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_expr_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_bool
aff_constant_is_negative(__isl_keep isl_set
*set
,
935 __isl_keep isl_aff
*aff
, void *user
)
940 v
= isl_aff_get_constant_val(aff
);
941 is_neg
= isl_val_is_neg(v
);
947 /* Does "pa" have a negative constant term over its entire domain?
949 static isl_bool
pw_aff_constant_is_negative(__isl_keep isl_pw_aff
*pa
,
952 return isl_pw_aff_every_piece(pa
, &aff_constant_is_negative
, NULL
);
955 /* Does each element in "list" have a negative constant term?
957 static int list_constant_is_negative(__isl_keep isl_pw_aff_list
*list
)
959 return isl_pw_aff_list_every(list
, &pw_aff_constant_is_negative
, NULL
);
962 /* Add 1 to each of the elements in "list", where each of these elements
963 * is defined over the internal schedule space of "build".
965 static __isl_give isl_pw_aff_list
*list_add_one(
966 __isl_take isl_pw_aff_list
*list
, __isl_keep isl_ast_build
*build
)
974 n
= isl_pw_aff_list_n_pw_aff(list
);
976 return isl_pw_aff_list_free(list
);
978 space
= isl_ast_build_get_space(build
, 1);
979 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
980 aff
= isl_aff_add_constant_si(aff
, 1);
981 one
= isl_pw_aff_from_aff(aff
);
983 for (i
= 0; i
< n
; ++i
) {
985 pa
= isl_pw_aff_list_get_pw_aff(list
, i
);
986 pa
= isl_pw_aff_add(pa
, isl_pw_aff_copy(one
));
987 list
= isl_pw_aff_list_set_pw_aff(list
, i
, pa
);
990 isl_pw_aff_free(one
);
995 /* Set the condition part of the for node graft->node in case
996 * the upper bound is represented as a list of piecewise affine expressions.
998 * In particular, set the condition to
1000 * iterator <= min(list of upper bounds)
1002 * If each of the upper bounds has a negative constant term, then
1003 * set the condition to
1005 * iterator < min(list of (upper bound + 1)s)
1008 static __isl_give isl_ast_graft
*set_for_cond_from_list(
1009 __isl_take isl_ast_graft
*graft
, __isl_keep isl_pw_aff_list
*list
,
1010 __isl_keep isl_ast_build
*build
)
1013 isl_ast_expr
*bound
, *iterator
, *cond
;
1014 enum isl_ast_expr_op_type type
= isl_ast_expr_op_le
;
1016 if (!graft
|| !list
)
1017 return isl_ast_graft_free(graft
);
1019 neg
= list_constant_is_negative(list
);
1021 return isl_ast_graft_free(graft
);
1022 list
= isl_pw_aff_list_copy(list
);
1024 list
= list_add_one(list
, build
);
1025 type
= isl_ast_expr_op_lt
;
1028 bound
= reduce_list(isl_ast_expr_op_min
, list
, build
);
1029 iterator
= isl_ast_expr_copy(graft
->node
->u
.f
.iterator
);
1030 cond
= isl_ast_expr_alloc_binary(type
, iterator
, bound
);
1031 graft
->node
->u
.f
.cond
= cond
;
1033 isl_pw_aff_list_free(list
);
1034 if (!graft
->node
->u
.f
.cond
)
1035 return isl_ast_graft_free(graft
);
1039 /* Set the condition part of the for node graft->node in case
1040 * the upper bound is represented as a set.
1042 static __isl_give isl_ast_graft
*set_for_cond_from_set(
1043 __isl_take isl_ast_graft
*graft
, __isl_keep isl_set
*set
,
1044 __isl_keep isl_ast_build
*build
)
1051 cond
= isl_ast_build_expr_from_set_internal(build
, isl_set_copy(set
));
1052 graft
->node
->u
.f
.cond
= cond
;
1053 if (!graft
->node
->u
.f
.cond
)
1054 return isl_ast_graft_free(graft
);
1058 /* Construct an isl_ast_expr for the increment (i.e., stride) of
1059 * the current dimension.
1061 static __isl_give isl_ast_expr
*for_inc(__isl_keep isl_ast_build
*build
)
1069 ctx
= isl_ast_build_get_ctx(build
);
1070 depth
= isl_ast_build_get_depth(build
);
1072 if (!isl_ast_build_has_stride(build
, depth
))
1073 return isl_ast_expr_alloc_int_si(ctx
, 1);
1075 v
= isl_ast_build_get_stride(build
, depth
);
1076 return isl_ast_expr_from_val(v
);
1079 /* Should we express the loop condition as
1081 * iterator <= min(list of upper bounds)
1083 * or as a conjunction of constraints?
1085 * The first is constructed from a list of upper bounds.
1086 * The second is constructed from a set.
1088 * If there are no upper bounds in "constraints", then this could mean
1089 * that "domain" simply doesn't have an upper bound or that we didn't
1090 * pick any upper bound. In the first case, we want to generate the
1091 * loop condition as a(n empty) conjunction of constraints
1092 * In the second case, we will compute
1093 * a single upper bound from "domain" and so we use the list form.
1095 * If there are upper bounds in "constraints",
1096 * then we use the list form iff the atomic_upper_bound option is set.
1098 static int use_upper_bound_list(isl_ctx
*ctx
, int n_upper
,
1099 __isl_keep isl_set
*domain
, int depth
)
1102 return isl_options_get_ast_build_atomic_upper_bound(ctx
);
1104 return isl_set_dim_has_upper_bound(domain
, isl_dim_set
, depth
);
1107 /* Fill in the expressions of the for node in graft->node.
1110 * - set the initialization part of the loop to the maximum of the lower bounds
1111 * - extract the increment from the stride of the current dimension
1112 * - construct the for condition either based on a list of upper bounds
1113 * or on a set of upper bound constraints.
1115 static __isl_give isl_ast_graft
*set_for_node_expressions(
1116 __isl_take isl_ast_graft
*graft
, __isl_keep isl_pw_aff_list
*lower
,
1117 int use_list
, __isl_keep isl_pw_aff_list
*upper_list
,
1118 __isl_keep isl_set
*upper_set
, __isl_keep isl_ast_build
*build
)
1125 build
= isl_ast_build_copy(build
);
1128 node
->u
.f
.init
= reduce_list(isl_ast_expr_op_max
, lower
, build
);
1129 node
->u
.f
.inc
= for_inc(build
);
1131 if (!node
->u
.f
.init
|| !node
->u
.f
.inc
)
1132 graft
= isl_ast_graft_free(graft
);
1135 graft
= set_for_cond_from_list(graft
, upper_list
, build
);
1137 graft
= set_for_cond_from_set(graft
, upper_set
, build
);
1139 isl_ast_build_free(build
);
1144 /* Update "graft" based on "bounds" and "domain" for the generic,
1145 * non-degenerate, case.
1147 * "c_lower" and "c_upper" contain the lower and upper bounds
1148 * that the loop node should express.
1149 * "domain" is the subset of the intersection of the constraints
1150 * for which some code is executed.
1152 * There may be zero lower bounds or zero upper bounds in "constraints"
1153 * in case the list of constraints was created
1154 * based on the atomic option or based on separation with explicit bounds.
1155 * In that case, we use "domain" to derive lower and/or upper bounds.
1157 * We first compute a list of one or more lower bounds.
1159 * Then we decide if we want to express the condition as
1161 * iterator <= min(list of upper bounds)
1163 * or as a conjunction of constraints.
1165 * The set of enforced constraints is then computed either based on
1166 * a list of upper bounds or on a set of upper bound constraints.
1167 * We do not compute any enforced constraints if we were forced
1168 * to compute a lower or upper bound using exact_bound. The domains
1169 * of the resulting expressions may imply some bounds on outer dimensions
1170 * that we do not want to appear in the enforced constraints since
1171 * they are not actually enforced by the corresponding code.
1173 * Finally, we fill in the expressions of the for node.
1175 static __isl_give isl_ast_graft
*refine_generic_bounds(
1176 __isl_take isl_ast_graft
*graft
,
1177 __isl_take isl_constraint_list
*c_lower
,
1178 __isl_take isl_constraint_list
*c_upper
,
1179 __isl_keep isl_set
*domain
, __isl_keep isl_ast_build
*build
)
1183 isl_pw_aff_list
*lower
;
1185 isl_set
*upper_set
= NULL
;
1186 isl_pw_aff_list
*upper_list
= NULL
;
1187 isl_size n_lower
, n_upper
;
1189 if (!graft
|| !c_lower
|| !c_upper
|| !build
)
1192 depth
= isl_ast_build_get_depth(build
);
1193 ctx
= isl_ast_graft_get_ctx(graft
);
1195 n_lower
= isl_constraint_list_n_constraint(c_lower
);
1196 n_upper
= isl_constraint_list_n_constraint(c_upper
);
1197 if (n_lower
< 0 || n_upper
< 0)
1200 use_list
= use_upper_bound_list(ctx
, n_upper
, domain
, depth
);
1202 lower
= lower_bounds(c_lower
, depth
, domain
, build
);
1205 upper_list
= upper_bounds(c_upper
, depth
, domain
, build
);
1206 else if (n_upper
> 0)
1207 upper_set
= intersect_constraints(c_upper
);
1209 upper_set
= isl_set_universe(isl_set_get_space(domain
));
1211 if (n_lower
== 0 || n_upper
== 0)
1214 graft
= set_enforced_from_list(graft
, lower
, upper_list
);
1216 graft
= set_enforced_from_set(graft
, lower
, depth
, upper_set
);
1218 graft
= set_for_node_expressions(graft
, lower
, use_list
, upper_list
,
1221 isl_pw_aff_list_free(lower
);
1222 isl_pw_aff_list_free(upper_list
);
1223 isl_set_free(upper_set
);
1224 isl_constraint_list_free(c_lower
);
1225 isl_constraint_list_free(c_upper
);
1229 isl_constraint_list_free(c_lower
);
1230 isl_constraint_list_free(c_upper
);
1231 return isl_ast_graft_free(graft
);
1234 /* Internal data structure used inside count_constraints to keep
1235 * track of the number of constraints that are independent of dimension "pos",
1236 * the lower bounds in "pos" and the upper bounds in "pos".
1238 struct isl_ast_count_constraints_data
{
1246 /* Increment data->n_indep, data->lower or data->upper depending
1247 * on whether "c" is independenct of dimensions data->pos,
1248 * a lower bound or an upper bound.
1250 static isl_stat
count_constraints(__isl_take isl_constraint
*c
, void *user
)
1252 struct isl_ast_count_constraints_data
*data
= user
;
1254 if (isl_constraint_is_lower_bound(c
, isl_dim_set
, data
->pos
))
1256 else if (isl_constraint_is_upper_bound(c
, isl_dim_set
, data
->pos
))
1261 isl_constraint_free(c
);
1266 /* Update "graft" based on "bounds" and "domain" for the generic,
1267 * non-degenerate, case.
1269 * "list" respresent the list of bounds that need to be encoded by
1270 * the for loop. Only the constraints that involve the iterator
1271 * are relevant here. The other constraints are taken care of by
1272 * the caller and are included in the generated constraints of "build".
1273 * "domain" is the subset of the intersection of the constraints
1274 * for which some code is executed.
1275 * "build" is the build in which graft->node was created.
1277 * We separate lower bounds, upper bounds and constraints that
1278 * are independent of the loop iterator.
1280 * The actual for loop bounds are generated in refine_generic_bounds.
1282 static __isl_give isl_ast_graft
*refine_generic_split(
1283 __isl_take isl_ast_graft
*graft
, __isl_take isl_constraint_list
*list
,
1284 __isl_keep isl_set
*domain
, __isl_keep isl_ast_build
*build
)
1286 struct isl_ast_count_constraints_data data
;
1287 isl_constraint_list
*lower
;
1288 isl_constraint_list
*upper
;
1291 return isl_ast_graft_free(graft
);
1293 data
.pos
= isl_ast_build_get_depth(build
);
1295 list
= isl_constraint_list_sort(list
, &cmp_constraint
, &data
.pos
);
1297 return isl_ast_graft_free(graft
);
1299 data
.n_indep
= data
.n_lower
= data
.n_upper
= 0;
1300 if (isl_constraint_list_foreach(list
, &count_constraints
, &data
) < 0) {
1301 isl_constraint_list_free(list
);
1302 return isl_ast_graft_free(graft
);
1305 lower
= isl_constraint_list_drop(list
, 0, data
.n_indep
);
1306 upper
= isl_constraint_list_copy(lower
);
1307 lower
= isl_constraint_list_drop(lower
, data
.n_lower
, data
.n_upper
);
1308 upper
= isl_constraint_list_drop(upper
, 0, data
.n_lower
);
1310 return refine_generic_bounds(graft
, lower
, upper
, domain
, build
);
1313 /* Update "graft" based on "bounds" and "domain" for the generic,
1314 * non-degenerate, case.
1316 * "bounds" respresent the bounds that need to be encoded by
1317 * the for loop (or a guard around the for loop).
1318 * "domain" is the subset of "bounds" for which some code is executed.
1319 * "build" is the build in which graft->node was created.
1321 * We break up "bounds" into a list of constraints and continue with
1322 * refine_generic_split.
1324 static __isl_give isl_ast_graft
*refine_generic(
1325 __isl_take isl_ast_graft
*graft
,
1326 __isl_keep isl_basic_set
*bounds
, __isl_keep isl_set
*domain
,
1327 __isl_keep isl_ast_build
*build
)
1329 isl_constraint_list
*list
;
1331 if (!build
|| !graft
)
1332 return isl_ast_graft_free(graft
);
1334 list
= isl_basic_set_get_constraint_list(bounds
);
1336 graft
= refine_generic_split(graft
, list
, domain
, build
);
1341 /* Create a for node for the current level.
1343 * Mark the for node degenerate if "degenerate" is set.
1345 static __isl_give isl_ast_node
*create_for(__isl_keep isl_ast_build
*build
,
1355 depth
= isl_ast_build_get_depth(build
);
1356 id
= isl_ast_build_get_iterator_id(build
, depth
);
1357 node
= isl_ast_node_alloc_for(id
);
1359 node
= isl_ast_node_for_mark_degenerate(node
);
1364 /* If the ast_build_exploit_nested_bounds option is set, then return
1365 * the constraints enforced by all elements in "list".
1366 * Otherwise, return the universe.
1368 static __isl_give isl_basic_set
*extract_shared_enforced(
1369 __isl_keep isl_ast_graft_list
*list
, __isl_keep isl_ast_build
*build
)
1377 ctx
= isl_ast_graft_list_get_ctx(list
);
1378 if (isl_options_get_ast_build_exploit_nested_bounds(ctx
))
1379 return isl_ast_graft_list_extract_shared_enforced(list
, build
);
1381 space
= isl_ast_build_get_space(build
, 1);
1382 return isl_basic_set_universe(space
);
1385 /* Return the pending constraints of "build" that are not already taken
1386 * care of (by a combination of "enforced" and the generated constraints
1389 static __isl_give isl_set
*extract_pending(__isl_keep isl_ast_build
*build
,
1390 __isl_keep isl_basic_set
*enforced
)
1392 isl_set
*guard
, *context
;
1394 guard
= isl_ast_build_get_pending(build
);
1395 context
= isl_set_from_basic_set(isl_basic_set_copy(enforced
));
1396 context
= isl_set_intersect(context
,
1397 isl_ast_build_get_generated(build
));
1398 return isl_set_gist(guard
, context
);
1401 /* Create an AST node for the current dimension based on
1402 * the schedule domain "bounds" and return the node encapsulated
1403 * in an isl_ast_graft.
1405 * "executed" is the current inverse schedule, taking into account
1406 * the bounds in "bounds"
1407 * "domain" is the domain of "executed", with inner dimensions projected out.
1408 * It may be a strict subset of "bounds" in case "bounds" was created
1409 * based on the atomic option or based on separation with explicit bounds.
1411 * "domain" may satisfy additional equalities that result
1412 * from intersecting "executed" with "bounds" in add_node.
1413 * It may also satisfy some global constraints that were dropped out because
1414 * we performed separation with explicit bounds.
1415 * The very first step is then to copy these constraints to "bounds".
1417 * Since we may be calling before_each_for and after_each_for
1418 * callbacks, we record the current inverse schedule in the build.
1420 * We consider three builds,
1421 * "build" is the one in which the current level is created,
1422 * "body_build" is the build in which the next level is created,
1423 * "sub_build" is essentially the same as "body_build", except that
1424 * the depth has not been increased yet.
1426 * "build" already contains information (in strides and offsets)
1427 * about the strides at the current level, but this information is not
1428 * reflected in the build->domain.
1429 * We first add this information and the "bounds" to the sub_build->domain.
1430 * isl_ast_build_set_loop_bounds adds the stride information and
1431 * checks whether the current dimension attains
1432 * only a single value and whether this single value can be represented using
1433 * a single affine expression.
1434 * In the first case, the current level is considered "degenerate".
1435 * In the second, sub-case, the current level is considered "eliminated".
1436 * Eliminated levels don't need to be reflected in the AST since we can
1437 * simply plug in the affine expression. For degenerate, but non-eliminated,
1438 * levels, we do introduce a for node, but mark is as degenerate so that
1439 * it can be printed as an assignment of the single value to the loop
1442 * If the current level is eliminated, we explicitly plug in the value
1443 * for the current level found by isl_ast_build_set_loop_bounds in the
1444 * inverse schedule. This ensures that if we are working on a slice
1445 * of the domain based on information available in the inverse schedule
1446 * and the build domain, that then this information is also reflected
1447 * in the inverse schedule. This operation also eliminates the current
1448 * dimension from the inverse schedule making sure no inner dimensions depend
1449 * on the current dimension. Otherwise, we create a for node, marking
1450 * it degenerate if appropriate. The initial for node is still incomplete
1451 * and will be completed in either refine_degenerate or refine_generic.
1453 * We then generate a sequence of grafts for the next level,
1454 * create a surrounding graft for the current level and insert
1455 * the for node we created (if the current level is not eliminated).
1456 * Before creating a graft for the current level, we first extract
1457 * hoistable constraints from the child guards and combine them
1458 * with the pending constraints in the build. These constraints
1459 * are used to simplify the child guards and then added to the guard
1460 * of the current graft to ensure that they will be generated.
1461 * If the hoisted guard is a disjunction, then we use it directly
1462 * to gist the guards on the children before intersect it with the
1463 * pending constraints. We do so because this disjunction is typically
1464 * identical to the guards on the children such that these guards
1465 * can be effectively removed completely. After the intersection,
1466 * the gist operation would have a harder time figuring this out.
1468 * Finally, we set the bounds of the for loop in either
1469 * refine_degenerate or refine_generic.
1470 * We do so in a context where the pending constraints of the build
1471 * have been replaced by the guard of the current graft.
1473 static __isl_give isl_ast_graft
*create_node_scaled(
1474 __isl_take isl_union_map
*executed
,
1475 __isl_take isl_basic_set
*bounds
, __isl_take isl_set
*domain
,
1476 __isl_take isl_ast_build
*build
)
1480 isl_bool eliminated
;
1482 isl_basic_set
*hull
;
1483 isl_basic_set
*enforced
;
1484 isl_set
*guard
, *hoisted
;
1485 isl_ast_node
*node
= NULL
;
1486 isl_ast_graft
*graft
;
1487 isl_ast_graft_list
*children
;
1488 isl_ast_build
*sub_build
;
1489 isl_ast_build
*body_build
;
1491 domain
= isl_ast_build_eliminate_divs(build
, domain
);
1492 domain
= isl_set_detect_equalities(domain
);
1493 hull
= isl_set_unshifted_simple_hull(isl_set_copy(domain
));
1494 bounds
= isl_basic_set_intersect(bounds
, hull
);
1495 build
= isl_ast_build_set_executed(build
, isl_union_map_copy(executed
));
1497 depth
= isl_ast_build_get_depth(build
);
1498 sub_build
= isl_ast_build_copy(build
);
1499 bounds
= isl_basic_set_remove_redundancies(bounds
);
1500 bounds
= isl_ast_build_specialize_basic_set(sub_build
, bounds
);
1501 sub_build
= isl_ast_build_set_loop_bounds(sub_build
,
1502 isl_basic_set_copy(bounds
));
1503 degenerate
= isl_ast_build_has_value(sub_build
);
1504 eliminated
= isl_ast_build_has_affine_value(sub_build
, depth
);
1505 if (degenerate
< 0 || eliminated
< 0)
1506 executed
= isl_union_map_free(executed
);
1508 bounds
= isl_ast_build_compute_gist_basic_set(build
, bounds
);
1509 sub_build
= isl_ast_build_set_pending_generated(sub_build
,
1510 isl_basic_set_copy(bounds
));
1512 executed
= plug_in_values(executed
, sub_build
);
1514 node
= create_for(build
, degenerate
);
1516 body_build
= isl_ast_build_copy(sub_build
);
1517 body_build
= isl_ast_build_increase_depth(body_build
);
1519 node
= before_each_for(node
, body_build
);
1520 children
= generate_next_level(executed
,
1521 isl_ast_build_copy(body_build
));
1523 enforced
= extract_shared_enforced(children
, build
);
1524 guard
= extract_pending(sub_build
, enforced
);
1525 hoisted
= isl_ast_graft_list_extract_hoistable_guard(children
, build
);
1526 n
= isl_set_n_basic_set(hoisted
);
1528 children
= isl_ast_graft_list_free(children
);
1530 children
= isl_ast_graft_list_gist_guards(children
,
1531 isl_set_copy(hoisted
));
1532 guard
= isl_set_intersect(guard
, hoisted
);
1534 guard
= add_implied_guards(guard
, degenerate
, bounds
, build
);
1536 graft
= isl_ast_graft_alloc_from_children(children
,
1537 isl_set_copy(guard
), enforced
, build
, sub_build
);
1540 isl_ast_build
*for_build
;
1542 graft
= isl_ast_graft_insert_for(graft
, node
);
1543 for_build
= isl_ast_build_copy(build
);
1544 for_build
= isl_ast_build_replace_pending_by_guard(for_build
,
1545 isl_set_copy(guard
));
1547 graft
= refine_degenerate(graft
, for_build
, sub_build
);
1549 graft
= refine_generic(graft
, bounds
,
1551 isl_ast_build_free(for_build
);
1553 isl_set_free(guard
);
1555 graft
= after_each_for(graft
, body_build
);
1557 isl_ast_build_free(body_build
);
1558 isl_ast_build_free(sub_build
);
1559 isl_ast_build_free(build
);
1560 isl_basic_set_free(bounds
);
1561 isl_set_free(domain
);
1566 /* Internal data structure for checking if all constraints involving
1567 * the input dimension "depth" are such that the other coefficients
1568 * are multiples of "m", reducing "m" if they are not.
1569 * If "m" is reduced all the way down to "1", then the check has failed
1570 * and we break out of the iteration.
1572 struct isl_check_scaled_data
{
1577 /* If constraint "c" involves the input dimension data->depth,
1578 * then make sure that all the other coefficients are multiples of data->m,
1579 * reducing data->m if needed.
1580 * Break out of the iteration if data->m has become equal to "1".
1582 static isl_stat
constraint_check_scaled(__isl_take isl_constraint
*c
,
1585 struct isl_check_scaled_data
*data
= user
;
1588 enum isl_dim_type t
[] = { isl_dim_param
, isl_dim_in
, isl_dim_out
,
1591 if (!isl_constraint_involves_dims(c
, isl_dim_in
, data
->depth
, 1)) {
1592 isl_constraint_free(c
);
1596 for (i
= 0; i
< 4; ++i
) {
1597 n
= isl_constraint_dim(c
, t
[i
]);
1600 for (j
= 0; j
< n
; ++j
) {
1603 if (t
[i
] == isl_dim_in
&& j
== data
->depth
)
1605 if (!isl_constraint_involves_dims(c
, t
[i
], j
, 1))
1607 d
= isl_constraint_get_coefficient_val(c
, t
[i
], j
);
1608 data
->m
= isl_val_gcd(data
->m
, d
);
1609 if (isl_val_is_one(data
->m
))
1616 isl_constraint_free(c
);
1618 return i
< 4 ? isl_stat_error
: isl_stat_ok
;
1621 /* For each constraint of "bmap" that involves the input dimension data->depth,
1622 * make sure that all the other coefficients are multiples of data->m,
1623 * reducing data->m if needed.
1624 * Break out of the iteration if data->m has become equal to "1".
1626 static isl_stat
basic_map_check_scaled(__isl_take isl_basic_map
*bmap
,
1631 r
= isl_basic_map_foreach_constraint(bmap
,
1632 &constraint_check_scaled
, user
);
1633 isl_basic_map_free(bmap
);
1638 /* For each constraint of "map" 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
map_check_scaled(__isl_take isl_map
*map
, void *user
)
1647 r
= isl_map_foreach_basic_map(map
, &basic_map_check_scaled
, user
);
1653 /* Create an AST node for the current dimension based on
1654 * the schedule domain "bounds" and return the node encapsulated
1655 * in an isl_ast_graft.
1657 * "executed" is the current inverse schedule, taking into account
1658 * the bounds in "bounds"
1659 * "domain" is the domain of "executed", with inner dimensions projected out.
1662 * Before moving on to the actual AST node construction in create_node_scaled,
1663 * we first check if the current dimension is strided and if we can scale
1664 * down this stride. Note that we only do this if the ast_build_scale_strides
1667 * In particular, let the current dimension take on values
1671 * with a an integer. We check if we can find an integer m that (obviously)
1672 * divides both f and s.
1674 * If so, we check if the current dimension only appears in constraints
1675 * where the coefficients of the other variables are multiples of m.
1676 * We perform this extra check to avoid the risk of introducing
1677 * divisions by scaling down the current dimension.
1679 * If so, we scale the current dimension down by a factor of m.
1680 * That is, we plug in
1684 * Note that in principle we could always scale down strided loops
1689 * but this may result in i' taking on larger values than the original i,
1690 * due to the shift by "f".
1691 * By constrast, the scaling in (1) can only reduce the (absolute) value "i".
1693 static __isl_give isl_ast_graft
*create_node(__isl_take isl_union_map
*executed
,
1694 __isl_take isl_basic_set
*bounds
, __isl_take isl_set
*domain
,
1695 __isl_take isl_ast_build
*build
)
1697 struct isl_check_scaled_data data
;
1702 ctx
= isl_ast_build_get_ctx(build
);
1703 if (!isl_options_get_ast_build_scale_strides(ctx
))
1704 return create_node_scaled(executed
, bounds
, domain
, build
);
1706 data
.depth
= isl_ast_build_get_depth(build
);
1707 if (!isl_ast_build_has_stride(build
, data
.depth
))
1708 return create_node_scaled(executed
, bounds
, domain
, build
);
1710 offset
= isl_ast_build_get_offset(build
, data
.depth
);
1711 data
.m
= isl_ast_build_get_stride(build
, data
.depth
);
1713 offset
= isl_aff_free(offset
);
1714 offset
= isl_aff_scale_down_val(offset
, isl_val_copy(data
.m
));
1715 d
= isl_aff_get_denominator_val(offset
);
1717 executed
= isl_union_map_free(executed
);
1719 if (executed
&& isl_val_is_divisible_by(data
.m
, d
))
1720 data
.m
= isl_val_div(data
.m
, d
);
1722 data
.m
= isl_val_set_si(data
.m
, 1);
1726 if (!isl_val_is_one(data
.m
)) {
1727 if (isl_union_map_foreach_map(executed
, &map_check_scaled
,
1729 !isl_val_is_one(data
.m
))
1730 executed
= isl_union_map_free(executed
);
1733 if (!isl_val_is_one(data
.m
)) {
1738 isl_union_map
*umap
;
1740 space
= isl_ast_build_get_space(build
, 1);
1741 space
= isl_space_map_from_set(space
);
1742 ma
= isl_multi_aff_identity(space
);
1743 aff
= isl_multi_aff_get_aff(ma
, data
.depth
);
1744 aff
= isl_aff_scale_val(aff
, isl_val_copy(data
.m
));
1745 ma
= isl_multi_aff_set_aff(ma
, data
.depth
, aff
);
1747 bounds
= isl_basic_set_preimage_multi_aff(bounds
,
1748 isl_multi_aff_copy(ma
));
1749 domain
= isl_set_preimage_multi_aff(domain
,
1750 isl_multi_aff_copy(ma
));
1751 map
= isl_map_reverse(isl_map_from_multi_aff(ma
));
1752 umap
= isl_union_map_from_map(map
);
1753 executed
= isl_union_map_apply_domain(executed
,
1754 isl_union_map_copy(umap
));
1755 build
= isl_ast_build_scale_down(build
, isl_val_copy(data
.m
),
1758 isl_aff_free(offset
);
1759 isl_val_free(data
.m
);
1761 return create_node_scaled(executed
, bounds
, domain
, build
);
1764 /* Add the basic set to the list that "user" points to.
1766 static isl_stat
collect_basic_set(__isl_take isl_basic_set
*bset
, void *user
)
1768 isl_basic_set_list
**list
= user
;
1770 *list
= isl_basic_set_list_add(*list
, bset
);
1775 /* Extract the basic sets of "set" and collect them in an isl_basic_set_list.
1777 static __isl_give isl_basic_set_list
*isl_basic_set_list_from_set(
1778 __isl_take isl_set
*set
)
1782 isl_basic_set_list
*list
;
1784 n
= isl_set_n_basic_set(set
);
1786 set
= isl_set_free(set
);
1790 ctx
= isl_set_get_ctx(set
);
1792 list
= isl_basic_set_list_alloc(ctx
, n
);
1793 if (isl_set_foreach_basic_set(set
, &collect_basic_set
, &list
) < 0)
1794 list
= isl_basic_set_list_free(list
);
1800 /* Generate code for the schedule domain "bounds"
1801 * and add the result to "list".
1803 * We mainly detect strides here and check if the bounds do not
1804 * conflict with the current build domain
1805 * and then pass over control to create_node.
1807 * "bounds" reflects the bounds on the current dimension and possibly
1808 * some extra conditions on outer dimensions.
1809 * It does not, however, include any divs involving the current dimension,
1810 * so it does not capture any stride constraints.
1811 * We therefore need to compute that part of the schedule domain that
1812 * intersects with "bounds" and derive the strides from the result.
1814 static __isl_give isl_ast_graft_list
*add_node(
1815 __isl_take isl_ast_graft_list
*list
, __isl_take isl_union_map
*executed
,
1816 __isl_take isl_basic_set
*bounds
, __isl_take isl_ast_build
*build
)
1818 isl_ast_graft
*graft
;
1819 isl_set
*domain
= NULL
;
1820 isl_union_set
*uset
;
1821 int empty
, disjoint
;
1823 uset
= isl_union_set_from_basic_set(isl_basic_set_copy(bounds
));
1824 executed
= isl_union_map_intersect_domain(executed
, uset
);
1825 empty
= isl_union_map_is_empty(executed
);
1831 uset
= isl_union_map_domain(isl_union_map_copy(executed
));
1832 domain
= isl_set_from_union_set(uset
);
1833 domain
= isl_ast_build_specialize(build
, domain
);
1835 domain
= isl_set_compute_divs(domain
);
1836 domain
= isl_ast_build_eliminate_inner(build
, domain
);
1837 disjoint
= isl_set_is_disjoint(domain
, build
->domain
);
1843 build
= isl_ast_build_detect_strides(build
, isl_set_copy(domain
));
1845 graft
= create_node(executed
, bounds
, domain
,
1846 isl_ast_build_copy(build
));
1847 list
= isl_ast_graft_list_add(list
, graft
);
1848 isl_ast_build_free(build
);
1851 list
= isl_ast_graft_list_free(list
);
1853 isl_set_free(domain
);
1854 isl_basic_set_free(bounds
);
1855 isl_union_map_free(executed
);
1856 isl_ast_build_free(build
);
1860 /* Does any element of i follow or coincide with any element of j
1861 * at the current depth for equal values of the outer dimensions?
1863 static isl_bool
domain_follows_at_depth(__isl_keep isl_basic_set
*i
,
1864 __isl_keep isl_basic_set
*j
, void *user
)
1866 int depth
= *(int *) user
;
1867 isl_basic_map
*test
;
1871 test
= isl_basic_map_from_domain_and_range(isl_basic_set_copy(i
),
1872 isl_basic_set_copy(j
));
1873 for (l
= 0; l
< depth
; ++l
)
1874 test
= isl_basic_map_equate(test
, isl_dim_in
, l
,
1876 test
= isl_basic_map_order_ge(test
, isl_dim_in
, depth
,
1877 isl_dim_out
, depth
);
1878 empty
= isl_basic_map_is_empty(test
);
1879 isl_basic_map_free(test
);
1881 return isl_bool_not(empty
);
1884 /* Split up each element of "list" into a part that is related to "bset"
1885 * according to "gt" and a part that is not.
1886 * Return a list that consist of "bset" and all the pieces.
1888 static __isl_give isl_basic_set_list
*add_split_on(
1889 __isl_take isl_basic_set_list
*list
, __isl_take isl_basic_set
*bset
,
1890 __isl_keep isl_basic_map
*gt
)
1894 isl_basic_set_list
*res
;
1896 n
= isl_basic_set_list_n_basic_set(list
);
1898 bset
= isl_basic_set_free(bset
);
1900 gt
= isl_basic_map_copy(gt
);
1901 gt
= isl_basic_map_intersect_domain(gt
, isl_basic_set_copy(bset
));
1902 res
= isl_basic_set_list_from_basic_set(bset
);
1903 for (i
= 0; res
&& i
< n
; ++i
) {
1904 isl_basic_set
*bset
;
1905 isl_set
*set1
, *set2
;
1906 isl_basic_map
*bmap
;
1909 bset
= isl_basic_set_list_get_basic_set(list
, i
);
1910 bmap
= isl_basic_map_copy(gt
);
1911 bmap
= isl_basic_map_intersect_range(bmap
, bset
);
1912 bset
= isl_basic_map_range(bmap
);
1913 empty
= isl_basic_set_is_empty(bset
);
1915 res
= isl_basic_set_list_free(res
);
1917 isl_basic_set_free(bset
);
1918 bset
= isl_basic_set_list_get_basic_set(list
, i
);
1919 res
= isl_basic_set_list_add(res
, bset
);
1923 res
= isl_basic_set_list_add(res
, isl_basic_set_copy(bset
));
1924 set1
= isl_set_from_basic_set(bset
);
1925 bset
= isl_basic_set_list_get_basic_set(list
, i
);
1926 set2
= isl_set_from_basic_set(bset
);
1927 set1
= isl_set_subtract(set2
, set1
);
1928 set1
= isl_set_make_disjoint(set1
);
1930 res
= isl_basic_set_list_concat(res
,
1931 isl_basic_set_list_from_set(set1
));
1933 isl_basic_map_free(gt
);
1934 isl_basic_set_list_free(list
);
1938 static __isl_give isl_ast_graft_list
*generate_sorted_domains(
1939 __isl_keep isl_basic_set_list
*domain_list
,
1940 __isl_keep isl_union_map
*executed
,
1941 __isl_keep isl_ast_build
*build
);
1943 /* Internal data structure for add_nodes.
1945 * "executed" and "build" are extra arguments to be passed to add_node.
1946 * "list" collects the results.
1948 struct isl_add_nodes_data
{
1949 isl_union_map
*executed
;
1950 isl_ast_build
*build
;
1952 isl_ast_graft_list
*list
;
1955 /* Generate code for the schedule domains in "scc"
1956 * and add the results to "list".
1958 * The domains in "scc" form a strongly connected component in the ordering.
1959 * If the number of domains in "scc" is larger than 1, then this means
1960 * that we cannot determine a valid ordering for the domains in the component.
1961 * This should be fairly rare because the individual domains
1962 * have been made disjoint first.
1963 * The problem is that the domains may be integrally disjoint but not
1964 * rationally disjoint. For example, we may have domains
1966 * { [i,i] : 0 <= i <= 1 } and { [i,1-i] : 0 <= i <= 1 }
1968 * These two domains have an empty intersection, but their rational
1969 * relaxations do intersect. It is impossible to order these domains
1970 * in the second dimension because the first should be ordered before
1971 * the second for outer dimension equal to 0, while it should be ordered
1972 * after for outer dimension equal to 1.
1974 * This may happen in particular in case of unrolling since the domain
1975 * of each slice is replaced by its simple hull.
1977 * For each basic set i in "scc" and for each of the following basic sets j,
1978 * we split off that part of the basic set i that shares the outer dimensions
1979 * with j and lies before j in the current dimension.
1980 * We collect all the pieces in a new list that replaces "scc".
1982 * While the elements in "scc" should be disjoint, we double-check
1983 * this property to avoid running into an infinite recursion in case
1984 * they intersect due to some internal error.
1986 static isl_stat
add_nodes(__isl_take isl_basic_set_list
*scc
, void *user
)
1988 struct isl_add_nodes_data
*data
= user
;
1991 isl_basic_set
*bset
, *first
;
1992 isl_basic_set_list
*list
;
1996 n
= isl_basic_set_list_n_basic_set(scc
);
1999 bset
= isl_basic_set_list_get_basic_set(scc
, 0);
2001 isl_basic_set_list_free(scc
);
2002 data
->list
= add_node(data
->list
,
2003 isl_union_map_copy(data
->executed
), bset
,
2004 isl_ast_build_copy(data
->build
));
2005 return data
->list
? isl_stat_ok
: isl_stat_error
;
2008 depth
= isl_ast_build_get_depth(data
->build
);
2009 space
= isl_basic_set_get_space(bset
);
2010 space
= isl_space_map_from_set(space
);
2011 gt
= isl_basic_map_universe(space
);
2012 for (i
= 0; i
< depth
; ++i
)
2013 gt
= isl_basic_map_equate(gt
, isl_dim_in
, i
, isl_dim_out
, i
);
2014 gt
= isl_basic_map_order_gt(gt
, isl_dim_in
, depth
, isl_dim_out
, depth
);
2016 first
= isl_basic_set_copy(bset
);
2017 list
= isl_basic_set_list_from_basic_set(bset
);
2018 for (i
= 1; i
< n
; ++i
) {
2021 bset
= isl_basic_set_list_get_basic_set(scc
, i
);
2023 disjoint
= isl_basic_set_is_disjoint(bset
, first
);
2025 list
= isl_basic_set_list_free(list
);
2027 isl_die(isl_basic_set_list_get_ctx(scc
),
2029 "basic sets in scc are assumed to be disjoint",
2030 list
= isl_basic_set_list_free(list
));
2032 list
= add_split_on(list
, bset
, gt
);
2034 isl_basic_set_free(first
);
2035 isl_basic_map_free(gt
);
2036 isl_basic_set_list_free(scc
);
2038 data
->list
= isl_ast_graft_list_concat(data
->list
,
2039 generate_sorted_domains(scc
, data
->executed
, data
->build
));
2040 isl_basic_set_list_free(scc
);
2042 return data
->list
? isl_stat_ok
: isl_stat_error
;
2044 isl_basic_set_list_free(scc
);
2045 return isl_stat_error
;
2048 /* Sort the domains in "domain_list" according to the execution order
2049 * at the current depth (for equal values of the outer dimensions),
2050 * generate code for each of them, collecting the results in a list.
2051 * If no code is generated (because the intersection of the inverse schedule
2052 * with the domains turns out to be empty), then an empty list is returned.
2054 * The caller is responsible for ensuring that the basic sets in "domain_list"
2055 * are pair-wise disjoint. It can, however, in principle happen that
2056 * two basic sets should be ordered one way for one value of the outer
2057 * dimensions and the other way for some other value of the outer dimensions.
2058 * We therefore play safe and look for strongly connected components.
2059 * The function add_nodes takes care of handling non-trivial components.
2061 static __isl_give isl_ast_graft_list
*generate_sorted_domains(
2062 __isl_keep isl_basic_set_list
*domain_list
,
2063 __isl_keep isl_union_map
*executed
, __isl_keep isl_ast_build
*build
)
2066 struct isl_add_nodes_data data
;
2070 n
= isl_basic_set_list_n_basic_set(domain_list
);
2074 ctx
= isl_basic_set_list_get_ctx(domain_list
);
2075 data
.list
= isl_ast_graft_list_alloc(ctx
, n
);
2079 return add_node(data
.list
, isl_union_map_copy(executed
),
2080 isl_basic_set_list_get_basic_set(domain_list
, 0),
2081 isl_ast_build_copy(build
));
2083 depth
= isl_ast_build_get_depth(build
);
2084 data
.executed
= executed
;
2086 if (isl_basic_set_list_foreach_scc(domain_list
,
2087 &domain_follows_at_depth
, &depth
,
2088 &add_nodes
, &data
) < 0)
2089 data
.list
= isl_ast_graft_list_free(data
.list
);
2094 /* Do i and j share any values for the outer dimensions?
2096 static isl_bool
shared_outer(__isl_keep isl_basic_set
*i
,
2097 __isl_keep isl_basic_set
*j
, void *user
)
2099 int depth
= *(int *) user
;
2100 isl_basic_map
*test
;
2104 test
= isl_basic_map_from_domain_and_range(isl_basic_set_copy(i
),
2105 isl_basic_set_copy(j
));
2106 for (l
= 0; l
< depth
; ++l
)
2107 test
= isl_basic_map_equate(test
, isl_dim_in
, l
,
2109 empty
= isl_basic_map_is_empty(test
);
2110 isl_basic_map_free(test
);
2112 return isl_bool_not(empty
);
2115 /* Internal data structure for generate_sorted_domains_wrap.
2117 * "n" is the total number of basic sets
2118 * "executed" and "build" are extra arguments to be passed
2119 * to generate_sorted_domains.
2121 * "single" is set to 1 by generate_sorted_domains_wrap if there
2122 * is only a single component.
2123 * "list" collects the results.
2125 struct isl_ast_generate_parallel_domains_data
{
2127 isl_union_map
*executed
;
2128 isl_ast_build
*build
;
2131 isl_ast_graft_list
*list
;
2134 /* Call generate_sorted_domains on "scc", fuse the result into a list
2135 * with either zero or one graft and collect the these single element
2136 * lists into data->list.
2138 * If there is only one component, i.e., if the number of basic sets
2139 * in the current component is equal to the total number of basic sets,
2140 * then data->single is set to 1 and the result of generate_sorted_domains
2143 static isl_stat
generate_sorted_domains_wrap(__isl_take isl_basic_set_list
*scc
,
2146 struct isl_ast_generate_parallel_domains_data
*data
= user
;
2147 isl_ast_graft_list
*list
;
2150 n
= isl_basic_set_list_n_basic_set(scc
);
2152 scc
= isl_basic_set_list_free(scc
);
2153 list
= generate_sorted_domains(scc
, data
->executed
, data
->build
);
2154 data
->single
= n
== data
->n
;
2156 list
= isl_ast_graft_list_fuse(list
, data
->build
);
2160 data
->list
= isl_ast_graft_list_concat(data
->list
, list
);
2162 isl_basic_set_list_free(scc
);
2164 return isl_stat_error
;
2169 /* Look for any (weakly connected) components in the "domain_list"
2170 * of domains that share some values of the outer dimensions.
2171 * That is, domains in different components do not share any values
2172 * of the outer dimensions. This means that these components
2173 * can be freely reordered.
2174 * Within each of the components, we sort the domains according
2175 * to the execution order at the current depth.
2177 * If there is more than one component, then generate_sorted_domains_wrap
2178 * fuses the result of each call to generate_sorted_domains
2179 * into a list with either zero or one graft and collects these (at most)
2180 * single element lists into a bigger list. This means that the elements of the
2181 * final list can be freely reordered. In particular, we sort them
2182 * according to an arbitrary but fixed ordering to ease merging of
2183 * graft lists from different components.
2185 static __isl_give isl_ast_graft_list
*generate_parallel_domains(
2186 __isl_keep isl_basic_set_list
*domain_list
,
2187 __isl_keep isl_union_map
*executed
, __isl_keep isl_ast_build
*build
)
2190 struct isl_ast_generate_parallel_domains_data data
;
2192 data
.n
= isl_basic_set_list_n_basic_set(domain_list
);
2197 return generate_sorted_domains(domain_list
, executed
, build
);
2199 depth
= isl_ast_build_get_depth(build
);
2201 data
.executed
= executed
;
2204 if (isl_basic_set_list_foreach_scc(domain_list
, &shared_outer
, &depth
,
2205 &generate_sorted_domains_wrap
,
2207 data
.list
= isl_ast_graft_list_free(data
.list
);
2210 data
.list
= isl_ast_graft_list_sort_guard(data
.list
);
2215 /* Internal data for separate_domain.
2217 * "explicit" is set if we only want to use explicit bounds.
2219 * "domain" collects the separated domains.
2221 struct isl_separate_domain_data
{
2222 isl_ast_build
*build
;
2227 /* Extract implicit bounds on the current dimension for the executed "map".
2229 * The domain of "map" may involve inner dimensions, so we
2230 * need to eliminate them.
2232 static __isl_give isl_set
*implicit_bounds(__isl_take isl_map
*map
,
2233 __isl_keep isl_ast_build
*build
)
2237 domain
= isl_map_domain(map
);
2238 domain
= isl_ast_build_eliminate(build
, domain
);
2243 /* Extract explicit bounds on the current dimension for the executed "map".
2245 * Rather than eliminating the inner dimensions as in implicit_bounds,
2246 * we simply drop any constraints involving those inner dimensions.
2247 * The idea is that most bounds that are implied by constraints on the
2248 * inner dimensions will be enforced by for loops and not by explicit guards.
2249 * There is then no need to separate along those bounds.
2251 static __isl_give isl_set
*explicit_bounds(__isl_take isl_map
*map
,
2252 __isl_keep isl_ast_build
*build
)
2258 dim
= isl_map_dim(map
, isl_dim_out
);
2260 return isl_map_domain(isl_map_free(map
));
2261 map
= isl_map_drop_constraints_involving_dims(map
, isl_dim_out
, 0, dim
);
2263 domain
= isl_map_domain(map
);
2264 depth
= isl_ast_build_get_depth(build
);
2265 dim
= isl_set_dim(domain
, isl_dim_set
);
2266 domain
= isl_set_detect_equalities(domain
);
2267 domain
= isl_set_drop_constraints_involving_dims(domain
,
2268 isl_dim_set
, depth
+ 1, dim
- (depth
+ 1));
2269 domain
= isl_set_remove_divs_involving_dims(domain
,
2270 isl_dim_set
, depth
, 1);
2271 domain
= isl_set_remove_unknown_divs(domain
);
2276 /* Split data->domain into pieces that intersect with the range of "map"
2277 * and pieces that do not intersect with the range of "map"
2278 * and then add that part of the range of "map" that does not intersect
2279 * with data->domain.
2281 static isl_stat
separate_domain(__isl_take isl_map
*map
, void *user
)
2283 struct isl_separate_domain_data
*data
= user
;
2288 domain
= explicit_bounds(map
, data
->build
);
2290 domain
= implicit_bounds(map
, data
->build
);
2292 domain
= isl_set_coalesce(domain
);
2293 domain
= isl_set_make_disjoint(domain
);
2294 d1
= isl_set_subtract(isl_set_copy(domain
), isl_set_copy(data
->domain
));
2295 d2
= isl_set_subtract(isl_set_copy(data
->domain
), isl_set_copy(domain
));
2296 data
->domain
= isl_set_intersect(data
->domain
, domain
);
2297 data
->domain
= isl_set_union(data
->domain
, d1
);
2298 data
->domain
= isl_set_union(data
->domain
, d2
);
2303 /* Separate the schedule domains of "executed".
2305 * That is, break up the domain of "executed" into basic sets,
2306 * such that for each basic set S, every element in S is associated with
2307 * the same domain spaces.
2309 * "space" is the (single) domain space of "executed".
2311 static __isl_give isl_set
*separate_schedule_domains(
2312 __isl_take isl_space
*space
, __isl_take isl_union_map
*executed
,
2313 __isl_keep isl_ast_build
*build
)
2315 struct isl_separate_domain_data data
= { build
};
2318 ctx
= isl_ast_build_get_ctx(build
);
2319 data
.explicit = isl_options_get_ast_build_separation_bounds(ctx
) ==
2320 ISL_AST_BUILD_SEPARATION_BOUNDS_EXPLICIT
;
2321 data
.domain
= isl_set_empty(space
);
2322 if (isl_union_map_foreach_map(executed
, &separate_domain
, &data
) < 0)
2323 data
.domain
= isl_set_free(data
.domain
);
2325 isl_union_map_free(executed
);
2329 /* Temporary data used during the search for a lower bound for unrolling.
2331 * "build" is the build in which the unrolling will be performed
2332 * "domain" is the original set for which to find a lower bound
2333 * "depth" is the dimension for which to find a lower boudn
2334 * "expansion" is the expansion that needs to be applied to "domain"
2335 * in the unrolling that will be performed
2337 * "lower" is the best lower bound found so far. It is NULL if we have not
2339 * "n" is the corresponding size. If lower is NULL, then the value of n
2341 * "n_div" is the maximal number of integer divisions in the first
2342 * unrolled iteration (after expansion). It is set to -1 if it hasn't
2343 * been computed yet.
2345 struct isl_find_unroll_data
{
2346 isl_ast_build
*build
;
2349 isl_basic_map
*expansion
;
2356 /* Return the constraint
2358 * i_"depth" = aff + offset
2360 static __isl_give isl_constraint
*at_offset(int depth
, __isl_keep isl_aff
*aff
,
2363 aff
= isl_aff_copy(aff
);
2364 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, depth
, -1);
2365 aff
= isl_aff_add_constant_si(aff
, offset
);
2366 return isl_equality_from_aff(aff
);
2369 /* Update *user to the number of integer divisions in the first element
2370 * of "ma", if it is larger than the current value.
2372 static isl_stat
update_n_div(__isl_take isl_set
*set
,
2373 __isl_take isl_multi_aff
*ma
, void *user
)
2379 aff
= isl_multi_aff_get_aff(ma
, 0);
2380 n_div
= isl_aff_dim(aff
, isl_dim_div
);
2382 isl_multi_aff_free(ma
);
2388 return n_div
>= 0 ? isl_stat_ok
: isl_stat_error
;
2391 /* Get the number of integer divisions in the expression for the iterator
2392 * value at the first slice in the unrolling based on lower bound "lower",
2393 * taking into account the expansion that needs to be performed on this slice.
2395 static int get_expanded_n_div(struct isl_find_unroll_data
*data
,
2396 __isl_keep isl_aff
*lower
)
2400 isl_map
*it_map
, *expansion
;
2401 isl_pw_multi_aff
*pma
;
2404 c
= at_offset(data
->depth
, lower
, 0);
2405 set
= isl_set_copy(data
->domain
);
2406 set
= isl_set_add_constraint(set
, c
);
2407 expansion
= isl_map_from_basic_map(isl_basic_map_copy(data
->expansion
));
2408 set
= isl_set_apply(set
, expansion
);
2409 it_map
= isl_ast_build_map_to_iterator(data
->build
, set
);
2410 pma
= isl_pw_multi_aff_from_map(it_map
);
2412 if (isl_pw_multi_aff_foreach_piece(pma
, &update_n_div
, &n
) < 0)
2414 isl_pw_multi_aff_free(pma
);
2419 /* Is the lower bound "lower" with corresponding iteration count "n"
2420 * better than the one stored in "data"?
2421 * If there is no upper bound on the iteration count ("n" is infinity) or
2422 * if the count is too large, then we cannot use this lower bound.
2423 * Otherwise, if there was no previous lower bound or
2424 * if the iteration count of the new lower bound is smaller than
2425 * the iteration count of the previous lower bound, then we consider
2426 * the new lower bound to be better.
2427 * If the iteration count is the same, then compare the number
2428 * of integer divisions that would be needed to express
2429 * the iterator value at the first slice in the unrolling
2430 * according to the lower bound. If we end up computing this
2431 * number, then store the lowest value in data->n_div.
2433 static int is_better_lower_bound(struct isl_find_unroll_data
*data
,
2434 __isl_keep isl_aff
*lower
, __isl_keep isl_val
*n
)
2441 if (isl_val_is_infty(n
))
2443 if (isl_val_cmp_si(n
, INT_MAX
) > 0)
2447 cmp
= isl_val_cmp_si(n
, *data
->n
);
2452 if (data
->n_div
< 0)
2453 data
->n_div
= get_expanded_n_div(data
, data
->lower
);
2454 if (data
->n_div
< 0)
2456 if (data
->n_div
== 0)
2458 n_div
= get_expanded_n_div(data
, lower
);
2461 if (n_div
>= data
->n_div
)
2463 data
->n_div
= n_div
;
2468 /* Check if we can use "c" as a lower bound and if it is better than
2469 * any previously found lower bound.
2471 * If "c" does not involve the dimension at the current depth,
2472 * then we cannot use it.
2473 * Otherwise, let "c" be of the form
2477 * We compute the maximal value of
2479 * -ceil(f(j)/a)) + i + 1
2481 * over the domain. If there is such a value "n", then we know
2483 * -ceil(f(j)/a)) + i + 1 <= n
2487 * i < ceil(f(j)/a)) + n
2489 * meaning that we can use ceil(f(j)/a)) as a lower bound for unrolling.
2490 * We just need to check if we have found any lower bound before and
2491 * if the new lower bound is better (smaller n or fewer integer divisions)
2492 * than the previously found lower bounds.
2494 static isl_stat
update_unrolling_lower_bound(struct isl_find_unroll_data
*data
,
2495 __isl_keep isl_constraint
*c
)
2497 isl_aff
*aff
, *lower
;
2501 if (!isl_constraint_is_lower_bound(c
, isl_dim_set
, data
->depth
))
2504 lower
= isl_constraint_get_bound(c
, isl_dim_set
, data
->depth
);
2505 lower
= isl_aff_ceil(lower
);
2506 aff
= isl_aff_copy(lower
);
2507 aff
= isl_aff_neg(aff
);
2508 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, data
->depth
, 1);
2509 aff
= isl_aff_add_constant_si(aff
, 1);
2510 max
= isl_set_max_val(data
->domain
, aff
);
2513 better
= is_better_lower_bound(data
, lower
, max
);
2514 if (better
< 0 || !better
) {
2516 isl_aff_free(lower
);
2517 return better
< 0 ? isl_stat_error
: isl_stat_ok
;
2520 isl_aff_free(data
->lower
);
2521 data
->lower
= lower
;
2522 *data
->n
= isl_val_get_num_si(max
);
2528 /* Check if we can use "c" as a lower bound and if it is better than
2529 * any previously found lower bound.
2531 static isl_stat
constraint_find_unroll(__isl_take isl_constraint
*c
, void *user
)
2533 struct isl_find_unroll_data
*data
;
2536 data
= (struct isl_find_unroll_data
*) user
;
2537 r
= update_unrolling_lower_bound(data
, c
);
2538 isl_constraint_free(c
);
2543 /* Look for a lower bound l(i) on the dimension at "depth"
2544 * and a size n such that "domain" is a subset of
2546 * { [i] : l(i) <= i_d < l(i) + n }
2548 * where d is "depth" and l(i) depends only on earlier dimensions.
2549 * Furthermore, try and find a lower bound such that n is as small as possible.
2550 * In particular, "n" needs to be finite.
2551 * "build" is the build in which the unrolling will be performed.
2552 * "expansion" is the expansion that needs to be applied to "domain"
2553 * in the unrolling that will be performed.
2555 * Inner dimensions have been eliminated from "domain" by the caller.
2557 * We first construct a collection of lower bounds on the input set
2558 * by computing its simple hull. We then iterate through them,
2559 * discarding those that we cannot use (either because they do not
2560 * involve the dimension at "depth" or because they have no corresponding
2561 * upper bound, meaning that "n" would be unbounded) and pick out the
2562 * best from the remaining ones.
2564 * If we cannot find a suitable lower bound, then we consider that
2567 static __isl_give isl_aff
*find_unroll_lower_bound(
2568 __isl_keep isl_ast_build
*build
, __isl_keep isl_set
*domain
,
2569 int depth
, __isl_keep isl_basic_map
*expansion
, int *n
)
2571 struct isl_find_unroll_data data
=
2572 { build
, domain
, depth
, expansion
, NULL
, n
, -1 };
2573 isl_basic_set
*hull
;
2575 hull
= isl_set_simple_hull(isl_set_copy(domain
));
2577 if (isl_basic_set_foreach_constraint(hull
,
2578 &constraint_find_unroll
, &data
) < 0)
2581 isl_basic_set_free(hull
);
2584 isl_die(isl_set_get_ctx(domain
), isl_error_invalid
,
2585 "cannot find lower bound for unrolling", return NULL
);
2589 isl_basic_set_free(hull
);
2590 return isl_aff_free(data
.lower
);
2593 /* Call "fn" on each iteration of the current dimension of "domain".
2594 * If "init" is not NULL, then it is called with the number of
2595 * iterations before any call to "fn".
2596 * Return -1 on failure.
2598 * Since we are going to be iterating over the individual values,
2599 * we first check if there are any strides on the current dimension.
2600 * If there is, we rewrite the current dimension i as
2602 * i = stride i' + offset
2604 * and then iterate over individual values of i' instead.
2606 * We then look for a lower bound on i' and a size such that the domain
2609 * { [j,i'] : l(j) <= i' < l(j) + n }
2611 * and then take slices of the domain at values of i'
2612 * between l(j) and l(j) + n - 1.
2614 * We compute the unshifted simple hull of each slice to ensure that
2615 * we have a single basic set per offset. The slicing constraint
2616 * may get simplified away before the unshifted simple hull is taken
2617 * and may therefore in some rare cases disappear from the result.
2618 * We therefore explicitly add the constraint back after computing
2619 * the unshifted simple hull to ensure that the basic sets
2620 * remain disjoint. The constraints that are dropped by taking the hull
2621 * will be taken into account at the next level, as in the case of the
2624 * Finally, we map i' back to i and call "fn".
2626 static int foreach_iteration(__isl_take isl_set
*domain
,
2627 __isl_keep isl_ast_build
*build
, int (*init
)(int n
, void *user
),
2628 int (*fn
)(__isl_take isl_basic_set
*bset
, void *user
), void *user
)
2633 isl_multi_aff
*expansion
;
2634 isl_basic_map
*bmap
;
2635 isl_aff
*lower
= NULL
;
2636 isl_ast_build
*stride_build
;
2638 depth
= isl_ast_build_get_depth(build
);
2640 domain
= isl_ast_build_eliminate_inner(build
, domain
);
2641 domain
= isl_set_intersect(domain
, isl_ast_build_get_domain(build
));
2642 stride_build
= isl_ast_build_copy(build
);
2643 stride_build
= isl_ast_build_detect_strides(stride_build
,
2644 isl_set_copy(domain
));
2645 expansion
= isl_ast_build_get_stride_expansion(stride_build
);
2647 domain
= isl_set_preimage_multi_aff(domain
,
2648 isl_multi_aff_copy(expansion
));
2649 domain
= isl_ast_build_eliminate_divs(stride_build
, domain
);
2650 isl_ast_build_free(stride_build
);
2652 bmap
= isl_basic_map_from_multi_aff(expansion
);
2654 empty
= isl_set_is_empty(domain
);
2660 lower
= find_unroll_lower_bound(build
, domain
, depth
, bmap
, &n
);
2664 if (n
>= 0 && init
&& init(n
, user
) < 0)
2666 for (i
= 0; i
< n
; ++i
) {
2668 isl_basic_set
*bset
;
2669 isl_constraint
*slice
;
2671 slice
= at_offset(depth
, lower
, i
);
2672 set
= isl_set_copy(domain
);
2673 set
= isl_set_add_constraint(set
, isl_constraint_copy(slice
));
2674 bset
= isl_set_unshifted_simple_hull(set
);
2675 bset
= isl_basic_set_add_constraint(bset
, slice
);
2676 bset
= isl_basic_set_apply(bset
, isl_basic_map_copy(bmap
));
2678 if (fn(bset
, user
) < 0)
2682 isl_aff_free(lower
);
2683 isl_set_free(domain
);
2684 isl_basic_map_free(bmap
);
2686 return n
< 0 || i
< n
? -1 : 0;
2689 /* Data structure for storing the results and the intermediate objects
2690 * of compute_domains.
2692 * "list" is the main result of the function and contains a list
2693 * of disjoint basic sets for which code should be generated.
2695 * "executed" and "build" are inputs to compute_domains.
2696 * "schedule_domain" is the domain of "executed".
2698 * "option" contains the domains at the current depth that should by
2699 * atomic, separated or unrolled. These domains are as specified by
2700 * the user, except that inner dimensions have been eliminated and
2701 * that they have been made pair-wise disjoint.
2703 * "sep_class" contains the user-specified split into separation classes
2704 * specialized to the current depth.
2705 * "done" contains the union of the separation domains that have already
2708 struct isl_codegen_domains
{
2709 isl_basic_set_list
*list
;
2711 isl_union_map
*executed
;
2712 isl_ast_build
*build
;
2713 isl_set
*schedule_domain
;
2721 /* Internal data structure for do_unroll.
2723 * "domains" stores the results of compute_domains.
2724 * "class_domain" is the original class domain passed to do_unroll.
2725 * "unroll_domain" collects the unrolled iterations.
2727 struct isl_ast_unroll_data
{
2728 struct isl_codegen_domains
*domains
;
2729 isl_set
*class_domain
;
2730 isl_set
*unroll_domain
;
2733 /* Given an iteration of an unrolled domain represented by "bset",
2734 * add it to data->domains->list.
2735 * Since we may have dropped some constraints, we intersect with
2736 * the class domain again to ensure that each element in the list
2737 * is disjoint from the other class domains.
2739 static int do_unroll_iteration(__isl_take isl_basic_set
*bset
, void *user
)
2741 struct isl_ast_unroll_data
*data
= user
;
2743 isl_basic_set_list
*list
;
2745 set
= isl_set_from_basic_set(bset
);
2746 data
->unroll_domain
= isl_set_union(data
->unroll_domain
,
2748 set
= isl_set_intersect(set
, isl_set_copy(data
->class_domain
));
2749 set
= isl_set_make_disjoint(set
);
2750 list
= isl_basic_set_list_from_set(set
);
2751 data
->domains
->list
= isl_basic_set_list_concat(data
->domains
->list
,
2757 /* Extend domains->list with a list of basic sets, one for each value
2758 * of the current dimension in "domain" and remove the corresponding
2759 * sets from the class domain. Return the updated class domain.
2760 * The divs that involve the current dimension have not been projected out
2763 * We call foreach_iteration to iterate over the individual values and
2764 * in do_unroll_iteration we collect the individual basic sets in
2765 * domains->list and their union in data->unroll_domain, which is then
2766 * used to update the class domain.
2768 static __isl_give isl_set
*do_unroll(struct isl_codegen_domains
*domains
,
2769 __isl_take isl_set
*domain
, __isl_take isl_set
*class_domain
)
2771 struct isl_ast_unroll_data data
;
2774 return isl_set_free(class_domain
);
2776 return isl_set_free(domain
);
2778 data
.domains
= domains
;
2779 data
.class_domain
= class_domain
;
2780 data
.unroll_domain
= isl_set_empty(isl_set_get_space(domain
));
2782 if (foreach_iteration(domain
, domains
->build
, NULL
,
2783 &do_unroll_iteration
, &data
) < 0)
2784 data
.unroll_domain
= isl_set_free(data
.unroll_domain
);
2786 class_domain
= isl_set_subtract(class_domain
, data
.unroll_domain
);
2788 return class_domain
;
2791 /* Add domains to domains->list for each individual value of the current
2792 * dimension, for that part of the schedule domain that lies in the
2793 * intersection of the option domain and the class domain.
2794 * Remove the corresponding sets from the class domain and
2795 * return the updated class domain.
2797 * We first break up the unroll option domain into individual pieces
2798 * and then handle each of them separately. The unroll option domain
2799 * has been made disjoint in compute_domains_init_options,
2801 * Note that we actively want to combine different pieces of the
2802 * schedule domain that have the same value at the current dimension.
2803 * We therefore need to break up the unroll option domain before
2804 * intersecting with class and schedule domain, hoping that the
2805 * unroll option domain specified by the user is relatively simple.
2807 static __isl_give isl_set
*compute_unroll_domains(
2808 struct isl_codegen_domains
*domains
, __isl_take isl_set
*class_domain
)
2810 isl_set
*unroll_domain
;
2811 isl_basic_set_list
*unroll_list
;
2816 empty
= isl_set_is_empty(domains
->option
[isl_ast_loop_unroll
]);
2818 return isl_set_free(class_domain
);
2820 return class_domain
;
2822 unroll_domain
= isl_set_copy(domains
->option
[isl_ast_loop_unroll
]);
2823 unroll_list
= isl_basic_set_list_from_set(unroll_domain
);
2825 n
= isl_basic_set_list_n_basic_set(unroll_list
);
2827 class_domain
= isl_set_free(class_domain
);
2828 for (i
= 0; i
< n
; ++i
) {
2829 isl_basic_set
*bset
;
2831 bset
= isl_basic_set_list_get_basic_set(unroll_list
, i
);
2832 unroll_domain
= isl_set_from_basic_set(bset
);
2833 unroll_domain
= isl_set_intersect(unroll_domain
,
2834 isl_set_copy(class_domain
));
2835 unroll_domain
= isl_set_intersect(unroll_domain
,
2836 isl_set_copy(domains
->schedule_domain
));
2838 empty
= isl_set_is_empty(unroll_domain
);
2839 if (empty
>= 0 && empty
) {
2840 isl_set_free(unroll_domain
);
2844 class_domain
= do_unroll(domains
, unroll_domain
, class_domain
);
2847 isl_basic_set_list_free(unroll_list
);
2849 return class_domain
;
2852 /* Try and construct a single basic set that includes the intersection of
2853 * the schedule domain, the atomic option domain and the class domain.
2854 * Add the resulting basic set(s) to domains->list and remove them
2855 * from class_domain. Return the updated class domain.
2857 * We construct a single domain rather than trying to combine
2858 * the schedule domains of individual domains because we are working
2859 * within a single component so that non-overlapping schedule domains
2860 * should already have been separated.
2861 * We do however need to make sure that this single domains is a subset
2862 * of the class domain so that it would not intersect with any other
2863 * class domains. This means that we may end up splitting up the atomic
2864 * domain in case separation classes are being used.
2866 * "domain" is the intersection of the schedule domain and the class domain,
2867 * with inner dimensions projected out.
2869 static __isl_give isl_set
*compute_atomic_domain(
2870 struct isl_codegen_domains
*domains
, __isl_take isl_set
*class_domain
)
2872 isl_basic_set
*bset
;
2873 isl_basic_set_list
*list
;
2874 isl_set
*domain
, *atomic_domain
;
2877 domain
= isl_set_copy(domains
->option
[isl_ast_loop_atomic
]);
2878 domain
= isl_set_intersect(domain
, isl_set_copy(class_domain
));
2879 domain
= isl_set_intersect(domain
,
2880 isl_set_copy(domains
->schedule_domain
));
2881 empty
= isl_set_is_empty(domain
);
2883 class_domain
= isl_set_free(class_domain
);
2885 isl_set_free(domain
);
2886 return class_domain
;
2889 domain
= isl_ast_build_eliminate(domains
->build
, domain
);
2890 domain
= isl_set_coalesce_preserve(domain
);
2891 bset
= isl_set_unshifted_simple_hull(domain
);
2892 domain
= isl_set_from_basic_set(bset
);
2893 atomic_domain
= isl_set_copy(domain
);
2894 domain
= isl_set_intersect(domain
, isl_set_copy(class_domain
));
2895 class_domain
= isl_set_subtract(class_domain
, atomic_domain
);
2896 domain
= isl_set_make_disjoint(domain
);
2897 list
= isl_basic_set_list_from_set(domain
);
2898 domains
->list
= isl_basic_set_list_concat(domains
->list
, list
);
2900 return class_domain
;
2903 /* Split up the schedule domain into uniform basic sets,
2904 * in the sense that each element in a basic set is associated to
2905 * elements of the same domains, and add the result to domains->list.
2906 * Do this for that part of the schedule domain that lies in the
2907 * intersection of "class_domain" and the separate option domain.
2909 * "class_domain" may or may not include the constraints
2910 * of the schedule domain, but this does not make a difference
2911 * since we are going to intersect it with the domain of the inverse schedule.
2912 * If it includes schedule domain constraints, then they may involve
2913 * inner dimensions, but we will eliminate them in separation_domain.
2915 static int compute_separate_domain(struct isl_codegen_domains
*domains
,
2916 __isl_keep isl_set
*class_domain
)
2920 isl_union_map
*executed
;
2921 isl_basic_set_list
*list
;
2924 domain
= isl_set_copy(domains
->option
[isl_ast_loop_separate
]);
2925 domain
= isl_set_intersect(domain
, isl_set_copy(class_domain
));
2926 executed
= isl_union_map_copy(domains
->executed
);
2927 executed
= isl_union_map_intersect_domain(executed
,
2928 isl_union_set_from_set(domain
));
2929 empty
= isl_union_map_is_empty(executed
);
2930 if (empty
< 0 || empty
) {
2931 isl_union_map_free(executed
);
2932 return empty
< 0 ? -1 : 0;
2935 space
= isl_set_get_space(class_domain
);
2936 domain
= separate_schedule_domains(space
, executed
, domains
->build
);
2938 list
= isl_basic_set_list_from_set(domain
);
2939 domains
->list
= isl_basic_set_list_concat(domains
->list
, list
);
2944 /* Split up the domain at the current depth into disjoint
2945 * basic sets for which code should be generated separately
2946 * for the given separation class domain.
2948 * If any separation classes have been defined, then "class_domain"
2949 * is the domain of the current class and does not refer to inner dimensions.
2950 * Otherwise, "class_domain" is the universe domain.
2952 * We first make sure that the class domain is disjoint from
2953 * previously considered class domains.
2955 * The separate domains can be computed directly from the "class_domain".
2957 * The unroll, atomic and remainder domains need the constraints
2958 * from the schedule domain.
2960 * For unrolling, the actual schedule domain is needed (with divs that
2961 * may refer to the current dimension) so that stride detection can be
2964 * For atomic and remainder domains, inner dimensions and divs involving
2965 * the current dimensions should be eliminated.
2966 * In case we are working within a separation class, we need to intersect
2967 * the result with the current "class_domain" to ensure that the domains
2968 * are disjoint from those generated from other class domains.
2970 * The domain that has been made atomic may be larger than specified
2971 * by the user since it needs to be representable as a single basic set.
2972 * This possibly larger domain is removed from class_domain by
2973 * compute_atomic_domain. It is computed first so that the extended domain
2974 * would not overlap with any domains computed before.
2975 * Similary, the unrolled domains may have some constraints removed and
2976 * may therefore also be larger than specified by the user.
2978 * If anything is left after handling separate, unroll and atomic,
2979 * we split it up into basic sets and append the basic sets to domains->list.
2981 static isl_stat
compute_partial_domains(struct isl_codegen_domains
*domains
,
2982 __isl_take isl_set
*class_domain
)
2984 isl_basic_set_list
*list
;
2987 class_domain
= isl_set_subtract(class_domain
,
2988 isl_set_copy(domains
->done
));
2989 domains
->done
= isl_set_union(domains
->done
,
2990 isl_set_copy(class_domain
));
2992 class_domain
= compute_atomic_domain(domains
, class_domain
);
2993 class_domain
= compute_unroll_domains(domains
, class_domain
);
2995 domain
= isl_set_copy(class_domain
);
2997 if (compute_separate_domain(domains
, domain
) < 0)
2999 domain
= isl_set_subtract(domain
,
3000 isl_set_copy(domains
->option
[isl_ast_loop_separate
]));
3002 domain
= isl_set_intersect(domain
,
3003 isl_set_copy(domains
->schedule_domain
));
3005 domain
= isl_ast_build_eliminate(domains
->build
, domain
);
3006 domain
= isl_set_intersect(domain
, isl_set_copy(class_domain
));
3008 domain
= isl_set_coalesce_preserve(domain
);
3009 domain
= isl_set_make_disjoint(domain
);
3011 list
= isl_basic_set_list_from_set(domain
);
3012 domains
->list
= isl_basic_set_list_concat(domains
->list
, list
);
3014 isl_set_free(class_domain
);
3018 isl_set_free(domain
);
3019 isl_set_free(class_domain
);
3020 return isl_stat_error
;
3023 /* Split up the domain at the current depth into disjoint
3024 * basic sets for which code should be generated separately
3025 * for the separation class identified by "pnt".
3027 * We extract the corresponding class domain from domains->sep_class,
3028 * eliminate inner dimensions and pass control to compute_partial_domains.
3030 static isl_stat
compute_class_domains(__isl_take isl_point
*pnt
, void *user
)
3032 struct isl_codegen_domains
*domains
= user
;
3037 class_set
= isl_set_from_point(pnt
);
3038 domain
= isl_map_domain(isl_map_intersect_range(
3039 isl_map_copy(domains
->sep_class
), class_set
));
3040 domain
= isl_ast_build_compute_gist(domains
->build
, domain
);
3041 domain
= isl_ast_build_eliminate(domains
->build
, domain
);
3043 disjoint
= isl_set_plain_is_disjoint(domain
, domains
->schedule_domain
);
3045 return isl_stat_error
;
3047 isl_set_free(domain
);
3051 return compute_partial_domains(domains
, domain
);
3054 /* Extract the domains at the current depth that should be atomic,
3055 * separated or unrolled and store them in option.
3057 * The domains specified by the user might overlap, so we make
3058 * them disjoint by subtracting earlier domains from later domains.
3060 static void compute_domains_init_options(isl_set
*option
[4],
3061 __isl_keep isl_ast_build
*build
)
3063 enum isl_ast_loop_type type
, type2
;
3066 for (type
= isl_ast_loop_atomic
;
3067 type
<= isl_ast_loop_separate
; ++type
) {
3068 option
[type
] = isl_ast_build_get_option_domain(build
, type
);
3069 for (type2
= isl_ast_loop_atomic
; type2
< type
; ++type2
)
3070 option
[type
] = isl_set_subtract(option
[type
],
3071 isl_set_copy(option
[type2
]));
3074 unroll
= option
[isl_ast_loop_unroll
];
3075 unroll
= isl_set_coalesce(unroll
);
3076 unroll
= isl_set_make_disjoint(unroll
);
3077 option
[isl_ast_loop_unroll
] = unroll
;
3080 /* Split up the domain at the current depth into disjoint
3081 * basic sets for which code should be generated separately,
3082 * based on the user-specified options.
3083 * Return the list of disjoint basic sets.
3085 * There are three kinds of domains that we need to keep track of.
3086 * - the "schedule domain" is the domain of "executed"
3087 * - the "class domain" is the domain corresponding to the currrent
3089 * - the "option domain" is the domain corresponding to one of the options
3090 * atomic, unroll or separate
3092 * We first consider the individial values of the separation classes
3093 * and split up the domain for each of them separately.
3094 * Finally, we consider the remainder. If no separation classes were
3095 * specified, then we call compute_partial_domains with the universe
3096 * "class_domain". Otherwise, we take the "schedule_domain" as "class_domain",
3097 * with inner dimensions removed. We do this because we want to
3098 * avoid computing the complement of the class domains (i.e., the difference
3099 * between the universe and domains->done).
3101 static __isl_give isl_basic_set_list
*compute_domains(
3102 __isl_keep isl_union_map
*executed
, __isl_keep isl_ast_build
*build
)
3104 struct isl_codegen_domains domains
;
3107 isl_union_set
*schedule_domain
;
3111 enum isl_ast_loop_type type
;
3117 ctx
= isl_union_map_get_ctx(executed
);
3118 domains
.list
= isl_basic_set_list_alloc(ctx
, 0);
3120 schedule_domain
= isl_union_map_domain(isl_union_map_copy(executed
));
3121 domain
= isl_set_from_union_set(schedule_domain
);
3123 compute_domains_init_options(domains
.option
, build
);
3125 domains
.sep_class
= isl_ast_build_get_separation_class(build
);
3126 classes
= isl_map_range(isl_map_copy(domains
.sep_class
));
3127 n_param
= isl_set_dim(classes
, isl_dim_param
);
3129 classes
= isl_set_free(classes
);
3130 classes
= isl_set_project_out(classes
, isl_dim_param
, 0, n_param
);
3132 space
= isl_set_get_space(domain
);
3133 domains
.build
= build
;
3134 domains
.schedule_domain
= isl_set_copy(domain
);
3135 domains
.executed
= executed
;
3136 domains
.done
= isl_set_empty(space
);
3138 if (isl_set_foreach_point(classes
, &compute_class_domains
, &domains
) < 0)
3139 domains
.list
= isl_basic_set_list_free(domains
.list
);
3140 isl_set_free(classes
);
3142 empty
= isl_set_is_empty(domains
.done
);
3144 domains
.list
= isl_basic_set_list_free(domains
.list
);
3145 domain
= isl_set_free(domain
);
3147 isl_set_free(domain
);
3148 domain
= isl_set_universe(isl_set_get_space(domains
.done
));
3150 domain
= isl_ast_build_eliminate(build
, domain
);
3152 if (compute_partial_domains(&domains
, domain
) < 0)
3153 domains
.list
= isl_basic_set_list_free(domains
.list
);
3155 isl_set_free(domains
.schedule_domain
);
3156 isl_set_free(domains
.done
);
3157 isl_map_free(domains
.sep_class
);
3158 for (type
= isl_ast_loop_atomic
; type
<= isl_ast_loop_separate
; ++type
)
3159 isl_set_free(domains
.option
[type
]);
3161 return domains
.list
;
3164 /* Generate code for a single component, after shifting (if any)
3165 * has been applied, in case the schedule was specified as a union map.
3167 * We first split up the domain at the current depth into disjoint
3168 * basic sets based on the user-specified options.
3169 * Then we generated code for each of them and concatenate the results.
3171 static __isl_give isl_ast_graft_list
*generate_shifted_component_flat(
3172 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
3174 isl_basic_set_list
*domain_list
;
3175 isl_ast_graft_list
*list
= NULL
;
3177 domain_list
= compute_domains(executed
, build
);
3178 list
= generate_parallel_domains(domain_list
, executed
, build
);
3180 isl_basic_set_list_free(domain_list
);
3181 isl_union_map_free(executed
);
3182 isl_ast_build_free(build
);
3187 /* Generate code for a single component, after shifting (if any)
3188 * has been applied, in case the schedule was specified as a schedule tree
3189 * and the separate option was specified.
3191 * We perform separation on the domain of "executed" and then generate
3192 * an AST for each of the resulting disjoint basic sets.
3194 static __isl_give isl_ast_graft_list
*generate_shifted_component_tree_separate(
3195 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
3199 isl_basic_set_list
*domain_list
;
3200 isl_ast_graft_list
*list
;
3202 space
= isl_ast_build_get_space(build
, 1);
3203 domain
= separate_schedule_domains(space
,
3204 isl_union_map_copy(executed
), build
);
3205 domain_list
= isl_basic_set_list_from_set(domain
);
3207 list
= generate_parallel_domains(domain_list
, executed
, build
);
3209 isl_basic_set_list_free(domain_list
);
3210 isl_union_map_free(executed
);
3211 isl_ast_build_free(build
);
3216 /* Internal data structure for generate_shifted_component_tree_unroll.
3218 * "executed" and "build" are inputs to generate_shifted_component_tree_unroll.
3219 * "list" collects the constructs grafts.
3221 struct isl_ast_unroll_tree_data
{
3222 isl_union_map
*executed
;
3223 isl_ast_build
*build
;
3224 isl_ast_graft_list
*list
;
3227 /* Initialize data->list to a list of "n" elements.
3229 static int init_unroll_tree(int n
, void *user
)
3231 struct isl_ast_unroll_tree_data
*data
= user
;
3234 ctx
= isl_ast_build_get_ctx(data
->build
);
3235 data
->list
= isl_ast_graft_list_alloc(ctx
, n
);
3240 /* Given an iteration of an unrolled domain represented by "bset",
3241 * generate the corresponding AST and add the result to data->list.
3243 static int do_unroll_tree_iteration(__isl_take isl_basic_set
*bset
, void *user
)
3245 struct isl_ast_unroll_tree_data
*data
= user
;
3247 data
->list
= add_node(data
->list
, isl_union_map_copy(data
->executed
),
3248 bset
, isl_ast_build_copy(data
->build
));
3253 /* Generate code for a single component, after shifting (if any)
3254 * has been applied, in case the schedule was specified as a schedule tree
3255 * and the unroll option was specified.
3257 * We call foreach_iteration to iterate over the individual values and
3258 * construct and collect the corresponding grafts in do_unroll_tree_iteration.
3260 static __isl_give isl_ast_graft_list
*generate_shifted_component_tree_unroll(
3261 __isl_take isl_union_map
*executed
, __isl_take isl_set
*domain
,
3262 __isl_take isl_ast_build
*build
)
3264 struct isl_ast_unroll_tree_data data
= { executed
, build
, NULL
};
3266 if (foreach_iteration(domain
, build
, &init_unroll_tree
,
3267 &do_unroll_tree_iteration
, &data
) < 0)
3268 data
.list
= isl_ast_graft_list_free(data
.list
);
3270 isl_union_map_free(executed
);
3271 isl_ast_build_free(build
);
3276 /* Does "domain" involve a disjunction that is purely based on
3277 * constraints involving only outer dimension?
3279 * In particular, is there a disjunction such that the constraints
3280 * involving the current and later dimensions are the same over
3281 * all the disjuncts?
3283 static isl_bool
has_pure_outer_disjunction(__isl_keep isl_set
*domain
,
3284 __isl_keep isl_ast_build
*build
)
3286 isl_basic_set
*hull
;
3287 isl_set
*shared
, *inner
;
3293 n
= isl_set_n_basic_set(domain
);
3295 return isl_bool_error
;
3297 return isl_bool_false
;
3298 dim
= isl_set_dim(domain
, isl_dim_set
);
3300 return isl_bool_error
;
3302 inner
= isl_set_copy(domain
);
3303 depth
= isl_ast_build_get_depth(build
);
3304 inner
= isl_set_drop_constraints_not_involving_dims(inner
,
3305 isl_dim_set
, depth
, dim
- depth
);
3306 hull
= isl_set_plain_unshifted_simple_hull(isl_set_copy(inner
));
3307 shared
= isl_set_from_basic_set(hull
);
3308 equal
= isl_set_plain_is_equal(inner
, shared
);
3309 isl_set_free(inner
);
3310 isl_set_free(shared
);
3315 /* Generate code for a single component, after shifting (if any)
3316 * has been applied, in case the schedule was specified as a schedule tree.
3317 * In particular, handle the base case where there is either no isolated
3318 * set or we are within the isolated set (in which case "isolated" is set)
3319 * or the iterations that precede or follow the isolated set.
3321 * The schedule domain is broken up or combined into basic sets
3322 * according to the AST generation option specified in the current
3323 * schedule node, which may be either atomic, separate, unroll or
3324 * unspecified. If the option is unspecified, then we currently simply
3325 * split the schedule domain into disjoint basic sets.
3327 * In case the separate option is specified, the AST generation is
3328 * handled by generate_shifted_component_tree_separate.
3329 * In the other cases, we need the global schedule domain.
3330 * In the unroll case, the AST generation is then handled by
3331 * generate_shifted_component_tree_unroll which needs the actual
3332 * schedule domain (with divs that may refer to the current dimension)
3333 * so that stride detection can be performed.
3334 * In the atomic or unspecified case, inner dimensions and divs involving
3335 * the current dimensions should be eliminated.
3336 * The result is then either combined into a single basic set or
3337 * split up into disjoint basic sets.
3338 * Finally an AST is generated for each basic set and the results are
3341 * If the schedule domain involves a disjunction that is purely based on
3342 * constraints involving only outer dimension, then it is treated as
3343 * if atomic was specified. This ensures that only a single loop
3344 * is generated instead of a sequence of identical loops with
3347 static __isl_give isl_ast_graft_list
*generate_shifted_component_tree_base(
3348 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
,
3351 isl_bool outer_disjunction
;
3352 isl_union_set
*schedule_domain
;
3354 isl_basic_set_list
*domain_list
;
3355 isl_ast_graft_list
*list
;
3356 enum isl_ast_loop_type type
;
3358 type
= isl_ast_build_get_loop_type(build
, isolated
);
3362 if (type
== isl_ast_loop_separate
)
3363 return generate_shifted_component_tree_separate(executed
,
3366 schedule_domain
= isl_union_map_domain(isl_union_map_copy(executed
));
3367 domain
= isl_set_from_union_set(schedule_domain
);
3369 if (type
== isl_ast_loop_unroll
)
3370 return generate_shifted_component_tree_unroll(executed
, domain
,
3373 domain
= isl_ast_build_eliminate(build
, domain
);
3374 domain
= isl_set_coalesce_preserve(domain
);
3376 outer_disjunction
= has_pure_outer_disjunction(domain
, build
);
3377 if (outer_disjunction
< 0)
3378 domain
= isl_set_free(domain
);
3380 if (outer_disjunction
|| type
== isl_ast_loop_atomic
) {
3381 isl_basic_set
*hull
;
3382 hull
= isl_set_unshifted_simple_hull(domain
);
3383 domain_list
= isl_basic_set_list_from_basic_set(hull
);
3385 domain
= isl_set_make_disjoint(domain
);
3386 domain_list
= isl_basic_set_list_from_set(domain
);
3389 list
= generate_parallel_domains(domain_list
, executed
, build
);
3391 isl_basic_set_list_free(domain_list
);
3392 isl_union_map_free(executed
);
3393 isl_ast_build_free(build
);
3397 isl_union_map_free(executed
);
3398 isl_ast_build_free(build
);
3402 /* Extract out the disjunction imposed by "domain" on the outer
3403 * schedule dimensions.
3405 * In particular, remove all inner dimensions from "domain" (including
3406 * the current dimension) and then remove the constraints that are shared
3407 * by all disjuncts in the result.
3409 static __isl_give isl_set
*extract_disjunction(__isl_take isl_set
*domain
,
3410 __isl_keep isl_ast_build
*build
)
3416 domain
= isl_ast_build_specialize(build
, domain
);
3417 depth
= isl_ast_build_get_depth(build
);
3418 dim
= isl_set_dim(domain
, isl_dim_set
);
3420 return isl_set_free(domain
);
3421 domain
= isl_set_eliminate(domain
, isl_dim_set
, depth
, dim
- depth
);
3422 domain
= isl_set_remove_unknown_divs(domain
);
3423 hull
= isl_set_copy(domain
);
3424 hull
= isl_set_from_basic_set(isl_set_unshifted_simple_hull(hull
));
3425 domain
= isl_set_gist(domain
, hull
);
3430 /* Add "guard" to the grafts in "list".
3431 * "build" is the outer AST build, while "sub_build" includes "guard"
3432 * in its generated domain.
3434 * First combine the grafts into a single graft and then add the guard.
3435 * If the list is empty, or if some error occurred, then simply return
3438 static __isl_give isl_ast_graft_list
*list_add_guard(
3439 __isl_take isl_ast_graft_list
*list
, __isl_keep isl_set
*guard
,
3440 __isl_keep isl_ast_build
*build
, __isl_keep isl_ast_build
*sub_build
)
3442 isl_ast_graft
*graft
;
3445 list
= isl_ast_graft_list_fuse(list
, sub_build
);
3447 n
= isl_ast_graft_list_n_ast_graft(list
);
3449 return isl_ast_graft_list_free(list
);
3453 graft
= isl_ast_graft_list_get_ast_graft(list
, 0);
3454 graft
= isl_ast_graft_add_guard(graft
, isl_set_copy(guard
), build
);
3455 list
= isl_ast_graft_list_set_ast_graft(list
, 0, graft
);
3460 /* Generate code for a single component, after shifting (if any)
3461 * has been applied, in case the schedule was specified as a schedule tree.
3462 * In particular, do so for the specified subset of the schedule domain.
3464 * If we are outside of the isolated part, then "domain" may include
3465 * a disjunction. Explicitly generate this disjunction at this point
3466 * instead of relying on the disjunction getting hoisted back up
3469 static __isl_give isl_ast_graft_list
*generate_shifted_component_tree_part(
3470 __isl_keep isl_union_map
*executed
, __isl_take isl_set
*domain
,
3471 __isl_keep isl_ast_build
*build
, int isolated
)
3473 isl_union_set
*uset
;
3474 isl_ast_graft_list
*list
;
3475 isl_ast_build
*sub_build
;
3478 uset
= isl_union_set_from_set(isl_set_copy(domain
));
3479 executed
= isl_union_map_copy(executed
);
3480 executed
= isl_union_map_intersect_domain(executed
, uset
);
3481 empty
= isl_union_map_is_empty(executed
);
3486 isl_union_map_free(executed
);
3487 isl_set_free(domain
);
3488 ctx
= isl_ast_build_get_ctx(build
);
3489 return isl_ast_graft_list_alloc(ctx
, 0);
3492 sub_build
= isl_ast_build_copy(build
);
3494 domain
= extract_disjunction(domain
, build
);
3495 sub_build
= isl_ast_build_restrict_generated(sub_build
,
3496 isl_set_copy(domain
));
3498 list
= generate_shifted_component_tree_base(executed
,
3499 isl_ast_build_copy(sub_build
), isolated
);
3501 list
= list_add_guard(list
, domain
, build
, sub_build
);
3502 isl_ast_build_free(sub_build
);
3503 isl_set_free(domain
);
3506 isl_union_map_free(executed
);
3507 isl_set_free(domain
);
3511 /* Generate code for a single component, after shifting (if any)
3512 * has been applied, in case the schedule was specified as a schedule tree.
3513 * In particular, do so for the specified sequence of subsets
3514 * of the schedule domain, "before", "isolated", "after" and "other",
3515 * where only the "isolated" part is considered to be isolated.
3517 static __isl_give isl_ast_graft_list
*generate_shifted_component_parts(
3518 __isl_take isl_union_map
*executed
, __isl_take isl_set
*before
,
3519 __isl_take isl_set
*isolated
, __isl_take isl_set
*after
,
3520 __isl_take isl_set
*other
, __isl_take isl_ast_build
*build
)
3522 isl_ast_graft_list
*list
, *res
;
3524 res
= generate_shifted_component_tree_part(executed
, before
, build
, 0);
3525 list
= generate_shifted_component_tree_part(executed
, isolated
,
3527 res
= isl_ast_graft_list_concat(res
, list
);
3528 list
= generate_shifted_component_tree_part(executed
, after
, build
, 0);
3529 res
= isl_ast_graft_list_concat(res
, list
);
3530 list
= generate_shifted_component_tree_part(executed
, other
, build
, 0);
3531 res
= isl_ast_graft_list_concat(res
, list
);
3533 isl_union_map_free(executed
);
3534 isl_ast_build_free(build
);
3539 /* Does "set" intersect "first", but not "second"?
3541 static isl_bool
only_intersects_first(__isl_keep isl_set
*set
,
3542 __isl_keep isl_set
*first
, __isl_keep isl_set
*second
)
3546 disjoint
= isl_set_is_disjoint(set
, first
);
3548 return isl_bool_error
;
3550 return isl_bool_false
;
3552 return isl_set_is_disjoint(set
, second
);
3555 /* Generate code for a single component, after shifting (if any)
3556 * has been applied, in case the schedule was specified as a schedule tree.
3557 * In particular, do so in case of isolation where there is
3558 * only an "isolated" part and an "after" part.
3559 * "dead1" and "dead2" are freed by this function in order to simplify
3562 * The "before" and "other" parts are set to empty sets.
3564 static __isl_give isl_ast_graft_list
*generate_shifted_component_only_after(
3565 __isl_take isl_union_map
*executed
, __isl_take isl_set
*isolated
,
3566 __isl_take isl_set
*after
, __isl_take isl_ast_build
*build
,
3567 __isl_take isl_set
*dead1
, __isl_take isl_set
*dead2
)
3571 empty
= isl_set_empty(isl_set_get_space(after
));
3572 isl_set_free(dead1
);
3573 isl_set_free(dead2
);
3574 return generate_shifted_component_parts(executed
, isl_set_copy(empty
),
3575 isolated
, after
, empty
, build
);
3578 /* Generate code for a single component, after shifting (if any)
3579 * has been applied, in case the schedule was specified as a schedule tree.
3581 * We first check if the user has specified an isolated schedule domain
3582 * and that we are not already outside of this isolated schedule domain.
3583 * If so, we break up the schedule domain into iterations that
3584 * precede the isolated domain, the isolated domain itself,
3585 * the iterations that follow the isolated domain and
3586 * the remaining iterations (those that are incomparable
3587 * to the isolated domain).
3588 * We generate an AST for each piece and concatenate the results.
3590 * If the isolated domain is not convex, then it is replaced
3591 * by a convex superset to ensure that the sets of preceding and
3592 * following iterations are properly defined and, in particular,
3593 * that there are no intermediate iterations that do not belong
3594 * to the isolated domain.
3596 * In the special case where at least one element of the schedule
3597 * domain that does not belong to the isolated domain needs
3598 * to be scheduled after this isolated domain, but none of those
3599 * elements need to be scheduled before, break up the schedule domain
3600 * in only two parts, the isolated domain, and a part that will be
3601 * scheduled after the isolated domain.
3603 * If no isolated set has been specified, then we generate an
3604 * AST for the entire inverse schedule.
3606 static __isl_give isl_ast_graft_list
*generate_shifted_component_tree(
3607 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
3610 int empty
, has_isolate
;
3612 isl_union_set
*schedule_domain
;
3614 isl_basic_set
*hull
;
3615 isl_set
*isolated
, *before
, *after
, *test
;
3619 build
= isl_ast_build_extract_isolated(build
);
3620 has_isolate
= isl_ast_build_has_isolated(build
);
3621 if (has_isolate
< 0)
3622 executed
= isl_union_map_free(executed
);
3623 else if (!has_isolate
)
3624 return generate_shifted_component_tree_base(executed
, build
, 0);
3626 schedule_domain
= isl_union_map_domain(isl_union_map_copy(executed
));
3627 domain
= isl_set_from_union_set(schedule_domain
);
3629 isolated
= isl_ast_build_get_isolated(build
);
3630 isolated
= isl_set_intersect(isolated
, isl_set_copy(domain
));
3631 test
= isl_ast_build_specialize(build
, isl_set_copy(isolated
));
3632 empty
= isl_set_is_empty(test
);
3637 isl_set_free(isolated
);
3638 isl_set_free(domain
);
3639 return generate_shifted_component_tree_base(executed
, build
, 0);
3641 isolated
= isl_ast_build_eliminate(build
, isolated
);
3642 hull
= isl_set_unshifted_simple_hull(isolated
);
3643 isolated
= isl_set_from_basic_set(hull
);
3645 depth
= isl_ast_build_get_depth(build
);
3646 space
= isl_space_map_from_set(isl_set_get_space(isolated
));
3647 gt
= isl_map_universe(space
);
3648 for (i
= 0; i
< depth
; ++i
)
3649 gt
= isl_map_equate(gt
, isl_dim_in
, i
, isl_dim_out
, i
);
3650 gt
= isl_map_order_gt(gt
, isl_dim_in
, depth
, isl_dim_out
, depth
);
3651 lt
= isl_map_reverse(isl_map_copy(gt
));
3652 before
= isl_set_apply(isl_set_copy(isolated
), gt
);
3653 after
= isl_set_apply(isl_set_copy(isolated
), lt
);
3655 domain
= isl_set_subtract(domain
, isl_set_copy(isolated
));
3656 pure
= only_intersects_first(domain
, after
, before
);
3658 executed
= isl_union_map_free(executed
);
3660 return generate_shifted_component_only_after(executed
, isolated
,
3661 domain
, build
, before
, after
);
3662 domain
= isl_set_subtract(domain
, isl_set_copy(before
));
3663 domain
= isl_set_subtract(domain
, isl_set_copy(after
));
3664 after
= isl_set_subtract(after
, isl_set_copy(isolated
));
3665 after
= isl_set_subtract(after
, isl_set_copy(before
));
3666 before
= isl_set_subtract(before
, isl_set_copy(isolated
));
3668 return generate_shifted_component_parts(executed
, before
, isolated
,
3669 after
, domain
, build
);
3671 isl_set_free(domain
);
3672 isl_set_free(isolated
);
3673 isl_union_map_free(executed
);
3674 isl_ast_build_free(build
);
3678 /* Generate code for a single component, after shifting (if any)
3681 * Call generate_shifted_component_tree or generate_shifted_component_flat
3682 * depending on whether the schedule was specified as a schedule tree.
3684 static __isl_give isl_ast_graft_list
*generate_shifted_component(
3685 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
3687 if (isl_ast_build_has_schedule_node(build
))
3688 return generate_shifted_component_tree(executed
, build
);
3690 return generate_shifted_component_flat(executed
, build
);
3693 struct isl_set_map_pair
{
3698 /* Given an array "domain" of isl_set_map_pairs and an array "order"
3699 * of indices into the "domain" array,
3700 * return the union of the "map" fields of the elements
3701 * indexed by the first "n" elements of "order".
3703 static __isl_give isl_union_map
*construct_component_executed(
3704 struct isl_set_map_pair
*domain
, int *order
, int n
)
3708 isl_union_map
*executed
;
3710 map
= isl_map_copy(domain
[order
[0]].map
);
3711 executed
= isl_union_map_from_map(map
);
3712 for (i
= 1; i
< n
; ++i
) {
3713 map
= isl_map_copy(domain
[order
[i
]].map
);
3714 executed
= isl_union_map_add_map(executed
, map
);
3720 /* Generate code for a single component, after shifting (if any)
3723 * The component inverse schedule is specified as the "map" fields
3724 * of the elements of "domain" indexed by the first "n" elements of "order".
3726 static __isl_give isl_ast_graft_list
*generate_shifted_component_from_list(
3727 struct isl_set_map_pair
*domain
, int *order
, int n
,
3728 __isl_take isl_ast_build
*build
)
3730 isl_union_map
*executed
;
3732 executed
= construct_component_executed(domain
, order
, n
);
3733 return generate_shifted_component(executed
, build
);
3736 /* Does set dimension "pos" of "set" have an obviously fixed value?
3738 static int dim_is_fixed(__isl_keep isl_set
*set
, int pos
)
3743 v
= isl_set_plain_get_val_if_fixed(set
, isl_dim_set
, pos
);
3746 fixed
= !isl_val_is_nan(v
);
3752 /* Given an array "domain" of isl_set_map_pairs and an array "order"
3753 * of indices into the "domain" array,
3754 * do all (except for at most one) of the "set" field of the elements
3755 * indexed by the first "n" elements of "order" have a fixed value
3756 * at position "depth"?
3758 static int at_most_one_non_fixed(struct isl_set_map_pair
*domain
,
3759 int *order
, int n
, int depth
)
3764 for (i
= 0; i
< n
; ++i
) {
3767 f
= dim_is_fixed(domain
[order
[i
]].set
, depth
);
3780 /* Given an array "domain" of isl_set_map_pairs and an array "order"
3781 * of indices into the "domain" array,
3782 * eliminate the inner dimensions from the "set" field of the elements
3783 * indexed by the first "n" elements of "order", provided the current
3784 * dimension does not have a fixed value.
3786 * Return the index of the first element in "order" with a corresponding
3787 * "set" field that does not have an (obviously) fixed value.
3789 static int eliminate_non_fixed(struct isl_set_map_pair
*domain
,
3790 int *order
, int n
, int depth
, __isl_keep isl_ast_build
*build
)
3795 for (i
= n
- 1; i
>= 0; --i
) {
3797 f
= dim_is_fixed(domain
[order
[i
]].set
, depth
);
3802 domain
[order
[i
]].set
= isl_ast_build_eliminate_inner(build
,
3803 domain
[order
[i
]].set
);
3810 /* Given an array "domain" of isl_set_map_pairs and an array "order"
3811 * of indices into the "domain" array,
3812 * find the element of "domain" (amongst those indexed by the first "n"
3813 * elements of "order") with the "set" field that has the smallest
3814 * value for the current iterator.
3816 * Note that the domain with the smallest value may depend on the parameters
3817 * and/or outer loop dimension. Since the result of this function is only
3818 * used as heuristic, we only make a reasonable attempt at finding the best
3819 * domain, one that should work in case a single domain provides the smallest
3820 * value for the current dimension over all values of the parameters
3821 * and outer dimensions.
3823 * In particular, we compute the smallest value of the first domain
3824 * and replace it by that of any later domain if that later domain
3825 * has a smallest value that is smaller for at least some value
3826 * of the parameters and outer dimensions.
3828 static int first_offset(struct isl_set_map_pair
*domain
, int *order
, int n
,
3829 __isl_keep isl_ast_build
*build
)
3835 min_first
= isl_ast_build_map_to_iterator(build
,
3836 isl_set_copy(domain
[order
[0]].set
));
3837 min_first
= isl_map_lexmin(min_first
);
3839 for (i
= 1; i
< n
; ++i
) {
3840 isl_map
*min
, *test
;
3843 min
= isl_ast_build_map_to_iterator(build
,
3844 isl_set_copy(domain
[order
[i
]].set
));
3845 min
= isl_map_lexmin(min
);
3846 test
= isl_map_copy(min
);
3847 test
= isl_map_apply_domain(isl_map_copy(min_first
), test
);
3848 test
= isl_map_order_lt(test
, isl_dim_in
, 0, isl_dim_out
, 0);
3849 empty
= isl_map_is_empty(test
);
3851 if (empty
>= 0 && !empty
) {
3852 isl_map_free(min_first
);
3862 isl_map_free(min_first
);
3864 return i
< n
? -1 : first
;
3867 /* Construct a shifted inverse schedule based on the original inverse schedule,
3868 * the stride and the offset.
3870 * The original inverse schedule is specified as the "map" fields
3871 * of the elements of "domain" indexed by the first "n" elements of "order".
3873 * "stride" and "offset" are such that the difference
3874 * between the values of the current dimension of domain "i"
3875 * and the values of the current dimension for some reference domain are
3878 * stride * integer + offset[i]
3880 * Moreover, 0 <= offset[i] < stride.
3882 * For each domain, we create a map
3884 * { [..., j, ...] -> [..., j - offset[i], offset[i], ....] }
3886 * where j refers to the current dimension and the other dimensions are
3887 * unchanged, and apply this map to the original schedule domain.
3889 * For example, for the original schedule
3891 * { A[i] -> [2i]: 0 <= i < 10; B[i] -> [2i+1] : 0 <= i < 10 }
3893 * and assuming the offset is 0 for the A domain and 1 for the B domain,
3894 * we apply the mapping
3898 * to the schedule of the "A" domain and the mapping
3900 * { [j - 1] -> [j, 1] }
3902 * to the schedule of the "B" domain.
3905 * Note that after the transformation, the differences between pairs
3906 * of values of the current dimension over all domains are multiples
3907 * of stride and that we have therefore exposed the stride.
3910 * To see that the mapping preserves the lexicographic order,
3911 * first note that each of the individual maps above preserves the order.
3912 * If the value of the current iterator is j1 in one domain and j2 in another,
3913 * then if j1 = j2, we know that the same map is applied to both domains
3914 * and the order is preserved.
3915 * Otherwise, let us assume, without loss of generality, that j1 < j2.
3916 * If c1 >= c2 (with c1 and c2 the corresponding offsets), then
3920 * and the order is preserved.
3921 * If c1 < c2, then we know
3927 * j2 - j1 = n * s + r
3929 * with n >= 0 and 0 <= r < s.
3930 * In other words, r = c2 - c1.
3941 * (j1 - c1, c1) << (j2 - c2, c2)
3943 * with "<<" the lexicographic order, proving that the order is preserved
3946 static __isl_give isl_union_map
*construct_shifted_executed(
3947 struct isl_set_map_pair
*domain
, int *order
, int n
,
3948 __isl_keep isl_val
*stride
, __isl_keep isl_multi_val
*offset
,
3949 __isl_take isl_ast_build
*build
)
3952 isl_union_map
*executed
;
3958 depth
= isl_ast_build_get_depth(build
);
3959 space
= isl_ast_build_get_space(build
, 1);
3960 executed
= isl_union_map_empty(isl_space_copy(space
));
3961 space
= isl_space_map_from_set(space
);
3962 map
= isl_map_identity(isl_space_copy(space
));
3963 map
= isl_map_eliminate(map
, isl_dim_out
, depth
, 1);
3964 map
= isl_map_insert_dims(map
, isl_dim_out
, depth
+ 1, 1);
3965 space
= isl_space_insert_dims(space
, isl_dim_out
, depth
+ 1, 1);
3967 c
= isl_constraint_alloc_equality(isl_local_space_from_space(space
));
3968 c
= isl_constraint_set_coefficient_si(c
, isl_dim_in
, depth
, 1);
3969 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
, depth
, -1);
3971 for (i
= 0; i
< n
; ++i
) {
3975 v
= isl_multi_val_get_val(offset
, i
);
3978 map_i
= isl_map_copy(map
);
3979 map_i
= isl_map_fix_val(map_i
, isl_dim_out
, depth
+ 1,
3982 c
= isl_constraint_set_constant_val(c
, v
);
3983 map_i
= isl_map_add_constraint(map_i
, isl_constraint_copy(c
));
3985 map_i
= isl_map_apply_domain(isl_map_copy(domain
[order
[i
]].map
),
3987 executed
= isl_union_map_add_map(executed
, map_i
);
3990 isl_constraint_free(c
);
3994 executed
= isl_union_map_free(executed
);
3999 /* Generate code for a single component, after exposing the stride,
4000 * given that the schedule domain is "shifted strided".
4002 * The component inverse schedule is specified as the "map" fields
4003 * of the elements of "domain" indexed by the first "n" elements of "order".
4005 * The schedule domain being "shifted strided" means that the differences
4006 * between the values of the current dimension of domain "i"
4007 * and the values of the current dimension for some reference domain are
4010 * stride * integer + offset[i]
4012 * We first look for the domain with the "smallest" value for the current
4013 * dimension and adjust the offsets such that the offset of the "smallest"
4014 * domain is equal to zero. The other offsets are reduced modulo stride.
4016 * Based on this information, we construct a new inverse schedule in
4017 * construct_shifted_executed that exposes the stride.
4018 * Since this involves the introduction of a new schedule dimension,
4019 * the build needs to be changed accordingly.
4020 * After computing the AST, the newly introduced dimension needs
4021 * to be removed again from the list of grafts. We do this by plugging
4022 * in a mapping that represents the new schedule domain in terms of the
4023 * old schedule domain.
4025 static __isl_give isl_ast_graft_list
*generate_shift_component(
4026 struct isl_set_map_pair
*domain
, int *order
, int n
,
4027 __isl_keep isl_val
*stride
, __isl_keep isl_multi_val
*offset
,
4028 __isl_take isl_ast_build
*build
)
4030 isl_ast_graft_list
*list
;
4036 isl_multi_aff
*ma
, *zero
;
4037 isl_union_map
*executed
;
4039 depth
= isl_ast_build_get_depth(build
);
4041 first
= first_offset(domain
, order
, n
, build
);
4045 mv
= isl_multi_val_copy(offset
);
4046 val
= isl_multi_val_get_val(offset
, first
);
4047 val
= isl_val_neg(val
);
4048 mv
= isl_multi_val_add_val(mv
, val
);
4049 mv
= isl_multi_val_mod_val(mv
, isl_val_copy(stride
));
4051 executed
= construct_shifted_executed(domain
, order
, n
, stride
, mv
,
4053 space
= isl_ast_build_get_space(build
, 1);
4054 space
= isl_space_map_from_set(space
);
4055 ma
= isl_multi_aff_identity(isl_space_copy(space
));
4056 space
= isl_space_from_domain(isl_space_domain(space
));
4057 space
= isl_space_add_dims(space
, isl_dim_out
, 1);
4058 zero
= isl_multi_aff_zero(space
);
4059 ma
= isl_multi_aff_range_splice(ma
, depth
+ 1, zero
);
4060 build
= isl_ast_build_insert_dim(build
, depth
+ 1);
4061 list
= generate_shifted_component(executed
, build
);
4063 list
= isl_ast_graft_list_preimage_multi_aff(list
, ma
);
4065 isl_multi_val_free(mv
);
4069 isl_ast_build_free(build
);
4073 /* Does any node in the schedule tree rooted at the current schedule node
4074 * of "build" depend on outer schedule nodes?
4076 static int has_anchored_subtree(__isl_keep isl_ast_build
*build
)
4078 isl_schedule_node
*node
;
4081 node
= isl_ast_build_get_schedule_node(build
);
4082 dependent
= isl_schedule_node_is_subtree_anchored(node
);
4083 isl_schedule_node_free(node
);
4088 /* Generate code for a single component.
4090 * The component inverse schedule is specified as the "map" fields
4091 * of the elements of "domain" indexed by the first "n" elements of "order".
4093 * This function may modify the "set" fields of "domain".
4095 * Before proceeding with the actual code generation for the component,
4096 * we first check if there are any "shifted" strides, meaning that
4097 * the schedule domains of the individual domains are all strided,
4098 * but that they have different offsets, resulting in the union
4099 * of schedule domains not being strided anymore.
4101 * The simplest example is the schedule
4103 * { A[i] -> [2i]: 0 <= i < 10; B[i] -> [2i+1] : 0 <= i < 10 }
4105 * Both schedule domains are strided, but their union is not.
4106 * This function detects such cases and then rewrites the schedule to
4108 * { A[i] -> [2i, 0]: 0 <= i < 10; B[i] -> [2i, 1] : 0 <= i < 10 }
4110 * In the new schedule, the schedule domains have the same offset (modulo
4111 * the stride), ensuring that the union of schedule domains is also strided.
4114 * If there is only a single domain in the component, then there is
4115 * nothing to do. Similarly, if the current schedule dimension has
4116 * a fixed value for almost all domains then there is nothing to be done.
4117 * In particular, we need at least two domains where the current schedule
4118 * dimension does not have a fixed value.
4119 * Finally, in case of a schedule map input,
4120 * if any of the options refer to the current schedule dimension,
4121 * then we bail out as well. It would be possible to reformulate the options
4122 * in terms of the new schedule domain, but that would introduce constraints
4123 * that separate the domains in the options and that is something we would
4125 * In the case of a schedule tree input, we bail out if any of
4126 * the descendants of the current schedule node refer to outer
4127 * schedule nodes in any way.
4130 * To see if there is any shifted stride, we look at the differences
4131 * between the values of the current dimension in pairs of domains
4132 * for equal values of outer dimensions. These differences should be
4137 * with "m" the stride and "r" a constant. Note that we cannot perform
4138 * this analysis on individual domains as the lower bound in each domain
4139 * may depend on parameters or outer dimensions and so the current dimension
4140 * itself may not have a fixed remainder on division by the stride.
4142 * In particular, we compare the first domain that does not have an
4143 * obviously fixed value for the current dimension to itself and all
4144 * other domains and collect the offsets and the gcd of the strides.
4145 * If the gcd becomes one, then we failed to find shifted strides.
4146 * If the gcd is zero, then the differences were all fixed, meaning
4147 * that some domains had non-obviously fixed values for the current dimension.
4148 * If all the offsets are the same (for those domains that do not have
4149 * an obviously fixed value for the current dimension), then we do not
4150 * apply the transformation.
4151 * If none of the domains were skipped, then there is nothing to do.
4152 * If some of them were skipped, then if we apply separation, the schedule
4153 * domain should get split in pieces with a (non-shifted) stride.
4155 * Otherwise, we apply a shift to expose the stride in
4156 * generate_shift_component.
4158 static __isl_give isl_ast_graft_list
*generate_component(
4159 struct isl_set_map_pair
*domain
, int *order
, int n
,
4160 __isl_take isl_ast_build
*build
)
4167 isl_val
*gcd
= NULL
;
4171 isl_ast_graft_list
*list
;
4174 depth
= isl_ast_build_get_depth(build
);
4177 if (skip
>= 0 && !skip
)
4178 skip
= at_most_one_non_fixed(domain
, order
, n
, depth
);
4179 if (skip
>= 0 && !skip
) {
4180 if (isl_ast_build_has_schedule_node(build
))
4181 skip
= has_anchored_subtree(build
);
4183 skip
= isl_ast_build_options_involve_depth(build
);
4188 return generate_shifted_component_from_list(domain
,
4191 base
= eliminate_non_fixed(domain
, order
, n
, depth
, build
);
4195 ctx
= isl_ast_build_get_ctx(build
);
4197 mv
= isl_multi_val_zero(isl_space_set_alloc(ctx
, 0, n
));
4200 for (i
= 0; i
< n
; ++i
) {
4203 map
= isl_map_from_domain_and_range(
4204 isl_set_copy(domain
[order
[base
]].set
),
4205 isl_set_copy(domain
[order
[i
]].set
));
4206 for (d
= 0; d
< depth
; ++d
)
4207 map
= isl_map_equate(map
, isl_dim_in
, d
,
4209 deltas
= isl_map_deltas(map
);
4210 res
= isl_set_dim_residue_class_val(deltas
, depth
, &m
, &r
);
4211 isl_set_free(deltas
);
4218 gcd
= isl_val_gcd(gcd
, m
);
4219 if (isl_val_is_one(gcd
)) {
4223 mv
= isl_multi_val_set_val(mv
, i
, r
);
4225 res
= dim_is_fixed(domain
[order
[i
]].set
, depth
);
4231 if (fixed
&& i
> base
) {
4233 a
= isl_multi_val_get_val(mv
, i
);
4234 b
= isl_multi_val_get_val(mv
, base
);
4235 if (isl_val_ne(a
, b
))
4242 if (res
< 0 || !gcd
) {
4243 isl_ast_build_free(build
);
4245 } else if (i
< n
|| fixed
|| isl_val_is_zero(gcd
)) {
4246 list
= generate_shifted_component_from_list(domain
,
4249 list
= generate_shift_component(domain
, order
, n
, gcd
, mv
,
4254 isl_multi_val_free(mv
);
4258 isl_ast_build_free(build
);
4262 /* Store both "map" itself and its domain in the
4263 * structure pointed to by *next and advance to the next array element.
4265 static isl_stat
extract_domain(__isl_take isl_map
*map
, void *user
)
4267 struct isl_set_map_pair
**next
= user
;
4269 (*next
)->map
= isl_map_copy(map
);
4270 (*next
)->set
= isl_map_domain(map
);
4276 static isl_bool
after_in_tree(__isl_keep isl_union_map
*umap
,
4277 __isl_keep isl_schedule_node
*node
);
4279 /* Is any domain element of "umap" scheduled after any of
4280 * the corresponding image elements by the tree rooted at
4281 * the child of "node"?
4283 static isl_bool
after_in_child(__isl_keep isl_union_map
*umap
,
4284 __isl_keep isl_schedule_node
*node
)
4286 isl_schedule_node
*child
;
4289 child
= isl_schedule_node_get_child(node
, 0);
4290 after
= after_in_tree(umap
, child
);
4291 isl_schedule_node_free(child
);
4296 /* Is any domain element of "umap" scheduled after any of
4297 * the corresponding image elements by the tree rooted at
4298 * the band node "node"?
4300 * We first check if any domain element is scheduled after any
4301 * of the corresponding image elements by the band node itself.
4302 * If not, we restrict "map" to those pairs of element that
4303 * are scheduled together by the band node and continue with
4304 * the child of the band node.
4305 * If there are no such pairs then the map passed to after_in_child
4306 * will be empty causing it to return 0.
4308 static isl_bool
after_in_band(__isl_keep isl_union_map
*umap
,
4309 __isl_keep isl_schedule_node
*node
)
4311 isl_multi_union_pw_aff
*mupa
;
4312 isl_union_map
*partial
, *test
, *gt
, *universe
, *umap1
, *umap2
;
4313 isl_union_set
*domain
, *range
;
4319 n
= isl_schedule_node_band_n_member(node
);
4321 return isl_bool_error
;
4323 return after_in_child(umap
, node
);
4325 mupa
= isl_schedule_node_band_get_partial_schedule(node
);
4326 space
= isl_multi_union_pw_aff_get_space(mupa
);
4327 partial
= isl_union_map_from_multi_union_pw_aff(mupa
);
4328 test
= isl_union_map_copy(umap
);
4329 test
= isl_union_map_apply_domain(test
, isl_union_map_copy(partial
));
4330 test
= isl_union_map_apply_range(test
, isl_union_map_copy(partial
));
4331 gt
= isl_union_map_from_map(isl_map_lex_gt(space
));
4332 test
= isl_union_map_intersect(test
, gt
);
4333 empty
= isl_union_map_is_empty(test
);
4334 isl_union_map_free(test
);
4336 if (empty
< 0 || !empty
) {
4337 isl_union_map_free(partial
);
4338 return isl_bool_not(empty
);
4341 universe
= isl_union_map_universe(isl_union_map_copy(umap
));
4342 domain
= isl_union_map_domain(isl_union_map_copy(universe
));
4343 range
= isl_union_map_range(universe
);
4344 umap1
= isl_union_map_copy(partial
);
4345 umap1
= isl_union_map_intersect_domain(umap1
, domain
);
4346 umap2
= isl_union_map_intersect_domain(partial
, range
);
4347 test
= isl_union_map_apply_range(umap1
, isl_union_map_reverse(umap2
));
4348 test
= isl_union_map_intersect(test
, isl_union_map_copy(umap
));
4349 after
= after_in_child(test
, node
);
4350 isl_union_map_free(test
);
4354 /* Is any domain element of "umap" scheduled after any of
4355 * the corresponding image elements by the tree rooted at
4356 * the context node "node"?
4358 * The context constraints apply to the schedule domain,
4359 * so we cannot apply them directly to "umap", which contains
4360 * pairs of statement instances. Instead, we add them
4361 * to the range of the prefix schedule for both domain and
4364 static isl_bool
after_in_context(__isl_keep isl_union_map
*umap
,
4365 __isl_keep isl_schedule_node
*node
)
4367 isl_union_map
*prefix
, *universe
, *umap1
, *umap2
;
4368 isl_union_set
*domain
, *range
;
4372 umap
= isl_union_map_copy(umap
);
4373 context
= isl_schedule_node_context_get_context(node
);
4374 prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
4375 universe
= isl_union_map_universe(isl_union_map_copy(umap
));
4376 domain
= isl_union_map_domain(isl_union_map_copy(universe
));
4377 range
= isl_union_map_range(universe
);
4378 umap1
= isl_union_map_copy(prefix
);
4379 umap1
= isl_union_map_intersect_domain(umap1
, domain
);
4380 umap2
= isl_union_map_intersect_domain(prefix
, range
);
4381 umap1
= isl_union_map_intersect_range(umap1
,
4382 isl_union_set_from_set(context
));
4383 umap1
= isl_union_map_apply_range(umap1
, isl_union_map_reverse(umap2
));
4384 umap
= isl_union_map_intersect(umap
, umap1
);
4386 after
= after_in_child(umap
, node
);
4388 isl_union_map_free(umap
);
4393 /* Is any domain element of "umap" scheduled after any of
4394 * the corresponding image elements by the tree rooted at
4395 * the expansion node "node"?
4397 * We apply the expansion to domain and range of "umap" and
4398 * continue with its child.
4400 static isl_bool
after_in_expansion(__isl_keep isl_union_map
*umap
,
4401 __isl_keep isl_schedule_node
*node
)
4403 isl_union_map
*expansion
;
4406 expansion
= isl_schedule_node_expansion_get_expansion(node
);
4407 umap
= isl_union_map_copy(umap
);
4408 umap
= isl_union_map_apply_domain(umap
, isl_union_map_copy(expansion
));
4409 umap
= isl_union_map_apply_range(umap
, expansion
);
4411 after
= after_in_child(umap
, node
);
4413 isl_union_map_free(umap
);
4418 /* Is any domain element of "umap" scheduled after any of
4419 * the corresponding image elements by the tree rooted at
4420 * the extension node "node"?
4422 * Since the extension node may add statement instances before or
4423 * after the pairs of statement instances in "umap", we return isl_bool_true
4424 * to ensure that these pairs are not broken up.
4426 static isl_bool
after_in_extension(__isl_keep isl_union_map
*umap
,
4427 __isl_keep isl_schedule_node
*node
)
4429 return isl_bool_true
;
4432 /* Is any domain element of "umap" scheduled after any of
4433 * the corresponding image elements by the tree rooted at
4434 * the filter node "node"?
4436 * We intersect domain and range of "umap" with the filter and
4437 * continue with its child.
4439 static isl_bool
after_in_filter(__isl_keep isl_union_map
*umap
,
4440 __isl_keep isl_schedule_node
*node
)
4442 isl_union_set
*filter
;
4445 umap
= isl_union_map_copy(umap
);
4446 filter
= isl_schedule_node_filter_get_filter(node
);
4447 umap
= isl_union_map_intersect_domain(umap
, isl_union_set_copy(filter
));
4448 umap
= isl_union_map_intersect_range(umap
, filter
);
4450 after
= after_in_child(umap
, node
);
4452 isl_union_map_free(umap
);
4457 /* Is any domain element of "umap" scheduled after any of
4458 * the corresponding image elements by the tree rooted at
4459 * the set node "node"?
4461 * This is only the case if this condition holds in any
4462 * of the (filter) children of the set node.
4463 * In particular, if the domain and the range of "umap"
4464 * are contained in different children, then the condition
4467 static isl_bool
after_in_set(__isl_keep isl_union_map
*umap
,
4468 __isl_keep isl_schedule_node
*node
)
4473 n
= isl_schedule_node_n_children(node
);
4475 return isl_bool_error
;
4476 for (i
= 0; i
< n
; ++i
) {
4477 isl_schedule_node
*child
;
4480 child
= isl_schedule_node_get_child(node
, i
);
4481 after
= after_in_tree(umap
, child
);
4482 isl_schedule_node_free(child
);
4484 if (after
< 0 || after
)
4488 return isl_bool_false
;
4491 /* Return the filter of child "i" of "node".
4493 static __isl_give isl_union_set
*child_filter(
4494 __isl_keep isl_schedule_node
*node
, int i
)
4496 isl_schedule_node
*child
;
4497 isl_union_set
*filter
;
4499 child
= isl_schedule_node_get_child(node
, i
);
4500 filter
= isl_schedule_node_filter_get_filter(child
);
4501 isl_schedule_node_free(child
);
4506 /* Is any domain element of "umap" scheduled after any of
4507 * the corresponding image elements by the tree rooted at
4508 * the sequence node "node"?
4510 * This happens in particular if any domain element is
4511 * contained in a later child than one containing a range element or
4512 * if the condition holds within a given child in the sequence.
4513 * The later part of the condition is checked by after_in_set.
4515 static isl_bool
after_in_sequence(__isl_keep isl_union_map
*umap
,
4516 __isl_keep isl_schedule_node
*node
)
4520 isl_union_map
*umap_i
;
4522 isl_bool after
= isl_bool_false
;
4524 n
= isl_schedule_node_n_children(node
);
4526 return isl_bool_error
;
4527 for (i
= 1; i
< n
; ++i
) {
4528 isl_union_set
*filter_i
;
4530 umap_i
= isl_union_map_copy(umap
);
4531 filter_i
= child_filter(node
, i
);
4532 umap_i
= isl_union_map_intersect_domain(umap_i
, filter_i
);
4533 empty
= isl_union_map_is_empty(umap_i
);
4537 isl_union_map_free(umap_i
);
4541 for (j
= 0; j
< i
; ++j
) {
4542 isl_union_set
*filter_j
;
4543 isl_union_map
*umap_ij
;
4545 umap_ij
= isl_union_map_copy(umap_i
);
4546 filter_j
= child_filter(node
, j
);
4547 umap_ij
= isl_union_map_intersect_range(umap_ij
,
4549 empty
= isl_union_map_is_empty(umap_ij
);
4550 isl_union_map_free(umap_ij
);
4555 after
= isl_bool_true
;
4560 isl_union_map_free(umap_i
);
4565 if (after
< 0 || after
)
4568 return after_in_set(umap
, node
);
4570 isl_union_map_free(umap_i
);
4571 return isl_bool_error
;
4574 /* Is any domain element of "umap" scheduled after any of
4575 * the corresponding image elements by the tree rooted at "node"?
4577 * If "umap" is empty, then clearly there is no such element.
4578 * Otherwise, consider the different types of nodes separately.
4580 static isl_bool
after_in_tree(__isl_keep isl_union_map
*umap
,
4581 __isl_keep isl_schedule_node
*node
)
4584 enum isl_schedule_node_type type
;
4586 empty
= isl_union_map_is_empty(umap
);
4588 return isl_bool_error
;
4590 return isl_bool_false
;
4592 return isl_bool_error
;
4594 type
= isl_schedule_node_get_type(node
);
4596 case isl_schedule_node_error
:
4597 return isl_bool_error
;
4598 case isl_schedule_node_leaf
:
4599 return isl_bool_false
;
4600 case isl_schedule_node_band
:
4601 return after_in_band(umap
, node
);
4602 case isl_schedule_node_domain
:
4603 isl_die(isl_schedule_node_get_ctx(node
), isl_error_internal
,
4604 "unexpected internal domain node",
4605 return isl_bool_error
);
4606 case isl_schedule_node_context
:
4607 return after_in_context(umap
, node
);
4608 case isl_schedule_node_expansion
:
4609 return after_in_expansion(umap
, node
);
4610 case isl_schedule_node_extension
:
4611 return after_in_extension(umap
, node
);
4612 case isl_schedule_node_filter
:
4613 return after_in_filter(umap
, node
);
4614 case isl_schedule_node_guard
:
4615 case isl_schedule_node_mark
:
4616 return after_in_child(umap
, node
);
4617 case isl_schedule_node_set
:
4618 return after_in_set(umap
, node
);
4619 case isl_schedule_node_sequence
:
4620 return after_in_sequence(umap
, node
);
4623 return isl_bool_true
;
4626 /* Is any domain element of "map1" scheduled after any domain
4627 * element of "map2" by the subtree underneath the current band node,
4628 * while at the same time being scheduled together by the current
4629 * band node, i.e., by "map1" and "map2?
4631 * If the child of the current band node is a leaf, then
4632 * no element can be scheduled after any other element.
4634 * Otherwise, we construct a relation between domain elements
4635 * of "map1" and domain elements of "map2" that are scheduled
4636 * together and then check if the subtree underneath the current
4637 * band node determines their relative order.
4639 static isl_bool
after_in_subtree(__isl_keep isl_ast_build
*build
,
4640 __isl_keep isl_map
*map1
, __isl_keep isl_map
*map2
)
4642 isl_schedule_node
*node
;
4644 isl_union_map
*umap
;
4647 node
= isl_ast_build_get_schedule_node(build
);
4649 return isl_bool_error
;
4650 node
= isl_schedule_node_child(node
, 0);
4651 if (isl_schedule_node_get_type(node
) == isl_schedule_node_leaf
) {
4652 isl_schedule_node_free(node
);
4653 return isl_bool_false
;
4655 map
= isl_map_copy(map2
);
4656 map
= isl_map_apply_domain(map
, isl_map_copy(map1
));
4657 umap
= isl_union_map_from_map(map
);
4658 after
= after_in_tree(umap
, node
);
4659 isl_union_map_free(umap
);
4660 isl_schedule_node_free(node
);
4664 /* Internal data for any_scheduled_after.
4666 * "build" is the build in which the AST is constructed.
4667 * "depth" is the number of loops that have already been generated
4668 * "group_coscheduled" is a local copy of options->ast_build_group_coscheduled
4669 * "domain" is an array of set-map pairs corresponding to the different
4670 * iteration domains. The set is the schedule domain, i.e., the domain
4671 * of the inverse schedule, while the map is the inverse schedule itself.
4673 struct isl_any_scheduled_after_data
{
4674 isl_ast_build
*build
;
4676 int group_coscheduled
;
4677 struct isl_set_map_pair
*domain
;
4680 /* Is any element of domain "i" scheduled after any element of domain "j"
4681 * (for a common iteration of the first data->depth loops)?
4683 * data->domain[i].set contains the domain of the inverse schedule
4684 * for domain "i", i.e., elements in the schedule domain.
4686 * If we are inside a band of a schedule tree and there is a pair
4687 * of elements in the two domains that is schedule together by
4688 * the current band, then we check if any element of "i" may be schedule
4689 * after element of "j" by the descendants of the band node.
4691 * If data->group_coscheduled is set, then we also return 1 if there
4692 * is any pair of elements in the two domains that are scheduled together.
4694 static isl_bool
any_scheduled_after(int i
, int j
, void *user
)
4696 struct isl_any_scheduled_after_data
*data
= user
;
4697 isl_size dim
= isl_set_dim(data
->domain
[i
].set
, isl_dim_set
);
4701 return isl_bool_error
;
4703 for (pos
= data
->depth
; pos
< dim
; ++pos
) {
4706 follows
= isl_set_follows_at(data
->domain
[i
].set
,
4707 data
->domain
[j
].set
, pos
);
4710 return isl_bool_error
;
4712 return isl_bool_true
;
4714 return isl_bool_false
;
4717 if (isl_ast_build_has_schedule_node(data
->build
)) {
4720 after
= after_in_subtree(data
->build
, data
->domain
[i
].map
,
4721 data
->domain
[j
].map
);
4722 if (after
< 0 || after
)
4726 return isl_bool_ok(data
->group_coscheduled
);
4729 /* Look for independent components at the current depth and generate code
4730 * for each component separately. The resulting lists of grafts are
4731 * merged in an attempt to combine grafts with identical guards.
4733 * Code for two domains can be generated separately if all the elements
4734 * of one domain are scheduled before (or together with) all the elements
4735 * of the other domain. We therefore consider the graph with as nodes
4736 * the domains and an edge between two nodes if any element of the first
4737 * node is scheduled after any element of the second node.
4738 * If the ast_build_group_coscheduled is set, then we also add an edge if
4739 * there is any pair of elements in the two domains that are scheduled
4741 * Code is then generated (by generate_component)
4742 * for each of the strongly connected components in this graph
4743 * in their topological order.
4745 * Since the test is performed on the domain of the inverse schedules of
4746 * the different domains, we precompute these domains and store
4747 * them in data.domain.
4749 static __isl_give isl_ast_graft_list
*generate_components(
4750 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
4753 isl_ctx
*ctx
= isl_ast_build_get_ctx(build
);
4754 isl_size n
= isl_union_map_n_map(executed
);
4755 struct isl_any_scheduled_after_data data
;
4756 struct isl_set_map_pair
*next
;
4757 struct isl_tarjan_graph
*g
= NULL
;
4758 isl_ast_graft_list
*list
= NULL
;
4764 data
.domain
= isl_calloc_array(ctx
, struct isl_set_map_pair
, n
);
4770 if (isl_union_map_foreach_map(executed
, &extract_domain
, &next
) < 0)
4776 data
.depth
= isl_ast_build_get_depth(build
);
4777 data
.group_coscheduled
= isl_options_get_ast_build_group_coscheduled(ctx
);
4778 g
= isl_tarjan_graph_init(ctx
, n
, &any_scheduled_after
, &data
);
4782 list
= isl_ast_graft_list_alloc(ctx
, 0);
4786 isl_ast_graft_list
*list_c
;
4789 if (g
->order
[i
] == -1)
4790 isl_die(ctx
, isl_error_internal
, "cannot happen",
4793 while (g
->order
[i
] != -1) {
4797 list_c
= generate_component(data
.domain
,
4798 g
->order
+ first
, i
- first
,
4799 isl_ast_build_copy(build
));
4800 list
= isl_ast_graft_list_merge(list
, list_c
, build
);
4806 error
: list
= isl_ast_graft_list_free(list
);
4807 isl_tarjan_graph_free(g
);
4808 for (i
= 0; i
< n_domain
; ++i
) {
4809 isl_map_free(data
.domain
[i
].map
);
4810 isl_set_free(data
.domain
[i
].set
);
4813 isl_union_map_free(executed
);
4814 isl_ast_build_free(build
);
4819 /* Generate code for the next level (and all inner levels).
4821 * If "executed" is empty, i.e., no code needs to be generated,
4822 * then we return an empty list.
4824 * If we have already generated code for all loop levels, then we pass
4825 * control to generate_inner_level.
4827 * If "executed" lives in a single space, i.e., if code needs to be
4828 * generated for a single domain, then there can only be a single
4829 * component and we go directly to generate_shifted_component.
4830 * Otherwise, we call generate_components to detect the components
4831 * and to call generate_component on each of them separately.
4833 static __isl_give isl_ast_graft_list
*generate_next_level(
4834 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
4840 if (!build
|| !executed
)
4843 if (isl_union_map_is_empty(executed
)) {
4844 isl_ctx
*ctx
= isl_ast_build_get_ctx(build
);
4845 isl_union_map_free(executed
);
4846 isl_ast_build_free(build
);
4847 return isl_ast_graft_list_alloc(ctx
, 0);
4850 depth
= isl_ast_build_get_depth(build
);
4851 dim
= isl_ast_build_dim(build
, isl_dim_set
);
4855 return generate_inner_level(executed
, build
);
4857 n
= isl_union_map_n_map(executed
);
4861 return generate_shifted_component(executed
, build
);
4863 return generate_components(executed
, build
);
4865 isl_union_map_free(executed
);
4866 isl_ast_build_free(build
);
4870 /* Internal data structure used by isl_ast_build_node_from_schedule_map.
4871 * internal, executed and build are the inputs to generate_code.
4872 * list collects the output.
4874 struct isl_generate_code_data
{
4876 isl_union_map
*executed
;
4877 isl_ast_build
*build
;
4879 isl_ast_graft_list
*list
;
4882 /* Given an inverse schedule in terms of the external build schedule, i.e.,
4886 * with E the external build schedule and S the additional schedule "space",
4887 * reformulate the inverse schedule in terms of the internal schedule domain,
4892 * We first obtain a mapping
4896 * take the inverse and the product with S -> S, resulting in
4898 * [I -> S] -> [E -> S]
4900 * Applying the map to the input produces the desired result.
4902 static __isl_give isl_union_map
*internal_executed(
4903 __isl_take isl_union_map
*executed
, __isl_keep isl_space
*space
,
4904 __isl_keep isl_ast_build
*build
)
4908 proj
= isl_ast_build_get_schedule_map(build
);
4909 proj
= isl_map_reverse(proj
);
4910 space
= isl_space_map_from_set(isl_space_copy(space
));
4911 id
= isl_map_identity(space
);
4912 proj
= isl_map_product(proj
, id
);
4913 executed
= isl_union_map_apply_domain(executed
,
4914 isl_union_map_from_map(proj
));
4918 /* Generate an AST that visits the elements in the range of data->executed
4919 * in the relative order specified by the corresponding domain element(s)
4920 * for those domain elements that belong to "set".
4921 * Add the result to data->list.
4923 * The caller ensures that "set" is a universe domain.
4924 * "space" is the space of the additional part of the schedule.
4925 * It is equal to the space of "set" if build->domain is parametric.
4926 * Otherwise, it is equal to the range of the wrapped space of "set".
4928 * If the build space is not parametric and
4929 * if isl_ast_build_node_from_schedule_map
4930 * was called from an outside user (data->internal not set), then
4931 * the (inverse) schedule refers to the external build domain and needs to
4932 * be transformed to refer to the internal build domain.
4934 * If the build space is parametric, then we add some of the parameter
4935 * constraints to the executed relation. Adding these constraints
4936 * allows for an earlier detection of conflicts in some cases.
4937 * However, we do not want to divide the executed relation into
4938 * more disjuncts than necessary. We therefore approximate
4939 * the constraints on the parameters by a single disjunct set.
4941 * The build is extended to include the additional part of the schedule.
4942 * If the original build space was not parametric, then the options
4943 * in data->build refer only to the additional part of the schedule
4944 * and they need to be adjusted to refer to the complete AST build
4947 * After having adjusted inverse schedule and build, we start generating
4948 * code with the outer loop of the current code generation
4949 * in generate_next_level.
4951 * If the original build space was not parametric, we undo the embedding
4952 * on the resulting isl_ast_node_list so that it can be used within
4953 * the outer AST build.
4955 static isl_stat
generate_code_in_space(struct isl_generate_code_data
*data
,
4956 __isl_take isl_set
*set
, __isl_take isl_space
*space
)
4958 isl_union_map
*executed
;
4959 isl_ast_build
*build
;
4960 isl_ast_graft_list
*list
;
4963 executed
= isl_union_map_copy(data
->executed
);
4964 executed
= isl_union_map_intersect_domain(executed
,
4965 isl_union_set_from_set(set
));
4967 embed
= !isl_set_is_params(data
->build
->domain
);
4968 if (embed
&& !data
->internal
)
4969 executed
= internal_executed(executed
, space
, data
->build
);
4972 domain
= isl_ast_build_get_domain(data
->build
);
4973 domain
= isl_set_from_basic_set(isl_set_simple_hull(domain
));
4974 executed
= isl_union_map_intersect_params(executed
, domain
);
4977 build
= isl_ast_build_copy(data
->build
);
4978 build
= isl_ast_build_product(build
, space
);
4980 list
= generate_next_level(executed
, build
);
4982 list
= isl_ast_graft_list_unembed(list
, embed
);
4984 data
->list
= isl_ast_graft_list_concat(data
->list
, list
);
4989 /* Generate an AST that visits the elements in the range of data->executed
4990 * in the relative order specified by the corresponding domain element(s)
4991 * for those domain elements that belong to "set".
4992 * Add the result to data->list.
4994 * The caller ensures that "set" is a universe domain.
4996 * If the build space S is not parametric, then the space of "set"
4997 * need to be a wrapped relation with S as domain. That is, it needs
5002 * Check this property and pass control to generate_code_in_space
5004 * If the build space is not parametric, then T is the space of "set".
5006 static isl_stat
generate_code_set(__isl_take isl_set
*set
, void *user
)
5008 struct isl_generate_code_data
*data
= user
;
5009 isl_space
*space
, *build_space
;
5012 space
= isl_set_get_space(set
);
5014 if (isl_set_is_params(data
->build
->domain
))
5015 return generate_code_in_space(data
, set
, space
);
5017 build_space
= isl_ast_build_get_space(data
->build
, data
->internal
);
5018 space
= isl_space_unwrap(space
);
5019 is_domain
= isl_space_is_domain(build_space
, space
);
5020 isl_space_free(build_space
);
5021 space
= isl_space_range(space
);
5026 isl_die(isl_set_get_ctx(set
), isl_error_invalid
,
5027 "invalid nested schedule space", goto error
);
5029 return generate_code_in_space(data
, set
, space
);
5032 isl_space_free(space
);
5033 return isl_stat_error
;
5036 /* Generate an AST that visits the elements in the range of "executed"
5037 * in the relative order specified by the corresponding domain element(s).
5039 * "build" is an isl_ast_build that has either been constructed by
5040 * isl_ast_build_from_context or passed to a callback set by
5041 * isl_ast_build_set_create_leaf.
5042 * In the first case, the space of the isl_ast_build is typically
5043 * a parametric space, although this is currently not enforced.
5044 * In the second case, the space is never a parametric space.
5045 * If the space S is not parametric, then the domain space(s) of "executed"
5046 * need to be wrapped relations with S as domain.
5048 * If the domain of "executed" consists of several spaces, then an AST
5049 * is generated for each of them (in arbitrary order) and the results
5052 * If "internal" is set, then the domain "S" above refers to the internal
5053 * schedule domain representation. Otherwise, it refers to the external
5054 * representation, as returned by isl_ast_build_get_schedule_space.
5056 * We essentially run over all the spaces in the domain of "executed"
5057 * and call generate_code_set on each of them.
5059 static __isl_give isl_ast_graft_list
*generate_code(
5060 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
,
5064 struct isl_generate_code_data data
= { 0 };
5066 isl_union_set
*schedule_domain
;
5067 isl_union_map
*universe
;
5071 space
= isl_ast_build_get_space(build
, 1);
5072 space
= isl_space_align_params(space
,
5073 isl_union_map_get_space(executed
));
5074 space
= isl_space_align_params(space
,
5075 isl_union_map_get_space(build
->options
));
5076 build
= isl_ast_build_align_params(build
, isl_space_copy(space
));
5077 executed
= isl_union_map_align_params(executed
, space
);
5078 if (!executed
|| !build
)
5081 ctx
= isl_ast_build_get_ctx(build
);
5083 data
.internal
= internal
;
5084 data
.executed
= executed
;
5086 data
.list
= isl_ast_graft_list_alloc(ctx
, 0);
5088 universe
= isl_union_map_universe(isl_union_map_copy(executed
));
5089 schedule_domain
= isl_union_map_domain(universe
);
5090 if (isl_union_set_foreach_set(schedule_domain
, &generate_code_set
,
5092 data
.list
= isl_ast_graft_list_free(data
.list
);
5094 isl_union_set_free(schedule_domain
);
5095 isl_union_map_free(executed
);
5097 isl_ast_build_free(build
);
5100 isl_union_map_free(executed
);
5101 isl_ast_build_free(build
);
5105 /* Generate an AST that visits the elements in the domain of "schedule"
5106 * in the relative order specified by the corresponding image element(s).
5108 * "build" is an isl_ast_build that has either been constructed by
5109 * isl_ast_build_from_context or passed to a callback set by
5110 * isl_ast_build_set_create_leaf.
5111 * In the first case, the space of the isl_ast_build is typically
5112 * a parametric space, although this is currently not enforced.
5113 * In the second case, the space is never a parametric space.
5114 * If the space S is not parametric, then the range space(s) of "schedule"
5115 * need to be wrapped relations with S as domain.
5117 * If the range of "schedule" consists of several spaces, then an AST
5118 * is generated for each of them (in arbitrary order) and the results
5121 * We first initialize the local copies of the relevant options.
5122 * We do this here rather than when the isl_ast_build is created
5123 * because the options may have changed between the construction
5124 * of the isl_ast_build and the call to isl_generate_code.
5126 * The main computation is performed on an inverse schedule (with
5127 * the schedule domain in the domain and the elements to be executed
5128 * in the range) called "executed".
5130 __isl_give isl_ast_node
*isl_ast_build_node_from_schedule_map(
5131 __isl_keep isl_ast_build
*build
, __isl_take isl_union_map
*schedule
)
5133 isl_ast_graft_list
*list
;
5135 isl_union_map
*executed
;
5137 build
= isl_ast_build_copy(build
);
5138 build
= isl_ast_build_set_single_valued(build
, 0);
5139 schedule
= isl_union_map_coalesce(schedule
);
5140 schedule
= isl_union_map_remove_redundancies(schedule
);
5141 executed
= isl_union_map_reverse(schedule
);
5142 list
= generate_code(executed
, isl_ast_build_copy(build
), 0);
5143 node
= isl_ast_node_from_graft_list(list
, build
);
5144 isl_ast_build_free(build
);
5149 /* The old name for isl_ast_build_node_from_schedule_map.
5150 * It is being kept for backward compatibility, but
5151 * it will be removed in the future.
5153 __isl_give isl_ast_node
*isl_ast_build_ast_from_schedule(
5154 __isl_keep isl_ast_build
*build
, __isl_take isl_union_map
*schedule
)
5156 return isl_ast_build_node_from_schedule_map(build
, schedule
);
5159 /* Generate an AST that visits the elements in the domain of "executed"
5160 * in the relative order specified by the leaf node "node".
5162 * The relation "executed" maps the outer generated loop iterators
5163 * to the domain elements executed by those iterations.
5165 * Simply pass control to generate_inner_level.
5166 * Note that the current build does not refer to any band node, so
5167 * that generate_inner_level will not try to visit the child of
5170 * If multiple statement instances reach a leaf,
5171 * then they can be executed in any order.
5172 * Group the list of grafts based on shared guards
5173 * such that identical guards are only generated once
5174 * when the list is eventually passed on to isl_ast_graft_list_fuse.
5176 static __isl_give isl_ast_graft_list
*build_ast_from_leaf(
5177 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5178 __isl_take isl_union_map
*executed
)
5180 isl_ast_graft_list
*list
;
5182 isl_schedule_node_free(node
);
5183 list
= generate_inner_level(executed
, isl_ast_build_copy(build
));
5184 list
= isl_ast_graft_list_group_on_guard(list
, build
);
5185 isl_ast_build_free(build
);
5190 /* Generate an AST that visits the elements in the domain of "executed"
5191 * in the relative order specified by the band node "node" and its descendants.
5193 * The relation "executed" maps the outer generated loop iterators
5194 * to the domain elements executed by those iterations.
5196 * If the band is empty, we continue with its descendants.
5197 * Otherwise, we extend the build and the inverse schedule with
5198 * the additional space/partial schedule and continue generating
5199 * an AST in generate_next_level.
5200 * As soon as we have extended the inverse schedule with the additional
5201 * partial schedule, we look for equalities that may exists between
5202 * the old and the new part.
5204 static __isl_give isl_ast_graft_list
*build_ast_from_band(
5205 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5206 __isl_take isl_union_map
*executed
)
5209 isl_multi_union_pw_aff
*extra
;
5210 isl_union_map
*extra_umap
;
5211 isl_ast_graft_list
*list
;
5215 n
= isl_schedule_node_band_n_member(node
);
5216 if (!build
|| n
< 0 || !executed
)
5220 return build_ast_from_child(build
, node
, executed
);
5222 extra
= isl_schedule_node_band_get_partial_schedule(node
);
5223 extra
= isl_multi_union_pw_aff_align_params(extra
,
5224 isl_ast_build_get_space(build
, 1));
5225 space
= isl_multi_union_pw_aff_get_space(extra
);
5227 extra_umap
= isl_union_map_from_multi_union_pw_aff(extra
);
5228 extra_umap
= isl_union_map_reverse(extra_umap
);
5230 executed
= isl_union_map_domain_product(executed
, extra_umap
);
5231 executed
= isl_union_map_detect_equalities(executed
);
5233 n1
= isl_ast_build_dim(build
, isl_dim_param
);
5234 build
= isl_ast_build_product(build
, space
);
5235 n2
= isl_ast_build_dim(build
, isl_dim_param
);
5236 if (n1
< 0 || n2
< 0)
5237 build
= isl_ast_build_free(build
);
5239 isl_die(isl_ast_build_get_ctx(build
), isl_error_invalid
,
5240 "band node is not allowed to introduce new parameters",
5241 build
= isl_ast_build_free(build
));
5242 build
= isl_ast_build_set_schedule_node(build
, node
);
5244 list
= generate_next_level(executed
, build
);
5246 list
= isl_ast_graft_list_unembed(list
, 1);
5250 isl_schedule_node_free(node
);
5251 isl_union_map_free(executed
);
5252 isl_ast_build_free(build
);
5256 /* Hoist a list of grafts (in practice containing a single graft)
5257 * from "sub_build" (which includes extra context information)
5260 * In particular, project out all additional parameters introduced
5261 * by the context node from the enforced constraints and the guard
5262 * of the single graft.
5264 static __isl_give isl_ast_graft_list
*hoist_out_of_context(
5265 __isl_take isl_ast_graft_list
*list
, __isl_keep isl_ast_build
*build
,
5266 __isl_keep isl_ast_build
*sub_build
)
5268 isl_ast_graft
*graft
;
5269 isl_basic_set
*enforced
;
5271 isl_size n_param
, extra_param
;
5273 n_param
= isl_ast_build_dim(build
, isl_dim_param
);
5274 extra_param
= isl_ast_build_dim(sub_build
, isl_dim_param
);
5275 if (n_param
< 0 || extra_param
< 0)
5276 return isl_ast_graft_list_free(list
);
5278 if (extra_param
== n_param
)
5281 extra_param
-= n_param
;
5282 enforced
= isl_ast_graft_list_extract_shared_enforced(list
, sub_build
);
5283 enforced
= isl_basic_set_project_out(enforced
, isl_dim_param
,
5284 n_param
, extra_param
);
5285 enforced
= isl_basic_set_remove_unknown_divs(enforced
);
5286 guard
= isl_ast_graft_list_extract_hoistable_guard(list
, sub_build
);
5287 guard
= isl_set_remove_divs_involving_dims(guard
, isl_dim_param
,
5288 n_param
, extra_param
);
5289 guard
= isl_set_project_out(guard
, isl_dim_param
, n_param
, extra_param
);
5290 guard
= isl_set_compute_divs(guard
);
5291 graft
= isl_ast_graft_alloc_from_children(list
, guard
, enforced
,
5293 list
= isl_ast_graft_list_from_ast_graft(graft
);
5298 /* Generate an AST that visits the elements in the domain of "executed"
5299 * in the relative order specified by the context node "node"
5300 * and its descendants.
5302 * The relation "executed" maps the outer generated loop iterators
5303 * to the domain elements executed by those iterations.
5305 * The context node may introduce additional parameters as well as
5306 * constraints on the outer schedule dimensions or original parameters.
5308 * We add the extra parameters to a new build and the context
5309 * constraints to both the build and (as a single disjunct)
5310 * to the domain of "executed". Since the context constraints
5311 * are specified in terms of the input schedule, we first need
5312 * to map them to the internal schedule domain.
5314 * After constructing the AST from the descendants of "node",
5315 * we combine the list of grafts into a single graft within
5316 * the new build, in order to be able to exploit the additional
5317 * context constraints during this combination.
5319 * Additionally, if the current node is the outermost node in
5320 * the schedule tree (apart from the root domain node), we generate
5321 * all pending guards, again to be able to exploit the additional
5322 * context constraints. We currently do not do this for internal
5323 * context nodes since we may still want to hoist conditions
5324 * to outer AST nodes.
5326 * If the context node introduced any new parameters, then they
5327 * are removed from the set of enforced constraints and guard
5328 * in hoist_out_of_context.
5330 static __isl_give isl_ast_graft_list
*build_ast_from_context(
5331 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5332 __isl_take isl_union_map
*executed
)
5336 isl_multi_aff
*internal2input
;
5337 isl_ast_build
*sub_build
;
5338 isl_ast_graft_list
*list
;
5342 depth
= isl_schedule_node_get_tree_depth(node
);
5344 build
= isl_ast_build_free(build
);
5345 space
= isl_ast_build_get_space(build
, 1);
5346 context
= isl_schedule_node_context_get_context(node
);
5347 context
= isl_set_align_params(context
, space
);
5348 sub_build
= isl_ast_build_copy(build
);
5349 space
= isl_set_get_space(context
);
5350 sub_build
= isl_ast_build_align_params(sub_build
, space
);
5351 internal2input
= isl_ast_build_get_internal2input(sub_build
);
5352 context
= isl_set_preimage_multi_aff(context
, internal2input
);
5353 sub_build
= isl_ast_build_restrict_generated(sub_build
,
5354 isl_set_copy(context
));
5355 context
= isl_set_from_basic_set(isl_set_simple_hull(context
));
5356 executed
= isl_union_map_intersect_domain(executed
,
5357 isl_union_set_from_set(context
));
5359 list
= build_ast_from_child(isl_ast_build_copy(sub_build
),
5361 n
= isl_ast_graft_list_n_ast_graft(list
);
5363 list
= isl_ast_graft_list_free(list
);
5365 list
= isl_ast_graft_list_fuse(list
, sub_build
);
5367 list
= isl_ast_graft_list_insert_pending_guard_nodes(list
,
5370 list
= hoist_out_of_context(list
, build
, sub_build
);
5372 isl_ast_build_free(build
);
5373 isl_ast_build_free(sub_build
);
5378 /* Generate an AST that visits the elements in the domain of "executed"
5379 * in the relative order specified by the expansion node "node" and
5382 * The relation "executed" maps the outer generated loop iterators
5383 * to the domain elements executed by those iterations.
5385 * We expand the domain elements by the expansion and
5386 * continue with the descendants of the node.
5388 static __isl_give isl_ast_graft_list
*build_ast_from_expansion(
5389 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5390 __isl_take isl_union_map
*executed
)
5392 isl_union_map
*expansion
;
5395 expansion
= isl_schedule_node_expansion_get_expansion(node
);
5396 expansion
= isl_union_map_align_params(expansion
,
5397 isl_union_map_get_space(executed
));
5399 n1
= isl_union_map_dim(executed
, isl_dim_param
);
5400 executed
= isl_union_map_apply_range(executed
, expansion
);
5401 n2
= isl_union_map_dim(executed
, isl_dim_param
);
5402 if (n1
< 0 || n2
< 0)
5405 isl_die(isl_ast_build_get_ctx(build
), isl_error_invalid
,
5406 "expansion node is not allowed to introduce "
5407 "new parameters", goto error
);
5409 return build_ast_from_child(build
, node
, executed
);
5411 isl_ast_build_free(build
);
5412 isl_schedule_node_free(node
);
5413 isl_union_map_free(executed
);
5417 /* Generate an AST that visits the elements in the domain of "executed"
5418 * in the relative order specified by the extension node "node" and
5421 * The relation "executed" maps the outer generated loop iterators
5422 * to the domain elements executed by those iterations.
5424 * Extend the inverse schedule with the extension applied to current
5425 * set of generated constraints. Since the extension if formulated
5426 * in terms of the input schedule, it first needs to be transformed
5427 * to refer to the internal schedule.
5429 static __isl_give isl_ast_graft_list
*build_ast_from_extension(
5430 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5431 __isl_take isl_union_map
*executed
)
5433 isl_union_set
*schedule_domain
;
5434 isl_union_map
*extension
;
5437 set
= isl_ast_build_get_generated(build
);
5438 set
= isl_set_from_basic_set(isl_set_simple_hull(set
));
5439 schedule_domain
= isl_union_set_from_set(set
);
5441 extension
= isl_schedule_node_extension_get_extension(node
);
5443 extension
= isl_union_map_preimage_domain_multi_aff(extension
,
5444 isl_multi_aff_copy(build
->internal2input
));
5445 extension
= isl_union_map_intersect_domain(extension
, schedule_domain
);
5446 extension
= isl_ast_build_substitute_values_union_map_domain(build
,
5448 executed
= isl_union_map_union(executed
, extension
);
5450 return build_ast_from_child(build
, node
, executed
);
5453 /* Generate an AST that visits the elements in the domain of "executed"
5454 * in the relative order specified by the filter node "node" and
5457 * The relation "executed" maps the outer generated loop iterators
5458 * to the domain elements executed by those iterations.
5460 * We simply intersect the iteration domain (i.e., the range of "executed")
5461 * with the filter and continue with the descendants of the node,
5462 * unless the resulting inverse schedule is empty, in which
5463 * case we return an empty list.
5465 * If the result of the intersection is equal to the original "executed"
5466 * relation, then keep the original representation since the intersection
5467 * may have unnecessarily broken up the relation into a greater number
5470 static __isl_give isl_ast_graft_list
*build_ast_from_filter(
5471 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5472 __isl_take isl_union_map
*executed
)
5475 isl_union_set
*filter
;
5476 isl_union_map
*orig
;
5477 isl_ast_graft_list
*list
;
5482 orig
= isl_union_map_copy(executed
);
5483 if (!build
|| !node
|| !executed
)
5486 filter
= isl_schedule_node_filter_get_filter(node
);
5487 filter
= isl_union_set_align_params(filter
,
5488 isl_union_map_get_space(executed
));
5489 n1
= isl_union_map_dim(executed
, isl_dim_param
);
5490 executed
= isl_union_map_intersect_range(executed
, filter
);
5491 n2
= isl_union_map_dim(executed
, isl_dim_param
);
5492 if (n1
< 0 || n2
< 0)
5495 isl_die(isl_ast_build_get_ctx(build
), isl_error_invalid
,
5496 "filter node is not allowed to introduce "
5497 "new parameters", goto error
);
5499 unchanged
= isl_union_map_is_subset(orig
, executed
);
5500 empty
= isl_union_map_is_empty(executed
);
5501 if (unchanged
< 0 || empty
< 0)
5504 isl_union_map_free(executed
);
5505 return build_ast_from_child(build
, node
, orig
);
5507 isl_union_map_free(orig
);
5509 return build_ast_from_child(build
, node
, executed
);
5511 ctx
= isl_ast_build_get_ctx(build
);
5512 list
= isl_ast_graft_list_alloc(ctx
, 0);
5513 isl_ast_build_free(build
);
5514 isl_schedule_node_free(node
);
5515 isl_union_map_free(executed
);
5518 isl_ast_build_free(build
);
5519 isl_schedule_node_free(node
);
5520 isl_union_map_free(executed
);
5521 isl_union_map_free(orig
);
5525 /* Generate an AST that visits the elements in the domain of "executed"
5526 * in the relative order specified by the guard node "node" and
5529 * The relation "executed" maps the outer generated loop iterators
5530 * to the domain elements executed by those iterations.
5532 * Ensure that the associated guard is enforced by the outer AST
5533 * constructs by adding it to the guard of the graft.
5534 * Since we know that we will enforce the guard, we can also include it
5535 * in the generated constraints used to construct an AST for
5536 * the descendant nodes.
5538 static __isl_give isl_ast_graft_list
*build_ast_from_guard(
5539 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5540 __isl_take isl_union_map
*executed
)
5543 isl_set
*guard
, *hoisted
;
5544 isl_basic_set
*enforced
;
5545 isl_ast_build
*sub_build
;
5546 isl_ast_graft
*graft
;
5547 isl_ast_graft_list
*list
;
5550 space
= isl_ast_build_get_space(build
, 1);
5551 guard
= isl_schedule_node_guard_get_guard(node
);
5552 n1
= isl_space_dim(space
, isl_dim_param
);
5553 guard
= isl_set_align_params(guard
, space
);
5554 n2
= isl_set_dim(guard
, isl_dim_param
);
5555 if (n1
< 0 || n2
< 0)
5556 guard
= isl_set_free(guard
);
5558 isl_die(isl_ast_build_get_ctx(build
), isl_error_invalid
,
5559 "guard node is not allowed to introduce "
5560 "new parameters", guard
= isl_set_free(guard
));
5561 guard
= isl_set_preimage_multi_aff(guard
,
5562 isl_multi_aff_copy(build
->internal2input
));
5563 guard
= isl_ast_build_specialize(build
, guard
);
5564 guard
= isl_set_gist(guard
, isl_set_copy(build
->generated
));
5566 sub_build
= isl_ast_build_copy(build
);
5567 sub_build
= isl_ast_build_restrict_generated(sub_build
,
5568 isl_set_copy(guard
));
5570 list
= build_ast_from_child(isl_ast_build_copy(sub_build
),
5573 hoisted
= isl_ast_graft_list_extract_hoistable_guard(list
, sub_build
);
5574 n
= isl_set_n_basic_set(hoisted
);
5576 list
= isl_ast_graft_list_free(list
);
5578 list
= isl_ast_graft_list_gist_guards(list
,
5579 isl_set_copy(hoisted
));
5580 guard
= isl_set_intersect(guard
, hoisted
);
5581 enforced
= extract_shared_enforced(list
, build
);
5582 graft
= isl_ast_graft_alloc_from_children(list
, guard
, enforced
,
5585 isl_ast_build_free(sub_build
);
5586 isl_ast_build_free(build
);
5587 return isl_ast_graft_list_from_ast_graft(graft
);
5590 /* Call the before_each_mark callback, if requested by the user.
5592 * Return 0 on success and -1 on error.
5594 * The caller is responsible for recording the current inverse schedule
5597 static isl_stat
before_each_mark(__isl_keep isl_id
*mark
,
5598 __isl_keep isl_ast_build
*build
)
5601 return isl_stat_error
;
5602 if (!build
->before_each_mark
)
5604 return build
->before_each_mark(mark
, build
,
5605 build
->before_each_mark_user
);
5608 /* Call the after_each_mark callback, if requested by the user.
5610 * The caller is responsible for recording the current inverse schedule
5613 static __isl_give isl_ast_graft
*after_each_mark(
5614 __isl_take isl_ast_graft
*graft
, __isl_keep isl_ast_build
*build
)
5616 if (!graft
|| !build
)
5617 return isl_ast_graft_free(graft
);
5618 if (!build
->after_each_mark
)
5620 graft
->node
= build
->after_each_mark(graft
->node
, build
,
5621 build
->after_each_mark_user
);
5623 return isl_ast_graft_free(graft
);
5628 /* Generate an AST that visits the elements in the domain of "executed"
5629 * in the relative order specified by the mark node "node" and
5632 * The relation "executed" maps the outer generated loop iterators
5633 * to the domain elements executed by those iterations.
5635 * Since we may be calling before_each_mark and after_each_mark
5636 * callbacks, we record the current inverse schedule in the build.
5638 * We generate an AST for the child of the mark node, combine
5639 * the graft list into a single graft and then insert the mark
5640 * in the AST of that single graft.
5642 static __isl_give isl_ast_graft_list
*build_ast_from_mark(
5643 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5644 __isl_take isl_union_map
*executed
)
5647 isl_ast_graft
*graft
;
5648 isl_ast_graft_list
*list
;
5651 build
= isl_ast_build_set_executed(build
, isl_union_map_copy(executed
));
5653 mark
= isl_schedule_node_mark_get_id(node
);
5654 if (before_each_mark(mark
, build
) < 0)
5655 node
= isl_schedule_node_free(node
);
5657 list
= build_ast_from_child(isl_ast_build_copy(build
), node
, executed
);
5658 list
= isl_ast_graft_list_fuse(list
, build
);
5659 n
= isl_ast_graft_list_n_ast_graft(list
);
5661 list
= isl_ast_graft_list_free(list
);
5665 graft
= isl_ast_graft_list_get_ast_graft(list
, 0);
5666 graft
= isl_ast_graft_insert_mark(graft
, mark
);
5667 graft
= after_each_mark(graft
, build
);
5668 list
= isl_ast_graft_list_set_ast_graft(list
, 0, graft
);
5670 isl_ast_build_free(build
);
5675 static __isl_give isl_ast_graft_list
*build_ast_from_schedule_node(
5676 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5677 __isl_take isl_union_map
*executed
);
5679 /* Generate an AST that visits the elements in the domain of "executed"
5680 * in the relative order specified by the sequence (or set) node "node" and
5683 * The relation "executed" maps the outer generated loop iterators
5684 * to the domain elements executed by those iterations.
5686 * We simply generate an AST for each of the children and concatenate
5689 static __isl_give isl_ast_graft_list
*build_ast_from_sequence(
5690 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5691 __isl_take isl_union_map
*executed
)
5696 isl_ast_graft_list
*list
;
5698 ctx
= isl_ast_build_get_ctx(build
);
5699 list
= isl_ast_graft_list_alloc(ctx
, 0);
5701 n
= isl_schedule_node_n_children(node
);
5703 list
= isl_ast_graft_list_free(list
);
5704 for (i
= 0; i
< n
; ++i
) {
5705 isl_schedule_node
*child
;
5706 isl_ast_graft_list
*list_i
;
5708 child
= isl_schedule_node_get_child(node
, i
);
5709 list_i
= build_ast_from_schedule_node(isl_ast_build_copy(build
),
5710 child
, isl_union_map_copy(executed
));
5711 list
= isl_ast_graft_list_concat(list
, list_i
);
5713 isl_ast_build_free(build
);
5714 isl_schedule_node_free(node
);
5715 isl_union_map_free(executed
);
5720 /* Generate an AST that visits the elements in the domain of "executed"
5721 * in the relative order specified by the node "node" and its descendants.
5723 * The relation "executed" maps the outer generated loop iterators
5724 * to the domain elements executed by those iterations.
5726 * The node types are handled in separate functions.
5727 * Set nodes are currently treated in the same way as sequence nodes.
5728 * The children of a set node may be executed in any order,
5729 * including the order of the children.
5731 static __isl_give isl_ast_graft_list
*build_ast_from_schedule_node(
5732 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5733 __isl_take isl_union_map
*executed
)
5735 enum isl_schedule_node_type type
;
5737 type
= isl_schedule_node_get_type(node
);
5740 case isl_schedule_node_error
:
5742 case isl_schedule_node_leaf
:
5743 return build_ast_from_leaf(build
, node
, executed
);
5744 case isl_schedule_node_band
:
5745 return build_ast_from_band(build
, node
, executed
);
5746 case isl_schedule_node_context
:
5747 return build_ast_from_context(build
, node
, executed
);
5748 case isl_schedule_node_domain
:
5749 isl_die(isl_schedule_node_get_ctx(node
), isl_error_unsupported
,
5750 "unexpected internal domain node", goto error
);
5751 case isl_schedule_node_expansion
:
5752 return build_ast_from_expansion(build
, node
, executed
);
5753 case isl_schedule_node_extension
:
5754 return build_ast_from_extension(build
, node
, executed
);
5755 case isl_schedule_node_filter
:
5756 return build_ast_from_filter(build
, node
, executed
);
5757 case isl_schedule_node_guard
:
5758 return build_ast_from_guard(build
, node
, executed
);
5759 case isl_schedule_node_mark
:
5760 return build_ast_from_mark(build
, node
, executed
);
5761 case isl_schedule_node_sequence
:
5762 case isl_schedule_node_set
:
5763 return build_ast_from_sequence(build
, node
, executed
);
5766 isl_die(isl_ast_build_get_ctx(build
), isl_error_internal
,
5767 "unhandled type", goto error
);
5769 isl_union_map_free(executed
);
5770 isl_schedule_node_free(node
);
5771 isl_ast_build_free(build
);
5776 /* Generate an AST that visits the elements in the domain of "executed"
5777 * in the relative order specified by the (single) child of "node" and
5780 * The relation "executed" maps the outer generated loop iterators
5781 * to the domain elements executed by those iterations.
5783 * This function is never called on a leaf, set or sequence node,
5784 * so the node always has exactly one child.
5786 static __isl_give isl_ast_graft_list
*build_ast_from_child(
5787 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5788 __isl_take isl_union_map
*executed
)
5790 node
= isl_schedule_node_child(node
, 0);
5791 return build_ast_from_schedule_node(build
, node
, executed
);
5794 /* Generate an AST that visits the elements in the domain of the domain
5795 * node "node" in the relative order specified by its descendants.
5797 * An initial inverse schedule is created that maps a zero-dimensional
5798 * schedule space to the node domain.
5799 * The input "build" is assumed to have a parametric domain and
5800 * is replaced by the same zero-dimensional schedule space.
5802 * We also add some of the parameter constraints in the build domain
5803 * to the executed relation. Adding these constraints
5804 * allows for an earlier detection of conflicts in some cases.
5805 * However, we do not want to divide the executed relation into
5806 * more disjuncts than necessary. We therefore approximate
5807 * the constraints on the parameters by a single disjunct set.
5809 static __isl_give isl_ast_node
*build_ast_from_domain(
5810 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
)
5813 isl_union_set
*domain
, *schedule_domain
;
5814 isl_union_map
*executed
;
5817 isl_ast_graft_list
*list
;
5824 ctx
= isl_ast_build_get_ctx(build
);
5825 space
= isl_ast_build_get_space(build
, 1);
5826 is_params
= isl_space_is_params(space
);
5827 isl_space_free(space
);
5831 isl_die(ctx
, isl_error_unsupported
,
5832 "expecting parametric initial context", goto error
);
5834 domain
= isl_schedule_node_domain_get_domain(node
);
5835 domain
= isl_union_set_coalesce(domain
);
5837 space
= isl_union_set_get_space(domain
);
5838 space
= isl_space_set_from_params(space
);
5839 build
= isl_ast_build_product(build
, space
);
5841 set
= isl_ast_build_get_domain(build
);
5842 set
= isl_set_from_basic_set(isl_set_simple_hull(set
));
5843 schedule_domain
= isl_union_set_from_set(set
);
5845 executed
= isl_union_map_from_domain_and_range(schedule_domain
, domain
);
5846 list
= build_ast_from_child(isl_ast_build_copy(build
), node
, executed
);
5847 ast
= isl_ast_node_from_graft_list(list
, build
);
5848 isl_ast_build_free(build
);
5852 isl_schedule_node_free(node
);
5853 isl_ast_build_free(build
);
5857 /* Generate an AST that visits the elements in the domain of "schedule"
5858 * in the relative order specified by the schedule tree.
5860 * "build" is an isl_ast_build that has been created using
5861 * isl_ast_build_alloc or isl_ast_build_from_context based
5862 * on a parametric set.
5864 * The construction starts at the root node of the schedule,
5865 * which is assumed to be a domain node.
5867 __isl_give isl_ast_node
*isl_ast_build_node_from_schedule(
5868 __isl_keep isl_ast_build
*build
, __isl_take isl_schedule
*schedule
)
5871 isl_schedule_node
*node
;
5873 if (!build
|| !schedule
)
5876 ctx
= isl_ast_build_get_ctx(build
);
5878 node
= isl_schedule_get_root(schedule
);
5881 isl_schedule_free(schedule
);
5883 build
= isl_ast_build_copy(build
);
5884 build
= isl_ast_build_set_single_valued(build
, 0);
5885 if (isl_schedule_node_get_type(node
) != isl_schedule_node_domain
)
5886 isl_die(ctx
, isl_error_unsupported
,
5887 "expecting root domain node",
5888 build
= isl_ast_build_free(build
));
5889 return build_ast_from_domain(build
, node
);
5891 isl_schedule_free(schedule
);