_set(lower, upper);
    922 	enforced = isl_set_simple_hull(cond);
    923 	graft = isl_ast_graft_enforce(graft, enforced);
    924 
    925 	return graft;
    926 }
    927 
    928 /* Does "aff" have a negative constant term?
    929  */
    930 static isl_bool aff_constant_is_negative(__isl_keep isl_set *set,
    931 	__isl_keep isl_aff *aff, void *user)
    932 {
    933 	isl_bool is_neg;
    934 	isl_val *v;
    935 
    936 	v = isl_aff_get_constant_val(aff);
    937 	is_neg = isl_val_is_neg(v);
    938 	isl_val_free(v);
    939 
    940 	return is_neg;
    941 }
    942 
    943 /* Does "pa" have a negative constant term over its entire domain?
    944  */
    945 static isl_bool pw_aff_constant_is_negative(__isl_keep isl_pw_aff *pa,
    946 	void *user)
    947 {
    948 	return isl_pw_aff_every_piece(pa, &aff_constant_is_negative, NULL);
    949 }
    950 
    951 /* Does each element in "list" have a negative constant term?
    952  */
    953 static int list_constant_is_negative(__isl_keep isl_pw_aff_list *list)
    954 {
    955 	return isl_pw_aff_list_every(list, &pw_aff_constant_is_negative, NULL);
    956 }
    957 
    958 /* Add 1 to each of the elements in "list", where each of these elements
    959  * is defined over the internal schedule space of "build".
    960  */
    961 static __isl_give isl_pw_aff_list *list_add_one(
    962 	__isl_take isl_pw_aff_list *list, __isl_keep isl_ast_build *build)
    963 {
    964 	int i;
    965 	isl_size n;
    966 	isl_space *space;
    967 	isl_aff *aff;
    968 	isl_pw_aff *one;
    969 
    970 	n = isl_pw_aff_list_n_pw_aff(list);
    971 	if (n < 0)
    972 		return isl_pw_aff_list_free(list);
    973 
    974 	space = isl_ast_build_get_space(build, 1);
    975 	aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
    976 	aff = isl_aff_add_constant_si(aff, 1);
    977 	one = isl_pw_aff_from_aff(aff);
    978 
    979 	for (i = 0; i < n; ++i) {
    980 		isl_pw_aff *pa;
    981 		pa = isl_pw_aff_list_get_pw_aff(list, i);
    982 		pa = isl_pw_aff_add(pa, isl_pw_aff_copy(one));
    983 		list = isl_pw_aff_list_set_pw_aff(list, i, pa);
    984 	}
    985 
    986 	isl_pw_aff_free(one);
    987 
    988 	return list;
    989 }
    990 
    991 /* Set the condition part of the for node graft->node in case
    992  * the upper bound is represented as a list of piecewise affine expressions.
    993  *
    994  * In particular, set the condition to
    995  *
    996  *	iterator <= min(list of upper bounds)
    997  *
    998  * If each of the upper bounds has a negative constant term, then
    999  * set the condition to
   1000  *
   1001  *	iterator < min(list of (upper bound + 1)s)
   1002  *
   1003  */
   1004 static __isl_give isl_ast_graft *set_for_cond_from_list(
   1005 	__isl_take isl_ast_graft *graft, __isl_keep isl_pw_aff_list *list,
   1006 	__isl_keep isl_ast_build *build)
   1007 {
   1008 	int neg;
   1009 	isl_ast_expr *bound, *iterator, *cond;
   1010 	enum isl_ast_expr_op_type type = isl_ast_expr_op_le;
   1011 
   1012 	if (!graft || !list)
   1013 		return isl_ast_graft_free(graft);
   1014 
   1015 	neg = list_constant_is_negative(list);
   1016 	if (neg < 0)
   1017 		return isl_ast_graft_free(graft);
   1018 	list = isl_pw_aff_list_copy(list);
   1019 	if (neg) {
   1020 		list = list_add_one(list, build);
   1021 		type = isl_ast_expr_op_lt;
   1022 	}
   1023 
   1024 	bound = reduce_list(isl_ast_expr_op_min, list, build);
   1025 	iterator = isl_ast_expr_copy(graft->node->u.f.iterator);
   1026 	cond = isl_ast_expr_alloc_binary(type, iterator, bound);
   1027 	graft->node = isl_ast_node_for_set_cond(graft->node, cond);
   1028 
   1029 	isl_pw_aff_list_free(list);
   1030 	if (!graft->node)
   1031 		return isl_ast_graft_free(graft);
   1032 	return graft;
   1033 }
   1034 
   1035 /* Set the condition part of the for node graft->node in case
   1036  * the upper bound is represented as a set.
   1037  */
   1038 static __isl_give isl_ast_graft *set_for_cond_from_set(
   1039 	__isl_take isl_ast_graft *graft, __isl_keep isl_set *set,
   1040 	__isl_keep isl_ast_build *build)
   1041 {
   1042 	isl_ast_expr *cond;
   1043 
   1044 	if (!graft)
   1045 		return NULL;
   1046 
   1047 	cond = isl_ast_build_expr_from_set_internal(build, isl_set_copy(set));
   1048 	graft->node = isl_ast_node_for_set_cond(graft->node, cond);
   1049 	if (!graft->node)
   1050 		return isl_ast_graft_free(graft);
   1051 	return graft;
   1052 }
   1053 
   1054 /* Construct an isl_ast_expr for the increment (i.e., stride) of
   1055  * the current dimension.
   1056  */
   1057 static __isl_give isl_ast_expr *for_inc(__isl_keep isl_ast_build *build)
   1058 {
   1059 	isl_size depth;
   1060 	isl_val *v;
   1061 	isl_ctx *ctx;
   1062 
   1063 	depth = isl_ast_build_get_depth(build);
   1064 	if (depth < 0)
   1065 		return NULL;
   1066 	ctx = isl_ast_build_get_ctx(build);
   1067 
   1068 	if (!isl_ast_build_has_stride(build, depth))
   1069 		return isl_ast_expr_alloc_int_si(ctx, 1);
   1070 
   1071 	v = isl_ast_build_get_stride(build, depth);
   1072 	return isl_ast_expr_from_val(v);
   1073 }
   1074 
   1075 /* Should we express the loop condition as
   1076  *
   1077  *	iterator <= min(list of upper bounds)
   1078  *
   1079  * or as a conjunction of constraints?
   1080  *
   1081  * The first is constructed from a list of upper bounds.
   1082  * The second is constructed from a set.
   1083  *
   1084  * If there are no upper bounds in "constraints", then this could mean
   1085  * that "domain" simply doesn't have an upper bound or that we didn't
   1086  * pick any upper bound.  In the first case, we want to generate the
   1087  * loop condition as a(n empty) conjunction of constraints
   1088  * In the second case, we will compute
   1089  * a single upper bound from "domain" and so we use the list form.
   1090  *
   1091  * If there are upper bounds in "constraints",
   1092  * then we use the list form iff the atomic_upper_bound option is set.
   1093  */
   1094 static int use_upper_bound_list(isl_ctx *ctx, int n_upper,
   1095 	__isl_keep isl_set *domain, int depth)
   1096 {
   1097 	if (n_upper > 0)
   1098 		return isl_options_get_ast_build_atomic_upper_bound(ctx);
   1099 	else
   1100 		return isl_set_dim_has_upper_bound(domain, isl_dim_set, depth);
   1101 }
   1102 
   1103 /* Fill in the expressions of the for node in graft->node.
   1104  *
   1105  * In particular,
   1106  * - set the initialization part of the loop to the maximum of the lower bounds
   1107  * - extract the increment from the stride of the current dimension
   1108  * - construct the for condition either based on a list of upper bounds
   1109  *	or on a set of upper bound constraints.
   1110  */
   1111 static __isl_give isl_ast_graft *set_for_node_expressions(
   1112 	__isl_take isl_ast_graft *graft, __isl_keep isl_pw_aff_list *lower,
   1113 	int use_list, __isl_keep isl_pw_aff_list *upper_list,
   1114 	__isl_keep isl_set *upper_set, __isl_keep isl_ast_build *build)
   1115 {
   1116 	isl_ast_expr *init;
   1117 
   1118 	if (!graft)
   1119 		return NULL;
   1120 
   1121 	init = reduce_list(isl_ast_expr_op_max, lower, build);
   1122 	graft->node = isl_ast_node_for_set_init(graft->node, init);
   1123 	graft->node = isl_ast_node_for_set_inc(graft->node, for_inc(build));
   1124 
   1125 	if (!graft->node)
   1126 		graft = isl_ast_graft_free(graft);
   1127 
   1128 	if (use_list)
   1129 		graft = set_for_cond_from_list(graft, upper_list, build);
   1130 	else
   1131 		graft = set_for_cond_from_set(graft, upper_set, build);
   1132 
   1133 	return graft;
   1134 }
   1135 
   1136 /* Update "graft" based on "bounds" and "domain" for the generic,
   1137  * non-degenerate, case.
   1138  *
   1139  * "c_lower" and "c_upper" contain the lower and upper bounds
   1140  * that the loop node should express.
   1141  * "domain" is the subset of the intersection of the constraints
   1142  * for which some code is executed.
   1143  *
   1144  * There may be zero lower bounds or zero upper bounds in "constraints"
   1145  * in case the list of constraints was created
   1146  * based on the atomic option or based on separation with explicit bounds.
   1147  * In that case, we use "domain" to derive lower and/or upper bounds.
   1148  *
   1149  * We first compute a list of one or more lower bounds.
   1150  *
   1151  * Then we decide if we want to express the condition as
   1152  *
   1153  *	iterator <= min(list of upper bounds)
   1154  *
   1155  * or as a conjunction of constraints.
   1156  *
   1157  * The set of enforced constraints is then computed either based on
   1158  * a list of upper bounds or on a set of upper bound constraints.
   1159  * We do not compute any enforced constraints if we were forced
   1160  * to compute a lower or upper bound using exact_bound.  The domains
   1161  * of the resulting expressions may imply some bounds on outer dimensions
   1162  * that we do not want to appear in the enforced constraints since
   1163  * they are not actually enforced by the corresponding code.
   1164  *
   1165  * Finally, we fill in the expressions of the for node.
   1166  */
   1167 static __isl_give isl_ast_graft *refine_generic_bounds(
   1168 	__isl_take isl_ast_graft *graft,
   1169 	__isl_take isl_constraint_list *c_lower,
   1170 	__isl_take isl_constraint_list *c_upper,
   1171 	__isl_keep isl_set *domain, __isl_keep isl_ast_build *build)
   1172 {
   1173 	isl_size depth;
   1174 	isl_ctx *ctx;
   1175 	isl_pw_aff_list *lower;
   1176 	int use_list;
   1177 	isl_set *upper_set = NULL;
   1178 	isl_pw_aff_list *upper_list = NULL;
   1179 	isl_size n_lower, n_upper;
   1180 
   1181 	depth = isl_ast_build_get_depth(build);
   1182 	if (!graft || !c_lower || !c_upper || depth < 0)
   1183 		goto error;
   1184 
   1185 	ctx = isl_ast_graft_get_ctx(graft);
   1186 
   1187 	n_lower = isl_constraint_list_n_constraint(c_lower);
   1188 	n_upper = isl_constraint_list_n_constraint(c_upper);
   1189 	if (n_lower < 0 || n_upper < 0)
   1190 		goto error;
   1191 
   1192 	use_list = use_upper_bound_list(ctx, n_upper, domain, depth);
   1193 
   1194 	lower = lower_bounds(c_lower, depth, domain, build);
   1195 
   1196 	if (use_list)
   1197 		upper_list = upper_bounds(c_upper, depth, domain, build);
   1198 	else if (n_upper > 0)
   1199 		upper_set = intersect_constraints(c_upper);
   1200 	else
   1201 		upper_set = isl_set_universe(isl_set_get_space(domain));
   1202 
   1203 	if (n_lower == 0 || n_upper == 0)
   1204 		;
   1205 	else if (use_list)
   1206 		graft = set_enforced_from_list(graft, lower, upper_list);
   1207 	else
   1208 		graft = set_enforced_from_set(graft, lower, depth, upper_set);
   1209 
   1210 	graft = set_for_node_expressions(graft, lower, use_list, upper_list,
   1211 					upper_set, build);
   1212 
   1213 	isl_pw_aff_list_free(lower);
   1214 	isl_pw_aff_list_free(upper_list);
   1215 	isl_set_free(upper_set);
   1216 	isl_constraint_list_free(c_lower);
   1217 	isl_constraint_list_free(c_upper);
   1218 
   1219 	return graft;
   1220 error:
   1221 	isl_constraint_list_free(c_lower);
   1222 	isl_constraint_list_free(c_upper);
   1223 	return isl_ast_graft_free(graft);
   1224 }
   1225 
   1226 /* Internal data structure used inside count_constraints to keep
   1227  * track of the number of constraints that are independent of dimension "pos",
   1228  * the lower bounds in "pos" and the upper bounds in "pos".
   1229  */
   1230 struct isl_ast_count_constraints_data {
   1231 	int pos;
   1232 
   1233 	int n_indep;
   1234 	int n_lower;
   1235 	int n_upper;
   1236 };
   1237 
   1238 /* Increment data->n_indep, data->lower or data->upper depending
   1239  * on whether "c" is independent of dimensions data->pos,
   1240  * a lower bound or an upper bound.
   1241  */
   1242 static isl_stat count_constraints(__isl_take isl_constraint *c, void *user)
   1243 {
   1244 	struct isl_ast_count_constraints_data *data = user;
   1245 
   1246 	if (isl_constraint_is_lower_bound(c, isl_dim_set, data->pos))
   1247 		data->n_lower++;
   1248 	else if (isl_constraint_is_upper_bound(c, isl_dim_set, data->pos))
   1249 		data->n_upper++;
   1250 	else
   1251 		data->n_indep++;
   1252 
   1253 	isl_constraint_free(c);
   1254 
   1255 	return isl_stat_ok;
   1256 }
   1257 
   1258 /* Update "graft" based on "bounds" and "domain" for the generic,
   1259  * non-degenerate, case.
   1260  *
   1261  * "list" respresent the list of bounds that need to be encoded by
   1262  * the for loop.  Only the constraints that involve the iterator
   1263  * are relevant here.  The other constraints are taken care of by
   1264  * the caller and are included in the generated constraints of "build".
   1265  * "domain" is the subset of the intersection of the constraints
   1266  * for which some code is executed.
   1267  * "build" is the build in which graft->node was created.
   1268  *
   1269  * We separate lower bounds, upper bounds and constraints that
   1270  * are independent of the loop iterator.
   1271  *
   1272  * The actual for loop bounds are generated in refine_generic_bounds.
   1273  */
   1274 static __isl_give isl_ast_graft *refine_generic_split(
   1275 	__isl_take isl_ast_graft *graft, __isl_take isl_constraint_list *list,
   1276 	__isl_keep isl_set *domain, __isl_keep isl_ast_build *build)
   1277 {
   1278 	struct isl_ast_count_constraints_data data;
   1279 	isl_size depth;
   1280 	isl_constraint_list *lower;
   1281 	isl_constraint_list *upper;
   1282 
   1283 	depth = isl_ast_build_get_depth(build);
   1284 	if (depth < 0)
   1285 		list = isl_constraint_list_free(list);
   1286 	if (!list)
   1287 		return isl_ast_graft_free(graft);
   1288 
   1289 	data.pos = depth;
   1290 
   1291 	list = isl_constraint_list_sort(list, &cmp_constraint, &data.pos);
   1292 	if (!list)
   1293 		return isl_ast_graft_free(graft);
   1294 
   1295 	data.n_indep = data.n_lower = data.n_upper = 0;
   1296 	if (isl_constraint_list_foreach(list, &count_constraints, &data) < 0) {
   1297 		isl_constraint_list_free(list);
   1298 		return isl_ast_graft_free(graft);
   1299 	}
   1300 
   1301 	lower = isl_constraint_list_drop(list, 0, data.n_indep);
   1302 	upper = isl_constraint_list_copy(lower);
   1303 	lower = isl_constraint_list_drop(lower, data.n_lower, data.n_upper);
   1304 	upper = isl_constraint_list_drop(upper, 0, data.n_lower);
   1305 
   1306 	return refine_generic_bounds(graft, lower, upper, domain, build);
   1307 }
   1308 
   1309 /* Update "graft" based on "bounds" and "domain" for the generic,
   1310  * non-degenerate, case.
   1311  *
   1312  * "bounds" respresent the bounds that need to be encoded by
   1313  * the for loop (or a guard around the for loop).
   1314  * "domain" is the subset of "bounds" for which some code is executed.
   1315  * "build" is the build in which graft->node was created.
   1316  *
   1317  * We break up "bounds" into a list of constraints and continue with
   1318  * refine_generic_split.
   1319  */
   1320 static __isl_give isl_ast_graft *refine_generic(
   1321 	__isl_take isl_ast_graft *graft,
   1322 	__isl_keep isl_basic_set *bounds, __isl_keep isl_set *domain,
   1323 	__isl_keep isl_ast_build *build)
   1324 {
   1325 	isl_constraint_list *list;
   1326 
   1327 	if (!build || !graft)
   1328 		return isl_ast_graft_free(graft);
   1329 
   1330 	list = isl_basic_set_get_constraint_list(bounds);
   1331 
   1332 	graft = refine_generic_split(graft, list, domain, build);
   1333 
   1334 	return graft;
   1335 }
   1336 
   1337 /* Create a for node for the current level.
   1338  *
   1339  * Mark the for node degenerate if "degenerate" is set.
   1340  */
   1341 static __isl_give isl_ast_node *create_for(__isl_keep isl_ast_build *build,
   1342 	int degenerate)
   1343 {
   1344 	isl_size depth;
   1345 	isl_id *id;
   1346 	isl_ast_node *node;
   1347 
   1348 	depth = isl_ast_build_get_depth(build);
   1349 	if (depth < 0)
   1350 		return NULL;
   1351 
   1352 	id = isl_ast_build_get_iterator_id(build, depth);
   1353 	node = isl_ast_node_alloc_for(id);
   1354 	if (degenerate)
   1355 		node = isl_ast_node_for_mark_degenerate(node);
   1356 
   1357 	return node;
   1358 }
   1359 
   1360 /* If the ast_build_exploit_nested_bounds option is set, then return
   1361  * the constraints enforced by all elements in "list".
   1362  * Otherwise, return the universe.
   1363  */
   1364 static __isl_give isl_basic_set *extract_shared_enforced(
   1365 	__isl_keep isl_ast_graft_list *list, __isl_keep isl_ast_build *build)
   1366 {
   1367 	isl_ctx *ctx;
   1368 	isl_space *space;
   1369 
   1370 	if (!list)
   1371 		return NULL;
   1372 
   1373 	ctx = isl_ast_graft_list_get_ctx(list);
   1374 	if (isl_options_get_ast_build_exploit_nested_bounds(ctx))
   1375 		return isl_ast_graft_list_extract_shared_enforced(list, build);
   1376 
   1377 	space = isl_ast_build_get_space(build, 1);
   1378 	return isl_basic_set_universe(space);
   1379 }
   1380 
   1381 /* Return the pending constraints of "build" that are not already taken
   1382  * care of (by a combination of "enforced" and the generated constraints
   1383  * of "build").
   1384  */
   1385 static __isl_give isl_set *extract_pending(__isl_keep isl_ast_build *build,
   1386 	__isl_keep isl_basic_set *enforced)
   1387 {
   1388 	isl_set *guard, *context;
   1389 
   1390 	guard = isl_ast_build_get_pending(build);
   1391 	context = isl_set_from_basic_set(isl_basic_set_copy(enforced));
   1392 	context = isl_set_intersect(context,
   1393 					isl_ast_build_get_generated(build));
   1394 	return isl_set_gist(guard, context);
   1395 }
   1396 
   1397 /* Create an AST node for the current dimension based on
   1398  * the schedule domain "bounds" and return the node encapsulated
   1399  * in an isl_ast_graft.
   1400  *
   1401  * "executed" is the current inverse schedule, taking into account
   1402  * the bounds in "bounds"
   1403  * "domain" is the domain of "executed", with inner dimensions projected out.
   1404  * It may be a strict subset of "bounds" in case "bounds" was created
   1405  * based on the atomic option or based on separation with explicit bounds.
   1406  *
   1407  * "domain" may satisfy additional equalities that result
   1408  * from intersecting "executed" with "bounds" in add_node.
   1409  * It may also satisfy some global constraints that were dropped out because
   1410  * we performed separation with explicit bounds.
   1411  * The very first step is then to copy these constraints to "bounds".
   1412  *
   1413  * Since we may be calling before_each_for and after_each_for
   1414  * callbacks, we record the current inverse schedule in the build.
   1415  *
   1416  * We consider three builds,
   1417  * "build" is the one in which the current level is created,
   1418  * "body_build" is the build in which the next level is created,
   1419  * "sub_build" is essentially the same as "body_build", except that
   1420  * the depth has not been increased yet.
   1421  *
   1422  * "build" already contains information (in strides and offsets)
   1423  * about the strides at the current level, but this information is not
   1424  * reflected in the build->domain.
   1425  * We first add this information and the "bounds" to the sub_build->domain.
   1426  * isl_ast_build_set_loop_bounds adds the stride information and
   1427  * checks whether the current dimension attains
   1428  * only a single value and whether this single value can be represented using
   1429  * a single affine expression.
   1430  * In the first case, the current level is considered "degenerate".
   1431  * In the second, sub-case, the current level is considered "eliminated".
   1432  * Eliminated levels don't need to be reflected in the AST since we can
   1433  * simply plug in the affine expression.  For degenerate, but non-eliminated,
   1434  * levels, we do introduce a for node, but mark is as degenerate so that
   1435  * it can be printed as an assignment of the single value to the loop
   1436  * "iterator".
   1437  *
   1438  * If the current level is eliminated, we explicitly plug in the value
   1439  * for the current level found by isl_ast_build_set_loop_bounds in the
   1440  * inverse schedule.  This ensures that if we are working on a slice
   1441  * of the domain based on information available in the inverse schedule
   1442  * and the build domain, that then this information is also reflected
   1443  * in the inverse schedule.  This operation also eliminates the current
   1444  * dimension from the inverse schedule making sure no inner dimensions depend
   1445  * on the current dimension.  Otherwise, we create a for node, marking
   1446  * it degenerate if appropriate.  The initial for node is still incomplete
   1447  * and will be completed in either refine_degenerate or refine_generic.
   1448  *
   1449  * We then generate a sequence of grafts for the next level,
   1450  * create a surrounding graft for the current level and insert
   1451  * the for node we created (if the current level is not eliminated).
   1452  * Before creating a graft for the current level, we first extract
   1453  * hoistable constraints from the child guards and combine them
   1454  * with the pending constraints in the build.  These constraints
   1455  * are used to simplify the child guards and then added to the guard
   1456  * of the current graft to ensure that they will be generated.
   1457  * If the hoisted guard is a disjunction, then we use it directly
   1458  * to gist the guards on the children before intersect it with the
   1459  * pending constraints.  We do so because this disjunction is typically
   1460  * identical to the guards on the children such that these guards
   1461  * can be effectively removed completely.  After the intersection,
   1462  * the gist operation would have a harder time figuring this out.
   1463  *
   1464  * Finally, we set the bounds of the for loop in either
   1465  * refine_degenerate or refine_generic.
   1466  * We do so in a context where the pending constraints of the build
   1467  * have been replaced by the guard of the current graft.
   1468  */
   1469 static __isl_give isl_ast_graft *create_node_scaled(
   1470 	__isl_take isl_union_map *executed,
   1471 	__isl_take isl_basic_set *bounds, __isl_take isl_set *domain,
   1472 	__isl_take isl_ast_build *build)
   1473 {
   1474 	isl_size depth;
   1475 	int degenerate;
   1476 	isl_bool eliminated;
   1477 	isl_size n;
   1478 	isl_basic_set *hull;
   1479 	isl_basic_set *enforced;
   1480 	isl_set *guard, *hoisted;
   1481 	isl_ast_node *node = NULL;
   1482 	isl_ast_graft *graft;
   1483 	isl_ast_graft_list *children;
   1484 	isl_ast_build *sub_build;
   1485 	isl_ast_build *body_build;
   1486 
   1487 	domain = isl_ast_build_eliminate_divs(build, domain);
   1488 	domain = isl_set_detect_equalities(domain);
   1489 	hull = isl_set_unshifted_simple_hull(isl_set_copy(domain));
   1490 	bounds = isl_basic_set_intersect(bounds, hull);
   1491 	build = isl_ast_build_set_executed(build, isl_union_map_copy(executed));
   1492 
   1493 	depth = isl_ast_build_get_depth(build);
   1494 	if (depth < 0)
   1495 		build = isl_ast_build_free(build);
   1496 	sub_build = isl_ast_build_copy(build);
   1497 	bounds = isl_basic_set_remove_redundancies(bounds);
   1498 	bounds = isl_ast_build_specialize_basic_set(sub_build, bounds);
   1499 	sub_build = isl_ast_build_set_loop_bounds(sub_build,
   1500 						isl_basic_set_copy(bounds));
   1501 	degenerate = isl_ast_build_has_value(sub_build);
   1502 	eliminated = isl_ast_build_has_affine_value(sub_build, depth);
   1503 	if (degenerate < 0 || eliminated < 0)
   1504 		executed = isl_union_map_free(executed);
   1505 	if (!degenerate)
   1506 		bounds = isl_ast_build_compute_gist_basic_set(build, bounds);
   1507 	sub_build = isl_ast_build_set_pending_generated(sub_build,
   1508 						isl_basic_set_copy(bounds));
   1509 	if (eliminated)
   1510 		executed = plug_in_values(executed, sub_build);
   1511 	else
   1512 		node = create_for(build, degenerate);
   1513 
   1514 	body_build = isl_ast_build_copy(sub_build);
   1515 	body_build = isl_ast_build_increase_depth(body_build);
   1516 	if (!eliminated)
   1517 		node = before_each_for(node, body_build);
   1518 	children = generate_next_level(executed,
   1519 				    isl_ast_build_copy(body_build));
   1520 
   1521 	enforced = extract_shared_enforced(children, build);
   1522 	guard = extract_pending(sub_build, enforced);
   1523 	hoisted = isl_ast_graft_list_extract_hoistable_guard(children, build);
   1524 	n = isl_set_n_basic_set(hoisted);
   1525 	if (n < 0)
   1526 		children = isl_ast_graft_list_free(children);
   1527 	if (n > 1)
   1528 		children = isl_ast_graft_list_gist_guards(children,
   1529 						    isl_set_copy(hoisted));
   1530 	guard = isl_set_intersect(guard, hoisted);
   1531 	if (!eliminated)
   1532 		guard = add_implied_guards(guard, degenerate, bounds, build);
   1533 
   1534 	graft = isl_ast_graft_alloc_from_children(children,
   1535 			    isl_set_copy(guard), enforced, build, sub_build);
   1536 
   1537 	if (!eliminated) {
   1538 		isl_ast_build *for_build;
   1539 
   1540 		graft = isl_ast_graft_insert_for(graft, node);
   1541 		for_build = isl_ast_build_copy(build);
   1542 		for_build = isl_ast_build_replace_pending_by_guard(for_build,
   1543 							isl_set_copy(guard));
   1544 		if (degenerate)
   1545 			graft = refine_degenerate(graft, for_build, sub_build);
   1546 		else
   1547 			graft = refine_generic(graft, bounds,
   1548 					domain, for_build);
   1549 		isl_ast_build_free(for_build);
   1550 	}
   1551 	isl_set_free(guard);
   1552 	if (!eliminated)
   1553 		graft = after_each_for(graft, body_build);
   1554 
   1555 	isl_ast_build_free(body_build);
   1556 	isl_ast_build_free(sub_build);
   1557 	isl_ast_build_free(build);
   1558 	isl_basic_set_free(bounds);
   1559 	isl_set_free(domain);
   1560 
   1561 	return graft;
   1562 }
   1563 
   1564 /* Internal data structure for checking if all constraints involving
   1565  * the input dimension "depth" are such that the other coefficients
   1566  * are multiples of "m", reducing "m" if they are not.
   1567  * If "m" is reduced all the way down to "1", then the check has failed
   1568  * and we break out of the iteration.
   1569  */
   1570 struct isl_check_scaled_data {
   1571 	int depth;
   1572 	isl_val *m;
   1573 };
   1574 
   1575 /* If constraint "c" involves the input dimension data->depth,
   1576  * then make sure that all the other coefficients are multiples of data->m,
   1577  * reducing data->m if needed.
   1578  * Break out of the iteration if data->m has become equal to "1".
   1579  */
   1580 static isl_stat constraint_check_scaled(__isl_take isl_constraint *c,
   1581 	void *user)
   1582 {
   1583 	struct isl_check_scaled_data *data = user;
   1584 	int i, j;
   1585 	isl_size n;
   1586 	enum isl_dim_type t[] = { isl_dim_param, isl_dim_in, isl_dim_out,
   1587 				    isl_dim_div };
   1588 
   1589 	if (!isl_constraint_involves_dims(c, isl_dim_in, data->depth, 1)) {
   1590 		isl_constraint_free(c);
   1591 		return isl_stat_ok;
   1592 	}
   1593 
   1594 	for (i = 0; i < 4; ++i) {
   1595 		n = isl_constraint_dim(c, t[i]);
   1596 		if (n < 0)
   1597 			break;
   1598 		for (j = 0; j < n; ++j) {
   1599 			isl_val *d;
   1600 
   1601 			if (t[i] == isl_dim_in && j == data->depth)
   1602 				continue;
   1603 			if (!isl_constraint_involves_dims(c, t[i], j, 1))
   1604 				continue;
   1605 			d = isl_constraint_get_coefficient_val(c, t[i], j);
   1606 			data->m = isl_val_gcd(data->m, d);
   1607 			if (isl_val_is_one(data->m))
   1608 				break;
   1609 		}
   1610 		if (j < n)
   1611 			break;
   1612 	}
   1613 
   1614 	isl_constraint_free(c);
   1615 
   1616 	return i < 4 ? isl_stat_error : isl_stat_ok;
   1617 }
   1618 
   1619 /* For each constraint of "bmap" that involves the input dimension data->depth,
   1620  * make sure that all the other coefficients are multiples of data->m,
   1621  * reducing data->m if needed.
   1622  * Break out of the iteration if data->m has become equal to "1".
   1623  */
   1624 static isl_stat basic_map_check_scaled(__isl_take isl_basic_map *bmap,
   1625 	void *user)
   1626 {
   1627 	isl_stat r;
   1628 
   1629 	r = isl_basic_map_foreach_constraint(bmap,
   1630 						&constraint_check_scaled, user);
   1631 	isl_basic_map_free(bmap);
   1632 
   1633 	return r;
   1634 }
   1635 
   1636 /* For each constraint of "map" that involves the input dimension data->depth,
   1637  * make sure that all the other coefficients are multiples of data->m,
   1638  * reducing data->m if needed.
   1639  * Break out of the iteration if data->m has become equal to "1".
   1640  */
   1641 static isl_stat map_check_scaled(__isl_take isl_map *map, void *user)
   1642 {
   1643 	isl_stat r;
   1644 
   1645 	r = isl_map_foreach_basic_map(map, &basic_map_check_scaled, user);
   1646 	isl_map_free(map);
   1647 
   1648 	return r;
   1649 }
   1650 
   1651 /* Create an AST node for the current dimension based on
   1652  * the schedule domain "bounds" and return the node encapsulated
   1653  * in an isl_ast_graft.
   1654  *
   1655  * "executed" is the current inverse schedule, taking into account
   1656  * the bounds in "bounds"
   1657  * "domain" is the domain of "executed", with inner dimensions projected out.
   1658  *
   1659  *
   1660  * Before moving on to the actual AST node construction in create_node_scaled,
   1661  * we first check if the current dimension is strided and if we can scale
   1662  * down this stride.  Note that we only do this if the ast_build_scale_strides
   1663  * option is set.
   1664  *
   1665  * In particular, let the current dimension take on values
   1666  *
   1667  *	f + s a
   1668  *
   1669  * with a an integer.  We check if we can find an integer m that (obviously)
   1670  * divides both f and s.
   1671  *
   1672  * If so, we check if the current dimension only appears in constraints
   1673  * where the coefficients of the other variables are multiples of m.
   1674  * We perform this extra check to avoid the risk of introducing
   1675  * divisions by scaling down the current dimension.
   1676  *
   1677  * If so, we scale the current dimension down by a factor of m.
   1678  * That is, we plug in
   1679  *
   1680  *	i = m i'							(1)
   1681  *
   1682  * Note that in principle we could always scale down strided loops
   1683  * by plugging in
   1684  *
   1685  *	i = f + s i'
   1686  *
   1687  * but this may result in i' taking on larger values than the original i,
   1688  * due to the shift by "f".
   1689  * By constrast, the scaling in (1) can only reduce the (absolute) value "i".
   1690  */
   1691 static __isl_give isl_ast_graft *create_node(__isl_take isl_union_map *executed,
   1692 	__isl_take isl_basic_set *bounds, __isl_take isl_set *domain,
   1693 	__isl_take isl_ast_build *build)
   1694 {
   1695 	struct isl_check_scaled_data data;
   1696 	isl_size depth;
   1697 	isl_ctx *ctx;
   1698 	isl_aff *offset;
   1699 	isl_val *d;
   1700 
   1701 	ctx = isl_ast_build_get_ctx(build);
   1702 	if (!isl_options_get_ast_build_scale_strides(ctx))
   1703 		return create_node_scaled(executed, bounds, domain, build);
   1704 
   1705 	depth = isl_ast_build_get_depth(build);
   1706 	if (depth < 0)
   1707 		build = isl_ast_build_free(build);
   1708 	data.depth = depth;
   1709 	if (!isl_ast_build_has_stride(build, data.depth))
   1710 		return create_node_scaled(executed, bounds, domain, build);
   1711 
   1712 	offset = isl_ast_build_get_offset(build, data.depth);
   1713 	data.m = isl_ast_build_get_stride(build, data.depth);
   1714 	if (!data.m)
   1715 		offset = isl_aff_free(offset);
   1716 	offset = isl_aff_scale_down_val(offset, isl_val_copy(data.m));
   1717 	d = isl_aff_get_denominator_val(offset);
   1718 	if (!d)
   1719 		executed = isl_union_map_free(executed);
   1720 
   1721 	if (executed && isl_val_is_divisible_by(data.m, d))
   1722 		data.m = isl_val_div(data.m, d);
   1723 	else {
   1724 		data.m = isl_val_set_si(data.m, 1);
   1725 		isl_val_free(d);
   1726 	}
   1727 
   1728 	if (!isl_val_is_one(data.m)) {
   1729 		if (isl_union_map_foreach_map(executed, &map_check_scaled,
   1730 						&data) < 0 &&
   1731 		    !isl_val_is_one(data.m))
   1732 			executed = isl_union_map_free(executed);
   1733 	}
   1734 
   1735 	if (!isl_val_is_one(data.m)) {
   1736 		isl_space *space;
   1737 		isl_multi_aff *ma;
   1738 		isl_aff *aff;
   1739 		isl_map *map;
   1740 		isl_union_map *umap;
   1741 
   1742 		space = isl_ast_build_get_space(build, 1);
   1743 		space = isl_space_map_from_set(space);
   1744 		ma = isl_multi_aff_identity(space);
   1745 		aff = isl_multi_aff_get_aff(ma, data.depth);
   1746 		aff = isl_aff_scale_val(aff, isl_val_copy(data.m));
   1747 		ma = isl_multi_aff_set_aff(ma, data.depth, aff);
   1748 
   1749 		bounds = isl_basic_set_preimage_multi_aff(bounds,
   1750 						isl_multi_aff_copy(ma));
   1751 		domain = isl_set_preimage_multi_aff(domain,
   1752 						isl_multi_aff_copy(ma));
   1753 		map = isl_map_reverse(isl_map_from_multi_aff(ma));
   1754 		umap = isl_union_map_from_map(map);
   1755 		executed = isl_union_map_apply_domain(executed,
   1756 						isl_union_map_copy(umap));
   1757 		build = isl_ast_build_scale_down(build, isl_val_copy(data.m),
   1758 						umap);
   1759 	}
   1760 	isl_aff_free(offset);
   1761 	isl_val_free(data.m);
   1762 
   1763 	return create_node_scaled(executed, bounds, domain, build);
   1764 }
   1765 
   1766 /* Add the basic set to the list that "user" points to.
   1767  */
   1768 static isl_stat collect_basic_set(__isl_take isl_basic_set *bset, void *user)
   1769 {
   1770 	isl_basic_set_list **list = user;
   1771 
   1772 	*list = isl_basic_set_list_add(*list, bset);
   1773 
   1774 	return isl_stat_ok;
   1775 }
   1776 
   1777 /* Extract the basic sets of "set" and collect them in an isl_basic_set_list.
   1778  */
   1779 static __isl_give isl_basic_set_list *isl_basic_set_list_from_set(
   1780 	__isl_take isl_set *set)
   1781 {
   1782 	isl_size n;
   1783 	isl_ctx *ctx;
   1784 	isl_basic_set_list *list;
   1785 
   1786 	n = isl_set_n_basic_set(set);
   1787 	if (n < 0)
   1788 		set = isl_set_free(set);
   1789 	if (!set)
   1790 		return NULL;
   1791 
   1792 	ctx = isl_set_get_ctx(set);
   1793 
   1794 	list = isl_basic_set_list_alloc(ctx, n);
   1795 	if (isl_set_foreach_basic_set(set, &collect_basic_set, &list) < 0)
   1796 		list = isl_basic_set_list_free(list);
   1797 
   1798 	isl_set_free(set);
   1799 	return list;
   1800 }
   1801 
   1802 /* Generate code for the schedule domain "bounds"
   1803  * and add the result to "list".
   1804  *
   1805  * We mainly detect strides here and check if the bounds do not
   1806  * conflict with the current build domain
   1807  * and then pass over control to create_node.
   1808  *
   1809  * "bounds" reflects the bounds on the current dimension and possibly
   1810  * some extra conditions on outer dimensions.
   1811  * It does not, however, include any divs involving the current dimension,
   1812  * so it does not capture any stride constraints.
   1813  * We therefore need to compute that part of the schedule domain that
   1814  * intersects with "bounds" and derive the strides from the result.
   1815  */
   1816 static __isl_give isl_ast_graft_list *add_node(
   1817 	__isl_take isl_ast_graft_list *list, __isl_take isl_union_map *executed,
   1818 	__isl_take isl_basic_set *bounds, __isl_take isl_ast_build *build)
   1819 {
   1820 	isl_ast_graft *graft;
   1821 	isl_set *domain = NULL;
   1822 	isl_union_set *uset;
   1823 	int empty, disjoint;
   1824 
   1825 	uset = isl_union_set_from_basic_set(isl_basic_set_copy(bounds));
   1826 	executed = isl_union_map_intersect_domain(executed, uset);
   1827 	empty = isl_union_map_is_empty(executed);
   1828 	if (empty < 0)
   1829 		goto error;
   1830 	if (empty)
   1831 		goto done;
   1832 
   1833 	uset = isl_union_map_domain(isl_union_map_copy(executed));
   1834 	domain = isl_set_from_union_set(uset);
   1835 	domain = isl_ast_build_specialize(build, domain);
   1836 
   1837 	domain = isl_set_compute_divs(domain);
   1838 	domain = isl_ast_build_eliminate_inner(build, domain);
   1839 	disjoint = isl_set_is_disjoint(domain, build->domain);
   1840 	if (disjoint < 0)
   1841 		goto error;
   1842 	if (disjoint)
   1843 		goto done;
   1844 
   1845 	build = isl_ast_build_detect_strides(build, isl_set_copy(domain));
   1846 
   1847 	graft = create_node(executed, bounds, domain,
   1848 				isl_ast_build_copy(build));
   1849 	list = isl_ast_graft_list_add(list, graft);
   1850 	isl_ast_build_free(build);
   1851 	return list;
   1852 error:
   1853 	list = isl_ast_graft_list_free(list);
   1854 done:
   1855 	isl_set_free(domain);
   1856 	isl_basic_set_free(bounds);
   1857 	isl_union_map_free(executed);
   1858 	isl_ast_build_free(build);
   1859 	return list;
   1860 }
   1861 
   1862 /* Does any element of i follow or coincide with any element of j
   1863  * at the current depth for equal values of the outer dimensions?
   1864  */
   1865 static isl_bool domain_follows_at_depth(__isl_keep isl_basic_set *i,
   1866 	__isl_keep isl_basic_set *j, void *user)
   1867 {
   1868 	int depth = *(int *) user;
   1869 	isl_basic_map *test;
   1870 	isl_bool empty;
   1871 	int l;
   1872 
   1873 	test = isl_basic_map_from_domain_and_range(isl_basic_set_copy(i),
   1874 						    isl_basic_set_copy(j));
   1875 	for (l = 0; l < depth; ++l)
   1876 		test = isl_basic_map_equate(test, isl_dim_in, l,
   1877 						isl_dim_out, l);
   1878 	test = isl_basic_map_order_ge(test, isl_dim_in, depth,
   1879 					isl_dim_out, depth);
   1880 	empty = isl_basic_map_is_empty(test);
   1881 	isl_basic_map_free(test);
   1882 
   1883 	return isl_bool_not(empty);
   1884 }
   1885 
   1886 /* Split up each element of "list" into a part that is related to "bset"
   1887  * according to "gt" and a part that is not.
   1888  * Return a list that consist of "bset" and all the pieces.
   1889  */
   1890 static __isl_give isl_basic_set_list *add_split_on(
   1891 	__isl_take isl_basic_set_list *list, __isl_take isl_basic_set *bset,
   1892 	__isl_keep isl_basic_map *gt)
   1893 {
   1894 	int i;
   1895 	isl_size n;
   1896 	isl_basic_set_list *res;
   1897 
   1898 	n = isl_basic_set_list_n_basic_set(list);
   1899 	if (n < 0)
   1900 		bset = isl_basic_set_free(bset);
   1901 
   1902 	gt = isl_basic_map_copy(gt);
   1903 	gt = isl_basic_map_intersect_domain(gt, isl_basic_set_copy(bset));
   1904 	res = isl_basic_set_list_from_basic_set(bset);
   1905 	for (i = 0; res && i < n; ++i) {
   1906 		isl_basic_set *bset;
   1907 		isl_set *set1, *set2;
   1908 		isl_basic_map *bmap;
   1909 		int empty;
   1910 
   1911 		bset = isl_basic_set_list_get_basic_set(list, i);
   1912 		bmap = isl_basic_map_copy(gt);
   1913 		bmap = isl_basic_map_intersect_range(bmap, bset);
   1914 		bset = isl_basic_map_range(bmap);
   1915 		empty = isl_basic_set_is_empty(bset);
   1916 		if (empty < 0)
   1917 			res = isl_basic_set_list_free(res);
   1918 		if (empty)  {
   1919 			isl_basic_set_free(bset);
   1920 			bset = isl_basic_set_list_get_basic_set(list, i);
   1921 			res = isl_basic_set_list_add(res, bset);
   1922 			continue;
   1923 		}
   1924 
   1925 		res = isl_basic_set_list_add(res, isl_basic_set_copy(bset));
   1926 		set1 = isl_set_from_basic_set(bset);
   1927 		bset = isl_basic_set_list_get_basic_set(list, i);
   1928 		set2 = isl_set_from_basic_set(bset);
   1929 		set1 = isl_set_subtract(set2, set1);
   1930 		set1 = isl_set_make_disjoint(set1);
   1931 
   1932 		res = isl_basic_set_list_concat(res,
   1933 					    isl_basic_set_list_from_set(set1));
   1934 	}
   1935 	isl_basic_map_free(gt);
   1936 	isl_basic_set_list_free(list);
   1937 	return res;
   1938 }
   1939 
   1940 static __isl_give isl_ast_graft_list *generate_sorted_domains(
   1941 	__isl_keep isl_basic_set_list *domain_list,
   1942 	__isl_keep isl_union_map *executed,
   1943 	__isl_keep isl_ast_build *build);
   1944 
   1945 /* Internal data structure for add_nodes.
   1946  *
   1947  * "executed" and "build" are extra arguments to be passed to add_node.
   1948  * "list" collects the results.
   1949  */
   1950 struct isl_add_nodes_data {
   1951 	isl_union_map *executed;
   1952 	isl_ast_build *build;
   1953 
   1954 	isl_ast_graft_list *list;
   1955 };
   1956 
   1957 /* Generate code for the schedule domains in "scc"
   1958  * and add the results to "list".
   1959  *
   1960  * The domains in "scc" form a strongly connected component in the ordering.
   1961  * If the number of domains in "scc" is larger than 1, then this means
   1962  * that we cannot determine a valid ordering for the domains in the component.
   1963  * This should be fairly rare because the individual domains
   1964  * have been made disjoint first.
   1965  * The problem is that the domains may be integrally disjoint but not
   1966  * rationally disjoint.  For example, we may have domains
   1967  *
   1968  *	{ [i,i] : 0 <= i <= 1 }		and	{ [i,1-i] : 0 <= i <= 1 }
   1969  *
   1970  * These two domains have an empty intersection, but their rational
   1971  * relaxations do intersect.  It is impossible to order these domains
   1972  * in the second dimension because the first should be ordered before
   1973  * the second for outer dimension equal to 0, while it should be ordered
   1974  * after for outer dimension equal to 1.
   1975  *
   1976  * This may happen in particular in case of unrolling since the domain
   1977  * of each slice is replaced by its simple hull.
   1978  *
   1979  * For each basic set i in "scc" and for each of the following basic sets j,
   1980  * we split off that part of the basic set i that shares the outer dimensions
   1981  * with j and lies before j in the current dimension.
   1982  * We collect all the pieces in a new list that replaces "scc".
   1983  *
   1984  * While the elements in "scc" should be disjoint, we double-check
   1985  * this property to avoid running into an infinite recursion in case
   1986  * they intersect due to some internal error.
   1987  */
   1988 static isl_stat add_nodes(__isl_take isl_basic_set_list *scc, void *user)
   1989 {
   1990 	struct isl_add_nodes_data *data = user;
   1991 	int i;
   1992 	isl_size depth;
   1993 	isl_size n;
   1994 	isl_basic_set *bset, *first;
   1995 	isl_basic_set_list *list;
   1996 	isl_space *space;
   1997 	isl_basic_map *gt;
   1998 
   1999 	n = isl_basic_set_list_n_basic_set(scc);
   2000 	if (n < 0)
   2001 		goto error;
   2002 	bset = isl_basic_set_list_get_basic_set(scc, 0);
   2003 	if (n == 1) {
   2004 		isl_basic_set_list_free(scc);
   2005 		data->list = add_node(data->list,
   2006 				isl_union_map_copy(data->executed), bset,
   2007 				isl_ast_build_copy(data->build));
   2008 		return data->list ? isl_stat_ok : isl_stat_error;
   2009 	}
   2010 
   2011 	depth = isl_ast_build_get_depth(data->build);
   2012 	if (depth < 0)
   2013 		bset = isl_basic_set_free(bset);
   2014 	space = isl_basic_set_get_space(bset);
   2015 	space = isl_space_map_from_set(space);
   2016 	gt = isl_basic_map_universe(space);
   2017 	for (i = 0; i < depth; ++i)
   2018 		gt = isl_basic_map_equate(gt, isl_dim_in, i, isl_dim_out, i);
   2019 	gt = isl_basic_map_order_gt(gt, isl_dim_in, depth, isl_dim_out, depth);
   2020 
   2021 	first = isl_basic_set_copy(bset);
   2022 	list = isl_basic_set_list_from_basic_set(bset);
   2023 	for (i = 1; i < n; ++i) {
   2024 		int disjoint;
   2025 
   2026 		bset = isl_basic_set_list_get_basic_set(scc, i);
   2027 
   2028 		disjoint = isl_basic_set_is_disjoint(bset, first);
   2029 		if (disjoint < 0)
   2030 			list = isl_basic_set_list_free(list);
   2031 		else if (!disjoint)
   2032 			isl_die(isl_basic_set_list_get_ctx(scc),
   2033 				isl_error_internal,
   2034 				"basic sets in scc are assumed to be disjoint",
   2035 				list = isl_basic_set_list_free(list));
   2036 
   2037 		list = add_split_on(list, bset, gt);
   2038 	}
   2039 	isl_basic_set_free(first);
   2040 	isl_basic_map_free(gt);
   2041 	isl_basic_set_list_free(scc);
   2042 	scc = list;
   2043 	data->list = isl_ast_graft_list_concat(data->list,
   2044 		    generate_sorted_domains(scc, data->executed, data->build));
   2045 	isl_basic_set_list_free(scc);
   2046 
   2047 	return data->list ? isl_stat_ok : isl_stat_error;
   2048 error:
   2049 	isl_basic_set_list_free(scc);
   2050 	return isl_stat_error;
   2051 }
   2052 
   2053 /* Sort the domains in "domain_list" according to the execution order
   2054  * at the current depth (for equal values of the outer dimensions),
   2055  * generate code for each of them, collecting the results in a list.
   2056  * If no code is generated (because the intersection of the inverse schedule
   2057  * with the domains turns out to be empty), then an empty list is returned.
   2058  *
   2059  * The caller is responsible for ensuring that the basic sets in "domain_list"
   2060  * are pair-wise disjoint.  It can, however, in principle happen that
   2061  * two basic sets should be ordered one way for one value of the outer
   2062  * dimensions and the other way for some other value of the outer dimensions.
   2063  * We therefore play safe and look for strongly connected components.
   2064  * The function add_nodes takes care of handling non-trivial components.
   2065  */
   2066 static __isl_give isl_ast_graft_list *generate_sorted_domains(
   2067 	__isl_keep isl_basic_set_list *domain_list,
   2068 	__isl_keep isl_union_map *executed, __isl_keep isl_ast_build *build)
   2069 {
   2070 	isl_ctx *ctx;
   2071 	struct isl_add_nodes_data data;
   2072 	isl_size depth;
   2073 	isl_size n;
   2074 
   2075 	n = isl_basic_set_list_n_basic_set(domain_list);
   2076 	if (n < 0)
   2077 		return NULL;
   2078 
   2079 	ctx = isl_basic_set_list_get_ctx(domain_list);
   2080 	data.list = isl_ast_graft_list_alloc(ctx, n);
   2081 	if (n == 0)
   2082 		return data.list;
   2083 	if (n == 1)
   2084 		return add_node(data.list, isl_union_map_copy(executed),
   2085 			isl_basic_set_list_get_basic_set(domain_list, 0),
   2086 			isl_ast_build_copy(build));
   2087 
   2088 	depth = isl_ast_build_get_depth(build);
   2089 	data.executed = executed;
   2090 	data.build = build;
   2091 	if (depth < 0 || isl_basic_set_list_foreach_scc(domain_list,
   2092 					&domain_follows_at_depth, &depth,
   2093 					&add_nodes, &data) < 0)
   2094 		data.list = isl_ast_graft_list_free(data.list);
   2095 
   2096 	return data.list;
   2097 }
   2098 
   2099 /* Do i and j share any values for the outer dimensions?
   2100  */
   2101 static isl_bool shared_outer(__isl_keep isl_basic_set *i,
   2102 	__isl_keep isl_basic_set *j, void *user)
   2103 {
   2104 	int depth = *(int *) user;
   2105 	isl_basic_map *test;
   2106 	isl_bool empty;
   2107 	int l;
   2108 
   2109 	test = isl_basic_map_from_domain_and_range(isl_basic_set_copy(i),
   2110 						    isl_basic_set_copy(j));
   2111 	for (l = 0; l < depth; ++l)
   2112 		test = isl_basic_map_equate(test, isl_dim_in, l,
   2113 						isl_dim_out, l);
   2114 	empty = isl_basic_map_is_empty(test);
   2115 	isl_basic_map_free(test);
   2116 
   2117 	return isl_bool_not(empty);
   2118 }
   2119 
   2120 /* Internal data structure for generate_sorted_domains_wrap.
   2121  *
   2122  * "n" is the total number of basic sets
   2123  * "executed" and "build" are extra arguments to be passed
   2124  *	to generate_sorted_domains.
   2125  *
   2126  * "single" is set to 1 by generate_sorted_domains_wrap if there
   2127  * is only a single component.
   2128  * "list" collects the results.
   2129  */
   2130 struct isl_ast_generate_parallel_domains_data {
   2131 	isl_size n;
   2132 	isl_union_map *executed;
   2133 	isl_ast_build *build;
   2134 
   2135 	int single;
   2136 	isl_ast_graft_list *list;
   2137 };
   2138 
   2139 /* Call generate_sorted_domains on "scc", fuse the result into a list
   2140  * with either zero or one graft and collect the these single element
   2141  * lists into data->list.
   2142  *
   2143  * If there is only one component, i.e., if the number of basic sets
   2144  * in the current component is equal to the total number of basic sets,
   2145  * then data->single is set to 1 and the result of generate_sorted_domains
   2146  * is not fused.
   2147  */
   2148 static isl_stat generate_sorted_domains_wrap(__isl_take isl_basic_set_list *scc,
   2149 	void *user)
   2150 {
   2151 	struct isl_ast_generate_parallel_domains_data *data = user;
   2152 	isl_ast_graft_list *list;
   2153 	isl_size n;
   2154 
   2155 	n = isl_basic_set_list_n_basic_set(scc);
   2156 	if (n < 0)
   2157 		scc = isl_basic_set_list_free(scc);
   2158 	list = generate_sorted_domains(scc, data->executed, data->build);
   2159 	data->single = n == data->n;
   2160 	if (!data->single)
   2161 		list = isl_ast_graft_list_fuse(list, data->build);
   2162 	if (!data->list)
   2163 		data->list = list;
   2164 	else
   2165 		data->list = isl_ast_graft_list_concat(data->list, list);
   2166 
   2167 	isl_basic_set_list_free(scc);
   2168 	if (!data->list)
   2169 		return isl_stat_error;
   2170 
   2171 	return isl_stat_ok;
   2172 }
   2173 
   2174 /* Look for any (weakly connected) components in the "domain_list"
   2175  * of domains that share some values of the outer dimensions.
   2176  * That is, domains in different components do not share any values
   2177  * of the outer dimensions.  This means that these components
   2178  * can be freely reordered.
   2179  * Within each of the components, we sort the domains according
   2180  * to the execution order at the current depth.
   2181  *
   2182  * If there is more than one component, then generate_sorted_domains_wrap
   2183  * fuses the result of each call to generate_sorted_domains
   2184  * into a list with either zero or one graft and collects these (at most)
   2185  * single element lists into a bigger list. This means that the elements of the
   2186  * final list can be freely reordered.  In particular, we sort them
   2187  * according to an arbitrary but fixed ordering to ease merging of
   2188  * graft lists from different components.
   2189  */
   2190 static __isl_give isl_ast_graft_list *generate_parallel_domains(
   2191 	__isl_keep isl_basic_set_list *domain_list,
   2192 	__isl_keep isl_union_map *executed, __isl_keep isl_ast_build *build)
   2193 {
   2194 	isl_size depth;
   2195 	struct isl_ast_generate_parallel_domains_data data;
   2196 
   2197 	data.n = isl_basic_set_list_n_basic_set(domain_list);
   2198 	if (data.n < 0)
   2199 		return NULL;
   2200 
   2201 	if (data.n <= 1)
   2202 		return generate_sorted_domains(domain_list, executed, build);
   2203 
   2204 	depth = isl_ast_build_get_depth(build);
   2205 	if (depth < 0)
   2206 		return NULL;
   2207 	data.list = NULL;
   2208 	data.executed = executed;
   2209 	data.build = build;
   2210 	data.single = 0;
   2211 	if (isl_basic_set_list_foreach_scc(domain_list, &shared_outer, &depth,
   2212 					    &generate_sorted_domains_wrap,
   2213 					    &data) < 0)
   2214 		data.list = isl_ast_graft_list_free(data.list);
   2215 
   2216 	if (!data.single)
   2217 		data.list = isl_ast_graft_list_sort_guard(data.list);
   2218 
   2219 	return data.list;
   2220 }
   2221 
   2222 /* Internal data for separate_domain.
   2223  *
   2224  * "explicit" is set if we only want to use explicit bounds.
   2225  *
   2226  * "domain" collects the separated domains.
   2227  */
   2228 struct isl_separate_domain_data {
   2229 	isl_ast_build *build;
   2230 	int explicit;
   2231 	isl_set *domain;
   2232 };
   2233 
   2234 /* Extract implicit bounds on the current dimension for the executed "map".
   2235  *
   2236  * The domain of "map" may involve inner dimensions, so we
   2237  * need to eliminate them.
   2238  */
   2239 static __isl_give isl_set *implicit_bounds(__isl_take isl_map *map,
   2240 	__isl_keep isl_ast_build *build)
   2241 {
   2242 	isl_set *domain;
   2243 
   2244 	domain = isl_map_domain(map);
   2245 	domain = isl_ast_build_eliminate(build, domain);
   2246 
   2247 	return domain;
   2248 }
   2249 
   2250 /* Extract explicit bounds on the current dimension for the executed "map".
   2251  *
   2252  * Rather than eliminating the inner dimensions as in implicit_bounds,
   2253  * we simply drop any constraints involving those inner dimensions.
   2254  * The idea is that most bounds that are implied by constraints on the
   2255  * inner dimensions will be enforced by for loops and not by explicit guards.
   2256  * There is then no need to separate along those bounds.
   2257  */
   2258 static __isl_give isl_set *explicit_bounds(__isl_take isl_map *map,
   2259 	__isl_keep isl_ast_build *build)
   2260 {
   2261 	isl_set *domain;
   2262 	isl_size depth;
   2263 	isl_size dim;
   2264 
   2265 	depth = isl_ast_build_get_depth(build);
   2266 	dim = isl_map_dim(map, isl_dim_out);
   2267 	if (depth < 0 || dim < 0)
   2268 		return isl_map_domain(isl_map_free(map));
   2269 	map = isl_map_drop_constraints_involving_dims(map, isl_dim_out, 0, dim);
   2270 
   2271 	domain = isl_map_domain(map);
   2272 	dim = isl_set_dim(domain, isl_dim_set);
   2273 	domain = isl_set_detect_equalities(domain);
   2274 	domain = isl_set_drop_constraints_involving_dims(domain,
   2275 				isl_dim_set, depth + 1, dim - (depth + 1));
   2276 	domain = isl_set_remove_divs_involving_dims(domain,
   2277 				isl_dim_set, depth, 1);
   2278 	domain = isl_set_remove_unknown_divs(domain);
   2279 
   2280 	return domain;
   2281 }
   2282 
   2283 /* Split data->domain into pieces that intersect with the range of "map"
   2284  * and pieces that do not intersect with the range of "map"
   2285  * and then add that part of the range of "map" that does not intersect
   2286  * with data->domain.
   2287  */
   2288 static isl_stat separate_domain(__isl_take isl_map *map, void *user)
   2289 {
   2290 	struct isl_separate_domain_data *data = user;
   2291 	isl_set *domain;
   2292 	isl_set *d1, *d2;
   2293 
   2294 	if (data->explicit)
   2295 		domain = explicit_bounds(map, data->build);
   2296 	else
   2297 		domain = implicit_bounds(map, data->build);
   2298 
   2299 	domain = isl_set_coalesce(domain);
   2300 	domain = isl_set_make_disjoint(domain);
   2301 	d1 = isl_set_subtract(isl_set_copy(domain), isl_set_copy(data->domain));
   2302 	d2 = isl_set_subtract(isl_set_copy(data->domain), isl_set_copy(domain));
   2303 	data->domain = isl_set_intersect(data->domain, domain);
   2304 	data->domain = isl_set_union(data->domain, d1);
   2305 	data->domain = isl_set_union(data->domain, d2);
   2306 
   2307 	return isl_stat_ok;
   2308 }
   2309 
   2310 /* Separate the schedule domains of "executed".
   2311  *
   2312  * That is, break up the domain of "executed" into basic sets,
   2313  * such that for each basic set S, every element in S is associated with
   2314  * the same domain spaces.
   2315  *
   2316  * "space" is the (single) domain space of "executed".
   2317  */
   2318 static __isl_give isl_set *separate_schedule_domains(
   2319 	__isl_take isl_space *space, __isl_take isl_union_map *executed,
   2320 	__isl_keep isl_ast_build *build)
   2321 {
   2322 	struct isl_separate_domain_data data = { build };
   2323 	isl_ctx *ctx;
   2324 
   2325 	ctx = isl_ast_build_get_ctx(build);
   2326 	data.explicit = isl_options_get_ast_build_separation_bounds(ctx) ==
   2327 				    ISL_AST_BUILD_SEPARATION_BOUNDS_EXPLICIT;
   2328 	data.domain = isl_set_empty(space);
   2329 	if (isl_union_map_foreach_map(executed, &separate_domain, &data) < 0)
   2330 		data.domain = isl_set_free(data.domain);
   2331 
   2332 	isl_union_map_free(executed);
   2333 	return data.domain;
   2334 }
   2335 
   2336 /* Temporary data used during the search for a lower bound for unrolling.
   2337  *
   2338  * "build" is the build in which the unrolling will be performed
   2339  * "domain" is the original set for which to find a lower bound
   2340  * "depth" is the dimension for which to find a lower boudn
   2341  * "expansion" is the expansion that needs to be applied to "domain"
   2342  * in the unrolling that will be performed
   2343  *
   2344  * "lower" is the best lower bound found so far.  It is NULL if we have not
   2345  * found any yet.
   2346  * "n" is the corresponding size.  If lower is NULL, then the value of n
   2347  * is undefined.
   2348  * "n_div" is the maximal number of integer divisions in the first
   2349  * unrolled iteration (after expansion).  It is set to -1 if it hasn't
   2350  * been computed yet.
   2351  */
   2352 struct isl_find_unroll_data {
   2353 	isl_ast_build *build;
   2354 	isl_set *domain;
   2355 	int depth;
   2356 	isl_basic_map *expansion;
   2357 
   2358 	isl_aff *lower;
   2359 	int *n;
   2360 	int n_div;
   2361 };
   2362 
   2363 /* Return the constraint
   2364  *
   2365  *	i_"depth" = aff + offset
   2366  */
   2367 static __isl_give isl_constraint *at_offset(int depth, __isl_keep isl_aff *aff,
   2368 	int offset)
   2369 {
   2370 	aff = isl_aff_copy(aff);
   2371 	aff = isl_aff_add_coefficient_si(aff, isl_dim_in, depth, -1);
   2372 	aff = isl_aff_add_constant_si(aff, offset);
   2373 	return isl_equality_from_aff(aff);
   2374 }
   2375 
   2376 /* Update *user to the number of integer divisions in the first element
   2377  * of "ma", if it is larger than the current value.
   2378  */
   2379 static isl_stat update_n_div(__isl_take isl_set *set,
   2380 	__isl_take isl_multi_aff *ma, void *user)
   2381 {
   2382 	isl_aff *aff;
   2383 	int *n = user;
   2384 	isl_size n_div;
   2385 
   2386 	aff = isl_multi_aff_get_aff(ma, 0);
   2387 	n_div = isl_aff_dim(aff, isl_dim_div);
   2388 	isl_aff_free(aff);
   2389 	isl_multi_aff_free(ma);
   2390 	isl_set_free(set);
   2391 
   2392 	if (n_div > *n)
   2393 		*n = n_div;
   2394 
   2395 	return n_div >= 0 ? isl_stat_ok : isl_stat_error;
   2396 }
   2397 
   2398 /* Get the number of integer divisions in the expression for the iterator
   2399  * value at the first slice in the unrolling based on lower bound "lower",
   2400  * taking into account the expansion that needs to be performed on this slice.
   2401  */
   2402 static int get_expanded_n_div(struct isl_find_unroll_data *data,
   2403 	__isl_keep isl_aff *lower)
   2404 {
   2405 	isl_constraint *c;
   2406 	isl_set *set;
   2407 	isl_map *it_map, *expansion;
   2408 	isl_pw_multi_aff *pma;
   2409 	int n;
   2410 
   2411 	c = at_offset(data->depth, lower, 0);
   2412 	set = isl_set_copy(data->domain);
   2413 	set = isl_set_add_constraint(set, c);
   2414 	expansion = isl_map_from_basic_map(isl_basic_map_copy(data->expansion));
   2415 	set = isl_set_apply(set, expansion);
   2416 	it_map = isl_ast_build_map_to_iterator(data->build, set);
   2417 	pma = isl_pw_multi_aff_from_map(it_map);
   2418 	n = 0;
   2419 	if (isl_pw_multi_aff_foreach_piece(pma, &update_n_div, &n) < 0)
   2420 		n = -1;
   2421 	isl_pw_multi_aff_free(pma);
   2422 
   2423 	return n;
   2424 }
   2425 
   2426 /* Is the lower bound "lower" with corresponding iteration count "n"
   2427  * better than the one stored in "data"?
   2428  * If there is no upper bound on the iteration count ("n" is infinity) or
   2429  * if the count is too large, then we cannot use this lower bound.
   2430  * Otherwise, if there was no previous lower bound or
   2431  * if the iteration count of the new lower bound is smaller than
   2432  * the iteration count of the previous lower bound, then we consider
   2433  * the new lower bound to be better.
   2434  * If the iteration count is the same, then compare the number
   2435  * of integer divisions that would be needed to express
   2436  * the iterator value at the first slice in the unrolling
   2437  * according to the lower bound.  If we end up computing this
   2438  * number, then store the lowest value in data->n_div.
   2439  */
   2440 static int is_better_lower_bound(struct isl_find_unroll_data *data,
   2441 	__isl_keep isl_aff *lower, __isl_keep isl_val *n)
   2442 {
   2443 	int cmp;
   2444 	int n_div;
   2445 
   2446 	if (!n)
   2447 		return -1;
   2448 	if (isl_val_is_infty(n))
   2449 		return 0;
   2450 	if (isl_val_cmp_si(n, INT_MAX) > 0)
   2451 		return 0;
   2452 	if (!data->lower)
   2453 		return 1;
   2454 	cmp = isl_val_cmp_si(n, *data->n);
   2455 	if (cmp < 0)
   2456 		return 1;
   2457 	if (cmp > 0)
   2458 		return 0;
   2459 	if (data->n_div < 0)
   2460 		data->n_div = get_expanded_n_div(data, data->lower);
   2461 	if (data->n_div < 0)
   2462 		return -1;
   2463 	if (data->n_div == 0)
   2464 		return 0;
   2465 	n_div = get_expanded_n_div(data, lower);
   2466 	if (n_div < 0)
   2467 		return -1;
   2468 	if (n_div >= data->n_div)
   2469 		return 0;
   2470 	data->n_div = n_div;
   2471 
   2472 	return 1;
   2473 }
   2474 
   2475 /* Check if we can use "c" as a lower bound and if it is better than
   2476  * any previously found lower bound.
   2477  *
   2478  * If "c" does not involve the dimension at the current depth,
   2479  * then we cannot use it.
   2480  * Otherwise, let "c" be of the form
   2481  *
   2482  *	i >= f(j)/a
   2483  *
   2484  * We compute the maximal value of
   2485  *
   2486  *	-ceil(f(j)/a)) + i + 1
   2487  *
   2488  * over the domain.  If there is such a value "n", then we know
   2489  *
   2490  *	-ceil(f(j)/a)) + i + 1 <= n
   2491  *
   2492  * or
   2493  *
   2494  *	i < ceil(f(j)/a)) + n
   2495  *
   2496  * meaning that we can use ceil(f(j)/a)) as a lower bound for unrolling.
   2497  * We just need to check if we have found any lower bound before and
   2498  * if the new lower bound is better (smaller n or fewer integer divisions)
   2499  * than the previously found lower bounds.
   2500  */
   2501 static isl_stat update_unrolling_lower_bound(struct isl_find_unroll_data *data,
   2502 	__isl_keep isl_constraint *c)
   2503 {
   2504 	isl_aff *aff, *lower;
   2505 	isl_val *max;
   2506 	int better;
   2507 
   2508 	if (!isl_constraint_is_lower_bound(c, isl_dim_set, data->depth))
   2509 		return isl_stat_ok;
   2510 
   2511 	lower = isl_constraint_get_bound(c, isl_dim_set, data->depth);
   2512 	lower = isl_aff_ceil(lower);
   2513 	aff = isl_aff_copy(lower);
   2514 	aff = isl_aff_neg(aff);
   2515 	aff = isl_aff_add_coefficient_si(aff, isl_dim_in, data->depth, 1);
   2516 	aff = isl_aff_add_constant_si(aff, 1);
   2517 	max = isl_set_max_val(data->domain, aff);
   2518 	isl_aff_free(aff);
   2519 
   2520 	better = is_better_lower_bound(data, lower, max);
   2521 	if (better < 0 || !better) {
   2522 		isl_val_free(max);
   2523 		isl_aff_free(lower);
   2524 		return better < 0 ? isl_stat_error : isl_stat_ok;
   2525 	}
   2526 
   2527 	isl_aff_free(data->lower);
   2528 	data->lower = lower;
   2529 	*data->n = isl_val_get_num_si(max);
   2530 	isl_val_free(max);
   2531 
   2532 	return isl_stat_ok;
   2533 }
   2534 
   2535 /* Check if we can use "c" as a lower bound and if it is better than
   2536  * any previously found lower bound.
   2537  */
   2538 static isl_stat constraint_find_unroll(__isl_take isl_constraint *c, void *user)
   2539 {
   2540 	struct isl_find_unroll_data *data;
   2541 	isl_stat r;
   2542 
   2543 	data = (struct isl_find_unroll_data *) user;
   2544 	r = update_unrolling_lower_bound(data, c);
   2545 	isl_constraint_free(c);
   2546 
   2547 	return r;
   2548 }
   2549 
   2550 /* Look for a lower bound l(i) on the dimension at "depth"
   2551  * and a size n such that "domain" is a subset of
   2552  *
   2553  *	{ [i] : l(i) <= i_d < l(i) + n }
   2554  *
   2555  * where d is "depth" and l(i) depends only on earlier dimensions.
   2556  * Furthermore, try and find a lower bound such that n is as small as possible.
   2557  * In particular, "n" needs to be finite.
   2558  * "build" is the build in which the unrolling will be performed.
   2559  * "expansion" is the expansion that needs to be applied to "domain"
   2560  * in the unrolling that will be performed.
   2561  *
   2562  * Inner dimensions have been eliminated from "domain" by the caller.
   2563  *
   2564  * We first construct a collection of lower bounds on the input set
   2565  * by computing its simple hull.  We then iterate through them,
   2566  * discarding those that we cannot use (either because they do not
   2567  * involve the dimension at "depth" or because they have no corresponding
   2568  * upper bound, meaning that "n" would be unbounded) and pick out the
   2569  * best from the remaining ones.
   2570  *
   2571  * If we cannot find a suitable lower bound, then we consider that
   2572  * to be an error.
   2573  */
   2574 static __isl_give isl_aff *find_unroll_lower_bound(
   2575 	__isl_keep isl_ast_build *build, __isl_keep isl_set *domain,
   2576 	int depth, __isl_keep isl_basic_map *expansion, int *n)
   2577 {
   2578 	struct isl_find_unroll_data data =
   2579 			{ build, domain, depth, expansion, NULL, n, -1 };
   2580 	isl_basic_set *hull;
   2581 
   2582 	hull = isl_set_simple_hull(isl_set_copy(domain));
   2583 
   2584 	if (isl_basic_set_foreach_constraint(hull,
   2585 					    &constraint_find_unroll, &data) < 0)
   2586 		goto error;
   2587 
   2588 	isl_basic_set_free(hull);
   2589 
   2590 	if (!data.lower)
   2591 		isl_die(isl_set_get_ctx(domain), isl_error_invalid,
   2592 			"cannot find lower bound for unrolling", return NULL);
   2593 
   2594 	return data.lower;
   2595 error:
   2596 	isl_basic_set_free(hull);
   2597 	return isl_aff_free(data.lower);
   2598 }
   2599 
   2600 /* Call "fn" on each iteration of the current dimension of "domain".
   2601  * If "init" is not NULL, then it is called with the number of
   2602  * iterations before any call to "fn".
   2603  * Return -1 on failure.
   2604  *
   2605  * Since we are going to be iterating over the individual values,
   2606  * we first check if there are any strides on the current dimension.
   2607  * If there is, we rewrite the current dimension i as
   2608  *
   2609  *		i = stride i' + offset
   2610  *
   2611  * and then iterate over individual values of i' instead.
   2612  *
   2613  * We then look for a lower bound on i' and a size such that the domain
   2614  * is a subset of
   2615  *
   2616  *	{ [j,i'] : l(j) <= i' < l(j) + n }
   2617  *
   2618  * and then take slices of the domain at values of i'
   2619  * between l(j) and l(j) + n - 1.
   2620  *
   2621  * We compute the unshifted simple hull of each slice to ensure that
   2622  * we have a single basic set per offset.  The slicing constraint
   2623  * may get simplified away before the unshifted simple hull is taken
   2624  * and may therefore in some rare cases disappear from the result.
   2625  * We therefore explicitly add the constraint back after computing
   2626  * the unshifted simple hull to ensure that the basic sets
   2627  * remain disjoint.  The constraints that are dropped by taking the hull
   2628  * will be taken into account at the next level, as in the case of the
   2629  * atomic option.
   2630  *
   2631  * Finally, we map i' back to i and call "fn".
   2632  */
   2633 static int foreach_iteration(__isl_take isl_set *domain,
   2634 	__isl_keep isl_ast_build *build, int (*init)(int n, void *user),
   2635 	int (*fn)(__isl_take isl_basic_set *bset, void *user), void *user)
   2636 {
   2637 	int i, n;
   2638 	isl_bool empty;
   2639 	isl_size depth;
   2640 	isl_multi_aff *expansion;
   2641 	isl_basic_map *bmap;
   2642 	isl_aff *lower = NULL;
   2643 	isl_ast_build *stride_build;
   2644 
   2645 	depth = isl_ast_build_get_depth(build);
   2646 	if (depth < 0)
   2647 		domain = isl_set_free(domain);
   2648 
   2649 	domain = isl_ast_build_eliminate_inner(build, domain);
   2650 	domain = isl_set_intersect(domain, isl_ast_build_get_domain(build));
   2651 	stride_build = isl_ast_build_copy(build);
   2652 	stride_build = isl_ast_build_detect_strides(stride_build,
   2653 							isl_set_copy(domain));
   2654 	expansion = isl_ast_build_get_stride_expansion(stride_build);
   2655 
   2656 	domain = isl_set_preimage_multi_aff(domain,
   2657 					    isl_multi_aff_copy(expansion));
   2658 	domain = isl_ast_build_eliminate_divs(stride_build, domain);
   2659 	isl_ast_build_free(stride_build);
   2660 
   2661 	bmap = isl_basic_map_from_multi_aff(expansion);
   2662 
   2663 	empty = isl_set_is_empty(domain);
   2664 	if (empty < 0) {
   2665 		n = -1;
   2666 	} else if (empty) {
   2667 		n = 0;
   2668 	} else {
   2669 		lower = find_unroll_lower_bound(build, domain, depth, bmap, &n);
   2670 		if (!lower)
   2671 			n = -1;
   2672 	}
   2673 	if (n >= 0 && init && init(n, user) < 0)
   2674 		n = -1;
   2675 	for (i = 0; i < n; ++i) {
   2676 		isl_set *set;
   2677 		isl_basic_set *bset;
   2678 		isl_constraint *slice;
   2679 
   2680 		slice = at_offset(depth, lower, i);
   2681 		set = isl_set_copy(domain);
   2682 		set = isl_set_add_constraint(set, isl_constraint_copy(slice));
   2683 		bset = isl_set_unshifted_simple_hull(set);
   2684 		bset = isl_basic_set_add_constraint(bset, slice);
   2685 		bset = isl_basic_set_apply(bset, isl_basic_map_copy(bmap));
   2686 
   2687 		if (fn(bset, user) < 0)
   2688 			break;
   2689 	}
   2690 
   2691 	isl_aff_free(lower);
   2692 	isl_set_free(domain);
   2693 	isl_basic_map_free(bmap);
   2694 
   2695 	return n < 0 || i < n ? -1 : 0;
   2696 }
   2697 
   2698 /* Data structure for storing the results and the intermediate objects
   2699  * of compute_domains.
   2700  *
   2701  * "list" is the main result of the function and contains a list
   2702  * of disjoint basic sets for which code should be generated.
   2703  *
   2704  * "executed" and "build" are inputs to compute_domains.
   2705  * "schedule_domain" is the domain of "executed".
   2706  *
   2707  * "option" contains the domains at the current depth that should by
   2708  * atomic, separated or unrolled.  These domains are as specified by
   2709  * the user, except that inner dimensions have been eliminated and
   2710  * that they have been made pair-wise disjoint.
   2711  *
   2712  * "sep_class" contains the user-specified split into separation classes
   2713  * specialized to the current depth.
   2714  * "done" contains the union of the separation domains that have already
   2715  * been handled.
   2716  */
   2717 struct isl_codegen_domains {
   2718 	isl_basic_set_list *list;
   2719 
   2720 	isl_union_map *executed;
   2721 	isl_ast_build *build;
   2722 	isl_set *schedule_domain;
   2723 
   2724 	isl_set *option[4];
   2725 
   2726 	isl_map *sep_class;
   2727 	isl_set *done;
   2728 };
   2729 
   2730 /* Internal data structure for do_unroll.
   2731  *
   2732  * "domains" stores the results of compute_domains.
   2733  * "class_domain" is the original class domain passed to do_unroll.
   2734  * "unroll_domain" collects the unrolled iterations.
   2735  */
   2736 struct isl_ast_unroll_data {
   2737 	struct isl_codegen_domains *domains;
   2738 	isl_set *class_domain;
   2739 	isl_set *unroll_domain;
   2740 };
   2741 
   2742 /* Given an iteration of an unrolled domain represented by "bset",
   2743  * add it to data->domains->list.
   2744  * Since we may have dropped some constraints, we intersect with
   2745  * the class domain again to ensure that each element in the list
   2746  * is disjoint from the other class domains.
   2747  */
   2748 static int do_unroll_iteration(__isl_take isl_basic_set *bset, void *user)
   2749 {
   2750 	struct isl_ast_unroll_data *data = user;
   2751 	isl_set *set;
   2752 	isl_basic_set_list *list;
   2753 
   2754 	set = isl_set_from_basic_set(bset);
   2755 	data->unroll_domain = isl_set_union(data->unroll_domain,
   2756 					    isl_set_copy(set));
   2757 	set = isl_set_intersect(set, isl_set_copy(data->class_domain));
   2758 	set = isl_set_make_disjoint(set);
   2759 	list = isl_basic_set_list_from_set(set);
   2760 	data->domains->list = isl_basic_set_list_concat(data->domains->list,
   2761 							list);
   2762 
   2763 	return 0;
   2764 }
   2765 
   2766 /* Extend domains->list with a list of basic sets, one for each value
   2767  * of the current dimension in "domain" and remove the corresponding
   2768  * sets from the class domain.  Return the updated class domain.
   2769  * The divs that involve the current dimension have not been projected out
   2770  * from this domain.
   2771  *
   2772  * We call foreach_iteration to iterate over the individual values and
   2773  * in do_unroll_iteration we collect the individual basic sets in
   2774  * domains->list and their union in data->unroll_domain, which is then
   2775  * used to update the class domain.
   2776  */
   2777 static __isl_give isl_set *do_unroll(struct isl_codegen_domains *domains,
   2778 	__isl_take isl_set *domain, __isl_take isl_set *class_domain)
   2779 {
   2780 	struct isl_ast_unroll_data data;
   2781 
   2782 	if (!domain)
   2783 		return isl_set_free(class_domain);
   2784 	if (!class_domain)
   2785 		return isl_set_free(domain);
   2786 
   2787 	data.domains = domains;
   2788 	data.class_domain = class_domain;
   2789 	data.unroll_domain = isl_set_empty(isl_set_get_space(domain));
   2790 
   2791 	if (foreach_iteration(domain, domains->build, NULL,
   2792 				&do_unroll_iteration, &data) < 0)
   2793 		data.unroll_domain = isl_set_free(data.unroll_domain);
   2794 
   2795 	class_domain = isl_set_subtract(class_domain, data.unroll_domain);
   2796 
   2797 	return class_domain;
   2798 }
   2799 
   2800 /* Add domains to domains->list for each individual value of the current
   2801  * dimension, for that part of the schedule domain that lies in the
   2802  * intersection of the option domain and the class domain.
   2803  * Remove the corresponding sets from the class domain and
   2804  * return the updated class domain.
   2805  *
   2806  * We first break up the unroll option domain into individual pieces
   2807  * and then handle each of them separately.  The unroll option domain
   2808  * has been made disjoint in compute_domains_init_options,
   2809  *
   2810  * Note that we actively want to combine different pieces of the
   2811  * schedule domain that have the same value at the current dimension.
   2812  * We therefore need to break up the unroll option domain before
   2813  * intersecting with class and schedule domain, hoping that the
   2814  * unroll option domain specified by the user is relatively simple.
   2815  */
   2816 static __isl_give isl_set *compute_unroll_domains(
   2817 	struct isl_codegen_domains *domains, __isl_take isl_set *class_domain)
   2818 {
   2819 	isl_set *unroll_domain;
   2820 	isl_basic_set_list *unroll_list;
   2821 	int i;
   2822 	isl_size n;
   2823 	isl_bool empty;
   2824 
   2825 	empty = isl_set_is_empty(domains->option[isl_ast_loop_unroll]);
   2826 	if (empty < 0)
   2827 		return isl_set_free(class_domain);
   2828 	if (empty)
   2829 		return class_domain;
   2830 
   2831 	unroll_domain = isl_set_copy(domains->option[isl_ast_loop_unroll]);
   2832 	unroll_list = isl_basic_set_list_from_set(unroll_domain);
   2833 
   2834 	n = isl_basic_set_list_n_basic_set(unroll_list);
   2835 	if (n < 0)
   2836 		class_domain = isl_set_free(class_domain);
   2837 	for (i = 0; i < n; ++i) {
   2838 		isl_basic_set *bset;
   2839 
   2840 		bset = isl_basic_set_list_get_basic_set(unroll_list, i);
   2841 		unroll_domain = isl_set_from_basic_set(bset);
   2842 		unroll_domain = isl_set_intersect(unroll_domain,
   2843 						    isl_set_copy(class_domain));
   2844 		unroll_domain = isl_set_intersect(unroll_domain,
   2845 					isl_set_copy(domains->schedule_domain));
   2846 
   2847 		empty = isl_set_is_empty(unroll_domain);
   2848 		if (empty >= 0 && empty) {
   2849 			isl_set_free(unroll_domain);
   2850 			continue;
   2851 		}
   2852 
   2853 		class_domain = do_unroll(domains, unroll_domain, class_domain);
   2854 	}
   2855 
   2856 	isl_basic_set_list_free(unroll_list);
   2857 
   2858 	return class_domain;
   2859 }
   2860 
   2861 /* Try and construct a single basic set that includes the intersection of
   2862  * the schedule domain, the atomic option domain and the class domain.
   2863  * Add the resulting basic set(s) to domains->list and remove them
   2864  * from class_domain.  Return the updated class domain.
   2865  *
   2866  * We construct a single domain rather than trying to combine
   2867  * the schedule domains of individual domains because we are working
   2868  * within a single component so that non-overlapping schedule domains
   2869  * should already have been separated.
   2870  * We do however need to make sure that this single domains is a subset
   2871  * of the class domain so that it would not intersect with any other
   2872  * class domains.  This means that we may end up splitting up the atomic
   2873  * domain in case separation classes are being used.
   2874  *
   2875  * "domain" is the intersection of the schedule domain and the class domain,
   2876  * with inner dimensions projected out.
   2877  */
   2878 static __isl_give isl_set *compute_atomic_domain(
   2879 	struct isl_codegen_domains *domains, __isl_take isl_set *class_domain)
   2880 {
   2881 	isl_basic_set *bset;
   2882 	isl_basic_set_list *list;
   2883 	isl_set *domain, *atomic_domain;
   2884 	int empty;
   2885 
   2886 	domain = isl_set_copy(domains->option[isl_ast_loop_atomic]);
   2887 	domain = isl_set_intersect(domain, isl_set_copy(class_domain));
   2888 	domain = isl_set_intersect(domain,
   2889 				isl_set_copy(domains->schedule_domain));
   2890 	empty = isl_set_is_empty(domain);
   2891 	if (empty < 0)
   2892 		class_domain = isl_set_free(class_domain);
   2893 	if (empty) {
   2894 		isl_set_free(domain);
   2895 		return class_domain;
   2896 	}
   2897 
   2898 	domain = isl_ast_build_eliminate(domains->build, domain);
   2899 	domain = isl_set_coalesce_preserve(domain);
   2900 	bset = isl_set_unshifted_simple_hull(domain);
   2901 	domain = isl_set_from_basic_set(bset);
   2902 	atomic_domain = isl_set_copy(domain);
   2903 	domain = isl_set_intersect(domain, isl_set_copy(class_domain));
   2904 	class_domain = isl_set_subtract(class_domain, atomic_domain);
   2905 	domain = isl_set_make_disjoint(domain);
   2906 	list = isl_basic_set_list_from_set(domain);
   2907 	domains->list = isl_basic_set_list_concat(domains->list, list);
   2908 
   2909 	return class_domain;
   2910 }
   2911 
   2912 /* Split up the schedule domain into uniform basic sets,
   2913  * in the sense that each element in a basic set is associated to
   2914  * elements of the same domains, and add the result to domains->list.
   2915  * Do this for that part of the schedule domain that lies in the
   2916  * intersection of "class_domain" and the separate option domain.
   2917  *
   2918  * "class_domain" may or may not include the constraints
   2919  * of the schedule domain, but this does not make a difference
   2920  * since we are going to intersect it with the domain of the inverse schedule.
   2921  * If it includes schedule domain constraints, then they may involve
   2922  * inner dimensions, but we will eliminate them in separation_domain.
   2923  */
   2924 static int compute_separate_domain(struct isl_codegen_domains *domains,
   2925 	__isl_keep isl_set *class_domain)
   2926 {
   2927 	isl_space *space;
   2928 	isl_set *domain;
   2929 	isl_union_map *executed;
   2930 	isl_basic_set_list *list;
   2931 	int empty;
   2932 
   2933 	domain = isl_set_copy(domains->option[isl_ast_loop_separate]);
   2934 	domain = isl_set_intersect(domain, isl_set_copy(class_domain));
   2935 	executed = isl_union_map_copy(domains->executed);
   2936 	executed = isl_union_map_intersect_domain(executed,
   2937 				    isl_union_set_from_set(domain));
   2938 	empty = isl_union_map_is_empty(executed);
   2939 	if (empty < 0 || empty) {
   2940 		isl_union_map_free(executed);
   2941 		return empty < 0 ? -1 : 0;
   2942 	}
   2943 
   2944 	space = isl_set_get_space(class_domain);
   2945 	domain = separate_schedule_domains(space, executed, domains->build);
   2946 
   2947 	list = isl_basic_set_list_from_set(domain);
   2948 	domains->list = isl_basic_set_list_concat(domains->list, list);
   2949 
   2950 	return 0;
   2951 }
   2952 
   2953 /* Split up the domain at the current depth into disjoint
   2954  * basic sets for which code should be generated separately
   2955  * for the given separation class domain.
   2956  *
   2957  * If any separation classes have been defined, then "class_domain"
   2958  * is the domain of the current class and does not refer to inner dimensions.
   2959  * Otherwise, "class_domain" is the universe domain.
   2960  *
   2961  * We first make sure that the class domain is disjoint from
   2962  * previously considered class domains.
   2963  *
   2964  * The separate domains can be computed directly from the "class_domain".
   2965  *
   2966  * The unroll, atomic and remainder domains need the constraints
   2967  * from the schedule domain.
   2968  *
   2969  * For unrolling, the actual schedule domain is needed (with divs that
   2970  * may refer to the current dimension) so that stride detection can be
   2971  * performed.
   2972  *
   2973  * For atomic and remainder domains, inner dimensions and divs involving
   2974  * the current dimensions should be eliminated.
   2975  * In case we are working within a separation class, we need to intersect
   2976  * the result with the current "class_domain" to ensure that the domains
   2977  * are disjoint from those generated from other class domains.
   2978  *
   2979  * The domain that has been made atomic may be larger than specified
   2980  * by the user since it needs to be representable as a single basic set.
   2981  * This possibly larger domain is removed from class_domain by
   2982  * compute_atomic_domain.  It is computed first so that the extended domain
   2983  * would not overlap with any domains computed before.
   2984  * Similary, the unrolled domains may have some constraints removed and
   2985  * may therefore also be larger than specified by the user.
   2986  *
   2987  * If anything is left after handling separate, unroll and atomic,
   2988  * we split it up into basic sets and append the basic sets to domains->list.
   2989  */
   2990 static isl_stat compute_partial_domains(struct isl_codegen_domains *domains,
   2991 	__isl_take isl_set *class_domain)
   2992 {
   2993 	isl_basic_set_list *list;
   2994 	isl_set *domain;
   2995 
   2996 	class_domain = isl_set_subtract(class_domain,
   2997 					isl_set_copy(domains->done));
   2998 	domains->done = isl_set_union(domains->done,
   2999 					isl_set_copy(class_domain));
   3000 
   3001 	class_domain = compute_atomic_domain(domains, class_domain);
   3002 	class_domain = compute_unroll_domains(domains, class_domain);
   3003 
   3004 	domain = isl_set_copy(class_domain);
   3005 
   3006 	if (compute_separate_domain(domains, domain) < 0)
   3007 		goto error;
   3008 	domain = isl_set_subtract(domain,
   3009 			isl_set_copy(domains->option[isl_ast_loop_separate]));
   3010 
   3011 	domain = isl_set_intersect(domain,
   3012 				isl_set_copy(domains->schedule_domain));
   3013 
   3014 	domain = isl_ast_build_eliminate(domains->build, domain);
   3015 	domain = isl_set_intersect(domain, isl_set_copy(class_domain));
   3016 
   3017 	domain = isl_set_coalesce_preserve(domain);
   3018 	domain = isl_set_make_disjoint(domain);
   3019 
   3020 	list = isl_basic_set_list_from_set(domain);
   3021 	domains->list = isl_basic_set_list_concat(domains->list, list);
   3022 
   3023 	isl_set_free(class_domain);
   3024 
   3025 	return isl_stat_ok;
   3026 error:
   3027 	isl_set_free(domain);
   3028 	isl_set_free(class_domain);
   3029 	return isl_stat_error;
   3030 }
   3031 
   3032 /* Split up the domain at the current depth into disjoint
   3033  * basic sets for which code should be generated separately
   3034  * for the separation class identified by "pnt".
   3035  *
   3036  * We extract the corresponding class domain from domains->sep_class,
   3037  * eliminate inner dimensions and pass control to compute_partial_domains.
   3038  */
   3039 static isl_stat compute_class_domains(__isl_take isl_point *pnt, void *user)
   3040 {
   3041 	struct isl_codegen_domains *domains = user;
   3042 	isl_set *class_set;
   3043 	isl_set *domain;
   3044 	int disjoint;
   3045 
   3046 	class_set = isl_set_from_point(pnt);
   3047 	domain = isl_map_domain(isl_map_intersect_range(
   3048 				isl_map_copy(domains->sep_class), class_set));
   3049 	domain = isl_ast_build_compute_gist(domains->build, domain);
   3050 	domain = isl_ast_build_eliminate(domains->build, domain);
   3051 
   3052 	disjoint = isl_set_plain_is_disjoint(domain, domains->schedule_domain);
   3053 	if (disjoint < 0)
   3054 		return isl_stat_error;
   3055 	if (disjoint) {
   3056 		isl_set_free(domain);
   3057 		return isl_stat_ok;
   3058 	}
   3059 
   3060 	return compute_partial_domains(domains, domain);
   3061 }
   3062 
   3063 /* Extract the domains at the current depth that should be atomic,
   3064  * separated or unrolled and store them in option.
   3065  *
   3066  * The domains specified by the user might overlap, so we make
   3067  * them disjoint by subtracting earlier domains from later domains.
   3068  */
   3069 static void compute_domains_init_options(isl_set *option[4],
   3070 	__isl_keep isl_ast_build *build)
   3071 {
   3072 	enum isl_ast_loop_type type, type2;
   3073 	isl_set *unroll;
   3074 
   3075 	for (type = isl_ast_loop_atomic;
   3076 	    type <= isl_ast_loop_separate; ++type) {
   3077 		option[type] = isl_ast_build_get_option_domain(build, type);
   3078 		for (type2 = isl_ast_loop_atomic; type2 < type; ++type2)
   3079 			option[type] = isl_set_subtract(option[type],
   3080 						isl_set_copy(option[type2]));
   3081 	}
   3082 
   3083 	unroll = option[isl_ast_loop_unroll];
   3084 	unroll = isl_set_coalesce(unroll);
   3085 	unroll = isl_set_make_disjoint(unroll);
   3086 	option[isl_ast_loop_unroll] = unroll;
   3087 }
   3088 
   3089 /* Split up the domain at the current depth into disjoint
   3090  * basic sets for which code should be generated separately,
   3091  * based on the user-specified options.
   3092  * Return the list of disjoint basic sets.
   3093  *
   3094  * There are three kinds of domains that we need to keep track of.
   3095  * - the "schedule domain" is the domain of "executed"
   3096  * - the "class domain" is the domain corresponding to the currrent
   3097  *	separation class
   3098  * - the "option domain" is the domain corresponding to one of the options
   3099  *	atomic, unroll or separate
   3100  *
   3101  * We first consider the individial values of the separation classes
   3102  * and split up the domain for each of them separately.
   3103  * Finally, we consider the remainder.  If no separation classes were
   3104  * specified, then we call compute_partial_domains with the universe
   3105  * "class_domain".  Otherwise, we take the "schedule_domain" as "class_domain",
   3106  * with inner dimensions removed.  We do this because we want to
   3107  * avoid computing the complement of the class domains (i.e., the difference
   3108  * between the universe and domains->done).
   3109  */
   3110 static __isl_give isl_basic_set_list *compute_domains(
   3111 	__isl_keep isl_union_map *executed, __isl_keep isl_ast_build *build)
   3112 {
   3113 	struct isl_codegen_domains domains;
   3114 	isl_ctx *ctx;
   3115 	isl_set *domain;
   3116 	isl_union_set *schedule_domain;
   3117 	isl_set *classes;
   3118 	isl_space *space;
   3119 	int n_param;
   3120 	enum isl_ast_loop_type type;
   3121 	isl_bool empty;
   3122 
   3123 	if (!executed)
   3124 		return NULL;
   3125 
   3126 	ctx = isl_union_map_get_ctx(executed);
   3127 	domains.list = isl_basic_set_list_alloc(ctx, 0);
   3128 
   3129 	schedule_domain = isl_union_map_domain(isl_union_map_copy(executed));
   3130 	domain = isl_set_from_union_set(schedule_domain);
   3131 
   3132 	compute_domains_init_options(domains.option, build);
   3133 
   3134 	domains.sep_class = isl_ast_build_get_separation_class(build);
   3135 	classes = isl_map_range(isl_map_copy(domains.sep_class));
   3136 	n_param = isl_set_dim(classes, isl_dim_param);
   3137 	if (n_param < 0)
   3138 		classes = isl_set_free(classes);
   3139 	classes = isl_set_project_out(classes, isl_dim_param, 0, n_param);
   3140 
   3141 	space = isl_set_get_space(domain);
   3142 	domains.build = build;
   3143 	domains.schedule_domain = isl_set_copy(domain);
   3144 	domains.executed = executed;
   3145 	domains.done = isl_set_empty(space);
   3146 
   3147 	if (isl_set_foreach_point(classes, &compute_class_domains, &domains) < 0)
   3148 		domains.list = isl_basic_set_list_free(domains.list);
   3149 	isl_set_free(classes);
   3150 
   3151 	empty = isl_set_is_empty(domains.done);
   3152 	if (empty < 0) {
   3153 		domains.list = isl_basic_set_list_free(domains.list);
   3154 		domain = isl_set_free(domain);
   3155 	} else if (empty) {
   3156 		isl_set_free(domain);
   3157 		domain = isl_set_universe(isl_set_get_space(domains.done));
   3158 	} else {
   3159 		domain = isl_ast_build_eliminate(build, domain);
   3160 	}
   3161 	if (compute_partial_domains(&domains, domain) < 0)
   3162 		domains.list = isl_basic_set_list_free(domains.list);
   3163 
   3164 	isl_set_free(domains.schedule_domain);
   3165 	isl_set_free(domains.done);
   3166 	isl_map_free(domains.sep_class);
   3167 	for (type = isl_ast_loop_atomic; type <= isl_ast_loop_separate; ++type)
   3168 		isl_set_free(domains.option[type]);
   3169 
   3170 	return domains.list;
   3171 }
   3172 
   3173 /* Generate code for a single component, after shifting (if any)
   3174  * has been applied, in case the schedule was specified as a union map.
   3175  *
   3176  * We first split up the domain at the current depth into disjoint
   3177  * basic sets based on the user-specified options.
   3178  * Then we generated code for each of them and concatenate the results.
   3179  */
   3180 static __isl_give isl_ast_graft_list *generate_shifted_component_flat(
   3181 	__isl_take isl_union_map *executed, __isl_take isl_ast_build *build)
   3182 {
   3183 	isl_basic_set_list *domain_list;
   3184 	isl_ast_graft_list *list = NULL;
   3185 
   3186 	domain_list = compute_domains(executed, build);
   3187 	list = generate_parallel_domains(domain_list, executed, build);
   3188 
   3189 	isl_basic_set_list_free(domain_list);
   3190 	isl_union_map_free(executed);
   3191 	isl_ast_build_free(build);
   3192 
   3193 	return list;
   3194 }
   3195 
   3196 /* Generate code for a single component, after shifting (if any)
   3197  * has been applied, in case the schedule was specified as a schedule tree
   3198  * and the separate option was specified.
   3199  *
   3200  * We perform separation on the domain of "executed" and then generate
   3201  * an AST for each of the resulting disjoint basic sets.
   3202  */
   3203 static __isl_give isl_ast_graft_list *generate_shifted_component_tree_separate(
   3204 	__isl_take isl_union_map *executed, __isl_take isl_ast_build *build)
   3205 {
   3206 	isl_space *space;
   3207 	isl_set *domain;
   3208 	isl_basic_set_list *domain_list;
   3209 	isl_ast_graft_list *list;
   3210 
   3211 	space = isl_ast_build_get_space(build, 1);
   3212 	domain = separate_schedule_domains(space,
   3213 					isl_union_map_copy(executed), build);
   3214 	domain_list = isl_basic_set_list_from_set(domain);
   3215 
   3216 	list = generate_parallel_domains(domain_list, executed, build);
   3217 
   3218 	isl_basic_set_list_free(domain_list);
   3219 	isl_union_map_free(executed);
   3220 	isl_ast_build_free(build);
   3221 
   3222 	return list;
   3223 }
   3224 
   3225 /* Internal data structure for generate_shifted_component_tree_unroll.
   3226  *
   3227  * "executed" and "build" are inputs to generate_shifted_component_tree_unroll.
   3228  * "list" collects the constructs grafts.
   3229  */
   3230 struct isl_ast_unroll_tree_data {
   3231 	isl_union_map *executed;
   3232 	isl_ast_build *build;
   3233 	isl_ast_graft_list *list;
   3234 };
   3235 
   3236 /* Initialize data->list to a list of "n" elements.
   3237  */
   3238 static int init_unroll_tree(int n, void *user)
   3239 {
   3240 	struct isl_ast_unroll_tree_data *data = user;
   3241 	isl_ctx *ctx;
   3242 
   3243 	ctx = isl_ast_build_get_ctx(data->build);
   3244 	data->list = isl_ast_graft_list_alloc(ctx, n);
   3245 
   3246 	return 0;
   3247 }
   3248 
   3249 /* Given an iteration of an unrolled domain represented by "bset",
   3250  * generate the corresponding AST and add the result to data->list.
   3251  */
   3252 static int do_unroll_tree_iteration(__isl_take isl_basic_set *bset, void *user)
   3253 {
   3254 	struct isl_ast_unroll_tree_data *data = user;
   3255 
   3256 	data->list = add_node(data->list, isl_union_map_copy(data->executed),
   3257 				bset, isl_ast_build_copy(data->build));
   3258 
   3259 	return 0;
   3260 }
   3261 
   3262 /* Generate code for a single component, after shifting (if any)
   3263  * has been applied, in case the schedule was specified as a schedule tree
   3264  * and the unroll option was specified.
   3265  *
   3266  * We call foreach_iteration to iterate over the individual values and
   3267  * construct and collect the corresponding grafts in do_unroll_tree_iteration.
   3268  */
   3269 static __isl_give isl_ast_graft_list *generate_shifted_component_tree_unroll(
   3270 	__isl_take isl_union_map *executed, __isl_take isl_set *domain,
   3271 	__isl_take isl_ast_build *build)
   3272 {
   3273 	struct isl_ast_unroll_tree_data data = { executed, build, NULL };
   3274 
   3275 	if (foreach_iteration(domain, build, &init_unroll_tree,
   3276 				&do_unroll_tree_iteration, &data) < 0)
   3277 		data.list = isl_ast_graft_list_free(data.list);
   3278 
   3279 	isl_union_map_free(executed);
   3280 	isl_ast_build_free(build);
   3281 
   3282 	return data.list;
   3283 }
   3284 
   3285 /* Does "domain" involve a disjunction that is purely based on
   3286  * constraints involving only outer dimension?
   3287  *
   3288  * In particular, is there a disjunction such that the constraints
   3289  * involving the current and later dimensions are the same over
   3290  * all the disjuncts?
   3291  */
   3292 static isl_bool has_pure_outer_disjunction(__isl_keep isl_set *domain,
   3293 	__isl_keep isl_ast_build *build)
   3294 {
   3295 	isl_basic_set *hull;
   3296 	isl_set *shared, *inner;
   3297 	isl_bool equal;
   3298 	isl_size depth;
   3299 	isl_size n;
   3300 	isl_size dim;
   3301 
   3302 	n = isl_set_n_basic_set(domain);
   3303 	if (n < 0)
   3304 		return isl_bool_error;
   3305 	if (n <= 1)
   3306 		return isl_bool_false;
   3307 	dim = isl_set_dim(domain, isl_dim_set);
   3308 	depth = isl_ast_build_get_depth(build);
   3309 	if (dim < 0 || depth < 0)
   3310 		return isl_bool_error;
   3311 
   3312 	inner = isl_set_copy(domain);
   3313 	inner = isl_set_drop_constraints_not_involving_dims(inner,
   3314 					    isl_dim_set, depth, dim - depth);
   3315 	hull = isl_set_plain_unshifted_simple_hull(isl_set_copy(inner));
   3316 	shared = isl_set_from_basic_set(hull);
   3317 	equal = isl_set_plain_is_equal(inner, shared);
   3318 	isl_set_free(inner);
   3319 	isl_set_free(shared);
   3320 
   3321 	return equal;
   3322 }
   3323 
   3324 /* Generate code for a single component, after shifting (if any)
   3325  * has been applied, in case the schedule was specified as a schedule tree.
   3326  * In particular, handle the base case where there is either no isolated
   3327  * set or we are within the isolated set (in which case "isolated" is set)
   3328  * or the iterations that precede or follow the isolated set.
   3329  *
   3330  * The schedule domain is broken up or combined into basic sets
   3331  * according to the AST generation option specified in the current
   3332  * schedule node, which may be either atomic, separate, unroll or
   3333  * unspecified.  If the option is unspecified, then we currently simply
   3334  * split the schedule domain into disjoint basic sets.
   3335  *
   3336  * In case the separate option is specified, the AST generation is
   3337  * handled by generate_shifted_component_tree_separate.
   3338  * In the other cases, we need the global schedule domain.
   3339  * In the unroll case, the AST generation is then handled by
   3340  * generate_shifted_component_tree_unroll which needs the actual
   3341  * schedule domain (with divs that may refer to the current dimension)
   3342  * so that stride detection can be performed.
   3343  * In the atomic or unspecified case, inner dimensions and divs involving
   3344  * the current dimensions should be eliminated.
   3345  * The result is then either combined into a single basic set or
   3346  * split up into disjoint basic sets.
   3347  * Finally an AST is generated for each basic set and the results are
   3348  * concatenated.
   3349  *
   3350  * If the schedule domain involves a disjunction that is purely based on
   3351  * constraints involving only outer dimension, then it is treated as
   3352  * if atomic was specified.  This ensures that only a single loop
   3353  * is generated instead of a sequence of identical loops with
   3354  * different guards.
   3355  */
   3356 static __isl_give isl_ast_graft_list *generate_shifted_component_tree_base(
   3357 	__isl_take isl_union_map *executed, __isl_take isl_ast_build *build,
   3358 	int isolated)
   3359 {
   3360 	isl_bool outer_disjunction;
   3361 	isl_union_set *schedule_domain;
   3362 	isl_set *domain;
   3363 	isl_basic_set_list *domain_list;
   3364 	isl_ast_graft_list *list;
   3365 	enum isl_ast_loop_type type;
   3366 
   3367 	type = isl_ast_build_get_loop_type(build, isolated);
   3368 	if (type < 0)
   3369 		goto error;
   3370 
   3371 	if (type == isl_ast_loop_separate)
   3372 		return generate_shifted_component_tree_separate(executed,
   3373 								build);
   3374 
   3375 	schedule_domain = isl_union_map_domain(isl_union_map_copy(executed));
   3376 	domain = isl_set_from_union_set(schedule_domain);
   3377 
   3378 	if (type == isl_ast_loop_unroll)
   3379 		return generate_shifted_component_tree_unroll(executed, domain,
   3380 								build);
   3381 
   3382 	domain = isl_ast_build_eliminate(build, domain);
   3383 	domain = isl_set_coalesce_preserve(domain);
   3384 
   3385 	outer_disjunction = has_pure_outer_disjunction(domain, build);
   3386 	if (outer_disjunction < 0)
   3387 		domain = isl_set_free(domain);
   3388 
   3389 	if (outer_disjunction || type == isl_ast_loop_atomic) {
   3390 		isl_basic_set *hull;
   3391 		hull = isl_set_unshifted_simple_hull(domain);
   3392 		domain_list = isl_basic_set_list_from_basic_set(hull);
   3393 	} else {
   3394 		domain = isl_set_make_disjoint(domain);
   3395 		domain_list = isl_basic_set_list_from_set(domain);
   3396 	}
   3397 
   3398 	list = generate_parallel_domains(domain_list, executed, build);
   3399 
   3400 	isl_basic_set_list_free(domain_list);
   3401 	isl_union_map_free(executed);
   3402 	isl_ast_build_free(build);
   3403 
   3404 	return list;
   3405 error:
   3406 	isl_union_map_free(executed);
   3407 	isl_ast_build_free(build);
   3408 	return NULL;
   3409 }
   3410 
   3411 /* Extract out the disjunction imposed by "domain" on the outer
   3412  * schedule dimensions.
   3413  *
   3414  * In particular, remove all inner dimensions from "domain" (including
   3415  * the current dimension) and then remove the constraints that are shared
   3416  * by all disjuncts in the result.
   3417  */
   3418 static __isl_give isl_set *extract_disjunction(__isl_take isl_set *domain,
   3419 	__isl_keep isl_ast_build *build)
   3420 {
   3421 	isl_set *hull;
   3422 	isl_size depth;
   3423 	isl_size dim;
   3424 
   3425 	domain = isl_ast_build_specialize(build, domain);
   3426 	depth = isl_ast_build_get_depth(build);
   3427 	dim = isl_set_dim(domain, isl_dim_set);
   3428 	if (depth < 0 || dim < 0)
   3429 		return isl_set_free(domain);
   3430 	domain = isl_set_eliminate(domain, isl_dim_set, depth, dim - depth);
   3431 	domain = isl_set_remove_unknown_divs(domain);
   3432 	hull = isl_set_copy(domain);
   3433 	hull = isl_set_from_basic_set(isl_set_unshifted_simple_hull(hull));
   3434 	domain = isl_set_gist(domain, hull);
   3435 
   3436 	return domain;
   3437 }
   3438 
   3439 /* Add "guard" to the grafts in "list".
   3440  * "build" is the outer AST build, while "sub_build" includes "guard"
   3441  * in its generated domain.
   3442  *
   3443  * First combine the grafts into a single graft and then add the guard.
   3444  * If the list is empty, or if some error occurred, then simply return
   3445  * the list.
   3446  */
   3447 static __isl_give isl_ast_graft_list *list_add_guard(
   3448 	__isl_take isl_ast_graft_list *list, __isl_keep isl_set *guard,
   3449 	__isl_keep isl_ast_build *build, __isl_keep isl_ast_build *sub_build)
   3450 {
   3451 	isl_ast_graft *graft;
   3452 	isl_size n;
   3453 
   3454 	list = isl_ast_graft_list_fuse(list, sub_build);
   3455 
   3456 	n = isl_ast_graft_list_n_ast_graft(list);
   3457 	if (n < 0)
   3458 		return isl_ast_graft_list_free(list);
   3459 	if (n != 1)
   3460 		return list;
   3461 
   3462 	graft = isl_ast_graft_list_get_ast_graft(list, 0);
   3463 	graft = isl_ast_graft_add_guard(graft, isl_set_copy(guard), build);
   3464 	list = isl_ast_graft_list_set_ast_graft(list, 0, graft);
   3465 
   3466 	return list;
   3467 }
   3468 
   3469 /* Generate code for a single component, after shifting (if any)
   3470  * has been applied, in case the schedule was specified as a schedule tree.
   3471  * In particular, do so for the specified subset of the schedule domain.
   3472  *
   3473  * If we are outside of the isolated part, then "domain" may include
   3474  * a disjunction.  Explicitly generate this disjunction at this point
   3475  * instead of relying on the disjunction getting hoisted back up
   3476  * to this level.
   3477  */
   3478 static __isl_give isl_ast_graft_list *generate_shifted_component_tree_part(
   3479 	__isl_keep isl_union_map *executed, __isl_take isl_set *domain,
   3480 	__isl_keep isl_ast_build *build, int isolated)
   3481 {
   3482 	isl_union_set *uset;
   3483 	isl_ast_graft_list *list;
   3484 	isl_ast_build *sub_build;
   3485 	int empty;
   3486 
   3487 	uset = isl_union_set_from_set(isl_set_copy(domain));
   3488 	executed = isl_union_map_copy(executed);
   3489 	executed = isl_union_map_intersect_domain(executed, uset);
   3490 	empty = isl_union_map_is_empty(executed);
   3491 	if (empty < 0)
   3492 		goto error;
   3493 	if (empty) {
   3494 		isl_ctx *ctx;
   3495 		isl_union_map_free(executed);
   3496 		isl_set_free(domain);
   3497 		ctx = isl_ast_build_get_ctx(build);
   3498 		return isl_ast_graft_list_alloc(ctx, 0);
   3499 	}
   3500 
   3501 	sub_build = isl_ast_build_copy(build);
   3502 	if (!isolated) {
   3503 		domain = extract_disjunction(domain, build);
   3504 		sub_build = isl_ast_build_restrict_generated(sub_build,
   3505 							isl_set_copy(domain));
   3506 	}
   3507 	list = generate_shifted_component_tree_base(executed,
   3508 				isl_ast_build_copy(sub_build), isolated);
   3509 	if (!isolated)
   3510 		list = list_add_guard(list, domain, build, sub_build);
   3511 	isl_ast_build_free(sub_build);
   3512 	isl_set_free(domain);
   3513 	return list;
   3514 error:
   3515 	isl_union_map_free(executed);
   3516 	isl_set_free(domain);
   3517 	return NULL;
   3518 }
   3519 
   3520 /* Generate code for a single component, after shifting (if any)
   3521  * has been applied, in case the schedule was specified as a schedule tree.
   3522  * In particular, do so for the specified sequence of subsets
   3523  * of the schedule domain, "before", "isolated", "after" and "other",
   3524  * where only the "isolated" part is considered to be isolated.
   3525  */
   3526 static __isl_give isl_ast_graft_list *generate_shifted_component_parts(
   3527 	__isl_take isl_union_map *executed, __isl_take isl_set *before,
   3528 	__isl_take isl_set *isolated, __isl_take isl_set *after,
   3529 	__isl_take isl_set *other, __isl_take isl_ast_build *build)
   3530 {
   3531 	isl_ast_graft_list *list, *res;
   3532 
   3533 	res = generate_shifted_component_tree_part(executed, before, build, 0);
   3534 	list = generate_shifted_component_tree_part(executed, isolated,
   3535 						    build, 1);
   3536 	res = isl_ast_graft_list_concat(res, list);
   3537 	list = generate_shifted_component_tree_part(executed, after, build, 0);
   3538 	res = isl_ast_graft_list_concat(res, list);
   3539 	list = generate_shifted_component_tree_part(executed, other, build, 0);
   3540 	res = isl_ast_graft_list_concat(res, list);
   3541 
   3542 	isl_union_map_free(executed);
   3543 	isl_ast_build_free(build);
   3544 
   3545 	return res;
   3546 }
   3547 
   3548 /* Does "set" intersect "first", but not "second"?
   3549  */
   3550 static isl_bool only_intersects_first(__isl_keep isl_set *set,
   3551 	__isl_keep isl_set *first, __isl_keep isl_set *second)
   3552 {
   3553 	isl_bool disjoint;
   3554 
   3555 	disjoint = isl_set_is_disjoint(set, first);
   3556 	if (disjoint < 0)
   3557 		return isl_bool_error;
   3558 	if (disjoint)
   3559 		return isl_bool_false;
   3560 
   3561 	return isl_set_is_disjoint(set, second);
   3562 }
   3563 
   3564 /* Generate code for a single component, after shifting (if any)
   3565  * has been applied, in case the schedule was specified as a schedule tree.
   3566  * In particular, do so in case of isolation where there is
   3567  * only an "isolated" part and an "after" part.
   3568  * "dead1" and "dead2" are freed by this function in order to simplify
   3569  * the caller.
   3570  *
   3571  * The "before" and "other" parts are set to empty sets.
   3572  */
   3573 static __isl_give isl_ast_graft_list *generate_shifted_component_only_after(
   3574 	__isl_take isl_union_map *executed, __isl_take isl_set *isolated,
   3575 	__isl_take isl_set *after, __isl_take isl_ast_build *build,
   3576 	__isl_take isl_set *dead1, __isl_take isl_set *dead2)
   3577 {
   3578 	isl_set *empty;
   3579 
   3580 	empty = isl_set_empty(isl_set_get_space(after));
   3581 	isl_set_free(dead1);
   3582 	isl_set_free(dead2);
   3583 	return generate_shifted_component_parts(executed, isl_set_copy(empty),
   3584 						isolated, after, empty, build);
   3585 }
   3586 
   3587 /* Generate code for a single component, after shifting (if any)
   3588  * has been applied, in case the schedule was specified as a schedule tree.
   3589  *
   3590  * We first check if the user has specified an isolated schedule domain
   3591  * and that we are not already outside of this isolated schedule domain.
   3592  * If so, we break up the schedule domain into iterations that
   3593  * precede the isolated domain, the isolated domain itself,
   3594  * the iterations that follow the isolated domain and
   3595  * the remaining iterations (those that are incomparable
   3596  * to the isolated domain).
   3597  * We generate an AST for each piece and concatenate the results.
   3598  *
   3599  * If the isolated domain is not convex, then it is replaced
   3600  * by a convex superset to ensure that the sets of preceding and
   3601  * following iterations are properly defined and, in particular,
   3602  * that there are no intermediate iterations that do not belong
   3603  * to the isolated domain.
   3604  *
   3605  * In the special case where at least one element of the schedule
   3606  * domain that does not belong to the isolated domain needs
   3607  * to be scheduled after this isolated domain, but none of those
   3608  * elements need to be scheduled before, break up the schedule domain
   3609  * in only two parts, the isolated domain, and a part that will be
   3610  * scheduled after the isolated domain.
   3611  *
   3612  * If no isolated set has been specified, then we generate an
   3613  * AST for the entire inverse schedule.
   3614  */
   3615 static __isl_give isl_ast_graft_list *generate_shifted_component_tree(
   3616 	__isl_take isl_union_map *executed, __isl_take isl_ast_build *build)
   3617 {
   3618 	int i;
   3619 	isl_size depth;
   3620 	int empty, has_isolate;
   3621 	isl_space *space;
   3622 	isl_union_set *schedule_domain;
   3623 	isl_set *domain;
   3624 	isl_basic_set *hull;
   3625 	isl_set *isolated, *before, *after, *test;
   3626 	isl_map *gt, *lt;
   3627 	isl_bool pure;
   3628 
   3629 	build = isl_ast_build_extract_isolated(build);
   3630 	has_isolate = isl_ast_build_has_isolated(build);
   3631 	if (has_isolate < 0)
   3632 		executed = isl_union_map_free(executed);
   3633 	else if (!has_isolate)
   3634 		return generate_shifted_component_tree_base(executed, build, 0);
   3635 
   3636 	schedule_domain = isl_union_map_domain(isl_union_map_copy(executed));
   3637 	domain = isl_set_from_union_set(schedule_domain);
   3638 
   3639 	isolated = isl_ast_build_get_isolated(build);
   3640 	isolated = isl_set_intersect(isolated, isl_set_copy(domain));
   3641 	test = isl_ast_build_specialize(build, isl_set_copy(isolated));
   3642 	empty = isl_set_is_empty(test);
   3643 	isl_set_free(test);
   3644 	if (empty < 0)
   3645 		goto error;
   3646 	if (empty) {
   3647 		isl_set_free(isolated);
   3648 		isl_set_free(domain);
   3649 		return generate_shifted_component_tree_base(executed, build, 0);
   3650 	}
   3651 	depth = isl_ast_build_get_depth(build);
   3652 	if (depth < 0)
   3653 		goto error;
   3654 
   3655 	isolated = isl_ast_build_eliminate(build, isolated);
   3656 	hull = isl_set_unshifted_simple_hull(isolated);
   3657 	isolated = isl_set_from_basic_set(hull);
   3658 
   3659 	space = isl_space_map_from_set(isl_set_get_space(isolated));
   3660 	gt = isl_map_universe(space);
   3661 	for (i = 0; i < depth; ++i)
   3662 		gt = isl_map_equate(gt, isl_dim_in, i, isl_dim_out, i);
   3663 	gt = isl_map_order_gt(gt, isl_dim_in, depth, isl_dim_out, depth);
   3664 	lt = isl_map_reverse(isl_map_copy(gt));
   3665 	before = isl_set_apply(isl_set_copy(isolated), gt);
   3666 	after = isl_set_apply(isl_set_copy(isolated), lt);
   3667 
   3668 	domain = isl_set_subtract(domain, isl_set_copy(isolated));
   3669 	pure = only_intersects_first(domain, after, before);
   3670 	if (pure < 0)
   3671 		executed = isl_union_map_free(executed);
   3672 	else if (pure)
   3673 		return generate_shifted_component_only_after(executed, isolated,
   3674 						domain, build, before, after);
   3675 	domain = isl_set_subtract(domain, isl_set_copy(before));
   3676 	domain = isl_set_subtract(domain, isl_set_copy(after));
   3677 	after = isl_set_subtract(after, isl_set_copy(isolated));
   3678 	after = isl_set_subtract(after, isl_set_copy(before));
   3679 	before = isl_set_subtract(before, isl_set_copy(isolated));
   3680 
   3681 	return generate_shifted_component_parts(executed, before, isolated,
   3682 						after, domain, build);
   3683 error:
   3684 	isl_set_free(domain);
   3685 	isl_set_free(isolated);
   3686 	isl_union_map_free(executed);
   3687 	isl_ast_build_free(build);
   3688 	return NULL;
   3689 }
   3690 
   3691 /* Generate code for a single component, after shifting (if any)
   3692  * has been applied.
   3693  *
   3694  * Call generate_shifted_component_tree or generate_shifted_component_flat
   3695  * depending on whether the schedule was specified as a schedule tree.
   3696  */
   3697 static __isl_give isl_ast_graft_list *generate_shifted_component(
   3698 	__isl_take isl_union_map *executed, __isl_take isl_ast_build *build)
   3699 {
   3700 	if (isl_ast_build_has_schedule_node(build))
   3701 		return generate_shifted_component_tree(executed, build);
   3702 	else
   3703 		return generate_shifted_component_flat(executed, build);
   3704 }
   3705 
   3706 struct isl_set_map_pair {
   3707 	isl_set *set;
   3708 	isl_map *map;
   3709 };
   3710 
   3711 /* Given an array "domain" of isl_set_map_pairs and an array "order"
   3712  * of indices into the "domain" array,
   3713  * return the union of the "map" fields of the elements
   3714  * indexed by the first "n" elements of "order".
   3715  */
   3716 static __isl_give isl_union_map *construct_component_executed(
   3717 	struct isl_set_map_pair *domain, int *order, int n)
   3718 {
   3719 	int i;
   3720 	isl_map *map;
   3721 	isl_union_map *executed;
   3722 
   3723 	map = isl_map_copy(domain[order[0]].map);
   3724 	executed = isl_union_map_from_map(map);
   3725 	for (i = 1; i < n; ++i) {
   3726 		map = isl_map_copy(domain[order[i]].map);
   3727 		executed = isl_union_map_add_map(executed, map);
   3728 	}
   3729 
   3730 	return executed;
   3731 }
   3732 
   3733 /* Generate code for a single component, after shifting (if any)
   3734  * has been applied.
   3735  *
   3736  * The component inverse schedule is specified as the "map" fields
   3737  * of the elements of "domain" indexed by the first "n" elements of "order".
   3738  */
   3739 static __isl_give isl_ast_graft_list *generate_shifted_component_from_list(
   3740 	struct isl_set_map_pair *domain, int *order, int n,
   3741 	__isl_take isl_ast_build *build)
   3742 {
   3743 	isl_union_map *executed;
   3744 
   3745 	executed = construct_component_executed(domain, order, n);
   3746 	return generate_shifted_component(executed, build);
   3747 }
   3748 
   3749 /* Does set dimension "pos" of "set" have an obviously fixed value?
   3750  */
   3751 static int dim_is_fixed(__isl_keep isl_set *set, int pos)
   3752 {
   3753 	int fixed;
   3754 	isl_val *v;
   3755 
   3756 	v = isl_set_plain_get_val_if_fixed(set, isl_dim_set, pos);
   3757 	if (!v)
   3758 		return -1;
   3759 	fixed = !isl_val_is_nan(v);
   3760 	isl_val_free(v);
   3761 
   3762 	return fixed;
   3763 }
   3764 
   3765 /* Given an array "domain" of isl_set_map_pairs and an array "order"
   3766  * of indices into the "domain" array,
   3767  * do all (except for at most one) of the "set" field of the elements
   3768  * indexed by the first "n" elements of "order" have a fixed value
   3769  * at position "depth"?
   3770  */
   3771 static int at_most_one_non_fixed(struct isl_set_map_pair *domain,
   3772 	int *order, int n, int depth)
   3773 {
   3774 	int i;
   3775 	int non_fixed = -1;
   3776 
   3777 	for (i = 0; i < n; ++i) {
   3778 		int f;
   3779 
   3780 		f = dim_is_fixed(domain[order[i]].set, depth);
   3781 		if (f < 0)
   3782 			return -1;
   3783 		if (f)
   3784 			continue;
   3785 		if (non_fixed >= 0)
   3786 			return 0;
   3787 		non_fixed = i;
   3788 	}
   3789 
   3790 	return 1;
   3791 }
   3792 
   3793 /* Given an array "domain" of isl_set_map_pairs and an array "order"
   3794  * of indices into the "domain" array,
   3795  * eliminate the inner dimensions from the "set" field of the elements
   3796  * indexed by the first "n" elements of "order", provided the current
   3797  * dimension does not have a fixed value.
   3798  *
   3799  * Return the index of the first element in "order" with a corresponding
   3800  * "set" field that does not have an (obviously) fixed value.
   3801  */
   3802 static int eliminate_non_fixed(struct isl_set_map_pair *domain,
   3803 	int *order, int n, int depth, __isl_keep isl_ast_build *build)
   3804 {
   3805 	int i;
   3806 	int base = -1;
   3807 
   3808 	for (i = n - 1; i >= 0; --i) {
   3809 		int f;
   3810 		f = dim_is_fixed(domain[order[i]].set, depth);
   3811 		if (f < 0)
   3812 			return -1;
   3813 		if (f)
   3814 			continue;
   3815 		domain[order[i]].set = isl_ast_build_eliminate_inner(build,
   3816 							domain[order[i]].set);
   3817 		base = i;
   3818 	}
   3819 
   3820 	return base;
   3821 }
   3822 
   3823 /* Given an array "domain" of isl_set_map_pairs and an array "order"
   3824  * of indices into the "domain" array,
   3825  * find the element of "domain" (amongst those indexed by the first "n"
   3826  * elements of "order") with the "set" field that has the smallest
   3827  * value for the current iterator.
   3828  *
   3829  * Note that the domain with the smallest value may depend on the parameters
   3830  * and/or outer loop dimension.  Since the result of this function is only
   3831  * used as heuristic, we only make a reasonable attempt at finding the best
   3832  * domain, one that should work in case a single domain provides the smallest
   3833  * value for the current dimension over all values of the parameters
   3834  * and outer dimensions.
   3835  *
   3836  * In particular, we compute the smallest value of the first domain
   3837  * and replace it by that of any later domain if that later domain
   3838  * has a smallest value that is smaller for at least some value
   3839  * of the parameters and outer dimensions.
   3840  */
   3841 static int first_offset(struct isl_set_map_pair *domain, int *order, int n,
   3842 	__isl_keep isl_ast_build *build)
   3843 {
   3844 	int i;
   3845 	isl_map *min_first;
   3846 	int first = 0;
   3847 
   3848 	min_first = isl_ast_build_map_to_iterator(build,
   3849 					isl_set_copy(domain[order[0]].set));
   3850 	min_first = isl_map_lexmin(min_first);
   3851 
   3852 	for (i = 1; i < n; ++i) {
   3853 		isl_map *min, *test;
   3854 		int empty;
   3855 
   3856 		min = isl_ast_build_map_to_iterator(build,
   3857 					isl_set_copy(domain[order[i]].set));
   3858 		min = isl_map_lexmin(min);
   3859 		test = isl_map_copy(min);
   3860 		test = isl_map_apply_domain(isl_map_copy(min_first), test);
   3861 		test = isl_map_order_lt(test, isl_dim_in, 0, isl_dim_out, 0);
   3862 		empty = isl_map_is_empty(test);
   3863 		isl_map_free(test);
   3864 		if (empty >= 0 && !empty) {
   3865 			isl_map_free(min_first);
   3866 			first = i;
   3867 			min_first = min;
   3868 		} else
   3869 			isl_map_free(min);
   3870 
   3871 		if (empty < 0)
   3872 			break;
   3873 	}
   3874 
   3875 	isl_map_free(min_first);
   3876 
   3877 	return i < n ? -1 : first;
   3878 }
   3879 
   3880 /* Construct a shifted inverse schedule based on the original inverse schedule,
   3881  * the stride and the offset.
   3882  *
   3883  * The original inverse schedule is specified as the "map" fields
   3884  * of the elements of "domain" indexed by the first "n" elements of "order".
   3885  *
   3886  * "stride" and "offset" are such that the difference
   3887  * between the values of the current dimension of domain "i"
   3888  * and the values of the current dimension for some reference domain are
   3889  * equal to
   3890  *
   3891  *	stride * integer + offset[i]
   3892  *
   3893  * Moreover, 0 <= offset[i] < stride.
   3894  *
   3895  * For each domain, we create a map
   3896  *
   3897  *	{ [..., j, ...] -> [..., j - offset[i], offset[i], ....] }
   3898  *
   3899  * where j refers to the current dimension and the other dimensions are
   3900  * unchanged, and apply this map to the original schedule domain.
   3901  *
   3902  * For example, for the original schedule
   3903  *
   3904  *	{ A[i] -> [2i]: 0 <= i < 10; B[i] -> [2i+1] : 0 <= i < 10 }
   3905  *
   3906  * and assuming the offset is 0 for the A domain and 1 for the B domain,
   3907  * we apply the mapping
   3908  *
   3909  *	{ [j] -> [j, 0] }
   3910  *
   3911  * to the schedule of the "A" domain and the mapping
   3912  *
   3913  *	{ [j - 1] -> [j, 1] }
   3914  *
   3915  * to the schedule of the "B" domain.
   3916  *
   3917  *
   3918  * Note that after the transformation, the differences between pairs
   3919  * of values of the current dimension over all domains are multiples
   3920  * of stride and that we have therefore exposed the stride.
   3921  *
   3922  *
   3923  * To see that the mapping preserves the lexicographic order,
   3924  * first note that each of the individual maps above preserves the order.
   3925  * If the value of the current iterator is j1 in one domain and j2 in another,
   3926  * then if j1 = j2, we know that the same map is applied to both domains
   3927  * and the order is preserved.
   3928  * Otherwise, let us assume, without loss of generality, that j1 < j2.
   3929  * If c1 >= c2 (with c1 and c2 the corresponding offsets), then
   3930  *
   3931  *	j1 - c1 < j2 - c2
   3932  *
   3933  * and the order is preserved.
   3934  * If c1 < c2, then we know
   3935  *
   3936  *	0 <= c2 - c1 < s
   3937  *
   3938  * We also have
   3939  *
   3940  *	j2 - j1 = n * s + r
   3941  *
   3942  * with n >= 0 and 0 <= r < s.
   3943  * In other words, r = c2 - c1.
   3944  * If n > 0, then
   3945  *
   3946  *	j1 - c1 < j2 - c2
   3947  *
   3948  * If n = 0, then
   3949  *
   3950  *	j1 - c1 = j2 - c2
   3951  *
   3952  * and so
   3953  *
   3954  *	(j1 - c1, c1) << (j2 - c2, c2)
   3955  *
   3956  * with "<<" the lexicographic order, proving that the order is preserved
   3957  * in all cases.
   3958  */
   3959 static __isl_give isl_union_map *construct_shifted_executed(
   3960 	struct isl_set_map_pair *domain, int *order, int n,
   3961 	__isl_keep isl_val *stride, __isl_keep isl_multi_val *offset,
   3962 	__isl_keep isl_ast_build *build)
   3963 {
   3964 	int i;
   3965 	isl_union_map *executed;
   3966 	isl_space *space;
   3967 	isl_map *map;
   3968 	isl_size depth;
   3969 	isl_constraint *c;
   3970 
   3971 	depth = isl_ast_build_get_depth(build);
   3972 	if (depth < 0)
   3973 		return NULL;
   3974 	space = isl_ast_build_get_space(build, 1);
   3975 	executed = isl_union_map_empty(isl_space_copy(space));
   3976 	space = isl_space_map_from_set(space);
   3977 	map = isl_map_identity(isl_space_copy(space));
   3978 	map = isl_map_eliminate(map, isl_dim_out, depth, 1);
   3979 	map = isl_map_insert_dims(map, isl_dim_out, depth + 1, 1);
   3980 	space = isl_space_insert_dims(space, isl_dim_out, depth + 1, 1);
   3981 
   3982 	c = isl_constraint_alloc_equality(isl_local_space_from_space(space));
   3983 	c = isl_constraint_set_coefficient_si(c, isl_dim_in, depth, 1);
   3984 	c = isl_constraint_set_coefficient_si(c, isl_dim_out, depth, -1);
   3985 
   3986 	for (i = 0; i < n; ++i) {
   3987 		isl_map *map_i;
   3988 		isl_val *v;
   3989 
   3990 		v = isl_multi_val_get_val(offset, i);
   3991 		if (!v)
   3992 			break;
   3993 		map_i = isl_map_copy(map);
   3994 		map_i = isl_map_fix_val(map_i, isl_dim_out, depth + 1,
   3995 					isl_val_copy(v));
   3996 		v = isl_val_neg(v);
   3997 		c = isl_constraint_set_constant_val(c, v);
   3998 		map_i = isl_map_add_constraint(map_i, isl_constraint_copy(c));
   3999 
   4000 		map_i = isl_map_apply_domain(isl_map_copy(domain[order[i]].map),
   4001 						map_i);
   4002 		executed = isl_union_map_add_map(executed, map_i);
   4003 	}
   4004 
   4005 	isl_constraint_free(c);
   4006 	isl_map_free(map);
   4007 
   4008 	if (i < n)
   4009 		executed = isl_union_map_free(executed);
   4010 
   4011 	return executed;
   4012 }
   4013 
   4014 /* Generate code for a single component, after exposing the stride,
   4015  * given that the schedule domain is "shifted strided".
   4016  *
   4017  * The component inverse schedule is specified as the "map" fields
   4018  * of the elements of "domain" indexed by the first "n" elements of "order".
   4019  *
   4020  * The schedule domain being "shifted strided" means that the differences
   4021  * between the values of the current dimension of domain "i"
   4022  * and the values of the current dimension for some reference domain are
   4023  * equal to
   4024  *
   4025  *	stride * integer + offset[i]
   4026  *
   4027  * We first look for the domain with the "smallest" value for the current
   4028  * dimension and adjust the offsets such that the offset of the "smallest"
   4029  * domain is equal to zero.  The other offsets are reduced modulo stride.
   4030  *
   4031  * Based on this information, we construct a new inverse schedule in
   4032  * construct_shifted_executed that exposes the stride.
   4033  * Since this involves the introduction of a new schedule dimension,
   4034  * the build needs to be changed accordingly.
   4035  * After computing the AST, the newly introduced dimension needs
   4036  * to be removed again from the list of grafts.  We do this by plugging
   4037  * in a mapping that represents the new schedule domain in terms of the
   4038  * old schedule domain.
   4039  */
   4040 static __isl_give isl_ast_graft_list *generate_shift_component(
   4041 	struct isl_set_map_pair *domain, int *order, int n,
   4042 	__isl_keep isl_val *stride, __isl_keep isl_multi_val *offset,
   4043 	__isl_take isl_ast_build *build)
   4044 {
   4045 	isl_ast_graft_list *list;
   4046 	int first;
   4047 	isl_size depth;
   4048 	isl_val *val;
   4049 	isl_multi_val *mv;
   4050 	isl_space *space;
   4051 	isl_multi_aff *ma, *zero;
   4052 	isl_union_map *executed;
   4053 
   4054 	depth = isl_ast_build_get_depth(build);
   4055 
   4056 	first = first_offset(domain, order, n, build);
   4057 	if (depth < 0 || first < 0)
   4058 		goto error;
   4059 
   4060 	mv = isl_multi_val_copy(offset);
   4061 	val = isl_multi_val_get_val(offset, first);
   4062 	val = isl_val_neg(val);
   4063 	mv = isl_multi_val_add_val(mv, val);
   4064 	mv = isl_multi_val_mod_val(mv, isl_val_copy(stride));
   4065 
   4066 	executed = construct_shifted_executed(domain, order, n, stride, mv,
   4067 						build);
   4068 	space = isl_ast_build_get_space(build, 1);
   4069 	space = isl_space_map_from_set(space);
   4070 	ma = isl_multi_aff_identity(isl_space_copy(space));
   4071 	space = isl_space_from_domain(isl_space_domain(space));
   4072 	space = isl_space_add_dims(space, isl_dim_out, 1);
   4073 	zero = isl_multi_aff_zero(space);
   4074 	ma = isl_multi_aff_range_splice(ma, depth + 1, zero);
   4075 	build = isl_ast_build_insert_dim(build, depth + 1);
   4076 	list = generate_shifted_component(executed, build);
   4077 
   4078 	list = isl_ast_graft_list_preimage_multi_aff(list, ma);
   4079 
   4080 	isl_multi_val_free(mv);
   4081 
   4082 	return list;
   4083 error:
   4084 	isl_ast_build_free(build);
   4085 	return NULL;
   4086 }
   4087 
   4088 /* Does any node in the schedule tree rooted at the current schedule node
   4089  * of "build" depend on outer schedule nodes?
   4090  */
   4091 static int has_anchored_subtree(__isl_keep isl_ast_build *build)
   4092 {
   4093 	isl_schedule_node *node;
   4094 	int dependent = 0;
   4095 
   4096 	node = isl_ast_build_get_schedule_node(build);
   4097 	dependent = isl_schedule_node_is_subtree_anchored(node);
   4098 	isl_schedule_node_free(node);
   4099 
   4100 	return dependent;
   4101 }
   4102 
   4103 /* Generate code for a single component.
   4104  *
   4105  * The component inverse schedule is specified as the "map" fields
   4106  * of the elements of "domain" indexed by the first "n" elements of "order".
   4107  *
   4108  * This function may modify the "set" fields of "domain".
   4109  *
   4110  * Before proceeding with the actual code generation for the component,
   4111  * we first check if there are any "shifted" strides, meaning that
   4112  * the schedule domains of the individual domains are all strided,
   4113  * but that they have different offsets, resulting in the union
   4114  * of schedule domains not being strided anymore.
   4115  *
   4116  * The simplest example is the schedule
   4117  *
   4118  *	{ A[i] -> [2i]: 0 <= i < 10; B[i] -> [2i+1] : 0 <= i < 10 }
   4119  *
   4120  * Both schedule domains are strided, but their union is not.
   4121  * This function detects such cases and then rewrites the schedule to
   4122  *
   4123  *	{ A[i] -> [2i, 0]: 0 <= i < 10; B[i] -> [2i, 1] : 0 <= i < 10 }
   4124  *
   4125  * In the new schedule, the schedule domains have the same offset (modulo
   4126  * the stride), ensuring that the union of schedule domains is also strided.
   4127  *
   4128  *
   4129  * If there is only a single domain in the component, then there is
   4130  * nothing to do.   Similarly, if the current schedule dimension has
   4131  * a fixed value for almost all domains then there is nothing to be done.
   4132  * In particular, we need at least two domains where the current schedule
   4133  * dimension does not have a fixed value.
   4134  * Finally, in case of a schedule map input,
   4135  * if any of the options refer to the current schedule dimension,
   4136  * then we bail out as well.  It would be possible to reformulate the options
   4137  * in terms of the new schedule domain, but that would introduce constraints
   4138  * that separate the domains in the options and that is something we would
   4139  * like to avoid.
   4140  * In the case of a schedule tree input, we bail out if any of
   4141  * the descendants of the current schedule node refer to outer
   4142  * schedule nodes in any way.
   4143  *
   4144  *
   4145  * To see if there is any shifted stride, we look at the differences
   4146  * between the values of the current dimension in pairs of domains
   4147  * for equal values of outer dimensions.  These differences should be
   4148  * of the form
   4149  *
   4150  *	m x + r
   4151  *
   4152  * with "m" the stride and "r" a constant.  Note that we cannot perform
   4153  * this analysis on individual domains as the lower bound in each domain
   4154  * may depend on parameters or outer dimensions and so the current dimension
   4155  * itself may not have a fixed remainder on division by the stride.
   4156  *
   4157  * In particular, we compare the first domain that does not have an
   4158  * obviously fixed value for the current dimension to itself and all
   4159  * other domains and collect the offsets and the gcd of the strides.
   4160  * If the gcd becomes one, then we failed to find shifted strides.
   4161  * If the gcd is zero, then the differences were all fixed, meaning
   4162  * that some domains had non-obviously fixed values for the current dimension.
   4163  * If all the offsets are the same (for those domains that do not have
   4164  * an obviously fixed value for the current dimension), then we do not
   4165  * apply the transformation.
   4166  * If none of the domains were skipped, then there is nothing to do.
   4167  * If some of them were skipped, then if we apply separation, the schedule
   4168  * domain should get split in pieces with a (non-shifted) stride.
   4169  *
   4170  * Otherwise, we apply a shift to expose the stride in
   4171  * generate_shift_component.
   4172  */
   4173 static __isl_give isl_ast_graft_list *generate_component(
   4174 	struct isl_set_map_pair *domain, int *order, int n,
   4175 	__isl_take isl_ast_build *build)
   4176 {
   4177 	int i, d;
   4178 	isl_size depth;
   4179 	isl_ctx *ctx;
   4180 	isl_map *map;
   4181 	isl_set *deltas;
   4182 	isl_val *gcd = NULL;
   4183 	isl_multi_val *mv;
   4184 	int fixed, skip;
   4185 	int base;
   4186 	isl_ast_graft_list *list;
   4187 	int res = 0;
   4188 
   4189 	depth = isl_ast_build_get_depth(build);
   4190 	if (depth < 0)
   4191 		goto error;
   4192 
   4193 	skip = n == 1;
   4194 	if (skip >= 0 && !skip)
   4195 		skip = at_most_one_non_fixed(domain, order, n, depth);
   4196 	if (skip >= 0 && !skip) {
   4197 		if (isl_ast_build_has_schedule_node(build))
   4198 			skip = has_anchored_subtree(build);
   4199 		else
   4200 			skip = isl_ast_build_options_involve_depth(build);
   4201 	}
   4202 	if (skip < 0)
   4203 		goto error;
   4204 	if (skip)
   4205 		return generate_shifted_component_from_list(domain,
   4206 							    order, n, build);
   4207 
   4208 	base = eliminate_non_fixed(domain, order, n, depth, build);
   4209 	if (base < 0)
   4210 		goto error;
   4211 
   4212 	ctx = isl_ast_build_get_ctx(build);
   4213 
   4214 	mv = isl_multi_val_zero(isl_space_set_alloc(ctx, 0, n));
   4215 
   4216 	fixed = 1;
   4217 	for (i = 0; i < n; ++i) {
   4218 		isl_val *r, *m;
   4219 
   4220 		map = isl_map_from_domain_and_range(
   4221 					isl_set_copy(domain[order[base]].set),
   4222 					isl_set_copy(domain[order[i]].set));
   4223 		for (d = 0; d < depth; ++d)
   4224 			map = isl_map_equate(map, isl_dim_in, d,
   4225 						    isl_dim_out, d);
   4226 		deltas = isl_map_deltas(map);
   4227 		res = isl_set_dim_residue_class_val(deltas, depth, &m, &r);
   4228 		isl_set_free(deltas);
   4229 		if (res < 0)
   4230 			break;
   4231 
   4232 		if (i == 0)
   4233 			gcd = m;
   4234 		else
   4235 			gcd = isl_val_gcd(gcd, m);
   4236 		if (isl_val_is_one(gcd)) {
   4237 			isl_val_free(r);
   4238 			break;
   4239 		}
   4240 		mv = isl_multi_val_set_val(mv, i, r);
   4241 
   4242 		res = dim_is_fixed(domain[order[i]].set, depth);
   4243 		if (res < 0)
   4244 			break;
   4245 		if (res)
   4246 			continue;
   4247 
   4248 		if (fixed && i > base) {
   4249 			isl_val *a, *b;
   4250 			a = isl_multi_val_get_val(mv, i);
   4251 			b = isl_multi_val_get_val(mv, base);
   4252 			if (isl_val_ne(a, b))
   4253 				fixed = 0;
   4254 			isl_val_free(a);
   4255 			isl_val_free(b);
   4256 		}
   4257 	}
   4258 
   4259 	if (res < 0 || !gcd) {
   4260 		isl_ast_build_free(build);
   4261 		list = NULL;
   4262 	} else if (i < n || fixed || isl_val_is_zero(gcd)) {
   4263 		list = generate_shifted_component_from_list(domain,
   4264 							    order, n, build);
   4265 	} else {
   4266 		list = generate_shift_component(domain, order, n, gcd, mv,
   4267 						build);
   4268 	}
   4269 
   4270 	isl_val_free(gcd);
   4271 	isl_multi_val_free(mv);
   4272 
   4273 	return list;
   4274 error:
   4275 	isl_ast_build_free(build);
   4276 	return NULL;
   4277 }
   4278 
   4279 /* Store both "map" itself and its domain in the
   4280  * structure pointed to by *next and advance to the next array element.
   4281  */
   4282 static isl_stat extract_domain(__isl_take isl_map *map, void *user)
   4283 {
   4284 	struct isl_set_map_pair **next = user;
   4285 
   4286 	(*next)->map = isl_map_copy(map);
   4287 	(*next)->set = isl_map_domain(map);
   4288 	(*next)++;
   4289 
   4290 	return isl_stat_ok;
   4291 }
   4292 
   4293 static isl_bool after_in_tree(__isl_keep isl_union_map *umap,
   4294 	__isl_keep isl_schedule_node *node);
   4295 
   4296 /* Is any domain element of "umap" scheduled after any of
   4297  * the corresponding image elements by the tree rooted at
   4298  * the child of "node"?
   4299  */
   4300 static isl_bool after_in_child(__isl_keep isl_union_map *umap,
   4301 	__isl_keep isl_schedule_node *node)
   4302 {
   4303 	isl_schedule_node *child;
   4304 	isl_bool after;
   4305 
   4306 	child = isl_schedule_node_get_child(node, 0);
   4307 	after = after_in_tree(umap, child);
   4308 	isl_schedule_node_free(child);
   4309 
   4310 	return after;
   4311 }
   4312 
   4313 /* Is any domain element of "umap" scheduled after any of
   4314  * the corresponding image elements by the tree rooted at
   4315  * the band node "node"?
   4316  *
   4317  * We first check if any domain element is scheduled after any
   4318  * of the corresponding image elements by the band node itself.
   4319  * If not, we restrict "map" to those pairs of element that
   4320  * are scheduled together by the band node and continue with
   4321  * the child of the band node.
   4322  * If there are no such pairs then the map passed to after_in_child
   4323  * will be empty causing it to return 0.
   4324  */
   4325 static isl_bool after_in_band(__isl_keep isl_union_map *umap,
   4326 	__isl_keep isl_schedule_node *node)
   4327 {
   4328 	isl_multi_union_pw_aff *mupa;
   4329 	isl_union_map *partial, *test, *gt, *universe, *umap1, *umap2;
   4330 	isl_union_set *domain, *range;
   4331 	isl_space *space;
   4332 	isl_bool empty;
   4333 	isl_bool after;
   4334 	isl_size n;
   4335 
   4336 	n = isl_schedule_node_band_n_member(node);
   4337 	if (n < 0)
   4338 		return isl_bool_error;
   4339 	if (n == 0)
   4340 		return after_in_child(umap, node);
   4341 
   4342 	mupa = isl_schedule_node_band_get_partial_schedule(node);
   4343 	space = isl_multi_union_pw_aff_get_space(mupa);
   4344 	partial = isl_union_map_from_multi_union_pw_aff(mupa);
   4345 	test = isl_union_map_copy(umap);
   4346 	test = isl_union_map_apply_domain(test, isl_union_map_copy(partial));
   4347 	test = isl_union_map_apply_range(test, isl_union_map_copy(partial));
   4348 	gt = isl_union_map_from_map(isl_map_lex_gt(space));
   4349 	test = isl_union_map_intersect(test, gt);
   4350 	empty = isl_union_map_is_empty(test);
   4351 	isl_union_map_free(test);
   4352 
   4353 	if (empty < 0 || !empty) {
   4354 		isl_union_map_free(partial);
   4355 		return isl_bool_not(empty);
   4356 	}
   4357 
   4358 	universe = isl_union_map_universe(isl_union_map_copy(umap));
   4359 	domain = isl_union_map_domain(isl_union_map_copy(universe));
   4360 	range = isl_union_map_range(universe);
   4361 	umap1 = isl_union_map_copy(partial);
   4362 	umap1 = isl_union_map_intersect_domain(umap1, domain);
   4363 	umap2 = isl_union_map_intersect_domain(partial, range);
   4364 	test = isl_union_map_apply_range(umap1, isl_union_map_reverse(umap2));
   4365 	test = isl_union_map_intersect(test, isl_union_map_copy(umap));
   4366 	after = after_in_child(test, node);
   4367 	isl_union_map_free(test);
   4368 	return after;
   4369 }
   4370 
   4371 /* Is any domain element of "umap" scheduled after any of
   4372  * the corresponding image elements by the tree rooted at
   4373  * the context node "node"?
   4374  *
   4375  * The context constraints apply to the schedule domain,
   4376  * so we cannot apply them directly to "umap", which contains
   4377  * pairs of statement instances.  Instead, we add them
   4378  * to the range of the prefix schedule for both domain and
   4379  * range of "umap".
   4380  */
   4381 static isl_bool after_in_context(__isl_keep isl_union_map *umap,
   4382 	__isl_keep isl_schedule_node *node)
   4383 {
   4384 	isl_union_map *prefix, *universe, *umap1, *umap2;
   4385 	isl_union_set *domain, *range;
   4386 	isl_set *context;
   4387 	isl_bool after;
   4388 
   4389 	umap = isl_union_map_copy(umap);
   4390 	context = isl_schedule_node_context_get_context(node);
   4391 	prefix = isl_schedule_node_get_prefix_schedule_union_map(node);
   4392 	universe = isl_union_map_universe(isl_union_map_copy(umap));
   4393 	domain = isl_union_map_domain(isl_union_map_copy(universe));
   4394 	range = isl_union_map_range(universe);
   4395 	umap1 = isl_union_map_copy(prefix);
   4396 	umap1 = isl_union_map_intersect_domain(umap1, domain);
   4397 	umap2 = isl_union_map_intersect_domain(prefix, range);
   4398 	umap1 = isl_union_map_intersect_range(umap1,
   4399 					    isl_union_set_from_set(context));
   4400 	umap1 = isl_union_map_apply_range(umap1, isl_union_map_reverse(umap2));
   4401 	umap = isl_union_map_intersect(umap, umap1);
   4402 
   4403 	after = after_in_child(umap, node);
   4404 
   4405 	isl_union_map_free(umap);
   4406 
   4407 	return after;
   4408 }
   4409 
   4410 /* Is any domain element of "umap" scheduled after any of
   4411  * the corresponding image elements by the tree rooted at
   4412  * the expansion node "node"?
   4413  *
   4414  * We apply the expansion to domain and range of "umap" and
   4415  * continue with its child.
   4416  */
   4417 static isl_bool after_in_expansion(__isl_keep isl_union_map *umap,
   4418 	__isl_keep isl_schedule_node *node)
   4419 {
   4420 	isl_union_map *expansion;
   4421 	isl_bool after;
   4422 
   4423 	expansion = isl_schedule_node_expansion_get_expansion(node);
   4424 	umap = isl_union_map_copy(umap);
   4425 	umap = isl_union_map_apply_domain(umap, isl_union_map_copy(expansion));
   4426 	umap = isl_union_map_apply_range(umap, expansion);
   4427 
   4428 	after = after_in_child(umap, node);
   4429 
   4430 	isl_union_map_free(umap);
   4431 
   4432 	return after;
   4433 }
   4434 
   4435 /* Is any domain element of "umap" scheduled after any of
   4436  * the corresponding image elements by the tree rooted at
   4437  * the extension node "node"?
   4438  *
   4439  * Since the extension node may add statement instances before or
   4440  * after the pairs of statement instances in "umap", we return isl_bool_true
   4441  * to ensure that these pairs are not broken up.
   4442  */
   4443 static isl_bool after_in_extension(__isl_keep isl_union_map *umap,
   4444 	__isl_keep isl_schedule_node *node)
   4445 {
   4446 	return isl_bool_true;
   4447 }
   4448 
   4449 /* Is any domain element of "umap" scheduled after any of
   4450  * the corresponding image elements by the tree rooted at
   4451  * the filter node "node"?
   4452  *
   4453  * We intersect domain and range of "umap" with the filter and
   4454  * continue with its child.
   4455  */
   4456 static isl_bool after_in_filter(__isl_keep isl_union_map *umap,
   4457 	__isl_keep isl_schedule_node *node)
   4458 {
   4459 	isl_union_set *filter;
   4460 	isl_bool after;
   4461 
   4462 	umap = isl_union_map_copy(umap);
   4463 	filter = isl_schedule_node_filter_get_filter(node);
   4464 	umap = isl_union_map_intersect_domain(umap, isl_union_set_copy(filter));
   4465 	umap = isl_union_map_intersect_range(umap, filter);
   4466 
   4467 	after = after_in_child(umap, node);
   4468 
   4469 	isl_union_map_free(umap);
   4470 
   4471 	return after;
   4472 }
   4473 
   4474 /* Is any domain element of "umap" scheduled after any of
   4475  * the corresponding image elements by the tree rooted at
   4476  * the set node "node"?
   4477  *
   4478  * This is only the case if this condition holds in any
   4479  * of the (filter) children of the set node.
   4480  * In particular, if the domain and the range of "umap"
   4481  * are contained in different children, then the condition
   4482  * does not hold.
   4483  */
   4484 static isl_bool after_in_set(__isl_keep isl_union_map *umap,
   4485 	__isl_keep isl_schedule_node *node)
   4486 {
   4487 	int i;
   4488 	isl_size n;
   4489 
   4490 	n = isl_schedule_node_n_children(node);
   4491 	if (n < 0)
   4492 		return isl_bool_error;
   4493 	for (i = 0; i < n; ++i) {
   4494 		isl_schedule_node *child;
   4495 		isl_bool after;
   4496 
   4497 		child = isl_schedule_node_get_child(node, i);
   4498 		after = after_in_tree(umap, child);
   4499 		isl_schedule_node_free(child);
   4500 
   4501 		if (after < 0 || after)
   4502 			return after;
   4503 	}
   4504 
   4505 	return isl_bool_false;
   4506 }
   4507 
   4508 /* Return the filter of child "i" of "node".
   4509  */
   4510 static __isl_give isl_union_set *child_filter(
   4511 	__isl_keep isl_schedule_node *node, int i)
   4512 {
   4513 	isl_schedule_node *child;
   4514 	isl_union_set *filter;
   4515 
   4516 	child = isl_schedule_node_get_child(node, i);
   4517 	filter = isl_schedule_node_filter_get_filter(child);
   4518 	isl_schedule_node_free(child);
   4519 
   4520 	return filter;
   4521 }
   4522 
   4523 /* Is any domain element of "umap" scheduled after any of
   4524  * the corresponding image elements by the tree rooted at
   4525  * the sequence node "node"?
   4526  *
   4527  * This happens in particular if any domain element is
   4528  * contained in a later child than one containing a range element or
   4529  * if the condition holds within a given child in the sequence.
   4530  * The later part of the condition is checked by after_in_set.
   4531  */
   4532 static isl_bool after_in_sequence(__isl_keep isl_union_map *umap,
   4533 	__isl_keep isl_schedule_node *node)
   4534 {
   4535 	int i, j;
   4536 	isl_size n;
   4537 	isl_union_map *umap_i;
   4538 	isl_bool empty;
   4539 	isl_bool after = isl_bool_false;
   4540 
   4541 	n = isl_schedule_node_n_children(node);
   4542 	if (n < 0)
   4543 		return isl_bool_error;
   4544 	for (i = 1; i < n; ++i) {
   4545 		isl_union_set *filter_i;
   4546 
   4547 		umap_i = isl_union_map_copy(umap);
   4548 		filter_i = child_filter(node, i);
   4549 		umap_i = isl_union_map_intersect_domain(umap_i, filter_i);
   4550 		empty = isl_union_map_is_empty(umap_i);
   4551 		if (empty < 0)
   4552 			goto error;
   4553 		if (empty) {
   4554 			isl_union_map_free(umap_i);
   4555 			continue;
   4556 		}
   4557 
   4558 		for (j = 0; j < i; ++j) {
   4559 			isl_union_set *filter_j;
   4560 			isl_union_map *umap_ij;
   4561 
   4562 			umap_ij = isl_union_map_copy(umap_i);
   4563 			filter_j = child_filter(node, j);
   4564 			umap_ij = isl_union_map_intersect_range(umap_ij,
   4565 								filter_j);
   4566 			empty = isl_union_map_is_empty(umap_ij);
   4567 			isl_union_map_free(umap_ij);
   4568 
   4569 			if (empty < 0)
   4570 				goto error;
   4571 			if (!empty)
   4572 				after = isl_bool_true;
   4573 			if (after)
   4574 				break;
   4575 		}
   4576 
   4577 		isl_union_map_free(umap_i);
   4578 		if (after)
   4579 			break;
   4580 	}
   4581 
   4582 	if (after < 0 || after)
   4583 		return after;
   4584 
   4585 	return after_in_set(umap, node);
   4586 error:
   4587 	isl_union_map_free(umap_i);
   4588 	return isl_bool_error;
   4589 }
   4590 
   4591 /* Is any domain element of "umap" scheduled after any of
   4592  * the corresponding image elements by the tree rooted at "node"?
   4593  *
   4594  * If "umap" is empty, then clearly there is no such element.
   4595  * Otherwise, consider the different types of nodes separately.
   4596  */
   4597 static isl_bool after_in_tree(__isl_keep isl_union_map *umap,
   4598 	__isl_keep isl_schedule_node *node)
   4599 {
   4600 	isl_bool empty;
   4601 	enum isl_schedule_node_type type;
   4602 
   4603 	empty = isl_union_map_is_empty(umap);
   4604 	if (empty < 0)
   4605 		return isl_bool_error;
   4606 	if (empty)
   4607 		return isl_bool_false;
   4608 	if (!node)
   4609 		return isl_bool_error;
   4610 
   4611 	type = isl_schedule_node_get_type(node);
   4612 	switch (type) {
   4613 	case isl_schedule_node_error:
   4614 		return isl_bool_error;
   4615 	case isl_schedule_node_leaf:
   4616 		return isl_bool_false;
   4617 	case isl_schedule_node_band:
   4618 		return after_in_band(umap, node);
   4619 	case isl_schedule_node_domain:
   4620 		isl_die(isl_schedule_node_get_ctx(node), isl_error_internal,
   4621 			"unexpected internal domain node",
   4622 			return isl_bool_error);
   4623 	case isl_schedule_node_context:
   4624 		return after_in_context(umap, node);
   4625 	case isl_schedule_node_expansion:
   4626 		return after_in_expansion(umap, node);
   4627 	case isl_schedule_node_extension:
   4628 		return after_in_extension(umap, node);
   4629 	case isl_schedule_node_filter:
   4630 		return after_in_filter(umap, node);
   4631 	case isl_schedule_node_guard:
   4632 	case isl_schedule_node_mark:
   4633 		return after_in_child(umap, node);
   4634 	case isl_schedule_node_set:
   4635 		return after_in_set(umap, node);
   4636 	case isl_schedule_node_sequence:
   4637 		return after_in_sequence(umap, node);
   4638 	}
   4639 
   4640 	return isl_bool_true;
   4641 }
   4642 
   4643 /* Is any domain element of "map1" scheduled after any domain
   4644  * element of "map2" by the subtree underneath the current band node,
   4645  * while at the same time being scheduled together by the current
   4646  * band node, i.e., by "map1" and "map2?
   4647  *
   4648  * If the child of the current band node is a leaf, then
   4649  * no element can be scheduled after any other element.
   4650  *
   4651  * Otherwise, we construct a relation between domain elements
   4652  * of "map1" and domain elements of "map2" that are scheduled
   4653  * together and then check if the subtree underneath the current
   4654  * band node determines their relative order.
   4655  */
   4656 static isl_bool after_in_subtree(__isl_keep isl_ast_build *build,
   4657 	__isl_keep isl_map *map1, __isl_keep isl_map *map2)
   4658 {
   4659 	isl_schedule_node *node;
   4660 	isl_map *map;
   4661 	isl_union_map *umap;
   4662 	isl_bool after;
   4663 
   4664 	node = isl_ast_build_get_schedule_node(build);
   4665 	if (!node)
   4666 		return isl_bool_error;
   4667 	node = isl_schedule_node_child(node, 0);
   4668 	if (isl_schedule_node_get_type(node) == isl_schedule_node_leaf) {
   4669 		isl_schedule_node_free(node);
   4670 		return isl_bool_false;
   4671 	}
   4672 	map = isl_map_copy(map2);
   4673 	map = isl_map_apply_domain(map, isl_map_copy(map1));
   4674 	umap = isl_union_map_from_map(map);
   4675 	after = after_in_tree(umap, node);
   4676 	isl_union_map_free(umap);
   4677 	isl_schedule_node_free(node);
   4678 	return after;
   4679 }
   4680 
   4681 /* Internal data for any_scheduled_after.
   4682  *
   4683  * "build" is the build in which the AST is constructed.
   4684  * "depth" is the number of loops that have already been generated
   4685  * "group_coscheduled" is a local copy of options->ast_build_group_coscheduled
   4686  * "domain" is an array of set-map pairs corresponding to the different
   4687  * iteration domains.  The set is the schedule domain, i.e., the domain
   4688  * of the inverse schedule, while the map is the inverse schedule itself.
   4689  */
   4690 struct isl_any_scheduled_after_data {
   4691 	isl_ast_build *build;
   4692 	int depth;
   4693 	int group_coscheduled;
   4694 	struct isl_set_map_pair *domain;
   4695 };
   4696 
   4697 /* Is any element of domain "i" scheduled after any element of domain "j"
   4698  * (for a common iteration of the first data->depth loops)?
   4699  *
   4700  * data->domain[i].set contains the domain of the inverse schedule
   4701  * for domain "i", i.e., elements in the schedule domain.
   4702  *
   4703  * If we are inside a band of a schedule tree and there is a pair
   4704  * of elements in the two domains that is schedule together by
   4705  * the current band, then we check if any element of "i" may be schedule
   4706  * after element of "j" by the descendants of the band node.
   4707  *
   4708  * If data->group_coscheduled is set, then we also return 1 if there
   4709  * is any pair of elements in the two domains that are scheduled together.
   4710  */
   4711 static isl_bool any_scheduled_after(int i, int j, void *user)
   4712 {
   4713 	struct isl_any_scheduled_after_data *data = user;
   4714 	isl_size dim = isl_set_dim(data->domain[i].set, isl_dim_set);
   4715 	int pos;
   4716 
   4717 	if (dim < 0)
   4718 		return isl_bool_error;
   4719 
   4720 	for (pos = data->depth; pos < dim; ++pos) {
   4721 		int follows;
   4722 
   4723 		follows = isl_set_follows_at(data->domain[i].set,
   4724 						data->domain[j].set, pos);
   4725 
   4726 		if (follows < -1)
   4727 			return isl_bool_error;
   4728 		if (follows > 0)
   4729 			return isl_bool_true;
   4730 		if (follows < 0)
   4731 			return isl_bool_false;
   4732 	}
   4733 
   4734 	if (isl_ast_build_has_schedule_node(data->build)) {
   4735 		isl_bool after;
   4736 
   4737 		after = after_in_subtree(data->build, data->domain[i].map,
   4738 					    data->domain[j].map);
   4739 		if (after < 0 || after)
   4740 			return after;
   4741 	}
   4742 
   4743 	return isl_bool_ok(data->group_coscheduled);
   4744 }
   4745 
   4746 /* Look for independent components at the current depth and generate code
   4747  * for each component separately.  The resulting lists of grafts are
   4748  * merged in an attempt to combine grafts with identical guards.
   4749  *
   4750  * Code for two domains can be generated separately if all the elements
   4751  * of one domain are scheduled before (or together with) all the elements
   4752  * of the other domain.  We therefore consider the graph with as nodes
   4753  * the domains and an edge between two nodes if any element of the first
   4754  * node is scheduled after any element of the second node.
   4755  * If the ast_build_group_coscheduled is set, then we also add an edge if
   4756  * there is any pair of elements in the two domains that are scheduled
   4757  * together.
   4758  * Code is then generated (by generate_component)
   4759  * for each of the strongly connected components in this graph
   4760  * in their topological order.
   4761  *
   4762  * Since the test is performed on the domain of the inverse schedules of
   4763  * the different domains, we precompute these domains and store
   4764  * them in data.domain.
   4765  */
   4766 static __isl_give isl_ast_graft_list *generate_components(
   4767 	__isl_take isl_union_map *executed, __isl_take isl_ast_build *build)
   4768 {
   4769 	int i;
   4770 	isl_ctx *ctx = isl_ast_build_get_ctx(build);
   4771 	isl_size n = isl_union_map_n_map(executed);
   4772 	isl_size depth;
   4773 	struct isl_any_scheduled_after_data data;
   4774 	struct isl_set_map_pair *next;
   4775 	struct isl_tarjan_graph *g = NULL;
   4776 	isl_ast_graft_list *list = NULL;
   4777 	int n_domain = 0;
   4778 
   4779 	data.domain = NULL;
   4780 	if (n < 0)
   4781 		goto error;
   4782 	data.domain = isl_calloc_array(ctx, struct isl_set_map_pair, n);
   4783 	if (!data.domain)
   4784 		goto error;
   4785 	n_domain = n;
   4786 
   4787 	next = data.domain;
   4788 	if (isl_union_map_foreach_map(executed, &extract_domain, &next) < 0)
   4789 		goto error;
   4790 
   4791 	depth = isl_ast_build_get_depth(build);
   4792 	if (depth < 0)
   4793 		goto error;
   4794 	data.build = build;
   4795 	data.depth = depth;
   4796 	data.group_coscheduled = isl_options_get_ast_build_group_coscheduled(ctx);
   4797 	g = isl_tarjan_graph_init(ctx, n, &any_scheduled_after, &data);
   4798 	if (!g)
   4799 		goto error;
   4800 
   4801 	list = isl_ast_graft_list_alloc(ctx, 0);
   4802 
   4803 	i = 0;
   4804 	while (list && n) {
   4805 		isl_ast_graft_list *list_c;
   4806 		int first = i;
   4807 
   4808 		if (g->order[i] == -1)
   4809 			isl_die(ctx, isl_error_internal, "cannot happen",
   4810 				goto error);
   4811 		++i; --n;
   4812 		while (g->order[i] != -1) {
   4813 			++i; --n;
   4814 		}
   4815 
   4816 		list_c = generate_component(data.domain,
   4817 					    g->order + first, i - first,
   4818 					    isl_ast_build_copy(build));
   4819 		list = isl_ast_graft_list_merge(list, list_c, build);
   4820 
   4821 		++i;
   4822 	}
   4823 
   4824 	if (0)
   4825 error:		list = isl_ast_graft_list_free(list);
   4826 	isl_tarjan_graph_free(g);
   4827 	for (i = 0; i < n_domain; ++i) {
   4828 		isl_map_free(data.domain[i].map);
   4829 		isl_set_free(data.domain[i].set);
   4830 	}
   4831 	free(data.domain);
   4832 	isl_union_map_free(executed);
   4833 	isl_ast_build_free(build);
   4834 
   4835 	return list;
   4836 }
   4837 
   4838 /* Generate code for the next level (and all inner levels).
   4839  *
   4840  * If "executed" is empty, i.e., no code needs to be generated,
   4841  * then we return an empty list.
   4842  *
   4843  * If we have already generated code for all loop levels, then we pass
   4844  * control to generate_inner_level.
   4845  *
   4846  * If "executed" lives in a single space, i.e., if code needs to be
   4847  * generated for a single domain, then there can only be a single
   4848  * component and we go directly to generate_shifted_component.
   4849  * Otherwise, we call generate_components to detect the components
   4850  * and to call generate_component on each of them separately.
   4851  */
   4852 static __isl_give isl_ast_graft_list *generate_next_level(
   4853 	__isl_take isl_union_map *executed, __isl_take isl_ast_build *build)
   4854 {
   4855 	isl_size depth;
   4856 	isl_size dim;
   4857 	isl_size n;
   4858 
   4859 	if (!build || !executed)
   4860 		goto error;
   4861 
   4862 	if (isl_union_map_is_empty(executed)) {
   4863 		isl_ctx *ctx = isl_ast_build_get_ctx(build);
   4864 		isl_union_map_free(executed);
   4865 		isl_ast_build_free(build);
   4866 		return isl_ast_graft_list_alloc(ctx, 0);
   4867 	}
   4868 
   4869 	depth = isl_ast_build_get_depth(build);
   4870 	dim = isl_ast_build_dim(build, isl_dim_set);
   4871 	if (depth < 0 || dim < 0)
   4872 		goto error;
   4873 	if (depth >= dim)
   4874 		return generate_inner_level(executed, build);
   4875 
   4876 	n = isl_union_map_n_map(executed);
   4877 	if (n < 0)
   4878 		goto error;
   4879 	if (n == 1)
   4880 		return generate_shifted_component(executed, build);
   4881 
   4882 	return generate_components(executed, build);
   4883 error:
   4884 	isl_union_map_free(executed);
   4885 	isl_ast_build_free(build);
   4886 	return NULL;
   4887 }
   4888 
   4889 /* Internal data structure used by isl_ast_build_node_from_schedule_map.
   4890  * internal, executed and build are the inputs to generate_code.
   4891  * list collects the output.
   4892  */
   4893 struct isl_generate_code_data {
   4894 	int internal;
   4895 	isl_union_map *executed;
   4896 	isl_ast_build *build;
   4897 
   4898 	isl_ast_graft_list *list;
   4899 };
   4900 
   4901 /* Given an inverse schedule in terms of the external build schedule, i.e.,
   4902  *
   4903  *	[E -> S] -> D
   4904  *
   4905  * with E the external build schedule and S the additional schedule "space",
   4906  * reformulate the inverse schedule in terms of the internal schedule domain,
   4907  * i.e., return
   4908  *
   4909  *	[I -> S] -> D
   4910  *
   4911  * We first obtain a mapping
   4912  *
   4913  *	I -> E
   4914  *
   4915  * take the inverse and the product with S -> S, resulting in
   4916  *
   4917  *	[I -> S] -> [E -> S]
   4918  *
   4919  * Applying the map to the input produces the desired result.
   4920  */
   4921 static __isl_give isl_union_map *internal_executed(
   4922 	__isl_take isl_union_map *executed, __isl_keep isl_space *space,
   4923 	__isl_keep isl_ast_build *build)
   4924 {
   4925 	isl_map *id, *proj;
   4926 
   4927 	proj = isl_ast_build_get_schedule_map(build);
   4928 	proj = isl_map_reverse(proj);
   4929 	space = isl_space_map_from_set(isl_space_copy(space));
   4930 	id = isl_map_identity(space);
   4931 	proj = isl_map_product(proj, id);
   4932 	executed = isl_union_map_apply_domain(executed,
   4933 						isl_union_map_from_map(proj));
   4934 	return executed;
   4935 }
   4936 
   4937 /* Generate an AST that visits the elements in the range of data->executed
   4938  * in the relative order specified by the corresponding domain element(s)
   4939  * for those domain elements that belong to "set".
   4940  * Add the result to data->list.
   4941  *
   4942  * The caller ensures that "set" is a universe domain.
   4943  * "space" is the space of the additional part of the schedule.
   4944  * It is equal to the space of "set" if build->domain is parametric.
   4945  * Otherwise, it is equal to the range of the wrapped space of "set".
   4946  *
   4947  * If the build space is not parametric and
   4948  * if isl_ast_build_node_from_schedule_map
   4949  * was called from an outside user (data->internal not set), then
   4950  * the (inverse) schedule refers to the external build domain and needs to
   4951  * be transformed to refer to the internal build domain.
   4952  *
   4953  * If the build space is parametric, then we add some of the parameter
   4954  * constraints to the executed relation.  Adding these constraints
   4955  * allows for an earlier detection of conflicts in some cases.
   4956  * However, we do not want to divide the executed relation into
   4957  * more disjuncts than necessary.  We therefore approximate
   4958  * the constraints on the parameters by a single disjunct set.
   4959  *
   4960  * The build is extended to include the additional part of the schedule.
   4961  * If the original build space was not parametric, then the options
   4962  * in data->build refer only to the additional part of the schedule
   4963  * and they need to be adjusted to refer to the complete AST build
   4964  * domain.
   4965  *
   4966  * After having adjusted inverse schedule and build, we start generating
   4967  * code with the outer loop of the current code generation
   4968  * in generate_next_level.
   4969  *
   4970  * If the original build space was not parametric, we undo the embedding
   4971  * on the resulting isl_ast_node_list so that it can be used within
   4972  * the outer AST build.
   4973  */
   4974 static isl_stat generate_code_in_space(struct isl_generate_code_data *data,
   4975 	__isl_take isl_set *set, __isl_take isl_space *space)
   4976 {
   4977 	isl_union_map *executed;
   4978 	isl_ast_build *build;
   4979 	isl_ast_graft_list *list;
   4980 	int embed;
   4981 
   4982 	executed = isl_union_map_copy(data->executed);
   4983 	executed = isl_union_map_intersect_domain(executed,
   4984 						 isl_union_set_from_set(set));
   4985 
   4986 	embed = !isl_set_is_params(data->build->domain);
   4987 	if (embed && !data->internal)
   4988 		executed = internal_executed(executed, space, data->build);
   4989 	if (!embed) {
   4990 		isl_set *domain;
   4991 		domain = isl_ast_build_get_domain(data->build);
   4992 		domain = isl_set_from_basic_set(isl_set_simple_hull(domain));
   4993 		executed = isl_union_map_intersect_params(executed, domain);
   4994 	}
   4995 
   4996 	build = isl_ast_build_copy(data->build);
   4997 	build = isl_ast_build_product(build, space);
   4998 
   4999 	list = generate_next_level(executed, build);
   5000 
   5001 	list = isl_ast_graft_list_unembed(list, embed);
   5002 
   5003 	data->list = isl_ast_graft_list_concat(data->list, list);
   5004 
   5005 	return isl_stat_ok;
   5006 }
   5007 
   5008 /* Generate an AST that visits the elements in the range of data->executed
   5009  * in the relative order specified by the corresponding domain element(s)
   5010  * for those domain elements that belong to "set".
   5011  * Add the result to data->list.
   5012  *
   5013  * The caller ensures that "set" is a universe domain.
   5014  *
   5015  * If the build space S is not parametric, then the space of "set"
   5016  * need to be a wrapped relation with S as domain.  That is, it needs
   5017  * to be of the form
   5018  *
   5019  *	[S -> T]
   5020  *
   5021  * Check this property and pass control to generate_code_in_space
   5022  * passing along T.
   5023  * If the build space is not parametric, then T is the space of "set".
   5024  */
   5025 static isl_stat generate_code_set(__isl_take isl_set *set, void *user)
   5026 {
   5027 	struct isl_generate_code_data *data = user;
   5028 	isl_space *space, *build_space;
   5029 	int is_domain;
   5030 
   5031 	space = isl_set_get_space(set);
   5032 
   5033 	if (isl_set_is_params(data->build->domain))
   5034 		return generate_code_in_space(data, set, space);
   5035 
   5036 	build_space = isl_ast_build_get_space(data->build, data->internal);
   5037 	space = isl_space_unwrap(space);
   5038 	is_domain = isl_space_is_domain(build_space, space);
   5039 	isl_space_free(build_space);
   5040 	space = isl_space_range(space);
   5041 
   5042 	if (is_domain < 0)
   5043 		goto error;
   5044 	if (!is_domain)
   5045 		isl_die(isl_set_get_ctx(set), isl_error_invalid,
   5046 			"invalid nested schedule space", goto error);
   5047 
   5048 	return generate_code_in_space(data, set, space);
   5049 error:
   5050 	isl_set_free(set);
   5051 	isl_space_free(space);
   5052 	return isl_stat_error;
   5053 }
   5054 
   5055 /* Generate an AST that visits the elements in the range of "executed"
   5056  * in the relative order specified by the corresponding domain element(s).
   5057  *
   5058  * "build" is an isl_ast_build that has either been constructed by
   5059  * isl_ast_build_from_context or passed to a callback set by
   5060  * isl_ast_build_set_create_leaf.
   5061  * In the first case, the space of the isl_ast_build is typically
   5062  * a parametric space, although this is currently not enforced.
   5063  * In the second case, the space is never a parametric space.
   5064  * If the space S is not parametric, then the domain space(s) of "executed"
   5065  * need to be wrapped relations with S as domain.
   5066  *
   5067  * If the domain of "executed" consists of several spaces, then an AST
   5068  * is generated for each of them (in arbitrary order) and the results
   5069  * are concatenated.
   5070  *
   5071  * If "internal" is set, then the domain "S" above refers to the internal
   5072  * schedule domain representation.  Otherwise, it refers to the external
   5073  * representation, as returned by isl_ast_build_get_schedule_space.
   5074  *
   5075  * We essentially run over all the spaces in the domain of "executed"
   5076  * and call generate_code_set on each of them.
   5077  */
   5078 static __isl_give isl_ast_graft_list *generate_code(
   5079 	__isl_take isl_union_map *executed, __isl_take isl_ast_build *build,
   5080 	int internal)
   5081 {
   5082 	isl_ctx *ctx;
   5083 	struct isl_generate_code_data data = { 0 };
   5084 	isl_space *space;
   5085 	isl_union_set *schedule_domain;
   5086 	isl_union_map *universe;
   5087 
   5088 	if (!build)
   5089 		goto error;
   5090 	space = isl_ast_build_get_space(build, 1);
   5091 	space = isl_space_align_params(space,
   5092 				    isl_union_map_get_space(executed));
   5093 	space = isl_space_align_params(space,
   5094 				    isl_union_map_get_space(build->options));
   5095 	build = isl_ast_build_align_params(build, isl_space_copy(space));
   5096 	executed = isl_union_map_align_params(executed, space);
   5097 	if (!executed || !build)
   5098 		goto error;
   5099 
   5100 	ctx = isl_ast_build_get_ctx(build);
   5101 
   5102 	data.internal = internal;
   5103 	data.executed = executed;
   5104 	data.build = build;
   5105 	data.list = isl_ast_graft_list_alloc(ctx, 0);
   5106 
   5107 	universe = isl_union_map_universe(isl_union_map_copy(executed));
   5108 	schedule_domain = isl_union_map_domain(universe);
   5109 	if (isl_union_set_foreach_set(schedule_domain, &generate_code_set,
   5110 					&data) < 0)
   5111 		data.list = isl_ast_graft_list_free(data.list);
   5112 
   5113 	isl_union_set_free(schedule_domain);
   5114 	isl_union_map_free(executed);
   5115 
   5116 	isl_ast_build_free(build);
   5117 	return data.list;
   5118 error:
   5119 	isl_union_map_free(executed);
   5120 	isl_ast_build_free(build);
   5121 	return NULL;
   5122 }
   5123 
   5124 /* Generate an AST that visits the elements in the domain of "schedule"
   5125  * in the relative order specified by the corresponding image element(s).
   5126  *
   5127  * "build" is an isl_ast_build that has either been constructed by
   5128  * isl_ast_build_from_context or passed to a callback set by
   5129  * isl_ast_build_set_create_leaf.
   5130  * In the first case, the space of the isl_ast_build is typically
   5131  * a parametric space, although this is currently not enforced.
   5132  * In the second case, the space is never a parametric space.
   5133  * If the space S is not parametric, then the range space(s) of "schedule"
   5134  * need to be wrapped relations with S as domain.
   5135  *
   5136  * If the range of "schedule" consists of several spaces, then an AST
   5137  * is generated for each of them (in arbitrary order) and the results
   5138  * are concatenated.
   5139  *
   5140  * We first initialize the local copies of the relevant options.
   5141  * We do this here rather than when the isl_ast_build is created
   5142  * because the options may have changed between the construction
   5143  * of the isl_ast_build and the call to isl_generate_code.
   5144  *
   5145  * The main computation is performed on an inverse schedule (with
   5146  * the schedule domain in the domain and the elements to be executed
   5147  * in the range) called "executed".
   5148  */
   5149 __isl_give isl_ast_node *isl_ast_build_node_from_schedule_map(
   5150 	__isl_keep isl_ast_build *build, __isl_take isl_union_map *schedule)
   5151 {
   5152 	isl_ast_graft_list *list;
   5153 	isl_ast_node *node;
   5154 	isl_union_map *executed;
   5155 
   5156 	build = isl_ast_build_copy(build);
   5157 	build = isl_ast_build_set_single_valued(build, 0);
   5158 	schedule = isl_union_map_coalesce(schedule);
   5159 	schedule = isl_union_map_remove_redundancies(schedule);
   5160 	executed = isl_union_map_reverse(schedule);
   5161 	list = generate_code(executed, isl_ast_build_copy(build), 0);
   5162 	node = isl_ast_node_from_graft_list(list, build);
   5163 	isl_ast_build_free(build);
   5164 
   5165 	return node;
   5166 }
   5167 
   5168 /* The old name for isl_ast_build_node_from_schedule_map.
   5169  * It is being kept for backward compatibility, but
   5170  * it will be removed in the future.
   5171  */
   5172 __isl_give isl_ast_node *isl_ast_build_ast_from_schedule(
   5173 	__isl_keep isl_ast_build *build, __isl_take isl_union_map *schedule)
   5174 {
   5175 	return isl_ast_build_node_from_schedule_map(build, schedule);
   5176 }
   5177 
   5178 /* Generate an AST that visits the elements in the domain of "executed"
   5179  * in the relative order specified by the leaf node "node".
   5180  *
   5181  * The relation "executed" maps the outer generated loop iterators
   5182  * to the domain elements executed by those iterations.
   5183  *
   5184  * Simply pass control to generate_inner_level.
   5185  * Note that the current build does not refer to any band node, so
   5186  * that generate_inner_level will not try to visit the child of
   5187  * the leaf node.
   5188  *
   5189  * If multiple statement instances reach a leaf,
   5190  * then they can be executed in any order.
   5191  * Group the list of grafts based on shared guards
   5192  * such that identical guards are only generated once
   5193  * when the list is eventually passed on to isl_ast_graft_list_fuse.
   5194  */
   5195 static __isl_give isl_ast_graft_list *build_ast_from_leaf(
   5196 	__isl_take isl_ast_build *build, __isl_take isl_schedule_node *node,
   5197 	__isl_take isl_union_map *executed)
   5198 {
   5199 	isl_ast_graft_list *list;
   5200 
   5201 	isl_schedule_node_free(node);
   5202 	list = generate_inner_level(executed, isl_ast_build_copy(build));
   5203 	list = isl_ast_graft_list_group_on_guard(list, build);
   5204 	isl_ast_build_free(build);
   5205 
   5206 	return list;
   5207 }
   5208 
   5209 /* Check that the band partial schedule "partial" does not filter out
   5210  * any statement instances, as specified by the range of "executed".
   5211  */
   5212 static isl_stat check_band_schedule_total_on_instances(
   5213 	__isl_keep isl_multi_union_pw_aff *partial,
   5214 	__isl_keep isl_union_map *executed)
   5215 {
   5216 	isl_bool subset;
   5217 	isl_union_set *domain, *instances;
   5218 
   5219 	instances = isl_union_map_range(isl_union_map_copy(executed));
   5220 	partial = isl_multi_union_pw_aff_copy(partial);
   5221 	domain = isl_multi_union_pw_aff_domain(partial);
   5222 	subset = isl_union_set_is_subset(instances, domain);
   5223 	isl_union_set_free(domain);
   5224 	isl_union_set_free(instances);
   5225 
   5226 	if (subset < 0)
   5227 		return isl_stat_error;
   5228 	if (!subset)
   5229 		isl_die(isl_union_map_get_ctx(executed), isl_error_invalid,
   5230 			"band node is not allowed to drop statement instances",
   5231 			return isl_stat_error);
   5232 	return isl_stat_ok;
   5233 }
   5234 
   5235 /* Generate an AST that visits the elements in the domain of "executed"
   5236  * in the relative order specified by the band node "node" and its descendants.
   5237  *
   5238  * The relation "executed" maps the outer generated loop iterators
   5239  * to the domain elements executed by those iterations.
   5240  *
   5241  * If the band is empty, we continue with its descendants.
   5242  * Otherwise, we extend the build and the inverse schedule with
   5243  * the additional space/partial schedule and continue generating
   5244  * an AST in generate_next_level.
   5245  * As soon as we have extended the inverse schedule with the additional
   5246  * partial schedule, we look for equalities that may exists between
   5247  * the old and the new part.
   5248  */
   5249 static __isl_give isl_ast_graft_list *build_ast_from_band(
   5250 	__isl_take isl_ast_build *build, __isl_take isl_schedule_node *node,
   5251 	__isl_take isl_union_map *executed)
   5252 {
   5253 	isl_space *space;
   5254 	isl_multi_union_pw_aff *extra;
   5255 	isl_union_map *extra_umap;
   5256 	isl_ast_graft_list *list;
   5257 	isl_size n1, n2;
   5258 	isl_size n;
   5259 
   5260 	n = isl_schedule_node_band_n_member(node);
   5261 	if (!build || n < 0 || !executed)
   5262 		goto error;
   5263 
   5264 	if (n == 0)
   5265 		return build_ast_from_child(build, node, executed);
   5266 
   5267 	extra = isl_schedule_node_band_get_partial_schedule(node);
   5268 	extra = isl_multi_union_pw_aff_align_params(extra,
   5269 				isl_ast_build_get_space(build, 1));
   5270 	space = isl_multi_union_pw_aff_get_space(extra);
   5271 
   5272 	if (check_band_schedule_total_on_instances(extra, executed) < 0)
   5273 		executed = isl_union_map_free(executed);
   5274 
   5275 	extra_umap = isl_union_map_from_multi_union_pw_aff(extra);
   5276 	extra_umap = isl_union_map_reverse(extra_umap);
   5277 
   5278 	executed = isl_union_map_domain_product(executed, extra_umap);
   5279 	executed = isl_union_map_detect_equalities(executed);
   5280 
   5281 	n1 = isl_ast_build_dim(build, isl_dim_param);
   5282 	build = isl_ast_build_product(build, space);
   5283 	n2 = isl_ast_build_dim(build, isl_dim_param);
   5284 	if (n1 < 0 || n2 < 0)
   5285 		build = isl_ast_build_free(build);
   5286 	else if (n2 > n1)
   5287 		isl_die(isl_ast_build_get_ctx(build), isl_error_invalid,
   5288 			"band node is not allowed to introduce new parameters",
   5289 			build = isl_ast_build_free(build));
   5290 	build = isl_ast_build_set_schedule_node(build, node);
   5291 
   5292 	list = generate_next_level(executed, build);
   5293 
   5294 	list = isl_ast_graft_list_unembed(list, 1);
   5295 
   5296 	return list;
   5297 error:
   5298 	isl_schedule_node_free(node);
   5299 	isl_union_map_free(executed);
   5300 	isl_ast_build_free(build);
   5301 	return NULL;
   5302 }
   5303 
   5304 /* Hoist a list of grafts (in practice containing a single graft)
   5305  * from "sub_build" (which includes extra context information)
   5306  * to "build".
   5307  *
   5308  * In particular, project out all additional parameters introduced
   5309  * by the context node from the enforced constraints and the guard
   5310  * of the single graft.
   5311  */
   5312 static __isl_give isl_ast_graft_list *hoist_out_of_context(
   5313 	__isl_take isl_ast_graft_list *list, __isl_keep isl_ast_build *build,
   5314 	__isl_keep isl_ast_build *sub_build)
   5315 {
   5316 	isl_ast_graft *graft;
   5317 	isl_basic_set *enforced;
   5318 	isl_set *guard;
   5319 	isl_size n_param, extra_param;
   5320 
   5321 	n_param = isl_ast_build_dim(build, isl_dim_param);
   5322 	extra_param = isl_ast_build_dim(sub_build, isl_dim_param);
   5323 	if (n_param < 0 || extra_param < 0)
   5324 		return isl_ast_graft_list_free(list);
   5325 
   5326 	if (extra_param == n_param)
   5327 		return list;
   5328 
   5329 	extra_param -= n_param;
   5330 	enforced = isl_ast_graft_list_extract_shared_enforced(list, sub_build);
   5331 	enforced = isl_basic_set_project_out(enforced, isl_dim_param,
   5332 							n_param, extra_param);
   5333 	enforced = isl_basic_set_remove_unknown_divs(enforced);
   5334 	guard = isl_ast_graft_list_extract_hoistable_guard(list, sub_build);
   5335 	guard = isl_set_remove_divs_involving_dims(guard, isl_dim_param,
   5336 							n_param, extra_param);
   5337 	guard = isl_set_project_out(guard, isl_dim_param, n_param, extra_param);
   5338 	guard = isl_set_compute_divs(guard);
   5339 	graft = isl_ast_graft_alloc_from_children(list, guard, enforced,
   5340 							build, sub_build);
   5341 	list = isl_ast_graft_list_from_ast_graft(graft);
   5342 
   5343 	return list;
   5344 }
   5345 
   5346 /* Generate an AST that visits the elements in the domain of "executed"
   5347  * in the relative order specified by the context node "node"
   5348  * and its descendants.
   5349  *
   5350  * The relation "executed" maps the outer generated loop iterators
   5351  * to the domain elements executed by those iterations.
   5352  *
   5353  * The context node may introduce additional parameters as well as
   5354  * constraints on the outer schedule dimensions or original parameters.
   5355  *
   5356  * We add the extra parameters to a new build and the context
   5357  * constraints to both the build and (as a single disjunct)
   5358  * to the domain of "executed".  Since the context constraints
   5359  * are specified in terms of the input schedule, we first need
   5360  * to map them to the internal schedule domain.
   5361  *
   5362  * After constructing the AST from the descendants of "node",
   5363  * we combine the list of grafts into a single graft within
   5364  * the new build, in order to be able to exploit the additional
   5365  * context constraints during this combination.
   5366  *
   5367  * Additionally, if the current node is the outermost node in
   5368  * the schedule tree (apart from the root domain node), we generate
   5369  * all pending guards, again to be able to exploit the additional
   5370  * context constraints.  We currently do not do this for internal
   5371  * context nodes since we may still want to hoist conditions
   5372  * to outer AST nodes.
   5373  *
   5374  * If the context node introduced any new parameters, then they
   5375  * are removed from the set of enforced constraints and guard
   5376  * in hoist_out_of_context.
   5377  */
   5378 static __isl_give isl_ast_graft_list *build_ast_from_context(
   5379 	__isl_take isl_ast_build *build, __isl_take isl_schedule_node *node,
   5380 	__isl_take isl_union_map *executed)
   5381 {
   5382 	isl_set *context;
   5383 	isl_space *space;
   5384 	isl_multi_aff *internal2input;
   5385 	isl_ast_build *sub_build;
   5386 	isl_ast_graft_list *list;
   5387 	isl_size n;
   5388 	isl_size depth;
   5389 
   5390 	depth = isl_schedule_node_get_tree_depth(node);
   5391 	if (depth < 0)
   5392 		build = isl_ast_build_free(build);
   5393 	space = isl_ast_build_get_space(build, 1);
   5394 	context = isl_schedule_node_context_get_context(node);
   5395 	context = isl_set_align_params(context, space);
   5396 	sub_build = isl_ast_build_copy(build);
   5397 	space = isl_set_get_space(context);
   5398 	sub_build = isl_ast_build_align_params(sub_build, space);
   5399 	internal2input = isl_ast_build_get_internal2input(sub_build);
   5400 	context = isl_set_preimage_multi_aff(context, internal2input);
   5401 	sub_build = isl_ast_build_restrict_generated(sub_build,
   5402 					isl_set_copy(context));
   5403 	context = isl_set_from_basic_set(isl_set_simple_hull(context));
   5404 	executed = isl_union_map_intersect_domain(executed,
   5405 					isl_union_set_from_set(context));
   5406 
   5407 	list = build_ast_from_child(isl_ast_build_copy(sub_build),
   5408 						node, executed);
   5409 	n = isl_ast_graft_list_n_ast_graft(list);
   5410 	if (n < 0)
   5411 		list = isl_ast_graft_list_free(list);
   5412 
   5413 	list = isl_ast_graft_list_fuse(list, sub_build);
   5414 	if (depth == 1)
   5415 		list = isl_ast_graft_list_insert_pending_guard_nodes(list,
   5416 								sub_build);
   5417 	if (n >= 1)
   5418 		list = hoist_out_of_context(list, build, sub_build);
   5419 
   5420 	isl_ast_build_free(build);
   5421 	isl_ast_build_free(sub_build);
   5422 
   5423 	return list;
   5424 }
   5425 
   5426 /* Generate an AST that visits the elements in the domain of "executed"
   5427  * in the relative order specified by the expansion node "node" and
   5428  * its descendants.
   5429  *
   5430  * The relation "executed" maps the outer generated loop iterators
   5431  * to the domain elements executed by those iterations.
   5432  *
   5433  * We expand the domain elements by the expansion and
   5434  * continue with the descendants of the node.
   5435  */
   5436 static __isl_give isl_ast_graft_list *build_ast_from_expansion(
   5437 	__isl_take isl_ast_build *build, __isl_take isl_schedule_node *node,
   5438 	__isl_take isl_union_map *executed)
   5439 {
   5440 	isl_union_map *expansion;
   5441 	isl_size n1, n2;
   5442 
   5443 	expansion = isl_schedule_node_expansion_get_expansion(node);
   5444 	expansion = isl_union_map_align_params(expansion,
   5445 				isl_union_map_get_space(executed));
   5446 
   5447 	n1 = isl_union_map_dim(executed, isl_dim_param);
   5448 	executed = isl_union_map_apply_range(executed, expansion);
   5449 	n2 = isl_union_map_dim(executed, isl_dim_param);
   5450 	if (n1 < 0 || n2 < 0)
   5451 		goto error;
   5452 	if (n2 > n1)
   5453 		isl_die(isl_ast_build_get_ctx(build), isl_error_invalid,
   5454 			"expansion node is not allowed to introduce "
   5455 			"new parameters", goto error);
   5456 
   5457 	return build_ast_from_child(build, node, executed);
   5458 error:
   5459 	isl_ast_build_free(build);
   5460 	isl_schedule_node_free(node);
   5461 	isl_union_map_free(executed);
   5462 	return NULL;
   5463 }
   5464 
   5465 /* Generate an AST that visits the elements in the domain of "executed"
   5466  * in the relative order specified by the extension node "node" and
   5467  * its descendants.
   5468  *
   5469  * The relation "executed" maps the outer generated loop iterators
   5470  * to the domain elements executed by those iterations.
   5471  *
   5472  * Extend the inverse schedule with the extension applied to current
   5473  * set of generated constraints.  Since the extension if formulated
   5474  * in terms of the input schedule, it first needs to be transformed
   5475  * to refer to the internal schedule.
   5476  */
   5477 static __isl_give isl_ast_graft_list *build_ast_from_extension(
   5478 	__isl_take isl_ast_build *build, __isl_take isl_schedule_node *node,
   5479 	__isl_take isl_union_map *executed)
   5480 {
   5481 	isl_union_set *schedule_domain;
   5482 	isl_union_map *extension;
   5483 	isl_set *set;
   5484 
   5485 	set = isl_ast_build_get_generated(build);
   5486 	set = isl_set_from_basic_set(isl_set_simple_hull(set));
   5487 	schedule_domain = isl_union_set_from_set(set);
   5488 
   5489 	extension = isl_schedule_node_extension_get_extension(node);
   5490 
   5491 	extension = isl_union_map_preimage_domain_multi_aff(extension,
   5492 			isl_multi_aff_copy(build->internal2input));
   5493 	extension = isl_union_map_intersect_domain(extension, schedule_domain);
   5494 	extension = isl_ast_build_substitute_values_union_map_domain(build,
   5495 								    extension);
   5496 	executed = isl_union_map_union(executed, extension);
   5497 
   5498 	return build_ast_from_child(build, node, executed);
   5499 }
   5500 
   5501 /* Generate an AST that visits the elements in the domain of "executed"
   5502  * in the relative order specified by the filter node "node" and
   5503  * its descendants.
   5504  *
   5505  * The relation "executed" maps the outer generated loop iterators
   5506  * to the domain elements executed by those iterations.
   5507  *
   5508  * We simply intersect the iteration domain (i.e., the range of "executed")
   5509  * with the filter and continue with the descendants of the node,
   5510  * unless the resulting inverse schedule is empty, in which
   5511  * case we return an empty list.
   5512  *
   5513  * If the result of the intersection is equal to the original "executed"
   5514  * relation, then keep the original representation since the intersection
   5515  * may have unnecessarily broken up the relation into a greater number
   5516  * of disjuncts.
   5517  */
   5518 static __isl_give isl_ast_graft_list *build_ast_from_filter(
   5519 	__isl_take isl_ast_build *build, __isl_take isl_schedule_node *node,
   5520 	__isl_take isl_union_map *executed)
   5521 {
   5522 	isl_ctx *ctx;
   5523 	isl_union_set *filter;
   5524 	isl_union_map *orig;
   5525 	isl_ast_graft_list *list;
   5526 	int empty;
   5527 	isl_bool unchanged;
   5528 	isl_size n1, n2;
   5529 
   5530 	orig = isl_union_map_copy(executed);
   5531 	if (!build || !node || !executed)
   5532 		goto error;
   5533 
   5534 	filter = isl_schedule_node_filter_get_filter(node);
   5535 	filter = isl_union_set_align_params(filter,
   5536 				isl_union_map_get_space(executed));
   5537 	n1 = isl_union_map_dim(executed, isl_dim_param);
   5538 	executed = isl_union_map_intersect_range(executed, filter);
   5539 	n2 = isl_union_map_dim(executed, isl_dim_param);
   5540 	if (n1 < 0 || n2 < 0)
   5541 		goto error;
   5542 	if (n2 > n1)
   5543 		isl_die(isl_ast_build_get_ctx(build), isl_error_invalid,
   5544 			"filter node is not allowed to introduce "
   5545 			"new parameters", goto error);
   5546 
   5547 	unchanged = isl_union_map_is_subset(orig, executed);
   5548 	empty = isl_union_map_is_empty(executed);
   5549 	if (unchanged < 0 || empty < 0)
   5550 		goto error;
   5551 	if (unchanged) {
   5552 		isl_union_map_free(executed);
   5553 		return build_ast_from_child(build, node, orig);
   5554 	}
   5555 	isl_union_map_free(orig);
   5556 	if (!empty)
   5557 		return build_ast_from_child(build, node, executed);
   5558 
   5559 	ctx = isl_ast_build_get_ctx(build);
   5560 	list = isl_ast_graft_list_alloc(ctx, 0);
   5561 	isl_ast_build_free(build);
   5562 	isl_schedule_node_free(node);
   5563 	isl_union_map_free(executed);
   5564 	return list;
   5565 error:
   5566 	isl_ast_build_free(build);
   5567 	isl_schedule_node_free(node);
   5568 	isl_union_map_free(executed);
   5569 	isl_union_map_free(orig);
   5570 	return NULL;
   5571 }
   5572 
   5573 /* Generate an AST that visits the elements in the domain of "executed"
   5574  * in the relative order specified by the guard node "node" and
   5575  * its descendants.
   5576  *
   5577  * The relation "executed" maps the outer generated loop iterators
   5578  * to the domain elements executed by those iterations.
   5579  *
   5580  * Ensure that the associated guard is enforced by the outer AST
   5581  * constructs by adding it to the guard of the graft.
   5582  * Since we know that we will enforce the guard, we can also include it
   5583  * in the generated constraints used to construct an AST for
   5584  * the descendant nodes.
   5585  */
   5586 static __isl_give isl_ast_graft_list *build_ast_from_guard(
   5587 	__isl_take isl_ast_build *build, __isl_take isl_schedule_node *node,
   5588 	__isl_take isl_union_map *executed)
   5589 {
   5590 	isl_space *space;
   5591 	isl_set *guard, *hoisted;
   5592 	isl_basic_set *enforced;
   5593 	isl_ast_build *sub_build;
   5594 	isl_ast_graft *graft;
   5595 	isl_ast_graft_list *list;
   5596 	isl_size n1, n2, n;
   5597 
   5598 	space = isl_ast_build_get_space(build, 1);
   5599 	guard = isl_schedule_node_guard_get_guard(node);
   5600 	n1 = isl_space_dim(space, isl_dim_param);
   5601 	guard = isl_set_align_params(guard, space);
   5602 	n2 = isl_set_dim(guard, isl_dim_param);
   5603 	if (n1 < 0 || n2 < 0)
   5604 		guard = isl_set_free(guard);
   5605 	else if (n2 > n1)
   5606 		isl_die(isl_ast_build_get_ctx(build), isl_error_invalid,
   5607 			"guard node is not allowed to introduce "
   5608 			"new parameters", guard = isl_set_free(guard));
   5609 	guard = isl_set_preimage_multi_aff(guard,
   5610 			isl_multi_aff_copy(build->internal2input));
   5611 	guard = isl_ast_build_specialize(build, guard);
   5612 	guard = isl_set_gist(guard, isl_set_copy(build->generated));
   5613 
   5614 	sub_build = isl_ast_build_copy(build);
   5615 	sub_build = isl_ast_build_restrict_generated(sub_build,
   5616 							isl_set_copy(guard));
   5617 
   5618 	list = build_ast_from_child(isl_ast_build_copy(sub_build),
   5619 							node, executed);
   5620 
   5621 	hoisted = isl_ast_graft_list_extract_hoistable_guard(list, sub_build);
   5622 	n = isl_set_n_basic_set(hoisted);
   5623 	if (n < 0)
   5624 		list = isl_ast_graft_list_free(list);
   5625 	if (n > 1)
   5626 		list = isl_ast_graft_list_gist_guards(list,
   5627 						    isl_set_copy(hoisted));
   5628 	guard = isl_set_intersect(guard, hoisted);
   5629 	enforced = extract_shared_enforced(list, build);
   5630 	graft = isl_ast_graft_alloc_from_children(list, guard, enforced,
   5631 						    build, sub_build);
   5632 
   5633 	isl_ast_build_free(sub_build);
   5634 	isl_ast_build_free(build);
   5635 	return isl_ast_graft_list_from_ast_graft(graft);
   5636 }
   5637 
   5638 /* Call the before_each_mark callback, if requested by the user.
   5639  *
   5640  * Return 0 on success and -1 on error.
   5641  *
   5642  * The caller is responsible for recording the current inverse schedule
   5643  * in "build".
   5644  */
   5645 static isl_stat before_each_mark(__isl_keep isl_id *mark,
   5646 	__isl_keep isl_ast_build *build)
   5647 {
   5648 	if (!build)
   5649 		return isl_stat_error;
   5650 	if (!build->before_each_mark)
   5651 		return isl_stat_ok;
   5652 	return build->before_each_mark(mark, build,
   5653 					build->before_each_mark_user);
   5654 }
   5655 
   5656 /* Call the after_each_mark callback, if requested by the user.
   5657  *
   5658  * The caller is responsible for recording the current inverse schedule
   5659  * in "build".
   5660  */
   5661 static __isl_give isl_ast_graft *after_each_mark(
   5662 	__isl_take isl_ast_graft *graft, __isl_keep isl_ast_build *build)
   5663 {
   5664 	if (!graft || !build)
   5665 		return isl_ast_graft_free(graft);
   5666 	if (!build->after_each_mark)
   5667 		return graft;
   5668 	graft->node = build->after_each_mark(graft->node, build,
   5669 						build->after_each_mark_user);
   5670 	if (!graft->node)
   5671 		return isl_ast_graft_free(graft);
   5672 	return graft;
   5673 }
   5674 
   5675 
   5676 /* Generate an AST that visits the elements in the domain of "executed"
   5677  * in the relative order specified by the mark node "node" and
   5678  * its descendants.
   5679  *
   5680  * The relation "executed" maps the outer generated loop iterators
   5681  * to the domain elements executed by those iterations.
   5682 
   5683  * Since we may be calling before_each_mark and after_each_mark
   5684  * callbacks, we record the current inverse schedule in the build.
   5685  *
   5686  * We generate an AST for the child of the mark node, combine
   5687  * the graft list into a single graft and then insert the mark
   5688  * in the AST of that single graft.
   5689  */
   5690 static __isl_give isl_ast_graft_list *build_ast_from_mark(
   5691 	__isl_take isl_ast_build *build, __isl_take isl_schedule_node *node,
   5692 	__isl_take isl_union_map *executed)
   5693 {
   5694 	isl_id *mark;
   5695 	isl_ast_graft *graft;
   5696 	isl_ast_graft_list *list;
   5697 	isl_size n;
   5698 
   5699 	build = isl_ast_build_set_executed(build, isl_union_map_copy(executed));
   5700 
   5701 	mark = isl_schedule_node_mark_get_id(node);
   5702 	if (before_each_mark(mark, build) < 0)
   5703 		node = isl_schedule_node_free(node);
   5704 
   5705 	list = build_ast_from_child(isl_ast_build_copy(build), node, executed);
   5706 	list = isl_ast_graft_list_fuse(list, build);
   5707 	n = isl_ast_graft_list_n_ast_graft(list);
   5708 	if (n < 0)
   5709 		list = isl_ast_graft_list_free(list);
   5710 	if (n == 0) {
   5711 		isl_id_free(mark);
   5712 	} else {
   5713 		graft = isl_ast_graft_list_get_ast_graft(list, 0);
   5714 		graft = isl_ast_graft_insert_mark(graft, mark);
   5715 		graft = after_each_mark(graft, build);
   5716 		list = isl_ast_graft_list_set_ast_graft(list, 0, graft);
   5717 	}
   5718 	isl_ast_build_free(build);
   5719 
   5720 	return list;
   5721 }
   5722 
   5723 static __isl_give isl_ast_graft_list *build_ast_from_schedule_node(
   5724 	__isl_take isl_ast_build *build, __isl_take isl_schedule_node *node,
   5725 	__isl_take isl_union_map *executed);
   5726 
   5727 /* Generate an AST that visits the elements in the domain of "executed"
   5728  * in the relative order specified by the sequence (or set) node "node" and
   5729  * its descendants.
   5730  *
   5731  * The relation "executed" maps the outer generated loop iterators
   5732  * to the domain elements executed by those iterations.
   5733  *
   5734  * We simply generate an AST for each of the children and concatenate
   5735  * the results.
   5736  */
   5737 static __isl_give isl_ast_graft_list *build_ast_from_sequence(
   5738 	__isl_take isl_ast_build *build, __isl_take isl_schedule_node *node,
   5739 	__isl_take isl_union_map *executed)
   5740 {
   5741 	int i;
   5742 	isl_size n;
   5743 	isl_ctx *ctx;
   5744 	isl_ast_graft_list *list;
   5745 
   5746 	ctx = isl_ast_build_get_ctx(build);
   5747 	list = isl_ast_graft_list_alloc(ctx, 0);
   5748 
   5749 	n = isl_schedule_node_n_children(node);
   5750 	if (n < 0)
   5751 		list = isl_ast_graft_list_free(list);
   5752 	for (i = 0; i < n; ++i) {
   5753 		isl_schedule_node *child;
   5754 		isl_ast_graft_list *list_i;
   5755 
   5756 		child = isl_schedule_node_get_child(node, i);
   5757 		list_i = build_ast_from_schedule_node(isl_ast_build_copy(build),
   5758 					child, isl_union_map_copy(executed));
   5759 		list = isl_ast_graft_list_concat(list, list_i);
   5760 	}
   5761 	isl_ast_build_free(build);
   5762 	isl_schedule_node_free(node);
   5763 	isl_union_map_free(executed);
   5764 
   5765 	return list;
   5766 }
   5767 
   5768 /* Generate an AST that visits the elements in the domain of "executed"
   5769  * in the relative order specified by the node "node" and its descendants.
   5770  *
   5771  * The relation "executed" maps the outer generated loop iterators
   5772  * to the domain elements executed by those iterations.
   5773  *
   5774  * The node types are handled in separate functions.
   5775  * Set nodes are currently treated in the same way as sequence nodes.
   5776  * The children of a set node may be executed in any order,
   5777  * including the order of the children.
   5778  */
   5779 static __isl_give isl_ast_graft_list *build_ast_from_schedule_node(
   5780 	__isl_take isl_ast_build *build, __isl_take isl_schedule_node *node,
   5781 	__isl_take isl_union_map *executed)
   5782 {
   5783 	enum isl_schedule_node_type type;
   5784 
   5785 	type = isl_schedule_node_get_type(node);
   5786 
   5787 	switch (type) {
   5788 	case isl_schedule_node_error:
   5789 		goto error;
   5790 	case isl_schedule_node_leaf:
   5791 		return build_ast_from_leaf(build, node, executed);
   5792 	case isl_schedule_node_band:
   5793 		return build_ast_from_band(build, node, executed);
   5794 	case isl_schedule_node_context:
   5795 		return build_ast_from_context(build, node, executed);
   5796 	case isl_schedule_node_domain:
   5797 		isl_die(isl_schedule_node_get_ctx(node), isl_error_unsupported,
   5798 			"unexpected internal domain node", goto error);
   5799 	case isl_schedule_node_expansion:
   5800 		return build_ast_from_expansion(build, node, executed);
   5801 	case isl_schedule_node_extension:
   5802 		return build_ast_from_extension(build, node, executed);
   5803 	case isl_schedule_node_filter:
   5804 		return build_ast_from_filter(build, node, executed);
   5805 	case isl_schedule_node_guard:
   5806 		return build_ast_from_guard(build, node, executed);
   5807 	case isl_schedule_node_mark:
   5808 		return build_ast_from_mark(build, node, executed);
   5809 	case isl_schedule_node_sequence:
   5810 	case isl_schedule_node_set:
   5811 		return build_ast_from_sequence(build, node, executed);
   5812 	}
   5813 
   5814 	isl_die(isl_ast_build_get_ctx(build), isl_error_internal,
   5815 		"unhandled type", goto error);
   5816 error:
   5817 	isl_union_map_free(executed);
   5818 	isl_schedule_node_free(node);
   5819 	isl_ast_build_free(build);
   5820 
   5821 	return NULL;
   5822 }
   5823 
   5824 /* Generate an AST that visits the elements in the domain of "executed"
   5825  * in the relative order specified by the (single) child of "node" and
   5826  * its descendants.
   5827  *
   5828  * The relation "executed" maps the outer generated loop iterators
   5829  * to the domain elements executed by those iterations.
   5830  *
   5831  * This function is never called on a leaf, set or sequence node,
   5832  * so the node always has exactly one child.
   5833  */
   5834 static __isl_give isl_ast_graft_list *build_ast_from_child(
   5835 	__isl_take isl_ast_build *build, __isl_take isl_schedule_node *node,
   5836 	__isl_take isl_union_map *executed)
   5837 {
   5838 	node = isl_schedule_node_child(node, 0);
   5839 	return build_ast_from_schedule_node(build, node, executed);
   5840 }
   5841 
   5842 /* Generate an AST that visits the elements in the domain of the domain
   5843  * node "node" in the relative order specified by its descendants.
   5844  *
   5845  * An initial inverse schedule is created that maps a zero-dimensional
   5846  * schedule space to the node domain.
   5847  * The input "build" is assumed to have a parametric domain and
   5848  * is replaced by the same zero-dimensional schedule space.
   5849  *
   5850  * We also add some of the parameter constraints in the build domain
   5851  * to the executed relation.  Adding these constraints
   5852  * allows for an earlier detection of conflicts in some cases.
   5853  * However, we do not want to divide the executed relation into
   5854  * more disjuncts than necessary.  We therefore approximate
   5855  * the constraints on the parameters by a single disjunct set.
   5856  */
   5857 static __isl_give isl_ast_node *build_ast_from_domain(
   5858 	__isl_take isl_ast_build *build, __isl_take isl_schedule_node *node)
   5859 {
   5860 	isl_ctx *ctx;
   5861 	isl_union_set *domain, *schedule_domain;
   5862 	isl_union_map *executed;
   5863 	isl_space *space;
   5864 	isl_set *set;
   5865 	isl_ast_graft_list *list;
   5866 	isl_ast_node *ast;
   5867 	int is_params;
   5868 
   5869 	if (!build)
   5870 		goto error;
   5871 
   5872 	ctx = isl_ast_build_get_ctx(build);
   5873 	space = isl_ast_build_get_space(build, 1);
   5874 	is_params = isl_space_is_params(space);
   5875 	isl_space_free(space);
   5876 	if (is_params < 0)
   5877 		goto error;
   5878 	if (!is_params)
   5879 		isl_die(ctx, isl_error_unsupported,
   5880 			"expecting parametric initial context", goto error);
   5881 
   5882 	domain = isl_schedule_node_domain_get_domain(node);
   5883 	domain = isl_union_set_coalesce(domain);
   5884 
   5885 	space = isl_union_set_get_space(domain);
   5886 	space = isl_space_set_from_params(space);
   5887 	build = isl_ast_build_product(build, space);
   5888 
   5889 	set = isl_ast_build_get_domain(build);
   5890 	set = isl_set_from_basic_set(isl_set_simple_hull(set));
   5891 	schedule_domain = isl_union_set_from_set(set);
   5892 
   5893 	executed = isl_union_map_from_domain_and_range(schedule_domain, domain);
   5894 	list = build_ast_from_child(isl_ast_build_copy(build), node, executed);
   5895 	ast = isl_ast_node_from_graft_list(list, build);
   5896 	isl_ast_build_free(build);
   5897 
   5898 	return ast;
   5899 error:
   5900 	isl_schedule_node_free(node);
   5901 	isl_ast_build_free(build);
   5902 	return NULL;
   5903 }
   5904 
   5905 /* Generate an AST that visits the elements in the domain of "schedule"
   5906  * in the relative order specified by the schedule tree.
   5907  *
   5908  * "build" is an isl_ast_build that has been created using
   5909  * isl_ast_build_alloc or isl_ast_build_from_context based
   5910  * on a parametric set.
   5911  *
   5912  * The construction starts at the root node of the schedule,
   5913  * which is assumed to be a domain node.
   5914  */
   5915 __isl_give isl_ast_node *isl_ast_build_node_from_schedule(
   5916 	__isl_keep isl_ast_build *build, __isl_take isl_schedule *schedule)
   5917 {
   5918 	isl_ctx *ctx;
   5919 	isl_schedule_node *node;
   5920 
   5921 	if (!build || !schedule)
   5922 		goto error;
   5923 
   5924 	ctx = isl_ast_build_get_ctx(build);
   5925 
   5926 	node = isl_schedule_get_root(schedule);
   5927 	if (!node)
   5928 		goto error;
   5929 	isl_schedule_free(schedule);
   5930 
   5931 	build = isl_ast_build_copy(build);
   5932 	build = isl_ast_build_set_single_valued(build, 0);
   5933 	if (isl_schedule_node_get_type(node) != isl_schedule_node_domain)
   5934 		isl_die(ctx, isl_error_unsupported,
   5935 			"expecting root domain node",
   5936 			build = isl_ast_build_free(build));
   5937 	return build_ast_from_domain(build, node);
   5938 error:
   5939 	isl_schedule_free(schedule);
   5940 	return NULL;
   5941 }
   5942