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 /* Collect all variable names that are in use in "scop".
92 * In particular, collect all parameters in the context and
93 * all the array names.
94 * Store these names in an isl_id_to_ast_expr by mapping
95 * them to a dummy value (0).
97 static __isl_give isl_id_to_ast_expr
*collect_names(struct pet_scop
*scop
)
102 isl_id_to_ast_expr
*names
;
104 ctx
= isl_set_get_ctx(scop
->context
);
106 n
= isl_set_dim(scop
->context
, isl_dim_param
);
108 names
= isl_id_to_ast_expr_alloc(ctx
, n
+ scop
->n_array
);
109 zero
= isl_ast_expr_from_val(isl_val_zero(ctx
));
111 for (i
= 0; i
< n
; ++i
) {
114 id
= isl_set_get_dim_id(scop
->context
, isl_dim_param
, i
);
115 names
= isl_id_to_ast_expr_set(names
,
116 id
, isl_ast_expr_copy(zero
));
119 for (i
= 0; i
< scop
->n_array
; ++i
) {
120 struct pet_array
*array
= scop
->arrays
[i
];
123 id
= isl_set_get_tuple_id(array
->extent
);
124 names
= isl_id_to_ast_expr_set(names
,
125 id
, isl_ast_expr_copy(zero
));
128 isl_ast_expr_free(zero
);
133 /* Return an isl_id called "prefix%d", with "%d" set to "i".
134 * If an isl_id with such a name already appears among the variable names
135 * of "scop", then adjust the name to "prefix%d_%d".
137 static __isl_give isl_id
*generate_name(struct ppcg_scop
*scop
,
138 const char *prefix
, int i
)
146 ctx
= isl_set_get_ctx(scop
->context
);
147 snprintf(name
, sizeof(name
), "%s%d", prefix
, i
);
148 id
= isl_id_alloc(ctx
, name
, NULL
);
151 while ((has_name
= isl_id_to_ast_expr_has(scop
->names
, id
)) == 1) {
153 snprintf(name
, sizeof(name
), "%s%d_%d", prefix
, i
, j
++);
154 id
= isl_id_alloc(ctx
, name
, NULL
);
157 return has_name
< 0 ? isl_id_free(id
) : id
;
160 /* Return a list of "n" isl_ids of the form "prefix%d".
161 * If an isl_id with such a name already appears among the variable names
162 * of "scop", then adjust the name to "prefix%d_%d".
164 __isl_give isl_id_list
*ppcg_scop_generate_names(struct ppcg_scop
*scop
,
165 int n
, const char *prefix
)
172 ctx
= isl_set_get_ctx(scop
->context
);
173 names
= isl_id_list_alloc(ctx
, n
);
174 for (i
= 0; i
< n
; ++i
) {
177 id
= generate_name(scop
, prefix
, i
);
178 names
= isl_id_list_add(names
, id
);
184 /* Is "stmt" not a kill statement?
186 static int is_not_kill(struct pet_stmt
*stmt
)
188 return !pet_stmt_is_kill(stmt
);
191 /* Collect the iteration domains of the statements in "scop" that
194 static __isl_give isl_union_set
*collect_domains(struct pet_scop
*scop
,
195 int (*pred
)(struct pet_stmt
*stmt
))
199 isl_union_set
*domain
;
204 domain
= isl_union_set_empty(isl_set_get_space(scop
->context
));
206 for (i
= 0; i
< scop
->n_stmt
; ++i
) {
207 struct pet_stmt
*stmt
= scop
->stmts
[i
];
213 isl_die(isl_union_set_get_ctx(domain
),
214 isl_error_unsupported
,
215 "data dependent conditions not supported",
216 return isl_union_set_free(domain
));
218 domain_i
= isl_set_copy(scop
->stmts
[i
]->domain
);
219 domain
= isl_union_set_add_set(domain
, domain_i
);
225 /* Collect the iteration domains of the statements in "scop",
226 * skipping kill statements.
228 static __isl_give isl_union_set
*collect_non_kill_domains(struct pet_scop
*scop
)
230 return collect_domains(scop
, &is_not_kill
);
233 /* This function is used as a callback to pet_expr_foreach_call_expr
234 * to detect if there is any call expression in the input expression.
235 * Assign the value 1 to the integer that "user" points to and
236 * abort the search since we have found what we were looking for.
238 static int set_has_call(__isl_keep pet_expr
*expr
, void *user
)
240 int *has_call
= user
;
247 /* Does "expr" contain any call expressions?
249 static int expr_has_call(__isl_keep pet_expr
*expr
)
253 if (pet_expr_foreach_call_expr(expr
, &set_has_call
, &has_call
) < 0 &&
260 /* This function is a callback for pet_tree_foreach_expr.
261 * If "expr" contains any call (sub)expressions, then set *has_call
262 * and abort the search.
264 static int check_call(__isl_keep pet_expr
*expr
, void *user
)
266 int *has_call
= user
;
268 if (expr_has_call(expr
))
271 return *has_call
? -1 : 0;
274 /* Does "stmt" contain any call expressions?
276 static int has_call(struct pet_stmt
*stmt
)
280 if (pet_tree_foreach_expr(stmt
->body
, &check_call
, &has_call
) < 0 &&
287 /* Collect the iteration domains of the statements in "scop"
288 * that contain a call expression.
290 static __isl_give isl_union_set
*collect_call_domains(struct pet_scop
*scop
)
292 return collect_domains(scop
, &has_call
);
295 /* Given a union of "tagged" access relations of the form
297 * [S_i[...] -> R_j[]] -> A_k[...]
299 * project out the "tags" (R_j[]).
300 * That is, return a union of relations of the form
302 * S_i[...] -> A_k[...]
304 static __isl_give isl_union_map
*project_out_tags(
305 __isl_take isl_union_map
*umap
)
307 return isl_union_map_domain_factor_domain(umap
);
310 /* Construct a function from tagged iteration domains to the corresponding
311 * untagged iteration domains with as range of the wrapped map in the domain
312 * the reference tags that appear in any of the reads, writes or kills.
313 * Store the result in ps->tagger.
315 * For example, if the statement with iteration space S[i,j]
316 * contains two array references R_1[] and R_2[], then ps->tagger will contain
318 * { [S[i,j] -> R_1[]] -> S[i,j]; [S[i,j] -> R_2[]] -> S[i,j] }
320 static void compute_tagger(struct ppcg_scop
*ps
)
322 isl_union_map
*tagged
;
323 isl_union_pw_multi_aff
*tagger
;
325 tagged
= isl_union_map_copy(ps
->tagged_reads
);
326 tagged
= isl_union_map_union(tagged
,
327 isl_union_map_copy(ps
->tagged_may_writes
));
328 tagged
= isl_union_map_union(tagged
,
329 isl_union_map_copy(ps
->tagged_must_kills
));
330 tagged
= isl_union_map_universe(tagged
);
331 tagged
= isl_union_set_unwrap(isl_union_map_domain(tagged
));
333 tagger
= isl_union_map_domain_map_union_pw_multi_aff(tagged
);
338 /* Compute the live out accesses, i.e., the writes that are
339 * potentially not killed by any kills or any other writes, and
340 * store them in ps->live_out.
342 * We compute the "dependence" of any "kill" (an explicit kill
343 * or a must write) on any may write.
344 * The may writes with a "depending" kill are definitely killed.
345 * The remaining may writes can potentially be live out.
347 static void compute_live_out(struct ppcg_scop
*ps
)
349 isl_union_pw_multi_aff
*tagger
;
350 isl_schedule
*schedule
;
351 isl_union_map
*kills
;
352 isl_union_map
*exposed
;
353 isl_union_map
*covering
;
354 isl_union_access_info
*access
;
355 isl_union_flow
*flow
;
357 tagger
= isl_union_pw_multi_aff_copy(ps
->tagger
);
358 schedule
= isl_schedule_copy(ps
->schedule
);
359 schedule
= isl_schedule_pullback_union_pw_multi_aff(schedule
, tagger
);
360 kills
= isl_union_map_union(isl_union_map_copy(ps
->tagged_must_writes
),
361 isl_union_map_copy(ps
->tagged_must_kills
));
362 access
= isl_union_access_info_from_sink(kills
);
363 access
= isl_union_access_info_set_may_source(access
,
364 isl_union_map_copy(ps
->tagged_may_writes
));
365 access
= isl_union_access_info_set_schedule(access
, schedule
);
366 flow
= isl_union_access_info_compute_flow(access
);
367 covering
= isl_union_flow_get_may_dependence(flow
);
368 isl_union_flow_free(flow
);
369 exposed
= isl_union_map_copy(ps
->tagged_may_writes
);
370 exposed
= isl_union_map_subtract_domain(exposed
,
371 isl_union_map_domain(covering
));
372 ps
->live_out
= project_out_tags(exposed
);
375 /* Compute the flow dependences and the live_in accesses and store
376 * the results in ps->dep_flow and ps->live_in.
377 * A copy of the flow dependences, tagged with the reference tags
378 * is stored in ps->tagged_dep_flow.
380 * We first compute ps->tagged_dep_flow, i.e., the tagged flow dependences
381 * and then project out the tags.
383 * We allow both the must writes and the must kills to serve as
384 * definite sources such that a subsequent read would not depend
385 * on any earlier write. The resulting flow dependences with
386 * a must kill as source reflect possibly uninitialized reads.
387 * No dependences need to be introduced to protect such reads
388 * (other than those imposed by potential flows from may writes
389 * that follow the kill). We therefore those flow dependences.
390 * This is also useful for the dead code elimination, which assumes
391 * the flow sources are non-kill instances.
393 static void compute_tagged_flow_dep(struct ppcg_scop
*ps
)
395 isl_union_pw_multi_aff
*tagger
;
396 isl_schedule
*schedule
;
397 isl_union_map
*live_in
;
398 isl_union_access_info
*access
;
399 isl_union_flow
*flow
;
400 isl_union_map
*must_source
;
401 isl_union_map
*kills
;
402 isl_union_map
*tagged_flow
;
404 tagger
= isl_union_pw_multi_aff_copy(ps
->tagger
);
405 schedule
= isl_schedule_copy(ps
->schedule
);
406 schedule
= isl_schedule_pullback_union_pw_multi_aff(schedule
, tagger
);
407 kills
= isl_union_map_copy(ps
->tagged_must_kills
);
408 must_source
= isl_union_map_copy(ps
->tagged_must_writes
);
409 must_source
= isl_union_map_union(must_source
,
410 isl_union_map_copy(kills
));
411 access
= isl_union_access_info_from_sink(
412 isl_union_map_copy(ps
->tagged_reads
));
413 access
= isl_union_access_info_set_must_source(access
, must_source
);
414 access
= isl_union_access_info_set_may_source(access
,
415 isl_union_map_copy(ps
->tagged_may_writes
));
416 access
= isl_union_access_info_set_schedule(access
, schedule
);
417 flow
= isl_union_access_info_compute_flow(access
);
418 tagged_flow
= isl_union_flow_get_may_dependence(flow
);
419 tagged_flow
= isl_union_map_subtract_domain(tagged_flow
,
420 isl_union_map_domain(kills
));
421 ps
->tagged_dep_flow
= tagged_flow
;
422 ps
->dep_flow
= isl_union_map_copy(ps
->tagged_dep_flow
);
423 ps
->dep_flow
= isl_union_map_factor_domain(ps
->dep_flow
);
424 live_in
= isl_union_flow_get_may_no_source(flow
);
425 ps
->live_in
= project_out_tags(live_in
);
426 isl_union_flow_free(flow
);
429 /* Compute the order dependences that prevent the potential live ranges
431 * "before" contains all pairs of statement iterations where
432 * the first is executed before the second according to the original schedule.
434 * In particular, construct a union of relations
436 * [R[...] -> R_1[]] -> [W[...] -> R_2[]]
438 * where [R[...] -> R_1[]] is the range of one or more live ranges
439 * (i.e., a read) and [W[...] -> R_2[]] is the domain of one or more
440 * live ranges (i.e., a write). Moreover, the read and the write
441 * access the same memory element and the read occurs before the write
442 * in the original schedule.
443 * The scheduler allows some of these dependences to be violated, provided
444 * the adjacent live ranges are all local (i.e., their domain and range
445 * are mapped to the same point by the current schedule band).
447 * Note that if a live range is not local, then we need to make
448 * sure it does not overlap with _any_ other live range, and not
449 * just with the "previous" and/or the "next" live range.
450 * We therefore add order dependences between reads and
451 * _any_ later potential write.
453 * We also need to be careful about writes without a corresponding read.
454 * They are already prevented from moving past non-local preceding
455 * intervals, but we also need to prevent them from moving past non-local
456 * following intervals. We therefore also add order dependences from
457 * potential writes that do not appear in any intervals
458 * to all later potential writes.
459 * Note that dead code elimination should have removed most of these
460 * dead writes, but the dead code elimination may not remove all dead writes,
461 * so we need to consider them to be safe.
463 static void compute_order_dependences(struct ppcg_scop
*ps
,
464 __isl_take isl_union_map
*before
)
466 isl_union_map
*reads
;
467 isl_union_map
*shared_access
;
468 isl_union_set
*matched
;
469 isl_union_map
*unmatched
;
470 isl_union_set
*domain
;
472 reads
= isl_union_map_copy(ps
->tagged_reads
);
473 matched
= isl_union_map_domain(isl_union_map_copy(ps
->tagged_dep_flow
));
474 unmatched
= isl_union_map_copy(ps
->tagged_may_writes
);
475 unmatched
= isl_union_map_subtract_domain(unmatched
, matched
);
476 reads
= isl_union_map_union(reads
, unmatched
);
477 shared_access
= isl_union_map_copy(ps
->tagged_may_writes
);
478 shared_access
= isl_union_map_reverse(shared_access
);
479 shared_access
= isl_union_map_apply_range(reads
, shared_access
);
480 shared_access
= isl_union_map_zip(shared_access
);
481 shared_access
= isl_union_map_intersect_domain(shared_access
,
482 isl_union_map_wrap(before
));
483 domain
= isl_union_map_domain(isl_union_map_copy(shared_access
));
484 shared_access
= isl_union_map_zip(shared_access
);
485 ps
->dep_order
= isl_union_set_unwrap(domain
);
486 ps
->tagged_dep_order
= shared_access
;
489 /* Compute those validity dependences of the program represented by "scop"
490 * that should be unconditionally enforced even when live-range reordering
492 * "before" contains all pairs of statement iterations where
493 * the first is executed before the second according to the original schedule.
495 * In particular, compute the external false dependences.
496 * The anti-dependences are already taken care of by the order dependences.
497 * The external false dependences are only used to ensure that live-in and
498 * live-out data is not overwritten by any writes inside the scop.
500 * In particular, the reads from live-in data need to precede any
501 * later write to the same memory element.
502 * As to live-out data, the last writes need to remain the last writes.
503 * That is, any earlier write in the original schedule needs to precede
504 * the last write to the same memory element in the computed schedule.
505 * The possible last writes have been computed by compute_live_out.
506 * They may include kills, but if the last access is a kill,
507 * then the corresponding dependences will effectively be ignored
508 * since we do not schedule any kill statements.
510 * Note that the set of live-in and live-out accesses may be
511 * an overapproximation. There may therefore be potential writes
512 * before a live-in access and after a live-out access.
514 static void compute_forced_dependences(struct ppcg_scop
*ps
,
515 __isl_take isl_union_map
*before
)
517 isl_union_map
*shared_access
;
518 isl_union_map
*exposed
;
519 isl_union_map
*live_in
;
521 exposed
= isl_union_map_copy(ps
->live_out
);
523 exposed
= isl_union_map_reverse(exposed
);
524 shared_access
= isl_union_map_copy(ps
->may_writes
);
525 shared_access
= isl_union_map_apply_range(shared_access
, exposed
);
527 ps
->dep_forced
= shared_access
;
529 live_in
= isl_union_map_apply_range(isl_union_map_copy(ps
->live_in
),
530 isl_union_map_reverse(isl_union_map_copy(ps
->may_writes
)));
532 ps
->dep_forced
= isl_union_map_union(ps
->dep_forced
, live_in
);
533 ps
->dep_forced
= isl_union_map_intersect(ps
->dep_forced
, before
);
536 /* Compute the dependences of the program represented by "scop"
537 * in case live range reordering is allowed.
539 * We compute the actual live ranges and the corresponding order
542 static void compute_live_range_reordering_dependences(struct ppcg_scop
*ps
)
544 isl_union_map
*before
;
545 isl_union_map
*schedule
;
547 schedule
= isl_schedule_get_map(ps
->schedule
);
548 before
= isl_union_map_lex_lt_union_map(schedule
,
549 isl_union_map_copy(schedule
));
551 compute_tagged_flow_dep(ps
);
552 compute_order_dependences(ps
, isl_union_map_copy(before
));
553 compute_forced_dependences(ps
, before
);
556 /* Compute the potential flow dependences and the potential live in
559 static void compute_flow_dep(struct ppcg_scop
*ps
)
561 isl_union_access_info
*access
;
562 isl_union_flow
*flow
;
564 access
= isl_union_access_info_from_sink(isl_union_map_copy(ps
->reads
));
565 access
= isl_union_access_info_set_must_source(access
,
566 isl_union_map_copy(ps
->must_writes
));
567 access
= isl_union_access_info_set_may_source(access
,
568 isl_union_map_copy(ps
->may_writes
));
569 access
= isl_union_access_info_set_schedule(access
,
570 isl_schedule_copy(ps
->schedule
));
571 flow
= isl_union_access_info_compute_flow(access
);
573 ps
->dep_flow
= isl_union_flow_get_may_dependence(flow
);
574 ps
->live_in
= isl_union_flow_get_may_no_source(flow
);
575 isl_union_flow_free(flow
);
578 /* Compute the dependences of the program represented by "scop".
579 * Store the computed potential flow dependences
580 * in scop->dep_flow and the reads with potentially no corresponding writes in
582 * Store the potential live out accesses in scop->live_out.
583 * Store the potential false (anti and output) dependences in scop->dep_false.
585 * If live range reordering is allowed, then we compute a separate
586 * set of order dependences and a set of external false dependences
587 * in compute_live_range_reordering_dependences.
589 static void compute_dependences(struct ppcg_scop
*scop
)
591 isl_union_map
*may_source
;
592 isl_union_access_info
*access
;
593 isl_union_flow
*flow
;
598 compute_live_out(scop
);
600 if (scop
->options
->live_range_reordering
)
601 compute_live_range_reordering_dependences(scop
);
602 else if (scop
->options
->target
!= PPCG_TARGET_C
)
603 compute_tagged_flow_dep(scop
);
605 compute_flow_dep(scop
);
607 may_source
= isl_union_map_union(isl_union_map_copy(scop
->may_writes
),
608 isl_union_map_copy(scop
->reads
));
609 access
= isl_union_access_info_from_sink(
610 isl_union_map_copy(scop
->may_writes
));
611 access
= isl_union_access_info_set_must_source(access
,
612 isl_union_map_copy(scop
->must_writes
));
613 access
= isl_union_access_info_set_may_source(access
, may_source
);
614 access
= isl_union_access_info_set_schedule(access
,
615 isl_schedule_copy(scop
->schedule
));
616 flow
= isl_union_access_info_compute_flow(access
);
618 scop
->dep_false
= isl_union_flow_get_may_dependence(flow
);
619 scop
->dep_false
= isl_union_map_coalesce(scop
->dep_false
);
620 isl_union_flow_free(flow
);
623 /* Eliminate dead code from ps->domain.
625 * In particular, intersect both ps->domain and the domain of
626 * ps->schedule with the (parts of) iteration
627 * domains that are needed to produce the output or for statement
628 * iterations that call functions.
629 * Also intersect the range of the dataflow dependences with
630 * this domain such that the removed instances will no longer
631 * be considered as targets of dataflow.
633 * We start with the iteration domains that call functions
634 * and the set of iterations that last write to an array
635 * (except those that are later killed).
637 * Then we add those statement iterations that produce
638 * something needed by the "live" statements iterations.
639 * We keep doing this until no more statement iterations can be added.
640 * To ensure that the procedure terminates, we compute the affine
641 * hull of the live iterations (bounded to the original iteration
642 * domains) each time we have added extra iterations.
644 static void eliminate_dead_code(struct ppcg_scop
*ps
)
648 isl_union_pw_multi_aff
*tagger
;
650 live
= isl_union_map_domain(isl_union_map_copy(ps
->live_out
));
651 if (!isl_union_set_is_empty(ps
->call
)) {
652 live
= isl_union_set_union(live
, isl_union_set_copy(ps
->call
));
653 live
= isl_union_set_coalesce(live
);
656 dep
= isl_union_map_copy(ps
->dep_flow
);
657 dep
= isl_union_map_reverse(dep
);
660 isl_union_set
*extra
;
662 extra
= isl_union_set_apply(isl_union_set_copy(live
),
663 isl_union_map_copy(dep
));
664 if (isl_union_set_is_subset(extra
, live
)) {
665 isl_union_set_free(extra
);
669 live
= isl_union_set_union(live
, extra
);
670 live
= isl_union_set_affine_hull(live
);
671 live
= isl_union_set_intersect(live
,
672 isl_union_set_copy(ps
->domain
));
675 isl_union_map_free(dep
);
677 ps
->domain
= isl_union_set_intersect(ps
->domain
,
678 isl_union_set_copy(live
));
679 ps
->schedule
= isl_schedule_intersect_domain(ps
->schedule
,
680 isl_union_set_copy(live
));
681 ps
->dep_flow
= isl_union_map_intersect_range(ps
->dep_flow
,
682 isl_union_set_copy(live
));
683 tagger
= isl_union_pw_multi_aff_copy(ps
->tagger
);
684 live
= isl_union_set_preimage_union_pw_multi_aff(live
, tagger
);
685 ps
->tagged_dep_flow
= isl_union_map_intersect_range(ps
->tagged_dep_flow
,
689 /* Intersect "set" with the set described by "str", taking the NULL
690 * string to represent the universal set.
692 static __isl_give isl_set
*set_intersect_str(__isl_take isl_set
*set
,
701 ctx
= isl_set_get_ctx(set
);
702 set2
= isl_set_read_from_str(ctx
, str
);
703 set
= isl_set_intersect(set
, set2
);
708 static void *ppcg_scop_free(struct ppcg_scop
*ps
)
713 isl_set_free(ps
->context
);
714 isl_union_set_free(ps
->domain
);
715 isl_union_set_free(ps
->call
);
716 isl_union_map_free(ps
->tagged_reads
);
717 isl_union_map_free(ps
->reads
);
718 isl_union_map_free(ps
->live_in
);
719 isl_union_map_free(ps
->tagged_may_writes
);
720 isl_union_map_free(ps
->tagged_must_writes
);
721 isl_union_map_free(ps
->may_writes
);
722 isl_union_map_free(ps
->must_writes
);
723 isl_union_map_free(ps
->live_out
);
724 isl_union_map_free(ps
->tagged_must_kills
);
725 isl_union_map_free(ps
->tagged_dep_flow
);
726 isl_union_map_free(ps
->dep_flow
);
727 isl_union_map_free(ps
->dep_false
);
728 isl_union_map_free(ps
->dep_forced
);
729 isl_union_map_free(ps
->tagged_dep_order
);
730 isl_union_map_free(ps
->dep_order
);
731 isl_schedule_free(ps
->schedule
);
732 isl_union_pw_multi_aff_free(ps
->tagger
);
733 isl_union_map_free(ps
->independence
);
734 isl_id_to_ast_expr_free(ps
->names
);
741 /* Extract a ppcg_scop from a pet_scop.
743 * The constructed ppcg_scop refers to elements from the pet_scop
744 * so the pet_scop should not be freed before the ppcg_scop.
746 static struct ppcg_scop
*ppcg_scop_from_pet_scop(struct pet_scop
*scop
,
747 struct ppcg_options
*options
)
751 struct ppcg_scop
*ps
;
756 ctx
= isl_set_get_ctx(scop
->context
);
758 ps
= isl_calloc_type(ctx
, struct ppcg_scop
);
762 ps
->names
= collect_names(scop
);
763 ps
->options
= options
;
764 ps
->start
= pet_loc_get_start(scop
->loc
);
765 ps
->end
= pet_loc_get_end(scop
->loc
);
766 ps
->context
= isl_set_copy(scop
->context
);
767 ps
->context
= set_intersect_str(ps
->context
, options
->ctx
);
768 ps
->domain
= collect_non_kill_domains(scop
);
769 ps
->call
= collect_call_domains(scop
);
770 ps
->tagged_reads
= pet_scop_collect_tagged_may_reads(scop
);
771 ps
->reads
= pet_scop_collect_may_reads(scop
);
772 ps
->tagged_may_writes
= pet_scop_collect_tagged_may_writes(scop
);
773 ps
->may_writes
= pet_scop_collect_may_writes(scop
);
774 ps
->tagged_must_writes
= pet_scop_collect_tagged_must_writes(scop
);
775 ps
->must_writes
= pet_scop_collect_must_writes(scop
);
776 ps
->tagged_must_kills
= pet_scop_collect_tagged_must_kills(scop
);
777 ps
->schedule
= isl_schedule_copy(scop
->schedule
);
779 ps
->independence
= isl_union_map_empty(isl_set_get_space(ps
->context
));
780 for (i
= 0; i
< scop
->n_independence
; ++i
)
781 ps
->independence
= isl_union_map_union(ps
->independence
,
782 isl_union_map_copy(scop
->independences
[i
]->filter
));
785 compute_dependences(ps
);
786 eliminate_dead_code(ps
);
788 if (!ps
->context
|| !ps
->domain
|| !ps
->call
|| !ps
->reads
||
789 !ps
->may_writes
|| !ps
->must_writes
|| !ps
->tagged_must_kills
||
790 !ps
->schedule
|| !ps
->independence
|| !ps
->names
)
791 return ppcg_scop_free(ps
);
796 /* Internal data structure for ppcg_transform.
798 struct ppcg_transform_data
{
799 struct ppcg_options
*options
;
800 __isl_give isl_printer
*(*transform
)(__isl_take isl_printer
*p
,
801 struct ppcg_scop
*scop
, void *user
);
805 /* Should we print the original code?
806 * That is, does "scop" involve any data dependent conditions or
807 * nested expressions that cannot be handled by pet_stmt_build_ast_exprs?
809 static int print_original(struct pet_scop
*scop
, struct ppcg_options
*options
)
811 if (!pet_scop_can_build_ast_exprs(scop
)) {
812 if (options
->debug
->verbose
)
813 fprintf(stdout
, "Printing original code because "
814 "some index expressions cannot currently "
819 if (pet_scop_has_data_dependent_conditions(scop
)) {
820 if (options
->debug
->verbose
)
821 fprintf(stdout
, "Printing original code because "
822 "input involves data dependent conditions\n");
829 /* Callback for pet_transform_C_source that transforms
830 * the given pet_scop to a ppcg_scop before calling the
831 * ppcg_transform callback.
833 * If "scop" contains any data dependent conditions or if we may
834 * not be able to print the transformed program, then just print
837 static __isl_give isl_printer
*transform(__isl_take isl_printer
*p
,
838 struct pet_scop
*scop
, void *user
)
840 struct ppcg_transform_data
*data
= user
;
841 struct ppcg_scop
*ps
;
843 if (print_original(scop
, data
->options
)) {
844 p
= pet_scop_print_original(scop
, p
);
849 scop
= pet_scop_align_params(scop
);
850 ps
= ppcg_scop_from_pet_scop(scop
, data
->options
);
852 p
= data
->transform(p
, ps
, data
->user
);
860 /* Transform the C source file "input" by rewriting each scop
861 * through a call to "transform".
862 * The transformed C code is written to "out".
864 * This is a wrapper around pet_transform_C_source that transforms
865 * the pet_scop to a ppcg_scop before calling "fn".
867 int ppcg_transform(isl_ctx
*ctx
, const char *input
, FILE *out
,
868 struct ppcg_options
*options
,
869 __isl_give isl_printer
*(*fn
)(__isl_take isl_printer
*p
,
870 struct ppcg_scop
*scop
, void *user
), void *user
)
872 struct ppcg_transform_data data
= { options
, fn
, user
};
873 return pet_transform_C_source(ctx
, input
, out
, &transform
, &data
);
876 /* Check consistency of options.
878 * Return -1 on error.
880 static int check_options(isl_ctx
*ctx
)
882 struct options
*options
;
884 options
= isl_ctx_peek_options(ctx
, &options_args
);
886 isl_die(ctx
, isl_error_internal
,
887 "unable to find options", return -1);
889 if (options
->ppcg
->openmp
&&
890 !isl_options_get_ast_build_atomic_upper_bound(ctx
))
891 isl_die(ctx
, isl_error_invalid
,
892 "OpenMP requires atomic bounds", return -1);
897 int main(int argc
, char **argv
)
901 struct options
*options
;
903 options
= options_new_with_defaults();
906 ctx
= isl_ctx_alloc_with_options(&options_args
, options
);
907 isl_options_set_schedule_outer_coincidence(ctx
, 1);
908 isl_options_set_schedule_maximize_band_depth(ctx
, 1);
909 pet_options_set_encapsulate_dynamic_control(ctx
, 1);
910 argc
= options_parse(options
, argc
, argv
, ISL_ARG_ALL
);
912 if (check_options(ctx
) < 0)
914 else if (options
->ppcg
->target
== PPCG_TARGET_CUDA
)
915 r
= generate_cuda(ctx
, options
->ppcg
, options
->input
);
916 else if (options
->ppcg
->target
== PPCG_TARGET_OPENCL
)
917 r
= generate_opencl(ctx
, options
->ppcg
, options
->input
,
920 r
= generate_cpu(ctx
, options
->ppcg
, options
->input
,