2 * Copyright 2012-2014 Ecole Normale Superieure
3 * Copyright 2014 INRIA Rocquencourt
5 * Use of this software is governed by the MIT license
7 * Written by Sven Verdoolaege,
8 * Ecole Normale Superieure, 45 rue d’Ulm, 75230 Paris, France
9 * and Inria Paris - Rocquencourt, Domaine de Voluceau - Rocquencourt,
10 * B.P. 105 - 78153 Le Chesnay, France
16 #include <isl/space.h>
18 #include <isl/constraint.h>
21 #include <isl/union_set.h>
22 #include <isl/union_map.h>
23 #include <isl/schedule_node.h>
24 #include <isl/options.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>
32 /* Try and reduce the number of disjuncts in the representation of "set",
33 * without dropping explicit representations of local variables.
35 static __isl_give isl_set
*isl_set_coalesce_preserve(__isl_take isl_set
*set
)
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
);
51 /* Data used in generate_domain.
53 * "build" is the input build.
54 * "list" collects the results.
56 struct isl_generate_domain_data
{
59 isl_ast_graft_list
*list
;
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
,
69 /* Generate an AST for a single domain based on
70 * the (non single valued) inverse schedule "executed".
72 * We extend the schedule with the iteration domain
73 * and continue generating through a call to generate_code.
75 * In particular, if executed has the form
79 * then we continue generating code on
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.
88 static isl_stat
generate_non_single_valued(__isl_take isl_map
*executed
,
89 struct isl_generate_domain_data
*data
)
93 isl_ast_graft_list
*list
;
95 build
= isl_ast_build_copy(data
->build
);
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);
101 list
= generate_code(isl_union_map_from_map(executed
), build
, 1);
103 data
->list
= isl_ast_graft_list_concat(data
->list
, list
);
108 /* Call the at_each_domain callback, if requested by the user,
109 * after recording the current inverse schedule in the build.
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
)
114 if (!graft
|| !build
)
115 return isl_ast_graft_free(graft
);
116 if (!build
->at_each_domain
)
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
)));
123 return isl_ast_graft_free(graft
);
125 graft
->node
= build
->at_each_domain(graft
->node
,
126 build
, build
->at_each_domain_user
);
127 isl_ast_build_free(build
);
130 graft
= isl_ast_graft_free(graft
);
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.
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.
150 * If the user has set an at_each_domain callback, it is called
151 * on the constructed call expression node.
153 static isl_stat
add_domain(__isl_take isl_map
*executed
,
154 __isl_take isl_map
*map
, struct isl_generate_domain_data
*data
)
156 isl_ast_build
*build
;
157 isl_ast_graft
*graft
;
158 isl_ast_graft_list
*list
;
159 isl_set
*guard
, *pending
;
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
);
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
);
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
);
177 list
= isl_ast_graft_list_from_ast_graft(graft
);
178 data
->list
= isl_ast_graft_list_concat(data
->list
, list
);
183 /* Generate an AST for a single domain based on
184 * the inverse schedule "executed" and add it to data->list.
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
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.
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
215 * Otherwise, call add_domain to generate a call expression (with guard) and
216 * to call the at_each_domain callback, if any.
218 static isl_stat
generate_domain(__isl_take isl_map
*executed
, void *user
)
220 struct isl_generate_domain_data
*data
= user
;
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
);
232 isl_map_free(executed
);
236 sv
= isl_map_plain_is_single_valued(executed
);
240 return add_domain(executed
, isl_map_copy(executed
), data
);
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
);
250 if (data
->build
->single_valued
)
251 map
= isl_map_copy(executed
);
253 return generate_non_single_valued(executed
, data
);
256 return add_domain(executed
, map
, data
);
259 isl_map_free(executed
);
260 return isl_stat_error
;
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.
267 * Note that the node returned by the user may be an entire tree.
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.
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.
279 static __isl_give isl_ast_graft_list
*call_create_leaf(
280 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
284 isl_ast_graft
*graft
;
285 isl_ast_build
*user_build
;
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
);
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
);
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
);
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.
311 * If we are inside a band node, then move on to the child of the band.
313 * If the user has specified a create_leaf callback, control
314 * is passed to the user in call_create_leaf.
316 * Otherwise, we generate one or more calls for each individual
317 * domain in generate_domain.
319 static __isl_give isl_ast_graft_list
*generate_inner_level(
320 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
323 struct isl_generate_domain_data data
= { build
};
325 if (!build
|| !executed
)
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
);
335 if (build
->create_leaf
)
336 return call_create_leaf(executed
, build
);
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
);
344 error
: data
.list
= NULL
;
345 isl_ast_build_free(build
);
346 isl_union_map_free(executed
);
350 /* Call the before_each_for callback, if requested by the user.
352 static __isl_give isl_ast_node
*before_each_for(__isl_take isl_ast_node
*node
,
353 __isl_keep isl_ast_build
*build
)
358 return isl_ast_node_free(node
);
359 if (!build
->before_each_for
)
361 id
= build
->before_each_for(build
, build
->before_each_for_user
);
362 node
= isl_ast_node_set_annotation(node
, id
);
366 /* Call the after_each_for callback, if requested by the user.
368 static __isl_give isl_ast_graft
*after_each_for(__isl_take isl_ast_graft
*graft
,
369 __isl_keep isl_ast_build
*build
)
371 if (!graft
|| !build
)
372 return isl_ast_graft_free(graft
);
373 if (!build
->after_each_for
)
375 graft
->node
= build
->after_each_for(graft
->node
, build
,
376 build
->after_each_for_user
);
378 return isl_ast_graft_free(graft
);
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.
388 static __isl_give isl_union_map
*plug_in_values(
389 __isl_take isl_union_map
*executed
, __isl_keep isl_ast_build
*build
)
391 return isl_ast_build_substitute_values_union_map_domain(build
,
395 /* Check if the constraint "c" is a lower bound on dimension "pos",
396 * an upper bound, or independent of dimension "pos".
398 static int constraint_type(isl_constraint
*c
, int pos
)
400 if (isl_constraint_is_lower_bound(c
, isl_dim_set
, pos
))
402 if (isl_constraint_is_upper_bound(c
, isl_dim_set
, pos
))
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".
412 static int cmp_constraint(__isl_keep isl_constraint
*a
,
413 __isl_keep isl_constraint
*b
, void *user
)
416 int t1
= constraint_type(a
, *depth
);
417 int t2
= constraint_type(b
, *depth
);
422 /* Extract a lower bound on dimension "pos" from constraint "c".
424 * If the constraint is of the form
428 * then we essentially return
430 * l = ceil(-f(...)/a)
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
437 * with f the offset and s the stride.
438 * We therefore compute
440 * f + s * ceil((l - f)/s)
442 static __isl_give isl_aff
*lower_bound(__isl_keep isl_constraint
*c
,
443 int pos
, __isl_keep isl_ast_build
*build
)
447 aff
= isl_constraint_get_bound(c
, isl_dim_set
, pos
);
448 aff
= isl_aff_ceil(aff
);
450 if (isl_ast_build_has_stride(build
, pos
)) {
454 offset
= isl_ast_build_get_offset(build
, pos
);
455 stride
= isl_ast_build_get_stride(build
, pos
);
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
);
464 aff
= isl_ast_build_compute_gist_aff(build
, aff
);
469 /* Return the exact lower bound (or upper bound if "upper" is set)
470 * of "domain" as a piecewise affine expression.
472 * If we are computing a lower bound (of a strided dimension), then
473 * we need to make sure it is of the form
477 * where f is the offset and s is the stride.
478 * We therefore need to include the stride constraint before computing
481 static __isl_give isl_pw_aff
*exact_bound(__isl_keep isl_set
*domain
,
482 __isl_keep isl_ast_build
*build
, int upper
)
487 isl_pw_multi_aff
*pma
;
489 domain
= isl_set_copy(domain
);
491 stride
= isl_ast_build_get_stride_constraint(build
);
492 domain
= isl_set_intersect(domain
, stride
);
494 it_map
= isl_ast_build_map_to_iterator(build
, domain
);
496 pma
= isl_map_lexmax_pw_multi_aff(it_map
);
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
);
507 /* Callback for sorting the isl_pw_aff_list passed to reduce_list and
508 * remove_redundant_lower_bounds.
510 static int reduce_list_cmp(__isl_keep isl_pw_aff
*a
, __isl_keep isl_pw_aff
*b
,
513 return isl_pw_aff_plain_cmp(a
, b
);
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".
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
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.
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
)
536 list
= isl_pw_aff_list_sort(list
, &reduce_list_cmp
, NULL
);
538 n
= isl_pw_aff_list_n_pw_aff(list
);
540 return isl_pw_aff_list_free(list
);
544 domain
= isl_ast_build_get_domain(build
);
546 for (i
= n
- 1; i
>= 0; --i
) {
551 domain_i
= isl_set_copy(domain
);
552 pa_i
= isl_pw_aff_list_get_pw_aff(list
, i
);
554 for (j
= 0; j
< n
; ++j
) {
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
);
566 empty
= isl_set_is_empty(domain_i
);
568 isl_set_free(domain_i
);
569 isl_pw_aff_free(pa_i
);
575 list
= isl_pw_aff_list_drop(list
, i
, 1);
579 isl_set_free(domain
);
583 isl_set_free(domain
);
584 return isl_pw_aff_list_free(list
);
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.
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.
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
)
603 isl_pw_aff_list
*list
;
610 n
= isl_constraint_list_n_constraint(constraints
);
615 pa
= exact_bound(domain
, build
, 0);
616 return isl_pw_aff_list_from_pw_aff(pa
);
619 ctx
= isl_ast_build_get_ctx(build
);
620 list
= isl_pw_aff_list_alloc(ctx
,n
);
622 for (i
= 0; i
< n
; ++i
) {
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
));
632 if (isl_ast_build_has_stride(build
, pos
))
633 list
= remove_redundant_lower_bounds(list
, build
);
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.
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
)
648 isl_pw_aff_list
*list
;
652 n
= isl_constraint_list_n_constraint(constraints
);
657 pa
= exact_bound(domain
, build
, 1);
658 return isl_pw_aff_list_from_pw_aff(pa
);
661 ctx
= isl_ast_build_get_ctx(build
);
662 list
= isl_pw_aff_list_alloc(ctx
,n
);
664 for (i
= 0; i
< n
; ++i
) {
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
));
678 /* Return an isl_ast_expr that performs the reduction of type "type"
679 * on AST expressions corresponding to the elements in "list".
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.
687 static __isl_give isl_ast_expr
*reduce_list(enum isl_ast_op_type type
,
688 __isl_keep isl_pw_aff_list
*list
, __isl_keep isl_ast_build
*build
)
695 n
= isl_pw_aff_list_n_pw_aff(list
);
700 return isl_ast_build_expr_from_pw_aff_internal(build
,
701 isl_pw_aff_list_get_pw_aff(list
, 0));
703 ctx
= isl_pw_aff_list_get_ctx(list
);
704 expr
= isl_ast_expr_alloc_op(ctx
, type
, n
);
708 list
= isl_pw_aff_list_copy(list
);
709 list
= isl_pw_aff_list_sort(list
, &reduce_list_cmp
, NULL
);
711 return isl_ast_expr_free(expr
);
713 for (i
= 0; i
< n
; ++i
) {
714 isl_ast_expr
*expr_i
;
716 expr_i
= isl_ast_build_expr_from_pw_aff_internal(build
,
717 isl_pw_aff_list_get_pw_aff(list
, i
));
720 expr
->u
.op
.args
[i
] = expr_i
;
723 isl_pw_aff_list_free(list
);
726 isl_pw_aff_list_free(list
);
727 isl_ast_expr_free(expr
);
731 /* Add guards implied by the "generated constraints",
732 * but not (necessarily) enforced by the generated AST to "guard".
733 * In particular, if there is any stride constraints,
734 * then add the guard implied by those constraints.
735 * If we have generated a degenerate loop, then add the guard
736 * implied by "bounds" on the outer dimensions, i.e., the guard
737 * that ensures that the single value actually exists.
738 * Since there may also be guards implied by a combination
739 * of these constraints, we first combine them before
740 * deriving the implied constraints.
742 static __isl_give isl_set
*add_implied_guards(__isl_take isl_set
*guard
,
743 int degenerate
, __isl_keep isl_basic_set
*bounds
,
744 __isl_keep isl_ast_build
*build
)
746 int depth
, has_stride
;
750 depth
= isl_ast_build_get_depth(build
);
751 has_stride
= isl_ast_build_has_stride(build
, depth
);
752 if (!has_stride
&& !degenerate
)
755 space
= isl_basic_set_get_space(bounds
);
756 dom
= isl_set_universe(space
);
759 bounds
= isl_basic_set_copy(bounds
);
760 bounds
= isl_basic_set_drop_constraints_not_involving_dims(
761 bounds
, isl_dim_set
, depth
, 1);
762 set
= isl_set_from_basic_set(bounds
);
763 dom
= isl_set_intersect(dom
, set
);
767 set
= isl_ast_build_get_stride_constraint(build
);
768 dom
= isl_set_intersect(dom
, set
);
771 dom
= isl_set_eliminate(dom
, isl_dim_set
, depth
, 1);
772 dom
= isl_ast_build_compute_gist(build
, dom
);
773 guard
= isl_set_intersect(guard
, dom
);
778 /* Update "graft" based on "sub_build" for the degenerate case.
780 * "build" is the build in which graft->node was created
781 * "sub_build" contains information about the current level itself,
782 * including the single value attained.
784 * We set the initialization part of the for loop to the single
785 * value attained by the current dimension.
786 * The increment and condition are not strictly needed as the are known
787 * to be "1" and "iterator <= value" respectively.
789 static __isl_give isl_ast_graft
*refine_degenerate(
790 __isl_take isl_ast_graft
*graft
, __isl_keep isl_ast_build
*build
,
791 __isl_keep isl_ast_build
*sub_build
)
795 if (!graft
|| !sub_build
)
796 return isl_ast_graft_free(graft
);
798 value
= isl_pw_aff_copy(sub_build
->value
);
800 graft
->node
->u
.f
.init
= isl_ast_build_expr_from_pw_aff_internal(build
,
802 if (!graft
->node
->u
.f
.init
)
803 return isl_ast_graft_free(graft
);
808 /* Return the intersection of constraints in "list" as a set.
810 static __isl_give isl_set
*intersect_constraints(
811 __isl_keep isl_constraint_list
*list
)
817 n
= isl_constraint_list_n_constraint(list
);
821 isl_die(isl_constraint_list_get_ctx(list
), isl_error_internal
,
822 "expecting at least one constraint", return NULL
);
824 bset
= isl_basic_set_from_constraint(
825 isl_constraint_list_get_constraint(list
, 0));
826 for (i
= 1; i
< n
; ++i
) {
827 isl_basic_set
*bset_i
;
829 bset_i
= isl_basic_set_from_constraint(
830 isl_constraint_list_get_constraint(list
, i
));
831 bset
= isl_basic_set_intersect(bset
, bset_i
);
834 return isl_set_from_basic_set(bset
);
837 /* Compute the constraints on the outer dimensions enforced by
838 * graft->node and add those constraints to graft->enforced,
839 * in case the upper bound is expressed as a set "upper".
841 * In particular, if l(...) is a lower bound in "lower", and
843 * -a i + f(...) >= 0 or a i <= f(...)
845 * is an upper bound ocnstraint on the current dimension i,
846 * then the for loop enforces the constraint
848 * -a l(...) + f(...) >= 0 or a l(...) <= f(...)
850 * We therefore simply take each lower bound in turn, plug it into
851 * the upper bounds and compute the intersection over all lower bounds.
853 * If a lower bound is a rational expression, then
854 * isl_basic_set_preimage_multi_aff will force this rational
855 * expression to have only integer values. However, the loop
856 * itself does not enforce this integrality constraint. We therefore
857 * use the ceil of the lower bounds instead of the lower bounds themselves.
858 * Other constraints will make sure that the for loop is only executed
859 * when each of the lower bounds attains an integral value.
860 * In particular, potentially rational values only occur in
861 * lower_bound if the offset is a (seemingly) rational expression,
862 * but then outer conditions will make sure that this rational expression
863 * only attains integer values.
865 static __isl_give isl_ast_graft
*set_enforced_from_set(
866 __isl_take isl_ast_graft
*graft
,
867 __isl_keep isl_pw_aff_list
*lower
, int pos
, __isl_keep isl_set
*upper
)
870 isl_basic_set
*enforced
;
871 isl_pw_multi_aff
*pma
;
875 n
= isl_pw_aff_list_n_pw_aff(lower
);
877 return isl_ast_graft_free(graft
);
879 space
= isl_set_get_space(upper
);
880 enforced
= isl_basic_set_universe(isl_space_copy(space
));
882 space
= isl_space_map_from_set(space
);
883 pma
= isl_pw_multi_aff_identity(space
);
885 for (i
= 0; i
< n
; ++i
) {
889 isl_pw_multi_aff
*pma_i
;
891 pa
= isl_pw_aff_list_get_pw_aff(lower
, i
);
892 pa
= isl_pw_aff_ceil(pa
);
893 pma_i
= isl_pw_multi_aff_copy(pma
);
894 pma_i
= isl_pw_multi_aff_set_pw_aff(pma_i
, pos
, pa
);
895 enforced_i
= isl_set_copy(upper
);
896 enforced_i
= isl_set_preimage_pw_multi_aff(enforced_i
, pma_i
);
897 hull
= isl_set_simple_hull(enforced_i
);
898 enforced
= isl_basic_set_intersect(enforced
, hull
);
901 isl_pw_multi_aff_free(pma
);
903 graft
= isl_ast_graft_enforce(graft
, enforced
);
908 /* Compute the constraints on the outer dimensions enforced by
909 * graft->node and add those constraints to graft->enforced,
910 * in case the upper bound is expressed as
911 * a list of affine expressions "upper".
913 * The enforced condition is that each lower bound expression is less
914 * than or equal to each upper bound expression.
916 static __isl_give isl_ast_graft
*set_enforced_from_list(
917 __isl_take isl_ast_graft
*graft
,
918 __isl_keep isl_pw_aff_list
*lower
, __isl_keep isl_pw_aff_list
*upper
)
921 isl_basic_set
*enforced
;
923 lower
= isl_pw_aff_list_copy(lower
);
924 upper
= isl_pw_aff_list_copy(upper
);
925 cond
= isl_pw_aff_list_le_set(lower
, upper
);
926 enforced
= isl_set_simple_hull(cond
);
927 graft
= isl_ast_graft_enforce(graft
, enforced
);
932 /* Does "aff" have a negative constant term?
934 static isl_stat
aff_constant_is_negative(__isl_take isl_set
*set
,
935 __isl_take isl_aff
*aff
, void *user
)
940 v
= isl_aff_get_constant_val(aff
);
941 *neg
= isl_val_is_neg(v
);
946 return *neg
? isl_stat_ok
: isl_stat_error
;
949 /* Does "pa" have a negative constant term over its entire domain?
951 static isl_stat
pw_aff_constant_is_negative(__isl_take isl_pw_aff
*pa
,
957 r
= isl_pw_aff_foreach_piece(pa
, &aff_constant_is_negative
, user
);
960 return (*neg
&& r
>= 0) ? isl_stat_ok
: isl_stat_error
;
963 /* Does each element in "list" have a negative constant term?
965 * The callback terminates the iteration as soon an element has been
966 * found that does not have a negative constant term.
968 static int list_constant_is_negative(__isl_keep isl_pw_aff_list
*list
)
972 if (isl_pw_aff_list_foreach(list
,
973 &pw_aff_constant_is_negative
, &neg
) < 0 && neg
)
979 /* Add 1 to each of the elements in "list", where each of these elements
980 * is defined over the internal schedule space of "build".
982 static __isl_give isl_pw_aff_list
*list_add_one(
983 __isl_take isl_pw_aff_list
*list
, __isl_keep isl_ast_build
*build
)
991 n
= isl_pw_aff_list_n_pw_aff(list
);
993 return isl_pw_aff_list_free(list
);
995 space
= isl_ast_build_get_space(build
, 1);
996 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
997 aff
= isl_aff_add_constant_si(aff
, 1);
998 one
= isl_pw_aff_from_aff(aff
);
1000 for (i
= 0; i
< n
; ++i
) {
1002 pa
= isl_pw_aff_list_get_pw_aff(list
, i
);
1003 pa
= isl_pw_aff_add(pa
, isl_pw_aff_copy(one
));
1004 list
= isl_pw_aff_list_set_pw_aff(list
, i
, pa
);
1007 isl_pw_aff_free(one
);
1012 /* Set the condition part of the for node graft->node in case
1013 * the upper bound is represented as a list of piecewise affine expressions.
1015 * In particular, set the condition to
1017 * iterator <= min(list of upper bounds)
1019 * If each of the upper bounds has a negative constant term, then
1020 * set the condition to
1022 * iterator < min(list of (upper bound + 1)s)
1025 static __isl_give isl_ast_graft
*set_for_cond_from_list(
1026 __isl_take isl_ast_graft
*graft
, __isl_keep isl_pw_aff_list
*list
,
1027 __isl_keep isl_ast_build
*build
)
1030 isl_ast_expr
*bound
, *iterator
, *cond
;
1031 enum isl_ast_op_type type
= isl_ast_op_le
;
1033 if (!graft
|| !list
)
1034 return isl_ast_graft_free(graft
);
1036 neg
= list_constant_is_negative(list
);
1038 return isl_ast_graft_free(graft
);
1039 list
= isl_pw_aff_list_copy(list
);
1041 list
= list_add_one(list
, build
);
1042 type
= isl_ast_op_lt
;
1045 bound
= reduce_list(isl_ast_op_min
, list
, build
);
1046 iterator
= isl_ast_expr_copy(graft
->node
->u
.f
.iterator
);
1047 cond
= isl_ast_expr_alloc_binary(type
, iterator
, bound
);
1048 graft
->node
->u
.f
.cond
= cond
;
1050 isl_pw_aff_list_free(list
);
1051 if (!graft
->node
->u
.f
.cond
)
1052 return isl_ast_graft_free(graft
);
1056 /* Set the condition part of the for node graft->node in case
1057 * the upper bound is represented as a set.
1059 static __isl_give isl_ast_graft
*set_for_cond_from_set(
1060 __isl_take isl_ast_graft
*graft
, __isl_keep isl_set
*set
,
1061 __isl_keep isl_ast_build
*build
)
1068 cond
= isl_ast_build_expr_from_set_internal(build
, isl_set_copy(set
));
1069 graft
->node
->u
.f
.cond
= cond
;
1070 if (!graft
->node
->u
.f
.cond
)
1071 return isl_ast_graft_free(graft
);
1075 /* Construct an isl_ast_expr for the increment (i.e., stride) of
1076 * the current dimension.
1078 static __isl_give isl_ast_expr
*for_inc(__isl_keep isl_ast_build
*build
)
1086 ctx
= isl_ast_build_get_ctx(build
);
1087 depth
= isl_ast_build_get_depth(build
);
1089 if (!isl_ast_build_has_stride(build
, depth
))
1090 return isl_ast_expr_alloc_int_si(ctx
, 1);
1092 v
= isl_ast_build_get_stride(build
, depth
);
1093 return isl_ast_expr_from_val(v
);
1096 /* Should we express the loop condition as
1098 * iterator <= min(list of upper bounds)
1100 * or as a conjunction of constraints?
1102 * The first is constructed from a list of upper bounds.
1103 * The second is constructed from a set.
1105 * If there are no upper bounds in "constraints", then this could mean
1106 * that "domain" simply doesn't have an upper bound or that we didn't
1107 * pick any upper bound. In the first case, we want to generate the
1108 * loop condition as a(n empty) conjunction of constraints
1109 * In the second case, we will compute
1110 * a single upper bound from "domain" and so we use the list form.
1112 * If there are upper bounds in "constraints",
1113 * then we use the list form iff the atomic_upper_bound option is set.
1115 static int use_upper_bound_list(isl_ctx
*ctx
, int n_upper
,
1116 __isl_keep isl_set
*domain
, int depth
)
1119 return isl_options_get_ast_build_atomic_upper_bound(ctx
);
1121 return isl_set_dim_has_upper_bound(domain
, isl_dim_set
, depth
);
1124 /* Fill in the expressions of the for node in graft->node.
1127 * - set the initialization part of the loop to the maximum of the lower bounds
1128 * - extract the increment from the stride of the current dimension
1129 * - construct the for condition either based on a list of upper bounds
1130 * or on a set of upper bound constraints.
1132 static __isl_give isl_ast_graft
*set_for_node_expressions(
1133 __isl_take isl_ast_graft
*graft
, __isl_keep isl_pw_aff_list
*lower
,
1134 int use_list
, __isl_keep isl_pw_aff_list
*upper_list
,
1135 __isl_keep isl_set
*upper_set
, __isl_keep isl_ast_build
*build
)
1142 build
= isl_ast_build_copy(build
);
1145 node
->u
.f
.init
= reduce_list(isl_ast_op_max
, lower
, build
);
1146 node
->u
.f
.inc
= for_inc(build
);
1148 if (!node
->u
.f
.init
|| !node
->u
.f
.inc
)
1149 graft
= isl_ast_graft_free(graft
);
1152 graft
= set_for_cond_from_list(graft
, upper_list
, build
);
1154 graft
= set_for_cond_from_set(graft
, upper_set
, build
);
1156 isl_ast_build_free(build
);
1161 /* Update "graft" based on "bounds" and "domain" for the generic,
1162 * non-degenerate, case.
1164 * "c_lower" and "c_upper" contain the lower and upper bounds
1165 * that the loop node should express.
1166 * "domain" is the subset of the intersection of the constraints
1167 * for which some code is executed.
1169 * There may be zero lower bounds or zero upper bounds in "constraints"
1170 * in case the list of constraints was created
1171 * based on the atomic option or based on separation with explicit bounds.
1172 * In that case, we use "domain" to derive lower and/or upper bounds.
1174 * We first compute a list of one or more lower bounds.
1176 * Then we decide if we want to express the condition as
1178 * iterator <= min(list of upper bounds)
1180 * or as a conjunction of constraints.
1182 * The set of enforced constraints is then computed either based on
1183 * a list of upper bounds or on a set of upper bound constraints.
1184 * We do not compute any enforced constraints if we were forced
1185 * to compute a lower or upper bound using exact_bound. The domains
1186 * of the resulting expressions may imply some bounds on outer dimensions
1187 * that we do not want to appear in the enforced constraints since
1188 * they are not actually enforced by the corresponding code.
1190 * Finally, we fill in the expressions of the for node.
1192 static __isl_give isl_ast_graft
*refine_generic_bounds(
1193 __isl_take isl_ast_graft
*graft
,
1194 __isl_take isl_constraint_list
*c_lower
,
1195 __isl_take isl_constraint_list
*c_upper
,
1196 __isl_keep isl_set
*domain
, __isl_keep isl_ast_build
*build
)
1200 isl_pw_aff_list
*lower
;
1202 isl_set
*upper_set
= NULL
;
1203 isl_pw_aff_list
*upper_list
= NULL
;
1204 isl_size n_lower
, n_upper
;
1206 if (!graft
|| !c_lower
|| !c_upper
|| !build
)
1209 depth
= isl_ast_build_get_depth(build
);
1210 ctx
= isl_ast_graft_get_ctx(graft
);
1212 n_lower
= isl_constraint_list_n_constraint(c_lower
);
1213 n_upper
= isl_constraint_list_n_constraint(c_upper
);
1214 if (n_lower
< 0 || n_upper
< 0)
1217 use_list
= use_upper_bound_list(ctx
, n_upper
, domain
, depth
);
1219 lower
= lower_bounds(c_lower
, depth
, domain
, build
);
1222 upper_list
= upper_bounds(c_upper
, depth
, domain
, build
);
1223 else if (n_upper
> 0)
1224 upper_set
= intersect_constraints(c_upper
);
1226 upper_set
= isl_set_universe(isl_set_get_space(domain
));
1228 if (n_lower
== 0 || n_upper
== 0)
1231 graft
= set_enforced_from_list(graft
, lower
, upper_list
);
1233 graft
= set_enforced_from_set(graft
, lower
, depth
, upper_set
);
1235 graft
= set_for_node_expressions(graft
, lower
, use_list
, upper_list
,
1238 isl_pw_aff_list_free(lower
);
1239 isl_pw_aff_list_free(upper_list
);
1240 isl_set_free(upper_set
);
1241 isl_constraint_list_free(c_lower
);
1242 isl_constraint_list_free(c_upper
);
1246 isl_constraint_list_free(c_lower
);
1247 isl_constraint_list_free(c_upper
);
1248 return isl_ast_graft_free(graft
);
1251 /* Internal data structure used inside count_constraints to keep
1252 * track of the number of constraints that are independent of dimension "pos",
1253 * the lower bounds in "pos" and the upper bounds in "pos".
1255 struct isl_ast_count_constraints_data
{
1263 /* Increment data->n_indep, data->lower or data->upper depending
1264 * on whether "c" is independenct of dimensions data->pos,
1265 * a lower bound or an upper bound.
1267 static isl_stat
count_constraints(__isl_take isl_constraint
*c
, void *user
)
1269 struct isl_ast_count_constraints_data
*data
= user
;
1271 if (isl_constraint_is_lower_bound(c
, isl_dim_set
, data
->pos
))
1273 else if (isl_constraint_is_upper_bound(c
, isl_dim_set
, data
->pos
))
1278 isl_constraint_free(c
);
1283 /* Update "graft" based on "bounds" and "domain" for the generic,
1284 * non-degenerate, case.
1286 * "list" respresent the list of bounds that need to be encoded by
1287 * the for loop. Only the constraints that involve the iterator
1288 * are relevant here. The other constraints are taken care of by
1289 * the caller and are included in the generated constraints of "build".
1290 * "domain" is the subset of the intersection of the constraints
1291 * for which some code is executed.
1292 * "build" is the build in which graft->node was created.
1294 * We separate lower bounds, upper bounds and constraints that
1295 * are independent of the loop iterator.
1297 * The actual for loop bounds are generated in refine_generic_bounds.
1299 static __isl_give isl_ast_graft
*refine_generic_split(
1300 __isl_take isl_ast_graft
*graft
, __isl_take isl_constraint_list
*list
,
1301 __isl_keep isl_set
*domain
, __isl_keep isl_ast_build
*build
)
1303 struct isl_ast_count_constraints_data data
;
1304 isl_constraint_list
*lower
;
1305 isl_constraint_list
*upper
;
1308 return isl_ast_graft_free(graft
);
1310 data
.pos
= isl_ast_build_get_depth(build
);
1312 list
= isl_constraint_list_sort(list
, &cmp_constraint
, &data
.pos
);
1314 return isl_ast_graft_free(graft
);
1316 data
.n_indep
= data
.n_lower
= data
.n_upper
= 0;
1317 if (isl_constraint_list_foreach(list
, &count_constraints
, &data
) < 0) {
1318 isl_constraint_list_free(list
);
1319 return isl_ast_graft_free(graft
);
1322 lower
= isl_constraint_list_drop(list
, 0, data
.n_indep
);
1323 upper
= isl_constraint_list_copy(lower
);
1324 lower
= isl_constraint_list_drop(lower
, data
.n_lower
, data
.n_upper
);
1325 upper
= isl_constraint_list_drop(upper
, 0, data
.n_lower
);
1327 return refine_generic_bounds(graft
, lower
, upper
, domain
, build
);
1330 /* Update "graft" based on "bounds" and "domain" for the generic,
1331 * non-degenerate, case.
1333 * "bounds" respresent the bounds that need to be encoded by
1334 * the for loop (or a guard around the for loop).
1335 * "domain" is the subset of "bounds" for which some code is executed.
1336 * "build" is the build in which graft->node was created.
1338 * We break up "bounds" into a list of constraints and continue with
1339 * refine_generic_split.
1341 static __isl_give isl_ast_graft
*refine_generic(
1342 __isl_take isl_ast_graft
*graft
,
1343 __isl_keep isl_basic_set
*bounds
, __isl_keep isl_set
*domain
,
1344 __isl_keep isl_ast_build
*build
)
1346 isl_constraint_list
*list
;
1348 if (!build
|| !graft
)
1349 return isl_ast_graft_free(graft
);
1351 list
= isl_basic_set_get_constraint_list(bounds
);
1353 graft
= refine_generic_split(graft
, list
, domain
, build
);
1358 /* Create a for node for the current level.
1360 * Mark the for node degenerate if "degenerate" is set.
1362 static __isl_give isl_ast_node
*create_for(__isl_keep isl_ast_build
*build
,
1372 depth
= isl_ast_build_get_depth(build
);
1373 id
= isl_ast_build_get_iterator_id(build
, depth
);
1374 node
= isl_ast_node_alloc_for(id
);
1376 node
= isl_ast_node_for_mark_degenerate(node
);
1381 /* If the ast_build_exploit_nested_bounds option is set, then return
1382 * the constraints enforced by all elements in "list".
1383 * Otherwise, return the universe.
1385 static __isl_give isl_basic_set
*extract_shared_enforced(
1386 __isl_keep isl_ast_graft_list
*list
, __isl_keep isl_ast_build
*build
)
1394 ctx
= isl_ast_graft_list_get_ctx(list
);
1395 if (isl_options_get_ast_build_exploit_nested_bounds(ctx
))
1396 return isl_ast_graft_list_extract_shared_enforced(list
, build
);
1398 space
= isl_ast_build_get_space(build
, 1);
1399 return isl_basic_set_universe(space
);
1402 /* Return the pending constraints of "build" that are not already taken
1403 * care of (by a combination of "enforced" and the generated constraints
1406 static __isl_give isl_set
*extract_pending(__isl_keep isl_ast_build
*build
,
1407 __isl_keep isl_basic_set
*enforced
)
1409 isl_set
*guard
, *context
;
1411 guard
= isl_ast_build_get_pending(build
);
1412 context
= isl_set_from_basic_set(isl_basic_set_copy(enforced
));
1413 context
= isl_set_intersect(context
,
1414 isl_ast_build_get_generated(build
));
1415 return isl_set_gist(guard
, context
);
1418 /* Create an AST node for the current dimension based on
1419 * the schedule domain "bounds" and return the node encapsulated
1420 * in an isl_ast_graft.
1422 * "executed" is the current inverse schedule, taking into account
1423 * the bounds in "bounds"
1424 * "domain" is the domain of "executed", with inner dimensions projected out.
1425 * It may be a strict subset of "bounds" in case "bounds" was created
1426 * based on the atomic option or based on separation with explicit bounds.
1428 * "domain" may satisfy additional equalities that result
1429 * from intersecting "executed" with "bounds" in add_node.
1430 * It may also satisfy some global constraints that were dropped out because
1431 * we performed separation with explicit bounds.
1432 * The very first step is then to copy these constraints to "bounds".
1434 * Since we may be calling before_each_for and after_each_for
1435 * callbacks, we record the current inverse schedule in the build.
1437 * We consider three builds,
1438 * "build" is the one in which the current level is created,
1439 * "body_build" is the build in which the next level is created,
1440 * "sub_build" is essentially the same as "body_build", except that
1441 * the depth has not been increased yet.
1443 * "build" already contains information (in strides and offsets)
1444 * about the strides at the current level, but this information is not
1445 * reflected in the build->domain.
1446 * We first add this information and the "bounds" to the sub_build->domain.
1447 * isl_ast_build_set_loop_bounds adds the stride information and
1448 * checks whether the current dimension attains
1449 * only a single value and whether this single value can be represented using
1450 * a single affine expression.
1451 * In the first case, the current level is considered "degenerate".
1452 * In the second, sub-case, the current level is considered "eliminated".
1453 * Eliminated levels don't need to be reflected in the AST since we can
1454 * simply plug in the affine expression. For degenerate, but non-eliminated,
1455 * levels, we do introduce a for node, but mark is as degenerate so that
1456 * it can be printed as an assignment of the single value to the loop
1459 * If the current level is eliminated, we explicitly plug in the value
1460 * for the current level found by isl_ast_build_set_loop_bounds in the
1461 * inverse schedule. This ensures that if we are working on a slice
1462 * of the domain based on information available in the inverse schedule
1463 * and the build domain, that then this information is also reflected
1464 * in the inverse schedule. This operation also eliminates the current
1465 * dimension from the inverse schedule making sure no inner dimensions depend
1466 * on the current dimension. Otherwise, we create a for node, marking
1467 * it degenerate if appropriate. The initial for node is still incomplete
1468 * and will be completed in either refine_degenerate or refine_generic.
1470 * We then generate a sequence of grafts for the next level,
1471 * create a surrounding graft for the current level and insert
1472 * the for node we created (if the current level is not eliminated).
1473 * Before creating a graft for the current level, we first extract
1474 * hoistable constraints from the child guards and combine them
1475 * with the pending constraints in the build. These constraints
1476 * are used to simplify the child guards and then added to the guard
1477 * of the current graft to ensure that they will be generated.
1478 * If the hoisted guard is a disjunction, then we use it directly
1479 * to gist the guards on the children before intersect it with the
1480 * pending constraints. We do so because this disjunction is typically
1481 * identical to the guards on the children such that these guards
1482 * can be effectively removed completely. After the intersection,
1483 * the gist operation would have a harder time figuring this out.
1485 * Finally, we set the bounds of the for loop in either
1486 * refine_degenerate or refine_generic.
1487 * We do so in a context where the pending constraints of the build
1488 * have been replaced by the guard of the current graft.
1490 static __isl_give isl_ast_graft
*create_node_scaled(
1491 __isl_take isl_union_map
*executed
,
1492 __isl_take isl_basic_set
*bounds
, __isl_take isl_set
*domain
,
1493 __isl_take isl_ast_build
*build
)
1497 isl_bool eliminated
;
1499 isl_basic_set
*hull
;
1500 isl_basic_set
*enforced
;
1501 isl_set
*guard
, *hoisted
;
1502 isl_ast_node
*node
= NULL
;
1503 isl_ast_graft
*graft
;
1504 isl_ast_graft_list
*children
;
1505 isl_ast_build
*sub_build
;
1506 isl_ast_build
*body_build
;
1508 domain
= isl_ast_build_eliminate_divs(build
, domain
);
1509 domain
= isl_set_detect_equalities(domain
);
1510 hull
= isl_set_unshifted_simple_hull(isl_set_copy(domain
));
1511 bounds
= isl_basic_set_intersect(bounds
, hull
);
1512 build
= isl_ast_build_set_executed(build
, isl_union_map_copy(executed
));
1514 depth
= isl_ast_build_get_depth(build
);
1515 sub_build
= isl_ast_build_copy(build
);
1516 bounds
= isl_basic_set_remove_redundancies(bounds
);
1517 bounds
= isl_ast_build_specialize_basic_set(sub_build
, bounds
);
1518 sub_build
= isl_ast_build_set_loop_bounds(sub_build
,
1519 isl_basic_set_copy(bounds
));
1520 degenerate
= isl_ast_build_has_value(sub_build
);
1521 eliminated
= isl_ast_build_has_affine_value(sub_build
, depth
);
1522 if (degenerate
< 0 || eliminated
< 0)
1523 executed
= isl_union_map_free(executed
);
1525 bounds
= isl_ast_build_compute_gist_basic_set(build
, bounds
);
1526 sub_build
= isl_ast_build_set_pending_generated(sub_build
,
1527 isl_basic_set_copy(bounds
));
1529 executed
= plug_in_values(executed
, sub_build
);
1531 node
= create_for(build
, degenerate
);
1533 body_build
= isl_ast_build_copy(sub_build
);
1534 body_build
= isl_ast_build_increase_depth(body_build
);
1536 node
= before_each_for(node
, body_build
);
1537 children
= generate_next_level(executed
,
1538 isl_ast_build_copy(body_build
));
1540 enforced
= extract_shared_enforced(children
, build
);
1541 guard
= extract_pending(sub_build
, enforced
);
1542 hoisted
= isl_ast_graft_list_extract_hoistable_guard(children
, build
);
1543 n
= isl_set_n_basic_set(hoisted
);
1545 children
= isl_ast_graft_list_free(children
);
1547 children
= isl_ast_graft_list_gist_guards(children
,
1548 isl_set_copy(hoisted
));
1549 guard
= isl_set_intersect(guard
, hoisted
);
1551 guard
= add_implied_guards(guard
, degenerate
, bounds
, build
);
1553 graft
= isl_ast_graft_alloc_from_children(children
,
1554 isl_set_copy(guard
), enforced
, build
, sub_build
);
1557 isl_ast_build
*for_build
;
1559 graft
= isl_ast_graft_insert_for(graft
, node
);
1560 for_build
= isl_ast_build_copy(build
);
1561 for_build
= isl_ast_build_replace_pending_by_guard(for_build
,
1562 isl_set_copy(guard
));
1564 graft
= refine_degenerate(graft
, for_build
, sub_build
);
1566 graft
= refine_generic(graft
, bounds
,
1568 isl_ast_build_free(for_build
);
1570 isl_set_free(guard
);
1572 graft
= after_each_for(graft
, body_build
);
1574 isl_ast_build_free(body_build
);
1575 isl_ast_build_free(sub_build
);
1576 isl_ast_build_free(build
);
1577 isl_basic_set_free(bounds
);
1578 isl_set_free(domain
);
1583 /* Internal data structure for checking if all constraints involving
1584 * the input dimension "depth" are such that the other coefficients
1585 * are multiples of "m", reducing "m" if they are not.
1586 * If "m" is reduced all the way down to "1", then the check has failed
1587 * and we break out of the iteration.
1589 struct isl_check_scaled_data
{
1594 /* If constraint "c" involves the input dimension data->depth,
1595 * then make sure that all the other coefficients are multiples of data->m,
1596 * reducing data->m if needed.
1597 * Break out of the iteration if data->m has become equal to "1".
1599 static isl_stat
constraint_check_scaled(__isl_take isl_constraint
*c
,
1602 struct isl_check_scaled_data
*data
= user
;
1605 enum isl_dim_type t
[] = { isl_dim_param
, isl_dim_in
, isl_dim_out
,
1608 if (!isl_constraint_involves_dims(c
, isl_dim_in
, data
->depth
, 1)) {
1609 isl_constraint_free(c
);
1613 for (i
= 0; i
< 4; ++i
) {
1614 n
= isl_constraint_dim(c
, t
[i
]);
1617 for (j
= 0; j
< n
; ++j
) {
1620 if (t
[i
] == isl_dim_in
&& j
== data
->depth
)
1622 if (!isl_constraint_involves_dims(c
, t
[i
], j
, 1))
1624 d
= isl_constraint_get_coefficient_val(c
, t
[i
], j
);
1625 data
->m
= isl_val_gcd(data
->m
, d
);
1626 if (isl_val_is_one(data
->m
))
1633 isl_constraint_free(c
);
1635 return i
< 4 ? isl_stat_error
: isl_stat_ok
;
1638 /* For each constraint of "bmap" that involves the input dimension data->depth,
1639 * make sure that all the other coefficients are multiples of data->m,
1640 * reducing data->m if needed.
1641 * Break out of the iteration if data->m has become equal to "1".
1643 static isl_stat
basic_map_check_scaled(__isl_take isl_basic_map
*bmap
,
1648 r
= isl_basic_map_foreach_constraint(bmap
,
1649 &constraint_check_scaled
, user
);
1650 isl_basic_map_free(bmap
);
1655 /* For each constraint of "map" that involves the input dimension data->depth,
1656 * make sure that all the other coefficients are multiples of data->m,
1657 * reducing data->m if needed.
1658 * Break out of the iteration if data->m has become equal to "1".
1660 static isl_stat
map_check_scaled(__isl_take isl_map
*map
, void *user
)
1664 r
= isl_map_foreach_basic_map(map
, &basic_map_check_scaled
, user
);
1670 /* Create an AST node for the current dimension based on
1671 * the schedule domain "bounds" and return the node encapsulated
1672 * in an isl_ast_graft.
1674 * "executed" is the current inverse schedule, taking into account
1675 * the bounds in "bounds"
1676 * "domain" is the domain of "executed", with inner dimensions projected out.
1679 * Before moving on to the actual AST node construction in create_node_scaled,
1680 * we first check if the current dimension is strided and if we can scale
1681 * down this stride. Note that we only do this if the ast_build_scale_strides
1684 * In particular, let the current dimension take on values
1688 * with a an integer. We check if we can find an integer m that (obviously)
1689 * divides both f and s.
1691 * If so, we check if the current dimension only appears in constraints
1692 * where the coefficients of the other variables are multiples of m.
1693 * We perform this extra check to avoid the risk of introducing
1694 * divisions by scaling down the current dimension.
1696 * If so, we scale the current dimension down by a factor of m.
1697 * That is, we plug in
1701 * Note that in principle we could always scale down strided loops
1706 * but this may result in i' taking on larger values than the original i,
1707 * due to the shift by "f".
1708 * By constrast, the scaling in (1) can only reduce the (absolute) value "i".
1710 static __isl_give isl_ast_graft
*create_node(__isl_take isl_union_map
*executed
,
1711 __isl_take isl_basic_set
*bounds
, __isl_take isl_set
*domain
,
1712 __isl_take isl_ast_build
*build
)
1714 struct isl_check_scaled_data data
;
1719 ctx
= isl_ast_build_get_ctx(build
);
1720 if (!isl_options_get_ast_build_scale_strides(ctx
))
1721 return create_node_scaled(executed
, bounds
, domain
, build
);
1723 data
.depth
= isl_ast_build_get_depth(build
);
1724 if (!isl_ast_build_has_stride(build
, data
.depth
))
1725 return create_node_scaled(executed
, bounds
, domain
, build
);
1727 offset
= isl_ast_build_get_offset(build
, data
.depth
);
1728 data
.m
= isl_ast_build_get_stride(build
, data
.depth
);
1730 offset
= isl_aff_free(offset
);
1731 offset
= isl_aff_scale_down_val(offset
, isl_val_copy(data
.m
));
1732 d
= isl_aff_get_denominator_val(offset
);
1734 executed
= isl_union_map_free(executed
);
1736 if (executed
&& isl_val_is_divisible_by(data
.m
, d
))
1737 data
.m
= isl_val_div(data
.m
, d
);
1739 data
.m
= isl_val_set_si(data
.m
, 1);
1743 if (!isl_val_is_one(data
.m
)) {
1744 if (isl_union_map_foreach_map(executed
, &map_check_scaled
,
1746 !isl_val_is_one(data
.m
))
1747 executed
= isl_union_map_free(executed
);
1750 if (!isl_val_is_one(data
.m
)) {
1755 isl_union_map
*umap
;
1757 space
= isl_ast_build_get_space(build
, 1);
1758 space
= isl_space_map_from_set(space
);
1759 ma
= isl_multi_aff_identity(space
);
1760 aff
= isl_multi_aff_get_aff(ma
, data
.depth
);
1761 aff
= isl_aff_scale_val(aff
, isl_val_copy(data
.m
));
1762 ma
= isl_multi_aff_set_aff(ma
, data
.depth
, aff
);
1764 bounds
= isl_basic_set_preimage_multi_aff(bounds
,
1765 isl_multi_aff_copy(ma
));
1766 domain
= isl_set_preimage_multi_aff(domain
,
1767 isl_multi_aff_copy(ma
));
1768 map
= isl_map_reverse(isl_map_from_multi_aff(ma
));
1769 umap
= isl_union_map_from_map(map
);
1770 executed
= isl_union_map_apply_domain(executed
,
1771 isl_union_map_copy(umap
));
1772 build
= isl_ast_build_scale_down(build
, isl_val_copy(data
.m
),
1775 isl_aff_free(offset
);
1776 isl_val_free(data
.m
);
1778 return create_node_scaled(executed
, bounds
, domain
, build
);
1781 /* Add the basic set to the list that "user" points to.
1783 static isl_stat
collect_basic_set(__isl_take isl_basic_set
*bset
, void *user
)
1785 isl_basic_set_list
**list
= user
;
1787 *list
= isl_basic_set_list_add(*list
, bset
);
1792 /* Extract the basic sets of "set" and collect them in an isl_basic_set_list.
1794 static __isl_give isl_basic_set_list
*isl_basic_set_list_from_set(
1795 __isl_take isl_set
*set
)
1799 isl_basic_set_list
*list
;
1801 n
= isl_set_n_basic_set(set
);
1803 set
= isl_set_free(set
);
1807 ctx
= isl_set_get_ctx(set
);
1809 list
= isl_basic_set_list_alloc(ctx
, n
);
1810 if (isl_set_foreach_basic_set(set
, &collect_basic_set
, &list
) < 0)
1811 list
= isl_basic_set_list_free(list
);
1817 /* Generate code for the schedule domain "bounds"
1818 * and add the result to "list".
1820 * We mainly detect strides here and check if the bounds do not
1821 * conflict with the current build domain
1822 * and then pass over control to create_node.
1824 * "bounds" reflects the bounds on the current dimension and possibly
1825 * some extra conditions on outer dimensions.
1826 * It does not, however, include any divs involving the current dimension,
1827 * so it does not capture any stride constraints.
1828 * We therefore need to compute that part of the schedule domain that
1829 * intersects with "bounds" and derive the strides from the result.
1831 static __isl_give isl_ast_graft_list
*add_node(
1832 __isl_take isl_ast_graft_list
*list
, __isl_take isl_union_map
*executed
,
1833 __isl_take isl_basic_set
*bounds
, __isl_take isl_ast_build
*build
)
1835 isl_ast_graft
*graft
;
1836 isl_set
*domain
= NULL
;
1837 isl_union_set
*uset
;
1838 int empty
, disjoint
;
1840 uset
= isl_union_set_from_basic_set(isl_basic_set_copy(bounds
));
1841 executed
= isl_union_map_intersect_domain(executed
, uset
);
1842 empty
= isl_union_map_is_empty(executed
);
1848 uset
= isl_union_map_domain(isl_union_map_copy(executed
));
1849 domain
= isl_set_from_union_set(uset
);
1850 domain
= isl_ast_build_specialize(build
, domain
);
1852 domain
= isl_set_compute_divs(domain
);
1853 domain
= isl_ast_build_eliminate_inner(build
, domain
);
1854 disjoint
= isl_set_is_disjoint(domain
, build
->domain
);
1860 build
= isl_ast_build_detect_strides(build
, isl_set_copy(domain
));
1862 graft
= create_node(executed
, bounds
, domain
,
1863 isl_ast_build_copy(build
));
1864 list
= isl_ast_graft_list_add(list
, graft
);
1865 isl_ast_build_free(build
);
1868 list
= isl_ast_graft_list_free(list
);
1870 isl_set_free(domain
);
1871 isl_basic_set_free(bounds
);
1872 isl_union_map_free(executed
);
1873 isl_ast_build_free(build
);
1877 /* Does any element of i follow or coincide with any element of j
1878 * at the current depth for equal values of the outer dimensions?
1880 static isl_bool
domain_follows_at_depth(__isl_keep isl_basic_set
*i
,
1881 __isl_keep isl_basic_set
*j
, void *user
)
1883 int depth
= *(int *) user
;
1884 isl_basic_map
*test
;
1888 test
= isl_basic_map_from_domain_and_range(isl_basic_set_copy(i
),
1889 isl_basic_set_copy(j
));
1890 for (l
= 0; l
< depth
; ++l
)
1891 test
= isl_basic_map_equate(test
, isl_dim_in
, l
,
1893 test
= isl_basic_map_order_ge(test
, isl_dim_in
, depth
,
1894 isl_dim_out
, depth
);
1895 empty
= isl_basic_map_is_empty(test
);
1896 isl_basic_map_free(test
);
1898 return isl_bool_not(empty
);
1901 /* Split up each element of "list" into a part that is related to "bset"
1902 * according to "gt" and a part that is not.
1903 * Return a list that consist of "bset" and all the pieces.
1905 static __isl_give isl_basic_set_list
*add_split_on(
1906 __isl_take isl_basic_set_list
*list
, __isl_take isl_basic_set
*bset
,
1907 __isl_keep isl_basic_map
*gt
)
1911 isl_basic_set_list
*res
;
1913 n
= isl_basic_set_list_n_basic_set(list
);
1915 bset
= isl_basic_set_free(bset
);
1917 gt
= isl_basic_map_copy(gt
);
1918 gt
= isl_basic_map_intersect_domain(gt
, isl_basic_set_copy(bset
));
1919 res
= isl_basic_set_list_from_basic_set(bset
);
1920 for (i
= 0; res
&& i
< n
; ++i
) {
1921 isl_basic_set
*bset
;
1922 isl_set
*set1
, *set2
;
1923 isl_basic_map
*bmap
;
1926 bset
= isl_basic_set_list_get_basic_set(list
, i
);
1927 bmap
= isl_basic_map_copy(gt
);
1928 bmap
= isl_basic_map_intersect_range(bmap
, bset
);
1929 bset
= isl_basic_map_range(bmap
);
1930 empty
= isl_basic_set_is_empty(bset
);
1932 res
= isl_basic_set_list_free(res
);
1934 isl_basic_set_free(bset
);
1935 bset
= isl_basic_set_list_get_basic_set(list
, i
);
1936 res
= isl_basic_set_list_add(res
, bset
);
1940 res
= isl_basic_set_list_add(res
, isl_basic_set_copy(bset
));
1941 set1
= isl_set_from_basic_set(bset
);
1942 bset
= isl_basic_set_list_get_basic_set(list
, i
);
1943 set2
= isl_set_from_basic_set(bset
);
1944 set1
= isl_set_subtract(set2
, set1
);
1945 set1
= isl_set_make_disjoint(set1
);
1947 res
= isl_basic_set_list_concat(res
,
1948 isl_basic_set_list_from_set(set1
));
1950 isl_basic_map_free(gt
);
1951 isl_basic_set_list_free(list
);
1955 static __isl_give isl_ast_graft_list
*generate_sorted_domains(
1956 __isl_keep isl_basic_set_list
*domain_list
,
1957 __isl_keep isl_union_map
*executed
,
1958 __isl_keep isl_ast_build
*build
);
1960 /* Internal data structure for add_nodes.
1962 * "executed" and "build" are extra arguments to be passed to add_node.
1963 * "list" collects the results.
1965 struct isl_add_nodes_data
{
1966 isl_union_map
*executed
;
1967 isl_ast_build
*build
;
1969 isl_ast_graft_list
*list
;
1972 /* Generate code for the schedule domains in "scc"
1973 * and add the results to "list".
1975 * The domains in "scc" form a strongly connected component in the ordering.
1976 * If the number of domains in "scc" is larger than 1, then this means
1977 * that we cannot determine a valid ordering for the domains in the component.
1978 * This should be fairly rare because the individual domains
1979 * have been made disjoint first.
1980 * The problem is that the domains may be integrally disjoint but not
1981 * rationally disjoint. For example, we may have domains
1983 * { [i,i] : 0 <= i <= 1 } and { [i,1-i] : 0 <= i <= 1 }
1985 * These two domains have an empty intersection, but their rational
1986 * relaxations do intersect. It is impossible to order these domains
1987 * in the second dimension because the first should be ordered before
1988 * the second for outer dimension equal to 0, while it should be ordered
1989 * after for outer dimension equal to 1.
1991 * This may happen in particular in case of unrolling since the domain
1992 * of each slice is replaced by its simple hull.
1994 * For each basic set i in "scc" and for each of the following basic sets j,
1995 * we split off that part of the basic set i that shares the outer dimensions
1996 * with j and lies before j in the current dimension.
1997 * We collect all the pieces in a new list that replaces "scc".
1999 * While the elements in "scc" should be disjoint, we double-check
2000 * this property to avoid running into an infinite recursion in case
2001 * they intersect due to some internal error.
2003 static isl_stat
add_nodes(__isl_take isl_basic_set_list
*scc
, void *user
)
2005 struct isl_add_nodes_data
*data
= user
;
2008 isl_basic_set
*bset
, *first
;
2009 isl_basic_set_list
*list
;
2013 n
= isl_basic_set_list_n_basic_set(scc
);
2016 bset
= isl_basic_set_list_get_basic_set(scc
, 0);
2018 isl_basic_set_list_free(scc
);
2019 data
->list
= add_node(data
->list
,
2020 isl_union_map_copy(data
->executed
), bset
,
2021 isl_ast_build_copy(data
->build
));
2022 return data
->list
? isl_stat_ok
: isl_stat_error
;
2025 depth
= isl_ast_build_get_depth(data
->build
);
2026 space
= isl_basic_set_get_space(bset
);
2027 space
= isl_space_map_from_set(space
);
2028 gt
= isl_basic_map_universe(space
);
2029 for (i
= 0; i
< depth
; ++i
)
2030 gt
= isl_basic_map_equate(gt
, isl_dim_in
, i
, isl_dim_out
, i
);
2031 gt
= isl_basic_map_order_gt(gt
, isl_dim_in
, depth
, isl_dim_out
, depth
);
2033 first
= isl_basic_set_copy(bset
);
2034 list
= isl_basic_set_list_from_basic_set(bset
);
2035 for (i
= 1; i
< n
; ++i
) {
2038 bset
= isl_basic_set_list_get_basic_set(scc
, i
);
2040 disjoint
= isl_basic_set_is_disjoint(bset
, first
);
2042 list
= isl_basic_set_list_free(list
);
2044 isl_die(isl_basic_set_list_get_ctx(scc
),
2046 "basic sets in scc are assumed to be disjoint",
2047 list
= isl_basic_set_list_free(list
));
2049 list
= add_split_on(list
, bset
, gt
);
2051 isl_basic_set_free(first
);
2052 isl_basic_map_free(gt
);
2053 isl_basic_set_list_free(scc
);
2055 data
->list
= isl_ast_graft_list_concat(data
->list
,
2056 generate_sorted_domains(scc
, data
->executed
, data
->build
));
2057 isl_basic_set_list_free(scc
);
2059 return data
->list
? isl_stat_ok
: isl_stat_error
;
2061 isl_basic_set_list_free(scc
);
2062 return isl_stat_error
;
2065 /* Sort the domains in "domain_list" according to the execution order
2066 * at the current depth (for equal values of the outer dimensions),
2067 * generate code for each of them, collecting the results in a list.
2068 * If no code is generated (because the intersection of the inverse schedule
2069 * with the domains turns out to be empty), then an empty list is returned.
2071 * The caller is responsible for ensuring that the basic sets in "domain_list"
2072 * are pair-wise disjoint. It can, however, in principle happen that
2073 * two basic sets should be ordered one way for one value of the outer
2074 * dimensions and the other way for some other value of the outer dimensions.
2075 * We therefore play safe and look for strongly connected components.
2076 * The function add_nodes takes care of handling non-trivial components.
2078 static __isl_give isl_ast_graft_list
*generate_sorted_domains(
2079 __isl_keep isl_basic_set_list
*domain_list
,
2080 __isl_keep isl_union_map
*executed
, __isl_keep isl_ast_build
*build
)
2083 struct isl_add_nodes_data data
;
2087 n
= isl_basic_set_list_n_basic_set(domain_list
);
2091 ctx
= isl_basic_set_list_get_ctx(domain_list
);
2092 data
.list
= isl_ast_graft_list_alloc(ctx
, n
);
2096 return add_node(data
.list
, isl_union_map_copy(executed
),
2097 isl_basic_set_list_get_basic_set(domain_list
, 0),
2098 isl_ast_build_copy(build
));
2100 depth
= isl_ast_build_get_depth(build
);
2101 data
.executed
= executed
;
2103 if (isl_basic_set_list_foreach_scc(domain_list
,
2104 &domain_follows_at_depth
, &depth
,
2105 &add_nodes
, &data
) < 0)
2106 data
.list
= isl_ast_graft_list_free(data
.list
);
2111 /* Do i and j share any values for the outer dimensions?
2113 static isl_bool
shared_outer(__isl_keep isl_basic_set
*i
,
2114 __isl_keep isl_basic_set
*j
, void *user
)
2116 int depth
= *(int *) user
;
2117 isl_basic_map
*test
;
2121 test
= isl_basic_map_from_domain_and_range(isl_basic_set_copy(i
),
2122 isl_basic_set_copy(j
));
2123 for (l
= 0; l
< depth
; ++l
)
2124 test
= isl_basic_map_equate(test
, isl_dim_in
, l
,
2126 empty
= isl_basic_map_is_empty(test
);
2127 isl_basic_map_free(test
);
2129 return isl_bool_not(empty
);
2132 /* Internal data structure for generate_sorted_domains_wrap.
2134 * "n" is the total number of basic sets
2135 * "executed" and "build" are extra arguments to be passed
2136 * to generate_sorted_domains.
2138 * "single" is set to 1 by generate_sorted_domains_wrap if there
2139 * is only a single component.
2140 * "list" collects the results.
2142 struct isl_ast_generate_parallel_domains_data
{
2144 isl_union_map
*executed
;
2145 isl_ast_build
*build
;
2148 isl_ast_graft_list
*list
;
2151 /* Call generate_sorted_domains on "scc", fuse the result into a list
2152 * with either zero or one graft and collect the these single element
2153 * lists into data->list.
2155 * If there is only one component, i.e., if the number of basic sets
2156 * in the current component is equal to the total number of basic sets,
2157 * then data->single is set to 1 and the result of generate_sorted_domains
2160 static isl_stat
generate_sorted_domains_wrap(__isl_take isl_basic_set_list
*scc
,
2163 struct isl_ast_generate_parallel_domains_data
*data
= user
;
2164 isl_ast_graft_list
*list
;
2167 n
= isl_basic_set_list_n_basic_set(scc
);
2169 scc
= isl_basic_set_list_free(scc
);
2170 list
= generate_sorted_domains(scc
, data
->executed
, data
->build
);
2171 data
->single
= n
== data
->n
;
2173 list
= isl_ast_graft_list_fuse(list
, data
->build
);
2177 data
->list
= isl_ast_graft_list_concat(data
->list
, list
);
2179 isl_basic_set_list_free(scc
);
2181 return isl_stat_error
;
2186 /* Look for any (weakly connected) components in the "domain_list"
2187 * of domains that share some values of the outer dimensions.
2188 * That is, domains in different components do not share any values
2189 * of the outer dimensions. This means that these components
2190 * can be freely reordered.
2191 * Within each of the components, we sort the domains according
2192 * to the execution order at the current depth.
2194 * If there is more than one component, then generate_sorted_domains_wrap
2195 * fuses the result of each call to generate_sorted_domains
2196 * into a list with either zero or one graft and collects these (at most)
2197 * single element lists into a bigger list. This means that the elements of the
2198 * final list can be freely reordered. In particular, we sort them
2199 * according to an arbitrary but fixed ordering to ease merging of
2200 * graft lists from different components.
2202 static __isl_give isl_ast_graft_list
*generate_parallel_domains(
2203 __isl_keep isl_basic_set_list
*domain_list
,
2204 __isl_keep isl_union_map
*executed
, __isl_keep isl_ast_build
*build
)
2207 struct isl_ast_generate_parallel_domains_data data
;
2209 data
.n
= isl_basic_set_list_n_basic_set(domain_list
);
2214 return generate_sorted_domains(domain_list
, executed
, build
);
2216 depth
= isl_ast_build_get_depth(build
);
2218 data
.executed
= executed
;
2221 if (isl_basic_set_list_foreach_scc(domain_list
, &shared_outer
, &depth
,
2222 &generate_sorted_domains_wrap
,
2224 data
.list
= isl_ast_graft_list_free(data
.list
);
2227 data
.list
= isl_ast_graft_list_sort_guard(data
.list
);
2232 /* Internal data for separate_domain.
2234 * "explicit" is set if we only want to use explicit bounds.
2236 * "domain" collects the separated domains.
2238 struct isl_separate_domain_data
{
2239 isl_ast_build
*build
;
2244 /* Extract implicit bounds on the current dimension for the executed "map".
2246 * The domain of "map" may involve inner dimensions, so we
2247 * need to eliminate them.
2249 static __isl_give isl_set
*implicit_bounds(__isl_take isl_map
*map
,
2250 __isl_keep isl_ast_build
*build
)
2254 domain
= isl_map_domain(map
);
2255 domain
= isl_ast_build_eliminate(build
, domain
);
2260 /* Extract explicit bounds on the current dimension for the executed "map".
2262 * Rather than eliminating the inner dimensions as in implicit_bounds,
2263 * we simply drop any constraints involving those inner dimensions.
2264 * The idea is that most bounds that are implied by constraints on the
2265 * inner dimensions will be enforced by for loops and not by explicit guards.
2266 * There is then no need to separate along those bounds.
2268 static __isl_give isl_set
*explicit_bounds(__isl_take isl_map
*map
,
2269 __isl_keep isl_ast_build
*build
)
2275 dim
= isl_map_dim(map
, isl_dim_out
);
2277 return isl_map_domain(isl_map_free(map
));
2278 map
= isl_map_drop_constraints_involving_dims(map
, isl_dim_out
, 0, dim
);
2280 domain
= isl_map_domain(map
);
2281 depth
= isl_ast_build_get_depth(build
);
2282 dim
= isl_set_dim(domain
, isl_dim_set
);
2283 domain
= isl_set_detect_equalities(domain
);
2284 domain
= isl_set_drop_constraints_involving_dims(domain
,
2285 isl_dim_set
, depth
+ 1, dim
- (depth
+ 1));
2286 domain
= isl_set_remove_divs_involving_dims(domain
,
2287 isl_dim_set
, depth
, 1);
2288 domain
= isl_set_remove_unknown_divs(domain
);
2293 /* Split data->domain into pieces that intersect with the range of "map"
2294 * and pieces that do not intersect with the range of "map"
2295 * and then add that part of the range of "map" that does not intersect
2296 * with data->domain.
2298 static isl_stat
separate_domain(__isl_take isl_map
*map
, void *user
)
2300 struct isl_separate_domain_data
*data
= user
;
2305 domain
= explicit_bounds(map
, data
->build
);
2307 domain
= implicit_bounds(map
, data
->build
);
2309 domain
= isl_set_coalesce(domain
);
2310 domain
= isl_set_make_disjoint(domain
);
2311 d1
= isl_set_subtract(isl_set_copy(domain
), isl_set_copy(data
->domain
));
2312 d2
= isl_set_subtract(isl_set_copy(data
->domain
), isl_set_copy(domain
));
2313 data
->domain
= isl_set_intersect(data
->domain
, domain
);
2314 data
->domain
= isl_set_union(data
->domain
, d1
);
2315 data
->domain
= isl_set_union(data
->domain
, d2
);
2320 /* Separate the schedule domains of "executed".
2322 * That is, break up the domain of "executed" into basic sets,
2323 * such that for each basic set S, every element in S is associated with
2324 * the same domain spaces.
2326 * "space" is the (single) domain space of "executed".
2328 static __isl_give isl_set
*separate_schedule_domains(
2329 __isl_take isl_space
*space
, __isl_take isl_union_map
*executed
,
2330 __isl_keep isl_ast_build
*build
)
2332 struct isl_separate_domain_data data
= { build
};
2335 ctx
= isl_ast_build_get_ctx(build
);
2336 data
.explicit = isl_options_get_ast_build_separation_bounds(ctx
) ==
2337 ISL_AST_BUILD_SEPARATION_BOUNDS_EXPLICIT
;
2338 data
.domain
= isl_set_empty(space
);
2339 if (isl_union_map_foreach_map(executed
, &separate_domain
, &data
) < 0)
2340 data
.domain
= isl_set_free(data
.domain
);
2342 isl_union_map_free(executed
);
2346 /* Temporary data used during the search for a lower bound for unrolling.
2348 * "build" is the build in which the unrolling will be performed
2349 * "domain" is the original set for which to find a lower bound
2350 * "depth" is the dimension for which to find a lower boudn
2351 * "expansion" is the expansion that needs to be applied to "domain"
2352 * in the unrolling that will be performed
2354 * "lower" is the best lower bound found so far. It is NULL if we have not
2356 * "n" is the corresponding size. If lower is NULL, then the value of n
2358 * "n_div" is the maximal number of integer divisions in the first
2359 * unrolled iteration (after expansion). It is set to -1 if it hasn't
2360 * been computed yet.
2362 struct isl_find_unroll_data
{
2363 isl_ast_build
*build
;
2366 isl_basic_map
*expansion
;
2373 /* Return the constraint
2375 * i_"depth" = aff + offset
2377 static __isl_give isl_constraint
*at_offset(int depth
, __isl_keep isl_aff
*aff
,
2380 aff
= isl_aff_copy(aff
);
2381 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, depth
, -1);
2382 aff
= isl_aff_add_constant_si(aff
, offset
);
2383 return isl_equality_from_aff(aff
);
2386 /* Update *user to the number of integer divisions in the first element
2387 * of "ma", if it is larger than the current value.
2389 static isl_stat
update_n_div(__isl_take isl_set
*set
,
2390 __isl_take isl_multi_aff
*ma
, void *user
)
2396 aff
= isl_multi_aff_get_aff(ma
, 0);
2397 n_div
= isl_aff_dim(aff
, isl_dim_div
);
2399 isl_multi_aff_free(ma
);
2405 return n_div
>= 0 ? isl_stat_ok
: isl_stat_error
;
2408 /* Get the number of integer divisions in the expression for the iterator
2409 * value at the first slice in the unrolling based on lower bound "lower",
2410 * taking into account the expansion that needs to be performed on this slice.
2412 static int get_expanded_n_div(struct isl_find_unroll_data
*data
,
2413 __isl_keep isl_aff
*lower
)
2417 isl_map
*it_map
, *expansion
;
2418 isl_pw_multi_aff
*pma
;
2421 c
= at_offset(data
->depth
, lower
, 0);
2422 set
= isl_set_copy(data
->domain
);
2423 set
= isl_set_add_constraint(set
, c
);
2424 expansion
= isl_map_from_basic_map(isl_basic_map_copy(data
->expansion
));
2425 set
= isl_set_apply(set
, expansion
);
2426 it_map
= isl_ast_build_map_to_iterator(data
->build
, set
);
2427 pma
= isl_pw_multi_aff_from_map(it_map
);
2429 if (isl_pw_multi_aff_foreach_piece(pma
, &update_n_div
, &n
) < 0)
2431 isl_pw_multi_aff_free(pma
);
2436 /* Is the lower bound "lower" with corresponding iteration count "n"
2437 * better than the one stored in "data"?
2438 * If there is no upper bound on the iteration count ("n" is infinity) or
2439 * if the count is too large, then we cannot use this lower bound.
2440 * Otherwise, if there was no previous lower bound or
2441 * if the iteration count of the new lower bound is smaller than
2442 * the iteration count of the previous lower bound, then we consider
2443 * the new lower bound to be better.
2444 * If the iteration count is the same, then compare the number
2445 * of integer divisions that would be needed to express
2446 * the iterator value at the first slice in the unrolling
2447 * according to the lower bound. If we end up computing this
2448 * number, then store the lowest value in data->n_div.
2450 static int is_better_lower_bound(struct isl_find_unroll_data
*data
,
2451 __isl_keep isl_aff
*lower
, __isl_keep isl_val
*n
)
2458 if (isl_val_is_infty(n
))
2460 if (isl_val_cmp_si(n
, INT_MAX
) > 0)
2464 cmp
= isl_val_cmp_si(n
, *data
->n
);
2469 if (data
->n_div
< 0)
2470 data
->n_div
= get_expanded_n_div(data
, data
->lower
);
2471 if (data
->n_div
< 0)
2473 if (data
->n_div
== 0)
2475 n_div
= get_expanded_n_div(data
, lower
);
2478 if (n_div
>= data
->n_div
)
2480 data
->n_div
= n_div
;
2485 /* Check if we can use "c" as a lower bound and if it is better than
2486 * any previously found lower bound.
2488 * If "c" does not involve the dimension at the current depth,
2489 * then we cannot use it.
2490 * Otherwise, let "c" be of the form
2494 * We compute the maximal value of
2496 * -ceil(f(j)/a)) + i + 1
2498 * over the domain. If there is such a value "n", then we know
2500 * -ceil(f(j)/a)) + i + 1 <= n
2504 * i < ceil(f(j)/a)) + n
2506 * meaning that we can use ceil(f(j)/a)) as a lower bound for unrolling.
2507 * We just need to check if we have found any lower bound before and
2508 * if the new lower bound is better (smaller n or fewer integer divisions)
2509 * than the previously found lower bounds.
2511 static isl_stat
update_unrolling_lower_bound(struct isl_find_unroll_data
*data
,
2512 __isl_keep isl_constraint
*c
)
2514 isl_aff
*aff
, *lower
;
2518 if (!isl_constraint_is_lower_bound(c
, isl_dim_set
, data
->depth
))
2521 lower
= isl_constraint_get_bound(c
, isl_dim_set
, data
->depth
);
2522 lower
= isl_aff_ceil(lower
);
2523 aff
= isl_aff_copy(lower
);
2524 aff
= isl_aff_neg(aff
);
2525 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, data
->depth
, 1);
2526 aff
= isl_aff_add_constant_si(aff
, 1);
2527 max
= isl_set_max_val(data
->domain
, aff
);
2530 better
= is_better_lower_bound(data
, lower
, max
);
2531 if (better
< 0 || !better
) {
2533 isl_aff_free(lower
);
2534 return better
< 0 ? isl_stat_error
: isl_stat_ok
;
2537 isl_aff_free(data
->lower
);
2538 data
->lower
= lower
;
2539 *data
->n
= isl_val_get_num_si(max
);
2545 /* Check if we can use "c" as a lower bound and if it is better than
2546 * any previously found lower bound.
2548 static isl_stat
constraint_find_unroll(__isl_take isl_constraint
*c
, void *user
)
2550 struct isl_find_unroll_data
*data
;
2553 data
= (struct isl_find_unroll_data
*) user
;
2554 r
= update_unrolling_lower_bound(data
, c
);
2555 isl_constraint_free(c
);
2560 /* Look for a lower bound l(i) on the dimension at "depth"
2561 * and a size n such that "domain" is a subset of
2563 * { [i] : l(i) <= i_d < l(i) + n }
2565 * where d is "depth" and l(i) depends only on earlier dimensions.
2566 * Furthermore, try and find a lower bound such that n is as small as possible.
2567 * In particular, "n" needs to be finite.
2568 * "build" is the build in which the unrolling will be performed.
2569 * "expansion" is the expansion that needs to be applied to "domain"
2570 * in the unrolling that will be performed.
2572 * Inner dimensions have been eliminated from "domain" by the caller.
2574 * We first construct a collection of lower bounds on the input set
2575 * by computing its simple hull. We then iterate through them,
2576 * discarding those that we cannot use (either because they do not
2577 * involve the dimension at "depth" or because they have no corresponding
2578 * upper bound, meaning that "n" would be unbounded) and pick out the
2579 * best from the remaining ones.
2581 * If we cannot find a suitable lower bound, then we consider that
2584 static __isl_give isl_aff
*find_unroll_lower_bound(
2585 __isl_keep isl_ast_build
*build
, __isl_keep isl_set
*domain
,
2586 int depth
, __isl_keep isl_basic_map
*expansion
, int *n
)
2588 struct isl_find_unroll_data data
=
2589 { build
, domain
, depth
, expansion
, NULL
, n
, -1 };
2590 isl_basic_set
*hull
;
2592 hull
= isl_set_simple_hull(isl_set_copy(domain
));
2594 if (isl_basic_set_foreach_constraint(hull
,
2595 &constraint_find_unroll
, &data
) < 0)
2598 isl_basic_set_free(hull
);
2601 isl_die(isl_set_get_ctx(domain
), isl_error_invalid
,
2602 "cannot find lower bound for unrolling", return NULL
);
2606 isl_basic_set_free(hull
);
2607 return isl_aff_free(data
.lower
);
2610 /* Call "fn" on each iteration of the current dimension of "domain".
2611 * If "init" is not NULL, then it is called with the number of
2612 * iterations before any call to "fn".
2613 * Return -1 on failure.
2615 * Since we are going to be iterating over the individual values,
2616 * we first check if there are any strides on the current dimension.
2617 * If there is, we rewrite the current dimension i as
2619 * i = stride i' + offset
2621 * and then iterate over individual values of i' instead.
2623 * We then look for a lower bound on i' and a size such that the domain
2626 * { [j,i'] : l(j) <= i' < l(j) + n }
2628 * and then take slices of the domain at values of i'
2629 * between l(j) and l(j) + n - 1.
2631 * We compute the unshifted simple hull of each slice to ensure that
2632 * we have a single basic set per offset. The slicing constraint
2633 * may get simplified away before the unshifted simple hull is taken
2634 * and may therefore in some rare cases disappear from the result.
2635 * We therefore explicitly add the constraint back after computing
2636 * the unshifted simple hull to ensure that the basic sets
2637 * remain disjoint. The constraints that are dropped by taking the hull
2638 * will be taken into account at the next level, as in the case of the
2641 * Finally, we map i' back to i and call "fn".
2643 static int foreach_iteration(__isl_take isl_set
*domain
,
2644 __isl_keep isl_ast_build
*build
, int (*init
)(int n
, void *user
),
2645 int (*fn
)(__isl_take isl_basic_set
*bset
, void *user
), void *user
)
2650 isl_multi_aff
*expansion
;
2651 isl_basic_map
*bmap
;
2652 isl_aff
*lower
= NULL
;
2653 isl_ast_build
*stride_build
;
2655 depth
= isl_ast_build_get_depth(build
);
2657 domain
= isl_ast_build_eliminate_inner(build
, domain
);
2658 domain
= isl_set_intersect(domain
, isl_ast_build_get_domain(build
));
2659 stride_build
= isl_ast_build_copy(build
);
2660 stride_build
= isl_ast_build_detect_strides(stride_build
,
2661 isl_set_copy(domain
));
2662 expansion
= isl_ast_build_get_stride_expansion(stride_build
);
2664 domain
= isl_set_preimage_multi_aff(domain
,
2665 isl_multi_aff_copy(expansion
));
2666 domain
= isl_ast_build_eliminate_divs(stride_build
, domain
);
2667 isl_ast_build_free(stride_build
);
2669 bmap
= isl_basic_map_from_multi_aff(expansion
);
2671 empty
= isl_set_is_empty(domain
);
2677 lower
= find_unroll_lower_bound(build
, domain
, depth
, bmap
, &n
);
2681 if (n
>= 0 && init
&& init(n
, user
) < 0)
2683 for (i
= 0; i
< n
; ++i
) {
2685 isl_basic_set
*bset
;
2686 isl_constraint
*slice
;
2688 slice
= at_offset(depth
, lower
, i
);
2689 set
= isl_set_copy(domain
);
2690 set
= isl_set_add_constraint(set
, isl_constraint_copy(slice
));
2691 bset
= isl_set_unshifted_simple_hull(set
);
2692 bset
= isl_basic_set_add_constraint(bset
, slice
);
2693 bset
= isl_basic_set_apply(bset
, isl_basic_map_copy(bmap
));
2695 if (fn(bset
, user
) < 0)
2699 isl_aff_free(lower
);
2700 isl_set_free(domain
);
2701 isl_basic_map_free(bmap
);
2703 return n
< 0 || i
< n
? -1 : 0;
2706 /* Data structure for storing the results and the intermediate objects
2707 * of compute_domains.
2709 * "list" is the main result of the function and contains a list
2710 * of disjoint basic sets for which code should be generated.
2712 * "executed" and "build" are inputs to compute_domains.
2713 * "schedule_domain" is the domain of "executed".
2715 * "option" contains the domains at the current depth that should by
2716 * atomic, separated or unrolled. These domains are as specified by
2717 * the user, except that inner dimensions have been eliminated and
2718 * that they have been made pair-wise disjoint.
2720 * "sep_class" contains the user-specified split into separation classes
2721 * specialized to the current depth.
2722 * "done" contains the union of the separation domains that have already
2725 struct isl_codegen_domains
{
2726 isl_basic_set_list
*list
;
2728 isl_union_map
*executed
;
2729 isl_ast_build
*build
;
2730 isl_set
*schedule_domain
;
2738 /* Internal data structure for do_unroll.
2740 * "domains" stores the results of compute_domains.
2741 * "class_domain" is the original class domain passed to do_unroll.
2742 * "unroll_domain" collects the unrolled iterations.
2744 struct isl_ast_unroll_data
{
2745 struct isl_codegen_domains
*domains
;
2746 isl_set
*class_domain
;
2747 isl_set
*unroll_domain
;
2750 /* Given an iteration of an unrolled domain represented by "bset",
2751 * add it to data->domains->list.
2752 * Since we may have dropped some constraints, we intersect with
2753 * the class domain again to ensure that each element in the list
2754 * is disjoint from the other class domains.
2756 static int do_unroll_iteration(__isl_take isl_basic_set
*bset
, void *user
)
2758 struct isl_ast_unroll_data
*data
= user
;
2760 isl_basic_set_list
*list
;
2762 set
= isl_set_from_basic_set(bset
);
2763 data
->unroll_domain
= isl_set_union(data
->unroll_domain
,
2765 set
= isl_set_intersect(set
, isl_set_copy(data
->class_domain
));
2766 set
= isl_set_make_disjoint(set
);
2767 list
= isl_basic_set_list_from_set(set
);
2768 data
->domains
->list
= isl_basic_set_list_concat(data
->domains
->list
,
2774 /* Extend domains->list with a list of basic sets, one for each value
2775 * of the current dimension in "domain" and remove the corresponding
2776 * sets from the class domain. Return the updated class domain.
2777 * The divs that involve the current dimension have not been projected out
2780 * We call foreach_iteration to iterate over the individual values and
2781 * in do_unroll_iteration we collect the individual basic sets in
2782 * domains->list and their union in data->unroll_domain, which is then
2783 * used to update the class domain.
2785 static __isl_give isl_set
*do_unroll(struct isl_codegen_domains
*domains
,
2786 __isl_take isl_set
*domain
, __isl_take isl_set
*class_domain
)
2788 struct isl_ast_unroll_data data
;
2791 return isl_set_free(class_domain
);
2793 return isl_set_free(domain
);
2795 data
.domains
= domains
;
2796 data
.class_domain
= class_domain
;
2797 data
.unroll_domain
= isl_set_empty(isl_set_get_space(domain
));
2799 if (foreach_iteration(domain
, domains
->build
, NULL
,
2800 &do_unroll_iteration
, &data
) < 0)
2801 data
.unroll_domain
= isl_set_free(data
.unroll_domain
);
2803 class_domain
= isl_set_subtract(class_domain
, data
.unroll_domain
);
2805 return class_domain
;
2808 /* Add domains to domains->list for each individual value of the current
2809 * dimension, for that part of the schedule domain that lies in the
2810 * intersection of the option domain and the class domain.
2811 * Remove the corresponding sets from the class domain and
2812 * return the updated class domain.
2814 * We first break up the unroll option domain into individual pieces
2815 * and then handle each of them separately. The unroll option domain
2816 * has been made disjoint in compute_domains_init_options,
2818 * Note that we actively want to combine different pieces of the
2819 * schedule domain that have the same value at the current dimension.
2820 * We therefore need to break up the unroll option domain before
2821 * intersecting with class and schedule domain, hoping that the
2822 * unroll option domain specified by the user is relatively simple.
2824 static __isl_give isl_set
*compute_unroll_domains(
2825 struct isl_codegen_domains
*domains
, __isl_take isl_set
*class_domain
)
2827 isl_set
*unroll_domain
;
2828 isl_basic_set_list
*unroll_list
;
2833 empty
= isl_set_is_empty(domains
->option
[isl_ast_loop_unroll
]);
2835 return isl_set_free(class_domain
);
2837 return class_domain
;
2839 unroll_domain
= isl_set_copy(domains
->option
[isl_ast_loop_unroll
]);
2840 unroll_list
= isl_basic_set_list_from_set(unroll_domain
);
2842 n
= isl_basic_set_list_n_basic_set(unroll_list
);
2844 class_domain
= isl_set_free(class_domain
);
2845 for (i
= 0; i
< n
; ++i
) {
2846 isl_basic_set
*bset
;
2848 bset
= isl_basic_set_list_get_basic_set(unroll_list
, i
);
2849 unroll_domain
= isl_set_from_basic_set(bset
);
2850 unroll_domain
= isl_set_intersect(unroll_domain
,
2851 isl_set_copy(class_domain
));
2852 unroll_domain
= isl_set_intersect(unroll_domain
,
2853 isl_set_copy(domains
->schedule_domain
));
2855 empty
= isl_set_is_empty(unroll_domain
);
2856 if (empty
>= 0 && empty
) {
2857 isl_set_free(unroll_domain
);
2861 class_domain
= do_unroll(domains
, unroll_domain
, class_domain
);
2864 isl_basic_set_list_free(unroll_list
);
2866 return class_domain
;
2869 /* Try and construct a single basic set that includes the intersection of
2870 * the schedule domain, the atomic option domain and the class domain.
2871 * Add the resulting basic set(s) to domains->list and remove them
2872 * from class_domain. Return the updated class domain.
2874 * We construct a single domain rather than trying to combine
2875 * the schedule domains of individual domains because we are working
2876 * within a single component so that non-overlapping schedule domains
2877 * should already have been separated.
2878 * We do however need to make sure that this single domains is a subset
2879 * of the class domain so that it would not intersect with any other
2880 * class domains. This means that we may end up splitting up the atomic
2881 * domain in case separation classes are being used.
2883 * "domain" is the intersection of the schedule domain and the class domain,
2884 * with inner dimensions projected out.
2886 static __isl_give isl_set
*compute_atomic_domain(
2887 struct isl_codegen_domains
*domains
, __isl_take isl_set
*class_domain
)
2889 isl_basic_set
*bset
;
2890 isl_basic_set_list
*list
;
2891 isl_set
*domain
, *atomic_domain
;
2894 domain
= isl_set_copy(domains
->option
[isl_ast_loop_atomic
]);
2895 domain
= isl_set_intersect(domain
, isl_set_copy(class_domain
));
2896 domain
= isl_set_intersect(domain
,
2897 isl_set_copy(domains
->schedule_domain
));
2898 empty
= isl_set_is_empty(domain
);
2900 class_domain
= isl_set_free(class_domain
);
2902 isl_set_free(domain
);
2903 return class_domain
;
2906 domain
= isl_ast_build_eliminate(domains
->build
, domain
);
2907 domain
= isl_set_coalesce_preserve(domain
);
2908 bset
= isl_set_unshifted_simple_hull(domain
);
2909 domain
= isl_set_from_basic_set(bset
);
2910 atomic_domain
= isl_set_copy(domain
);
2911 domain
= isl_set_intersect(domain
, isl_set_copy(class_domain
));
2912 class_domain
= isl_set_subtract(class_domain
, atomic_domain
);
2913 domain
= isl_set_make_disjoint(domain
);
2914 list
= isl_basic_set_list_from_set(domain
);
2915 domains
->list
= isl_basic_set_list_concat(domains
->list
, list
);
2917 return class_domain
;
2920 /* Split up the schedule domain into uniform basic sets,
2921 * in the sense that each element in a basic set is associated to
2922 * elements of the same domains, and add the result to domains->list.
2923 * Do this for that part of the schedule domain that lies in the
2924 * intersection of "class_domain" and the separate option domain.
2926 * "class_domain" may or may not include the constraints
2927 * of the schedule domain, but this does not make a difference
2928 * since we are going to intersect it with the domain of the inverse schedule.
2929 * If it includes schedule domain constraints, then they may involve
2930 * inner dimensions, but we will eliminate them in separation_domain.
2932 static int compute_separate_domain(struct isl_codegen_domains
*domains
,
2933 __isl_keep isl_set
*class_domain
)
2937 isl_union_map
*executed
;
2938 isl_basic_set_list
*list
;
2941 domain
= isl_set_copy(domains
->option
[isl_ast_loop_separate
]);
2942 domain
= isl_set_intersect(domain
, isl_set_copy(class_domain
));
2943 executed
= isl_union_map_copy(domains
->executed
);
2944 executed
= isl_union_map_intersect_domain(executed
,
2945 isl_union_set_from_set(domain
));
2946 empty
= isl_union_map_is_empty(executed
);
2947 if (empty
< 0 || empty
) {
2948 isl_union_map_free(executed
);
2949 return empty
< 0 ? -1 : 0;
2952 space
= isl_set_get_space(class_domain
);
2953 domain
= separate_schedule_domains(space
, executed
, domains
->build
);
2955 list
= isl_basic_set_list_from_set(domain
);
2956 domains
->list
= isl_basic_set_list_concat(domains
->list
, list
);
2961 /* Split up the domain at the current depth into disjoint
2962 * basic sets for which code should be generated separately
2963 * for the given separation class domain.
2965 * If any separation classes have been defined, then "class_domain"
2966 * is the domain of the current class and does not refer to inner dimensions.
2967 * Otherwise, "class_domain" is the universe domain.
2969 * We first make sure that the class domain is disjoint from
2970 * previously considered class domains.
2972 * The separate domains can be computed directly from the "class_domain".
2974 * The unroll, atomic and remainder domains need the constraints
2975 * from the schedule domain.
2977 * For unrolling, the actual schedule domain is needed (with divs that
2978 * may refer to the current dimension) so that stride detection can be
2981 * For atomic and remainder domains, inner dimensions and divs involving
2982 * the current dimensions should be eliminated.
2983 * In case we are working within a separation class, we need to intersect
2984 * the result with the current "class_domain" to ensure that the domains
2985 * are disjoint from those generated from other class domains.
2987 * The domain that has been made atomic may be larger than specified
2988 * by the user since it needs to be representable as a single basic set.
2989 * This possibly larger domain is removed from class_domain by
2990 * compute_atomic_domain. It is computed first so that the extended domain
2991 * would not overlap with any domains computed before.
2992 * Similary, the unrolled domains may have some constraints removed and
2993 * may therefore also be larger than specified by the user.
2995 * If anything is left after handling separate, unroll and atomic,
2996 * we split it up into basic sets and append the basic sets to domains->list.
2998 static isl_stat
compute_partial_domains(struct isl_codegen_domains
*domains
,
2999 __isl_take isl_set
*class_domain
)
3001 isl_basic_set_list
*list
;
3004 class_domain
= isl_set_subtract(class_domain
,
3005 isl_set_copy(domains
->done
));
3006 domains
->done
= isl_set_union(domains
->done
,
3007 isl_set_copy(class_domain
));
3009 class_domain
= compute_atomic_domain(domains
, class_domain
);
3010 class_domain
= compute_unroll_domains(domains
, class_domain
);
3012 domain
= isl_set_copy(class_domain
);
3014 if (compute_separate_domain(domains
, domain
) < 0)
3016 domain
= isl_set_subtract(domain
,
3017 isl_set_copy(domains
->option
[isl_ast_loop_separate
]));
3019 domain
= isl_set_intersect(domain
,
3020 isl_set_copy(domains
->schedule_domain
));
3022 domain
= isl_ast_build_eliminate(domains
->build
, domain
);
3023 domain
= isl_set_intersect(domain
, isl_set_copy(class_domain
));
3025 domain
= isl_set_coalesce_preserve(domain
);
3026 domain
= isl_set_make_disjoint(domain
);
3028 list
= isl_basic_set_list_from_set(domain
);
3029 domains
->list
= isl_basic_set_list_concat(domains
->list
, list
);
3031 isl_set_free(class_domain
);
3035 isl_set_free(domain
);
3036 isl_set_free(class_domain
);
3037 return isl_stat_error
;
3040 /* Split up the domain at the current depth into disjoint
3041 * basic sets for which code should be generated separately
3042 * for the separation class identified by "pnt".
3044 * We extract the corresponding class domain from domains->sep_class,
3045 * eliminate inner dimensions and pass control to compute_partial_domains.
3047 static isl_stat
compute_class_domains(__isl_take isl_point
*pnt
, void *user
)
3049 struct isl_codegen_domains
*domains
= user
;
3054 class_set
= isl_set_from_point(pnt
);
3055 domain
= isl_map_domain(isl_map_intersect_range(
3056 isl_map_copy(domains
->sep_class
), class_set
));
3057 domain
= isl_ast_build_compute_gist(domains
->build
, domain
);
3058 domain
= isl_ast_build_eliminate(domains
->build
, domain
);
3060 disjoint
= isl_set_plain_is_disjoint(domain
, domains
->schedule_domain
);
3062 return isl_stat_error
;
3064 isl_set_free(domain
);
3068 return compute_partial_domains(domains
, domain
);
3071 /* Extract the domains at the current depth that should be atomic,
3072 * separated or unrolled and store them in option.
3074 * The domains specified by the user might overlap, so we make
3075 * them disjoint by subtracting earlier domains from later domains.
3077 static void compute_domains_init_options(isl_set
*option
[4],
3078 __isl_keep isl_ast_build
*build
)
3080 enum isl_ast_loop_type type
, type2
;
3083 for (type
= isl_ast_loop_atomic
;
3084 type
<= isl_ast_loop_separate
; ++type
) {
3085 option
[type
] = isl_ast_build_get_option_domain(build
, type
);
3086 for (type2
= isl_ast_loop_atomic
; type2
< type
; ++type2
)
3087 option
[type
] = isl_set_subtract(option
[type
],
3088 isl_set_copy(option
[type2
]));
3091 unroll
= option
[isl_ast_loop_unroll
];
3092 unroll
= isl_set_coalesce(unroll
);
3093 unroll
= isl_set_make_disjoint(unroll
);
3094 option
[isl_ast_loop_unroll
] = unroll
;
3097 /* Split up the domain at the current depth into disjoint
3098 * basic sets for which code should be generated separately,
3099 * based on the user-specified options.
3100 * Return the list of disjoint basic sets.
3102 * There are three kinds of domains that we need to keep track of.
3103 * - the "schedule domain" is the domain of "executed"
3104 * - the "class domain" is the domain corresponding to the currrent
3106 * - the "option domain" is the domain corresponding to one of the options
3107 * atomic, unroll or separate
3109 * We first consider the individial values of the separation classes
3110 * and split up the domain for each of them separately.
3111 * Finally, we consider the remainder. If no separation classes were
3112 * specified, then we call compute_partial_domains with the universe
3113 * "class_domain". Otherwise, we take the "schedule_domain" as "class_domain",
3114 * with inner dimensions removed. We do this because we want to
3115 * avoid computing the complement of the class domains (i.e., the difference
3116 * between the universe and domains->done).
3118 static __isl_give isl_basic_set_list
*compute_domains(
3119 __isl_keep isl_union_map
*executed
, __isl_keep isl_ast_build
*build
)
3121 struct isl_codegen_domains domains
;
3124 isl_union_set
*schedule_domain
;
3128 enum isl_ast_loop_type type
;
3134 ctx
= isl_union_map_get_ctx(executed
);
3135 domains
.list
= isl_basic_set_list_alloc(ctx
, 0);
3137 schedule_domain
= isl_union_map_domain(isl_union_map_copy(executed
));
3138 domain
= isl_set_from_union_set(schedule_domain
);
3140 compute_domains_init_options(domains
.option
, build
);
3142 domains
.sep_class
= isl_ast_build_get_separation_class(build
);
3143 classes
= isl_map_range(isl_map_copy(domains
.sep_class
));
3144 n_param
= isl_set_dim(classes
, isl_dim_param
);
3146 classes
= isl_set_free(classes
);
3147 classes
= isl_set_project_out(classes
, isl_dim_param
, 0, n_param
);
3149 space
= isl_set_get_space(domain
);
3150 domains
.build
= build
;
3151 domains
.schedule_domain
= isl_set_copy(domain
);
3152 domains
.executed
= executed
;
3153 domains
.done
= isl_set_empty(space
);
3155 if (isl_set_foreach_point(classes
, &compute_class_domains
, &domains
) < 0)
3156 domains
.list
= isl_basic_set_list_free(domains
.list
);
3157 isl_set_free(classes
);
3159 empty
= isl_set_is_empty(domains
.done
);
3161 domains
.list
= isl_basic_set_list_free(domains
.list
);
3162 domain
= isl_set_free(domain
);
3164 isl_set_free(domain
);
3165 domain
= isl_set_universe(isl_set_get_space(domains
.done
));
3167 domain
= isl_ast_build_eliminate(build
, domain
);
3169 if (compute_partial_domains(&domains
, domain
) < 0)
3170 domains
.list
= isl_basic_set_list_free(domains
.list
);
3172 isl_set_free(domains
.schedule_domain
);
3173 isl_set_free(domains
.done
);
3174 isl_map_free(domains
.sep_class
);
3175 for (type
= isl_ast_loop_atomic
; type
<= isl_ast_loop_separate
; ++type
)
3176 isl_set_free(domains
.option
[type
]);
3178 return domains
.list
;
3181 /* Generate code for a single component, after shifting (if any)
3182 * has been applied, in case the schedule was specified as a union map.
3184 * We first split up the domain at the current depth into disjoint
3185 * basic sets based on the user-specified options.
3186 * Then we generated code for each of them and concatenate the results.
3188 static __isl_give isl_ast_graft_list
*generate_shifted_component_flat(
3189 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
3191 isl_basic_set_list
*domain_list
;
3192 isl_ast_graft_list
*list
= NULL
;
3194 domain_list
= compute_domains(executed
, build
);
3195 list
= generate_parallel_domains(domain_list
, executed
, build
);
3197 isl_basic_set_list_free(domain_list
);
3198 isl_union_map_free(executed
);
3199 isl_ast_build_free(build
);
3204 /* Generate code for a single component, after shifting (if any)
3205 * has been applied, in case the schedule was specified as a schedule tree
3206 * and the separate option was specified.
3208 * We perform separation on the domain of "executed" and then generate
3209 * an AST for each of the resulting disjoint basic sets.
3211 static __isl_give isl_ast_graft_list
*generate_shifted_component_tree_separate(
3212 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
3216 isl_basic_set_list
*domain_list
;
3217 isl_ast_graft_list
*list
;
3219 space
= isl_ast_build_get_space(build
, 1);
3220 domain
= separate_schedule_domains(space
,
3221 isl_union_map_copy(executed
), build
);
3222 domain_list
= isl_basic_set_list_from_set(domain
);
3224 list
= generate_parallel_domains(domain_list
, executed
, build
);
3226 isl_basic_set_list_free(domain_list
);
3227 isl_union_map_free(executed
);
3228 isl_ast_build_free(build
);
3233 /* Internal data structure for generate_shifted_component_tree_unroll.
3235 * "executed" and "build" are inputs to generate_shifted_component_tree_unroll.
3236 * "list" collects the constructs grafts.
3238 struct isl_ast_unroll_tree_data
{
3239 isl_union_map
*executed
;
3240 isl_ast_build
*build
;
3241 isl_ast_graft_list
*list
;
3244 /* Initialize data->list to a list of "n" elements.
3246 static int init_unroll_tree(int n
, void *user
)
3248 struct isl_ast_unroll_tree_data
*data
= user
;
3251 ctx
= isl_ast_build_get_ctx(data
->build
);
3252 data
->list
= isl_ast_graft_list_alloc(ctx
, n
);
3257 /* Given an iteration of an unrolled domain represented by "bset",
3258 * generate the corresponding AST and add the result to data->list.
3260 static int do_unroll_tree_iteration(__isl_take isl_basic_set
*bset
, void *user
)
3262 struct isl_ast_unroll_tree_data
*data
= user
;
3264 data
->list
= add_node(data
->list
, isl_union_map_copy(data
->executed
),
3265 bset
, isl_ast_build_copy(data
->build
));
3270 /* Generate code for a single component, after shifting (if any)
3271 * has been applied, in case the schedule was specified as a schedule tree
3272 * and the unroll option was specified.
3274 * We call foreach_iteration to iterate over the individual values and
3275 * construct and collect the corresponding grafts in do_unroll_tree_iteration.
3277 static __isl_give isl_ast_graft_list
*generate_shifted_component_tree_unroll(
3278 __isl_take isl_union_map
*executed
, __isl_take isl_set
*domain
,
3279 __isl_take isl_ast_build
*build
)
3281 struct isl_ast_unroll_tree_data data
= { executed
, build
, NULL
};
3283 if (foreach_iteration(domain
, build
, &init_unroll_tree
,
3284 &do_unroll_tree_iteration
, &data
) < 0)
3285 data
.list
= isl_ast_graft_list_free(data
.list
);
3287 isl_union_map_free(executed
);
3288 isl_ast_build_free(build
);
3293 /* Does "domain" involve a disjunction that is purely based on
3294 * constraints involving only outer dimension?
3296 * In particular, is there a disjunction such that the constraints
3297 * involving the current and later dimensions are the same over
3298 * all the disjuncts?
3300 static isl_bool
has_pure_outer_disjunction(__isl_keep isl_set
*domain
,
3301 __isl_keep isl_ast_build
*build
)
3303 isl_basic_set
*hull
;
3304 isl_set
*shared
, *inner
;
3310 n
= isl_set_n_basic_set(domain
);
3312 return isl_bool_error
;
3314 return isl_bool_false
;
3315 dim
= isl_set_dim(domain
, isl_dim_set
);
3317 return isl_bool_error
;
3319 inner
= isl_set_copy(domain
);
3320 depth
= isl_ast_build_get_depth(build
);
3321 inner
= isl_set_drop_constraints_not_involving_dims(inner
,
3322 isl_dim_set
, depth
, dim
- depth
);
3323 hull
= isl_set_plain_unshifted_simple_hull(isl_set_copy(inner
));
3324 shared
= isl_set_from_basic_set(hull
);
3325 equal
= isl_set_plain_is_equal(inner
, shared
);
3326 isl_set_free(inner
);
3327 isl_set_free(shared
);
3332 /* Generate code for a single component, after shifting (if any)
3333 * has been applied, in case the schedule was specified as a schedule tree.
3334 * In particular, handle the base case where there is either no isolated
3335 * set or we are within the isolated set (in which case "isolated" is set)
3336 * or the iterations that precede or follow the isolated set.
3338 * The schedule domain is broken up or combined into basic sets
3339 * according to the AST generation option specified in the current
3340 * schedule node, which may be either atomic, separate, unroll or
3341 * unspecified. If the option is unspecified, then we currently simply
3342 * split the schedule domain into disjoint basic sets.
3344 * In case the separate option is specified, the AST generation is
3345 * handled by generate_shifted_component_tree_separate.
3346 * In the other cases, we need the global schedule domain.
3347 * In the unroll case, the AST generation is then handled by
3348 * generate_shifted_component_tree_unroll which needs the actual
3349 * schedule domain (with divs that may refer to the current dimension)
3350 * so that stride detection can be performed.
3351 * In the atomic or unspecified case, inner dimensions and divs involving
3352 * the current dimensions should be eliminated.
3353 * The result is then either combined into a single basic set or
3354 * split up into disjoint basic sets.
3355 * Finally an AST is generated for each basic set and the results are
3358 * If the schedule domain involves a disjunction that is purely based on
3359 * constraints involving only outer dimension, then it is treated as
3360 * if atomic was specified. This ensures that only a single loop
3361 * is generated instead of a sequence of identical loops with
3364 static __isl_give isl_ast_graft_list
*generate_shifted_component_tree_base(
3365 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
,
3368 isl_bool outer_disjunction
;
3369 isl_union_set
*schedule_domain
;
3371 isl_basic_set_list
*domain_list
;
3372 isl_ast_graft_list
*list
;
3373 enum isl_ast_loop_type type
;
3375 type
= isl_ast_build_get_loop_type(build
, isolated
);
3379 if (type
== isl_ast_loop_separate
)
3380 return generate_shifted_component_tree_separate(executed
,
3383 schedule_domain
= isl_union_map_domain(isl_union_map_copy(executed
));
3384 domain
= isl_set_from_union_set(schedule_domain
);
3386 if (type
== isl_ast_loop_unroll
)
3387 return generate_shifted_component_tree_unroll(executed
, domain
,
3390 domain
= isl_ast_build_eliminate(build
, domain
);
3391 domain
= isl_set_coalesce_preserve(domain
);
3393 outer_disjunction
= has_pure_outer_disjunction(domain
, build
);
3394 if (outer_disjunction
< 0)
3395 domain
= isl_set_free(domain
);
3397 if (outer_disjunction
|| type
== isl_ast_loop_atomic
) {
3398 isl_basic_set
*hull
;
3399 hull
= isl_set_unshifted_simple_hull(domain
);
3400 domain_list
= isl_basic_set_list_from_basic_set(hull
);
3402 domain
= isl_set_make_disjoint(domain
);
3403 domain_list
= isl_basic_set_list_from_set(domain
);
3406 list
= generate_parallel_domains(domain_list
, executed
, build
);
3408 isl_basic_set_list_free(domain_list
);
3409 isl_union_map_free(executed
);
3410 isl_ast_build_free(build
);
3414 isl_union_map_free(executed
);
3415 isl_ast_build_free(build
);
3419 /* Extract out the disjunction imposed by "domain" on the outer
3420 * schedule dimensions.
3422 * In particular, remove all inner dimensions from "domain" (including
3423 * the current dimension) and then remove the constraints that are shared
3424 * by all disjuncts in the result.
3426 static __isl_give isl_set
*extract_disjunction(__isl_take isl_set
*domain
,
3427 __isl_keep isl_ast_build
*build
)
3433 domain
= isl_ast_build_specialize(build
, domain
);
3434 depth
= isl_ast_build_get_depth(build
);
3435 dim
= isl_set_dim(domain
, isl_dim_set
);
3437 return isl_set_free(domain
);
3438 domain
= isl_set_eliminate(domain
, isl_dim_set
, depth
, dim
- depth
);
3439 domain
= isl_set_remove_unknown_divs(domain
);
3440 hull
= isl_set_copy(domain
);
3441 hull
= isl_set_from_basic_set(isl_set_unshifted_simple_hull(hull
));
3442 domain
= isl_set_gist(domain
, hull
);
3447 /* Add "guard" to the grafts in "list".
3448 * "build" is the outer AST build, while "sub_build" includes "guard"
3449 * in its generated domain.
3451 * First combine the grafts into a single graft and then add the guard.
3452 * If the list is empty, or if some error occurred, then simply return
3455 static __isl_give isl_ast_graft_list
*list_add_guard(
3456 __isl_take isl_ast_graft_list
*list
, __isl_keep isl_set
*guard
,
3457 __isl_keep isl_ast_build
*build
, __isl_keep isl_ast_build
*sub_build
)
3459 isl_ast_graft
*graft
;
3462 list
= isl_ast_graft_list_fuse(list
, sub_build
);
3464 n
= isl_ast_graft_list_n_ast_graft(list
);
3466 return isl_ast_graft_list_free(list
);
3470 graft
= isl_ast_graft_list_get_ast_graft(list
, 0);
3471 graft
= isl_ast_graft_add_guard(graft
, isl_set_copy(guard
), build
);
3472 list
= isl_ast_graft_list_set_ast_graft(list
, 0, graft
);
3477 /* Generate code for a single component, after shifting (if any)
3478 * has been applied, in case the schedule was specified as a schedule tree.
3479 * In particular, do so for the specified subset of the schedule domain.
3481 * If we are outside of the isolated part, then "domain" may include
3482 * a disjunction. Explicitly generate this disjunction at this point
3483 * instead of relying on the disjunction getting hoisted back up
3486 static __isl_give isl_ast_graft_list
*generate_shifted_component_tree_part(
3487 __isl_keep isl_union_map
*executed
, __isl_take isl_set
*domain
,
3488 __isl_keep isl_ast_build
*build
, int isolated
)
3490 isl_union_set
*uset
;
3491 isl_ast_graft_list
*list
;
3492 isl_ast_build
*sub_build
;
3495 uset
= isl_union_set_from_set(isl_set_copy(domain
));
3496 executed
= isl_union_map_copy(executed
);
3497 executed
= isl_union_map_intersect_domain(executed
, uset
);
3498 empty
= isl_union_map_is_empty(executed
);
3503 isl_union_map_free(executed
);
3504 isl_set_free(domain
);
3505 ctx
= isl_ast_build_get_ctx(build
);
3506 return isl_ast_graft_list_alloc(ctx
, 0);
3509 sub_build
= isl_ast_build_copy(build
);
3511 domain
= extract_disjunction(domain
, build
);
3512 sub_build
= isl_ast_build_restrict_generated(sub_build
,
3513 isl_set_copy(domain
));
3515 list
= generate_shifted_component_tree_base(executed
,
3516 isl_ast_build_copy(sub_build
), isolated
);
3518 list
= list_add_guard(list
, domain
, build
, sub_build
);
3519 isl_ast_build_free(sub_build
);
3520 isl_set_free(domain
);
3523 isl_union_map_free(executed
);
3524 isl_set_free(domain
);
3528 /* Generate code for a single component, after shifting (if any)
3529 * has been applied, in case the schedule was specified as a schedule tree.
3530 * In particular, do so for the specified sequence of subsets
3531 * of the schedule domain, "before", "isolated", "after" and "other",
3532 * where only the "isolated" part is considered to be isolated.
3534 static __isl_give isl_ast_graft_list
*generate_shifted_component_parts(
3535 __isl_take isl_union_map
*executed
, __isl_take isl_set
*before
,
3536 __isl_take isl_set
*isolated
, __isl_take isl_set
*after
,
3537 __isl_take isl_set
*other
, __isl_take isl_ast_build
*build
)
3539 isl_ast_graft_list
*list
, *res
;
3541 res
= generate_shifted_component_tree_part(executed
, before
, build
, 0);
3542 list
= generate_shifted_component_tree_part(executed
, isolated
,
3544 res
= isl_ast_graft_list_concat(res
, list
);
3545 list
= generate_shifted_component_tree_part(executed
, after
, build
, 0);
3546 res
= isl_ast_graft_list_concat(res
, list
);
3547 list
= generate_shifted_component_tree_part(executed
, other
, build
, 0);
3548 res
= isl_ast_graft_list_concat(res
, list
);
3550 isl_union_map_free(executed
);
3551 isl_ast_build_free(build
);
3556 /* Does "set" intersect "first", but not "second"?
3558 static isl_bool
only_intersects_first(__isl_keep isl_set
*set
,
3559 __isl_keep isl_set
*first
, __isl_keep isl_set
*second
)
3563 disjoint
= isl_set_is_disjoint(set
, first
);
3565 return isl_bool_error
;
3567 return isl_bool_false
;
3569 return isl_set_is_disjoint(set
, second
);
3572 /* Generate code for a single component, after shifting (if any)
3573 * has been applied, in case the schedule was specified as a schedule tree.
3574 * In particular, do so in case of isolation where there is
3575 * only an "isolated" part and an "after" part.
3576 * "dead1" and "dead2" are freed by this function in order to simplify
3579 * The "before" and "other" parts are set to empty sets.
3581 static __isl_give isl_ast_graft_list
*generate_shifted_component_only_after(
3582 __isl_take isl_union_map
*executed
, __isl_take isl_set
*isolated
,
3583 __isl_take isl_set
*after
, __isl_take isl_ast_build
*build
,
3584 __isl_take isl_set
*dead1
, __isl_take isl_set
*dead2
)
3588 empty
= isl_set_empty(isl_set_get_space(after
));
3589 isl_set_free(dead1
);
3590 isl_set_free(dead2
);
3591 return generate_shifted_component_parts(executed
, isl_set_copy(empty
),
3592 isolated
, after
, empty
, build
);
3595 /* Generate code for a single component, after shifting (if any)
3596 * has been applied, in case the schedule was specified as a schedule tree.
3598 * We first check if the user has specified an isolated schedule domain
3599 * and that we are not already outside of this isolated schedule domain.
3600 * If so, we break up the schedule domain into iterations that
3601 * precede the isolated domain, the isolated domain itself,
3602 * the iterations that follow the isolated domain and
3603 * the remaining iterations (those that are incomparable
3604 * to the isolated domain).
3605 * We generate an AST for each piece and concatenate the results.
3607 * If the isolated domain is not convex, then it is replaced
3608 * by a convex superset to ensure that the sets of preceding and
3609 * following iterations are properly defined and, in particular,
3610 * that there are no intermediate iterations that do not belong
3611 * to the isolated domain.
3613 * In the special case where at least one element of the schedule
3614 * domain that does not belong to the isolated domain needs
3615 * to be scheduled after this isolated domain, but none of those
3616 * elements need to be scheduled before, break up the schedule domain
3617 * in only two parts, the isolated domain, and a part that will be
3618 * scheduled after the isolated domain.
3620 * If no isolated set has been specified, then we generate an
3621 * AST for the entire inverse schedule.
3623 static __isl_give isl_ast_graft_list
*generate_shifted_component_tree(
3624 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
3627 int empty
, has_isolate
;
3629 isl_union_set
*schedule_domain
;
3631 isl_basic_set
*hull
;
3632 isl_set
*isolated
, *before
, *after
, *test
;
3636 build
= isl_ast_build_extract_isolated(build
);
3637 has_isolate
= isl_ast_build_has_isolated(build
);
3638 if (has_isolate
< 0)
3639 executed
= isl_union_map_free(executed
);
3640 else if (!has_isolate
)
3641 return generate_shifted_component_tree_base(executed
, build
, 0);
3643 schedule_domain
= isl_union_map_domain(isl_union_map_copy(executed
));
3644 domain
= isl_set_from_union_set(schedule_domain
);
3646 isolated
= isl_ast_build_get_isolated(build
);
3647 isolated
= isl_set_intersect(isolated
, isl_set_copy(domain
));
3648 test
= isl_ast_build_specialize(build
, isl_set_copy(isolated
));
3649 empty
= isl_set_is_empty(test
);
3654 isl_set_free(isolated
);
3655 isl_set_free(domain
);
3656 return generate_shifted_component_tree_base(executed
, build
, 0);
3658 isolated
= isl_ast_build_eliminate(build
, isolated
);
3659 hull
= isl_set_unshifted_simple_hull(isolated
);
3660 isolated
= isl_set_from_basic_set(hull
);
3662 depth
= isl_ast_build_get_depth(build
);
3663 space
= isl_space_map_from_set(isl_set_get_space(isolated
));
3664 gt
= isl_map_universe(space
);
3665 for (i
= 0; i
< depth
; ++i
)
3666 gt
= isl_map_equate(gt
, isl_dim_in
, i
, isl_dim_out
, i
);
3667 gt
= isl_map_order_gt(gt
, isl_dim_in
, depth
, isl_dim_out
, depth
);
3668 lt
= isl_map_reverse(isl_map_copy(gt
));
3669 before
= isl_set_apply(isl_set_copy(isolated
), gt
);
3670 after
= isl_set_apply(isl_set_copy(isolated
), lt
);
3672 domain
= isl_set_subtract(domain
, isl_set_copy(isolated
));
3673 pure
= only_intersects_first(domain
, after
, before
);
3675 executed
= isl_union_map_free(executed
);
3677 return generate_shifted_component_only_after(executed
, isolated
,
3678 domain
, build
, before
, after
);
3679 domain
= isl_set_subtract(domain
, isl_set_copy(before
));
3680 domain
= isl_set_subtract(domain
, isl_set_copy(after
));
3681 after
= isl_set_subtract(after
, isl_set_copy(isolated
));
3682 after
= isl_set_subtract(after
, isl_set_copy(before
));
3683 before
= isl_set_subtract(before
, isl_set_copy(isolated
));
3685 return generate_shifted_component_parts(executed
, before
, isolated
,
3686 after
, domain
, build
);
3688 isl_set_free(domain
);
3689 isl_set_free(isolated
);
3690 isl_union_map_free(executed
);
3691 isl_ast_build_free(build
);
3695 /* Generate code for a single component, after shifting (if any)
3698 * Call generate_shifted_component_tree or generate_shifted_component_flat
3699 * depending on whether the schedule was specified as a schedule tree.
3701 static __isl_give isl_ast_graft_list
*generate_shifted_component(
3702 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
3704 if (isl_ast_build_has_schedule_node(build
))
3705 return generate_shifted_component_tree(executed
, build
);
3707 return generate_shifted_component_flat(executed
, build
);
3710 struct isl_set_map_pair
{
3715 /* Given an array "domain" of isl_set_map_pairs and an array "order"
3716 * of indices into the "domain" array,
3717 * return the union of the "map" fields of the elements
3718 * indexed by the first "n" elements of "order".
3720 static __isl_give isl_union_map
*construct_component_executed(
3721 struct isl_set_map_pair
*domain
, int *order
, int n
)
3725 isl_union_map
*executed
;
3727 map
= isl_map_copy(domain
[order
[0]].map
);
3728 executed
= isl_union_map_from_map(map
);
3729 for (i
= 1; i
< n
; ++i
) {
3730 map
= isl_map_copy(domain
[order
[i
]].map
);
3731 executed
= isl_union_map_add_map(executed
, map
);
3737 /* Generate code for a single component, after shifting (if any)
3740 * The component inverse schedule is specified as the "map" fields
3741 * of the elements of "domain" indexed by the first "n" elements of "order".
3743 static __isl_give isl_ast_graft_list
*generate_shifted_component_from_list(
3744 struct isl_set_map_pair
*domain
, int *order
, int n
,
3745 __isl_take isl_ast_build
*build
)
3747 isl_union_map
*executed
;
3749 executed
= construct_component_executed(domain
, order
, n
);
3750 return generate_shifted_component(executed
, build
);
3753 /* Does set dimension "pos" of "set" have an obviously fixed value?
3755 static int dim_is_fixed(__isl_keep isl_set
*set
, int pos
)
3760 v
= isl_set_plain_get_val_if_fixed(set
, isl_dim_set
, pos
);
3763 fixed
= !isl_val_is_nan(v
);
3769 /* Given an array "domain" of isl_set_map_pairs and an array "order"
3770 * of indices into the "domain" array,
3771 * do all (except for at most one) of the "set" field of the elements
3772 * indexed by the first "n" elements of "order" have a fixed value
3773 * at position "depth"?
3775 static int at_most_one_non_fixed(struct isl_set_map_pair
*domain
,
3776 int *order
, int n
, int depth
)
3781 for (i
= 0; i
< n
; ++i
) {
3784 f
= dim_is_fixed(domain
[order
[i
]].set
, depth
);
3797 /* Given an array "domain" of isl_set_map_pairs and an array "order"
3798 * of indices into the "domain" array,
3799 * eliminate the inner dimensions from the "set" field of the elements
3800 * indexed by the first "n" elements of "order", provided the current
3801 * dimension does not have a fixed value.
3803 * Return the index of the first element in "order" with a corresponding
3804 * "set" field that does not have an (obviously) fixed value.
3806 static int eliminate_non_fixed(struct isl_set_map_pair
*domain
,
3807 int *order
, int n
, int depth
, __isl_keep isl_ast_build
*build
)
3812 for (i
= n
- 1; i
>= 0; --i
) {
3814 f
= dim_is_fixed(domain
[order
[i
]].set
, depth
);
3819 domain
[order
[i
]].set
= isl_ast_build_eliminate_inner(build
,
3820 domain
[order
[i
]].set
);
3827 /* Given an array "domain" of isl_set_map_pairs and an array "order"
3828 * of indices into the "domain" array,
3829 * find the element of "domain" (amongst those indexed by the first "n"
3830 * elements of "order") with the "set" field that has the smallest
3831 * value for the current iterator.
3833 * Note that the domain with the smallest value may depend on the parameters
3834 * and/or outer loop dimension. Since the result of this function is only
3835 * used as heuristic, we only make a reasonable attempt at finding the best
3836 * domain, one that should work in case a single domain provides the smallest
3837 * value for the current dimension over all values of the parameters
3838 * and outer dimensions.
3840 * In particular, we compute the smallest value of the first domain
3841 * and replace it by that of any later domain if that later domain
3842 * has a smallest value that is smaller for at least some value
3843 * of the parameters and outer dimensions.
3845 static int first_offset(struct isl_set_map_pair
*domain
, int *order
, int n
,
3846 __isl_keep isl_ast_build
*build
)
3852 min_first
= isl_ast_build_map_to_iterator(build
,
3853 isl_set_copy(domain
[order
[0]].set
));
3854 min_first
= isl_map_lexmin(min_first
);
3856 for (i
= 1; i
< n
; ++i
) {
3857 isl_map
*min
, *test
;
3860 min
= isl_ast_build_map_to_iterator(build
,
3861 isl_set_copy(domain
[order
[i
]].set
));
3862 min
= isl_map_lexmin(min
);
3863 test
= isl_map_copy(min
);
3864 test
= isl_map_apply_domain(isl_map_copy(min_first
), test
);
3865 test
= isl_map_order_lt(test
, isl_dim_in
, 0, isl_dim_out
, 0);
3866 empty
= isl_map_is_empty(test
);
3868 if (empty
>= 0 && !empty
) {
3869 isl_map_free(min_first
);
3879 isl_map_free(min_first
);
3881 return i
< n
? -1 : first
;
3884 /* Construct a shifted inverse schedule based on the original inverse schedule,
3885 * the stride and the offset.
3887 * The original inverse schedule is specified as the "map" fields
3888 * of the elements of "domain" indexed by the first "n" elements of "order".
3890 * "stride" and "offset" are such that the difference
3891 * between the values of the current dimension of domain "i"
3892 * and the values of the current dimension for some reference domain are
3895 * stride * integer + offset[i]
3897 * Moreover, 0 <= offset[i] < stride.
3899 * For each domain, we create a map
3901 * { [..., j, ...] -> [..., j - offset[i], offset[i], ....] }
3903 * where j refers to the current dimension and the other dimensions are
3904 * unchanged, and apply this map to the original schedule domain.
3906 * For example, for the original schedule
3908 * { A[i] -> [2i]: 0 <= i < 10; B[i] -> [2i+1] : 0 <= i < 10 }
3910 * and assuming the offset is 0 for the A domain and 1 for the B domain,
3911 * we apply the mapping
3915 * to the schedule of the "A" domain and the mapping
3917 * { [j - 1] -> [j, 1] }
3919 * to the schedule of the "B" domain.
3922 * Note that after the transformation, the differences between pairs
3923 * of values of the current dimension over all domains are multiples
3924 * of stride and that we have therefore exposed the stride.
3927 * To see that the mapping preserves the lexicographic order,
3928 * first note that each of the individual maps above preserves the order.
3929 * If the value of the current iterator is j1 in one domain and j2 in another,
3930 * then if j1 = j2, we know that the same map is applied to both domains
3931 * and the order is preserved.
3932 * Otherwise, let us assume, without loss of generality, that j1 < j2.
3933 * If c1 >= c2 (with c1 and c2 the corresponding offsets), then
3937 * and the order is preserved.
3938 * If c1 < c2, then we know
3944 * j2 - j1 = n * s + r
3946 * with n >= 0 and 0 <= r < s.
3947 * In other words, r = c2 - c1.
3958 * (j1 - c1, c1) << (j2 - c2, c2)
3960 * with "<<" the lexicographic order, proving that the order is preserved
3963 static __isl_give isl_union_map
*construct_shifted_executed(
3964 struct isl_set_map_pair
*domain
, int *order
, int n
,
3965 __isl_keep isl_val
*stride
, __isl_keep isl_multi_val
*offset
,
3966 __isl_take isl_ast_build
*build
)
3969 isl_union_map
*executed
;
3975 depth
= isl_ast_build_get_depth(build
);
3976 space
= isl_ast_build_get_space(build
, 1);
3977 executed
= isl_union_map_empty(isl_space_copy(space
));
3978 space
= isl_space_map_from_set(space
);
3979 map
= isl_map_identity(isl_space_copy(space
));
3980 map
= isl_map_eliminate(map
, isl_dim_out
, depth
, 1);
3981 map
= isl_map_insert_dims(map
, isl_dim_out
, depth
+ 1, 1);
3982 space
= isl_space_insert_dims(space
, isl_dim_out
, depth
+ 1, 1);
3984 c
= isl_constraint_alloc_equality(isl_local_space_from_space(space
));
3985 c
= isl_constraint_set_coefficient_si(c
, isl_dim_in
, depth
, 1);
3986 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
, depth
, -1);
3988 for (i
= 0; i
< n
; ++i
) {
3992 v
= isl_multi_val_get_val(offset
, i
);
3995 map_i
= isl_map_copy(map
);
3996 map_i
= isl_map_fix_val(map_i
, isl_dim_out
, depth
+ 1,
3999 c
= isl_constraint_set_constant_val(c
, v
);
4000 map_i
= isl_map_add_constraint(map_i
, isl_constraint_copy(c
));
4002 map_i
= isl_map_apply_domain(isl_map_copy(domain
[order
[i
]].map
),
4004 executed
= isl_union_map_add_map(executed
, map_i
);
4007 isl_constraint_free(c
);
4011 executed
= isl_union_map_free(executed
);
4016 /* Generate code for a single component, after exposing the stride,
4017 * given that the schedule domain is "shifted strided".
4019 * The component inverse schedule is specified as the "map" fields
4020 * of the elements of "domain" indexed by the first "n" elements of "order".
4022 * The schedule domain being "shifted strided" means that the differences
4023 * between the values of the current dimension of domain "i"
4024 * and the values of the current dimension for some reference domain are
4027 * stride * integer + offset[i]
4029 * We first look for the domain with the "smallest" value for the current
4030 * dimension and adjust the offsets such that the offset of the "smallest"
4031 * domain is equal to zero. The other offsets are reduced modulo stride.
4033 * Based on this information, we construct a new inverse schedule in
4034 * construct_shifted_executed that exposes the stride.
4035 * Since this involves the introduction of a new schedule dimension,
4036 * the build needs to be changed accordingly.
4037 * After computing the AST, the newly introduced dimension needs
4038 * to be removed again from the list of grafts. We do this by plugging
4039 * in a mapping that represents the new schedule domain in terms of the
4040 * old schedule domain.
4042 static __isl_give isl_ast_graft_list
*generate_shift_component(
4043 struct isl_set_map_pair
*domain
, int *order
, int n
,
4044 __isl_keep isl_val
*stride
, __isl_keep isl_multi_val
*offset
,
4045 __isl_take isl_ast_build
*build
)
4047 isl_ast_graft_list
*list
;
4053 isl_multi_aff
*ma
, *zero
;
4054 isl_union_map
*executed
;
4056 depth
= isl_ast_build_get_depth(build
);
4058 first
= first_offset(domain
, order
, n
, build
);
4062 mv
= isl_multi_val_copy(offset
);
4063 val
= isl_multi_val_get_val(offset
, first
);
4064 val
= isl_val_neg(val
);
4065 mv
= isl_multi_val_add_val(mv
, val
);
4066 mv
= isl_multi_val_mod_val(mv
, isl_val_copy(stride
));
4068 executed
= construct_shifted_executed(domain
, order
, n
, stride
, mv
,
4070 space
= isl_ast_build_get_space(build
, 1);
4071 space
= isl_space_map_from_set(space
);
4072 ma
= isl_multi_aff_identity(isl_space_copy(space
));
4073 space
= isl_space_from_domain(isl_space_domain(space
));
4074 space
= isl_space_add_dims(space
, isl_dim_out
, 1);
4075 zero
= isl_multi_aff_zero(space
);
4076 ma
= isl_multi_aff_range_splice(ma
, depth
+ 1, zero
);
4077 build
= isl_ast_build_insert_dim(build
, depth
+ 1);
4078 list
= generate_shifted_component(executed
, build
);
4080 list
= isl_ast_graft_list_preimage_multi_aff(list
, ma
);
4082 isl_multi_val_free(mv
);
4086 isl_ast_build_free(build
);
4090 /* Does any node in the schedule tree rooted at the current schedule node
4091 * of "build" depend on outer schedule nodes?
4093 static int has_anchored_subtree(__isl_keep isl_ast_build
*build
)
4095 isl_schedule_node
*node
;
4098 node
= isl_ast_build_get_schedule_node(build
);
4099 dependent
= isl_schedule_node_is_subtree_anchored(node
);
4100 isl_schedule_node_free(node
);
4105 /* Generate code for a single component.
4107 * The component inverse schedule is specified as the "map" fields
4108 * of the elements of "domain" indexed by the first "n" elements of "order".
4110 * This function may modify the "set" fields of "domain".
4112 * Before proceeding with the actual code generation for the component,
4113 * we first check if there are any "shifted" strides, meaning that
4114 * the schedule domains of the individual domains are all strided,
4115 * but that they have different offsets, resulting in the union
4116 * of schedule domains not being strided anymore.
4118 * The simplest example is the schedule
4120 * { A[i] -> [2i]: 0 <= i < 10; B[i] -> [2i+1] : 0 <= i < 10 }
4122 * Both schedule domains are strided, but their union is not.
4123 * This function detects such cases and then rewrites the schedule to
4125 * { A[i] -> [2i, 0]: 0 <= i < 10; B[i] -> [2i, 1] : 0 <= i < 10 }
4127 * In the new schedule, the schedule domains have the same offset (modulo
4128 * the stride), ensuring that the union of schedule domains is also strided.
4131 * If there is only a single domain in the component, then there is
4132 * nothing to do. Similarly, if the current schedule dimension has
4133 * a fixed value for almost all domains then there is nothing to be done.
4134 * In particular, we need at least two domains where the current schedule
4135 * dimension does not have a fixed value.
4136 * Finally, in case of a schedule map input,
4137 * if any of the options refer to the current schedule dimension,
4138 * then we bail out as well. It would be possible to reformulate the options
4139 * in terms of the new schedule domain, but that would introduce constraints
4140 * that separate the domains in the options and that is something we would
4142 * In the case of a schedule tree input, we bail out if any of
4143 * the descendants of the current schedule node refer to outer
4144 * schedule nodes in any way.
4147 * To see if there is any shifted stride, we look at the differences
4148 * between the values of the current dimension in pairs of domains
4149 * for equal values of outer dimensions. These differences should be
4154 * with "m" the stride and "r" a constant. Note that we cannot perform
4155 * this analysis on individual domains as the lower bound in each domain
4156 * may depend on parameters or outer dimensions and so the current dimension
4157 * itself may not have a fixed remainder on division by the stride.
4159 * In particular, we compare the first domain that does not have an
4160 * obviously fixed value for the current dimension to itself and all
4161 * other domains and collect the offsets and the gcd of the strides.
4162 * If the gcd becomes one, then we failed to find shifted strides.
4163 * If the gcd is zero, then the differences were all fixed, meaning
4164 * that some domains had non-obviously fixed values for the current dimension.
4165 * If all the offsets are the same (for those domains that do not have
4166 * an obviously fixed value for the current dimension), then we do not
4167 * apply the transformation.
4168 * If none of the domains were skipped, then there is nothing to do.
4169 * If some of them were skipped, then if we apply separation, the schedule
4170 * domain should get split in pieces with a (non-shifted) stride.
4172 * Otherwise, we apply a shift to expose the stride in
4173 * generate_shift_component.
4175 static __isl_give isl_ast_graft_list
*generate_component(
4176 struct isl_set_map_pair
*domain
, int *order
, int n
,
4177 __isl_take isl_ast_build
*build
)
4184 isl_val
*gcd
= NULL
;
4188 isl_ast_graft_list
*list
;
4191 depth
= isl_ast_build_get_depth(build
);
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
);
4200 skip
= isl_ast_build_options_involve_depth(build
);
4205 return generate_shifted_component_from_list(domain
,
4208 base
= eliminate_non_fixed(domain
, order
, n
, depth
, build
);
4212 ctx
= isl_ast_build_get_ctx(build
);
4214 mv
= isl_multi_val_zero(isl_space_set_alloc(ctx
, 0, n
));
4217 for (i
= 0; i
< n
; ++i
) {
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
,
4226 deltas
= isl_map_deltas(map
);
4227 res
= isl_set_dim_residue_class_val(deltas
, depth
, &m
, &r
);
4228 isl_set_free(deltas
);
4235 gcd
= isl_val_gcd(gcd
, m
);
4236 if (isl_val_is_one(gcd
)) {
4240 mv
= isl_multi_val_set_val(mv
, i
, r
);
4242 res
= dim_is_fixed(domain
[order
[i
]].set
, depth
);
4248 if (fixed
&& i
> base
) {
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
))
4259 if (res
< 0 || !gcd
) {
4260 isl_ast_build_free(build
);
4262 } else if (i
< n
|| fixed
|| isl_val_is_zero(gcd
)) {
4263 list
= generate_shifted_component_from_list(domain
,
4266 list
= generate_shift_component(domain
, order
, n
, gcd
, mv
,
4271 isl_multi_val_free(mv
);
4275 isl_ast_build_free(build
);
4279 /* Store both "map" itself and its domain in the
4280 * structure pointed to by *next and advance to the next array element.
4282 static isl_stat
extract_domain(__isl_take isl_map
*map
, void *user
)
4284 struct isl_set_map_pair
**next
= user
;
4286 (*next
)->map
= isl_map_copy(map
);
4287 (*next
)->set
= isl_map_domain(map
);
4293 static isl_bool
after_in_tree(__isl_keep isl_union_map
*umap
,
4294 __isl_keep isl_schedule_node
*node
);
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"?
4300 static isl_bool
after_in_child(__isl_keep isl_union_map
*umap
,
4301 __isl_keep isl_schedule_node
*node
)
4303 isl_schedule_node
*child
;
4306 child
= isl_schedule_node_get_child(node
, 0);
4307 after
= after_in_tree(umap
, child
);
4308 isl_schedule_node_free(child
);
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"?
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.
4325 static isl_bool
after_in_band(__isl_keep isl_union_map
*umap
,
4326 __isl_keep isl_schedule_node
*node
)
4328 isl_multi_union_pw_aff
*mupa
;
4329 isl_union_map
*partial
, *test
, *gt
, *universe
, *umap1
, *umap2
;
4330 isl_union_set
*domain
, *range
;
4336 n
= isl_schedule_node_band_n_member(node
);
4338 return isl_bool_error
;
4340 return after_in_child(umap
, node
);
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
);
4353 if (empty
< 0 || !empty
) {
4354 isl_union_map_free(partial
);
4355 return isl_bool_not(empty
);
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
);
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"?
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
4381 static isl_bool
after_in_context(__isl_keep isl_union_map
*umap
,
4382 __isl_keep isl_schedule_node
*node
)
4384 isl_union_map
*prefix
, *universe
, *umap1
, *umap2
;
4385 isl_union_set
*domain
, *range
;
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
);
4403 after
= after_in_child(umap
, node
);
4405 isl_union_map_free(umap
);
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"?
4414 * We apply the expansion to domain and range of "umap" and
4415 * continue with its child.
4417 static isl_bool
after_in_expansion(__isl_keep isl_union_map
*umap
,
4418 __isl_keep isl_schedule_node
*node
)
4420 isl_union_map
*expansion
;
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
);
4428 after
= after_in_child(umap
, node
);
4430 isl_union_map_free(umap
);
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"?
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.
4443 static isl_bool
after_in_extension(__isl_keep isl_union_map
*umap
,
4444 __isl_keep isl_schedule_node
*node
)
4446 return isl_bool_true
;
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"?
4453 * We intersect domain and range of "umap" with the filter and
4454 * continue with its child.
4456 static isl_bool
after_in_filter(__isl_keep isl_union_map
*umap
,
4457 __isl_keep isl_schedule_node
*node
)
4459 isl_union_set
*filter
;
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
);
4467 after
= after_in_child(umap
, node
);
4469 isl_union_map_free(umap
);
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"?
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
4484 static isl_bool
after_in_set(__isl_keep isl_union_map
*umap
,
4485 __isl_keep isl_schedule_node
*node
)
4490 n
= isl_schedule_node_n_children(node
);
4492 return isl_bool_error
;
4493 for (i
= 0; i
< n
; ++i
) {
4494 isl_schedule_node
*child
;
4497 child
= isl_schedule_node_get_child(node
, i
);
4498 after
= after_in_tree(umap
, child
);
4499 isl_schedule_node_free(child
);
4501 if (after
< 0 || after
)
4505 return isl_bool_false
;
4508 /* Return the filter of child "i" of "node".
4510 static __isl_give isl_union_set
*child_filter(
4511 __isl_keep isl_schedule_node
*node
, int i
)
4513 isl_schedule_node
*child
;
4514 isl_union_set
*filter
;
4516 child
= isl_schedule_node_get_child(node
, i
);
4517 filter
= isl_schedule_node_filter_get_filter(child
);
4518 isl_schedule_node_free(child
);
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"?
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.
4532 static isl_bool
after_in_sequence(__isl_keep isl_union_map
*umap
,
4533 __isl_keep isl_schedule_node
*node
)
4537 isl_union_map
*umap_i
;
4539 isl_bool after
= isl_bool_false
;
4541 n
= isl_schedule_node_n_children(node
);
4543 return isl_bool_error
;
4544 for (i
= 1; i
< n
; ++i
) {
4545 isl_union_set
*filter_i
;
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
);
4554 isl_union_map_free(umap_i
);
4558 for (j
= 0; j
< i
; ++j
) {
4559 isl_union_set
*filter_j
;
4560 isl_union_map
*umap_ij
;
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
,
4566 empty
= isl_union_map_is_empty(umap_ij
);
4567 isl_union_map_free(umap_ij
);
4572 after
= isl_bool_true
;
4577 isl_union_map_free(umap_i
);
4582 if (after
< 0 || after
)
4585 return after_in_set(umap
, node
);
4587 isl_union_map_free(umap_i
);
4588 return isl_bool_error
;
4591 /* Is any domain element of "umap" scheduled after any of
4592 * the corresponding image elements by the tree rooted at "node"?
4594 * If "umap" is empty, then clearly there is no such element.
4595 * Otherwise, consider the different types of nodes separately.
4597 static isl_bool
after_in_tree(__isl_keep isl_union_map
*umap
,
4598 __isl_keep isl_schedule_node
*node
)
4601 enum isl_schedule_node_type type
;
4603 empty
= isl_union_map_is_empty(umap
);
4605 return isl_bool_error
;
4607 return isl_bool_false
;
4609 return isl_bool_error
;
4611 type
= isl_schedule_node_get_type(node
);
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
);
4640 return isl_bool_true
;
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?
4648 * If the child of the current band node is a leaf, then
4649 * no element can be scheduled after any other element.
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.
4656 static isl_bool
after_in_subtree(__isl_keep isl_ast_build
*build
,
4657 __isl_keep isl_map
*map1
, __isl_keep isl_map
*map2
)
4659 isl_schedule_node
*node
;
4661 isl_union_map
*umap
;
4664 node
= isl_ast_build_get_schedule_node(build
);
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
;
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
);
4681 /* Internal data for any_scheduled_after.
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.
4690 struct isl_any_scheduled_after_data
{
4691 isl_ast_build
*build
;
4693 int group_coscheduled
;
4694 struct isl_set_map_pair
*domain
;
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)?
4700 * data->domain[i].set contains the domain of the inverse schedule
4701 * for domain "i", i.e., elements in the schedule domain.
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.
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.
4711 static isl_bool
any_scheduled_after(int i
, int j
, void *user
)
4713 struct isl_any_scheduled_after_data
*data
= user
;
4714 isl_size dim
= isl_set_dim(data
->domain
[i
].set
, isl_dim_set
);
4718 return isl_bool_error
;
4720 for (pos
= data
->depth
; pos
< dim
; ++pos
) {
4723 follows
= isl_set_follows_at(data
->domain
[i
].set
,
4724 data
->domain
[j
].set
, pos
);
4727 return isl_bool_error
;
4729 return isl_bool_true
;
4731 return isl_bool_false
;
4734 if (isl_ast_build_has_schedule_node(data
->build
)) {
4737 after
= after_in_subtree(data
->build
, data
->domain
[i
].map
,
4738 data
->domain
[j
].map
);
4739 if (after
< 0 || after
)
4743 return isl_bool_ok(data
->group_coscheduled
);
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.
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
4758 * Code is then generated (by generate_component)
4759 * for each of the strongly connected components in this graph
4760 * in their topological order.
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.
4766 static __isl_give isl_ast_graft_list
*generate_components(
4767 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
4770 isl_ctx
*ctx
= isl_ast_build_get_ctx(build
);
4771 isl_size n
= isl_union_map_n_map(executed
);
4772 struct isl_any_scheduled_after_data data
;
4773 struct isl_set_map_pair
*next
;
4774 struct isl_tarjan_graph
*g
= NULL
;
4775 isl_ast_graft_list
*list
= NULL
;
4781 data
.domain
= isl_calloc_array(ctx
, struct isl_set_map_pair
, n
);
4787 if (isl_union_map_foreach_map(executed
, &extract_domain
, &next
) < 0)
4793 data
.depth
= isl_ast_build_get_depth(build
);
4794 data
.group_coscheduled
= isl_options_get_ast_build_group_coscheduled(ctx
);
4795 g
= isl_tarjan_graph_init(ctx
, n
, &any_scheduled_after
, &data
);
4799 list
= isl_ast_graft_list_alloc(ctx
, 0);
4803 isl_ast_graft_list
*list_c
;
4806 if (g
->order
[i
] == -1)
4807 isl_die(ctx
, isl_error_internal
, "cannot happen",
4810 while (g
->order
[i
] != -1) {
4814 list_c
= generate_component(data
.domain
,
4815 g
->order
+ first
, i
- first
,
4816 isl_ast_build_copy(build
));
4817 list
= isl_ast_graft_list_merge(list
, list_c
, build
);
4823 error
: list
= isl_ast_graft_list_free(list
);
4824 isl_tarjan_graph_free(g
);
4825 for (i
= 0; i
< n_domain
; ++i
) {
4826 isl_map_free(data
.domain
[i
].map
);
4827 isl_set_free(data
.domain
[i
].set
);
4830 isl_union_map_free(executed
);
4831 isl_ast_build_free(build
);
4836 /* Generate code for the next level (and all inner levels).
4838 * If "executed" is empty, i.e., no code needs to be generated,
4839 * then we return an empty list.
4841 * If we have already generated code for all loop levels, then we pass
4842 * control to generate_inner_level.
4844 * If "executed" lives in a single space, i.e., if code needs to be
4845 * generated for a single domain, then there can only be a single
4846 * component and we go directly to generate_shifted_component.
4847 * Otherwise, we call generate_components to detect the components
4848 * and to call generate_component on each of them separately.
4850 static __isl_give isl_ast_graft_list
*generate_next_level(
4851 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
4857 if (!build
|| !executed
)
4860 if (isl_union_map_is_empty(executed
)) {
4861 isl_ctx
*ctx
= isl_ast_build_get_ctx(build
);
4862 isl_union_map_free(executed
);
4863 isl_ast_build_free(build
);
4864 return isl_ast_graft_list_alloc(ctx
, 0);
4867 depth
= isl_ast_build_get_depth(build
);
4868 dim
= isl_ast_build_dim(build
, isl_dim_set
);
4872 return generate_inner_level(executed
, build
);
4874 n
= isl_union_map_n_map(executed
);
4878 return generate_shifted_component(executed
, build
);
4880 return generate_components(executed
, build
);
4882 isl_union_map_free(executed
);
4883 isl_ast_build_free(build
);
4887 /* Internal data structure used by isl_ast_build_node_from_schedule_map.
4888 * internal, executed and build are the inputs to generate_code.
4889 * list collects the output.
4891 struct isl_generate_code_data
{
4893 isl_union_map
*executed
;
4894 isl_ast_build
*build
;
4896 isl_ast_graft_list
*list
;
4899 /* Given an inverse schedule in terms of the external build schedule, i.e.,
4903 * with E the external build schedule and S the additional schedule "space",
4904 * reformulate the inverse schedule in terms of the internal schedule domain,
4909 * We first obtain a mapping
4913 * take the inverse and the product with S -> S, resulting in
4915 * [I -> S] -> [E -> S]
4917 * Applying the map to the input produces the desired result.
4919 static __isl_give isl_union_map
*internal_executed(
4920 __isl_take isl_union_map
*executed
, __isl_keep isl_space
*space
,
4921 __isl_keep isl_ast_build
*build
)
4925 proj
= isl_ast_build_get_schedule_map(build
);
4926 proj
= isl_map_reverse(proj
);
4927 space
= isl_space_map_from_set(isl_space_copy(space
));
4928 id
= isl_map_identity(space
);
4929 proj
= isl_map_product(proj
, id
);
4930 executed
= isl_union_map_apply_domain(executed
,
4931 isl_union_map_from_map(proj
));
4935 /* Generate an AST that visits the elements in the range of data->executed
4936 * in the relative order specified by the corresponding domain element(s)
4937 * for those domain elements that belong to "set".
4938 * Add the result to data->list.
4940 * The caller ensures that "set" is a universe domain.
4941 * "space" is the space of the additional part of the schedule.
4942 * It is equal to the space of "set" if build->domain is parametric.
4943 * Otherwise, it is equal to the range of the wrapped space of "set".
4945 * If the build space is not parametric and
4946 * if isl_ast_build_node_from_schedule_map
4947 * was called from an outside user (data->internal not set), then
4948 * the (inverse) schedule refers to the external build domain and needs to
4949 * be transformed to refer to the internal build domain.
4951 * If the build space is parametric, then we add some of the parameter
4952 * constraints to the executed relation. Adding these constraints
4953 * allows for an earlier detection of conflicts in some cases.
4954 * However, we do not want to divide the executed relation into
4955 * more disjuncts than necessary. We therefore approximate
4956 * the constraints on the parameters by a single disjunct set.
4958 * The build is extended to include the additional part of the schedule.
4959 * If the original build space was not parametric, then the options
4960 * in data->build refer only to the additional part of the schedule
4961 * and they need to be adjusted to refer to the complete AST build
4964 * After having adjusted inverse schedule and build, we start generating
4965 * code with the outer loop of the current code generation
4966 * in generate_next_level.
4968 * If the original build space was not parametric, we undo the embedding
4969 * on the resulting isl_ast_node_list so that it can be used within
4970 * the outer AST build.
4972 static isl_stat
generate_code_in_space(struct isl_generate_code_data
*data
,
4973 __isl_take isl_set
*set
, __isl_take isl_space
*space
)
4975 isl_union_map
*executed
;
4976 isl_ast_build
*build
;
4977 isl_ast_graft_list
*list
;
4980 executed
= isl_union_map_copy(data
->executed
);
4981 executed
= isl_union_map_intersect_domain(executed
,
4982 isl_union_set_from_set(set
));
4984 embed
= !isl_set_is_params(data
->build
->domain
);
4985 if (embed
&& !data
->internal
)
4986 executed
= internal_executed(executed
, space
, data
->build
);
4989 domain
= isl_ast_build_get_domain(data
->build
);
4990 domain
= isl_set_from_basic_set(isl_set_simple_hull(domain
));
4991 executed
= isl_union_map_intersect_params(executed
, domain
);
4994 build
= isl_ast_build_copy(data
->build
);
4995 build
= isl_ast_build_product(build
, space
);
4997 list
= generate_next_level(executed
, build
);
4999 list
= isl_ast_graft_list_unembed(list
, embed
);
5001 data
->list
= isl_ast_graft_list_concat(data
->list
, list
);
5006 /* Generate an AST that visits the elements in the range of data->executed
5007 * in the relative order specified by the corresponding domain element(s)
5008 * for those domain elements that belong to "set".
5009 * Add the result to data->list.
5011 * The caller ensures that "set" is a universe domain.
5013 * If the build space S is not parametric, then the space of "set"
5014 * need to be a wrapped relation with S as domain. That is, it needs
5019 * Check this property and pass control to generate_code_in_space
5021 * If the build space is not parametric, then T is the space of "set".
5023 static isl_stat
generate_code_set(__isl_take isl_set
*set
, void *user
)
5025 struct isl_generate_code_data
*data
= user
;
5026 isl_space
*space
, *build_space
;
5029 space
= isl_set_get_space(set
);
5031 if (isl_set_is_params(data
->build
->domain
))
5032 return generate_code_in_space(data
, set
, space
);
5034 build_space
= isl_ast_build_get_space(data
->build
, data
->internal
);
5035 space
= isl_space_unwrap(space
);
5036 is_domain
= isl_space_is_domain(build_space
, space
);
5037 isl_space_free(build_space
);
5038 space
= isl_space_range(space
);
5043 isl_die(isl_set_get_ctx(set
), isl_error_invalid
,
5044 "invalid nested schedule space", goto error
);
5046 return generate_code_in_space(data
, set
, space
);
5049 isl_space_free(space
);
5050 return isl_stat_error
;
5053 /* Generate an AST that visits the elements in the range of "executed"
5054 * in the relative order specified by the corresponding domain element(s).
5056 * "build" is an isl_ast_build that has either been constructed by
5057 * isl_ast_build_from_context or passed to a callback set by
5058 * isl_ast_build_set_create_leaf.
5059 * In the first case, the space of the isl_ast_build is typically
5060 * a parametric space, although this is currently not enforced.
5061 * In the second case, the space is never a parametric space.
5062 * If the space S is not parametric, then the domain space(s) of "executed"
5063 * need to be wrapped relations with S as domain.
5065 * If the domain of "executed" consists of several spaces, then an AST
5066 * is generated for each of them (in arbitrary order) and the results
5069 * If "internal" is set, then the domain "S" above refers to the internal
5070 * schedule domain representation. Otherwise, it refers to the external
5071 * representation, as returned by isl_ast_build_get_schedule_space.
5073 * We essentially run over all the spaces in the domain of "executed"
5074 * and call generate_code_set on each of them.
5076 static __isl_give isl_ast_graft_list
*generate_code(
5077 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
,
5081 struct isl_generate_code_data data
= { 0 };
5083 isl_union_set
*schedule_domain
;
5084 isl_union_map
*universe
;
5088 space
= isl_ast_build_get_space(build
, 1);
5089 space
= isl_space_align_params(space
,
5090 isl_union_map_get_space(executed
));
5091 space
= isl_space_align_params(space
,
5092 isl_union_map_get_space(build
->options
));
5093 build
= isl_ast_build_align_params(build
, isl_space_copy(space
));
5094 executed
= isl_union_map_align_params(executed
, space
);
5095 if (!executed
|| !build
)
5098 ctx
= isl_ast_build_get_ctx(build
);
5100 data
.internal
= internal
;
5101 data
.executed
= executed
;
5103 data
.list
= isl_ast_graft_list_alloc(ctx
, 0);
5105 universe
= isl_union_map_universe(isl_union_map_copy(executed
));
5106 schedule_domain
= isl_union_map_domain(universe
);
5107 if (isl_union_set_foreach_set(schedule_domain
, &generate_code_set
,
5109 data
.list
= isl_ast_graft_list_free(data
.list
);
5111 isl_union_set_free(schedule_domain
);
5112 isl_union_map_free(executed
);
5114 isl_ast_build_free(build
);
5117 isl_union_map_free(executed
);
5118 isl_ast_build_free(build
);
5122 /* Generate an AST that visits the elements in the domain of "schedule"
5123 * in the relative order specified by the corresponding image element(s).
5125 * "build" is an isl_ast_build that has either been constructed by
5126 * isl_ast_build_from_context or passed to a callback set by
5127 * isl_ast_build_set_create_leaf.
5128 * In the first case, the space of the isl_ast_build is typically
5129 * a parametric space, although this is currently not enforced.
5130 * In the second case, the space is never a parametric space.
5131 * If the space S is not parametric, then the range space(s) of "schedule"
5132 * need to be wrapped relations with S as domain.
5134 * If the range of "schedule" consists of several spaces, then an AST
5135 * is generated for each of them (in arbitrary order) and the results
5138 * We first initialize the local copies of the relevant options.
5139 * We do this here rather than when the isl_ast_build is created
5140 * because the options may have changed between the construction
5141 * of the isl_ast_build and the call to isl_generate_code.
5143 * The main computation is performed on an inverse schedule (with
5144 * the schedule domain in the domain and the elements to be executed
5145 * in the range) called "executed".
5147 __isl_give isl_ast_node
*isl_ast_build_node_from_schedule_map(
5148 __isl_keep isl_ast_build
*build
, __isl_take isl_union_map
*schedule
)
5150 isl_ast_graft_list
*list
;
5152 isl_union_map
*executed
;
5154 build
= isl_ast_build_copy(build
);
5155 build
= isl_ast_build_set_single_valued(build
, 0);
5156 schedule
= isl_union_map_coalesce(schedule
);
5157 schedule
= isl_union_map_remove_redundancies(schedule
);
5158 executed
= isl_union_map_reverse(schedule
);
5159 list
= generate_code(executed
, isl_ast_build_copy(build
), 0);
5160 node
= isl_ast_node_from_graft_list(list
, build
);
5161 isl_ast_build_free(build
);
5166 /* The old name for isl_ast_build_node_from_schedule_map.
5167 * It is being kept for backward compatibility, but
5168 * it will be removed in the future.
5170 __isl_give isl_ast_node
*isl_ast_build_ast_from_schedule(
5171 __isl_keep isl_ast_build
*build
, __isl_take isl_union_map
*schedule
)
5173 return isl_ast_build_node_from_schedule_map(build
, schedule
);
5176 /* Generate an AST that visits the elements in the domain of "executed"
5177 * in the relative order specified by the band node "node" and its descendants.
5179 * The relation "executed" maps the outer generated loop iterators
5180 * to the domain elements executed by those iterations.
5182 * If the band is empty, we continue with its descendants.
5183 * Otherwise, we extend the build and the inverse schedule with
5184 * the additional space/partial schedule and continue generating
5185 * an AST in generate_next_level.
5186 * As soon as we have extended the inverse schedule with the additional
5187 * partial schedule, we look for equalities that may exists between
5188 * the old and the new part.
5190 static __isl_give isl_ast_graft_list
*build_ast_from_band(
5191 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5192 __isl_take isl_union_map
*executed
)
5195 isl_multi_union_pw_aff
*extra
;
5196 isl_union_map
*extra_umap
;
5197 isl_ast_graft_list
*list
;
5201 n
= isl_schedule_node_band_n_member(node
);
5202 if (!build
|| n
< 0 || !executed
)
5206 return build_ast_from_child(build
, node
, executed
);
5208 extra
= isl_schedule_node_band_get_partial_schedule(node
);
5209 extra
= isl_multi_union_pw_aff_align_params(extra
,
5210 isl_ast_build_get_space(build
, 1));
5211 space
= isl_multi_union_pw_aff_get_space(extra
);
5213 extra_umap
= isl_union_map_from_multi_union_pw_aff(extra
);
5214 extra_umap
= isl_union_map_reverse(extra_umap
);
5216 executed
= isl_union_map_domain_product(executed
, extra_umap
);
5217 executed
= isl_union_map_detect_equalities(executed
);
5219 n1
= isl_ast_build_dim(build
, isl_dim_param
);
5220 build
= isl_ast_build_product(build
, space
);
5221 n2
= isl_ast_build_dim(build
, isl_dim_param
);
5222 if (n1
< 0 || n2
< 0)
5223 build
= isl_ast_build_free(build
);
5225 isl_die(isl_ast_build_get_ctx(build
), isl_error_invalid
,
5226 "band node is not allowed to introduce new parameters",
5227 build
= isl_ast_build_free(build
));
5228 build
= isl_ast_build_set_schedule_node(build
, node
);
5230 list
= generate_next_level(executed
, build
);
5232 list
= isl_ast_graft_list_unembed(list
, 1);
5236 isl_schedule_node_free(node
);
5237 isl_union_map_free(executed
);
5238 isl_ast_build_free(build
);
5242 /* Hoist a list of grafts (in practice containing a single graft)
5243 * from "sub_build" (which includes extra context information)
5246 * In particular, project out all additional parameters introduced
5247 * by the context node from the enforced constraints and the guard
5248 * of the single graft.
5250 static __isl_give isl_ast_graft_list
*hoist_out_of_context(
5251 __isl_take isl_ast_graft_list
*list
, __isl_keep isl_ast_build
*build
,
5252 __isl_keep isl_ast_build
*sub_build
)
5254 isl_ast_graft
*graft
;
5255 isl_basic_set
*enforced
;
5257 isl_size n_param
, extra_param
;
5259 n_param
= isl_ast_build_dim(build
, isl_dim_param
);
5260 extra_param
= isl_ast_build_dim(sub_build
, isl_dim_param
);
5261 if (n_param
< 0 || extra_param
< 0)
5262 return isl_ast_graft_list_free(list
);
5264 if (extra_param
== n_param
)
5267 extra_param
-= n_param
;
5268 enforced
= isl_ast_graft_list_extract_shared_enforced(list
, sub_build
);
5269 enforced
= isl_basic_set_project_out(enforced
, isl_dim_param
,
5270 n_param
, extra_param
);
5271 enforced
= isl_basic_set_remove_unknown_divs(enforced
);
5272 guard
= isl_ast_graft_list_extract_hoistable_guard(list
, sub_build
);
5273 guard
= isl_set_remove_divs_involving_dims(guard
, isl_dim_param
,
5274 n_param
, extra_param
);
5275 guard
= isl_set_project_out(guard
, isl_dim_param
, n_param
, extra_param
);
5276 guard
= isl_set_compute_divs(guard
);
5277 graft
= isl_ast_graft_alloc_from_children(list
, guard
, enforced
,
5279 list
= isl_ast_graft_list_from_ast_graft(graft
);
5284 /* Generate an AST that visits the elements in the domain of "executed"
5285 * in the relative order specified by the context node "node"
5286 * and its descendants.
5288 * The relation "executed" maps the outer generated loop iterators
5289 * to the domain elements executed by those iterations.
5291 * The context node may introduce additional parameters as well as
5292 * constraints on the outer schedule dimensions or original parameters.
5294 * We add the extra parameters to a new build and the context
5295 * constraints to both the build and (as a single disjunct)
5296 * to the domain of "executed". Since the context constraints
5297 * are specified in terms of the input schedule, we first need
5298 * to map them to the internal schedule domain.
5300 * After constructing the AST from the descendants of "node",
5301 * we combine the list of grafts into a single graft within
5302 * the new build, in order to be able to exploit the additional
5303 * context constraints during this combination.
5305 * Additionally, if the current node is the outermost node in
5306 * the schedule tree (apart from the root domain node), we generate
5307 * all pending guards, again to be able to exploit the additional
5308 * context constraints. We currently do not do this for internal
5309 * context nodes since we may still want to hoist conditions
5310 * to outer AST nodes.
5312 * If the context node introduced any new parameters, then they
5313 * are removed from the set of enforced constraints and guard
5314 * in hoist_out_of_context.
5316 static __isl_give isl_ast_graft_list
*build_ast_from_context(
5317 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5318 __isl_take isl_union_map
*executed
)
5322 isl_multi_aff
*internal2input
;
5323 isl_ast_build
*sub_build
;
5324 isl_ast_graft_list
*list
;
5328 depth
= isl_schedule_node_get_tree_depth(node
);
5330 build
= isl_ast_build_free(build
);
5331 space
= isl_ast_build_get_space(build
, 1);
5332 context
= isl_schedule_node_context_get_context(node
);
5333 context
= isl_set_align_params(context
, space
);
5334 sub_build
= isl_ast_build_copy(build
);
5335 space
= isl_set_get_space(context
);
5336 sub_build
= isl_ast_build_align_params(sub_build
, space
);
5337 internal2input
= isl_ast_build_get_internal2input(sub_build
);
5338 context
= isl_set_preimage_multi_aff(context
, internal2input
);
5339 sub_build
= isl_ast_build_restrict_generated(sub_build
,
5340 isl_set_copy(context
));
5341 context
= isl_set_from_basic_set(isl_set_simple_hull(context
));
5342 executed
= isl_union_map_intersect_domain(executed
,
5343 isl_union_set_from_set(context
));
5345 list
= build_ast_from_child(isl_ast_build_copy(sub_build
),
5347 n
= isl_ast_graft_list_n_ast_graft(list
);
5349 list
= isl_ast_graft_list_free(list
);
5351 list
= isl_ast_graft_list_fuse(list
, sub_build
);
5353 list
= isl_ast_graft_list_insert_pending_guard_nodes(list
,
5356 list
= hoist_out_of_context(list
, build
, sub_build
);
5358 isl_ast_build_free(build
);
5359 isl_ast_build_free(sub_build
);
5364 /* Generate an AST that visits the elements in the domain of "executed"
5365 * in the relative order specified by the expansion node "node" and
5368 * The relation "executed" maps the outer generated loop iterators
5369 * to the domain elements executed by those iterations.
5371 * We expand the domain elements by the expansion and
5372 * continue with the descendants of the node.
5374 static __isl_give isl_ast_graft_list
*build_ast_from_expansion(
5375 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5376 __isl_take isl_union_map
*executed
)
5378 isl_union_map
*expansion
;
5381 expansion
= isl_schedule_node_expansion_get_expansion(node
);
5382 expansion
= isl_union_map_align_params(expansion
,
5383 isl_union_map_get_space(executed
));
5385 n1
= isl_union_map_dim(executed
, isl_dim_param
);
5386 executed
= isl_union_map_apply_range(executed
, expansion
);
5387 n2
= isl_union_map_dim(executed
, isl_dim_param
);
5388 if (n1
< 0 || n2
< 0)
5391 isl_die(isl_ast_build_get_ctx(build
), isl_error_invalid
,
5392 "expansion node is not allowed to introduce "
5393 "new parameters", goto error
);
5395 return build_ast_from_child(build
, node
, executed
);
5397 isl_ast_build_free(build
);
5398 isl_schedule_node_free(node
);
5399 isl_union_map_free(executed
);
5403 /* Generate an AST that visits the elements in the domain of "executed"
5404 * in the relative order specified by the extension node "node" and
5407 * The relation "executed" maps the outer generated loop iterators
5408 * to the domain elements executed by those iterations.
5410 * Extend the inverse schedule with the extension applied to current
5411 * set of generated constraints. Since the extension if formulated
5412 * in terms of the input schedule, it first needs to be transformed
5413 * to refer to the internal schedule.
5415 static __isl_give isl_ast_graft_list
*build_ast_from_extension(
5416 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5417 __isl_take isl_union_map
*executed
)
5419 isl_union_set
*schedule_domain
;
5420 isl_union_map
*extension
;
5423 set
= isl_ast_build_get_generated(build
);
5424 set
= isl_set_from_basic_set(isl_set_simple_hull(set
));
5425 schedule_domain
= isl_union_set_from_set(set
);
5427 extension
= isl_schedule_node_extension_get_extension(node
);
5429 extension
= isl_union_map_preimage_domain_multi_aff(extension
,
5430 isl_multi_aff_copy(build
->internal2input
));
5431 extension
= isl_union_map_intersect_domain(extension
, schedule_domain
);
5432 extension
= isl_ast_build_substitute_values_union_map_domain(build
,
5434 executed
= isl_union_map_union(executed
, extension
);
5436 return build_ast_from_child(build
, node
, executed
);
5439 /* Generate an AST that visits the elements in the domain of "executed"
5440 * in the relative order specified by the filter node "node" and
5443 * The relation "executed" maps the outer generated loop iterators
5444 * to the domain elements executed by those iterations.
5446 * We simply intersect the iteration domain (i.e., the range of "executed")
5447 * with the filter and continue with the descendants of the node,
5448 * unless the resulting inverse schedule is empty, in which
5449 * case we return an empty list.
5451 * If the result of the intersection is equal to the original "executed"
5452 * relation, then keep the original representation since the intersection
5453 * may have unnecessarily broken up the relation into a greater number
5456 static __isl_give isl_ast_graft_list
*build_ast_from_filter(
5457 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5458 __isl_take isl_union_map
*executed
)
5461 isl_union_set
*filter
;
5462 isl_union_map
*orig
;
5463 isl_ast_graft_list
*list
;
5468 orig
= isl_union_map_copy(executed
);
5469 if (!build
|| !node
|| !executed
)
5472 filter
= isl_schedule_node_filter_get_filter(node
);
5473 filter
= isl_union_set_align_params(filter
,
5474 isl_union_map_get_space(executed
));
5475 n1
= isl_union_map_dim(executed
, isl_dim_param
);
5476 executed
= isl_union_map_intersect_range(executed
, filter
);
5477 n2
= isl_union_map_dim(executed
, isl_dim_param
);
5478 if (n1
< 0 || n2
< 0)
5481 isl_die(isl_ast_build_get_ctx(build
), isl_error_invalid
,
5482 "filter node is not allowed to introduce "
5483 "new parameters", goto error
);
5485 unchanged
= isl_union_map_is_subset(orig
, executed
);
5486 empty
= isl_union_map_is_empty(executed
);
5487 if (unchanged
< 0 || empty
< 0)
5490 isl_union_map_free(executed
);
5491 return build_ast_from_child(build
, node
, orig
);
5493 isl_union_map_free(orig
);
5495 return build_ast_from_child(build
, node
, executed
);
5497 ctx
= isl_ast_build_get_ctx(build
);
5498 list
= isl_ast_graft_list_alloc(ctx
, 0);
5499 isl_ast_build_free(build
);
5500 isl_schedule_node_free(node
);
5501 isl_union_map_free(executed
);
5504 isl_ast_build_free(build
);
5505 isl_schedule_node_free(node
);
5506 isl_union_map_free(executed
);
5507 isl_union_map_free(orig
);
5511 /* Generate an AST that visits the elements in the domain of "executed"
5512 * in the relative order specified by the guard node "node" and
5515 * The relation "executed" maps the outer generated loop iterators
5516 * to the domain elements executed by those iterations.
5518 * Ensure that the associated guard is enforced by the outer AST
5519 * constructs by adding it to the guard of the graft.
5520 * Since we know that we will enforce the guard, we can also include it
5521 * in the generated constraints used to construct an AST for
5522 * the descendant nodes.
5524 static __isl_give isl_ast_graft_list
*build_ast_from_guard(
5525 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5526 __isl_take isl_union_map
*executed
)
5529 isl_set
*guard
, *hoisted
;
5530 isl_basic_set
*enforced
;
5531 isl_ast_build
*sub_build
;
5532 isl_ast_graft
*graft
;
5533 isl_ast_graft_list
*list
;
5536 space
= isl_ast_build_get_space(build
, 1);
5537 guard
= isl_schedule_node_guard_get_guard(node
);
5538 n1
= isl_space_dim(space
, isl_dim_param
);
5539 guard
= isl_set_align_params(guard
, space
);
5540 n2
= isl_set_dim(guard
, isl_dim_param
);
5541 if (n1
< 0 || n2
< 0)
5542 guard
= isl_set_free(guard
);
5544 isl_die(isl_ast_build_get_ctx(build
), isl_error_invalid
,
5545 "guard node is not allowed to introduce "
5546 "new parameters", guard
= isl_set_free(guard
));
5547 guard
= isl_set_preimage_multi_aff(guard
,
5548 isl_multi_aff_copy(build
->internal2input
));
5549 guard
= isl_ast_build_specialize(build
, guard
);
5550 guard
= isl_set_gist(guard
, isl_set_copy(build
->generated
));
5552 sub_build
= isl_ast_build_copy(build
);
5553 sub_build
= isl_ast_build_restrict_generated(sub_build
,
5554 isl_set_copy(guard
));
5556 list
= build_ast_from_child(isl_ast_build_copy(sub_build
),
5559 hoisted
= isl_ast_graft_list_extract_hoistable_guard(list
, sub_build
);
5560 n
= isl_set_n_basic_set(hoisted
);
5562 list
= isl_ast_graft_list_free(list
);
5564 list
= isl_ast_graft_list_gist_guards(list
,
5565 isl_set_copy(hoisted
));
5566 guard
= isl_set_intersect(guard
, hoisted
);
5567 enforced
= extract_shared_enforced(list
, build
);
5568 graft
= isl_ast_graft_alloc_from_children(list
, guard
, enforced
,
5571 isl_ast_build_free(sub_build
);
5572 isl_ast_build_free(build
);
5573 return isl_ast_graft_list_from_ast_graft(graft
);
5576 /* Call the before_each_mark callback, if requested by the user.
5578 * Return 0 on success and -1 on error.
5580 * The caller is responsible for recording the current inverse schedule
5583 static isl_stat
before_each_mark(__isl_keep isl_id
*mark
,
5584 __isl_keep isl_ast_build
*build
)
5587 return isl_stat_error
;
5588 if (!build
->before_each_mark
)
5590 return build
->before_each_mark(mark
, build
,
5591 build
->before_each_mark_user
);
5594 /* Call the after_each_mark callback, if requested by the user.
5596 * The caller is responsible for recording the current inverse schedule
5599 static __isl_give isl_ast_graft
*after_each_mark(
5600 __isl_take isl_ast_graft
*graft
, __isl_keep isl_ast_build
*build
)
5602 if (!graft
|| !build
)
5603 return isl_ast_graft_free(graft
);
5604 if (!build
->after_each_mark
)
5606 graft
->node
= build
->after_each_mark(graft
->node
, build
,
5607 build
->after_each_mark_user
);
5609 return isl_ast_graft_free(graft
);
5614 /* Generate an AST that visits the elements in the domain of "executed"
5615 * in the relative order specified by the mark node "node" and
5618 * The relation "executed" maps the outer generated loop iterators
5619 * to the domain elements executed by those iterations.
5621 * Since we may be calling before_each_mark and after_each_mark
5622 * callbacks, we record the current inverse schedule in the build.
5624 * We generate an AST for the child of the mark node, combine
5625 * the graft list into a single graft and then insert the mark
5626 * in the AST of that single graft.
5628 static __isl_give isl_ast_graft_list
*build_ast_from_mark(
5629 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5630 __isl_take isl_union_map
*executed
)
5633 isl_ast_graft
*graft
;
5634 isl_ast_graft_list
*list
;
5637 build
= isl_ast_build_set_executed(build
, isl_union_map_copy(executed
));
5639 mark
= isl_schedule_node_mark_get_id(node
);
5640 if (before_each_mark(mark
, build
) < 0)
5641 node
= isl_schedule_node_free(node
);
5643 list
= build_ast_from_child(isl_ast_build_copy(build
), node
, executed
);
5644 list
= isl_ast_graft_list_fuse(list
, build
);
5645 n
= isl_ast_graft_list_n_ast_graft(list
);
5647 list
= isl_ast_graft_list_free(list
);
5651 graft
= isl_ast_graft_list_get_ast_graft(list
, 0);
5652 graft
= isl_ast_graft_insert_mark(graft
, mark
);
5653 graft
= after_each_mark(graft
, build
);
5654 list
= isl_ast_graft_list_set_ast_graft(list
, 0, graft
);
5656 isl_ast_build_free(build
);
5661 static __isl_give isl_ast_graft_list
*build_ast_from_schedule_node(
5662 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5663 __isl_take isl_union_map
*executed
);
5665 /* Generate an AST that visits the elements in the domain of "executed"
5666 * in the relative order specified by the sequence (or set) node "node" and
5669 * The relation "executed" maps the outer generated loop iterators
5670 * to the domain elements executed by those iterations.
5672 * We simply generate an AST for each of the children and concatenate
5675 static __isl_give isl_ast_graft_list
*build_ast_from_sequence(
5676 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5677 __isl_take isl_union_map
*executed
)
5682 isl_ast_graft_list
*list
;
5684 ctx
= isl_ast_build_get_ctx(build
);
5685 list
= isl_ast_graft_list_alloc(ctx
, 0);
5687 n
= isl_schedule_node_n_children(node
);
5689 list
= isl_ast_graft_list_free(list
);
5690 for (i
= 0; i
< n
; ++i
) {
5691 isl_schedule_node
*child
;
5692 isl_ast_graft_list
*list_i
;
5694 child
= isl_schedule_node_get_child(node
, i
);
5695 list_i
= build_ast_from_schedule_node(isl_ast_build_copy(build
),
5696 child
, isl_union_map_copy(executed
));
5697 list
= isl_ast_graft_list_concat(list
, list_i
);
5699 isl_ast_build_free(build
);
5700 isl_schedule_node_free(node
);
5701 isl_union_map_free(executed
);
5706 /* Generate an AST that visits the elements in the domain of "executed"
5707 * in the relative order specified by the node "node" and its descendants.
5709 * The relation "executed" maps the outer generated loop iterators
5710 * to the domain elements executed by those iterations.
5712 * If the node is a leaf, then we pass control to generate_inner_level.
5713 * Note that the current build does not refer to any band node, so
5714 * that generate_inner_level will not try to visit the child of
5717 * The other node types are handled in separate functions.
5718 * Set nodes are currently treated in the same way as sequence nodes.
5719 * The children of a set node may be executed in any order,
5720 * including the order of the children.
5722 static __isl_give isl_ast_graft_list
*build_ast_from_schedule_node(
5723 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5724 __isl_take isl_union_map
*executed
)
5726 enum isl_schedule_node_type type
;
5728 type
= isl_schedule_node_get_type(node
);
5731 case isl_schedule_node_error
:
5733 case isl_schedule_node_leaf
:
5734 isl_schedule_node_free(node
);
5735 return generate_inner_level(executed
, build
);
5736 case isl_schedule_node_band
:
5737 return build_ast_from_band(build
, node
, executed
);
5738 case isl_schedule_node_context
:
5739 return build_ast_from_context(build
, node
, executed
);
5740 case isl_schedule_node_domain
:
5741 isl_die(isl_schedule_node_get_ctx(node
), isl_error_unsupported
,
5742 "unexpected internal domain node", goto error
);
5743 case isl_schedule_node_expansion
:
5744 return build_ast_from_expansion(build
, node
, executed
);
5745 case isl_schedule_node_extension
:
5746 return build_ast_from_extension(build
, node
, executed
);
5747 case isl_schedule_node_filter
:
5748 return build_ast_from_filter(build
, node
, executed
);
5749 case isl_schedule_node_guard
:
5750 return build_ast_from_guard(build
, node
, executed
);
5751 case isl_schedule_node_mark
:
5752 return build_ast_from_mark(build
, node
, executed
);
5753 case isl_schedule_node_sequence
:
5754 case isl_schedule_node_set
:
5755 return build_ast_from_sequence(build
, node
, executed
);
5758 isl_die(isl_ast_build_get_ctx(build
), isl_error_internal
,
5759 "unhandled type", goto error
);
5761 isl_union_map_free(executed
);
5762 isl_schedule_node_free(node
);
5763 isl_ast_build_free(build
);
5768 /* Generate an AST that visits the elements in the domain of "executed"
5769 * in the relative order specified by the (single) child of "node" and
5772 * The relation "executed" maps the outer generated loop iterators
5773 * to the domain elements executed by those iterations.
5775 * This function is never called on a leaf, set or sequence node,
5776 * so the node always has exactly one child.
5778 static __isl_give isl_ast_graft_list
*build_ast_from_child(
5779 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5780 __isl_take isl_union_map
*executed
)
5782 node
= isl_schedule_node_child(node
, 0);
5783 return build_ast_from_schedule_node(build
, node
, executed
);
5786 /* Generate an AST that visits the elements in the domain of the domain
5787 * node "node" in the relative order specified by its descendants.
5789 * An initial inverse schedule is created that maps a zero-dimensional
5790 * schedule space to the node domain.
5791 * The input "build" is assumed to have a parametric domain and
5792 * is replaced by the same zero-dimensional schedule space.
5794 * We also add some of the parameter constraints in the build domain
5795 * to the executed relation. Adding these constraints
5796 * allows for an earlier detection of conflicts in some cases.
5797 * However, we do not want to divide the executed relation into
5798 * more disjuncts than necessary. We therefore approximate
5799 * the constraints on the parameters by a single disjunct set.
5801 static __isl_give isl_ast_node
*build_ast_from_domain(
5802 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
)
5805 isl_union_set
*domain
, *schedule_domain
;
5806 isl_union_map
*executed
;
5809 isl_ast_graft_list
*list
;
5816 ctx
= isl_ast_build_get_ctx(build
);
5817 space
= isl_ast_build_get_space(build
, 1);
5818 is_params
= isl_space_is_params(space
);
5819 isl_space_free(space
);
5823 isl_die(ctx
, isl_error_unsupported
,
5824 "expecting parametric initial context", goto error
);
5826 domain
= isl_schedule_node_domain_get_domain(node
);
5827 domain
= isl_union_set_coalesce(domain
);
5829 space
= isl_union_set_get_space(domain
);
5830 space
= isl_space_set_from_params(space
);
5831 build
= isl_ast_build_product(build
, space
);
5833 set
= isl_ast_build_get_domain(build
);
5834 set
= isl_set_from_basic_set(isl_set_simple_hull(set
));
5835 schedule_domain
= isl_union_set_from_set(set
);
5837 executed
= isl_union_map_from_domain_and_range(schedule_domain
, domain
);
5838 list
= build_ast_from_child(isl_ast_build_copy(build
), node
, executed
);
5839 ast
= isl_ast_node_from_graft_list(list
, build
);
5840 isl_ast_build_free(build
);
5844 isl_schedule_node_free(node
);
5845 isl_ast_build_free(build
);
5849 /* Generate an AST that visits the elements in the domain of "schedule"
5850 * in the relative order specified by the schedule tree.
5852 * "build" is an isl_ast_build that has been created using
5853 * isl_ast_build_alloc or isl_ast_build_from_context based
5854 * on a parametric set.
5856 * The construction starts at the root node of the schedule,
5857 * which is assumed to be a domain node.
5859 __isl_give isl_ast_node
*isl_ast_build_node_from_schedule(
5860 __isl_keep isl_ast_build
*build
, __isl_take isl_schedule
*schedule
)
5863 isl_schedule_node
*node
;
5865 if (!build
|| !schedule
)
5868 ctx
= isl_ast_build_get_ctx(build
);
5870 node
= isl_schedule_get_root(schedule
);
5873 isl_schedule_free(schedule
);
5875 build
= isl_ast_build_copy(build
);
5876 build
= isl_ast_build_set_single_valued(build
, 0);
5877 if (isl_schedule_node_get_type(node
) != isl_schedule_node_domain
)
5878 isl_die(ctx
, isl_error_unsupported
,
5879 "expecting root domain node",
5880 build
= isl_ast_build_free(build
));
5881 return build_ast_from_domain(build
, node
);
5883 isl_schedule_free(schedule
);