• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2013      Ecole Normale Superieure
3  *
4  * Use of this software is governed by the MIT license
5  *
6  * Written by Sven Verdoolaege,
7  * Ecole Normale Superieure, 45 rue d'Ulm, 75230 Paris, France
8  */
9 
10 #include <string.h>
11 
12 #include <isl/set.h>
13 #include <isl/union_set.h>
14 #include <isl/space.h>
15 
16 #include "gpu_tree.h"
17 
18 /* The functions in this file are used to navigate part of a schedule tree
19  * that is mapped to blocks.  Initially, this part consists of a linear
20  * branch segment with a mark node with name "kernel" on the outer end
21  * and a mark node with name "thread" on the inner end.
22  * During the mapping to blocks, branching may be introduced, but only
23  * one of the elements in each sequence contains the "thread" mark.
24  * The filter of this element (and only this filter) contains
25  * domain elements identified by the "core" argument of the functions
26  * that move down this tree.
27  *
28  * Synchronization statements have a name that starts with "sync" and
29  * a user pointer pointing to the kernel that contains the synchronization.
30  * The functions inserting or detecting synchronizations take a ppcg_kernel
31  * argument to be able to create or identify such statements.
32  * They may also use two fields in this structure, the "core" field
33  * to move around in the tree and the "n_sync" field to make sure that
34  * each synchronization has a different name (within the kernel).
35  */
36 
37 /* Is "node" a mark node with an identifier called "name"?
38  */
is_marked(__isl_keep isl_schedule_node * node,const char * name)39 static int is_marked(__isl_keep isl_schedule_node *node, const char *name)
40 {
41 	isl_id *mark;
42 	int has_name;
43 
44 	if (!node)
45 		return -1;
46 
47 	if (isl_schedule_node_get_type(node) != isl_schedule_node_mark)
48 		return 0;
49 
50 	mark = isl_schedule_node_mark_get_id(node);
51 	if (!mark)
52 		return -1;
53 
54 	has_name = !strcmp(isl_id_get_name(mark), name);
55 	isl_id_free(mark);
56 
57 	return has_name;
58 }
59 
60 /* Is "node" a mark node with an identifier called "kernel"?
61  */
gpu_tree_node_is_kernel(__isl_keep isl_schedule_node * node)62 int gpu_tree_node_is_kernel(__isl_keep isl_schedule_node *node)
63 {
64 	return is_marked(node, "kernel");
65 }
66 
67 /* Is "node" a mark node with an identifier called "shared"?
68  */
node_is_shared(__isl_keep isl_schedule_node * node)69 static int node_is_shared(__isl_keep isl_schedule_node *node)
70 {
71 	return is_marked(node, "shared");
72 }
73 
74 /* Is "node" a mark node with an identifier called "thread"?
75  */
node_is_thread(__isl_keep isl_schedule_node * node)76 static int node_is_thread(__isl_keep isl_schedule_node *node)
77 {
78 	return is_marked(node, "thread");
79 }
80 
81 /* Insert a mark node with identifier "shared" in front of "node".
82  */
insert_shared(__isl_take isl_schedule_node * node)83 static __isl_give isl_schedule_node *insert_shared(
84 	__isl_take isl_schedule_node *node)
85 {
86 	isl_ctx *ctx;
87 	isl_id *id;
88 
89 	ctx = isl_schedule_node_get_ctx(node);
90 	id = isl_id_alloc(ctx, "shared", NULL);
91 	node = isl_schedule_node_insert_mark(node, id);
92 
93 	return node;
94 }
95 
96 /* Insert a "shared" mark in front of the "thread" mark
97  * provided the linear branch between "node" and the "thread" mark
98  * does not contain such a "shared" mark already.
99  *
100  * As a side effect, this function checks that the subtree at "node"
101  * actually contains a "thread" mark and that there is no branching
102  * in between "node" and this "thread" mark.
103  */
gpu_tree_insert_shared_before_thread(__isl_take isl_schedule_node * node)104 __isl_give isl_schedule_node *gpu_tree_insert_shared_before_thread(
105 	__isl_take isl_schedule_node *node)
106 {
107 	int depth0, depth;
108 	int any_shared = 0;
109 
110 	if (!node)
111 		return NULL;
112 
113 	depth0 = isl_schedule_node_get_tree_depth(node);
114 
115 	for (;;) {
116 		int is_thread;
117 		int n;
118 
119 		if (!any_shared) {
120 			any_shared = node_is_shared(node);
121 			if (any_shared < 0)
122 				return isl_schedule_node_free(node);
123 		}
124 		is_thread = node_is_thread(node);
125 		if (is_thread < 0)
126 			return isl_schedule_node_free(node);
127 		if (is_thread)
128 			break;
129 		n = isl_schedule_node_n_children(node);
130 		if (n == 0)
131 			isl_die(isl_schedule_node_get_ctx(node),
132 				isl_error_invalid,
133 				"no thread marker found",
134 				return isl_schedule_node_free(node));
135 		if (n > 1)
136 			isl_die(isl_schedule_node_get_ctx(node),
137 				isl_error_invalid,
138 				"expecting single thread marker",
139 				return isl_schedule_node_free(node));
140 
141 		node = isl_schedule_node_child(node, 0);
142 	}
143 
144 	if (!any_shared)
145 		node = insert_shared(node);
146 	depth = isl_schedule_node_get_tree_depth(node);
147 	node = isl_schedule_node_ancestor(node, depth - depth0);
148 
149 	return node;
150 }
151 
152 /* Assuming "node" is a filter node, does it correspond to the branch
153  * that contains the "thread" mark, i.e., does it contain any elements
154  * in "core"?
155  */
node_is_core(__isl_keep isl_schedule_node * node,__isl_keep isl_union_set * core)156 static int node_is_core(__isl_keep isl_schedule_node *node,
157 	__isl_keep isl_union_set *core)
158 {
159 	int disjoint;
160 	isl_union_set *filter;
161 
162 	filter = isl_schedule_node_filter_get_filter(node);
163 	disjoint = isl_union_set_is_disjoint(filter, core);
164 	isl_union_set_free(filter);
165 	if (disjoint < 0)
166 		return -1;
167 
168 	return !disjoint;
169 }
170 
171 /* Move to the only child of "node" that has the "thread" mark as descendant,
172  * where the branch containing this mark is identified by the domain elements
173  * in "core".
174  *
175  * If "node" is not a sequence, then it only has one child and we move
176  * to that single child.
177  * Otherwise, we check each of the filters in the children, pick
178  * the one that corresponds to "core" and return a pointer to the child
179  * of the filter node.
180  */
core_child(__isl_take isl_schedule_node * node,__isl_keep isl_union_set * core)181 static __isl_give isl_schedule_node *core_child(
182 	__isl_take isl_schedule_node *node, __isl_keep isl_union_set *core)
183 {
184 	int i, n;
185 
186 	if (isl_schedule_node_get_type(node) != isl_schedule_node_sequence)
187 		return isl_schedule_node_child(node, 0);
188 
189 	n = isl_schedule_node_n_children(node);
190 	for (i = 0; i < n; ++i) {
191 		int is_core;
192 
193 		node = isl_schedule_node_child(node, i);
194 		is_core = node_is_core(node, core);
195 
196 		if (is_core < 0)
197 			return isl_schedule_node_free(node);
198 		if (is_core)
199 			return isl_schedule_node_child(node, 0);
200 
201 		node = isl_schedule_node_parent(node);
202 	}
203 
204 	isl_die(isl_schedule_node_get_ctx(node), isl_error_internal,
205 		"core child not found", return isl_schedule_node_free(node));
206 }
207 
208 /* Move down the branch between "kernel" and "thread" until
209  * the "shared" mark is reached, where the branch containing the "shared"
210  * mark is identified by the domain elements in "core".
211  */
gpu_tree_move_down_to_shared(__isl_take isl_schedule_node * node,__isl_keep isl_union_set * core)212 __isl_give isl_schedule_node *gpu_tree_move_down_to_shared(
213 	__isl_take isl_schedule_node *node, __isl_keep isl_union_set *core)
214 {
215 	int is_shared;
216 
217 	while ((is_shared = node_is_shared(node)) == 0)
218 		node = core_child(node, core);
219 	if (is_shared < 0)
220 		node = isl_schedule_node_free(node);
221 
222 	return node;
223 }
224 
225 /* Move down the branch between "kernel" and "thread" until
226  * the "thread" mark is reached, where the branch containing the "thread"
227  * mark is identified by the domain elements in "core".
228  */
gpu_tree_move_down_to_thread(__isl_take isl_schedule_node * node,__isl_keep isl_union_set * core)229 __isl_give isl_schedule_node *gpu_tree_move_down_to_thread(
230 	__isl_take isl_schedule_node *node, __isl_keep isl_union_set *core)
231 {
232 	int is_thread;
233 
234 	while ((is_thread = node_is_thread(node)) == 0)
235 		node = core_child(node, core);
236 	if (is_thread < 0)
237 		node = isl_schedule_node_free(node);
238 
239 	return node;
240 }
241 
242 /* Move up the tree underneath the "thread" mark until
243  * the "thread" mark is reached.
244  */
gpu_tree_move_up_to_thread(__isl_take isl_schedule_node * node)245 __isl_give isl_schedule_node *gpu_tree_move_up_to_thread(
246 	__isl_take isl_schedule_node *node)
247 {
248 	int is_thread;
249 
250 	while ((is_thread = node_is_thread(node)) == 0)
251 		node = isl_schedule_node_parent(node);
252 	if (is_thread < 0)
253 		node = isl_schedule_node_free(node);
254 
255 	return node;
256 }
257 
258 /* Move up the tree underneath the "kernel" mark until
259  * the "kernel" mark is reached.
260  */
gpu_tree_move_up_to_kernel(__isl_take isl_schedule_node * node)261 __isl_give isl_schedule_node *gpu_tree_move_up_to_kernel(
262 	__isl_take isl_schedule_node *node)
263 {
264 	int is_kernel;
265 
266 	while ((is_kernel = gpu_tree_node_is_kernel(node)) == 0)
267 		node = isl_schedule_node_parent(node);
268 	if (is_kernel < 0)
269 		node = isl_schedule_node_free(node);
270 
271 	return node;
272 }
273 
274 /* Move down from the "kernel" mark (or at least a node with schedule
275  * depth smaller than or equal to "depth") to a band node at schedule
276  * depth "depth".  The "thread" mark is assumed to have a schedule
277  * depth greater than or equal to "depth".  The branch containing the
278  * "thread" mark is identified by the domain elements in "core".
279  *
280  * If the desired schedule depth is in the middle of band node,
281  * then the band node is split into two pieces, the second piece
282  * at the desired schedule depth.
283  */
gpu_tree_move_down_to_depth(__isl_take isl_schedule_node * node,int depth,__isl_keep isl_union_set * core)284 __isl_give isl_schedule_node *gpu_tree_move_down_to_depth(
285 	__isl_take isl_schedule_node *node, int depth,
286 	__isl_keep isl_union_set *core)
287 {
288 	int is_shared;
289 	int is_thread = 0;
290 
291 	while (node && isl_schedule_node_get_schedule_depth(node) < depth) {
292 		if (isl_schedule_node_get_type(node) ==
293 						    isl_schedule_node_band) {
294 			int node_depth, node_dim;
295 			node_depth = isl_schedule_node_get_schedule_depth(node);
296 			node_dim = isl_schedule_node_band_n_member(node);
297 			if (node_depth + node_dim > depth)
298 				node = isl_schedule_node_band_split(node,
299 							depth - node_depth);
300 		}
301 		node = core_child(node, core);
302 	}
303 	while ((is_shared = node_is_shared(node)) == 0 &&
304 	    (is_thread = node_is_thread(node)) == 0 &&
305 	    isl_schedule_node_get_type(node) != isl_schedule_node_band)
306 		node = core_child(node, core);
307 	if (is_shared < 0 || is_thread < 0)
308 		node = isl_schedule_node_free(node);
309 
310 	return node;
311 }
312 
313 /* Create a union set containing a single set with a tuple identifier
314  * called "syncX" and user pointer equal to "kernel".
315  */
create_sync_domain(struct ppcg_kernel * kernel)316 static __isl_give isl_union_set *create_sync_domain(struct ppcg_kernel *kernel)
317 {
318 	isl_space *space;
319 	isl_id *id;
320 	char name[40];
321 
322 	space = isl_space_set_alloc(kernel->ctx, 0, 0);
323 	snprintf(name, sizeof(name), "sync%d", kernel->n_sync++);
324 	id = isl_id_alloc(kernel->ctx, name, kernel);
325 	space = isl_space_set_tuple_id(space, isl_dim_set, id);
326 	return isl_union_set_from_set(isl_set_universe(space));
327 }
328 
329 /* Is "id" the identifier of a synchronization statement inside "kernel"?
330  * That is, does its name start with "sync" and does it point to "kernel"?
331  */
gpu_tree_id_is_sync(__isl_keep isl_id * id,struct ppcg_kernel * kernel)332 int gpu_tree_id_is_sync(__isl_keep isl_id *id, struct ppcg_kernel *kernel)
333 {
334 	const char *name;
335 
336 	name = isl_id_get_name(id);
337 	if (!name)
338 		return 0;
339 	else if (strncmp(name, "sync", 4))
340 		return 0;
341 	return isl_id_get_user(id) == kernel;
342 }
343 
344 /* Does "domain" consist of a single set with a tuple identifier
345  * corresponding to a synchronization for "kernel"?
346  */
domain_is_sync(__isl_keep isl_union_set * domain,struct ppcg_kernel * kernel)347 static int domain_is_sync(__isl_keep isl_union_set *domain,
348 	struct ppcg_kernel *kernel)
349 {
350 	int is_sync;
351 	isl_id *id;
352 	isl_set *set;
353 
354 	if (isl_union_set_n_set(domain) != 1)
355 		return 0;
356 	set = isl_set_from_union_set(isl_union_set_copy(domain));
357 	id = isl_set_get_tuple_id(set);
358 	is_sync = gpu_tree_id_is_sync(id, kernel);
359 	isl_id_free(id);
360 	isl_set_free(set);
361 
362 	return is_sync;
363 }
364 
365 /* Does "node" point to a filter selecting a synchronization statement
366  * for "kernel"?
367  */
node_is_sync_filter(__isl_keep isl_schedule_node * node,struct ppcg_kernel * kernel)368 static int node_is_sync_filter(__isl_keep isl_schedule_node *node,
369 	struct ppcg_kernel *kernel)
370 {
371 	int is_sync;
372 	enum isl_schedule_node_type type;
373 	isl_union_set *domain;
374 
375 	if (!node)
376 		return -1;
377 	type = isl_schedule_node_get_type(node);
378 	if (type != isl_schedule_node_filter)
379 		return 0;
380 	domain = isl_schedule_node_filter_get_filter(node);
381 	is_sync = domain_is_sync(domain, kernel);
382 	isl_union_set_free(domain);
383 
384 	return is_sync;
385 }
386 
387 /* Is "node" part of a sequence with a previous synchronization statement
388  * for "kernel"?
389  * That is, is the parent of "node" a filter such that there is
390  * a previous filter that picks out exactly such a synchronization statement?
391  */
has_preceding_sync(__isl_keep isl_schedule_node * node,struct ppcg_kernel * kernel)392 static int has_preceding_sync(__isl_keep isl_schedule_node *node,
393 	struct ppcg_kernel *kernel)
394 {
395 	int found = 0;
396 
397 	node = isl_schedule_node_copy(node);
398 	node = isl_schedule_node_parent(node);
399 	while (!found && isl_schedule_node_has_previous_sibling(node)) {
400 		node = isl_schedule_node_previous_sibling(node);
401 		if (!node)
402 			break;
403 		found = node_is_sync_filter(node, kernel);
404 	}
405 	if (!node)
406 		found = -1;
407 	isl_schedule_node_free(node);
408 
409 	return found;
410 }
411 
412 /* Is "node" part of a sequence with a subsequent synchronization statement
413  * for "kernel"?
414  * That is, is the parent of "node" a filter such that there is
415  * a subsequent filter that picks out exactly such a synchronization statement?
416  */
has_following_sync(__isl_keep isl_schedule_node * node,struct ppcg_kernel * kernel)417 static int has_following_sync(__isl_keep isl_schedule_node *node,
418 	struct ppcg_kernel *kernel)
419 {
420 	int found = 0;
421 
422 	node = isl_schedule_node_copy(node);
423 	node = isl_schedule_node_parent(node);
424 	while (!found && isl_schedule_node_has_next_sibling(node)) {
425 		node = isl_schedule_node_next_sibling(node);
426 		if (!node)
427 			break;
428 		found = node_is_sync_filter(node, kernel);
429 	}
430 	if (!node)
431 		found = -1;
432 	isl_schedule_node_free(node);
433 
434 	return found;
435 }
436 
437 /* Does the subtree rooted at "node" (which is a band node) contain
438  * any synchronization statement for "kernel" that precedes
439  * the core computation of "kernel" (identified by the elements
440  * in kernel->core)?
441  */
has_sync_before_core(__isl_keep isl_schedule_node * node,struct ppcg_kernel * kernel)442 static int has_sync_before_core(__isl_keep isl_schedule_node *node,
443 	struct ppcg_kernel *kernel)
444 {
445 	int has_sync = 0;
446 	int is_thread;
447 
448 	node = isl_schedule_node_copy(node);
449 	while ((is_thread = node_is_thread(node)) == 0) {
450 		node = core_child(node, kernel->core);
451 		has_sync = has_preceding_sync(node, kernel);
452 		if (has_sync < 0 || has_sync)
453 			break;
454 	}
455 	if (is_thread < 0 || !node)
456 		has_sync = -1;
457 	isl_schedule_node_free(node);
458 
459 	return has_sync;
460 }
461 
462 /* Does the subtree rooted at "node" (which is a band node) contain
463  * any synchronization statement for "kernel" that follows
464  * the core computation of "kernel" (identified by the elements
465  * in kernel->core)?
466  */
has_sync_after_core(__isl_keep isl_schedule_node * node,struct ppcg_kernel * kernel)467 static int has_sync_after_core(__isl_keep isl_schedule_node *node,
468 	struct ppcg_kernel *kernel)
469 {
470 	int has_sync = 0;
471 	int is_thread;
472 
473 	node = isl_schedule_node_copy(node);
474 	while ((is_thread = node_is_thread(node)) == 0) {
475 		node = core_child(node, kernel->core);
476 		has_sync = has_following_sync(node, kernel);
477 		if (has_sync < 0 || has_sync)
478 			break;
479 	}
480 	if (is_thread < 0 || !node)
481 		has_sync = -1;
482 	isl_schedule_node_free(node);
483 
484 	return has_sync;
485 }
486 
487 /* Insert (or extend) an extension on top of "node" that puts
488  * a synchronization node for "kernel" before "node".
489  * Return a pointer to the original node in the updated schedule tree.
490  */
insert_sync_before(__isl_take isl_schedule_node * node,struct ppcg_kernel * kernel)491 static __isl_give isl_schedule_node *insert_sync_before(
492 	__isl_take isl_schedule_node *node, struct ppcg_kernel *kernel)
493 {
494 	isl_union_set *domain;
495 	isl_schedule_node *graft;
496 
497 	if (!node)
498 		return NULL;
499 
500 	domain = create_sync_domain(kernel);
501 	graft = isl_schedule_node_from_domain(domain);
502 	node = isl_schedule_node_graft_before(node, graft);
503 
504 	return node;
505 }
506 
507 /* Insert (or extend) an extension on top of "node" that puts
508  * a synchronization node for "kernel" afater "node".
509  * Return a pointer to the original node in the updated schedule tree.
510  */
insert_sync_after(__isl_take isl_schedule_node * node,struct ppcg_kernel * kernel)511 static __isl_give isl_schedule_node *insert_sync_after(
512 	__isl_take isl_schedule_node *node, struct ppcg_kernel *kernel)
513 {
514 	isl_union_set *domain;
515 	isl_schedule_node *graft;
516 
517 	if (!node)
518 		return NULL;
519 
520 	domain = create_sync_domain(kernel);
521 	graft = isl_schedule_node_from_domain(domain);
522 	node = isl_schedule_node_graft_after(node, graft);
523 
524 	return node;
525 }
526 
527 /* Insert an extension on top of "node" that puts a synchronization node
528  * for "kernel" before "node" unless there already is
529  * such a synchronization node.
530  */
gpu_tree_ensure_preceding_sync(__isl_take isl_schedule_node * node,struct ppcg_kernel * kernel)531 __isl_give isl_schedule_node *gpu_tree_ensure_preceding_sync(
532 	__isl_take isl_schedule_node *node, struct ppcg_kernel *kernel)
533 {
534 	int has_sync;
535 
536 	has_sync = has_preceding_sync(node, kernel);
537 	if (has_sync < 0)
538 		return isl_schedule_node_free(node);
539 	if (has_sync)
540 		return node;
541 	return insert_sync_before(node, kernel);
542 }
543 
544 /* Insert an extension on top of "node" that puts a synchronization node
545  * for "kernel" after "node" unless there already is
546  * such a synchronization node.
547  */
gpu_tree_ensure_following_sync(__isl_take isl_schedule_node * node,struct ppcg_kernel * kernel)548 __isl_give isl_schedule_node *gpu_tree_ensure_following_sync(
549 	__isl_take isl_schedule_node *node, struct ppcg_kernel *kernel)
550 {
551 	int has_sync;
552 
553 	has_sync = has_following_sync(node, kernel);
554 	if (has_sync < 0)
555 		return isl_schedule_node_free(node);
556 	if (has_sync)
557 		return node;
558 	return insert_sync_after(node, kernel);
559 }
560 
561 /* Insert an extension on top of "node" that puts a synchronization node
562  * for "kernel" after "node" unless there already is such a sync node or
563  * "node" itself already * contains a synchronization node following
564  * the core computation of "kernel".
565  */
gpu_tree_ensure_sync_after_core(__isl_take isl_schedule_node * node,struct ppcg_kernel * kernel)566 __isl_give isl_schedule_node *gpu_tree_ensure_sync_after_core(
567 	__isl_take isl_schedule_node *node, struct ppcg_kernel *kernel)
568 {
569 	int has_sync;
570 
571 	has_sync = has_sync_after_core(node, kernel);
572 	if (has_sync < 0)
573 		return isl_schedule_node_free(node);
574 	if (has_sync)
575 		return node;
576 	has_sync = has_following_sync(node, kernel);
577 	if (has_sync < 0)
578 		return isl_schedule_node_free(node);
579 	if (has_sync)
580 		return node;
581 	return insert_sync_after(node, kernel);
582 }
583 
584 /* Move left in the sequence on top of "node" to a synchronization node
585  * for "kernel".
586  * If "node" itself contains a synchronization node preceding
587  * the core computation of "kernel", then return "node" itself.
588  * Otherwise, if "node" does not have a preceding synchronization node,
589  * then create one first.
590  */
gpu_tree_move_left_to_sync(__isl_take isl_schedule_node * node,struct ppcg_kernel * kernel)591 __isl_give isl_schedule_node *gpu_tree_move_left_to_sync(
592 	__isl_take isl_schedule_node *node, struct ppcg_kernel *kernel)
593 {
594 	int has_sync;
595 	int is_sync;
596 
597 	has_sync = has_sync_before_core(node, kernel);
598 	if (has_sync < 0)
599 		return isl_schedule_node_free(node);
600 	if (has_sync)
601 		return node;
602 	node = gpu_tree_ensure_preceding_sync(node, kernel);
603 	node = isl_schedule_node_parent(node);
604 	while ((is_sync = node_is_sync_filter(node, kernel)) == 0)
605 		node = isl_schedule_node_previous_sibling(node);
606 	if (is_sync < 0)
607 		node = isl_schedule_node_free(node);
608 	node = isl_schedule_node_child(node, 0);
609 
610 	return node;
611 }
612 
613 /* Move right in the sequence on top of "node" to a synchronization node
614  * for "kernel".
615  * If "node" itself contains a synchronization node following
616  * the core computation of "kernel", then return "node" itself.
617  * Otherwise, if "node" does not have a following synchronization node,
618  * then create one first.
619  */
gpu_tree_move_right_to_sync(__isl_take isl_schedule_node * node,struct ppcg_kernel * kernel)620 __isl_give isl_schedule_node *gpu_tree_move_right_to_sync(
621 	__isl_take isl_schedule_node *node, struct ppcg_kernel *kernel)
622 {
623 	int has_sync;
624 	int is_sync;
625 
626 	has_sync = has_sync_after_core(node, kernel);
627 	if (has_sync < 0)
628 		return isl_schedule_node_free(node);
629 	if (has_sync)
630 		return node;
631 	node = gpu_tree_ensure_following_sync(node, kernel);
632 	node = isl_schedule_node_parent(node);
633 	while ((is_sync = node_is_sync_filter(node, kernel)) == 0)
634 		node = isl_schedule_node_next_sibling(node);
635 	if (is_sync < 0)
636 		node = isl_schedule_node_free(node);
637 	node = isl_schedule_node_child(node, 0);
638 
639 	return node;
640 }
641