2 * Copyright 2011 INRIA Saclay
3 * Copyright 2013 Ecole Normale Superieure
4 * Copyright 2015 Sven Verdoolaege
6 * Use of this software is governed by the MIT license
8 * Written by Sven Verdoolaege, INRIA Saclay - Ile-de-France,
9 * Parc Club Orsay Universite, ZAC des vignes, 4 rue Jacques Monod,
11 * and Ecole Normale Superieure, 45 rue d'Ulm, 75230 Paris, France
20 #include <isl/options.h>
21 #include <isl/schedule.h>
22 #include <isl/ast_build.h>
23 #include <isl/schedule.h>
26 #include "ppcg_options.h"
32 struct isl_options
*isl
;
33 struct pet_options
*pet
;
34 struct ppcg_options
*ppcg
;
39 const char *ppcg_version(void);
40 static void print_version(void)
42 printf("%s", ppcg_version());
45 ISL_ARGS_START(struct options
, options_args
)
46 ISL_ARG_CHILD(struct options
, isl
, "isl", &isl_options_args
, "isl options")
47 ISL_ARG_CHILD(struct options
, pet
, "pet", &pet_options_args
, "pet options")
48 ISL_ARG_CHILD(struct options
, ppcg
, NULL
, &ppcg_options_args
, "ppcg options")
49 ISL_ARG_STR(struct options
, output
, 'o', NULL
,
50 "filename", NULL
, "output filename (c and opencl targets)")
51 ISL_ARG_ARG(struct options
, input
, "input", NULL
)
52 ISL_ARG_VERSION(print_version
)
55 ISL_ARG_DEF(options
, struct options
, options_args
)
57 /* Return a pointer to the final path component of "filename" or
58 * to "filename" itself if it does not contain any components.
60 const char *ppcg_base_name(const char *filename
)
64 base
= strrchr(filename
, '/');
71 /* Copy the base name of "input" to "name" and return its length.
72 * "name" is not NULL terminated.
74 * In particular, remove all leading directory components and
75 * the final extension, if any.
77 int ppcg_extract_base_name(char *name
, const char *input
)
83 base
= ppcg_base_name(input
);
84 ext
= strrchr(base
, '.');
85 len
= ext
? ext
- base
: strlen(base
);
87 memcpy(name
, base
, len
);
92 /* Does "scop" refer to any arrays that are declared, but not
93 * exposed to the code after the scop?
95 int ppcg_scop_any_hidden_declarations(struct ppcg_scop
*scop
)
102 for (i
= 0; i
< scop
->pet
->n_array
; ++i
)
103 if (scop
->pet
->arrays
[i
]->declared
&&
104 !scop
->pet
->arrays
[i
]->exposed
)
110 /* Collect all variable names that are in use in "scop".
111 * In particular, collect all parameters in the context and
112 * all the array names.
113 * Store these names in an isl_id_to_ast_expr by mapping
114 * them to a dummy value (0).
116 static __isl_give isl_id_to_ast_expr
*collect_names(struct pet_scop
*scop
)
121 isl_id_to_ast_expr
*names
;
123 ctx
= isl_set_get_ctx(scop
->context
);
125 n
= isl_set_dim(scop
->context
, isl_dim_param
);
127 names
= isl_id_to_ast_expr_alloc(ctx
, n
+ scop
->n_array
);
128 zero
= isl_ast_expr_from_val(isl_val_zero(ctx
));
130 for (i
= 0; i
< n
; ++i
) {
133 id
= isl_set_get_dim_id(scop
->context
, isl_dim_param
, i
);
134 names
= isl_id_to_ast_expr_set(names
,
135 id
, isl_ast_expr_copy(zero
));
138 for (i
= 0; i
< scop
->n_array
; ++i
) {
139 struct pet_array
*array
= scop
->arrays
[i
];
142 id
= isl_set_get_tuple_id(array
->extent
);
143 names
= isl_id_to_ast_expr_set(names
,
144 id
, isl_ast_expr_copy(zero
));
147 isl_ast_expr_free(zero
);
152 /* Return an isl_id called "prefix%d", with "%d" set to "i".
153 * If an isl_id with such a name already appears among the variable names
154 * of "scop", then adjust the name to "prefix%d_%d".
156 static __isl_give isl_id
*generate_name(struct ppcg_scop
*scop
,
157 const char *prefix
, int i
)
165 ctx
= isl_set_get_ctx(scop
->context
);
166 snprintf(name
, sizeof(name
), "%s%d", prefix
, i
);
167 id
= isl_id_alloc(ctx
, name
, NULL
);
170 while ((has_name
= isl_id_to_ast_expr_has(scop
->names
, id
)) == 1) {
172 snprintf(name
, sizeof(name
), "%s%d_%d", prefix
, i
, j
++);
173 id
= isl_id_alloc(ctx
, name
, NULL
);
176 return has_name
< 0 ? isl_id_free(id
) : id
;
179 /* Return a list of "n" isl_ids of the form "prefix%d".
180 * If an isl_id with such a name already appears among the variable names
181 * of "scop", then adjust the name to "prefix%d_%d".
183 __isl_give isl_id_list
*ppcg_scop_generate_names(struct ppcg_scop
*scop
,
184 int n
, const char *prefix
)
191 ctx
= isl_set_get_ctx(scop
->context
);
192 names
= isl_id_list_alloc(ctx
, n
);
193 for (i
= 0; i
< n
; ++i
) {
196 id
= generate_name(scop
, prefix
, i
);
197 names
= isl_id_list_add(names
, id
);
203 /* Is "stmt" not a kill statement?
205 static int is_not_kill(struct pet_stmt
*stmt
)
207 return !pet_stmt_is_kill(stmt
);
210 /* Collect the iteration domains of the statements in "scop" that
213 static __isl_give isl_union_set
*collect_domains(struct pet_scop
*scop
,
214 int (*pred
)(struct pet_stmt
*stmt
))
218 isl_union_set
*domain
;
223 domain
= isl_union_set_empty(isl_set_get_space(scop
->context
));
225 for (i
= 0; i
< scop
->n_stmt
; ++i
) {
226 struct pet_stmt
*stmt
= scop
->stmts
[i
];
232 isl_die(isl_union_set_get_ctx(domain
),
233 isl_error_unsupported
,
234 "data dependent conditions not supported",
235 return isl_union_set_free(domain
));
237 domain_i
= isl_set_copy(scop
->stmts
[i
]->domain
);
238 domain
= isl_union_set_add_set(domain
, domain_i
);
244 /* Collect the iteration domains of the statements in "scop",
245 * skipping kill statements.
247 static __isl_give isl_union_set
*collect_non_kill_domains(struct pet_scop
*scop
)
249 return collect_domains(scop
, &is_not_kill
);
252 /* This function is used as a callback to pet_expr_foreach_call_expr
253 * to detect if there is any call expression in the input expression.
254 * Assign the value 1 to the integer that "user" points to and
255 * abort the search since we have found what we were looking for.
257 static int set_has_call(__isl_keep pet_expr
*expr
, void *user
)
259 int *has_call
= user
;
266 /* Does "expr" contain any call expressions?
268 static int expr_has_call(__isl_keep pet_expr
*expr
)
272 if (pet_expr_foreach_call_expr(expr
, &set_has_call
, &has_call
) < 0 &&
279 /* This function is a callback for pet_tree_foreach_expr.
280 * If "expr" contains any call (sub)expressions, then set *has_call
281 * and abort the search.
283 static int check_call(__isl_keep pet_expr
*expr
, void *user
)
285 int *has_call
= user
;
287 if (expr_has_call(expr
))
290 return *has_call
? -1 : 0;
293 /* Does "stmt" contain any call expressions?
295 static int has_call(struct pet_stmt
*stmt
)
299 if (pet_tree_foreach_expr(stmt
->body
, &check_call
, &has_call
) < 0 &&
306 /* Collect the iteration domains of the statements in "scop"
307 * that contain a call expression.
309 static __isl_give isl_union_set
*collect_call_domains(struct pet_scop
*scop
)
311 return collect_domains(scop
, &has_call
);
314 /* Given a union of "tagged" access relations of the form
316 * [S_i[...] -> R_j[]] -> A_k[...]
318 * project out the "tags" (R_j[]).
319 * That is, return a union of relations of the form
321 * S_i[...] -> A_k[...]
323 static __isl_give isl_union_map
*project_out_tags(
324 __isl_take isl_union_map
*umap
)
326 return isl_union_map_domain_factor_domain(umap
);
329 /* Construct a function from tagged iteration domains to the corresponding
330 * untagged iteration domains with as range of the wrapped map in the domain
331 * the reference tags that appear in any of the reads, writes or kills.
332 * Store the result in ps->tagger.
334 * For example, if the statement with iteration space S[i,j]
335 * contains two array references R_1[] and R_2[], then ps->tagger will contain
337 * { [S[i,j] -> R_1[]] -> S[i,j]; [S[i,j] -> R_2[]] -> S[i,j] }
339 static void compute_tagger(struct ppcg_scop
*ps
)
341 isl_union_map
*tagged
;
342 isl_union_pw_multi_aff
*tagger
;
344 tagged
= isl_union_map_copy(ps
->tagged_reads
);
345 tagged
= isl_union_map_union(tagged
,
346 isl_union_map_copy(ps
->tagged_may_writes
));
347 tagged
= isl_union_map_union(tagged
,
348 isl_union_map_copy(ps
->tagged_must_kills
));
349 tagged
= isl_union_map_universe(tagged
);
350 tagged
= isl_union_set_unwrap(isl_union_map_domain(tagged
));
352 tagger
= isl_union_map_domain_map_union_pw_multi_aff(tagged
);
357 /* Compute the live out accesses, i.e., the writes that are
358 * potentially not killed by any kills or any other writes, and
359 * store them in ps->live_out.
361 * We compute the "dependence" of any "kill" (an explicit kill
362 * or a must write) on any may write.
363 * The may writes with a "depending" kill are definitely killed.
364 * The remaining may writes can potentially be live out.
366 static void compute_live_out(struct ppcg_scop
*ps
)
368 isl_union_pw_multi_aff
*tagger
;
369 isl_schedule
*schedule
;
370 isl_union_map
*kills
;
371 isl_union_map
*exposed
;
372 isl_union_map
*covering
;
373 isl_union_access_info
*access
;
374 isl_union_flow
*flow
;
376 tagger
= isl_union_pw_multi_aff_copy(ps
->tagger
);
377 schedule
= isl_schedule_copy(ps
->schedule
);
378 schedule
= isl_schedule_pullback_union_pw_multi_aff(schedule
, tagger
);
379 kills
= isl_union_map_union(isl_union_map_copy(ps
->tagged_must_writes
),
380 isl_union_map_copy(ps
->tagged_must_kills
));
381 access
= isl_union_access_info_from_sink(kills
);
382 access
= isl_union_access_info_set_may_source(access
,
383 isl_union_map_copy(ps
->tagged_may_writes
));
384 access
= isl_union_access_info_set_schedule(access
, schedule
);
385 flow
= isl_union_access_info_compute_flow(access
);
386 covering
= isl_union_flow_get_may_dependence(flow
);
387 isl_union_flow_free(flow
);
388 exposed
= isl_union_map_copy(ps
->tagged_may_writes
);
389 exposed
= isl_union_map_subtract_domain(exposed
,
390 isl_union_map_domain(covering
));
391 ps
->live_out
= project_out_tags(exposed
);
394 /* Compute the flow dependences and the live_in accesses and store
395 * the results in ps->dep_flow and ps->live_in.
396 * A copy of the flow dependences, tagged with the reference tags
397 * is stored in ps->tagged_dep_flow.
399 * We first compute ps->tagged_dep_flow, i.e., the tagged flow dependences
400 * and then project out the tags.
402 * We allow both the must writes and the must kills to serve as
403 * definite sources such that a subsequent read would not depend
404 * on any earlier write. The resulting flow dependences with
405 * a must kill as source reflect possibly uninitialized reads.
406 * No dependences need to be introduced to protect such reads
407 * (other than those imposed by potential flows from may writes
408 * that follow the kill). We therefore those flow dependences.
409 * This is also useful for the dead code elimination, which assumes
410 * the flow sources are non-kill instances.
412 static void compute_tagged_flow_dep(struct ppcg_scop
*ps
)
414 isl_union_pw_multi_aff
*tagger
;
415 isl_schedule
*schedule
;
416 isl_union_map
*live_in
;
417 isl_union_access_info
*access
;
418 isl_union_flow
*flow
;
419 isl_union_map
*must_source
;
420 isl_union_map
*kills
;
421 isl_union_map
*tagged_flow
;
423 tagger
= isl_union_pw_multi_aff_copy(ps
->tagger
);
424 schedule
= isl_schedule_copy(ps
->schedule
);
425 schedule
= isl_schedule_pullback_union_pw_multi_aff(schedule
, tagger
);
426 kills
= isl_union_map_copy(ps
->tagged_must_kills
);
427 must_source
= isl_union_map_copy(ps
->tagged_must_writes
);
428 must_source
= isl_union_map_union(must_source
,
429 isl_union_map_copy(kills
));
430 access
= isl_union_access_info_from_sink(
431 isl_union_map_copy(ps
->tagged_reads
));
432 access
= isl_union_access_info_set_must_source(access
, must_source
);
433 access
= isl_union_access_info_set_may_source(access
,
434 isl_union_map_copy(ps
->tagged_may_writes
));
435 access
= isl_union_access_info_set_schedule(access
, schedule
);
436 flow
= isl_union_access_info_compute_flow(access
);
437 tagged_flow
= isl_union_flow_get_may_dependence(flow
);
438 tagged_flow
= isl_union_map_subtract_domain(tagged_flow
,
439 isl_union_map_domain(kills
));
440 ps
->tagged_dep_flow
= tagged_flow
;
441 ps
->dep_flow
= isl_union_map_copy(ps
->tagged_dep_flow
);
442 ps
->dep_flow
= isl_union_map_factor_domain(ps
->dep_flow
);
443 live_in
= isl_union_flow_get_may_no_source(flow
);
444 ps
->live_in
= project_out_tags(live_in
);
445 isl_union_flow_free(flow
);
448 /* Compute the order dependences that prevent the potential live ranges
450 * "before" contains all pairs of statement iterations where
451 * the first is executed before the second according to the original schedule.
453 * In particular, construct a union of relations
455 * [R[...] -> R_1[]] -> [W[...] -> R_2[]]
457 * where [R[...] -> R_1[]] is the range of one or more live ranges
458 * (i.e., a read) and [W[...] -> R_2[]] is the domain of one or more
459 * live ranges (i.e., a write). Moreover, the read and the write
460 * access the same memory element and the read occurs before the write
461 * in the original schedule.
462 * The scheduler allows some of these dependences to be violated, provided
463 * the adjacent live ranges are all local (i.e., their domain and range
464 * are mapped to the same point by the current schedule band).
466 * Note that if a live range is not local, then we need to make
467 * sure it does not overlap with _any_ other live range, and not
468 * just with the "previous" and/or the "next" live range.
469 * We therefore add order dependences between reads and
470 * _any_ later potential write.
472 * We also need to be careful about writes without a corresponding read.
473 * They are already prevented from moving past non-local preceding
474 * intervals, but we also need to prevent them from moving past non-local
475 * following intervals. We therefore also add order dependences from
476 * potential writes that do not appear in any intervals
477 * to all later potential writes.
478 * Note that dead code elimination should have removed most of these
479 * dead writes, but the dead code elimination may not remove all dead writes,
480 * so we need to consider them to be safe.
482 static void compute_order_dependences(struct ppcg_scop
*ps
,
483 __isl_take isl_union_map
*before
)
485 isl_union_map
*reads
;
486 isl_union_map
*shared_access
;
487 isl_union_set
*matched
;
488 isl_union_map
*unmatched
;
489 isl_union_set
*domain
;
491 reads
= isl_union_map_copy(ps
->tagged_reads
);
492 matched
= isl_union_map_domain(isl_union_map_copy(ps
->tagged_dep_flow
));
493 unmatched
= isl_union_map_copy(ps
->tagged_may_writes
);
494 unmatched
= isl_union_map_subtract_domain(unmatched
, matched
);
495 reads
= isl_union_map_union(reads
, unmatched
);
496 shared_access
= isl_union_map_copy(ps
->tagged_may_writes
);
497 shared_access
= isl_union_map_reverse(shared_access
);
498 shared_access
= isl_union_map_apply_range(reads
, shared_access
);
499 shared_access
= isl_union_map_zip(shared_access
);
500 shared_access
= isl_union_map_intersect_domain(shared_access
,
501 isl_union_map_wrap(before
));
502 domain
= isl_union_map_domain(isl_union_map_copy(shared_access
));
503 shared_access
= isl_union_map_zip(shared_access
);
504 ps
->dep_order
= isl_union_set_unwrap(domain
);
505 ps
->tagged_dep_order
= shared_access
;
508 /* Compute those validity dependences of the program represented by "scop"
509 * that should be unconditionally enforced even when live-range reordering
511 * "before" contains all pairs of statement iterations where
512 * the first is executed before the second according to the original schedule.
514 * In particular, compute the external false dependences
515 * as well as order dependences between sources with the same sink.
516 * The anti-dependences are already taken care of by the order dependences.
517 * The external false dependences are only used to ensure that live-in and
518 * live-out data is not overwritten by any writes inside the scop.
520 * In particular, the reads from live-in data need to precede any
521 * later write to the same memory element.
522 * As to live-out data, the last writes need to remain the last writes.
523 * That is, any earlier write in the original schedule needs to precede
524 * the last write to the same memory element in the computed schedule.
525 * The possible last writes have been computed by compute_live_out.
526 * They may include kills, but if the last access is a kill,
527 * then the corresponding dependences will effectively be ignored
528 * since we do not schedule any kill statements.
530 * Note that the set of live-in and live-out accesses may be
531 * an overapproximation. There may therefore be potential writes
532 * before a live-in access and after a live-out access.
534 * In the presence of may-writes, there may be multiple live-ranges
535 * with the same sink, accessing the same memory element.
536 * The sources of these live-ranges need to be executed
537 * in the same relative order as in the original program
538 * since we do not know which of the may-writes will actually
539 * perform a write. Consider all sources that share a sink and
540 * that may write to the same memory element and compute
541 * the order dependences among them.
543 static void compute_forced_dependences(struct ppcg_scop
*ps
,
544 __isl_take isl_union_map
*before
)
546 isl_union_map
*shared_access
;
547 isl_union_map
*exposed
;
548 isl_union_map
*live_in
;
549 isl_union_map
*sink_access
;
550 isl_union_map
*shared_sink
;
551 isl_union_access_info
*access
;
552 isl_union_flow
*flow
;
553 isl_schedule
*schedule
;
555 exposed
= isl_union_map_copy(ps
->live_out
);
557 exposed
= isl_union_map_reverse(exposed
);
558 shared_access
= isl_union_map_copy(ps
->may_writes
);
559 shared_access
= isl_union_map_apply_range(shared_access
, exposed
);
561 ps
->dep_forced
= shared_access
;
563 live_in
= isl_union_map_apply_range(isl_union_map_copy(ps
->live_in
),
564 isl_union_map_reverse(isl_union_map_copy(ps
->may_writes
)));
566 ps
->dep_forced
= isl_union_map_union(ps
->dep_forced
, live_in
);
567 ps
->dep_forced
= isl_union_map_intersect(ps
->dep_forced
, before
);
569 schedule
= isl_schedule_copy(ps
->schedule
);
570 sink_access
= isl_union_map_copy(ps
->tagged_dep_flow
);
571 sink_access
= isl_union_map_range_product(sink_access
,
572 isl_union_map_copy(ps
->tagged_may_writes
));
573 sink_access
= isl_union_map_domain_factor_domain(sink_access
);
574 access
= isl_union_access_info_from_sink(
575 isl_union_map_copy(sink_access
));
576 access
= isl_union_access_info_set_may_source(access
, sink_access
);
577 access
= isl_union_access_info_set_schedule(access
, schedule
);
578 flow
= isl_union_access_info_compute_flow(access
);
579 shared_sink
= isl_union_flow_get_may_dependence(flow
);
580 isl_union_flow_free(flow
);
581 ps
->dep_forced
= isl_union_map_union(ps
->dep_forced
, shared_sink
);
584 /* Compute the dependences of the program represented by "scop"
585 * in case live range reordering is allowed.
587 * We compute the actual live ranges and the corresponding order
590 static void compute_live_range_reordering_dependences(struct ppcg_scop
*ps
)
592 isl_union_map
*before
;
593 isl_union_map
*schedule
;
595 schedule
= isl_schedule_get_map(ps
->schedule
);
596 before
= isl_union_map_lex_lt_union_map(schedule
,
597 isl_union_map_copy(schedule
));
599 compute_tagged_flow_dep(ps
);
600 compute_order_dependences(ps
, isl_union_map_copy(before
));
601 compute_forced_dependences(ps
, before
);
604 /* Compute the potential flow dependences and the potential live in
607 static void compute_flow_dep(struct ppcg_scop
*ps
)
609 isl_union_access_info
*access
;
610 isl_union_flow
*flow
;
612 access
= isl_union_access_info_from_sink(isl_union_map_copy(ps
->reads
));
613 access
= isl_union_access_info_set_must_source(access
,
614 isl_union_map_copy(ps
->must_writes
));
615 access
= isl_union_access_info_set_may_source(access
,
616 isl_union_map_copy(ps
->may_writes
));
617 access
= isl_union_access_info_set_schedule(access
,
618 isl_schedule_copy(ps
->schedule
));
619 flow
= isl_union_access_info_compute_flow(access
);
621 ps
->dep_flow
= isl_union_flow_get_may_dependence(flow
);
622 ps
->live_in
= isl_union_flow_get_may_no_source(flow
);
623 isl_union_flow_free(flow
);
626 /* Compute the dependences of the program represented by "scop".
627 * Store the computed potential flow dependences
628 * in scop->dep_flow and the reads with potentially no corresponding writes in
630 * Store the potential live out accesses in scop->live_out.
631 * Store the potential false (anti and output) dependences in scop->dep_false.
633 * If live range reordering is allowed, then we compute a separate
634 * set of order dependences and a set of external false dependences
635 * in compute_live_range_reordering_dependences.
637 static void compute_dependences(struct ppcg_scop
*scop
)
639 isl_union_map
*may_source
;
640 isl_union_access_info
*access
;
641 isl_union_flow
*flow
;
646 compute_live_out(scop
);
648 if (scop
->options
->live_range_reordering
)
649 compute_live_range_reordering_dependences(scop
);
650 else if (scop
->options
->target
!= PPCG_TARGET_C
)
651 compute_tagged_flow_dep(scop
);
653 compute_flow_dep(scop
);
655 may_source
= isl_union_map_union(isl_union_map_copy(scop
->may_writes
),
656 isl_union_map_copy(scop
->reads
));
657 access
= isl_union_access_info_from_sink(
658 isl_union_map_copy(scop
->may_writes
));
659 access
= isl_union_access_info_set_must_source(access
,
660 isl_union_map_copy(scop
->must_writes
));
661 access
= isl_union_access_info_set_may_source(access
, may_source
);
662 access
= isl_union_access_info_set_schedule(access
,
663 isl_schedule_copy(scop
->schedule
));
664 flow
= isl_union_access_info_compute_flow(access
);
666 scop
->dep_false
= isl_union_flow_get_may_dependence(flow
);
667 scop
->dep_false
= isl_union_map_coalesce(scop
->dep_false
);
668 isl_union_flow_free(flow
);
671 /* Eliminate dead code from ps->domain.
673 * In particular, intersect both ps->domain and the domain of
674 * ps->schedule with the (parts of) iteration
675 * domains that are needed to produce the output or for statement
676 * iterations that call functions.
677 * Also intersect the range of the dataflow dependences with
678 * this domain such that the removed instances will no longer
679 * be considered as targets of dataflow.
681 * We start with the iteration domains that call functions
682 * and the set of iterations that last write to an array
683 * (except those that are later killed).
685 * Then we add those statement iterations that produce
686 * something needed by the "live" statements iterations.
687 * We keep doing this until no more statement iterations can be added.
688 * To ensure that the procedure terminates, we compute the affine
689 * hull of the live iterations (bounded to the original iteration
690 * domains) each time we have added extra iterations.
692 static void eliminate_dead_code(struct ppcg_scop
*ps
)
696 isl_union_pw_multi_aff
*tagger
;
698 live
= isl_union_map_domain(isl_union_map_copy(ps
->live_out
));
699 if (!isl_union_set_is_empty(ps
->call
)) {
700 live
= isl_union_set_union(live
, isl_union_set_copy(ps
->call
));
701 live
= isl_union_set_coalesce(live
);
704 dep
= isl_union_map_copy(ps
->dep_flow
);
705 dep
= isl_union_map_reverse(dep
);
708 isl_union_set
*extra
;
710 extra
= isl_union_set_apply(isl_union_set_copy(live
),
711 isl_union_map_copy(dep
));
712 if (isl_union_set_is_subset(extra
, live
)) {
713 isl_union_set_free(extra
);
717 live
= isl_union_set_union(live
, extra
);
718 live
= isl_union_set_affine_hull(live
);
719 live
= isl_union_set_intersect(live
,
720 isl_union_set_copy(ps
->domain
));
723 isl_union_map_free(dep
);
725 ps
->domain
= isl_union_set_intersect(ps
->domain
,
726 isl_union_set_copy(live
));
727 ps
->schedule
= isl_schedule_intersect_domain(ps
->schedule
,
728 isl_union_set_copy(live
));
729 ps
->dep_flow
= isl_union_map_intersect_range(ps
->dep_flow
,
730 isl_union_set_copy(live
));
731 tagger
= isl_union_pw_multi_aff_copy(ps
->tagger
);
732 live
= isl_union_set_preimage_union_pw_multi_aff(live
, tagger
);
733 ps
->tagged_dep_flow
= isl_union_map_intersect_range(ps
->tagged_dep_flow
,
737 /* Intersect "set" with the set described by "str", taking the NULL
738 * string to represent the universal set.
740 static __isl_give isl_set
*set_intersect_str(__isl_take isl_set
*set
,
749 ctx
= isl_set_get_ctx(set
);
750 set2
= isl_set_read_from_str(ctx
, str
);
751 set
= isl_set_intersect(set
, set2
);
756 static void *ppcg_scop_free(struct ppcg_scop
*ps
)
761 isl_set_free(ps
->context
);
762 isl_union_set_free(ps
->domain
);
763 isl_union_set_free(ps
->call
);
764 isl_union_map_free(ps
->tagged_reads
);
765 isl_union_map_free(ps
->reads
);
766 isl_union_map_free(ps
->live_in
);
767 isl_union_map_free(ps
->tagged_may_writes
);
768 isl_union_map_free(ps
->tagged_must_writes
);
769 isl_union_map_free(ps
->may_writes
);
770 isl_union_map_free(ps
->must_writes
);
771 isl_union_map_free(ps
->live_out
);
772 isl_union_map_free(ps
->tagged_must_kills
);
773 isl_union_map_free(ps
->tagged_dep_flow
);
774 isl_union_map_free(ps
->dep_flow
);
775 isl_union_map_free(ps
->dep_false
);
776 isl_union_map_free(ps
->dep_forced
);
777 isl_union_map_free(ps
->tagged_dep_order
);
778 isl_union_map_free(ps
->dep_order
);
779 isl_schedule_free(ps
->schedule
);
780 isl_union_pw_multi_aff_free(ps
->tagger
);
781 isl_union_map_free(ps
->independence
);
782 isl_id_to_ast_expr_free(ps
->names
);
789 /* Extract a ppcg_scop from a pet_scop.
791 * The constructed ppcg_scop refers to elements from the pet_scop
792 * so the pet_scop should not be freed before the ppcg_scop.
794 static struct ppcg_scop
*ppcg_scop_from_pet_scop(struct pet_scop
*scop
,
795 struct ppcg_options
*options
)
799 struct ppcg_scop
*ps
;
804 ctx
= isl_set_get_ctx(scop
->context
);
806 ps
= isl_calloc_type(ctx
, struct ppcg_scop
);
810 ps
->names
= collect_names(scop
);
811 ps
->options
= options
;
812 ps
->start
= pet_loc_get_start(scop
->loc
);
813 ps
->end
= pet_loc_get_end(scop
->loc
);
814 ps
->context
= isl_set_copy(scop
->context
);
815 ps
->context
= set_intersect_str(ps
->context
, options
->ctx
);
816 if (options
->non_negative_parameters
) {
817 isl_space
*space
= isl_set_get_space(ps
->context
);
818 isl_set
*nn
= isl_set_nat_universe(space
);
819 ps
->context
= isl_set_intersect(ps
->context
, nn
);
821 ps
->domain
= collect_non_kill_domains(scop
);
822 ps
->call
= collect_call_domains(scop
);
823 ps
->tagged_reads
= pet_scop_collect_tagged_may_reads(scop
);
824 ps
->reads
= pet_scop_collect_may_reads(scop
);
825 ps
->tagged_may_writes
= pet_scop_collect_tagged_may_writes(scop
);
826 ps
->may_writes
= pet_scop_collect_may_writes(scop
);
827 ps
->tagged_must_writes
= pet_scop_collect_tagged_must_writes(scop
);
828 ps
->must_writes
= pet_scop_collect_must_writes(scop
);
829 ps
->tagged_must_kills
= pet_scop_collect_tagged_must_kills(scop
);
830 ps
->schedule
= isl_schedule_copy(scop
->schedule
);
832 ps
->independence
= isl_union_map_empty(isl_set_get_space(ps
->context
));
833 for (i
= 0; i
< scop
->n_independence
; ++i
)
834 ps
->independence
= isl_union_map_union(ps
->independence
,
835 isl_union_map_copy(scop
->independences
[i
]->filter
));
838 compute_dependences(ps
);
839 eliminate_dead_code(ps
);
841 if (!ps
->context
|| !ps
->domain
|| !ps
->call
|| !ps
->reads
||
842 !ps
->may_writes
|| !ps
->must_writes
|| !ps
->tagged_must_kills
||
843 !ps
->schedule
|| !ps
->independence
|| !ps
->names
)
844 return ppcg_scop_free(ps
);
849 /* Internal data structure for ppcg_transform.
851 struct ppcg_transform_data
{
852 struct ppcg_options
*options
;
853 __isl_give isl_printer
*(*transform
)(__isl_take isl_printer
*p
,
854 struct ppcg_scop
*scop
, void *user
);
858 /* Should we print the original code?
859 * That is, does "scop" involve any data dependent conditions or
860 * nested expressions that cannot be handled by pet_stmt_build_ast_exprs?
862 static int print_original(struct pet_scop
*scop
, struct ppcg_options
*options
)
864 if (!pet_scop_can_build_ast_exprs(scop
)) {
865 if (options
->debug
->verbose
)
866 fprintf(stdout
, "Printing original code because "
867 "some index expressions cannot currently "
872 if (pet_scop_has_data_dependent_conditions(scop
)) {
873 if (options
->debug
->verbose
)
874 fprintf(stdout
, "Printing original code because "
875 "input involves data dependent conditions\n");
882 /* Callback for pet_transform_C_source that transforms
883 * the given pet_scop to a ppcg_scop before calling the
884 * ppcg_transform callback.
886 * If "scop" contains any data dependent conditions or if we may
887 * not be able to print the transformed program, then just print
890 static __isl_give isl_printer
*transform(__isl_take isl_printer
*p
,
891 struct pet_scop
*scop
, void *user
)
893 struct ppcg_transform_data
*data
= user
;
894 struct ppcg_scop
*ps
;
896 if (print_original(scop
, data
->options
)) {
897 p
= pet_scop_print_original(scop
, p
);
902 scop
= pet_scop_align_params(scop
);
903 ps
= ppcg_scop_from_pet_scop(scop
, data
->options
);
905 p
= data
->transform(p
, ps
, data
->user
);
913 /* Transform the C source file "input" by rewriting each scop
914 * through a call to "transform".
915 * The transformed C code is written to "out".
917 * This is a wrapper around pet_transform_C_source that transforms
918 * the pet_scop to a ppcg_scop before calling "fn".
920 int ppcg_transform(isl_ctx
*ctx
, const char *input
, FILE *out
,
921 struct ppcg_options
*options
,
922 __isl_give isl_printer
*(*fn
)(__isl_take isl_printer
*p
,
923 struct ppcg_scop
*scop
, void *user
), void *user
)
925 struct ppcg_transform_data data
= { options
, fn
, user
};
926 return pet_transform_C_source(ctx
, input
, out
, &transform
, &data
);
929 /* Check consistency of options.
931 * Return -1 on error.
933 static int check_options(isl_ctx
*ctx
)
935 struct options
*options
;
937 options
= isl_ctx_peek_options(ctx
, &options_args
);
939 isl_die(ctx
, isl_error_internal
,
940 "unable to find options", return -1);
942 if (options
->ppcg
->openmp
&&
943 !isl_options_get_ast_build_atomic_upper_bound(ctx
))
944 isl_die(ctx
, isl_error_invalid
,
945 "OpenMP requires atomic bounds", return -1);
950 int main(int argc
, char **argv
)
954 struct options
*options
;
956 options
= options_new_with_defaults();
959 ctx
= isl_ctx_alloc_with_options(&options_args
, options
);
960 isl_options_set_schedule_outer_coincidence(ctx
, 1);
961 isl_options_set_schedule_maximize_band_depth(ctx
, 1);
962 pet_options_set_encapsulate_dynamic_control(ctx
, 1);
963 argc
= options_parse(options
, argc
, argv
, ISL_ARG_ALL
);
965 if (check_options(ctx
) < 0)
967 else if (options
->ppcg
->target
== PPCG_TARGET_CUDA
)
968 r
= generate_cuda(ctx
, options
->ppcg
, options
->input
);
969 else if (options
->ppcg
->target
== PPCG_TARGET_OPENCL
)
970 r
= generate_opencl(ctx
, options
->ppcg
, options
->input
,
973 r
= generate_cpu(ctx
, options
->ppcg
, options
->input
,