2 * Copyright 2011 INRIA Saclay
3 * Copyright 2013 Ecole Normale Superieure
5 * Use of this software is governed by the MIT license
7 * Written by Sven Verdoolaege, INRIA Saclay - Ile-de-France,
8 * Parc Club Orsay Universite, ZAC des vignes, 4 rue Jacques Monod,
10 * and Ecole Normale Superieure, 45 rue d'Ulm, 75230 Paris, France
19 #include <isl/options.h>
20 #include <isl/schedule.h>
21 #include <isl/ast_build.h>
22 #include <isl/schedule.h>
25 #include "ppcg_options.h"
31 struct isl_options
*isl
;
32 struct pet_options
*pet
;
33 struct ppcg_options
*ppcg
;
38 const char *ppcg_version(void);
39 static void print_version(void)
41 printf("%s", ppcg_version());
44 ISL_ARGS_START(struct options
, options_args
)
45 ISL_ARG_CHILD(struct options
, isl
, "isl", &isl_options_args
, "isl options")
46 ISL_ARG_CHILD(struct options
, pet
, "pet", &pet_options_args
, "pet options")
47 ISL_ARG_CHILD(struct options
, ppcg
, NULL
, &ppcg_options_args
, "ppcg options")
48 ISL_ARG_STR(struct options
, output
, 'o', NULL
,
49 "filename", NULL
, "output filename (c and opencl targets)")
50 ISL_ARG_ARG(struct options
, input
, "input", NULL
)
51 ISL_ARG_VERSION(print_version
)
54 ISL_ARG_DEF(options
, struct options
, options_args
)
56 /* Return a pointer to the final path component of "filename" or
57 * to "filename" itself if it does not contain any components.
59 const char *ppcg_base_name(const char *filename
)
63 base
= strrchr(filename
, '/');
70 /* Copy the base name of "input" to "name" and return its length.
71 * "name" is not NULL terminated.
73 * In particular, remove all leading directory components and
74 * the final extension, if any.
76 int ppcg_extract_base_name(char *name
, const char *input
)
82 base
= ppcg_base_name(input
);
83 ext
= strrchr(base
, '.');
84 len
= ext
? ext
- base
: strlen(base
);
86 memcpy(name
, base
, len
);
91 /* Is "stmt" not a kill statement?
93 static int is_not_kill(struct pet_stmt
*stmt
)
95 return !pet_stmt_is_kill(stmt
);
98 /* Collect the iteration domains of the statements in "scop" that
101 static __isl_give isl_union_set
*collect_domains(struct pet_scop
*scop
,
102 int (*pred
)(struct pet_stmt
*stmt
))
106 isl_union_set
*domain
;
111 domain
= isl_union_set_empty(isl_set_get_space(scop
->context
));
113 for (i
= 0; i
< scop
->n_stmt
; ++i
) {
114 struct pet_stmt
*stmt
= scop
->stmts
[i
];
120 isl_die(isl_union_set_get_ctx(domain
),
121 isl_error_unsupported
,
122 "data dependent conditions not supported",
123 return isl_union_set_free(domain
));
125 domain_i
= isl_set_copy(scop
->stmts
[i
]->domain
);
126 domain
= isl_union_set_add_set(domain
, domain_i
);
132 /* Collect the iteration domains of the statements in "scop",
133 * skipping kill statements.
135 static __isl_give isl_union_set
*collect_non_kill_domains(struct pet_scop
*scop
)
137 return collect_domains(scop
, &is_not_kill
);
140 /* This function is used as a callback to pet_expr_foreach_call_expr
141 * to detect if there is any call expression in the input expression.
142 * Assign the value 1 to the integer that "user" points to and
143 * abort the search since we have found what we were looking for.
145 static int set_has_call(__isl_keep pet_expr
*expr
, void *user
)
147 int *has_call
= user
;
154 /* Does "expr" contain any call expressions?
156 static int expr_has_call(__isl_keep pet_expr
*expr
)
160 if (pet_expr_foreach_call_expr(expr
, &set_has_call
, &has_call
) < 0 &&
167 /* This function is a callback for pet_tree_foreach_expr.
168 * If "expr" contains any call (sub)expressions, then set *has_call
169 * and abort the search.
171 static int check_call(__isl_keep pet_expr
*expr
, void *user
)
173 int *has_call
= user
;
175 if (expr_has_call(expr
))
178 return *has_call
? -1 : 0;
181 /* Does "stmt" contain any call expressions?
183 static int has_call(struct pet_stmt
*stmt
)
187 if (pet_tree_foreach_expr(stmt
->body
, &check_call
, &has_call
) < 0 &&
194 /* Collect the iteration domains of the statements in "scop"
195 * that contain a call expression.
197 static __isl_give isl_union_set
*collect_call_domains(struct pet_scop
*scop
)
199 return collect_domains(scop
, &has_call
);
202 /* Given a union of "tagged" access relations of the form
204 * [S_i[...] -> R_j[]] -> A_k[...]
206 * project out the "tags" (R_j[]).
207 * That is, return a union of relations of the form
209 * S_i[...] -> A_k[...]
211 static __isl_give isl_union_map
*project_out_tags(
212 __isl_take isl_union_map
*umap
)
216 proj
= isl_union_map_universe(isl_union_map_copy(umap
));
217 proj
= isl_union_set_unwrap(isl_union_map_domain(proj
));
218 proj
= isl_union_map_domain_map(proj
);
220 umap
= isl_union_map_apply_domain(umap
, proj
);
225 /* Construct a relation from the iteration domains to tagged iteration
226 * domains with as range the reference tags that appear
227 * in any of the reads, writes or kills.
228 * Store the result in ps->tagger.
230 * For example, if the statement with iteration space S[i,j]
231 * contains two array references R_1[] and R_2[], then ps->tagger will contain
233 * { S[i,j] -> [S[i,j] -> R_1[]]; S[i,j] -> [S[i,j] -> R_2[]] }
235 static void compute_tagger(struct ppcg_scop
*ps
)
237 isl_union_map
*tagged
, *tagger
;
239 tagged
= isl_union_map_copy(ps
->tagged_reads
);
240 tagged
= isl_union_map_union(tagged
,
241 isl_union_map_copy(ps
->tagged_may_writes
));
242 tagged
= isl_union_map_union(tagged
,
243 isl_union_map_copy(ps
->tagged_must_kills
));
245 tagger
= isl_union_map_universe(tagged
);
246 tagger
= isl_union_set_unwrap(isl_union_map_domain(tagger
));
247 tagger
= isl_union_map_reverse(isl_union_map_domain_map(tagger
));
252 /* Compute the live out accesses, i.e., the writes that are
253 * potentially not killed by any kills or any other writes, and
254 * store them in ps->live_out.
256 * We compute the "dependence" of any "kill" (an explicit kill
257 * or a must write) on any may write.
258 * The may writes with a "depending" kill are definitely killed.
259 * The remaining may writes can potentially be live out.
261 static void compute_live_out(struct ppcg_scop
*ps
)
263 isl_union_map
*tagger
;
264 isl_union_map
*schedule
;
265 isl_union_map
*empty
;
266 isl_union_map
*kills
;
267 isl_union_map
*exposed
;
268 isl_union_map
*covering
;
270 tagger
= isl_union_map_copy(ps
->tagger
);
271 schedule
= isl_union_map_copy(ps
->schedule
);
272 schedule
= isl_union_map_apply_domain(schedule
,
273 isl_union_map_copy(tagger
));
274 empty
= isl_union_map_empty(isl_union_set_get_space(ps
->domain
));
275 kills
= isl_union_map_union(isl_union_map_copy(ps
->tagged_must_writes
),
276 isl_union_map_copy(ps
->tagged_must_kills
));
277 isl_union_map_compute_flow(kills
, empty
,
278 isl_union_map_copy(ps
->tagged_may_writes
),
279 schedule
, NULL
, &covering
, NULL
, NULL
);
280 exposed
= isl_union_map_copy(ps
->tagged_may_writes
);
281 exposed
= isl_union_map_subtract_domain(exposed
,
282 isl_union_map_domain(covering
));
283 exposed
= isl_union_map_apply_range(tagger
, exposed
);
284 ps
->live_out
= exposed
;
287 /* Compute the flow dependences and the live_in accesses and store
288 * the results in ps->dep_flow and ps->live_in.
289 * A copy of the flow dependences, tagged with the reference tags
290 * is stored in ps->tagged_dep_flow.
292 * We first compute ps->tagged_dep_flow, i.e., the tagged flow dependences
293 * and then project out the tags.
295 * We allow both the must writes and the must kills to serve as
296 * definite sources such that a subsequent read would not depend
297 * on any earlier write. The resulting flow dependences with
298 * a must kill as source reflect possibly uninitialized reads.
299 * No dependences need to be introduced to protect such reads
300 * (other than those imposed by potential flows from may writes
301 * that follow the kill). We therefore those flow dependences.
302 * This is also useful for the dead code elimination, which assumes
303 * the flow sources are non-kill instances.
305 static void compute_tagged_flow_dep(struct ppcg_scop
*ps
)
307 isl_union_map
*tagger
;
308 isl_union_map
*schedule
;
309 isl_union_map
*may_flow
;
310 isl_union_map
*live_in
, *may_live_in
;
311 isl_union_map
*must_source
;
312 isl_union_map
*kills
;
313 isl_union_map
*tagged_flow
;
315 tagger
= isl_union_map_copy(ps
->tagger
);
316 schedule
= isl_union_map_copy(ps
->schedule
);
317 schedule
= isl_union_map_apply_domain(schedule
, tagger
);
318 kills
= isl_union_map_copy(ps
->tagged_must_kills
);
319 must_source
= isl_union_map_copy(ps
->tagged_must_writes
);
320 must_source
= isl_union_map_union(must_source
,
321 isl_union_map_copy(kills
));
322 isl_union_map_compute_flow(isl_union_map_copy(ps
->tagged_reads
),
324 isl_union_map_copy(ps
->tagged_may_writes
),
325 schedule
, &ps
->tagged_dep_flow
, &may_flow
,
326 &live_in
, &may_live_in
);
327 tagged_flow
= isl_union_map_union(ps
->tagged_dep_flow
, may_flow
);
328 tagged_flow
= isl_union_map_subtract_domain(tagged_flow
,
329 isl_union_map_domain(kills
));
330 ps
->tagged_dep_flow
= tagged_flow
;
331 ps
->dep_flow
= isl_union_map_copy(ps
->tagged_dep_flow
);
332 ps
->dep_flow
= isl_union_map_zip(ps
->dep_flow
);
333 ps
->dep_flow
= isl_union_set_unwrap(isl_union_map_domain(ps
->dep_flow
));
334 live_in
= isl_union_map_union(live_in
, may_live_in
);
335 ps
->live_in
= project_out_tags(live_in
);
338 /* Compute the order dependences that prevent the potential live ranges
340 * "before" contains all pairs of statement iterations where
341 * the first is executed before the second according to the original schedule.
343 * In particular, construct a union of relations
345 * [R[...] -> R_1[]] -> [W[...] -> R_2[]]
347 * where [R[...] -> R_1[]] is the range of one or more live ranges
348 * (i.e., a read) and [W[...] -> R_2[]] is the domain of one or more
349 * live ranges (i.e., a write). Moreover, the read and the write
350 * access the same memory element and the read occurs before the write
351 * in the original schedule.
352 * The scheduler allows some of these dependences to be violated, provided
353 * the adjacent live ranges are all local (i.e., their domain and range
354 * are mapped to the same point by the current schedule band).
356 * Note that if a live range is not local, then we need to make
357 * sure it does not overlap with _any_ other live range, and not
358 * just with the "previous" and/or the "next" live range.
359 * We therefore add order dependences between reads and
360 * _any_ later potential write.
362 * We also need to be careful about writes without a corresponding read.
363 * They are already prevented from moving past non-local preceding
364 * intervals, but we also need to prevent them from moving past non-local
365 * following intervals. We therefore also add order dependences from
366 * potential writes that do not appear in any intervals
367 * to all later potential writes.
368 * Note that dead code elimination should have removed most of these
369 * dead writes, but the dead code elimination may not remove all dead writes,
370 * so we need to consider them to be safe.
372 static void compute_order_dependences(struct ppcg_scop
*ps
,
373 __isl_take isl_union_map
*before
)
375 isl_union_map
*reads
;
376 isl_union_map
*shared_access
;
377 isl_union_set
*matched
;
378 isl_union_map
*unmatched
;
379 isl_union_set
*domain
;
381 reads
= isl_union_map_copy(ps
->tagged_reads
);
382 matched
= isl_union_map_domain(isl_union_map_copy(ps
->tagged_dep_flow
));
383 unmatched
= isl_union_map_copy(ps
->tagged_may_writes
);
384 unmatched
= isl_union_map_subtract_domain(unmatched
, matched
);
385 reads
= isl_union_map_union(reads
, unmatched
);
386 shared_access
= isl_union_map_copy(ps
->tagged_may_writes
);
387 shared_access
= isl_union_map_reverse(shared_access
);
388 shared_access
= isl_union_map_apply_range(reads
, shared_access
);
389 shared_access
= isl_union_map_zip(shared_access
);
390 shared_access
= isl_union_map_intersect_domain(shared_access
,
391 isl_union_map_wrap(before
));
392 domain
= isl_union_map_domain(isl_union_map_copy(shared_access
));
393 shared_access
= isl_union_map_zip(shared_access
);
394 ps
->dep_order
= isl_union_set_unwrap(domain
);
395 ps
->tagged_dep_order
= shared_access
;
398 /* Compute the external false dependences of the program represented by "scop"
399 * in case live range reordering is allowed.
400 * "before" contains all pairs of statement iterations where
401 * the first is executed before the second according to the original schedule.
403 * The anti-dependences are already taken care of by the order dependences.
404 * The external false dependences are only used to ensure that live-in and
405 * live-out data is not overwritten by any writes inside the scop.
407 * In particular, the reads from live-in data need to precede any
408 * later write to the same memory element.
409 * As to live-out data, the last writes need to remain the last writes.
410 * That is, any earlier write in the original schedule needs to precede
411 * the last write to the same memory element in the computed schedule.
412 * The possible last writes have been computed by compute_live_out.
413 * They may include kills, but if the last access is a kill,
414 * then the corresponding dependences will effectively be ignored
415 * since we do not schedule any kill statements.
417 * Note that the set of live-in and live-out accesses may be
418 * an overapproximation. There may therefore be potential writes
419 * before a live-in access and after a live-out access.
421 static void compute_external_false_dependences(struct ppcg_scop
*ps
,
422 __isl_take isl_union_map
*before
)
424 isl_union_map
*shared_access
;
425 isl_union_map
*exposed
;
426 isl_union_map
*live_in
;
428 exposed
= isl_union_map_copy(ps
->live_out
);
430 exposed
= isl_union_map_reverse(exposed
);
431 shared_access
= isl_union_map_copy(ps
->may_writes
);
432 shared_access
= isl_union_map_apply_range(shared_access
, exposed
);
434 ps
->dep_external
= shared_access
;
436 live_in
= isl_union_map_apply_range(isl_union_map_copy(ps
->live_in
),
437 isl_union_map_reverse(isl_union_map_copy(ps
->may_writes
)));
439 ps
->dep_external
= isl_union_map_union(ps
->dep_external
, live_in
);
440 ps
->dep_external
= isl_union_map_intersect(ps
->dep_external
, before
);
443 /* Compute the dependences of the program represented by "scop"
444 * in case live range reordering is allowed.
446 * We compute the actual live ranges and the corresponding order
449 static void compute_live_range_reordering_dependences(struct ppcg_scop
*ps
)
451 isl_union_map
*before
;
453 before
= isl_union_map_lex_lt_union_map(
454 isl_union_map_copy(ps
->schedule
),
455 isl_union_map_copy(ps
->schedule
));
457 compute_tagged_flow_dep(ps
);
458 compute_order_dependences(ps
, isl_union_map_copy(before
));
459 compute_external_false_dependences(ps
, before
);
462 /* Compute the potential flow dependences and the potential live in
465 static void compute_flow_dep(struct ppcg_scop
*ps
)
467 isl_union_map
*may_flow
;
468 isl_union_map
*may_live_in
;
470 isl_union_map_compute_flow(isl_union_map_copy(ps
->reads
),
471 isl_union_map_copy(ps
->must_writes
),
472 isl_union_map_copy(ps
->may_writes
),
473 isl_union_map_copy(ps
->schedule
),
474 &ps
->dep_flow
, &may_flow
,
475 &ps
->live_in
, &may_live_in
);
477 ps
->dep_flow
= isl_union_map_union(ps
->dep_flow
, may_flow
);
478 ps
->live_in
= isl_union_map_union(ps
->live_in
, may_live_in
);
481 /* Compute the dependences of the program represented by "scop".
482 * Store the computed potential flow dependences
483 * in scop->dep_flow and the reads with potentially no corresponding writes in
485 * Store the potential live out accesses in scop->live_out.
486 * Store the potential false (anti and output) dependences in scop->dep_false.
488 * If live range reordering is allowed, then we compute a separate
489 * set of order dependences and a set of external false dependences
490 * in compute_live_range_reordering_dependences.
492 static void compute_dependences(struct ppcg_scop
*scop
)
494 isl_union_map
*dep1
, *dep2
;
495 isl_union_map
*may_source
;
500 compute_live_out(scop
);
502 if (scop
->options
->live_range_reordering
)
503 compute_live_range_reordering_dependences(scop
);
504 else if (scop
->options
->target
!= PPCG_TARGET_C
)
505 compute_tagged_flow_dep(scop
);
507 compute_flow_dep(scop
);
509 may_source
= isl_union_map_union(isl_union_map_copy(scop
->may_writes
),
510 isl_union_map_copy(scop
->reads
));
511 isl_union_map_compute_flow(isl_union_map_copy(scop
->may_writes
),
512 isl_union_map_copy(scop
->must_writes
),
513 may_source
, isl_union_map_copy(scop
->schedule
),
514 &dep1
, &dep2
, NULL
, NULL
);
516 scop
->dep_false
= isl_union_map_union(dep1
, dep2
);
517 scop
->dep_false
= isl_union_map_coalesce(scop
->dep_false
);
520 /* Eliminate dead code from ps->domain.
522 * In particular, intersect ps->domain with the (parts of) iteration
523 * domains that are needed to produce the output or for statement
524 * iterations that call functions.
525 * Also intersect the range of the dataflow dependences with
526 * this domain such that the removed instances will no longer
527 * be considered as targets of dataflow.
529 * We start with the iteration domains that call functions
530 * and the set of iterations that last write to an array
531 * (except those that are later killed).
533 * Then we add those statement iterations that produce
534 * something needed by the "live" statements iterations.
535 * We keep doing this until no more statement iterations can be added.
536 * To ensure that the procedure terminates, we compute the affine
537 * hull of the live iterations (bounded to the original iteration
538 * domains) each time we have added extra iterations.
540 static void eliminate_dead_code(struct ppcg_scop
*ps
)
544 isl_union_map
*tagger
;
546 live
= isl_union_map_domain(isl_union_map_copy(ps
->live_out
));
547 if (!isl_union_set_is_empty(ps
->call
)) {
548 live
= isl_union_set_union(live
, isl_union_set_copy(ps
->call
));
549 live
= isl_union_set_coalesce(live
);
552 dep
= isl_union_map_copy(ps
->dep_flow
);
553 dep
= isl_union_map_reverse(dep
);
556 isl_union_set
*extra
;
558 extra
= isl_union_set_apply(isl_union_set_copy(live
),
559 isl_union_map_copy(dep
));
560 if (isl_union_set_is_subset(extra
, live
)) {
561 isl_union_set_free(extra
);
565 live
= isl_union_set_union(live
, extra
);
566 live
= isl_union_set_affine_hull(live
);
567 live
= isl_union_set_intersect(live
,
568 isl_union_set_copy(ps
->domain
));
571 isl_union_map_free(dep
);
573 ps
->domain
= isl_union_set_intersect(ps
->domain
,
574 isl_union_set_copy(live
));
575 ps
->dep_flow
= isl_union_map_intersect_range(ps
->dep_flow
,
576 isl_union_set_copy(live
));
577 tagger
= isl_union_map_copy(ps
->tagger
);
578 live
= isl_union_set_apply(live
, tagger
);
579 ps
->tagged_dep_flow
= isl_union_map_intersect_range(ps
->tagged_dep_flow
,
583 /* Intersect "set" with the set described by "str", taking the NULL
584 * string to represent the universal set.
586 static __isl_give isl_set
*set_intersect_str(__isl_take isl_set
*set
,
595 ctx
= isl_set_get_ctx(set
);
596 set2
= isl_set_read_from_str(ctx
, str
);
597 set
= isl_set_intersect(set
, set2
);
602 static void *ppcg_scop_free(struct ppcg_scop
*ps
)
607 isl_set_free(ps
->context
);
608 isl_union_set_free(ps
->domain
);
609 isl_union_set_free(ps
->call
);
610 isl_union_map_free(ps
->tagged_reads
);
611 isl_union_map_free(ps
->reads
);
612 isl_union_map_free(ps
->live_in
);
613 isl_union_map_free(ps
->tagged_may_writes
);
614 isl_union_map_free(ps
->tagged_must_writes
);
615 isl_union_map_free(ps
->may_writes
);
616 isl_union_map_free(ps
->must_writes
);
617 isl_union_map_free(ps
->live_out
);
618 isl_union_map_free(ps
->tagged_must_kills
);
619 isl_union_map_free(ps
->tagged_dep_flow
);
620 isl_union_map_free(ps
->dep_flow
);
621 isl_union_map_free(ps
->dep_false
);
622 isl_union_map_free(ps
->dep_external
);
623 isl_union_map_free(ps
->tagged_dep_order
);
624 isl_union_map_free(ps
->dep_order
);
625 isl_union_map_free(ps
->schedule
);
626 isl_union_map_free(ps
->tagger
);
627 isl_union_map_free(ps
->independence
);
634 /* Extract a ppcg_scop from a pet_scop.
636 * The constructed ppcg_scop refers to elements from the pet_scop
637 * so the pet_scop should not be freed before the ppcg_scop.
639 static struct ppcg_scop
*ppcg_scop_from_pet_scop(struct pet_scop
*scop
,
640 struct ppcg_options
*options
)
644 struct ppcg_scop
*ps
;
649 ctx
= isl_set_get_ctx(scop
->context
);
651 ps
= isl_calloc_type(ctx
, struct ppcg_scop
);
655 ps
->options
= options
;
656 ps
->start
= pet_loc_get_start(scop
->loc
);
657 ps
->end
= pet_loc_get_end(scop
->loc
);
658 ps
->context
= isl_set_copy(scop
->context
);
659 ps
->context
= set_intersect_str(ps
->context
, options
->ctx
);
660 ps
->domain
= collect_non_kill_domains(scop
);
661 ps
->call
= collect_call_domains(scop
);
662 ps
->tagged_reads
= pet_scop_collect_tagged_may_reads(scop
);
663 ps
->reads
= pet_scop_collect_may_reads(scop
);
664 ps
->tagged_may_writes
= pet_scop_collect_tagged_may_writes(scop
);
665 ps
->may_writes
= pet_scop_collect_may_writes(scop
);
666 ps
->tagged_must_writes
= pet_scop_collect_tagged_must_writes(scop
);
667 ps
->must_writes
= pet_scop_collect_must_writes(scop
);
668 ps
->tagged_must_kills
= pet_scop_collect_tagged_must_kills(scop
);
669 ps
->schedule
= pet_scop_collect_schedule(scop
);
671 ps
->independence
= isl_union_map_empty(isl_set_get_space(ps
->context
));
672 for (i
= 0; i
< scop
->n_independence
; ++i
)
673 ps
->independence
= isl_union_map_union(ps
->independence
,
674 isl_union_map_copy(scop
->independences
[i
]->filter
));
677 compute_dependences(ps
);
678 eliminate_dead_code(ps
);
680 if (!ps
->context
|| !ps
->domain
|| !ps
->call
|| !ps
->reads
||
681 !ps
->may_writes
|| !ps
->must_writes
|| !ps
->tagged_must_kills
||
682 !ps
->schedule
|| !ps
->independence
)
683 return ppcg_scop_free(ps
);
688 /* Internal data structure for ppcg_transform.
690 struct ppcg_transform_data
{
691 struct ppcg_options
*options
;
692 __isl_give isl_printer
*(*transform
)(__isl_take isl_printer
*p
,
693 struct ppcg_scop
*scop
, void *user
);
697 /* Should we print the original code?
698 * That is, does "scop" involve any data dependent conditions or
699 * nested expressions that cannot be handled by pet_stmt_build_ast_exprs?
701 static int print_original(struct pet_scop
*scop
, struct ppcg_options
*options
)
703 if (!pet_scop_can_build_ast_exprs(scop
)) {
704 if (options
->debug
->verbose
)
705 fprintf(stdout
, "Printing original code because "
706 "some index expressions cannot currently "
711 if (pet_scop_has_data_dependent_conditions(scop
)) {
712 if (options
->debug
->verbose
)
713 fprintf(stdout
, "Printing original code because "
714 "input involves data dependent conditions\n");
721 /* Callback for pet_transform_C_source that transforms
722 * the given pet_scop to a ppcg_scop before calling the
723 * ppcg_transform callback.
725 * If "scop" contains any data dependent conditions or if we may
726 * not be able to print the transformed program, then just print
729 static __isl_give isl_printer
*transform(__isl_take isl_printer
*p
,
730 struct pet_scop
*scop
, void *user
)
732 struct ppcg_transform_data
*data
= user
;
733 struct ppcg_scop
*ps
;
735 if (print_original(scop
, data
->options
)) {
736 p
= pet_scop_print_original(scop
, p
);
741 scop
= pet_scop_align_params(scop
);
742 ps
= ppcg_scop_from_pet_scop(scop
, data
->options
);
744 p
= data
->transform(p
, ps
, data
->user
);
752 /* Transform the C source file "input" by rewriting each scop
753 * through a call to "transform".
754 * The transformed C code is written to "out".
756 * This is a wrapper around pet_transform_C_source that transforms
757 * the pet_scop to a ppcg_scop before calling "fn".
759 int ppcg_transform(isl_ctx
*ctx
, const char *input
, FILE *out
,
760 struct ppcg_options
*options
,
761 __isl_give isl_printer
*(*fn
)(__isl_take isl_printer
*p
,
762 struct ppcg_scop
*scop
, void *user
), void *user
)
764 struct ppcg_transform_data data
= { options
, fn
, user
};
765 return pet_transform_C_source(ctx
, input
, out
, &transform
, &data
);
768 /* Check consistency of options.
770 * Return -1 on error.
772 static int check_options(isl_ctx
*ctx
)
774 struct options
*options
;
776 options
= isl_ctx_peek_options(ctx
, &options_args
);
778 isl_die(ctx
, isl_error_internal
,
779 "unable to find options", return -1);
781 if (options
->ppcg
->openmp
&&
782 !isl_options_get_ast_build_atomic_upper_bound(ctx
))
783 isl_die(ctx
, isl_error_invalid
,
784 "OpenMP requires atomic bounds", return -1);
789 int main(int argc
, char **argv
)
793 struct options
*options
;
795 options
= options_new_with_defaults();
798 ctx
= isl_ctx_alloc_with_options(&options_args
, options
);
799 isl_options_set_schedule_outer_coincidence(ctx
, 1);
800 isl_options_set_schedule_maximize_band_depth(ctx
, 1);
801 pet_options_set_encapsulate_dynamic_control(ctx
, 1);
802 argc
= options_parse(options
, argc
, argv
, ISL_ARG_ALL
);
804 if (check_options(ctx
) < 0)
806 else if (options
->ppcg
->target
== PPCG_TARGET_CUDA
)
807 r
= generate_cuda(ctx
, options
->ppcg
, options
->input
);
808 else if (options
->ppcg
->target
== PPCG_TARGET_OPENCL
)
809 r
= generate_opencl(ctx
, options
->ppcg
, options
->input
,
812 r
= generate_cpu(ctx
, options
->ppcg
, options
->input
,