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
);
100 list
= generate_code(isl_union_map_from_map(executed
), build
, 1);
102 data
->list
= isl_ast_graft_list_concat(data
->list
, list
);
107 /* Call the at_each_domain callback, if requested by the user,
108 * after recording the current inverse schedule in the build.
110 static __isl_give isl_ast_graft
*at_each_domain(__isl_take isl_ast_graft
*graft
,
111 __isl_keep isl_map
*executed
, __isl_keep isl_ast_build
*build
)
113 if (!graft
|| !build
)
114 return isl_ast_graft_free(graft
);
115 if (!build
->at_each_domain
)
118 build
= isl_ast_build_copy(build
);
119 build
= isl_ast_build_set_executed(build
,
120 isl_union_map_from_map(isl_map_copy(executed
)));
122 return isl_ast_graft_free(graft
);
124 graft
->node
= build
->at_each_domain(graft
->node
,
125 build
, build
->at_each_domain_user
);
126 isl_ast_build_free(build
);
129 graft
= isl_ast_graft_free(graft
);
134 /* Generate a call expression for the single executed
135 * domain element "map" and put a guard around it based on its (simplified)
136 * domain. "executed" is the original inverse schedule from which "map"
137 * has been derived. In particular, "map" is identical to "executed".
138 * "executed" is only used if there is an at_each_domain callback.
140 * At this stage, any pending constraints in the build can no longer
141 * be simplified with respect to any enforced constraints since
142 * the call node does not have any enforced constraints.
143 * Since all pending constraints not covered by any enforced constraints
144 * will be added as a guard to the graft in create_node_scaled,
145 * even in the eliminated case, the pending constraints
146 * can be considered to have been generated by outer constructs.
148 * If the user has set an at_each_domain callback, it is called
149 * on the constructed call expression node.
151 static isl_stat
add_domain(__isl_take isl_map
*executed
,
152 __isl_take isl_map
*map
, struct isl_generate_domain_data
*data
)
154 isl_ast_build
*build
;
155 isl_ast_graft
*graft
;
156 isl_ast_graft_list
*list
;
157 isl_set
*guard
, *pending
;
159 build
= isl_ast_build_copy(data
->build
);
160 pending
= isl_ast_build_get_pending(build
);
161 build
= isl_ast_build_replace_pending_by_guard(build
, pending
);
163 guard
= isl_map_domain(isl_map_copy(map
));
164 guard
= isl_set_compute_divs(guard
);
165 guard
= isl_set_coalesce_preserve(guard
);
166 guard
= isl_set_gist(guard
, isl_ast_build_get_generated(build
));
167 guard
= isl_ast_build_specialize(build
, guard
);
169 graft
= isl_ast_graft_alloc_domain(map
, build
);
170 graft
= at_each_domain(graft
, executed
, build
);
171 isl_ast_build_free(build
);
172 isl_map_free(executed
);
173 graft
= isl_ast_graft_add_guard(graft
, guard
, data
->build
);
175 list
= isl_ast_graft_list_from_ast_graft(graft
);
176 data
->list
= isl_ast_graft_list_concat(data
->list
, list
);
181 /* Generate an AST for a single domain based on
182 * the inverse schedule "executed" and add it to data->list.
184 * If there is more than one domain element associated to the current
185 * schedule "time", then we need to continue the generation process
186 * in generate_non_single_valued.
187 * Note that the inverse schedule being single-valued may depend
188 * on constraints that are only available in the original context
189 * domain specified by the user. We therefore first introduce
190 * some of the constraints of data->build->domain. In particular,
191 * we intersect with a single-disjunct approximation of this set.
192 * We perform this approximation to avoid further splitting up
193 * the executed relation, possibly introducing a disjunctive guard
196 * Otherwise, call add_domain to generate a call expression (with guard) and
197 * to call the at_each_domain callback, if any.
199 * Coalesce the inverse schedule before checking for single-valuedness.
200 * Skip this if the inverse schedule is obviously single-valued.
202 static isl_stat
generate_domain(__isl_take isl_map
*executed
, void *user
)
204 struct isl_generate_domain_data
*data
= user
;
208 domain
= isl_ast_build_get_domain(data
->build
);
209 domain
= isl_set_from_basic_set(isl_set_simple_hull(domain
));
210 executed
= isl_map_intersect_domain(executed
, domain
);
211 empty
= isl_map_is_empty(executed
);
215 isl_map_free(executed
);
219 sv
= isl_map_plain_is_single_valued(executed
);
223 return add_domain(executed
, isl_map_copy(executed
), data
);
225 executed
= isl_map_coalesce(executed
);
226 sv
= isl_map_is_single_valued(executed
);
230 return generate_non_single_valued(executed
, data
);
232 return add_domain(executed
, isl_map_copy(executed
), data
);
234 isl_map_free(executed
);
235 return isl_stat_error
;
238 /* Call build->create_leaf to a create "leaf" node in the AST,
239 * encapsulate the result in an isl_ast_graft and return the result
240 * as a 1-element list.
242 * Note that the node returned by the user may be an entire tree.
244 * Since the node itself cannot enforce any constraints, we turn
245 * all pending constraints into guards and add them to the resulting
246 * graft to ensure that they will be generated.
248 * Before we pass control to the user, we first clear some information
249 * from the build that is (presumbably) only meaningful
250 * for the current code generation.
251 * This includes the create_leaf callback itself, so we make a copy
252 * of the build first.
254 static __isl_give isl_ast_graft_list
*call_create_leaf(
255 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
259 isl_ast_graft
*graft
;
260 isl_ast_build
*user_build
;
262 guard
= isl_ast_build_get_pending(build
);
263 user_build
= isl_ast_build_copy(build
);
264 user_build
= isl_ast_build_replace_pending_by_guard(user_build
,
265 isl_set_copy(guard
));
266 user_build
= isl_ast_build_set_executed(user_build
, executed
);
267 user_build
= isl_ast_build_clear_local_info(user_build
);
271 node
= build
->create_leaf(user_build
, build
->create_leaf_user
);
272 graft
= isl_ast_graft_alloc(node
, build
);
273 graft
= isl_ast_graft_add_guard(graft
, guard
, build
);
274 isl_ast_build_free(build
);
275 return isl_ast_graft_list_from_ast_graft(graft
);
278 static __isl_give isl_ast_graft_list
*build_ast_from_child(
279 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
280 __isl_take isl_union_map
*executed
);
282 /* Generate an AST after having handled the complete schedule
283 * of this call to the code generator or the complete band
284 * if we are generating an AST from a schedule tree.
286 * If we are inside a band node, then move on to the child of the band.
288 * If the user has specified a create_leaf callback, control
289 * is passed to the user in call_create_leaf.
291 * Otherwise, we generate one or more calls for each individual
292 * domain in generate_domain.
294 static __isl_give isl_ast_graft_list
*generate_inner_level(
295 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
298 struct isl_generate_domain_data data
= { build
};
300 if (!build
|| !executed
)
303 if (isl_ast_build_has_schedule_node(build
)) {
304 isl_schedule_node
*node
;
305 node
= isl_ast_build_get_schedule_node(build
);
306 build
= isl_ast_build_reset_schedule_node(build
);
307 return build_ast_from_child(build
, node
, executed
);
310 if (build
->create_leaf
)
311 return call_create_leaf(executed
, build
);
313 ctx
= isl_union_map_get_ctx(executed
);
314 data
.list
= isl_ast_graft_list_alloc(ctx
, 0);
315 if (isl_union_map_foreach_map(executed
, &generate_domain
, &data
) < 0)
316 data
.list
= isl_ast_graft_list_free(data
.list
);
319 error
: data
.list
= NULL
;
320 isl_ast_build_free(build
);
321 isl_union_map_free(executed
);
325 /* Call the before_each_for callback, if requested by the user.
327 static __isl_give isl_ast_node
*before_each_for(__isl_take isl_ast_node
*node
,
328 __isl_keep isl_ast_build
*build
)
333 return isl_ast_node_free(node
);
334 if (!build
->before_each_for
)
336 id
= build
->before_each_for(build
, build
->before_each_for_user
);
337 node
= isl_ast_node_set_annotation(node
, id
);
341 /* Call the after_each_for callback, if requested by the user.
343 static __isl_give isl_ast_graft
*after_each_for(__isl_take isl_ast_graft
*graft
,
344 __isl_keep isl_ast_build
*build
)
346 if (!graft
|| !build
)
347 return isl_ast_graft_free(graft
);
348 if (!build
->after_each_for
)
350 graft
->node
= build
->after_each_for(graft
->node
, build
,
351 build
->after_each_for_user
);
353 return isl_ast_graft_free(graft
);
357 /* Plug in all the know values of the current and outer dimensions
358 * in the domain of "executed". In principle, we only need to plug
359 * in the known value of the current dimension since the values of
360 * outer dimensions have been plugged in already.
361 * However, it turns out to be easier to just plug in all known values.
363 static __isl_give isl_union_map
*plug_in_values(
364 __isl_take isl_union_map
*executed
, __isl_keep isl_ast_build
*build
)
366 return isl_ast_build_substitute_values_union_map_domain(build
,
370 /* Check if the constraint "c" is a lower bound on dimension "pos",
371 * an upper bound, or independent of dimension "pos".
373 static int constraint_type(isl_constraint
*c
, int pos
)
375 if (isl_constraint_is_lower_bound(c
, isl_dim_set
, pos
))
377 if (isl_constraint_is_upper_bound(c
, isl_dim_set
, pos
))
382 /* Compare the types of the constraints "a" and "b",
383 * resulting in constraints that are independent of "depth"
384 * to be sorted before the lower bounds on "depth", which in
385 * turn are sorted before the upper bounds on "depth".
387 static int cmp_constraint(__isl_keep isl_constraint
*a
,
388 __isl_keep isl_constraint
*b
, void *user
)
391 int t1
= constraint_type(a
, *depth
);
392 int t2
= constraint_type(b
, *depth
);
397 /* Extract a lower bound on dimension "pos" from constraint "c".
399 * If the constraint is of the form
403 * then we essentially return
405 * l = ceil(-f(...)/a)
407 * However, if the current dimension is strided, then we need to make
408 * sure that the lower bound we construct is of the form
412 * with f the offset and s the stride.
413 * We therefore compute
415 * f + s * ceil((l - f)/s)
417 static __isl_give isl_aff
*lower_bound(__isl_keep isl_constraint
*c
,
418 int pos
, __isl_keep isl_ast_build
*build
)
422 aff
= isl_constraint_get_bound(c
, isl_dim_set
, pos
);
423 aff
= isl_aff_ceil(aff
);
425 if (isl_ast_build_has_stride(build
, pos
)) {
429 offset
= isl_ast_build_get_offset(build
, pos
);
430 stride
= isl_ast_build_get_stride(build
, pos
);
432 aff
= isl_aff_sub(aff
, isl_aff_copy(offset
));
433 aff
= isl_aff_scale_down_val(aff
, isl_val_copy(stride
));
434 aff
= isl_aff_ceil(aff
);
435 aff
= isl_aff_scale_val(aff
, stride
);
436 aff
= isl_aff_add(aff
, offset
);
439 aff
= isl_ast_build_compute_gist_aff(build
, aff
);
444 /* Return the exact lower bound (or upper bound if "upper" is set)
445 * of "domain" as a piecewise affine expression.
447 * If we are computing a lower bound (of a strided dimension), then
448 * we need to make sure it is of the form
452 * where f is the offset and s is the stride.
453 * We therefore need to include the stride constraint before computing
456 static __isl_give isl_pw_aff
*exact_bound(__isl_keep isl_set
*domain
,
457 __isl_keep isl_ast_build
*build
, int upper
)
462 isl_pw_multi_aff
*pma
;
464 domain
= isl_set_copy(domain
);
466 stride
= isl_ast_build_get_stride_constraint(build
);
467 domain
= isl_set_intersect(domain
, stride
);
469 it_map
= isl_ast_build_map_to_iterator(build
, domain
);
471 pma
= isl_map_lexmax_pw_multi_aff(it_map
);
473 pma
= isl_map_lexmin_pw_multi_aff(it_map
);
474 pa
= isl_pw_multi_aff_get_pw_aff(pma
, 0);
475 isl_pw_multi_aff_free(pma
);
476 pa
= isl_ast_build_compute_gist_pw_aff(build
, pa
);
477 pa
= isl_pw_aff_coalesce(pa
);
482 /* Callback for sorting the isl_pw_aff_list passed to reduce_list and
483 * remove_redundant_lower_bounds.
485 static int reduce_list_cmp(__isl_keep isl_pw_aff
*a
, __isl_keep isl_pw_aff
*b
,
488 return isl_pw_aff_plain_cmp(a
, b
);
491 /* Given a list of lower bounds "list", remove those that are redundant
492 * with respect to the other bounds in "list" and the domain of "build".
494 * We first sort the bounds in the same way as they would be sorted
495 * by set_for_node_expressions so that we can try and remove the last
498 * For a lower bound to be effective, there needs to be at least
499 * one domain element for which it is larger than all other lower bounds.
500 * For each lower bound we therefore intersect the domain with
501 * the conditions that it is larger than all other bounds and
502 * check whether the result is empty. If so, the bound can be removed.
504 static __isl_give isl_pw_aff_list
*remove_redundant_lower_bounds(
505 __isl_take isl_pw_aff_list
*list
, __isl_keep isl_ast_build
*build
)
511 list
= isl_pw_aff_list_sort(list
, &reduce_list_cmp
, NULL
);
513 n
= isl_pw_aff_list_n_pw_aff(list
);
515 return isl_pw_aff_list_free(list
);
519 domain
= isl_ast_build_get_domain(build
);
521 for (i
= n
- 1; i
>= 0; --i
) {
526 domain_i
= isl_set_copy(domain
);
527 pa_i
= isl_pw_aff_list_get_pw_aff(list
, i
);
529 for (j
= 0; j
< n
; ++j
) {
536 pa_j
= isl_pw_aff_list_get_pw_aff(list
, j
);
537 better
= isl_pw_aff_gt_set(isl_pw_aff_copy(pa_i
), pa_j
);
538 domain_i
= isl_set_intersect(domain_i
, better
);
541 empty
= isl_set_is_empty(domain_i
);
543 isl_set_free(domain_i
);
544 isl_pw_aff_free(pa_i
);
550 list
= isl_pw_aff_list_drop(list
, i
, 1);
554 isl_set_free(domain
);
558 isl_set_free(domain
);
559 return isl_pw_aff_list_free(list
);
562 /* Extract a lower bound on dimension "pos" from each constraint
563 * in "constraints" and return the list of lower bounds.
564 * If "constraints" has zero elements, then we extract a lower bound
565 * from "domain" instead.
567 * If the current dimension is strided, then the lower bound
568 * is adjusted by lower_bound to match the stride information.
569 * This modification may make one or more lower bounds redundant
570 * with respect to the other lower bounds. We therefore check
571 * for this condition and remove the redundant lower bounds.
573 static __isl_give isl_pw_aff_list
*lower_bounds(
574 __isl_keep isl_constraint_list
*constraints
, int pos
,
575 __isl_keep isl_set
*domain
, __isl_keep isl_ast_build
*build
)
578 isl_pw_aff_list
*list
;
585 n
= isl_constraint_list_n_constraint(constraints
);
590 pa
= exact_bound(domain
, build
, 0);
591 return isl_pw_aff_list_from_pw_aff(pa
);
594 ctx
= isl_ast_build_get_ctx(build
);
595 list
= isl_pw_aff_list_alloc(ctx
,n
);
597 for (i
= 0; i
< n
; ++i
) {
601 c
= isl_constraint_list_get_constraint(constraints
, i
);
602 aff
= lower_bound(c
, pos
, build
);
603 isl_constraint_free(c
);
604 list
= isl_pw_aff_list_add(list
, isl_pw_aff_from_aff(aff
));
607 if (isl_ast_build_has_stride(build
, pos
))
608 list
= remove_redundant_lower_bounds(list
, build
);
613 /* Extract an upper bound on dimension "pos" from each constraint
614 * in "constraints" and return the list of upper bounds.
615 * If "constraints" has zero elements, then we extract an upper bound
616 * from "domain" instead.
618 static __isl_give isl_pw_aff_list
*upper_bounds(
619 __isl_keep isl_constraint_list
*constraints
, int pos
,
620 __isl_keep isl_set
*domain
, __isl_keep isl_ast_build
*build
)
623 isl_pw_aff_list
*list
;
627 n
= isl_constraint_list_n_constraint(constraints
);
632 pa
= exact_bound(domain
, build
, 1);
633 return isl_pw_aff_list_from_pw_aff(pa
);
636 ctx
= isl_ast_build_get_ctx(build
);
637 list
= isl_pw_aff_list_alloc(ctx
,n
);
639 for (i
= 0; i
< n
; ++i
) {
643 c
= isl_constraint_list_get_constraint(constraints
, i
);
644 aff
= isl_constraint_get_bound(c
, isl_dim_set
, pos
);
645 isl_constraint_free(c
);
646 aff
= isl_aff_floor(aff
);
647 list
= isl_pw_aff_list_add(list
, isl_pw_aff_from_aff(aff
));
653 /* Return an isl_ast_expr that performs the reduction of type "type"
654 * on AST expressions corresponding to the elements in "list".
656 * The list is assumed to contain at least one element.
657 * If the list contains exactly one element, then the returned isl_ast_expr
658 * simply computes that affine expression.
659 * If the list contains more than one element, then we sort it
660 * using a fairly arbitrary but hopefully reasonably stable order.
662 static __isl_give isl_ast_expr
*reduce_list(enum isl_ast_expr_op_type type
,
663 __isl_keep isl_pw_aff_list
*list
, __isl_keep isl_ast_build
*build
)
670 n
= isl_pw_aff_list_n_pw_aff(list
);
675 return isl_ast_build_expr_from_pw_aff_internal(build
,
676 isl_pw_aff_list_get_pw_aff(list
, 0));
678 ctx
= isl_pw_aff_list_get_ctx(list
);
679 expr
= isl_ast_expr_alloc_op(ctx
, type
, n
);
681 list
= isl_pw_aff_list_copy(list
);
682 list
= isl_pw_aff_list_sort(list
, &reduce_list_cmp
, NULL
);
684 return isl_ast_expr_free(expr
);
686 for (i
= 0; i
< n
; ++i
) {
687 isl_ast_expr
*expr_i
;
689 expr_i
= isl_ast_build_expr_from_pw_aff_internal(build
,
690 isl_pw_aff_list_get_pw_aff(list
, i
));
691 expr
= isl_ast_expr_op_add_arg(expr
, expr_i
);
694 isl_pw_aff_list_free(list
);
698 /* Add guards implied by the "generated constraints",
699 * but not (necessarily) enforced by the generated AST to "guard".
700 * In particular, if there is any stride constraints,
701 * then add the guard implied by those constraints.
702 * If we have generated a degenerate loop, then add the guard
703 * implied by "bounds" on the outer dimensions, i.e., the guard
704 * that ensures that the single value actually exists.
705 * Since there may also be guards implied by a combination
706 * of these constraints, we first combine them before
707 * deriving the implied constraints.
709 static __isl_give isl_set
*add_implied_guards(__isl_take isl_set
*guard
,
710 int degenerate
, __isl_keep isl_basic_set
*bounds
,
711 __isl_keep isl_ast_build
*build
)
718 depth
= isl_ast_build_get_depth(build
);
719 has_stride
= isl_ast_build_has_stride(build
, depth
);
720 if (depth
< 0 || has_stride
< 0)
721 return isl_set_free(guard
);
722 if (!has_stride
&& !degenerate
)
725 space
= isl_basic_set_get_space(bounds
);
726 dom
= isl_set_universe(space
);
729 bounds
= isl_basic_set_copy(bounds
);
730 bounds
= isl_basic_set_drop_constraints_not_involving_dims(
731 bounds
, isl_dim_set
, depth
, 1);
732 set
= isl_set_from_basic_set(bounds
);
733 dom
= isl_set_intersect(dom
, set
);
737 set
= isl_ast_build_get_stride_constraint(build
);
738 dom
= isl_set_intersect(dom
, set
);
741 dom
= isl_set_eliminate(dom
, isl_dim_set
, depth
, 1);
742 dom
= isl_ast_build_compute_gist(build
, dom
);
743 guard
= isl_set_intersect(guard
, dom
);
748 /* Update "graft" based on "sub_build" for the degenerate case.
750 * "build" is the build in which graft->node was created
751 * "sub_build" contains information about the current level itself,
752 * including the single value attained.
754 * We set the initialization part of the for loop to the single
755 * value attained by the current dimension.
756 * The increment and condition are not strictly needed as they are known
757 * to be "1" and "iterator <= value" respectively.
759 static __isl_give isl_ast_graft
*refine_degenerate(
760 __isl_take isl_ast_graft
*graft
, __isl_keep isl_ast_build
*build
,
761 __isl_keep isl_ast_build
*sub_build
)
766 if (!graft
|| !sub_build
)
767 return isl_ast_graft_free(graft
);
769 value
= isl_pw_aff_copy(sub_build
->value
);
771 init
= isl_ast_build_expr_from_pw_aff_internal(build
, value
);
772 graft
->node
= isl_ast_node_for_set_init(graft
->node
, init
);
774 return isl_ast_graft_free(graft
);
779 /* Return the intersection of constraints in "list" as a set.
781 static __isl_give isl_set
*intersect_constraints(
782 __isl_keep isl_constraint_list
*list
)
788 n
= isl_constraint_list_n_constraint(list
);
792 isl_die(isl_constraint_list_get_ctx(list
), isl_error_internal
,
793 "expecting at least one constraint", return NULL
);
795 bset
= isl_basic_set_from_constraint(
796 isl_constraint_list_get_constraint(list
, 0));
797 for (i
= 1; i
< n
; ++i
) {
798 isl_basic_set
*bset_i
;
800 bset_i
= isl_basic_set_from_constraint(
801 isl_constraint_list_get_constraint(list
, i
));
802 bset
= isl_basic_set_intersect(bset
, bset_i
);
805 return isl_set_from_basic_set(bset
);
808 /* Compute the constraints on the outer dimensions enforced by
809 * graft->node and add those constraints to graft->enforced,
810 * in case the upper bound is expressed as a set "upper".
812 * In particular, if l(...) is a lower bound in "lower", and
814 * -a i + f(...) >= 0 or a i <= f(...)
816 * is an upper bound ocnstraint on the current dimension i,
817 * then the for loop enforces the constraint
819 * -a l(...) + f(...) >= 0 or a l(...) <= f(...)
821 * We therefore simply take each lower bound in turn, plug it into
822 * the upper bounds and compute the intersection over all lower bounds.
824 * If a lower bound is a rational expression, then
825 * isl_basic_set_preimage_multi_aff will force this rational
826 * expression to have only integer values. However, the loop
827 * itself does not enforce this integrality constraint. We therefore
828 * use the ceil of the lower bounds instead of the lower bounds themselves.
829 * Other constraints will make sure that the for loop is only executed
830 * when each of the lower bounds attains an integral value.
831 * In particular, potentially rational values only occur in
832 * lower_bound if the offset is a (seemingly) rational expression,
833 * but then outer conditions will make sure that this rational expression
834 * only attains integer values.
836 static __isl_give isl_ast_graft
*set_enforced_from_set(
837 __isl_take isl_ast_graft
*graft
,
838 __isl_keep isl_pw_aff_list
*lower
, int pos
, __isl_keep isl_set
*upper
)
841 isl_basic_set
*enforced
;
842 isl_pw_multi_aff
*pma
;
846 n
= isl_pw_aff_list_n_pw_aff(lower
);
848 return isl_ast_graft_free(graft
);
850 space
= isl_set_get_space(upper
);
851 enforced
= isl_basic_set_universe(isl_space_copy(space
));
853 space
= isl_space_map_from_set(space
);
854 pma
= isl_pw_multi_aff_identity(space
);
856 for (i
= 0; i
< n
; ++i
) {
860 isl_pw_multi_aff
*pma_i
;
862 pa
= isl_pw_aff_list_get_pw_aff(lower
, i
);
863 pa
= isl_pw_aff_ceil(pa
);
864 pma_i
= isl_pw_multi_aff_copy(pma
);
865 pma_i
= isl_pw_multi_aff_set_pw_aff(pma_i
, pos
, pa
);
866 enforced_i
= isl_set_copy(upper
);
867 enforced_i
= isl_set_preimage_pw_multi_aff(enforced_i
, pma_i
);
868 hull
= isl_set_simple_hull(enforced_i
);
869 enforced
= isl_basic_set_intersect(enforced
, hull
);
872 isl_pw_multi_aff_free(pma
);
874 graft
= isl_ast_graft_enforce(graft
, enforced
);
879 /* Compute the constraints on the outer dimensions enforced by
880 * graft->node and add those constraints to graft->enforced,
881 * in case the upper bound is expressed as
882 * a list of affine expressions "upper".
884 * The enforced condition is that each lower bound expression is less
885 * than or equal to each upper bound expression.
887 static __isl_give isl_ast_graft
*set_enforced_from_list(
888 __isl_take isl_ast_graft
*graft
,
889 __isl_keep isl_pw_aff_list
*lower
, __isl_keep isl_pw_aff_list
*upper
)
892 isl_basic_set
*enforced
;
894 lower
= isl_pw_aff_list_copy(lower
);
895 upper
= isl_pw_aff_list_copy(upper
);
896 cond
= isl_pw_aff_list_le_set(lower
, upper
);
897 enforced
= isl_set_simple_hull(cond
);
898 graft
= isl_ast_graft_enforce(graft
, enforced
);
903 /* Does "aff" have a negative constant term?
905 static isl_bool
aff_constant_is_negative(__isl_keep isl_set
*set
,
906 __isl_keep isl_aff
*aff
, void *user
)
911 v
= isl_aff_get_constant_val(aff
);
912 is_neg
= isl_val_is_neg(v
);
918 /* Does "pa" have a negative constant term over its entire domain?
920 static isl_bool
pw_aff_constant_is_negative(__isl_keep isl_pw_aff
*pa
,
923 return isl_pw_aff_every_piece(pa
, &aff_constant_is_negative
, NULL
);
926 /* Does each element in "list" have a negative constant term?
928 static int list_constant_is_negative(__isl_keep isl_pw_aff_list
*list
)
930 return isl_pw_aff_list_every(list
, &pw_aff_constant_is_negative
, NULL
);
933 /* Add 1 to each of the elements in "list", where each of these elements
934 * is defined over the internal schedule space of "build".
936 static __isl_give isl_pw_aff_list
*list_add_one(
937 __isl_take isl_pw_aff_list
*list
, __isl_keep isl_ast_build
*build
)
945 n
= isl_pw_aff_list_n_pw_aff(list
);
947 return isl_pw_aff_list_free(list
);
949 space
= isl_ast_build_get_space(build
, 1);
950 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
951 aff
= isl_aff_add_constant_si(aff
, 1);
952 one
= isl_pw_aff_from_aff(aff
);
954 for (i
= 0; i
< n
; ++i
) {
956 pa
= isl_pw_aff_list_get_pw_aff(list
, i
);
957 pa
= isl_pw_aff_add(pa
, isl_pw_aff_copy(one
));
958 list
= isl_pw_aff_list_set_pw_aff(list
, i
, pa
);
961 isl_pw_aff_free(one
);
966 /* Set the condition part of the for node graft->node in case
967 * the upper bound is represented as a list of piecewise affine expressions.
969 * In particular, set the condition to
971 * iterator <= min(list of upper bounds)
973 * If each of the upper bounds has a negative constant term, then
974 * set the condition to
976 * iterator < min(list of (upper bound + 1)s)
979 static __isl_give isl_ast_graft
*set_for_cond_from_list(
980 __isl_take isl_ast_graft
*graft
, __isl_keep isl_pw_aff_list
*list
,
981 __isl_keep isl_ast_build
*build
)
984 isl_ast_expr
*bound
, *iterator
, *cond
;
985 enum isl_ast_expr_op_type type
= isl_ast_expr_op_le
;
988 return isl_ast_graft_free(graft
);
990 neg
= list_constant_is_negative(list
);
992 return isl_ast_graft_free(graft
);
993 list
= isl_pw_aff_list_copy(list
);
995 list
= list_add_one(list
, build
);
996 type
= isl_ast_expr_op_lt
;
999 bound
= reduce_list(isl_ast_expr_op_min
, list
, build
);
1000 iterator
= isl_ast_expr_copy(graft
->node
->u
.f
.iterator
);
1001 cond
= isl_ast_expr_alloc_binary(type
, iterator
, bound
);
1002 graft
->node
= isl_ast_node_for_set_cond(graft
->node
, cond
);
1004 isl_pw_aff_list_free(list
);
1006 return isl_ast_graft_free(graft
);
1010 /* Set the condition part of the for node graft->node in case
1011 * the upper bound is represented as a set.
1013 static __isl_give isl_ast_graft
*set_for_cond_from_set(
1014 __isl_take isl_ast_graft
*graft
, __isl_keep isl_set
*set
,
1015 __isl_keep isl_ast_build
*build
)
1022 cond
= isl_ast_build_expr_from_set_internal(build
, isl_set_copy(set
));
1023 graft
->node
= isl_ast_node_for_set_cond(graft
->node
, cond
);
1025 return isl_ast_graft_free(graft
);
1029 /* Construct an isl_ast_expr for the increment (i.e., stride) of
1030 * the current dimension.
1032 static __isl_give isl_ast_expr
*for_inc(__isl_keep isl_ast_build
*build
)
1038 depth
= isl_ast_build_get_depth(build
);
1041 ctx
= isl_ast_build_get_ctx(build
);
1043 if (!isl_ast_build_has_stride(build
, depth
))
1044 return isl_ast_expr_alloc_int_si(ctx
, 1);
1046 v
= isl_ast_build_get_stride(build
, depth
);
1047 return isl_ast_expr_from_val(v
);
1050 /* Should we express the loop condition as
1052 * iterator <= min(list of upper bounds)
1054 * or as a conjunction of constraints?
1056 * The first is constructed from a list of upper bounds.
1057 * The second is constructed from a set.
1059 * If there are no upper bounds in "constraints", then this could mean
1060 * that "domain" simply doesn't have an upper bound or that we didn't
1061 * pick any upper bound. In the first case, we want to generate the
1062 * loop condition as a(n empty) conjunction of constraints
1063 * In the second case, we will compute
1064 * a single upper bound from "domain" and so we use the list form.
1066 * If there are upper bounds in "constraints",
1067 * then we use the list form iff the atomic_upper_bound option is set.
1069 static int use_upper_bound_list(isl_ctx
*ctx
, int n_upper
,
1070 __isl_keep isl_set
*domain
, int depth
)
1073 return isl_options_get_ast_build_atomic_upper_bound(ctx
);
1075 return isl_set_dim_has_upper_bound(domain
, isl_dim_set
, depth
);
1078 /* Fill in the expressions of the for node in graft->node.
1081 * - set the initialization part of the loop to the maximum of the lower bounds
1082 * - extract the increment from the stride of the current dimension
1083 * - construct the for condition either based on a list of upper bounds
1084 * or on a set of upper bound constraints.
1086 static __isl_give isl_ast_graft
*set_for_node_expressions(
1087 __isl_take isl_ast_graft
*graft
, __isl_keep isl_pw_aff_list
*lower
,
1088 int use_list
, __isl_keep isl_pw_aff_list
*upper_list
,
1089 __isl_keep isl_set
*upper_set
, __isl_keep isl_ast_build
*build
)
1096 init
= reduce_list(isl_ast_expr_op_max
, lower
, build
);
1097 graft
->node
= isl_ast_node_for_set_init(graft
->node
, init
);
1098 graft
->node
= isl_ast_node_for_set_inc(graft
->node
, for_inc(build
));
1101 graft
= isl_ast_graft_free(graft
);
1104 graft
= set_for_cond_from_list(graft
, upper_list
, build
);
1106 graft
= set_for_cond_from_set(graft
, upper_set
, build
);
1111 /* Update "graft" based on "bounds" and "domain" for the generic,
1112 * non-degenerate, case.
1114 * "c_lower" and "c_upper" contain the lower and upper bounds
1115 * that the loop node should express.
1116 * "domain" is the subset of the intersection of the constraints
1117 * for which some code is executed.
1119 * There may be zero lower bounds or zero upper bounds in "constraints"
1120 * in case the list of constraints was created
1121 * based on the atomic option or based on separation with explicit bounds.
1122 * In that case, we use "domain" to derive lower and/or upper bounds.
1124 * We first compute a list of one or more lower bounds.
1126 * Then we decide if we want to express the condition as
1128 * iterator <= min(list of upper bounds)
1130 * or as a conjunction of constraints.
1132 * The set of enforced constraints is then computed either based on
1133 * a list of upper bounds or on a set of upper bound constraints.
1134 * We do not compute any enforced constraints if we were forced
1135 * to compute a lower or upper bound using exact_bound. The domains
1136 * of the resulting expressions may imply some bounds on outer dimensions
1137 * that we do not want to appear in the enforced constraints since
1138 * they are not actually enforced by the corresponding code.
1140 * Finally, we fill in the expressions of the for node.
1142 static __isl_give isl_ast_graft
*refine_generic_bounds(
1143 __isl_take isl_ast_graft
*graft
,
1144 __isl_take isl_constraint_list
*c_lower
,
1145 __isl_take isl_constraint_list
*c_upper
,
1146 __isl_keep isl_set
*domain
, __isl_keep isl_ast_build
*build
)
1150 isl_pw_aff_list
*lower
;
1152 isl_set
*upper_set
= NULL
;
1153 isl_pw_aff_list
*upper_list
= NULL
;
1154 isl_size n_lower
, n_upper
;
1156 depth
= isl_ast_build_get_depth(build
);
1157 if (!graft
|| !c_lower
|| !c_upper
|| depth
< 0)
1160 ctx
= isl_ast_graft_get_ctx(graft
);
1162 n_lower
= isl_constraint_list_n_constraint(c_lower
);
1163 n_upper
= isl_constraint_list_n_constraint(c_upper
);
1164 if (n_lower
< 0 || n_upper
< 0)
1167 use_list
= use_upper_bound_list(ctx
, n_upper
, domain
, depth
);
1169 lower
= lower_bounds(c_lower
, depth
, domain
, build
);
1172 upper_list
= upper_bounds(c_upper
, depth
, domain
, build
);
1173 else if (n_upper
> 0)
1174 upper_set
= intersect_constraints(c_upper
);
1176 upper_set
= isl_set_universe(isl_set_get_space(domain
));
1178 if (n_lower
== 0 || n_upper
== 0)
1181 graft
= set_enforced_from_list(graft
, lower
, upper_list
);
1183 graft
= set_enforced_from_set(graft
, lower
, depth
, upper_set
);
1185 graft
= set_for_node_expressions(graft
, lower
, use_list
, upper_list
,
1188 isl_pw_aff_list_free(lower
);
1189 isl_pw_aff_list_free(upper_list
);
1190 isl_set_free(upper_set
);
1191 isl_constraint_list_free(c_lower
);
1192 isl_constraint_list_free(c_upper
);
1196 isl_constraint_list_free(c_lower
);
1197 isl_constraint_list_free(c_upper
);
1198 return isl_ast_graft_free(graft
);
1201 /* Internal data structure used inside count_constraints to keep
1202 * track of the number of constraints that are independent of dimension "pos",
1203 * the lower bounds in "pos" and the upper bounds in "pos".
1205 struct isl_ast_count_constraints_data
{
1213 /* Increment data->n_indep, data->lower or data->upper depending
1214 * on whether "c" is independent of dimensions data->pos,
1215 * a lower bound or an upper bound.
1217 static isl_stat
count_constraints(__isl_take isl_constraint
*c
, void *user
)
1219 struct isl_ast_count_constraints_data
*data
= user
;
1221 if (isl_constraint_is_lower_bound(c
, isl_dim_set
, data
->pos
))
1223 else if (isl_constraint_is_upper_bound(c
, isl_dim_set
, data
->pos
))
1228 isl_constraint_free(c
);
1233 /* Update "graft" based on "bounds" and "domain" for the generic,
1234 * non-degenerate, case.
1236 * "list" respresent the list of bounds that need to be encoded by
1237 * the for loop. Only the constraints that involve the iterator
1238 * are relevant here. The other constraints are taken care of by
1239 * the caller and are included in the generated constraints of "build".
1240 * "domain" is the subset of the intersection of the constraints
1241 * for which some code is executed.
1242 * "build" is the build in which graft->node was created.
1244 * We separate lower bounds, upper bounds and constraints that
1245 * are independent of the loop iterator.
1247 * The actual for loop bounds are generated in refine_generic_bounds.
1249 static __isl_give isl_ast_graft
*refine_generic_split(
1250 __isl_take isl_ast_graft
*graft
, __isl_take isl_constraint_list
*list
,
1251 __isl_keep isl_set
*domain
, __isl_keep isl_ast_build
*build
)
1253 struct isl_ast_count_constraints_data data
;
1255 isl_constraint_list
*lower
;
1256 isl_constraint_list
*upper
;
1258 depth
= isl_ast_build_get_depth(build
);
1260 list
= isl_constraint_list_free(list
);
1262 return isl_ast_graft_free(graft
);
1266 list
= isl_constraint_list_sort(list
, &cmp_constraint
, &data
.pos
);
1268 return isl_ast_graft_free(graft
);
1270 data
.n_indep
= data
.n_lower
= data
.n_upper
= 0;
1271 if (isl_constraint_list_foreach(list
, &count_constraints
, &data
) < 0) {
1272 isl_constraint_list_free(list
);
1273 return isl_ast_graft_free(graft
);
1276 lower
= isl_constraint_list_drop(list
, 0, data
.n_indep
);
1277 upper
= isl_constraint_list_copy(lower
);
1278 lower
= isl_constraint_list_drop(lower
, data
.n_lower
, data
.n_upper
);
1279 upper
= isl_constraint_list_drop(upper
, 0, data
.n_lower
);
1281 return refine_generic_bounds(graft
, lower
, upper
, domain
, build
);
1284 /* Update "graft" based on "bounds" and "domain" for the generic,
1285 * non-degenerate, case.
1287 * "bounds" respresent the bounds that need to be encoded by
1288 * the for loop (or a guard around the for loop).
1289 * "domain" is the subset of "bounds" for which some code is executed.
1290 * "build" is the build in which graft->node was created.
1292 * We break up "bounds" into a list of constraints and continue with
1293 * refine_generic_split.
1295 static __isl_give isl_ast_graft
*refine_generic(
1296 __isl_take isl_ast_graft
*graft
,
1297 __isl_keep isl_basic_set
*bounds
, __isl_keep isl_set
*domain
,
1298 __isl_keep isl_ast_build
*build
)
1300 isl_constraint_list
*list
;
1302 if (!build
|| !graft
)
1303 return isl_ast_graft_free(graft
);
1305 list
= isl_basic_set_get_constraint_list(bounds
);
1307 graft
= refine_generic_split(graft
, list
, domain
, build
);
1312 /* Create a for node for the current level.
1314 * Mark the for node degenerate if "degenerate" is set.
1316 static __isl_give isl_ast_node
*create_for(__isl_keep isl_ast_build
*build
,
1323 depth
= isl_ast_build_get_depth(build
);
1327 id
= isl_ast_build_get_iterator_id(build
, depth
);
1328 node
= isl_ast_node_alloc_for(id
);
1330 node
= isl_ast_node_for_mark_degenerate(node
);
1335 /* If the ast_build_exploit_nested_bounds option is set, then return
1336 * the constraints enforced by all elements in "list".
1337 * Otherwise, return the universe.
1339 static __isl_give isl_basic_set
*extract_shared_enforced(
1340 __isl_keep isl_ast_graft_list
*list
, __isl_keep isl_ast_build
*build
)
1348 ctx
= isl_ast_graft_list_get_ctx(list
);
1349 if (isl_options_get_ast_build_exploit_nested_bounds(ctx
))
1350 return isl_ast_graft_list_extract_shared_enforced(list
, build
);
1352 space
= isl_ast_build_get_space(build
, 1);
1353 return isl_basic_set_universe(space
);
1356 /* Return the pending constraints of "build" that are not already taken
1357 * care of (by a combination of "enforced" and the generated constraints
1360 static __isl_give isl_set
*extract_pending(__isl_keep isl_ast_build
*build
,
1361 __isl_keep isl_basic_set
*enforced
)
1363 isl_set
*guard
, *context
;
1365 guard
= isl_ast_build_get_pending(build
);
1366 context
= isl_set_from_basic_set(isl_basic_set_copy(enforced
));
1367 context
= isl_set_intersect(context
,
1368 isl_ast_build_get_generated(build
));
1369 return isl_set_gist(guard
, context
);
1372 /* Create an AST node for the current dimension based on
1373 * the schedule domain "bounds" and return the node encapsulated
1374 * in an isl_ast_graft.
1376 * "executed" is the current inverse schedule, taking into account
1377 * the bounds in "bounds"
1378 * "domain" is the domain of "executed", with inner dimensions projected out.
1379 * It may be a strict subset of "bounds" in case "bounds" was created
1380 * based on the atomic option or based on separation with explicit bounds.
1382 * "domain" may satisfy additional equalities that result
1383 * from intersecting "executed" with "bounds" in add_node.
1384 * It may also satisfy some global constraints that were dropped out because
1385 * we performed separation with explicit bounds.
1386 * The very first step is then to copy these constraints to "bounds".
1388 * Since we may be calling before_each_for and after_each_for
1389 * callbacks, we record the current inverse schedule in the build.
1391 * We consider three builds,
1392 * "build" is the one in which the current level is created,
1393 * "body_build" is the build in which the next level is created,
1394 * "sub_build" is essentially the same as "body_build", except that
1395 * the depth has not been increased yet.
1397 * "build" already contains information (in strides and offsets)
1398 * about the strides at the current level, but this information is not
1399 * reflected in the build->domain.
1400 * We first add this information and the "bounds" to the sub_build->domain.
1401 * isl_ast_build_set_loop_bounds adds the stride information and
1402 * checks whether the current dimension attains
1403 * only a single value and whether this single value can be represented using
1404 * a single affine expression.
1405 * In the first case, the current level is considered "degenerate".
1406 * In the second, sub-case, the current level is considered "eliminated".
1407 * Eliminated levels don't need to be reflected in the AST since we can
1408 * simply plug in the affine expression. For degenerate, but non-eliminated,
1409 * levels, we do introduce a for node, but mark is as degenerate so that
1410 * it can be printed as an assignment of the single value to the loop
1413 * If the current level is eliminated, we explicitly plug in the value
1414 * for the current level found by isl_ast_build_set_loop_bounds in the
1415 * inverse schedule. This ensures that if we are working on a slice
1416 * of the domain based on information available in the inverse schedule
1417 * and the build domain, that then this information is also reflected
1418 * in the inverse schedule. This operation also eliminates the current
1419 * dimension from the inverse schedule making sure no inner dimensions depend
1420 * on the current dimension. Otherwise, we create a for node, marking
1421 * it degenerate if appropriate. The initial for node is still incomplete
1422 * and will be completed in either refine_degenerate or refine_generic.
1424 * We then generate a sequence of grafts for the next level,
1425 * create a surrounding graft for the current level and insert
1426 * the for node we created (if the current level is not eliminated).
1427 * Before creating a graft for the current level, we first extract
1428 * hoistable constraints from the child guards and combine them
1429 * with the pending constraints in the build. These constraints
1430 * are used to simplify the child guards and then added to the guard
1431 * of the current graft to ensure that they will be generated.
1432 * If the hoisted guard is a disjunction, then we use it directly
1433 * to gist the guards on the children before intersect it with the
1434 * pending constraints. We do so because this disjunction is typically
1435 * identical to the guards on the children such that these guards
1436 * can be effectively removed completely. After the intersection,
1437 * the gist operation would have a harder time figuring this out.
1439 * Finally, we set the bounds of the for loop in either
1440 * refine_degenerate or refine_generic.
1441 * We do so in a context where the pending constraints of the build
1442 * have been replaced by the guard of the current graft.
1444 static __isl_give isl_ast_graft
*create_node_scaled(
1445 __isl_take isl_union_map
*executed
,
1446 __isl_take isl_basic_set
*bounds
, __isl_take isl_set
*domain
,
1447 __isl_take isl_ast_build
*build
)
1451 isl_bool eliminated
;
1453 isl_basic_set
*hull
;
1454 isl_basic_set
*enforced
;
1455 isl_set
*guard
, *hoisted
;
1456 isl_ast_node
*node
= NULL
;
1457 isl_ast_graft
*graft
;
1458 isl_ast_graft_list
*children
;
1459 isl_ast_build
*sub_build
;
1460 isl_ast_build
*body_build
;
1462 domain
= isl_ast_build_eliminate_divs(build
, domain
);
1463 domain
= isl_set_detect_equalities(domain
);
1464 hull
= isl_set_unshifted_simple_hull(isl_set_copy(domain
));
1465 bounds
= isl_basic_set_intersect(bounds
, hull
);
1466 build
= isl_ast_build_set_executed(build
, isl_union_map_copy(executed
));
1468 depth
= isl_ast_build_get_depth(build
);
1470 build
= isl_ast_build_free(build
);
1471 sub_build
= isl_ast_build_copy(build
);
1472 bounds
= isl_basic_set_remove_redundancies(bounds
);
1473 bounds
= isl_ast_build_specialize_basic_set(sub_build
, bounds
);
1474 sub_build
= isl_ast_build_set_loop_bounds(sub_build
,
1475 isl_basic_set_copy(bounds
));
1476 degenerate
= isl_ast_build_has_value(sub_build
);
1477 eliminated
= isl_ast_build_has_affine_value(sub_build
, depth
);
1478 if (degenerate
< 0 || eliminated
< 0)
1479 executed
= isl_union_map_free(executed
);
1481 bounds
= isl_ast_build_compute_gist_basic_set(build
, bounds
);
1482 sub_build
= isl_ast_build_set_pending_generated(sub_build
,
1483 isl_basic_set_copy(bounds
));
1485 executed
= plug_in_values(executed
, sub_build
);
1487 node
= create_for(build
, degenerate
);
1489 body_build
= isl_ast_build_copy(sub_build
);
1490 body_build
= isl_ast_build_increase_depth(body_build
);
1492 node
= before_each_for(node
, body_build
);
1493 children
= generate_next_level(executed
,
1494 isl_ast_build_copy(body_build
));
1496 enforced
= extract_shared_enforced(children
, build
);
1497 guard
= extract_pending(sub_build
, enforced
);
1498 hoisted
= isl_ast_graft_list_extract_hoistable_guard(children
, build
);
1499 n
= isl_set_n_basic_set(hoisted
);
1501 children
= isl_ast_graft_list_free(children
);
1503 children
= isl_ast_graft_list_gist_guards(children
,
1504 isl_set_copy(hoisted
));
1505 guard
= isl_set_intersect(guard
, hoisted
);
1507 guard
= add_implied_guards(guard
, degenerate
, bounds
, build
);
1509 graft
= isl_ast_graft_alloc_from_children(children
,
1510 isl_set_copy(guard
), enforced
, build
, sub_build
);
1513 isl_ast_build
*for_build
;
1515 graft
= isl_ast_graft_insert_for(graft
, node
);
1516 for_build
= isl_ast_build_copy(build
);
1517 for_build
= isl_ast_build_replace_pending_by_guard(for_build
,
1518 isl_set_copy(guard
));
1520 graft
= refine_degenerate(graft
, for_build
, sub_build
);
1522 graft
= refine_generic(graft
, bounds
,
1524 isl_ast_build_free(for_build
);
1526 isl_set_free(guard
);
1528 graft
= after_each_for(graft
, body_build
);
1530 isl_ast_build_free(body_build
);
1531 isl_ast_build_free(sub_build
);
1532 isl_ast_build_free(build
);
1533 isl_basic_set_free(bounds
);
1534 isl_set_free(domain
);
1539 /* Internal data structure for checking if all constraints involving
1540 * the input dimension "depth" are such that the other coefficients
1541 * are multiples of "m", reducing "m" if they are not.
1542 * If "m" is reduced all the way down to "1", then the check has failed
1543 * and we break out of the iteration.
1545 struct isl_check_scaled_data
{
1550 /* If constraint "c" involves the input dimension data->depth,
1551 * then make sure that all the other coefficients are multiples of data->m,
1552 * reducing data->m if needed.
1553 * Break out of the iteration if data->m has become equal to "1".
1555 static isl_stat
constraint_check_scaled(__isl_take isl_constraint
*c
,
1558 struct isl_check_scaled_data
*data
= user
;
1561 enum isl_dim_type t
[] = { isl_dim_param
, isl_dim_in
, isl_dim_out
,
1564 if (!isl_constraint_involves_dims(c
, isl_dim_in
, data
->depth
, 1)) {
1565 isl_constraint_free(c
);
1569 for (i
= 0; i
< 4; ++i
) {
1570 n
= isl_constraint_dim(c
, t
[i
]);
1573 for (j
= 0; j
< n
; ++j
) {
1576 if (t
[i
] == isl_dim_in
&& j
== data
->depth
)
1578 if (!isl_constraint_involves_dims(c
, t
[i
], j
, 1))
1580 d
= isl_constraint_get_coefficient_val(c
, t
[i
], j
);
1581 data
->m
= isl_val_gcd(data
->m
, d
);
1582 if (isl_val_is_one(data
->m
))
1589 isl_constraint_free(c
);
1591 return i
< 4 ? isl_stat_error
: isl_stat_ok
;
1594 /* For each constraint of "bmap" that involves the input dimension data->depth,
1595 * make sure that all the other coefficients are multiples of data->m,
1596 * reducing data->m if needed.
1597 * Break out of the iteration if data->m has become equal to "1".
1599 static isl_stat
basic_map_check_scaled(__isl_take isl_basic_map
*bmap
,
1604 r
= isl_basic_map_foreach_constraint(bmap
,
1605 &constraint_check_scaled
, user
);
1606 isl_basic_map_free(bmap
);
1611 /* For each constraint of "map" that involves the input dimension data->depth,
1612 * make sure that all the other coefficients are multiples of data->m,
1613 * reducing data->m if needed.
1614 * Break out of the iteration if data->m has become equal to "1".
1616 static isl_stat
map_check_scaled(__isl_take isl_map
*map
, void *user
)
1620 r
= isl_map_foreach_basic_map(map
, &basic_map_check_scaled
, user
);
1626 /* Create an AST node for the current dimension based on
1627 * the schedule domain "bounds" and return the node encapsulated
1628 * in an isl_ast_graft.
1630 * "executed" is the current inverse schedule, taking into account
1631 * the bounds in "bounds"
1632 * "domain" is the domain of "executed", with inner dimensions projected out.
1635 * Before moving on to the actual AST node construction in create_node_scaled,
1636 * we first check if the current dimension is strided and if we can scale
1637 * down this stride. Note that we only do this if the ast_build_scale_strides
1640 * In particular, let the current dimension take on values
1644 * with a an integer. We check if we can find an integer m that (obviously)
1645 * divides both f and s.
1647 * If so, we check if the current dimension only appears in constraints
1648 * where the coefficients of the other variables are multiples of m.
1649 * We perform this extra check to avoid the risk of introducing
1650 * divisions by scaling down the current dimension.
1652 * If so, we scale the current dimension down by a factor of m.
1653 * That is, we plug in
1657 * Note that in principle we could always scale down strided loops
1662 * but this may result in i' taking on larger values than the original i,
1663 * due to the shift by "f".
1664 * By constrast, the scaling in (1) can only reduce the (absolute) value "i".
1666 static __isl_give isl_ast_graft
*create_node(__isl_take isl_union_map
*executed
,
1667 __isl_take isl_basic_set
*bounds
, __isl_take isl_set
*domain
,
1668 __isl_take isl_ast_build
*build
)
1670 struct isl_check_scaled_data data
;
1676 ctx
= isl_ast_build_get_ctx(build
);
1677 if (!isl_options_get_ast_build_scale_strides(ctx
))
1678 return create_node_scaled(executed
, bounds
, domain
, build
);
1680 depth
= isl_ast_build_get_depth(build
);
1682 build
= isl_ast_build_free(build
);
1684 if (!isl_ast_build_has_stride(build
, data
.depth
))
1685 return create_node_scaled(executed
, bounds
, domain
, build
);
1687 offset
= isl_ast_build_get_offset(build
, data
.depth
);
1688 data
.m
= isl_ast_build_get_stride(build
, data
.depth
);
1690 offset
= isl_aff_free(offset
);
1691 offset
= isl_aff_scale_down_val(offset
, isl_val_copy(data
.m
));
1692 d
= isl_aff_get_denominator_val(offset
);
1694 executed
= isl_union_map_free(executed
);
1696 if (executed
&& isl_val_is_divisible_by(data
.m
, d
))
1697 data
.m
= isl_val_div(data
.m
, d
);
1699 data
.m
= isl_val_set_si(data
.m
, 1);
1703 if (!isl_val_is_one(data
.m
)) {
1704 if (isl_union_map_foreach_map(executed
, &map_check_scaled
,
1706 !isl_val_is_one(data
.m
))
1707 executed
= isl_union_map_free(executed
);
1710 if (!isl_val_is_one(data
.m
)) {
1715 isl_union_map
*umap
;
1717 space
= isl_ast_build_get_space(build
, 1);
1718 space
= isl_space_map_from_set(space
);
1719 ma
= isl_multi_aff_identity(space
);
1720 aff
= isl_multi_aff_get_aff(ma
, data
.depth
);
1721 aff
= isl_aff_scale_val(aff
, isl_val_copy(data
.m
));
1722 ma
= isl_multi_aff_set_aff(ma
, data
.depth
, aff
);
1724 bounds
= isl_basic_set_preimage_multi_aff(bounds
,
1725 isl_multi_aff_copy(ma
));
1726 domain
= isl_set_preimage_multi_aff(domain
,
1727 isl_multi_aff_copy(ma
));
1728 map
= isl_map_reverse(isl_map_from_multi_aff(ma
));
1729 umap
= isl_union_map_from_map(map
);
1730 executed
= isl_union_map_apply_domain(executed
,
1731 isl_union_map_copy(umap
));
1732 build
= isl_ast_build_scale_down(build
, isl_val_copy(data
.m
),
1735 isl_aff_free(offset
);
1736 isl_val_free(data
.m
);
1738 return create_node_scaled(executed
, bounds
, domain
, build
);
1741 /* Add the basic set to the list that "user" points to.
1743 static isl_stat
collect_basic_set(__isl_take isl_basic_set
*bset
, void *user
)
1745 isl_basic_set_list
**list
= user
;
1747 *list
= isl_basic_set_list_add(*list
, bset
);
1752 /* Extract the basic sets of "set" and collect them in an isl_basic_set_list.
1754 static __isl_give isl_basic_set_list
*isl_basic_set_list_from_set(
1755 __isl_take isl_set
*set
)
1759 isl_basic_set_list
*list
;
1761 n
= isl_set_n_basic_set(set
);
1763 set
= isl_set_free(set
);
1767 ctx
= isl_set_get_ctx(set
);
1769 list
= isl_basic_set_list_alloc(ctx
, n
);
1770 if (isl_set_foreach_basic_set(set
, &collect_basic_set
, &list
) < 0)
1771 list
= isl_basic_set_list_free(list
);
1777 /* Generate code for the schedule domain "bounds"
1778 * and add the result to "list".
1780 * We mainly detect strides here and check if the bounds do not
1781 * conflict with the current build domain
1782 * and then pass over control to create_node.
1784 * "bounds" reflects the bounds on the current dimension and possibly
1785 * some extra conditions on outer dimensions.
1786 * It does not, however, include any divs involving the current dimension,
1787 * so it does not capture any stride constraints.
1788 * We therefore need to compute that part of the schedule domain that
1789 * intersects with "bounds" and derive the strides from the result.
1791 static __isl_give isl_ast_graft_list
*add_node(
1792 __isl_take isl_ast_graft_list
*list
, __isl_take isl_union_map
*executed
,
1793 __isl_take isl_basic_set
*bounds
, __isl_take isl_ast_build
*build
)
1795 isl_ast_graft
*graft
;
1796 isl_set
*domain
= NULL
;
1797 isl_union_set
*uset
;
1798 int empty
, disjoint
;
1800 uset
= isl_union_set_from_basic_set(isl_basic_set_copy(bounds
));
1801 executed
= isl_union_map_intersect_domain(executed
, uset
);
1802 empty
= isl_union_map_is_empty(executed
);
1808 uset
= isl_union_map_domain(isl_union_map_copy(executed
));
1809 domain
= isl_set_from_union_set(uset
);
1810 domain
= isl_ast_build_specialize(build
, domain
);
1812 domain
= isl_set_compute_divs(domain
);
1813 domain
= isl_ast_build_eliminate_inner(build
, domain
);
1814 disjoint
= isl_set_is_disjoint(domain
, build
->domain
);
1820 build
= isl_ast_build_detect_strides(build
, isl_set_copy(domain
));
1822 graft
= create_node(executed
, bounds
, domain
,
1823 isl_ast_build_copy(build
));
1824 list
= isl_ast_graft_list_add(list
, graft
);
1825 isl_ast_build_free(build
);
1828 list
= isl_ast_graft_list_free(list
);
1830 isl_set_free(domain
);
1831 isl_basic_set_free(bounds
);
1832 isl_union_map_free(executed
);
1833 isl_ast_build_free(build
);
1837 /* Does any element of i follow or coincide with any element of j
1838 * at the current depth for equal values of the outer dimensions?
1840 static isl_bool
domain_follows_at_depth(__isl_keep isl_basic_set
*i
,
1841 __isl_keep isl_basic_set
*j
, void *user
)
1843 int depth
= *(int *) user
;
1844 isl_basic_map
*test
;
1848 test
= isl_basic_map_from_domain_and_range(isl_basic_set_copy(i
),
1849 isl_basic_set_copy(j
));
1850 for (l
= 0; l
< depth
; ++l
)
1851 test
= isl_basic_map_equate(test
, isl_dim_in
, l
,
1853 test
= isl_basic_map_order_ge(test
, isl_dim_in
, depth
,
1854 isl_dim_out
, depth
);
1855 empty
= isl_basic_map_is_empty(test
);
1856 isl_basic_map_free(test
);
1858 return isl_bool_not(empty
);
1861 /* Split up each element of "list" into a part that is related to "bset"
1862 * according to "gt" and a part that is not.
1863 * Return a list that consist of "bset" and all the pieces.
1865 static __isl_give isl_basic_set_list
*add_split_on(
1866 __isl_take isl_basic_set_list
*list
, __isl_take isl_basic_set
*bset
,
1867 __isl_keep isl_basic_map
*gt
)
1871 isl_basic_set_list
*res
;
1873 n
= isl_basic_set_list_n_basic_set(list
);
1875 bset
= isl_basic_set_free(bset
);
1877 gt
= isl_basic_map_copy(gt
);
1878 gt
= isl_basic_map_intersect_domain(gt
, isl_basic_set_copy(bset
));
1879 res
= isl_basic_set_list_from_basic_set(bset
);
1880 for (i
= 0; res
&& i
< n
; ++i
) {
1881 isl_basic_set
*bset
;
1882 isl_set
*set1
, *set2
;
1883 isl_basic_map
*bmap
;
1886 bset
= isl_basic_set_list_get_basic_set(list
, i
);
1887 bmap
= isl_basic_map_copy(gt
);
1888 bmap
= isl_basic_map_intersect_range(bmap
, bset
);
1889 bset
= isl_basic_map_range(bmap
);
1890 empty
= isl_basic_set_is_empty(bset
);
1892 res
= isl_basic_set_list_free(res
);
1894 isl_basic_set_free(bset
);
1895 bset
= isl_basic_set_list_get_basic_set(list
, i
);
1896 res
= isl_basic_set_list_add(res
, bset
);
1900 res
= isl_basic_set_list_add(res
, isl_basic_set_copy(bset
));
1901 set1
= isl_set_from_basic_set(bset
);
1902 bset
= isl_basic_set_list_get_basic_set(list
, i
);
1903 set2
= isl_set_from_basic_set(bset
);
1904 set1
= isl_set_subtract(set2
, set1
);
1905 set1
= isl_set_make_disjoint(set1
);
1907 res
= isl_basic_set_list_concat(res
,
1908 isl_basic_set_list_from_set(set1
));
1910 isl_basic_map_free(gt
);
1911 isl_basic_set_list_free(list
);
1915 static __isl_give isl_ast_graft_list
*generate_sorted_domains(
1916 __isl_keep isl_basic_set_list
*domain_list
,
1917 __isl_keep isl_union_map
*executed
,
1918 __isl_keep isl_ast_build
*build
);
1920 /* Internal data structure for add_nodes.
1922 * "executed" and "build" are extra arguments to be passed to add_node.
1923 * "list" collects the results.
1925 struct isl_add_nodes_data
{
1926 isl_union_map
*executed
;
1927 isl_ast_build
*build
;
1929 isl_ast_graft_list
*list
;
1932 /* Generate code for the schedule domains in "scc"
1933 * and add the results to "list".
1935 * The domains in "scc" form a strongly connected component in the ordering.
1936 * If the number of domains in "scc" is larger than 1, then this means
1937 * that we cannot determine a valid ordering for the domains in the component.
1938 * This should be fairly rare because the individual domains
1939 * have been made disjoint first.
1940 * The problem is that the domains may be integrally disjoint but not
1941 * rationally disjoint. For example, we may have domains
1943 * { [i,i] : 0 <= i <= 1 } and { [i,1-i] : 0 <= i <= 1 }
1945 * These two domains have an empty intersection, but their rational
1946 * relaxations do intersect. It is impossible to order these domains
1947 * in the second dimension because the first should be ordered before
1948 * the second for outer dimension equal to 0, while it should be ordered
1949 * after for outer dimension equal to 1.
1951 * This may happen in particular in case of unrolling since the domain
1952 * of each slice is replaced by its simple hull.
1954 * For each basic set i in "scc" and for each of the following basic sets j,
1955 * we split off that part of the basic set i that shares the outer dimensions
1956 * with j and lies before j in the current dimension.
1957 * We collect all the pieces in a new list that replaces "scc".
1959 * While the elements in "scc" should be disjoint, we double-check
1960 * this property to avoid running into an infinite recursion in case
1961 * they intersect due to some internal error.
1963 static isl_stat
add_nodes(__isl_take isl_basic_set_list
*scc
, void *user
)
1965 struct isl_add_nodes_data
*data
= user
;
1969 isl_basic_set
*bset
, *first
;
1970 isl_basic_set_list
*list
;
1974 n
= isl_basic_set_list_n_basic_set(scc
);
1977 bset
= isl_basic_set_list_get_basic_set(scc
, 0);
1979 isl_basic_set_list_free(scc
);
1980 data
->list
= add_node(data
->list
,
1981 isl_union_map_copy(data
->executed
), bset
,
1982 isl_ast_build_copy(data
->build
));
1983 return data
->list
? isl_stat_ok
: isl_stat_error
;
1986 depth
= isl_ast_build_get_depth(data
->build
);
1988 bset
= isl_basic_set_free(bset
);
1989 space
= isl_basic_set_get_space(bset
);
1990 space
= isl_space_map_from_set(space
);
1991 gt
= isl_basic_map_universe(space
);
1992 for (i
= 0; i
< depth
; ++i
)
1993 gt
= isl_basic_map_equate(gt
, isl_dim_in
, i
, isl_dim_out
, i
);
1994 gt
= isl_basic_map_order_gt(gt
, isl_dim_in
, depth
, isl_dim_out
, depth
);
1996 first
= isl_basic_set_copy(bset
);
1997 list
= isl_basic_set_list_from_basic_set(bset
);
1998 for (i
= 1; i
< n
; ++i
) {
2001 bset
= isl_basic_set_list_get_basic_set(scc
, i
);
2003 disjoint
= isl_basic_set_is_disjoint(bset
, first
);
2005 list
= isl_basic_set_list_free(list
);
2007 isl_die(isl_basic_set_list_get_ctx(scc
),
2009 "basic sets in scc are assumed to be disjoint",
2010 list
= isl_basic_set_list_free(list
));
2012 list
= add_split_on(list
, bset
, gt
);
2014 isl_basic_set_free(first
);
2015 isl_basic_map_free(gt
);
2016 isl_basic_set_list_free(scc
);
2018 data
->list
= isl_ast_graft_list_concat(data
->list
,
2019 generate_sorted_domains(scc
, data
->executed
, data
->build
));
2020 isl_basic_set_list_free(scc
);
2022 return data
->list
? isl_stat_ok
: isl_stat_error
;
2024 isl_basic_set_list_free(scc
);
2025 return isl_stat_error
;
2028 /* Sort the domains in "domain_list" according to the execution order
2029 * at the current depth (for equal values of the outer dimensions),
2030 * generate code for each of them, collecting the results in a list.
2031 * If no code is generated (because the intersection of the inverse schedule
2032 * with the domains turns out to be empty), then an empty list is returned.
2034 * The caller is responsible for ensuring that the basic sets in "domain_list"
2035 * are pair-wise disjoint. It can, however, in principle happen that
2036 * two basic sets should be ordered one way for one value of the outer
2037 * dimensions and the other way for some other value of the outer dimensions.
2038 * We therefore play safe and look for strongly connected components.
2039 * The function add_nodes takes care of handling non-trivial components.
2041 static __isl_give isl_ast_graft_list
*generate_sorted_domains(
2042 __isl_keep isl_basic_set_list
*domain_list
,
2043 __isl_keep isl_union_map
*executed
, __isl_keep isl_ast_build
*build
)
2046 struct isl_add_nodes_data data
;
2050 n
= isl_basic_set_list_n_basic_set(domain_list
);
2054 ctx
= isl_basic_set_list_get_ctx(domain_list
);
2055 data
.list
= isl_ast_graft_list_alloc(ctx
, n
);
2059 return add_node(data
.list
, isl_union_map_copy(executed
),
2060 isl_basic_set_list_get_basic_set(domain_list
, 0),
2061 isl_ast_build_copy(build
));
2063 depth
= isl_ast_build_get_depth(build
);
2064 data
.executed
= executed
;
2066 if (depth
< 0 || isl_basic_set_list_foreach_scc(domain_list
,
2067 &domain_follows_at_depth
, &depth
,
2068 &add_nodes
, &data
) < 0)
2069 data
.list
= isl_ast_graft_list_free(data
.list
);
2074 /* Do i and j share any values for the outer dimensions?
2076 static isl_bool
shared_outer(__isl_keep isl_basic_set
*i
,
2077 __isl_keep isl_basic_set
*j
, void *user
)
2079 int depth
= *(int *) user
;
2080 isl_basic_map
*test
;
2084 test
= isl_basic_map_from_domain_and_range(isl_basic_set_copy(i
),
2085 isl_basic_set_copy(j
));
2086 for (l
= 0; l
< depth
; ++l
)
2087 test
= isl_basic_map_equate(test
, isl_dim_in
, l
,
2089 empty
= isl_basic_map_is_empty(test
);
2090 isl_basic_map_free(test
);
2092 return isl_bool_not(empty
);
2095 /* Internal data structure for generate_sorted_domains_wrap.
2097 * "n" is the total number of basic sets
2098 * "executed" and "build" are extra arguments to be passed
2099 * to generate_sorted_domains.
2101 * "single" is set to 1 by generate_sorted_domains_wrap if there
2102 * is only a single component.
2103 * "list" collects the results.
2105 struct isl_ast_generate_parallel_domains_data
{
2107 isl_union_map
*executed
;
2108 isl_ast_build
*build
;
2111 isl_ast_graft_list
*list
;
2114 /* Call generate_sorted_domains on "scc", fuse the result into a list
2115 * with either zero or one graft and collect the these single element
2116 * lists into data->list.
2118 * If there is only one component, i.e., if the number of basic sets
2119 * in the current component is equal to the total number of basic sets,
2120 * then data->single is set to 1 and the result of generate_sorted_domains
2123 static isl_stat
generate_sorted_domains_wrap(__isl_take isl_basic_set_list
*scc
,
2126 struct isl_ast_generate_parallel_domains_data
*data
= user
;
2127 isl_ast_graft_list
*list
;
2130 n
= isl_basic_set_list_n_basic_set(scc
);
2132 scc
= isl_basic_set_list_free(scc
);
2133 list
= generate_sorted_domains(scc
, data
->executed
, data
->build
);
2134 data
->single
= n
== data
->n
;
2136 list
= isl_ast_graft_list_fuse(list
, data
->build
);
2140 data
->list
= isl_ast_graft_list_concat(data
->list
, list
);
2142 isl_basic_set_list_free(scc
);
2144 return isl_stat_error
;
2149 /* Look for any (weakly connected) components in the "domain_list"
2150 * of domains that share some values of the outer dimensions.
2151 * That is, domains in different components do not share any values
2152 * of the outer dimensions. This means that these components
2153 * can be freely reordered.
2154 * Within each of the components, we sort the domains according
2155 * to the execution order at the current depth.
2157 * If there is more than one component, then generate_sorted_domains_wrap
2158 * fuses the result of each call to generate_sorted_domains
2159 * into a list with either zero or one graft and collects these (at most)
2160 * single element lists into a bigger list. This means that the elements of the
2161 * final list can be freely reordered. In particular, we sort them
2162 * according to an arbitrary but fixed ordering to ease merging of
2163 * graft lists from different components.
2165 static __isl_give isl_ast_graft_list
*generate_parallel_domains(
2166 __isl_keep isl_basic_set_list
*domain_list
,
2167 __isl_keep isl_union_map
*executed
, __isl_keep isl_ast_build
*build
)
2170 struct isl_ast_generate_parallel_domains_data data
;
2172 data
.n
= isl_basic_set_list_n_basic_set(domain_list
);
2177 return generate_sorted_domains(domain_list
, executed
, build
);
2179 depth
= isl_ast_build_get_depth(build
);
2183 data
.executed
= executed
;
2186 if (isl_basic_set_list_foreach_scc(domain_list
, &shared_outer
, &depth
,
2187 &generate_sorted_domains_wrap
,
2189 data
.list
= isl_ast_graft_list_free(data
.list
);
2192 data
.list
= isl_ast_graft_list_sort_guard(data
.list
);
2197 /* Internal data for separate_domain.
2199 * "explicit" is set if we only want to use explicit bounds.
2201 * "domain" collects the separated domains.
2203 struct isl_separate_domain_data
{
2204 isl_ast_build
*build
;
2209 /* Extract implicit bounds on the current dimension for the executed "map".
2211 * The domain of "map" may involve inner dimensions, so we
2212 * need to eliminate them.
2214 static __isl_give isl_set
*implicit_bounds(__isl_take isl_map
*map
,
2215 __isl_keep isl_ast_build
*build
)
2219 domain
= isl_map_domain(map
);
2220 domain
= isl_ast_build_eliminate(build
, domain
);
2225 /* Extract explicit bounds on the current dimension for the executed "map".
2227 * Rather than eliminating the inner dimensions as in implicit_bounds,
2228 * we simply drop any constraints involving those inner dimensions.
2229 * The idea is that most bounds that are implied by constraints on the
2230 * inner dimensions will be enforced by for loops and not by explicit guards.
2231 * There is then no need to separate along those bounds.
2233 static __isl_give isl_set
*explicit_bounds(__isl_take isl_map
*map
,
2234 __isl_keep isl_ast_build
*build
)
2240 depth
= isl_ast_build_get_depth(build
);
2241 dim
= isl_map_dim(map
, isl_dim_out
);
2242 if (depth
< 0 || dim
< 0)
2243 return isl_map_domain(isl_map_free(map
));
2244 map
= isl_map_drop_constraints_involving_dims(map
, isl_dim_out
, 0, dim
);
2246 domain
= isl_map_domain(map
);
2247 dim
= isl_set_dim(domain
, isl_dim_set
);
2248 domain
= isl_set_detect_equalities(domain
);
2249 domain
= isl_set_drop_constraints_involving_dims(domain
,
2250 isl_dim_set
, depth
+ 1, dim
- (depth
+ 1));
2251 domain
= isl_set_remove_divs_involving_dims(domain
,
2252 isl_dim_set
, depth
, 1);
2253 domain
= isl_set_remove_unknown_divs(domain
);
2258 /* Split data->domain into pieces that intersect with the range of "map"
2259 * and pieces that do not intersect with the range of "map"
2260 * and then add that part of the range of "map" that does not intersect
2261 * with data->domain.
2263 static isl_stat
separate_domain(__isl_take isl_map
*map
, void *user
)
2265 struct isl_separate_domain_data
*data
= user
;
2270 domain
= explicit_bounds(map
, data
->build
);
2272 domain
= implicit_bounds(map
, data
->build
);
2274 domain
= isl_set_coalesce(domain
);
2275 domain
= isl_set_make_disjoint(domain
);
2276 d1
= isl_set_subtract(isl_set_copy(domain
), isl_set_copy(data
->domain
));
2277 d2
= isl_set_subtract(isl_set_copy(data
->domain
), isl_set_copy(domain
));
2278 data
->domain
= isl_set_intersect(data
->domain
, domain
);
2279 data
->domain
= isl_set_union(data
->domain
, d1
);
2280 data
->domain
= isl_set_union(data
->domain
, d2
);
2285 /* Separate the schedule domains of "executed".
2287 * That is, break up the domain of "executed" into basic sets,
2288 * such that for each basic set S, every element in S is associated with
2289 * the same domain spaces.
2291 * "space" is the (single) domain space of "executed".
2293 static __isl_give isl_set
*separate_schedule_domains(
2294 __isl_take isl_space
*space
, __isl_take isl_union_map
*executed
,
2295 __isl_keep isl_ast_build
*build
)
2297 struct isl_separate_domain_data data
= { build
};
2300 ctx
= isl_ast_build_get_ctx(build
);
2301 data
.explicit = isl_options_get_ast_build_separation_bounds(ctx
) ==
2302 ISL_AST_BUILD_SEPARATION_BOUNDS_EXPLICIT
;
2303 data
.domain
= isl_set_empty(space
);
2304 if (isl_union_map_foreach_map(executed
, &separate_domain
, &data
) < 0)
2305 data
.domain
= isl_set_free(data
.domain
);
2307 isl_union_map_free(executed
);
2311 /* Temporary data used during the search for a lower bound for unrolling.
2313 * "build" is the build in which the unrolling will be performed
2314 * "domain" is the original set for which to find a lower bound
2315 * "depth" is the dimension for which to find a lower boudn
2316 * "expansion" is the expansion that needs to be applied to "domain"
2317 * in the unrolling that will be performed
2319 * "lower" is the best lower bound found so far. It is NULL if we have not
2321 * "n" is the corresponding size. If lower is NULL, then the value of n
2323 * "n_div" is the maximal number of integer divisions in the first
2324 * unrolled iteration (after expansion). It is set to -1 if it hasn't
2325 * been computed yet.
2327 struct isl_find_unroll_data
{
2328 isl_ast_build
*build
;
2331 isl_basic_map
*expansion
;
2338 /* Return the constraint
2340 * i_"depth" = aff + offset
2342 static __isl_give isl_constraint
*at_offset(int depth
, __isl_keep isl_aff
*aff
,
2345 aff
= isl_aff_copy(aff
);
2346 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, depth
, -1);
2347 aff
= isl_aff_add_constant_si(aff
, offset
);
2348 return isl_equality_from_aff(aff
);
2351 /* Update *user to the number of integer divisions in the first element
2352 * of "ma", if it is larger than the current value.
2354 static isl_stat
update_n_div(__isl_take isl_set
*set
,
2355 __isl_take isl_multi_aff
*ma
, void *user
)
2361 aff
= isl_multi_aff_get_aff(ma
, 0);
2362 n_div
= isl_aff_dim(aff
, isl_dim_div
);
2364 isl_multi_aff_free(ma
);
2370 return n_div
>= 0 ? isl_stat_ok
: isl_stat_error
;
2373 /* Get the number of integer divisions in the expression for the iterator
2374 * value at the first slice in the unrolling based on lower bound "lower",
2375 * taking into account the expansion that needs to be performed on this slice.
2377 static int get_expanded_n_div(struct isl_find_unroll_data
*data
,
2378 __isl_keep isl_aff
*lower
)
2382 isl_map
*it_map
, *expansion
;
2383 isl_pw_multi_aff
*pma
;
2386 c
= at_offset(data
->depth
, lower
, 0);
2387 set
= isl_set_copy(data
->domain
);
2388 set
= isl_set_add_constraint(set
, c
);
2389 expansion
= isl_map_from_basic_map(isl_basic_map_copy(data
->expansion
));
2390 set
= isl_set_apply(set
, expansion
);
2391 it_map
= isl_ast_build_map_to_iterator(data
->build
, set
);
2392 pma
= isl_pw_multi_aff_from_map(it_map
);
2394 if (isl_pw_multi_aff_foreach_piece(pma
, &update_n_div
, &n
) < 0)
2396 isl_pw_multi_aff_free(pma
);
2401 /* Is the lower bound "lower" with corresponding iteration count "n"
2402 * better than the one stored in "data"?
2403 * If there is no upper bound on the iteration count ("n" is infinity) or
2404 * if the count is too large, then we cannot use this lower bound.
2405 * Otherwise, if there was no previous lower bound or
2406 * if the iteration count of the new lower bound is smaller than
2407 * the iteration count of the previous lower bound, then we consider
2408 * the new lower bound to be better.
2409 * If the iteration count is the same, then compare the number
2410 * of integer divisions that would be needed to express
2411 * the iterator value at the first slice in the unrolling
2412 * according to the lower bound. If we end up computing this
2413 * number, then store the lowest value in data->n_div.
2415 static int is_better_lower_bound(struct isl_find_unroll_data
*data
,
2416 __isl_keep isl_aff
*lower
, __isl_keep isl_val
*n
)
2423 if (isl_val_is_infty(n
))
2425 if (isl_val_cmp_si(n
, INT_MAX
) > 0)
2429 cmp
= isl_val_cmp_si(n
, *data
->n
);
2434 if (data
->n_div
< 0)
2435 data
->n_div
= get_expanded_n_div(data
, data
->lower
);
2436 if (data
->n_div
< 0)
2438 if (data
->n_div
== 0)
2440 n_div
= get_expanded_n_div(data
, lower
);
2443 if (n_div
>= data
->n_div
)
2445 data
->n_div
= n_div
;
2450 /* Check if we can use "c" as a lower bound and if it is better than
2451 * any previously found lower bound.
2453 * If "c" does not involve the dimension at the current depth,
2454 * then we cannot use it.
2455 * Otherwise, let "c" be of the form
2459 * We compute the maximal value of
2461 * -ceil(f(j)/a)) + i + 1
2463 * over the domain. If there is such a value "n", then we know
2465 * -ceil(f(j)/a)) + i + 1 <= n
2469 * i < ceil(f(j)/a)) + n
2471 * meaning that we can use ceil(f(j)/a)) as a lower bound for unrolling.
2472 * We just need to check if we have found any lower bound before and
2473 * if the new lower bound is better (smaller n or fewer integer divisions)
2474 * than the previously found lower bounds.
2476 static isl_stat
update_unrolling_lower_bound(struct isl_find_unroll_data
*data
,
2477 __isl_keep isl_constraint
*c
)
2479 isl_aff
*aff
, *lower
;
2483 if (!isl_constraint_is_lower_bound(c
, isl_dim_set
, data
->depth
))
2486 lower
= isl_constraint_get_bound(c
, isl_dim_set
, data
->depth
);
2487 lower
= isl_aff_ceil(lower
);
2488 aff
= isl_aff_copy(lower
);
2489 aff
= isl_aff_neg(aff
);
2490 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, data
->depth
, 1);
2491 aff
= isl_aff_add_constant_si(aff
, 1);
2492 max
= isl_set_max_val(data
->domain
, aff
);
2495 better
= is_better_lower_bound(data
, lower
, max
);
2496 if (better
< 0 || !better
) {
2498 isl_aff_free(lower
);
2499 return better
< 0 ? isl_stat_error
: isl_stat_ok
;
2502 isl_aff_free(data
->lower
);
2503 data
->lower
= lower
;
2504 *data
->n
= isl_val_get_num_si(max
);
2510 /* Check if we can use "c" as a lower bound and if it is better than
2511 * any previously found lower bound.
2513 static isl_stat
constraint_find_unroll(__isl_take isl_constraint
*c
, void *user
)
2515 struct isl_find_unroll_data
*data
;
2518 data
= (struct isl_find_unroll_data
*) user
;
2519 r
= update_unrolling_lower_bound(data
, c
);
2520 isl_constraint_free(c
);
2525 /* Look for a lower bound l(i) on the dimension at "depth"
2526 * and a size n such that "domain" is a subset of
2528 * { [i] : l(i) <= i_d < l(i) + n }
2530 * where d is "depth" and l(i) depends only on earlier dimensions.
2531 * Furthermore, try and find a lower bound such that n is as small as possible.
2532 * In particular, "n" needs to be finite.
2533 * "build" is the build in which the unrolling will be performed.
2534 * "expansion" is the expansion that needs to be applied to "domain"
2535 * in the unrolling that will be performed.
2537 * Inner dimensions have been eliminated from "domain" by the caller.
2539 * We first construct a collection of lower bounds on the input set
2540 * by computing its simple hull. We then iterate through them,
2541 * discarding those that we cannot use (either because they do not
2542 * involve the dimension at "depth" or because they have no corresponding
2543 * upper bound, meaning that "n" would be unbounded) and pick out the
2544 * best from the remaining ones.
2546 * If we cannot find a suitable lower bound, then we consider that
2549 static __isl_give isl_aff
*find_unroll_lower_bound(
2550 __isl_keep isl_ast_build
*build
, __isl_keep isl_set
*domain
,
2551 int depth
, __isl_keep isl_basic_map
*expansion
, int *n
)
2553 struct isl_find_unroll_data data
=
2554 { build
, domain
, depth
, expansion
, NULL
, n
, -1 };
2555 isl_basic_set
*hull
;
2557 hull
= isl_set_simple_hull(isl_set_copy(domain
));
2559 if (isl_basic_set_foreach_constraint(hull
,
2560 &constraint_find_unroll
, &data
) < 0)
2563 isl_basic_set_free(hull
);
2566 isl_die(isl_set_get_ctx(domain
), isl_error_invalid
,
2567 "cannot find lower bound for unrolling", return NULL
);
2571 isl_basic_set_free(hull
);
2572 return isl_aff_free(data
.lower
);
2575 /* Call "fn" on each iteration of the current dimension of "domain".
2576 * If "init" is not NULL, then it is called with the number of
2577 * iterations before any call to "fn".
2578 * Return -1 on failure.
2580 * Since we are going to be iterating over the individual values,
2581 * we first check if there are any strides on the current dimension.
2582 * If there is, we rewrite the current dimension i as
2584 * i = stride i' + offset
2586 * and then iterate over individual values of i' instead.
2588 * We then look for a lower bound on i' and a size such that the domain
2591 * { [j,i'] : l(j) <= i' < l(j) + n }
2593 * and then take slices of the domain at values of i'
2594 * between l(j) and l(j) + n - 1.
2596 * We compute the unshifted simple hull of each slice to ensure that
2597 * we have a single basic set per offset. The slicing constraint
2598 * may get simplified away before the unshifted simple hull is taken
2599 * and may therefore in some rare cases disappear from the result.
2600 * We therefore explicitly add the constraint back after computing
2601 * the unshifted simple hull to ensure that the basic sets
2602 * remain disjoint. The constraints that are dropped by taking the hull
2603 * will be taken into account at the next level, as in the case of the
2606 * Finally, we map i' back to i and call "fn".
2608 static int foreach_iteration(__isl_take isl_set
*domain
,
2609 __isl_keep isl_ast_build
*build
, int (*init
)(int n
, void *user
),
2610 int (*fn
)(__isl_take isl_basic_set
*bset
, void *user
), void *user
)
2615 isl_multi_aff
*expansion
;
2616 isl_basic_map
*bmap
;
2617 isl_aff
*lower
= NULL
;
2618 isl_ast_build
*stride_build
;
2620 depth
= isl_ast_build_get_depth(build
);
2622 domain
= isl_set_free(domain
);
2624 domain
= isl_ast_build_eliminate_inner(build
, domain
);
2625 domain
= isl_set_intersect(domain
, isl_ast_build_get_domain(build
));
2626 stride_build
= isl_ast_build_copy(build
);
2627 stride_build
= isl_ast_build_detect_strides(stride_build
,
2628 isl_set_copy(domain
));
2629 expansion
= isl_ast_build_get_stride_expansion(stride_build
);
2631 domain
= isl_set_preimage_multi_aff(domain
,
2632 isl_multi_aff_copy(expansion
));
2633 domain
= isl_ast_build_eliminate_divs(stride_build
, domain
);
2634 isl_ast_build_free(stride_build
);
2636 bmap
= isl_basic_map_from_multi_aff(expansion
);
2638 empty
= isl_set_is_empty(domain
);
2644 lower
= find_unroll_lower_bound(build
, domain
, depth
, bmap
, &n
);
2648 if (n
>= 0 && init
&& init(n
, user
) < 0)
2650 for (i
= 0; i
< n
; ++i
) {
2652 isl_basic_set
*bset
;
2653 isl_constraint
*slice
;
2655 slice
= at_offset(depth
, lower
, i
);
2656 set
= isl_set_copy(domain
);
2657 set
= isl_set_add_constraint(set
, isl_constraint_copy(slice
));
2658 bset
= isl_set_unshifted_simple_hull(set
);
2659 bset
= isl_basic_set_add_constraint(bset
, slice
);
2660 bset
= isl_basic_set_apply(bset
, isl_basic_map_copy(bmap
));
2662 if (fn(bset
, user
) < 0)
2666 isl_aff_free(lower
);
2667 isl_set_free(domain
);
2668 isl_basic_map_free(bmap
);
2670 return n
< 0 || i
< n
? -1 : 0;
2673 /* Data structure for storing the results and the intermediate objects
2674 * of compute_domains.
2676 * "list" is the main result of the function and contains a list
2677 * of disjoint basic sets for which code should be generated.
2679 * "executed" and "build" are inputs to compute_domains.
2680 * "schedule_domain" is the domain of "executed".
2682 * "option" contains the domains at the current depth that should by
2683 * atomic, separated or unrolled. These domains are as specified by
2684 * the user, except that inner dimensions have been eliminated and
2685 * that they have been made pair-wise disjoint.
2687 * "sep_class" contains the user-specified split into separation classes
2688 * specialized to the current depth.
2689 * "done" contains the union of the separation domains that have already
2692 struct isl_codegen_domains
{
2693 isl_basic_set_list
*list
;
2695 isl_union_map
*executed
;
2696 isl_ast_build
*build
;
2697 isl_set
*schedule_domain
;
2705 /* Internal data structure for do_unroll.
2707 * "domains" stores the results of compute_domains.
2708 * "class_domain" is the original class domain passed to do_unroll.
2709 * "unroll_domain" collects the unrolled iterations.
2711 struct isl_ast_unroll_data
{
2712 struct isl_codegen_domains
*domains
;
2713 isl_set
*class_domain
;
2714 isl_set
*unroll_domain
;
2717 /* Given an iteration of an unrolled domain represented by "bset",
2718 * add it to data->domains->list.
2719 * Since we may have dropped some constraints, we intersect with
2720 * the class domain again to ensure that each element in the list
2721 * is disjoint from the other class domains.
2723 static int do_unroll_iteration(__isl_take isl_basic_set
*bset
, void *user
)
2725 struct isl_ast_unroll_data
*data
= user
;
2727 isl_basic_set_list
*list
;
2729 set
= isl_set_from_basic_set(bset
);
2730 data
->unroll_domain
= isl_set_union(data
->unroll_domain
,
2732 set
= isl_set_intersect(set
, isl_set_copy(data
->class_domain
));
2733 set
= isl_set_make_disjoint(set
);
2734 list
= isl_basic_set_list_from_set(set
);
2735 data
->domains
->list
= isl_basic_set_list_concat(data
->domains
->list
,
2741 /* Extend domains->list with a list of basic sets, one for each value
2742 * of the current dimension in "domain" and remove the corresponding
2743 * sets from the class domain. Return the updated class domain.
2744 * The divs that involve the current dimension have not been projected out
2747 * We call foreach_iteration to iterate over the individual values and
2748 * in do_unroll_iteration we collect the individual basic sets in
2749 * domains->list and their union in data->unroll_domain, which is then
2750 * used to update the class domain.
2752 static __isl_give isl_set
*do_unroll(struct isl_codegen_domains
*domains
,
2753 __isl_take isl_set
*domain
, __isl_take isl_set
*class_domain
)
2755 struct isl_ast_unroll_data data
;
2758 return isl_set_free(class_domain
);
2760 return isl_set_free(domain
);
2762 data
.domains
= domains
;
2763 data
.class_domain
= class_domain
;
2764 data
.unroll_domain
= isl_set_empty(isl_set_get_space(domain
));
2766 if (foreach_iteration(domain
, domains
->build
, NULL
,
2767 &do_unroll_iteration
, &data
) < 0)
2768 data
.unroll_domain
= isl_set_free(data
.unroll_domain
);
2770 class_domain
= isl_set_subtract(class_domain
, data
.unroll_domain
);
2772 return class_domain
;
2775 /* Add domains to domains->list for each individual value of the current
2776 * dimension, for that part of the schedule domain that lies in the
2777 * intersection of the option domain and the class domain.
2778 * Remove the corresponding sets from the class domain and
2779 * return the updated class domain.
2781 * We first break up the unroll option domain into individual pieces
2782 * and then handle each of them separately. The unroll option domain
2783 * has been made disjoint in compute_domains_init_options,
2785 * Note that we actively want to combine different pieces of the
2786 * schedule domain that have the same value at the current dimension.
2787 * We therefore need to break up the unroll option domain before
2788 * intersecting with class and schedule domain, hoping that the
2789 * unroll option domain specified by the user is relatively simple.
2791 static __isl_give isl_set
*compute_unroll_domains(
2792 struct isl_codegen_domains
*domains
, __isl_take isl_set
*class_domain
)
2794 isl_set
*unroll_domain
;
2795 isl_basic_set_list
*unroll_list
;
2800 empty
= isl_set_is_empty(domains
->option
[isl_ast_loop_unroll
]);
2802 return isl_set_free(class_domain
);
2804 return class_domain
;
2806 unroll_domain
= isl_set_copy(domains
->option
[isl_ast_loop_unroll
]);
2807 unroll_list
= isl_basic_set_list_from_set(unroll_domain
);
2809 n
= isl_basic_set_list_n_basic_set(unroll_list
);
2811 class_domain
= isl_set_free(class_domain
);
2812 for (i
= 0; i
< n
; ++i
) {
2813 isl_basic_set
*bset
;
2815 bset
= isl_basic_set_list_get_basic_set(unroll_list
, i
);
2816 unroll_domain
= isl_set_from_basic_set(bset
);
2817 unroll_domain
= isl_set_intersect(unroll_domain
,
2818 isl_set_copy(class_domain
));
2819 unroll_domain
= isl_set_intersect(unroll_domain
,
2820 isl_set_copy(domains
->schedule_domain
));
2822 empty
= isl_set_is_empty(unroll_domain
);
2823 if (empty
>= 0 && empty
) {
2824 isl_set_free(unroll_domain
);
2828 class_domain
= do_unroll(domains
, unroll_domain
, class_domain
);
2831 isl_basic_set_list_free(unroll_list
);
2833 return class_domain
;
2836 /* Try and construct a single basic set that includes the intersection of
2837 * the schedule domain, the atomic option domain and the class domain.
2838 * Add the resulting basic set(s) to domains->list and remove them
2839 * from class_domain. Return the updated class domain.
2841 * We construct a single domain rather than trying to combine
2842 * the schedule domains of individual domains because we are working
2843 * within a single component so that non-overlapping schedule domains
2844 * should already have been separated.
2845 * We do however need to make sure that this single domains is a subset
2846 * of the class domain so that it would not intersect with any other
2847 * class domains. This means that we may end up splitting up the atomic
2848 * domain in case separation classes are being used.
2850 * "domain" is the intersection of the schedule domain and the class domain,
2851 * with inner dimensions projected out.
2853 static __isl_give isl_set
*compute_atomic_domain(
2854 struct isl_codegen_domains
*domains
, __isl_take isl_set
*class_domain
)
2856 isl_basic_set
*bset
;
2857 isl_basic_set_list
*list
;
2858 isl_set
*domain
, *atomic_domain
;
2861 domain
= isl_set_copy(domains
->option
[isl_ast_loop_atomic
]);
2862 domain
= isl_set_intersect(domain
, isl_set_copy(class_domain
));
2863 domain
= isl_set_intersect(domain
,
2864 isl_set_copy(domains
->schedule_domain
));
2865 empty
= isl_set_is_empty(domain
);
2867 class_domain
= isl_set_free(class_domain
);
2869 isl_set_free(domain
);
2870 return class_domain
;
2873 domain
= isl_ast_build_eliminate(domains
->build
, domain
);
2874 domain
= isl_set_coalesce_preserve(domain
);
2875 bset
= isl_set_unshifted_simple_hull(domain
);
2876 domain
= isl_set_from_basic_set(bset
);
2877 atomic_domain
= isl_set_copy(domain
);
2878 domain
= isl_set_intersect(domain
, isl_set_copy(class_domain
));
2879 class_domain
= isl_set_subtract(class_domain
, atomic_domain
);
2880 domain
= isl_set_make_disjoint(domain
);
2881 list
= isl_basic_set_list_from_set(domain
);
2882 domains
->list
= isl_basic_set_list_concat(domains
->list
, list
);
2884 return class_domain
;
2887 /* Split up the schedule domain into uniform basic sets,
2888 * in the sense that each element in a basic set is associated to
2889 * elements of the same domains, and add the result to domains->list.
2890 * Do this for that part of the schedule domain that lies in the
2891 * intersection of "class_domain" and the separate option domain.
2893 * "class_domain" may or may not include the constraints
2894 * of the schedule domain, but this does not make a difference
2895 * since we are going to intersect it with the domain of the inverse schedule.
2896 * If it includes schedule domain constraints, then they may involve
2897 * inner dimensions, but we will eliminate them in separation_domain.
2899 static int compute_separate_domain(struct isl_codegen_domains
*domains
,
2900 __isl_keep isl_set
*class_domain
)
2904 isl_union_map
*executed
;
2905 isl_basic_set_list
*list
;
2908 domain
= isl_set_copy(domains
->option
[isl_ast_loop_separate
]);
2909 domain
= isl_set_intersect(domain
, isl_set_copy(class_domain
));
2910 executed
= isl_union_map_copy(domains
->executed
);
2911 executed
= isl_union_map_intersect_domain(executed
,
2912 isl_union_set_from_set(domain
));
2913 empty
= isl_union_map_is_empty(executed
);
2914 if (empty
< 0 || empty
) {
2915 isl_union_map_free(executed
);
2916 return empty
< 0 ? -1 : 0;
2919 space
= isl_set_get_space(class_domain
);
2920 domain
= separate_schedule_domains(space
, executed
, domains
->build
);
2922 list
= isl_basic_set_list_from_set(domain
);
2923 domains
->list
= isl_basic_set_list_concat(domains
->list
, list
);
2928 /* Split up the domain at the current depth into disjoint
2929 * basic sets for which code should be generated separately
2930 * for the given separation class domain.
2932 * If any separation classes have been defined, then "class_domain"
2933 * is the domain of the current class and does not refer to inner dimensions.
2934 * Otherwise, "class_domain" is the universe domain.
2936 * We first make sure that the class domain is disjoint from
2937 * previously considered class domains.
2939 * The separate domains can be computed directly from the "class_domain".
2941 * The unroll, atomic and remainder domains need the constraints
2942 * from the schedule domain.
2944 * For unrolling, the actual schedule domain is needed (with divs that
2945 * may refer to the current dimension) so that stride detection can be
2948 * For atomic and remainder domains, inner dimensions and divs involving
2949 * the current dimensions should be eliminated.
2950 * In case we are working within a separation class, we need to intersect
2951 * the result with the current "class_domain" to ensure that the domains
2952 * are disjoint from those generated from other class domains.
2954 * The domain that has been made atomic may be larger than specified
2955 * by the user since it needs to be representable as a single basic set.
2956 * This possibly larger domain is removed from class_domain by
2957 * compute_atomic_domain. It is computed first so that the extended domain
2958 * would not overlap with any domains computed before.
2959 * Similary, the unrolled domains may have some constraints removed and
2960 * may therefore also be larger than specified by the user.
2962 * If anything is left after handling separate, unroll and atomic,
2963 * we split it up into basic sets and append the basic sets to domains->list.
2965 static isl_stat
compute_partial_domains(struct isl_codegen_domains
*domains
,
2966 __isl_take isl_set
*class_domain
)
2968 isl_basic_set_list
*list
;
2971 class_domain
= isl_set_subtract(class_domain
,
2972 isl_set_copy(domains
->done
));
2973 domains
->done
= isl_set_union(domains
->done
,
2974 isl_set_copy(class_domain
));
2976 class_domain
= compute_atomic_domain(domains
, class_domain
);
2977 class_domain
= compute_unroll_domains(domains
, class_domain
);
2979 domain
= isl_set_copy(class_domain
);
2981 if (compute_separate_domain(domains
, domain
) < 0)
2983 domain
= isl_set_subtract(domain
,
2984 isl_set_copy(domains
->option
[isl_ast_loop_separate
]));
2986 domain
= isl_set_intersect(domain
,
2987 isl_set_copy(domains
->schedule_domain
));
2989 domain
= isl_ast_build_eliminate(domains
->build
, domain
);
2990 domain
= isl_set_intersect(domain
, isl_set_copy(class_domain
));
2992 domain
= isl_set_coalesce_preserve(domain
);
2993 domain
= isl_set_make_disjoint(domain
);
2995 list
= isl_basic_set_list_from_set(domain
);
2996 domains
->list
= isl_basic_set_list_concat(domains
->list
, list
);
2998 isl_set_free(class_domain
);
3002 isl_set_free(domain
);
3003 isl_set_free(class_domain
);
3004 return isl_stat_error
;
3007 /* Split up the domain at the current depth into disjoint
3008 * basic sets for which code should be generated separately
3009 * for the separation class identified by "pnt".
3011 * We extract the corresponding class domain from domains->sep_class,
3012 * eliminate inner dimensions and pass control to compute_partial_domains.
3014 static isl_stat
compute_class_domains(__isl_take isl_point
*pnt
, void *user
)
3016 struct isl_codegen_domains
*domains
= user
;
3021 class_set
= isl_set_from_point(pnt
);
3022 domain
= isl_map_domain(isl_map_intersect_range(
3023 isl_map_copy(domains
->sep_class
), class_set
));
3024 domain
= isl_ast_build_compute_gist(domains
->build
, domain
);
3025 domain
= isl_ast_build_eliminate(domains
->build
, domain
);
3027 disjoint
= isl_set_plain_is_disjoint(domain
, domains
->schedule_domain
);
3029 return isl_stat_error
;
3031 isl_set_free(domain
);
3035 return compute_partial_domains(domains
, domain
);
3038 /* Extract the domains at the current depth that should be atomic,
3039 * separated or unrolled and store them in option.
3041 * The domains specified by the user might overlap, so we make
3042 * them disjoint by subtracting earlier domains from later domains.
3044 static void compute_domains_init_options(isl_set
*option
[4],
3045 __isl_keep isl_ast_build
*build
)
3047 enum isl_ast_loop_type type
, type2
;
3050 for (type
= isl_ast_loop_atomic
;
3051 type
<= isl_ast_loop_separate
; ++type
) {
3052 option
[type
] = isl_ast_build_get_option_domain(build
, type
);
3053 for (type2
= isl_ast_loop_atomic
; type2
< type
; ++type2
)
3054 option
[type
] = isl_set_subtract(option
[type
],
3055 isl_set_copy(option
[type2
]));
3058 unroll
= option
[isl_ast_loop_unroll
];
3059 unroll
= isl_set_coalesce(unroll
);
3060 unroll
= isl_set_make_disjoint(unroll
);
3061 option
[isl_ast_loop_unroll
] = unroll
;
3064 /* Split up the domain at the current depth into disjoint
3065 * basic sets for which code should be generated separately,
3066 * based on the user-specified options.
3067 * Return the list of disjoint basic sets.
3069 * There are three kinds of domains that we need to keep track of.
3070 * - the "schedule domain" is the domain of "executed"
3071 * - the "class domain" is the domain corresponding to the currrent
3073 * - the "option domain" is the domain corresponding to one of the options
3074 * atomic, unroll or separate
3076 * We first consider the individial values of the separation classes
3077 * and split up the domain for each of them separately.
3078 * Finally, we consider the remainder. If no separation classes were
3079 * specified, then we call compute_partial_domains with the universe
3080 * "class_domain". Otherwise, we take the "schedule_domain" as "class_domain",
3081 * with inner dimensions removed. We do this because we want to
3082 * avoid computing the complement of the class domains (i.e., the difference
3083 * between the universe and domains->done).
3085 static __isl_give isl_basic_set_list
*compute_domains(
3086 __isl_keep isl_union_map
*executed
, __isl_keep isl_ast_build
*build
)
3088 struct isl_codegen_domains domains
;
3091 isl_union_set
*schedule_domain
;
3095 enum isl_ast_loop_type type
;
3101 ctx
= isl_union_map_get_ctx(executed
);
3102 domains
.list
= isl_basic_set_list_alloc(ctx
, 0);
3104 schedule_domain
= isl_union_map_domain(isl_union_map_copy(executed
));
3105 domain
= isl_set_from_union_set(schedule_domain
);
3107 compute_domains_init_options(domains
.option
, build
);
3109 domains
.sep_class
= isl_ast_build_get_separation_class(build
);
3110 classes
= isl_map_range(isl_map_copy(domains
.sep_class
));
3111 n_param
= isl_set_dim(classes
, isl_dim_param
);
3113 classes
= isl_set_free(classes
);
3114 classes
= isl_set_project_out(classes
, isl_dim_param
, 0, n_param
);
3116 space
= isl_set_get_space(domain
);
3117 domains
.build
= build
;
3118 domains
.schedule_domain
= isl_set_copy(domain
);
3119 domains
.executed
= executed
;
3120 domains
.done
= isl_set_empty(space
);
3122 if (isl_set_foreach_point(classes
, &compute_class_domains
, &domains
) < 0)
3123 domains
.list
= isl_basic_set_list_free(domains
.list
);
3124 isl_set_free(classes
);
3126 empty
= isl_set_is_empty(domains
.done
);
3128 domains
.list
= isl_basic_set_list_free(domains
.list
);
3129 domain
= isl_set_free(domain
);
3131 isl_set_free(domain
);
3132 domain
= isl_set_universe(isl_set_get_space(domains
.done
));
3134 domain
= isl_ast_build_eliminate(build
, domain
);
3136 if (compute_partial_domains(&domains
, domain
) < 0)
3137 domains
.list
= isl_basic_set_list_free(domains
.list
);
3139 isl_set_free(domains
.schedule_domain
);
3140 isl_set_free(domains
.done
);
3141 isl_map_free(domains
.sep_class
);
3142 for (type
= isl_ast_loop_atomic
; type
<= isl_ast_loop_separate
; ++type
)
3143 isl_set_free(domains
.option
[type
]);
3145 return domains
.list
;
3148 /* Generate code for a single component, after shifting (if any)
3149 * has been applied, in case the schedule was specified as a union map.
3151 * We first split up the domain at the current depth into disjoint
3152 * basic sets based on the user-specified options.
3153 * Then we generated code for each of them and concatenate the results.
3155 static __isl_give isl_ast_graft_list
*generate_shifted_component_flat(
3156 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
3158 isl_basic_set_list
*domain_list
;
3159 isl_ast_graft_list
*list
= NULL
;
3161 domain_list
= compute_domains(executed
, build
);
3162 list
= generate_parallel_domains(domain_list
, executed
, build
);
3164 isl_basic_set_list_free(domain_list
);
3165 isl_union_map_free(executed
);
3166 isl_ast_build_free(build
);
3171 /* Generate code for a single component, after shifting (if any)
3172 * has been applied, in case the schedule was specified as a schedule tree
3173 * and the separate option was specified.
3175 * We perform separation on the domain of "executed" and then generate
3176 * an AST for each of the resulting disjoint basic sets.
3178 static __isl_give isl_ast_graft_list
*generate_shifted_component_tree_separate(
3179 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
3183 isl_basic_set_list
*domain_list
;
3184 isl_ast_graft_list
*list
;
3186 space
= isl_ast_build_get_space(build
, 1);
3187 domain
= separate_schedule_domains(space
,
3188 isl_union_map_copy(executed
), build
);
3189 domain_list
= isl_basic_set_list_from_set(domain
);
3191 list
= generate_parallel_domains(domain_list
, executed
, build
);
3193 isl_basic_set_list_free(domain_list
);
3194 isl_union_map_free(executed
);
3195 isl_ast_build_free(build
);
3200 /* Internal data structure for generate_shifted_component_tree_unroll.
3202 * "executed" and "build" are inputs to generate_shifted_component_tree_unroll.
3203 * "list" collects the constructs grafts.
3205 struct isl_ast_unroll_tree_data
{
3206 isl_union_map
*executed
;
3207 isl_ast_build
*build
;
3208 isl_ast_graft_list
*list
;
3211 /* Initialize data->list to a list of "n" elements.
3213 static int init_unroll_tree(int n
, void *user
)
3215 struct isl_ast_unroll_tree_data
*data
= user
;
3218 ctx
= isl_ast_build_get_ctx(data
->build
);
3219 data
->list
= isl_ast_graft_list_alloc(ctx
, n
);
3224 /* Given an iteration of an unrolled domain represented by "bset",
3225 * generate the corresponding AST and add the result to data->list.
3227 static int do_unroll_tree_iteration(__isl_take isl_basic_set
*bset
, void *user
)
3229 struct isl_ast_unroll_tree_data
*data
= user
;
3231 data
->list
= add_node(data
->list
, isl_union_map_copy(data
->executed
),
3232 bset
, isl_ast_build_copy(data
->build
));
3237 /* Generate code for a single component, after shifting (if any)
3238 * has been applied, in case the schedule was specified as a schedule tree
3239 * and the unroll option was specified.
3241 * We call foreach_iteration to iterate over the individual values and
3242 * construct and collect the corresponding grafts in do_unroll_tree_iteration.
3244 static __isl_give isl_ast_graft_list
*generate_shifted_component_tree_unroll(
3245 __isl_take isl_union_map
*executed
, __isl_take isl_set
*domain
,
3246 __isl_take isl_ast_build
*build
)
3248 struct isl_ast_unroll_tree_data data
= { executed
, build
, NULL
};
3250 if (foreach_iteration(domain
, build
, &init_unroll_tree
,
3251 &do_unroll_tree_iteration
, &data
) < 0)
3252 data
.list
= isl_ast_graft_list_free(data
.list
);
3254 isl_union_map_free(executed
);
3255 isl_ast_build_free(build
);
3260 /* Does "domain" involve a disjunction that is purely based on
3261 * constraints involving only outer dimension?
3263 * In particular, is there a disjunction such that the constraints
3264 * involving the current and later dimensions are the same over
3265 * all the disjuncts?
3267 static isl_bool
has_pure_outer_disjunction(__isl_keep isl_set
*domain
,
3268 __isl_keep isl_ast_build
*build
)
3270 isl_basic_set
*hull
;
3271 isl_set
*shared
, *inner
;
3277 n
= isl_set_n_basic_set(domain
);
3279 return isl_bool_error
;
3281 return isl_bool_false
;
3282 dim
= isl_set_dim(domain
, isl_dim_set
);
3283 depth
= isl_ast_build_get_depth(build
);
3284 if (dim
< 0 || depth
< 0)
3285 return isl_bool_error
;
3287 inner
= isl_set_copy(domain
);
3288 inner
= isl_set_drop_constraints_not_involving_dims(inner
,
3289 isl_dim_set
, depth
, dim
- depth
);
3290 hull
= isl_set_plain_unshifted_simple_hull(isl_set_copy(inner
));
3291 shared
= isl_set_from_basic_set(hull
);
3292 equal
= isl_set_plain_is_equal(inner
, shared
);
3293 isl_set_free(inner
);
3294 isl_set_free(shared
);
3299 /* Generate code for a single component, after shifting (if any)
3300 * has been applied, in case the schedule was specified as a schedule tree.
3301 * In particular, handle the base case where there is either no isolated
3302 * set or we are within the isolated set (in which case "isolated" is set)
3303 * or the iterations that precede or follow the isolated set.
3305 * The schedule domain is broken up or combined into basic sets
3306 * according to the AST generation option specified in the current
3307 * schedule node, which may be either atomic, separate, unroll or
3308 * unspecified. If the option is unspecified, then we currently simply
3309 * split the schedule domain into disjoint basic sets.
3311 * In case the separate option is specified, the AST generation is
3312 * handled by generate_shifted_component_tree_separate.
3313 * In the other cases, we need the global schedule domain.
3314 * In the unroll case, the AST generation is then handled by
3315 * generate_shifted_component_tree_unroll which needs the actual
3316 * schedule domain (with divs that may refer to the current dimension)
3317 * so that stride detection can be performed.
3318 * In the atomic or unspecified case, inner dimensions and divs involving
3319 * the current dimensions should be eliminated.
3320 * The result is then either combined into a single basic set or
3321 * split up into disjoint basic sets.
3322 * Finally an AST is generated for each basic set and the results are
3325 * If the schedule domain involves a disjunction that is purely based on
3326 * constraints involving only outer dimension, then it is treated as
3327 * if atomic was specified. This ensures that only a single loop
3328 * is generated instead of a sequence of identical loops with
3331 static __isl_give isl_ast_graft_list
*generate_shifted_component_tree_base(
3332 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
,
3335 isl_bool outer_disjunction
;
3336 isl_union_set
*schedule_domain
;
3338 isl_basic_set_list
*domain_list
;
3339 isl_ast_graft_list
*list
;
3340 enum isl_ast_loop_type type
;
3342 type
= isl_ast_build_get_loop_type(build
, isolated
);
3346 if (type
== isl_ast_loop_separate
)
3347 return generate_shifted_component_tree_separate(executed
,
3350 schedule_domain
= isl_union_map_domain(isl_union_map_copy(executed
));
3351 domain
= isl_set_from_union_set(schedule_domain
);
3353 if (type
== isl_ast_loop_unroll
)
3354 return generate_shifted_component_tree_unroll(executed
, domain
,
3357 domain
= isl_ast_build_eliminate(build
, domain
);
3358 domain
= isl_set_coalesce_preserve(domain
);
3360 outer_disjunction
= has_pure_outer_disjunction(domain
, build
);
3361 if (outer_disjunction
< 0)
3362 domain
= isl_set_free(domain
);
3364 if (outer_disjunction
|| type
== isl_ast_loop_atomic
) {
3365 isl_basic_set
*hull
;
3366 hull
= isl_set_unshifted_simple_hull(domain
);
3367 domain_list
= isl_basic_set_list_from_basic_set(hull
);
3369 domain
= isl_set_make_disjoint(domain
);
3370 domain_list
= isl_basic_set_list_from_set(domain
);
3373 list
= generate_parallel_domains(domain_list
, executed
, build
);
3375 isl_basic_set_list_free(domain_list
);
3376 isl_union_map_free(executed
);
3377 isl_ast_build_free(build
);
3381 isl_union_map_free(executed
);
3382 isl_ast_build_free(build
);
3386 /* Extract out the disjunction imposed by "domain" on the outer
3387 * schedule dimensions.
3389 * In particular, remove all inner dimensions from "domain" (including
3390 * the current dimension) and then remove the constraints that are shared
3391 * by all disjuncts in the result.
3393 static __isl_give isl_set
*extract_disjunction(__isl_take isl_set
*domain
,
3394 __isl_keep isl_ast_build
*build
)
3400 domain
= isl_ast_build_specialize(build
, domain
);
3401 depth
= isl_ast_build_get_depth(build
);
3402 dim
= isl_set_dim(domain
, isl_dim_set
);
3403 if (depth
< 0 || dim
< 0)
3404 return isl_set_free(domain
);
3405 domain
= isl_set_eliminate(domain
, isl_dim_set
, depth
, dim
- depth
);
3406 domain
= isl_set_remove_unknown_divs(domain
);
3407 hull
= isl_set_copy(domain
);
3408 hull
= isl_set_from_basic_set(isl_set_unshifted_simple_hull(hull
));
3409 domain
= isl_set_gist(domain
, hull
);
3414 /* Add "guard" to the grafts in "list".
3415 * "build" is the outer AST build, while "sub_build" includes "guard"
3416 * in its generated domain.
3418 * First combine the grafts into a single graft and then add the guard.
3419 * If the list is empty, or if some error occurred, then simply return
3422 static __isl_give isl_ast_graft_list
*list_add_guard(
3423 __isl_take isl_ast_graft_list
*list
, __isl_keep isl_set
*guard
,
3424 __isl_keep isl_ast_build
*build
, __isl_keep isl_ast_build
*sub_build
)
3426 isl_ast_graft
*graft
;
3429 list
= isl_ast_graft_list_fuse(list
, sub_build
);
3431 n
= isl_ast_graft_list_n_ast_graft(list
);
3433 return isl_ast_graft_list_free(list
);
3437 graft
= isl_ast_graft_list_get_ast_graft(list
, 0);
3438 graft
= isl_ast_graft_add_guard(graft
, isl_set_copy(guard
), build
);
3439 list
= isl_ast_graft_list_set_ast_graft(list
, 0, graft
);
3444 /* Generate code for a single component, after shifting (if any)
3445 * has been applied, in case the schedule was specified as a schedule tree.
3446 * In particular, do so for the specified subset of the schedule domain.
3448 * If we are outside of the isolated part, then "domain" may include
3449 * a disjunction. Explicitly generate this disjunction at this point
3450 * instead of relying on the disjunction getting hoisted back up
3453 static __isl_give isl_ast_graft_list
*generate_shifted_component_tree_part(
3454 __isl_keep isl_union_map
*executed
, __isl_take isl_set
*domain
,
3455 __isl_keep isl_ast_build
*build
, int isolated
)
3457 isl_union_set
*uset
;
3458 isl_ast_graft_list
*list
;
3459 isl_ast_build
*sub_build
;
3462 uset
= isl_union_set_from_set(isl_set_copy(domain
));
3463 executed
= isl_union_map_copy(executed
);
3464 executed
= isl_union_map_intersect_domain(executed
, uset
);
3465 empty
= isl_union_map_is_empty(executed
);
3470 isl_union_map_free(executed
);
3471 isl_set_free(domain
);
3472 ctx
= isl_ast_build_get_ctx(build
);
3473 return isl_ast_graft_list_alloc(ctx
, 0);
3476 sub_build
= isl_ast_build_copy(build
);
3478 domain
= extract_disjunction(domain
, build
);
3479 sub_build
= isl_ast_build_restrict_generated(sub_build
,
3480 isl_set_copy(domain
));
3482 list
= generate_shifted_component_tree_base(executed
,
3483 isl_ast_build_copy(sub_build
), isolated
);
3485 list
= list_add_guard(list
, domain
, build
, sub_build
);
3486 isl_ast_build_free(sub_build
);
3487 isl_set_free(domain
);
3490 isl_union_map_free(executed
);
3491 isl_set_free(domain
);
3495 /* Generate code for a single component, after shifting (if any)
3496 * has been applied, in case the schedule was specified as a schedule tree.
3497 * In particular, do so for the specified sequence of subsets
3498 * of the schedule domain, "before", "isolated", "after" and "other",
3499 * where only the "isolated" part is considered to be isolated.
3501 static __isl_give isl_ast_graft_list
*generate_shifted_component_parts(
3502 __isl_take isl_union_map
*executed
, __isl_take isl_set
*before
,
3503 __isl_take isl_set
*isolated
, __isl_take isl_set
*after
,
3504 __isl_take isl_set
*other
, __isl_take isl_ast_build
*build
)
3506 isl_ast_graft_list
*list
, *res
;
3508 res
= generate_shifted_component_tree_part(executed
, before
, build
, 0);
3509 list
= generate_shifted_component_tree_part(executed
, isolated
,
3511 res
= isl_ast_graft_list_concat(res
, list
);
3512 list
= generate_shifted_component_tree_part(executed
, after
, build
, 0);
3513 res
= isl_ast_graft_list_concat(res
, list
);
3514 list
= generate_shifted_component_tree_part(executed
, other
, build
, 0);
3515 res
= isl_ast_graft_list_concat(res
, list
);
3517 isl_union_map_free(executed
);
3518 isl_ast_build_free(build
);
3523 /* Does "set" intersect "first", but not "second"?
3525 static isl_bool
only_intersects_first(__isl_keep isl_set
*set
,
3526 __isl_keep isl_set
*first
, __isl_keep isl_set
*second
)
3530 disjoint
= isl_set_is_disjoint(set
, first
);
3532 return isl_bool_error
;
3534 return isl_bool_false
;
3536 return isl_set_is_disjoint(set
, second
);
3539 /* Generate code for a single component, after shifting (if any)
3540 * has been applied, in case the schedule was specified as a schedule tree.
3541 * In particular, do so in case of isolation where there is
3542 * only an "isolated" part and an "after" part.
3543 * "dead1" and "dead2" are freed by this function in order to simplify
3546 * The "before" and "other" parts are set to empty sets.
3548 static __isl_give isl_ast_graft_list
*generate_shifted_component_only_after(
3549 __isl_take isl_union_map
*executed
, __isl_take isl_set
*isolated
,
3550 __isl_take isl_set
*after
, __isl_take isl_ast_build
*build
,
3551 __isl_take isl_set
*dead1
, __isl_take isl_set
*dead2
)
3555 empty
= isl_set_empty(isl_set_get_space(after
));
3556 isl_set_free(dead1
);
3557 isl_set_free(dead2
);
3558 return generate_shifted_component_parts(executed
, isl_set_copy(empty
),
3559 isolated
, after
, empty
, build
);
3562 /* Generate code for a single component, after shifting (if any)
3563 * has been applied, in case the schedule was specified as a schedule tree.
3565 * We first check if the user has specified an isolated schedule domain
3566 * and that we are not already outside of this isolated schedule domain.
3567 * If so, we break up the schedule domain into iterations that
3568 * precede the isolated domain, the isolated domain itself,
3569 * the iterations that follow the isolated domain and
3570 * the remaining iterations (those that are incomparable
3571 * to the isolated domain).
3572 * We generate an AST for each piece and concatenate the results.
3574 * If the isolated domain is not convex, then it is replaced
3575 * by a convex superset to ensure that the sets of preceding and
3576 * following iterations are properly defined and, in particular,
3577 * that there are no intermediate iterations that do not belong
3578 * to the isolated domain.
3580 * In the special case where at least one element of the schedule
3581 * domain that does not belong to the isolated domain needs
3582 * to be scheduled after this isolated domain, but none of those
3583 * elements need to be scheduled before, break up the schedule domain
3584 * in only two parts, the isolated domain, and a part that will be
3585 * scheduled after the isolated domain.
3587 * If no isolated set has been specified, then we generate an
3588 * AST for the entire inverse schedule.
3590 static __isl_give isl_ast_graft_list
*generate_shifted_component_tree(
3591 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
3595 int empty
, has_isolate
;
3597 isl_union_set
*schedule_domain
;
3599 isl_basic_set
*hull
;
3600 isl_set
*isolated
, *before
, *after
, *test
;
3604 build
= isl_ast_build_extract_isolated(build
);
3605 has_isolate
= isl_ast_build_has_isolated(build
);
3606 if (has_isolate
< 0)
3607 executed
= isl_union_map_free(executed
);
3608 else if (!has_isolate
)
3609 return generate_shifted_component_tree_base(executed
, build
, 0);
3611 schedule_domain
= isl_union_map_domain(isl_union_map_copy(executed
));
3612 domain
= isl_set_from_union_set(schedule_domain
);
3614 isolated
= isl_ast_build_get_isolated(build
);
3615 isolated
= isl_set_intersect(isolated
, isl_set_copy(domain
));
3616 test
= isl_ast_build_specialize(build
, isl_set_copy(isolated
));
3617 empty
= isl_set_is_empty(test
);
3622 isl_set_free(isolated
);
3623 isl_set_free(domain
);
3624 return generate_shifted_component_tree_base(executed
, build
, 0);
3626 depth
= isl_ast_build_get_depth(build
);
3630 isolated
= isl_ast_build_eliminate(build
, isolated
);
3631 hull
= isl_set_unshifted_simple_hull(isolated
);
3632 isolated
= isl_set_from_basic_set(hull
);
3634 space
= isl_space_map_from_set(isl_set_get_space(isolated
));
3635 gt
= isl_map_universe(space
);
3636 for (i
= 0; i
< depth
; ++i
)
3637 gt
= isl_map_equate(gt
, isl_dim_in
, i
, isl_dim_out
, i
);
3638 gt
= isl_map_order_gt(gt
, isl_dim_in
, depth
, isl_dim_out
, depth
);
3639 lt
= isl_map_reverse(isl_map_copy(gt
));
3640 before
= isl_set_apply(isl_set_copy(isolated
), gt
);
3641 after
= isl_set_apply(isl_set_copy(isolated
), lt
);
3643 domain
= isl_set_subtract(domain
, isl_set_copy(isolated
));
3644 pure
= only_intersects_first(domain
, after
, before
);
3646 executed
= isl_union_map_free(executed
);
3648 return generate_shifted_component_only_after(executed
, isolated
,
3649 domain
, build
, before
, after
);
3650 domain
= isl_set_subtract(domain
, isl_set_copy(before
));
3651 domain
= isl_set_subtract(domain
, isl_set_copy(after
));
3652 after
= isl_set_subtract(after
, isl_set_copy(isolated
));
3653 after
= isl_set_subtract(after
, isl_set_copy(before
));
3654 before
= isl_set_subtract(before
, isl_set_copy(isolated
));
3656 return generate_shifted_component_parts(executed
, before
, isolated
,
3657 after
, domain
, build
);
3659 isl_set_free(domain
);
3660 isl_set_free(isolated
);
3661 isl_union_map_free(executed
);
3662 isl_ast_build_free(build
);
3666 /* Generate code for a single component, after shifting (if any)
3669 * Call generate_shifted_component_tree or generate_shifted_component_flat
3670 * depending on whether the schedule was specified as a schedule tree.
3672 static __isl_give isl_ast_graft_list
*generate_shifted_component(
3673 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
3675 if (isl_ast_build_has_schedule_node(build
))
3676 return generate_shifted_component_tree(executed
, build
);
3678 return generate_shifted_component_flat(executed
, build
);
3681 struct isl_set_map_pair
{
3686 /* Given an array "domain" of isl_set_map_pairs and an array "order"
3687 * of indices into the "domain" array,
3688 * return the union of the "map" fields of the elements
3689 * indexed by the first "n" elements of "order".
3691 static __isl_give isl_union_map
*construct_component_executed(
3692 struct isl_set_map_pair
*domain
, int *order
, int n
)
3696 isl_union_map
*executed
;
3698 map
= isl_map_copy(domain
[order
[0]].map
);
3699 executed
= isl_union_map_from_map(map
);
3700 for (i
= 1; i
< n
; ++i
) {
3701 map
= isl_map_copy(domain
[order
[i
]].map
);
3702 executed
= isl_union_map_add_map(executed
, map
);
3708 /* Generate code for a single component, after shifting (if any)
3711 * The component inverse schedule is specified as the "map" fields
3712 * of the elements of "domain" indexed by the first "n" elements of "order".
3714 static __isl_give isl_ast_graft_list
*generate_shifted_component_from_list(
3715 struct isl_set_map_pair
*domain
, int *order
, int n
,
3716 __isl_take isl_ast_build
*build
)
3718 isl_union_map
*executed
;
3720 executed
= construct_component_executed(domain
, order
, n
);
3721 return generate_shifted_component(executed
, build
);
3724 /* Does set dimension "pos" of "set" have an obviously fixed value?
3726 static int dim_is_fixed(__isl_keep isl_set
*set
, int pos
)
3731 v
= isl_set_plain_get_val_if_fixed(set
, isl_dim_set
, pos
);
3734 fixed
= !isl_val_is_nan(v
);
3740 /* Given an array "domain" of isl_set_map_pairs and an array "order"
3741 * of indices into the "domain" array,
3742 * do all (except for at most one) of the "set" field of the elements
3743 * indexed by the first "n" elements of "order" have a fixed value
3744 * at position "depth"?
3746 static int at_most_one_non_fixed(struct isl_set_map_pair
*domain
,
3747 int *order
, int n
, int depth
)
3752 for (i
= 0; i
< n
; ++i
) {
3755 f
= dim_is_fixed(domain
[order
[i
]].set
, depth
);
3768 /* Given an array "domain" of isl_set_map_pairs and an array "order"
3769 * of indices into the "domain" array,
3770 * eliminate the inner dimensions from the "set" field of the elements
3771 * indexed by the first "n" elements of "order", provided the current
3772 * dimension does not have a fixed value.
3774 * Return the index of the first element in "order" with a corresponding
3775 * "set" field that does not have an (obviously) fixed value.
3777 static int eliminate_non_fixed(struct isl_set_map_pair
*domain
,
3778 int *order
, int n
, int depth
, __isl_keep isl_ast_build
*build
)
3783 for (i
= n
- 1; i
>= 0; --i
) {
3785 f
= dim_is_fixed(domain
[order
[i
]].set
, depth
);
3790 domain
[order
[i
]].set
= isl_ast_build_eliminate_inner(build
,
3791 domain
[order
[i
]].set
);
3798 /* Given an array "domain" of isl_set_map_pairs and an array "order"
3799 * of indices into the "domain" array,
3800 * find the element of "domain" (amongst those indexed by the first "n"
3801 * elements of "order") with the "set" field that has the smallest
3802 * value for the current iterator.
3804 * Note that the domain with the smallest value may depend on the parameters
3805 * and/or outer loop dimension. Since the result of this function is only
3806 * used as heuristic, we only make a reasonable attempt at finding the best
3807 * domain, one that should work in case a single domain provides the smallest
3808 * value for the current dimension over all values of the parameters
3809 * and outer dimensions.
3811 * In particular, we compute the smallest value of the first domain
3812 * and replace it by that of any later domain if that later domain
3813 * has a smallest value that is smaller for at least some value
3814 * of the parameters and outer dimensions.
3816 static int first_offset(struct isl_set_map_pair
*domain
, int *order
, int n
,
3817 __isl_keep isl_ast_build
*build
)
3823 min_first
= isl_ast_build_map_to_iterator(build
,
3824 isl_set_copy(domain
[order
[0]].set
));
3825 min_first
= isl_map_lexmin(min_first
);
3827 for (i
= 1; i
< n
; ++i
) {
3828 isl_map
*min
, *test
;
3831 min
= isl_ast_build_map_to_iterator(build
,
3832 isl_set_copy(domain
[order
[i
]].set
));
3833 min
= isl_map_lexmin(min
);
3834 test
= isl_map_copy(min
);
3835 test
= isl_map_apply_domain(isl_map_copy(min_first
), test
);
3836 test
= isl_map_order_lt(test
, isl_dim_in
, 0, isl_dim_out
, 0);
3837 empty
= isl_map_is_empty(test
);
3839 if (empty
>= 0 && !empty
) {
3840 isl_map_free(min_first
);
3850 isl_map_free(min_first
);
3852 return i
< n
? -1 : first
;
3855 /* Construct a shifted inverse schedule based on the original inverse schedule,
3856 * the stride and the offset.
3858 * The original inverse schedule is specified as the "map" fields
3859 * of the elements of "domain" indexed by the first "n" elements of "order".
3861 * "stride" and "offset" are such that the difference
3862 * between the values of the current dimension of domain "i"
3863 * and the values of the current dimension for some reference domain are
3866 * stride * integer + offset[i]
3868 * Moreover, 0 <= offset[i] < stride.
3870 * For each domain, we create a map
3872 * { [..., j, ...] -> [..., j - offset[i], offset[i], ....] }
3874 * where j refers to the current dimension and the other dimensions are
3875 * unchanged, and apply this map to the original schedule domain.
3877 * For example, for the original schedule
3879 * { A[i] -> [2i]: 0 <= i < 10; B[i] -> [2i+1] : 0 <= i < 10 }
3881 * and assuming the offset is 0 for the A domain and 1 for the B domain,
3882 * we apply the mapping
3886 * to the schedule of the "A" domain and the mapping
3888 * { [j - 1] -> [j, 1] }
3890 * to the schedule of the "B" domain.
3893 * Note that after the transformation, the differences between pairs
3894 * of values of the current dimension over all domains are multiples
3895 * of stride and that we have therefore exposed the stride.
3898 * To see that the mapping preserves the lexicographic order,
3899 * first note that each of the individual maps above preserves the order.
3900 * If the value of the current iterator is j1 in one domain and j2 in another,
3901 * then if j1 = j2, we know that the same map is applied to both domains
3902 * and the order is preserved.
3903 * Otherwise, let us assume, without loss of generality, that j1 < j2.
3904 * If c1 >= c2 (with c1 and c2 the corresponding offsets), then
3908 * and the order is preserved.
3909 * If c1 < c2, then we know
3915 * j2 - j1 = n * s + r
3917 * with n >= 0 and 0 <= r < s.
3918 * In other words, r = c2 - c1.
3929 * (j1 - c1, c1) << (j2 - c2, c2)
3931 * with "<<" the lexicographic order, proving that the order is preserved
3934 static __isl_give isl_union_map
*construct_shifted_executed(
3935 struct isl_set_map_pair
*domain
, int *order
, int n
,
3936 __isl_keep isl_val
*stride
, __isl_keep isl_multi_val
*offset
,
3937 __isl_keep isl_ast_build
*build
)
3940 isl_union_map
*executed
;
3946 depth
= isl_ast_build_get_depth(build
);
3949 space
= isl_ast_build_get_space(build
, 1);
3950 executed
= isl_union_map_empty(isl_space_copy(space
));
3951 space
= isl_space_map_from_set(space
);
3952 map
= isl_map_identity(isl_space_copy(space
));
3953 map
= isl_map_eliminate(map
, isl_dim_out
, depth
, 1);
3954 map
= isl_map_insert_dims(map
, isl_dim_out
, depth
+ 1, 1);
3955 space
= isl_space_insert_dims(space
, isl_dim_out
, depth
+ 1, 1);
3957 c
= isl_constraint_alloc_equality(isl_local_space_from_space(space
));
3958 c
= isl_constraint_set_coefficient_si(c
, isl_dim_in
, depth
, 1);
3959 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
, depth
, -1);
3961 for (i
= 0; i
< n
; ++i
) {
3965 v
= isl_multi_val_get_val(offset
, i
);
3968 map_i
= isl_map_copy(map
);
3969 map_i
= isl_map_fix_val(map_i
, isl_dim_out
, depth
+ 1,
3972 c
= isl_constraint_set_constant_val(c
, v
);
3973 map_i
= isl_map_add_constraint(map_i
, isl_constraint_copy(c
));
3975 map_i
= isl_map_apply_domain(isl_map_copy(domain
[order
[i
]].map
),
3977 executed
= isl_union_map_add_map(executed
, map_i
);
3980 isl_constraint_free(c
);
3984 executed
= isl_union_map_free(executed
);
3989 /* Generate code for a single component, after exposing the stride,
3990 * given that the schedule domain is "shifted strided".
3992 * The component inverse schedule is specified as the "map" fields
3993 * of the elements of "domain" indexed by the first "n" elements of "order".
3995 * The schedule domain being "shifted strided" means that the differences
3996 * between the values of the current dimension of domain "i"
3997 * and the values of the current dimension for some reference domain are
4000 * stride * integer + offset[i]
4002 * We first look for the domain with the "smallest" value for the current
4003 * dimension and adjust the offsets such that the offset of the "smallest"
4004 * domain is equal to zero. The other offsets are reduced modulo stride.
4006 * Based on this information, we construct a new inverse schedule in
4007 * construct_shifted_executed that exposes the stride.
4008 * Since this involves the introduction of a new schedule dimension,
4009 * the build needs to be changed accordingly.
4010 * After computing the AST, the newly introduced dimension needs
4011 * to be removed again from the list of grafts. We do this by plugging
4012 * in a mapping that represents the new schedule domain in terms of the
4013 * old schedule domain.
4015 static __isl_give isl_ast_graft_list
*generate_shift_component(
4016 struct isl_set_map_pair
*domain
, int *order
, int n
,
4017 __isl_keep isl_val
*stride
, __isl_keep isl_multi_val
*offset
,
4018 __isl_take isl_ast_build
*build
)
4020 isl_ast_graft_list
*list
;
4026 isl_multi_aff
*ma
, *zero
;
4027 isl_union_map
*executed
;
4029 depth
= isl_ast_build_get_depth(build
);
4031 first
= first_offset(domain
, order
, n
, build
);
4032 if (depth
< 0 || first
< 0)
4035 mv
= isl_multi_val_copy(offset
);
4036 val
= isl_multi_val_get_val(offset
, first
);
4037 val
= isl_val_neg(val
);
4038 mv
= isl_multi_val_add_val(mv
, val
);
4039 mv
= isl_multi_val_mod_val(mv
, isl_val_copy(stride
));
4041 executed
= construct_shifted_executed(domain
, order
, n
, stride
, mv
,
4043 space
= isl_ast_build_get_space(build
, 1);
4044 space
= isl_space_map_from_set(space
);
4045 ma
= isl_multi_aff_identity(isl_space_copy(space
));
4046 space
= isl_space_from_domain(isl_space_domain(space
));
4047 space
= isl_space_add_dims(space
, isl_dim_out
, 1);
4048 zero
= isl_multi_aff_zero(space
);
4049 ma
= isl_multi_aff_range_splice(ma
, depth
+ 1, zero
);
4050 build
= isl_ast_build_insert_dim(build
, depth
+ 1);
4051 list
= generate_shifted_component(executed
, build
);
4053 list
= isl_ast_graft_list_preimage_multi_aff(list
, ma
);
4055 isl_multi_val_free(mv
);
4059 isl_ast_build_free(build
);
4063 /* Does any node in the schedule tree rooted at the current schedule node
4064 * of "build" depend on outer schedule nodes?
4066 static int has_anchored_subtree(__isl_keep isl_ast_build
*build
)
4068 isl_schedule_node
*node
;
4071 node
= isl_ast_build_get_schedule_node(build
);
4072 dependent
= isl_schedule_node_is_subtree_anchored(node
);
4073 isl_schedule_node_free(node
);
4078 /* Generate code for a single component.
4080 * The component inverse schedule is specified as the "map" fields
4081 * of the elements of "domain" indexed by the first "n" elements of "order".
4083 * This function may modify the "set" fields of "domain".
4085 * Before proceeding with the actual code generation for the component,
4086 * we first check if there are any "shifted" strides, meaning that
4087 * the schedule domains of the individual domains are all strided,
4088 * but that they have different offsets, resulting in the union
4089 * of schedule domains not being strided anymore.
4091 * The simplest example is the schedule
4093 * { A[i] -> [2i]: 0 <= i < 10; B[i] -> [2i+1] : 0 <= i < 10 }
4095 * Both schedule domains are strided, but their union is not.
4096 * This function detects such cases and then rewrites the schedule to
4098 * { A[i] -> [2i, 0]: 0 <= i < 10; B[i] -> [2i, 1] : 0 <= i < 10 }
4100 * In the new schedule, the schedule domains have the same offset (modulo
4101 * the stride), ensuring that the union of schedule domains is also strided.
4104 * If there is only a single domain in the component, then there is
4105 * nothing to do. Similarly, if the current schedule dimension has
4106 * a fixed value for almost all domains then there is nothing to be done.
4107 * In particular, we need at least two domains where the current schedule
4108 * dimension does not have a fixed value.
4109 * Finally, in case of a schedule map input,
4110 * if any of the options refer to the current schedule dimension,
4111 * then we bail out as well. It would be possible to reformulate the options
4112 * in terms of the new schedule domain, but that would introduce constraints
4113 * that separate the domains in the options and that is something we would
4115 * In the case of a schedule tree input, we bail out if any of
4116 * the descendants of the current schedule node refer to outer
4117 * schedule nodes in any way.
4120 * To see if there is any shifted stride, we look at the differences
4121 * between the values of the current dimension in pairs of domains
4122 * for equal values of outer dimensions. These differences should be
4127 * with "m" the stride and "r" a constant. Note that we cannot perform
4128 * this analysis on individual domains as the lower bound in each domain
4129 * may depend on parameters or outer dimensions and so the current dimension
4130 * itself may not have a fixed remainder on division by the stride.
4132 * In particular, we compare the first domain that does not have an
4133 * obviously fixed value for the current dimension to itself and all
4134 * other domains and collect the offsets and the gcd of the strides.
4135 * If the gcd becomes one, then we failed to find shifted strides.
4136 * If the gcd is zero, then the differences were all fixed, meaning
4137 * that some domains had non-obviously fixed values for the current dimension.
4138 * If all the offsets are the same (for those domains that do not have
4139 * an obviously fixed value for the current dimension), then we do not
4140 * apply the transformation.
4141 * If none of the domains were skipped, then there is nothing to do.
4142 * If some of them were skipped, then if we apply separation, the schedule
4143 * domain should get split in pieces with a (non-shifted) stride.
4145 * Otherwise, we apply a shift to expose the stride in
4146 * generate_shift_component.
4148 static __isl_give isl_ast_graft_list
*generate_component(
4149 struct isl_set_map_pair
*domain
, int *order
, int n
,
4150 __isl_take isl_ast_build
*build
)
4157 isl_val
*gcd
= NULL
;
4161 isl_ast_graft_list
*list
;
4164 depth
= isl_ast_build_get_depth(build
);
4169 if (skip
>= 0 && !skip
)
4170 skip
= at_most_one_non_fixed(domain
, order
, n
, depth
);
4171 if (skip
>= 0 && !skip
) {
4172 if (isl_ast_build_has_schedule_node(build
))
4173 skip
= has_anchored_subtree(build
);
4175 skip
= isl_ast_build_options_involve_depth(build
);
4180 return generate_shifted_component_from_list(domain
,
4183 base
= eliminate_non_fixed(domain
, order
, n
, depth
, build
);
4187 ctx
= isl_ast_build_get_ctx(build
);
4189 mv
= isl_multi_val_zero(isl_space_set_alloc(ctx
, 0, n
));
4192 for (i
= 0; i
< n
; ++i
) {
4195 map
= isl_map_from_domain_and_range(
4196 isl_set_copy(domain
[order
[base
]].set
),
4197 isl_set_copy(domain
[order
[i
]].set
));
4198 for (d
= 0; d
< depth
; ++d
)
4199 map
= isl_map_equate(map
, isl_dim_in
, d
,
4201 deltas
= isl_map_deltas(map
);
4202 res
= isl_set_dim_residue_class_val(deltas
, depth
, &m
, &r
);
4203 isl_set_free(deltas
);
4210 gcd
= isl_val_gcd(gcd
, m
);
4211 if (isl_val_is_one(gcd
)) {
4215 mv
= isl_multi_val_set_val(mv
, i
, r
);
4217 res
= dim_is_fixed(domain
[order
[i
]].set
, depth
);
4223 if (fixed
&& i
> base
) {
4225 a
= isl_multi_val_get_val(mv
, i
);
4226 b
= isl_multi_val_get_val(mv
, base
);
4227 if (isl_val_ne(a
, b
))
4234 if (res
< 0 || !gcd
) {
4235 isl_ast_build_free(build
);
4237 } else if (i
< n
|| fixed
|| isl_val_is_zero(gcd
)) {
4238 list
= generate_shifted_component_from_list(domain
,
4241 list
= generate_shift_component(domain
, order
, n
, gcd
, mv
,
4246 isl_multi_val_free(mv
);
4250 isl_ast_build_free(build
);
4254 /* Store both "map" itself and its domain in the
4255 * structure pointed to by *next and advance to the next array element.
4257 static isl_stat
extract_domain(__isl_take isl_map
*map
, void *user
)
4259 struct isl_set_map_pair
**next
= user
;
4261 (*next
)->map
= isl_map_copy(map
);
4262 (*next
)->set
= isl_map_domain(map
);
4268 static isl_bool
after_in_tree(__isl_keep isl_union_map
*umap
,
4269 __isl_keep isl_schedule_node
*node
);
4271 /* Is any domain element of "umap" scheduled after any of
4272 * the corresponding image elements by the tree rooted at
4273 * the child of "node"?
4275 static isl_bool
after_in_child(__isl_keep isl_union_map
*umap
,
4276 __isl_keep isl_schedule_node
*node
)
4278 isl_schedule_node
*child
;
4281 child
= isl_schedule_node_get_child(node
, 0);
4282 after
= after_in_tree(umap
, child
);
4283 isl_schedule_node_free(child
);
4288 /* Is any domain element of "umap" scheduled after any of
4289 * the corresponding image elements by the tree rooted at
4290 * the band node "node"?
4292 * We first check if any domain element is scheduled after any
4293 * of the corresponding image elements by the band node itself.
4294 * If not, we restrict "map" to those pairs of element that
4295 * are scheduled together by the band node and continue with
4296 * the child of the band node.
4297 * If there are no such pairs then the map passed to after_in_child
4298 * will be empty causing it to return 0.
4300 static isl_bool
after_in_band(__isl_keep isl_union_map
*umap
,
4301 __isl_keep isl_schedule_node
*node
)
4303 isl_multi_union_pw_aff
*mupa
;
4304 isl_union_map
*partial
, *test
, *gt
, *universe
, *umap1
, *umap2
;
4305 isl_union_set
*domain
, *range
;
4311 n
= isl_schedule_node_band_n_member(node
);
4313 return isl_bool_error
;
4315 return after_in_child(umap
, node
);
4317 mupa
= isl_schedule_node_band_get_partial_schedule(node
);
4318 space
= isl_multi_union_pw_aff_get_space(mupa
);
4319 partial
= isl_union_map_from_multi_union_pw_aff(mupa
);
4320 test
= isl_union_map_copy(umap
);
4321 test
= isl_union_map_apply_domain(test
, isl_union_map_copy(partial
));
4322 test
= isl_union_map_apply_range(test
, isl_union_map_copy(partial
));
4323 gt
= isl_union_map_from_map(isl_map_lex_gt(space
));
4324 test
= isl_union_map_intersect(test
, gt
);
4325 empty
= isl_union_map_is_empty(test
);
4326 isl_union_map_free(test
);
4328 if (empty
< 0 || !empty
) {
4329 isl_union_map_free(partial
);
4330 return isl_bool_not(empty
);
4333 universe
= isl_union_map_universe(isl_union_map_copy(umap
));
4334 domain
= isl_union_map_domain(isl_union_map_copy(universe
));
4335 range
= isl_union_map_range(universe
);
4336 umap1
= isl_union_map_copy(partial
);
4337 umap1
= isl_union_map_intersect_domain(umap1
, domain
);
4338 umap2
= isl_union_map_intersect_domain(partial
, range
);
4339 test
= isl_union_map_apply_range(umap1
, isl_union_map_reverse(umap2
));
4340 test
= isl_union_map_intersect(test
, isl_union_map_copy(umap
));
4341 after
= after_in_child(test
, node
);
4342 isl_union_map_free(test
);
4346 /* Is any domain element of "umap" scheduled after any of
4347 * the corresponding image elements by the tree rooted at
4348 * the context node "node"?
4350 * The context constraints apply to the schedule domain,
4351 * so we cannot apply them directly to "umap", which contains
4352 * pairs of statement instances. Instead, we add them
4353 * to the range of the prefix schedule for both domain and
4356 static isl_bool
after_in_context(__isl_keep isl_union_map
*umap
,
4357 __isl_keep isl_schedule_node
*node
)
4359 isl_union_map
*prefix
, *universe
, *umap1
, *umap2
;
4360 isl_union_set
*domain
, *range
;
4364 umap
= isl_union_map_copy(umap
);
4365 context
= isl_schedule_node_context_get_context(node
);
4366 prefix
= isl_schedule_node_get_prefix_schedule_union_map(node
);
4367 universe
= isl_union_map_universe(isl_union_map_copy(umap
));
4368 domain
= isl_union_map_domain(isl_union_map_copy(universe
));
4369 range
= isl_union_map_range(universe
);
4370 umap1
= isl_union_map_copy(prefix
);
4371 umap1
= isl_union_map_intersect_domain(umap1
, domain
);
4372 umap2
= isl_union_map_intersect_domain(prefix
, range
);
4373 umap1
= isl_union_map_intersect_range(umap1
,
4374 isl_union_set_from_set(context
));
4375 umap1
= isl_union_map_apply_range(umap1
, isl_union_map_reverse(umap2
));
4376 umap
= isl_union_map_intersect(umap
, umap1
);
4378 after
= after_in_child(umap
, node
);
4380 isl_union_map_free(umap
);
4385 /* Is any domain element of "umap" scheduled after any of
4386 * the corresponding image elements by the tree rooted at
4387 * the expansion node "node"?
4389 * We apply the expansion to domain and range of "umap" and
4390 * continue with its child.
4392 static isl_bool
after_in_expansion(__isl_keep isl_union_map
*umap
,
4393 __isl_keep isl_schedule_node
*node
)
4395 isl_union_map
*expansion
;
4398 expansion
= isl_schedule_node_expansion_get_expansion(node
);
4399 umap
= isl_union_map_copy(umap
);
4400 umap
= isl_union_map_apply_domain(umap
, isl_union_map_copy(expansion
));
4401 umap
= isl_union_map_apply_range(umap
, expansion
);
4403 after
= after_in_child(umap
, node
);
4405 isl_union_map_free(umap
);
4410 /* Is any domain element of "umap" scheduled after any of
4411 * the corresponding image elements by the tree rooted at
4412 * the extension node "node"?
4414 * Since the extension node may add statement instances before or
4415 * after the pairs of statement instances in "umap", we return isl_bool_true
4416 * to ensure that these pairs are not broken up.
4418 static isl_bool
after_in_extension(__isl_keep isl_union_map
*umap
,
4419 __isl_keep isl_schedule_node
*node
)
4421 return isl_bool_true
;
4424 /* Is any domain element of "umap" scheduled after any of
4425 * the corresponding image elements by the tree rooted at
4426 * the filter node "node"?
4428 * We intersect domain and range of "umap" with the filter and
4429 * continue with its child.
4431 static isl_bool
after_in_filter(__isl_keep isl_union_map
*umap
,
4432 __isl_keep isl_schedule_node
*node
)
4434 isl_union_set
*filter
;
4437 umap
= isl_union_map_copy(umap
);
4438 filter
= isl_schedule_node_filter_get_filter(node
);
4439 umap
= isl_union_map_intersect_domain(umap
, isl_union_set_copy(filter
));
4440 umap
= isl_union_map_intersect_range(umap
, filter
);
4442 after
= after_in_child(umap
, node
);
4444 isl_union_map_free(umap
);
4449 /* Is any domain element of "umap" scheduled after any of
4450 * the corresponding image elements by the tree rooted at
4451 * the set node "node"?
4453 * This is only the case if this condition holds in any
4454 * of the (filter) children of the set node.
4455 * In particular, if the domain and the range of "umap"
4456 * are contained in different children, then the condition
4459 static isl_bool
after_in_set(__isl_keep isl_union_map
*umap
,
4460 __isl_keep isl_schedule_node
*node
)
4465 n
= isl_schedule_node_n_children(node
);
4467 return isl_bool_error
;
4468 for (i
= 0; i
< n
; ++i
) {
4469 isl_schedule_node
*child
;
4472 child
= isl_schedule_node_get_child(node
, i
);
4473 after
= after_in_tree(umap
, child
);
4474 isl_schedule_node_free(child
);
4476 if (after
< 0 || after
)
4480 return isl_bool_false
;
4483 /* Return the filter of child "i" of "node".
4485 static __isl_give isl_union_set
*child_filter(
4486 __isl_keep isl_schedule_node
*node
, int i
)
4488 isl_schedule_node
*child
;
4489 isl_union_set
*filter
;
4491 child
= isl_schedule_node_get_child(node
, i
);
4492 filter
= isl_schedule_node_filter_get_filter(child
);
4493 isl_schedule_node_free(child
);
4498 /* Is any domain element of "umap" scheduled after any of
4499 * the corresponding image elements by the tree rooted at
4500 * the sequence node "node"?
4502 * This happens in particular if any domain element is
4503 * contained in a later child than one containing a range element or
4504 * if the condition holds within a given child in the sequence.
4505 * The later part of the condition is checked by after_in_set.
4507 static isl_bool
after_in_sequence(__isl_keep isl_union_map
*umap
,
4508 __isl_keep isl_schedule_node
*node
)
4512 isl_union_map
*umap_i
;
4514 isl_bool after
= isl_bool_false
;
4516 n
= isl_schedule_node_n_children(node
);
4518 return isl_bool_error
;
4519 for (i
= 1; i
< n
; ++i
) {
4520 isl_union_set
*filter_i
;
4522 umap_i
= isl_union_map_copy(umap
);
4523 filter_i
= child_filter(node
, i
);
4524 umap_i
= isl_union_map_intersect_domain(umap_i
, filter_i
);
4525 empty
= isl_union_map_is_empty(umap_i
);
4529 isl_union_map_free(umap_i
);
4533 for (j
= 0; j
< i
; ++j
) {
4534 isl_union_set
*filter_j
;
4535 isl_union_map
*umap_ij
;
4537 umap_ij
= isl_union_map_copy(umap_i
);
4538 filter_j
= child_filter(node
, j
);
4539 umap_ij
= isl_union_map_intersect_range(umap_ij
,
4541 empty
= isl_union_map_is_empty(umap_ij
);
4542 isl_union_map_free(umap_ij
);
4547 after
= isl_bool_true
;
4552 isl_union_map_free(umap_i
);
4557 if (after
< 0 || after
)
4560 return after_in_set(umap
, node
);
4562 isl_union_map_free(umap_i
);
4563 return isl_bool_error
;
4566 /* Is any domain element of "umap" scheduled after any of
4567 * the corresponding image elements by the tree rooted at "node"?
4569 * If "umap" is empty, then clearly there is no such element.
4570 * Otherwise, consider the different types of nodes separately.
4572 static isl_bool
after_in_tree(__isl_keep isl_union_map
*umap
,
4573 __isl_keep isl_schedule_node
*node
)
4576 enum isl_schedule_node_type type
;
4578 empty
= isl_union_map_is_empty(umap
);
4580 return isl_bool_error
;
4582 return isl_bool_false
;
4584 return isl_bool_error
;
4586 type
= isl_schedule_node_get_type(node
);
4588 case isl_schedule_node_error
:
4589 return isl_bool_error
;
4590 case isl_schedule_node_leaf
:
4591 return isl_bool_false
;
4592 case isl_schedule_node_band
:
4593 return after_in_band(umap
, node
);
4594 case isl_schedule_node_domain
:
4595 isl_die(isl_schedule_node_get_ctx(node
), isl_error_internal
,
4596 "unexpected internal domain node",
4597 return isl_bool_error
);
4598 case isl_schedule_node_context
:
4599 return after_in_context(umap
, node
);
4600 case isl_schedule_node_expansion
:
4601 return after_in_expansion(umap
, node
);
4602 case isl_schedule_node_extension
:
4603 return after_in_extension(umap
, node
);
4604 case isl_schedule_node_filter
:
4605 return after_in_filter(umap
, node
);
4606 case isl_schedule_node_guard
:
4607 case isl_schedule_node_mark
:
4608 return after_in_child(umap
, node
);
4609 case isl_schedule_node_set
:
4610 return after_in_set(umap
, node
);
4611 case isl_schedule_node_sequence
:
4612 return after_in_sequence(umap
, node
);
4615 return isl_bool_true
;
4618 /* Is any domain element of "map1" scheduled after any domain
4619 * element of "map2" by the subtree underneath the current band node,
4620 * while at the same time being scheduled together by the current
4621 * band node, i.e., by "map1" and "map2?
4623 * If the child of the current band node is a leaf, then
4624 * no element can be scheduled after any other element.
4626 * Otherwise, we construct a relation between domain elements
4627 * of "map1" and domain elements of "map2" that are scheduled
4628 * together and then check if the subtree underneath the current
4629 * band node determines their relative order.
4631 static isl_bool
after_in_subtree(__isl_keep isl_ast_build
*build
,
4632 __isl_keep isl_map
*map1
, __isl_keep isl_map
*map2
)
4634 isl_schedule_node
*node
;
4636 isl_union_map
*umap
;
4639 node
= isl_ast_build_get_schedule_node(build
);
4641 return isl_bool_error
;
4642 node
= isl_schedule_node_child(node
, 0);
4643 if (isl_schedule_node_get_type(node
) == isl_schedule_node_leaf
) {
4644 isl_schedule_node_free(node
);
4645 return isl_bool_false
;
4647 map
= isl_map_copy(map2
);
4648 map
= isl_map_apply_domain(map
, isl_map_copy(map1
));
4649 umap
= isl_union_map_from_map(map
);
4650 after
= after_in_tree(umap
, node
);
4651 isl_union_map_free(umap
);
4652 isl_schedule_node_free(node
);
4656 /* Internal data for any_scheduled_after.
4658 * "build" is the build in which the AST is constructed.
4659 * "depth" is the number of loops that have already been generated
4660 * "group_coscheduled" is a local copy of options->ast_build_group_coscheduled
4661 * "domain" is an array of set-map pairs corresponding to the different
4662 * iteration domains. The set is the schedule domain, i.e., the domain
4663 * of the inverse schedule, while the map is the inverse schedule itself.
4665 struct isl_any_scheduled_after_data
{
4666 isl_ast_build
*build
;
4668 int group_coscheduled
;
4669 struct isl_set_map_pair
*domain
;
4672 /* Is any element of domain "i" scheduled after any element of domain "j"
4673 * (for a common iteration of the first data->depth loops)?
4675 * data->domain[i].set contains the domain of the inverse schedule
4676 * for domain "i", i.e., elements in the schedule domain.
4678 * If we are inside a band of a schedule tree and there is a pair
4679 * of elements in the two domains that is schedule together by
4680 * the current band, then we check if any element of "i" may be schedule
4681 * after element of "j" by the descendants of the band node.
4683 * If data->group_coscheduled is set, then we also return 1 if there
4684 * is any pair of elements in the two domains that are scheduled together.
4686 static isl_bool
any_scheduled_after(int i
, int j
, void *user
)
4688 struct isl_any_scheduled_after_data
*data
= user
;
4689 isl_size dim
= isl_set_dim(data
->domain
[i
].set
, isl_dim_set
);
4693 return isl_bool_error
;
4695 for (pos
= data
->depth
; pos
< dim
; ++pos
) {
4698 follows
= isl_set_follows_at(data
->domain
[i
].set
,
4699 data
->domain
[j
].set
, pos
);
4702 return isl_bool_error
;
4704 return isl_bool_true
;
4706 return isl_bool_false
;
4709 if (isl_ast_build_has_schedule_node(data
->build
)) {
4712 after
= after_in_subtree(data
->build
, data
->domain
[i
].map
,
4713 data
->domain
[j
].map
);
4714 if (after
< 0 || after
)
4718 return isl_bool_ok(data
->group_coscheduled
);
4721 /* Look for independent components at the current depth and generate code
4722 * for each component separately. The resulting lists of grafts are
4723 * merged in an attempt to combine grafts with identical guards.
4725 * Code for two domains can be generated separately if all the elements
4726 * of one domain are scheduled before (or together with) all the elements
4727 * of the other domain. We therefore consider the graph with as nodes
4728 * the domains and an edge between two nodes if any element of the first
4729 * node is scheduled after any element of the second node.
4730 * If the ast_build_group_coscheduled is set, then we also add an edge if
4731 * there is any pair of elements in the two domains that are scheduled
4733 * Code is then generated (by generate_component)
4734 * for each of the strongly connected components in this graph
4735 * in their topological order.
4737 * Since the test is performed on the domain of the inverse schedules of
4738 * the different domains, we precompute these domains and store
4739 * them in data.domain.
4741 static __isl_give isl_ast_graft_list
*generate_components(
4742 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
4745 isl_ctx
*ctx
= isl_ast_build_get_ctx(build
);
4746 isl_size n
= isl_union_map_n_map(executed
);
4748 struct isl_any_scheduled_after_data data
;
4749 struct isl_set_map_pair
*next
;
4750 struct isl_tarjan_graph
*g
= NULL
;
4751 isl_ast_graft_list
*list
= NULL
;
4757 data
.domain
= isl_calloc_array(ctx
, struct isl_set_map_pair
, n
);
4763 if (isl_union_map_foreach_map(executed
, &extract_domain
, &next
) < 0)
4766 depth
= isl_ast_build_get_depth(build
);
4771 data
.group_coscheduled
= isl_options_get_ast_build_group_coscheduled(ctx
);
4772 g
= isl_tarjan_graph_init(ctx
, n
, &any_scheduled_after
, &data
);
4776 list
= isl_ast_graft_list_alloc(ctx
, 0);
4780 isl_ast_graft_list
*list_c
;
4783 if (g
->order
[i
] == -1)
4784 isl_die(ctx
, isl_error_internal
, "cannot happen",
4787 while (g
->order
[i
] != -1) {
4791 list_c
= generate_component(data
.domain
,
4792 g
->order
+ first
, i
- first
,
4793 isl_ast_build_copy(build
));
4794 list
= isl_ast_graft_list_merge(list
, list_c
, build
);
4800 error
: list
= isl_ast_graft_list_free(list
);
4801 isl_tarjan_graph_free(g
);
4802 for (i
= 0; i
< n_domain
; ++i
) {
4803 isl_map_free(data
.domain
[i
].map
);
4804 isl_set_free(data
.domain
[i
].set
);
4807 isl_union_map_free(executed
);
4808 isl_ast_build_free(build
);
4813 /* Generate code for the next level (and all inner levels).
4815 * If "executed" is empty, i.e., no code needs to be generated,
4816 * then we return an empty list.
4818 * If we have already generated code for all loop levels, then we pass
4819 * control to generate_inner_level.
4821 * If "executed" lives in a single space, i.e., if code needs to be
4822 * generated for a single domain, then there can only be a single
4823 * component and we go directly to generate_shifted_component.
4824 * Otherwise, we call generate_components to detect the components
4825 * and to call generate_component on each of them separately.
4827 static __isl_give isl_ast_graft_list
*generate_next_level(
4828 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
)
4834 if (!build
|| !executed
)
4837 if (isl_union_map_is_empty(executed
)) {
4838 isl_ctx
*ctx
= isl_ast_build_get_ctx(build
);
4839 isl_union_map_free(executed
);
4840 isl_ast_build_free(build
);
4841 return isl_ast_graft_list_alloc(ctx
, 0);
4844 depth
= isl_ast_build_get_depth(build
);
4845 dim
= isl_ast_build_dim(build
, isl_dim_set
);
4846 if (depth
< 0 || dim
< 0)
4849 return generate_inner_level(executed
, build
);
4851 n
= isl_union_map_n_map(executed
);
4855 return generate_shifted_component(executed
, build
);
4857 return generate_components(executed
, build
);
4859 isl_union_map_free(executed
);
4860 isl_ast_build_free(build
);
4864 /* Internal data structure used by isl_ast_build_node_from_schedule_map.
4865 * internal, executed and build are the inputs to generate_code.
4866 * list collects the output.
4868 struct isl_generate_code_data
{
4870 isl_union_map
*executed
;
4871 isl_ast_build
*build
;
4873 isl_ast_graft_list
*list
;
4876 /* Given an inverse schedule in terms of the external build schedule, i.e.,
4880 * with E the external build schedule and S the additional schedule "space",
4881 * reformulate the inverse schedule in terms of the internal schedule domain,
4886 * We first obtain a mapping
4890 * take the inverse and the product with S -> S, resulting in
4892 * [I -> S] -> [E -> S]
4894 * Applying the map to the input produces the desired result.
4896 static __isl_give isl_union_map
*internal_executed(
4897 __isl_take isl_union_map
*executed
, __isl_keep isl_space
*space
,
4898 __isl_keep isl_ast_build
*build
)
4902 proj
= isl_ast_build_get_schedule_map(build
);
4903 proj
= isl_map_reverse(proj
);
4904 space
= isl_space_map_from_set(isl_space_copy(space
));
4905 id
= isl_map_identity(space
);
4906 proj
= isl_map_product(proj
, id
);
4907 executed
= isl_union_map_apply_domain(executed
,
4908 isl_union_map_from_map(proj
));
4912 /* Generate an AST that visits the elements in the range of data->executed
4913 * in the relative order specified by the corresponding domain element(s)
4914 * for those domain elements that belong to "set".
4915 * Add the result to data->list.
4917 * The caller ensures that "set" is a universe domain.
4918 * "space" is the space of the additional part of the schedule.
4919 * It is equal to the space of "set" if build->domain is parametric.
4920 * Otherwise, it is equal to the range of the wrapped space of "set".
4922 * If the build space is not parametric and
4923 * if isl_ast_build_node_from_schedule_map
4924 * was called from an outside user (data->internal not set), then
4925 * the (inverse) schedule refers to the external build domain and needs to
4926 * be transformed to refer to the internal build domain.
4928 * If the build space is parametric, then we add some of the parameter
4929 * constraints to the executed relation. Adding these constraints
4930 * allows for an earlier detection of conflicts in some cases.
4931 * However, we do not want to divide the executed relation into
4932 * more disjuncts than necessary. We therefore approximate
4933 * the constraints on the parameters by a single disjunct set.
4935 * The build is extended to include the additional part of the schedule.
4936 * If the original build space was not parametric, then the options
4937 * in data->build refer only to the additional part of the schedule
4938 * and they need to be adjusted to refer to the complete AST build
4941 * After having adjusted inverse schedule and build, we start generating
4942 * code with the outer loop of the current code generation
4943 * in generate_next_level.
4945 * If the original build space was not parametric, we undo the embedding
4946 * on the resulting isl_ast_node_list so that it can be used within
4947 * the outer AST build.
4949 static isl_stat
generate_code_in_space(struct isl_generate_code_data
*data
,
4950 __isl_take isl_set
*set
, __isl_take isl_space
*space
)
4952 isl_union_map
*executed
;
4953 isl_ast_build
*build
;
4954 isl_ast_graft_list
*list
;
4957 executed
= isl_union_map_copy(data
->executed
);
4958 executed
= isl_union_map_intersect_domain(executed
,
4959 isl_union_set_from_set(set
));
4961 embed
= !isl_set_is_params(data
->build
->domain
);
4962 if (embed
&& !data
->internal
)
4963 executed
= internal_executed(executed
, space
, data
->build
);
4966 domain
= isl_ast_build_get_domain(data
->build
);
4967 domain
= isl_set_from_basic_set(isl_set_simple_hull(domain
));
4968 executed
= isl_union_map_intersect_params(executed
, domain
);
4971 build
= isl_ast_build_copy(data
->build
);
4972 build
= isl_ast_build_product(build
, space
);
4974 list
= generate_next_level(executed
, build
);
4976 list
= isl_ast_graft_list_unembed(list
, embed
);
4978 data
->list
= isl_ast_graft_list_concat(data
->list
, list
);
4983 /* Generate an AST that visits the elements in the range of data->executed
4984 * in the relative order specified by the corresponding domain element(s)
4985 * for those domain elements that belong to "set".
4986 * Add the result to data->list.
4988 * The caller ensures that "set" is a universe domain.
4990 * If the build space S is not parametric, then the space of "set"
4991 * need to be a wrapped relation with S as domain. That is, it needs
4996 * Check this property and pass control to generate_code_in_space
4998 * If the build space is not parametric, then T is the space of "set".
5000 static isl_stat
generate_code_set(__isl_take isl_set
*set
, void *user
)
5002 struct isl_generate_code_data
*data
= user
;
5003 isl_space
*space
, *build_space
;
5006 space
= isl_set_get_space(set
);
5008 if (isl_set_is_params(data
->build
->domain
))
5009 return generate_code_in_space(data
, set
, space
);
5011 build_space
= isl_ast_build_get_space(data
->build
, data
->internal
);
5012 space
= isl_space_unwrap(space
);
5013 is_domain
= isl_space_is_domain(build_space
, space
);
5014 isl_space_free(build_space
);
5015 space
= isl_space_range(space
);
5020 isl_die(isl_set_get_ctx(set
), isl_error_invalid
,
5021 "invalid nested schedule space", goto error
);
5023 return generate_code_in_space(data
, set
, space
);
5026 isl_space_free(space
);
5027 return isl_stat_error
;
5030 /* Generate an AST that visits the elements in the range of "executed"
5031 * in the relative order specified by the corresponding domain element(s).
5033 * "build" is an isl_ast_build that has either been constructed by
5034 * isl_ast_build_from_context or passed to a callback set by
5035 * isl_ast_build_set_create_leaf.
5036 * In the first case, the space of the isl_ast_build is typically
5037 * a parametric space, although this is currently not enforced.
5038 * In the second case, the space is never a parametric space.
5039 * If the space S is not parametric, then the domain space(s) of "executed"
5040 * need to be wrapped relations with S as domain.
5042 * If the domain of "executed" consists of several spaces, then an AST
5043 * is generated for each of them (in arbitrary order) and the results
5046 * If "internal" is set, then the domain "S" above refers to the internal
5047 * schedule domain representation. Otherwise, it refers to the external
5048 * representation, as returned by isl_ast_build_get_schedule_space.
5050 * We essentially run over all the spaces in the domain of "executed"
5051 * and call generate_code_set on each of them.
5053 static __isl_give isl_ast_graft_list
*generate_code(
5054 __isl_take isl_union_map
*executed
, __isl_take isl_ast_build
*build
,
5058 struct isl_generate_code_data data
= { 0 };
5060 isl_union_set
*schedule_domain
;
5061 isl_union_map
*universe
;
5065 space
= isl_ast_build_get_space(build
, 1);
5066 space
= isl_space_align_params(space
,
5067 isl_union_map_get_space(executed
));
5068 space
= isl_space_align_params(space
,
5069 isl_union_map_get_space(build
->options
));
5070 build
= isl_ast_build_align_params(build
, isl_space_copy(space
));
5071 executed
= isl_union_map_align_params(executed
, space
);
5072 if (!executed
|| !build
)
5075 ctx
= isl_ast_build_get_ctx(build
);
5077 data
.internal
= internal
;
5078 data
.executed
= executed
;
5080 data
.list
= isl_ast_graft_list_alloc(ctx
, 0);
5082 universe
= isl_union_map_universe(isl_union_map_copy(executed
));
5083 schedule_domain
= isl_union_map_domain(universe
);
5084 if (isl_union_set_foreach_set(schedule_domain
, &generate_code_set
,
5086 data
.list
= isl_ast_graft_list_free(data
.list
);
5088 isl_union_set_free(schedule_domain
);
5089 isl_union_map_free(executed
);
5091 isl_ast_build_free(build
);
5094 isl_union_map_free(executed
);
5095 isl_ast_build_free(build
);
5099 /* Generate an AST that visits the elements in the domain of "schedule"
5100 * in the relative order specified by the corresponding image element(s).
5102 * "build" is an isl_ast_build that has either been constructed by
5103 * isl_ast_build_from_context or passed to a callback set by
5104 * isl_ast_build_set_create_leaf.
5105 * In the first case, the space of the isl_ast_build is typically
5106 * a parametric space, although this is currently not enforced.
5107 * In the second case, the space is never a parametric space.
5108 * If the space S is not parametric, then the range space(s) of "schedule"
5109 * need to be wrapped relations with S as domain.
5111 * If the range of "schedule" consists of several spaces, then an AST
5112 * is generated for each of them (in arbitrary order) and the results
5115 * We first initialize the local copies of the relevant options.
5116 * We do this here rather than when the isl_ast_build is created
5117 * because the options may have changed between the construction
5118 * of the isl_ast_build and the call to isl_generate_code.
5120 * The main computation is performed on an inverse schedule (with
5121 * the schedule domain in the domain and the elements to be executed
5122 * in the range) called "executed".
5124 __isl_give isl_ast_node
*isl_ast_build_node_from_schedule_map(
5125 __isl_keep isl_ast_build
*build
, __isl_take isl_union_map
*schedule
)
5127 isl_ast_graft_list
*list
;
5129 isl_union_map
*executed
;
5131 schedule
= isl_union_map_coalesce(schedule
);
5132 schedule
= isl_union_map_remove_redundancies(schedule
);
5133 executed
= isl_union_map_reverse(schedule
);
5134 list
= generate_code(executed
, isl_ast_build_copy(build
), 0);
5135 node
= isl_ast_node_from_graft_list(list
, build
);
5140 /* The old name for isl_ast_build_node_from_schedule_map.
5141 * It is being kept for backward compatibility, but
5142 * it will be removed in the future.
5144 __isl_give isl_ast_node
*isl_ast_build_ast_from_schedule(
5145 __isl_keep isl_ast_build
*build
, __isl_take isl_union_map
*schedule
)
5147 return isl_ast_build_node_from_schedule_map(build
, schedule
);
5150 /* Generate an AST that visits the elements in the domain of "executed"
5151 * in the relative order specified by the leaf node "node".
5153 * The relation "executed" maps the outer generated loop iterators
5154 * to the domain elements executed by those iterations.
5156 * Simply pass control to generate_inner_level.
5157 * Note that the current build does not refer to any band node, so
5158 * that generate_inner_level will not try to visit the child of
5161 * If multiple statement instances reach a leaf,
5162 * then they can be executed in any order.
5163 * Group the list of grafts based on shared guards
5164 * such that identical guards are only generated once
5165 * when the list is eventually passed on to isl_ast_graft_list_fuse.
5167 static __isl_give isl_ast_graft_list
*build_ast_from_leaf(
5168 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5169 __isl_take isl_union_map
*executed
)
5171 isl_ast_graft_list
*list
;
5173 isl_schedule_node_free(node
);
5174 list
= generate_inner_level(executed
, isl_ast_build_copy(build
));
5175 list
= isl_ast_graft_list_group_on_guard(list
, build
);
5176 isl_ast_build_free(build
);
5181 /* Check that the band partial schedule "partial" does not filter out
5182 * any statement instances, as specified by the range of "executed".
5184 static isl_stat
check_band_schedule_total_on_instances(
5185 __isl_keep isl_multi_union_pw_aff
*partial
,
5186 __isl_keep isl_union_map
*executed
)
5189 isl_union_set
*domain
, *instances
;
5191 instances
= isl_union_map_range(isl_union_map_copy(executed
));
5192 partial
= isl_multi_union_pw_aff_copy(partial
);
5193 domain
= isl_multi_union_pw_aff_domain(partial
);
5194 subset
= isl_union_set_is_subset(instances
, domain
);
5195 isl_union_set_free(domain
);
5196 isl_union_set_free(instances
);
5199 return isl_stat_error
;
5201 isl_die(isl_union_map_get_ctx(executed
), isl_error_invalid
,
5202 "band node is not allowed to drop statement instances",
5203 return isl_stat_error
);
5207 /* Generate an AST that visits the elements in the domain of "executed"
5208 * in the relative order specified by the band node "node" and its descendants.
5210 * The relation "executed" maps the outer generated loop iterators
5211 * to the domain elements executed by those iterations.
5213 * If the band is empty, we continue with its descendants.
5214 * Otherwise, we extend the build and the inverse schedule with
5215 * the additional space/partial schedule and continue generating
5216 * an AST in generate_next_level.
5217 * As soon as we have extended the inverse schedule with the additional
5218 * partial schedule, we look for equalities that may exists between
5219 * the old and the new part.
5221 static __isl_give isl_ast_graft_list
*build_ast_from_band(
5222 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5223 __isl_take isl_union_map
*executed
)
5226 isl_multi_union_pw_aff
*extra
;
5227 isl_union_map
*extra_umap
;
5228 isl_ast_graft_list
*list
;
5232 n
= isl_schedule_node_band_n_member(node
);
5233 if (!build
|| n
< 0 || !executed
)
5237 return build_ast_from_child(build
, node
, executed
);
5239 extra
= isl_schedule_node_band_get_partial_schedule(node
);
5240 extra
= isl_multi_union_pw_aff_align_params(extra
,
5241 isl_ast_build_get_space(build
, 1));
5242 space
= isl_multi_union_pw_aff_get_space(extra
);
5244 if (check_band_schedule_total_on_instances(extra
, executed
) < 0)
5245 executed
= isl_union_map_free(executed
);
5247 extra_umap
= isl_union_map_from_multi_union_pw_aff(extra
);
5248 extra_umap
= isl_union_map_reverse(extra_umap
);
5250 executed
= isl_union_map_domain_product(executed
, extra_umap
);
5251 executed
= isl_union_map_detect_equalities(executed
);
5253 n1
= isl_ast_build_dim(build
, isl_dim_param
);
5254 build
= isl_ast_build_product(build
, space
);
5255 n2
= isl_ast_build_dim(build
, isl_dim_param
);
5256 if (n1
< 0 || n2
< 0)
5257 build
= isl_ast_build_free(build
);
5259 isl_die(isl_ast_build_get_ctx(build
), isl_error_invalid
,
5260 "band node is not allowed to introduce new parameters",
5261 build
= isl_ast_build_free(build
));
5262 build
= isl_ast_build_set_schedule_node(build
, node
);
5264 list
= generate_next_level(executed
, build
);
5266 list
= isl_ast_graft_list_unembed(list
, 1);
5270 isl_schedule_node_free(node
);
5271 isl_union_map_free(executed
);
5272 isl_ast_build_free(build
);
5276 /* Hoist a list of grafts (in practice containing a single graft)
5277 * from "sub_build" (which includes extra context information)
5280 * In particular, project out all additional parameters introduced
5281 * by the context node from the enforced constraints and the guard
5282 * of the single graft.
5284 static __isl_give isl_ast_graft_list
*hoist_out_of_context(
5285 __isl_take isl_ast_graft_list
*list
, __isl_keep isl_ast_build
*build
,
5286 __isl_keep isl_ast_build
*sub_build
)
5288 isl_ast_graft
*graft
;
5289 isl_basic_set
*enforced
;
5291 isl_size n_param
, extra_param
;
5293 n_param
= isl_ast_build_dim(build
, isl_dim_param
);
5294 extra_param
= isl_ast_build_dim(sub_build
, isl_dim_param
);
5295 if (n_param
< 0 || extra_param
< 0)
5296 return isl_ast_graft_list_free(list
);
5298 if (extra_param
== n_param
)
5301 extra_param
-= n_param
;
5302 enforced
= isl_ast_graft_list_extract_shared_enforced(list
, sub_build
);
5303 enforced
= isl_basic_set_project_out(enforced
, isl_dim_param
,
5304 n_param
, extra_param
);
5305 enforced
= isl_basic_set_remove_unknown_divs(enforced
);
5306 guard
= isl_ast_graft_list_extract_hoistable_guard(list
, sub_build
);
5307 guard
= isl_set_remove_divs_involving_dims(guard
, isl_dim_param
,
5308 n_param
, extra_param
);
5309 guard
= isl_set_project_out(guard
, isl_dim_param
, n_param
, extra_param
);
5310 guard
= isl_set_compute_divs(guard
);
5311 graft
= isl_ast_graft_alloc_from_children(list
, guard
, enforced
,
5313 list
= isl_ast_graft_list_from_ast_graft(graft
);
5318 /* Generate an AST that visits the elements in the domain of "executed"
5319 * in the relative order specified by the context node "node"
5320 * and its descendants.
5322 * The relation "executed" maps the outer generated loop iterators
5323 * to the domain elements executed by those iterations.
5325 * The context node may introduce additional parameters as well as
5326 * constraints on the outer schedule dimensions or original parameters.
5328 * We add the extra parameters to a new build and the context
5329 * constraints to both the build and (as a single disjunct)
5330 * to the domain of "executed". Since the context constraints
5331 * are specified in terms of the input schedule, we first need
5332 * to map them to the internal schedule domain.
5334 * After constructing the AST from the descendants of "node",
5335 * we combine the list of grafts into a single graft within
5336 * the new build, in order to be able to exploit the additional
5337 * context constraints during this combination.
5339 * Additionally, if the current node is the outermost node in
5340 * the schedule tree (apart from the root domain node), we generate
5341 * all pending guards, again to be able to exploit the additional
5342 * context constraints. We currently do not do this for internal
5343 * context nodes since we may still want to hoist conditions
5344 * to outer AST nodes.
5346 * If the context node introduced any new parameters, then they
5347 * are removed from the set of enforced constraints and guard
5348 * in hoist_out_of_context.
5350 static __isl_give isl_ast_graft_list
*build_ast_from_context(
5351 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5352 __isl_take isl_union_map
*executed
)
5356 isl_multi_aff
*internal2input
;
5357 isl_ast_build
*sub_build
;
5358 isl_ast_graft_list
*list
;
5362 depth
= isl_schedule_node_get_tree_depth(node
);
5364 build
= isl_ast_build_free(build
);
5365 space
= isl_ast_build_get_space(build
, 1);
5366 context
= isl_schedule_node_context_get_context(node
);
5367 context
= isl_set_align_params(context
, space
);
5368 sub_build
= isl_ast_build_copy(build
);
5369 space
= isl_set_get_space(context
);
5370 sub_build
= isl_ast_build_align_params(sub_build
, space
);
5371 internal2input
= isl_ast_build_get_internal2input(sub_build
);
5372 context
= isl_set_preimage_multi_aff(context
, internal2input
);
5373 sub_build
= isl_ast_build_restrict_generated(sub_build
,
5374 isl_set_copy(context
));
5375 context
= isl_set_from_basic_set(isl_set_simple_hull(context
));
5376 executed
= isl_union_map_intersect_domain(executed
,
5377 isl_union_set_from_set(context
));
5379 list
= build_ast_from_child(isl_ast_build_copy(sub_build
),
5381 n
= isl_ast_graft_list_n_ast_graft(list
);
5383 list
= isl_ast_graft_list_free(list
);
5385 list
= isl_ast_graft_list_fuse(list
, sub_build
);
5387 list
= isl_ast_graft_list_insert_pending_guard_nodes(list
,
5390 list
= hoist_out_of_context(list
, build
, sub_build
);
5392 isl_ast_build_free(build
);
5393 isl_ast_build_free(sub_build
);
5398 /* Generate an AST that visits the elements in the domain of "executed"
5399 * in the relative order specified by the expansion node "node" and
5402 * The relation "executed" maps the outer generated loop iterators
5403 * to the domain elements executed by those iterations.
5405 * We expand the domain elements by the expansion and
5406 * continue with the descendants of the node.
5408 static __isl_give isl_ast_graft_list
*build_ast_from_expansion(
5409 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5410 __isl_take isl_union_map
*executed
)
5412 isl_union_map
*expansion
;
5415 expansion
= isl_schedule_node_expansion_get_expansion(node
);
5416 expansion
= isl_union_map_align_params(expansion
,
5417 isl_union_map_get_space(executed
));
5419 n1
= isl_union_map_dim(executed
, isl_dim_param
);
5420 executed
= isl_union_map_apply_range(executed
, expansion
);
5421 n2
= isl_union_map_dim(executed
, isl_dim_param
);
5422 if (n1
< 0 || n2
< 0)
5425 isl_die(isl_ast_build_get_ctx(build
), isl_error_invalid
,
5426 "expansion node is not allowed to introduce "
5427 "new parameters", goto error
);
5429 return build_ast_from_child(build
, node
, executed
);
5431 isl_ast_build_free(build
);
5432 isl_schedule_node_free(node
);
5433 isl_union_map_free(executed
);
5437 /* Generate an AST that visits the elements in the domain of "executed"
5438 * in the relative order specified by the extension node "node" and
5441 * The relation "executed" maps the outer generated loop iterators
5442 * to the domain elements executed by those iterations.
5444 * Extend the inverse schedule with the extension applied to current
5445 * set of generated constraints. Since the extension if formulated
5446 * in terms of the input schedule, it first needs to be transformed
5447 * to refer to the internal schedule.
5449 static __isl_give isl_ast_graft_list
*build_ast_from_extension(
5450 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5451 __isl_take isl_union_map
*executed
)
5453 isl_union_set
*schedule_domain
;
5454 isl_union_map
*extension
;
5457 set
= isl_ast_build_get_generated(build
);
5458 set
= isl_set_from_basic_set(isl_set_simple_hull(set
));
5459 schedule_domain
= isl_union_set_from_set(set
);
5461 extension
= isl_schedule_node_extension_get_extension(node
);
5463 extension
= isl_union_map_preimage_domain_multi_aff(extension
,
5464 isl_multi_aff_copy(build
->internal2input
));
5465 extension
= isl_union_map_intersect_domain(extension
, schedule_domain
);
5466 extension
= isl_ast_build_substitute_values_union_map_domain(build
,
5468 executed
= isl_union_map_union(executed
, extension
);
5470 return build_ast_from_child(build
, node
, executed
);
5473 /* Generate an AST that visits the elements in the domain of "executed"
5474 * in the relative order specified by the filter node "node" and
5477 * The relation "executed" maps the outer generated loop iterators
5478 * to the domain elements executed by those iterations.
5480 * We simply intersect the iteration domain (i.e., the range of "executed")
5481 * with the filter and continue with the descendants of the node,
5482 * unless the resulting inverse schedule is empty, in which
5483 * case we return an empty list.
5485 * If the result of the intersection is equal to the original "executed"
5486 * relation, then keep the original representation since the intersection
5487 * may have unnecessarily broken up the relation into a greater number
5490 static __isl_give isl_ast_graft_list
*build_ast_from_filter(
5491 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5492 __isl_take isl_union_map
*executed
)
5495 isl_union_set
*filter
;
5496 isl_union_map
*orig
;
5497 isl_ast_graft_list
*list
;
5502 orig
= isl_union_map_copy(executed
);
5503 if (!build
|| !node
|| !executed
)
5506 filter
= isl_schedule_node_filter_get_filter(node
);
5507 filter
= isl_union_set_align_params(filter
,
5508 isl_union_map_get_space(executed
));
5509 n1
= isl_union_map_dim(executed
, isl_dim_param
);
5510 executed
= isl_union_map_intersect_range(executed
, filter
);
5511 n2
= isl_union_map_dim(executed
, isl_dim_param
);
5512 if (n1
< 0 || n2
< 0)
5515 isl_die(isl_ast_build_get_ctx(build
), isl_error_invalid
,
5516 "filter node is not allowed to introduce "
5517 "new parameters", goto error
);
5519 unchanged
= isl_union_map_is_subset(orig
, executed
);
5520 empty
= isl_union_map_is_empty(executed
);
5521 if (unchanged
< 0 || empty
< 0)
5524 isl_union_map_free(executed
);
5525 return build_ast_from_child(build
, node
, orig
);
5527 isl_union_map_free(orig
);
5529 return build_ast_from_child(build
, node
, executed
);
5531 ctx
= isl_ast_build_get_ctx(build
);
5532 list
= isl_ast_graft_list_alloc(ctx
, 0);
5533 isl_ast_build_free(build
);
5534 isl_schedule_node_free(node
);
5535 isl_union_map_free(executed
);
5538 isl_ast_build_free(build
);
5539 isl_schedule_node_free(node
);
5540 isl_union_map_free(executed
);
5541 isl_union_map_free(orig
);
5545 /* Generate an AST that visits the elements in the domain of "executed"
5546 * in the relative order specified by the guard node "node" and
5549 * The relation "executed" maps the outer generated loop iterators
5550 * to the domain elements executed by those iterations.
5552 * Ensure that the associated guard is enforced by the outer AST
5553 * constructs by adding it to the guard of the graft.
5554 * Since we know that we will enforce the guard, we can also include it
5555 * in the generated constraints used to construct an AST for
5556 * the descendant nodes.
5558 static __isl_give isl_ast_graft_list
*build_ast_from_guard(
5559 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5560 __isl_take isl_union_map
*executed
)
5563 isl_set
*guard
, *hoisted
;
5564 isl_basic_set
*enforced
;
5565 isl_ast_build
*sub_build
;
5566 isl_ast_graft
*graft
;
5567 isl_ast_graft_list
*list
;
5570 space
= isl_ast_build_get_space(build
, 1);
5571 guard
= isl_schedule_node_guard_get_guard(node
);
5572 n1
= isl_space_dim(space
, isl_dim_param
);
5573 guard
= isl_set_align_params(guard
, space
);
5574 n2
= isl_set_dim(guard
, isl_dim_param
);
5575 if (n1
< 0 || n2
< 0)
5576 guard
= isl_set_free(guard
);
5578 isl_die(isl_ast_build_get_ctx(build
), isl_error_invalid
,
5579 "guard node is not allowed to introduce "
5580 "new parameters", guard
= isl_set_free(guard
));
5581 guard
= isl_set_preimage_multi_aff(guard
,
5582 isl_multi_aff_copy(build
->internal2input
));
5583 guard
= isl_ast_build_specialize(build
, guard
);
5584 guard
= isl_set_gist(guard
, isl_set_copy(build
->generated
));
5586 sub_build
= isl_ast_build_copy(build
);
5587 sub_build
= isl_ast_build_restrict_generated(sub_build
,
5588 isl_set_copy(guard
));
5590 list
= build_ast_from_child(isl_ast_build_copy(sub_build
),
5593 hoisted
= isl_ast_graft_list_extract_hoistable_guard(list
, sub_build
);
5594 n
= isl_set_n_basic_set(hoisted
);
5596 list
= isl_ast_graft_list_free(list
);
5598 list
= isl_ast_graft_list_gist_guards(list
,
5599 isl_set_copy(hoisted
));
5600 guard
= isl_set_intersect(guard
, hoisted
);
5601 enforced
= extract_shared_enforced(list
, build
);
5602 graft
= isl_ast_graft_alloc_from_children(list
, guard
, enforced
,
5605 isl_ast_build_free(sub_build
);
5606 isl_ast_build_free(build
);
5607 return isl_ast_graft_list_from_ast_graft(graft
);
5610 /* Call the before_each_mark callback, if requested by the user.
5612 * Return 0 on success and -1 on error.
5614 * The caller is responsible for recording the current inverse schedule
5617 static isl_stat
before_each_mark(__isl_keep isl_id
*mark
,
5618 __isl_keep isl_ast_build
*build
)
5621 return isl_stat_error
;
5622 if (!build
->before_each_mark
)
5624 return build
->before_each_mark(mark
, build
,
5625 build
->before_each_mark_user
);
5628 /* Call the after_each_mark callback, if requested by the user.
5630 * The caller is responsible for recording the current inverse schedule
5633 static __isl_give isl_ast_graft
*after_each_mark(
5634 __isl_take isl_ast_graft
*graft
, __isl_keep isl_ast_build
*build
)
5636 if (!graft
|| !build
)
5637 return isl_ast_graft_free(graft
);
5638 if (!build
->after_each_mark
)
5640 graft
->node
= build
->after_each_mark(graft
->node
, build
,
5641 build
->after_each_mark_user
);
5643 return isl_ast_graft_free(graft
);
5648 /* Generate an AST that visits the elements in the domain of "executed"
5649 * in the relative order specified by the mark node "node" and
5652 * The relation "executed" maps the outer generated loop iterators
5653 * to the domain elements executed by those iterations.
5655 * Since we may be calling before_each_mark and after_each_mark
5656 * callbacks, we record the current inverse schedule in the build.
5658 * We generate an AST for the child of the mark node, combine
5659 * the graft list into a single graft and then insert the mark
5660 * in the AST of that single graft.
5662 static __isl_give isl_ast_graft_list
*build_ast_from_mark(
5663 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5664 __isl_take isl_union_map
*executed
)
5667 isl_ast_graft
*graft
;
5668 isl_ast_graft_list
*list
;
5671 build
= isl_ast_build_set_executed(build
, isl_union_map_copy(executed
));
5673 mark
= isl_schedule_node_mark_get_id(node
);
5674 if (before_each_mark(mark
, build
) < 0)
5675 node
= isl_schedule_node_free(node
);
5677 list
= build_ast_from_child(isl_ast_build_copy(build
), node
, executed
);
5678 list
= isl_ast_graft_list_fuse(list
, build
);
5679 n
= isl_ast_graft_list_n_ast_graft(list
);
5681 list
= isl_ast_graft_list_free(list
);
5685 graft
= isl_ast_graft_list_get_ast_graft(list
, 0);
5686 graft
= isl_ast_graft_insert_mark(graft
, mark
);
5687 graft
= after_each_mark(graft
, build
);
5688 list
= isl_ast_graft_list_set_ast_graft(list
, 0, graft
);
5690 isl_ast_build_free(build
);
5695 static __isl_give isl_ast_graft_list
*build_ast_from_schedule_node(
5696 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5697 __isl_take isl_union_map
*executed
);
5699 /* Generate an AST that visits the elements in the domain of "executed"
5700 * in the relative order specified by the sequence (or set) node "node" and
5703 * The relation "executed" maps the outer generated loop iterators
5704 * to the domain elements executed by those iterations.
5706 * We simply generate an AST for each of the children and concatenate
5709 static __isl_give isl_ast_graft_list
*build_ast_from_sequence(
5710 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5711 __isl_take isl_union_map
*executed
)
5716 isl_ast_graft_list
*list
;
5718 ctx
= isl_ast_build_get_ctx(build
);
5719 list
= isl_ast_graft_list_alloc(ctx
, 0);
5721 n
= isl_schedule_node_n_children(node
);
5723 list
= isl_ast_graft_list_free(list
);
5724 for (i
= 0; i
< n
; ++i
) {
5725 isl_schedule_node
*child
;
5726 isl_ast_graft_list
*list_i
;
5728 child
= isl_schedule_node_get_child(node
, i
);
5729 list_i
= build_ast_from_schedule_node(isl_ast_build_copy(build
),
5730 child
, isl_union_map_copy(executed
));
5731 list
= isl_ast_graft_list_concat(list
, list_i
);
5733 isl_ast_build_free(build
);
5734 isl_schedule_node_free(node
);
5735 isl_union_map_free(executed
);
5740 /* Generate an AST that visits the elements in the domain of "executed"
5741 * in the relative order specified by the node "node" and its descendants.
5743 * The relation "executed" maps the outer generated loop iterators
5744 * to the domain elements executed by those iterations.
5746 * The node types are handled in separate functions.
5747 * Set nodes are currently treated in the same way as sequence nodes.
5748 * The children of a set node may be executed in any order,
5749 * including the order of the children.
5751 static __isl_give isl_ast_graft_list
*build_ast_from_schedule_node(
5752 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5753 __isl_take isl_union_map
*executed
)
5755 enum isl_schedule_node_type type
;
5757 type
= isl_schedule_node_get_type(node
);
5760 case isl_schedule_node_error
:
5762 case isl_schedule_node_leaf
:
5763 return build_ast_from_leaf(build
, node
, executed
);
5764 case isl_schedule_node_band
:
5765 return build_ast_from_band(build
, node
, executed
);
5766 case isl_schedule_node_context
:
5767 return build_ast_from_context(build
, node
, executed
);
5768 case isl_schedule_node_domain
:
5769 isl_die(isl_schedule_node_get_ctx(node
), isl_error_unsupported
,
5770 "unexpected internal domain node", goto error
);
5771 case isl_schedule_node_expansion
:
5772 return build_ast_from_expansion(build
, node
, executed
);
5773 case isl_schedule_node_extension
:
5774 return build_ast_from_extension(build
, node
, executed
);
5775 case isl_schedule_node_filter
:
5776 return build_ast_from_filter(build
, node
, executed
);
5777 case isl_schedule_node_guard
:
5778 return build_ast_from_guard(build
, node
, executed
);
5779 case isl_schedule_node_mark
:
5780 return build_ast_from_mark(build
, node
, executed
);
5781 case isl_schedule_node_sequence
:
5782 case isl_schedule_node_set
:
5783 return build_ast_from_sequence(build
, node
, executed
);
5786 isl_die(isl_ast_build_get_ctx(build
), isl_error_internal
,
5787 "unhandled type", goto error
);
5789 isl_union_map_free(executed
);
5790 isl_schedule_node_free(node
);
5791 isl_ast_build_free(build
);
5796 /* Generate an AST that visits the elements in the domain of "executed"
5797 * in the relative order specified by the (single) child of "node" and
5800 * The relation "executed" maps the outer generated loop iterators
5801 * to the domain elements executed by those iterations.
5803 * This function is never called on a leaf, set or sequence node,
5804 * so the node always has exactly one child.
5806 static __isl_give isl_ast_graft_list
*build_ast_from_child(
5807 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
,
5808 __isl_take isl_union_map
*executed
)
5810 node
= isl_schedule_node_child(node
, 0);
5811 return build_ast_from_schedule_node(build
, node
, executed
);
5814 /* Generate an AST that visits the elements in the domain of the domain
5815 * node "node" in the relative order specified by its descendants.
5817 * An initial inverse schedule is created that maps a zero-dimensional
5818 * schedule space to the node domain.
5819 * The input "build" is assumed to have a parametric domain and
5820 * is replaced by the same zero-dimensional schedule space.
5822 * We also add some of the parameter constraints in the build domain
5823 * to the executed relation. Adding these constraints
5824 * allows for an earlier detection of conflicts in some cases.
5825 * However, we do not want to divide the executed relation into
5826 * more disjuncts than necessary. We therefore approximate
5827 * the constraints on the parameters by a single disjunct set.
5829 static __isl_give isl_ast_node
*build_ast_from_domain(
5830 __isl_take isl_ast_build
*build
, __isl_take isl_schedule_node
*node
)
5833 isl_union_set
*domain
, *schedule_domain
;
5834 isl_union_map
*executed
;
5837 isl_ast_graft_list
*list
;
5844 ctx
= isl_ast_build_get_ctx(build
);
5845 space
= isl_ast_build_get_space(build
, 1);
5846 is_params
= isl_space_is_params(space
);
5847 isl_space_free(space
);
5851 isl_die(ctx
, isl_error_unsupported
,
5852 "expecting parametric initial context", goto error
);
5854 domain
= isl_schedule_node_domain_get_domain(node
);
5855 domain
= isl_union_set_coalesce(domain
);
5857 space
= isl_union_set_get_space(domain
);
5858 space
= isl_space_set_from_params(space
);
5859 build
= isl_ast_build_product(build
, space
);
5861 set
= isl_ast_build_get_domain(build
);
5862 set
= isl_set_from_basic_set(isl_set_simple_hull(set
));
5863 schedule_domain
= isl_union_set_from_set(set
);
5865 executed
= isl_union_map_from_domain_and_range(schedule_domain
, domain
);
5866 list
= build_ast_from_child(isl_ast_build_copy(build
), node
, executed
);
5867 ast
= isl_ast_node_from_graft_list(list
, build
);
5868 isl_ast_build_free(build
);
5872 isl_schedule_node_free(node
);
5873 isl_ast_build_free(build
);
5877 /* Generate an AST that visits the elements in the domain of "schedule"
5878 * in the relative order specified by the schedule tree.
5880 * "build" is an isl_ast_build that has been created using
5881 * isl_ast_build_alloc or isl_ast_build_from_context based
5882 * on a parametric set.
5884 * The construction starts at the root node of the schedule,
5885 * which is assumed to be a domain node.
5887 __isl_give isl_ast_node
*isl_ast_build_node_from_schedule(
5888 __isl_keep isl_ast_build
*build
, __isl_take isl_schedule
*schedule
)
5891 isl_schedule_node
*node
;
5893 if (!build
|| !schedule
)
5896 ctx
= isl_ast_build_get_ctx(build
);
5898 node
= isl_schedule_get_root(schedule
);
5901 isl_schedule_free(schedule
);
5903 build
= isl_ast_build_copy(build
);
5904 if (isl_schedule_node_get_type(node
) != isl_schedule_node_domain
)
5905 isl_die(ctx
, isl_error_unsupported
,
5906 "expecting root domain node",
5907 build
= isl_ast_build_free(build
));
5908 return build_ast_from_domain(build
, node
);
5910 isl_schedule_free(schedule
);