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
)
309 proj
= isl_union_map_universe(isl_union_map_copy(umap
));
310 proj
= isl_union_set_unwrap(isl_union_map_domain(proj
));
311 proj
= isl_union_map_domain_map(proj
);
313 umap
= isl_union_map_apply_domain(umap
, proj
);
318 /* Construct a relation from the iteration domains to tagged iteration
319 * domains with as range the reference tags that appear
320 * in any of the reads, writes or kills.
321 * Store the result in ps->tagger.
323 * For example, if the statement with iteration space S[i,j]
324 * contains two array references R_1[] and R_2[], then ps->tagger will contain
326 * { S[i,j] -> [S[i,j] -> R_1[]]; S[i,j] -> [S[i,j] -> R_2[]] }
328 static void compute_tagger(struct ppcg_scop
*ps
)
330 isl_union_map
*tagged
, *tagger
;
332 tagged
= isl_union_map_copy(ps
->tagged_reads
);
333 tagged
= isl_union_map_union(tagged
,
334 isl_union_map_copy(ps
->tagged_may_writes
));
335 tagged
= isl_union_map_union(tagged
,
336 isl_union_map_copy(ps
->tagged_must_kills
));
338 tagger
= isl_union_map_universe(tagged
);
339 tagger
= isl_union_set_unwrap(isl_union_map_domain(tagger
));
340 tagger
= isl_union_map_reverse(isl_union_map_domain_map(tagger
));
345 /* Compute the live out accesses, i.e., the writes that are
346 * potentially not killed by any kills or any other writes, and
347 * store them in ps->live_out.
349 * We compute the "dependence" of any "kill" (an explicit kill
350 * or a must write) on any may write.
351 * The may writes with a "depending" kill are definitely killed.
352 * The remaining may writes can potentially be live out.
354 static void compute_live_out(struct ppcg_scop
*ps
)
356 isl_union_map
*tagger
;
357 isl_union_map
*schedule
;
358 isl_union_map
*empty
;
359 isl_union_map
*kills
;
360 isl_union_map
*exposed
;
361 isl_union_map
*covering
;
363 tagger
= isl_union_map_copy(ps
->tagger
);
364 schedule
= isl_union_map_copy(ps
->schedule
);
365 schedule
= isl_union_map_apply_domain(schedule
,
366 isl_union_map_copy(tagger
));
367 empty
= isl_union_map_empty(isl_union_set_get_space(ps
->domain
));
368 kills
= isl_union_map_union(isl_union_map_copy(ps
->tagged_must_writes
),
369 isl_union_map_copy(ps
->tagged_must_kills
));
370 isl_union_map_compute_flow(kills
, empty
,
371 isl_union_map_copy(ps
->tagged_may_writes
),
372 schedule
, NULL
, &covering
, NULL
, NULL
);
373 exposed
= isl_union_map_copy(ps
->tagged_may_writes
);
374 exposed
= isl_union_map_subtract_domain(exposed
,
375 isl_union_map_domain(covering
));
376 exposed
= isl_union_map_apply_range(tagger
, exposed
);
377 ps
->live_out
= exposed
;
380 /* Compute the flow dependences and the live_in accesses and store
381 * the results in ps->dep_flow and ps->live_in.
382 * A copy of the flow dependences, tagged with the reference tags
383 * is stored in ps->tagged_dep_flow.
385 * We first compute ps->tagged_dep_flow, i.e., the tagged flow dependences
386 * and then project out the tags.
388 * We allow both the must writes and the must kills to serve as
389 * definite sources such that a subsequent read would not depend
390 * on any earlier write. The resulting flow dependences with
391 * a must kill as source reflect possibly uninitialized reads.
392 * No dependences need to be introduced to protect such reads
393 * (other than those imposed by potential flows from may writes
394 * that follow the kill). We therefore those flow dependences.
395 * This is also useful for the dead code elimination, which assumes
396 * the flow sources are non-kill instances.
398 static void compute_tagged_flow_dep(struct ppcg_scop
*ps
)
400 isl_union_map
*tagger
;
401 isl_union_map
*schedule
;
402 isl_union_map
*may_flow
;
403 isl_union_map
*live_in
, *may_live_in
;
404 isl_union_map
*must_source
;
405 isl_union_map
*kills
;
406 isl_union_map
*tagged_flow
;
408 tagger
= isl_union_map_copy(ps
->tagger
);
409 schedule
= isl_union_map_copy(ps
->schedule
);
410 schedule
= isl_union_map_apply_domain(schedule
, tagger
);
411 kills
= isl_union_map_copy(ps
->tagged_must_kills
);
412 must_source
= isl_union_map_copy(ps
->tagged_must_writes
);
413 must_source
= isl_union_map_union(must_source
,
414 isl_union_map_copy(kills
));
415 isl_union_map_compute_flow(isl_union_map_copy(ps
->tagged_reads
),
417 isl_union_map_copy(ps
->tagged_may_writes
),
418 schedule
, &ps
->tagged_dep_flow
, &may_flow
,
419 &live_in
, &may_live_in
);
420 tagged_flow
= isl_union_map_union(ps
->tagged_dep_flow
, may_flow
);
421 tagged_flow
= isl_union_map_subtract_domain(tagged_flow
,
422 isl_union_map_domain(kills
));
423 ps
->tagged_dep_flow
= tagged_flow
;
424 ps
->dep_flow
= isl_union_map_copy(ps
->tagged_dep_flow
);
425 ps
->dep_flow
= isl_union_map_factor_domain(ps
->dep_flow
);
426 live_in
= isl_union_map_union(live_in
, may_live_in
);
427 ps
->live_in
= project_out_tags(live_in
);
430 /* Compute the order dependences that prevent the potential live ranges
432 * "before" contains all pairs of statement iterations where
433 * the first is executed before the second according to the original schedule.
435 * In particular, construct a union of relations
437 * [R[...] -> R_1[]] -> [W[...] -> R_2[]]
439 * where [R[...] -> R_1[]] is the range of one or more live ranges
440 * (i.e., a read) and [W[...] -> R_2[]] is the domain of one or more
441 * live ranges (i.e., a write). Moreover, the read and the write
442 * access the same memory element and the read occurs before the write
443 * in the original schedule.
444 * The scheduler allows some of these dependences to be violated, provided
445 * the adjacent live ranges are all local (i.e., their domain and range
446 * are mapped to the same point by the current schedule band).
448 * Note that if a live range is not local, then we need to make
449 * sure it does not overlap with _any_ other live range, and not
450 * just with the "previous" and/or the "next" live range.
451 * We therefore add order dependences between reads and
452 * _any_ later potential write.
454 * We also need to be careful about writes without a corresponding read.
455 * They are already prevented from moving past non-local preceding
456 * intervals, but we also need to prevent them from moving past non-local
457 * following intervals. We therefore also add order dependences from
458 * potential writes that do not appear in any intervals
459 * to all later potential writes.
460 * Note that dead code elimination should have removed most of these
461 * dead writes, but the dead code elimination may not remove all dead writes,
462 * so we need to consider them to be safe.
464 static void compute_order_dependences(struct ppcg_scop
*ps
,
465 __isl_take isl_union_map
*before
)
467 isl_union_map
*reads
;
468 isl_union_map
*shared_access
;
469 isl_union_set
*matched
;
470 isl_union_map
*unmatched
;
471 isl_union_set
*domain
;
473 reads
= isl_union_map_copy(ps
->tagged_reads
);
474 matched
= isl_union_map_domain(isl_union_map_copy(ps
->tagged_dep_flow
));
475 unmatched
= isl_union_map_copy(ps
->tagged_may_writes
);
476 unmatched
= isl_union_map_subtract_domain(unmatched
, matched
);
477 reads
= isl_union_map_union(reads
, unmatched
);
478 shared_access
= isl_union_map_copy(ps
->tagged_may_writes
);
479 shared_access
= isl_union_map_reverse(shared_access
);
480 shared_access
= isl_union_map_apply_range(reads
, shared_access
);
481 shared_access
= isl_union_map_zip(shared_access
);
482 shared_access
= isl_union_map_intersect_domain(shared_access
,
483 isl_union_map_wrap(before
));
484 domain
= isl_union_map_domain(isl_union_map_copy(shared_access
));
485 shared_access
= isl_union_map_zip(shared_access
);
486 ps
->dep_order
= isl_union_set_unwrap(domain
);
487 ps
->tagged_dep_order
= shared_access
;
490 /* Compute the external false dependences of the program represented by "scop"
491 * in case live range reordering is allowed.
492 * "before" contains all pairs of statement iterations where
493 * the first is executed before the second according to the original schedule.
495 * The anti-dependences are already taken care of by the order dependences.
496 * The external false dependences are only used to ensure that live-in and
497 * live-out data is not overwritten by any writes inside the scop.
499 * In particular, the reads from live-in data need to precede any
500 * later write to the same memory element.
501 * As to live-out data, the last writes need to remain the last writes.
502 * That is, any earlier write in the original schedule needs to precede
503 * the last write to the same memory element in the computed schedule.
504 * The possible last writes have been computed by compute_live_out.
505 * They may include kills, but if the last access is a kill,
506 * then the corresponding dependences will effectively be ignored
507 * since we do not schedule any kill statements.
509 * Note that the set of live-in and live-out accesses may be
510 * an overapproximation. There may therefore be potential writes
511 * before a live-in access and after a live-out access.
513 static void compute_external_false_dependences(struct ppcg_scop
*ps
,
514 __isl_take isl_union_map
*before
)
516 isl_union_map
*shared_access
;
517 isl_union_map
*exposed
;
518 isl_union_map
*live_in
;
520 exposed
= isl_union_map_copy(ps
->live_out
);
522 exposed
= isl_union_map_reverse(exposed
);
523 shared_access
= isl_union_map_copy(ps
->may_writes
);
524 shared_access
= isl_union_map_apply_range(shared_access
, exposed
);
526 ps
->dep_external
= shared_access
;
528 live_in
= isl_union_map_apply_range(isl_union_map_copy(ps
->live_in
),
529 isl_union_map_reverse(isl_union_map_copy(ps
->may_writes
)));
531 ps
->dep_external
= isl_union_map_union(ps
->dep_external
, live_in
);
532 ps
->dep_external
= isl_union_map_intersect(ps
->dep_external
, before
);
535 /* Compute the dependences of the program represented by "scop"
536 * in case live range reordering is allowed.
538 * We compute the actual live ranges and the corresponding order
541 static void compute_live_range_reordering_dependences(struct ppcg_scop
*ps
)
543 isl_union_map
*before
;
545 before
= isl_union_map_lex_lt_union_map(
546 isl_union_map_copy(ps
->schedule
),
547 isl_union_map_copy(ps
->schedule
));
549 compute_tagged_flow_dep(ps
);
550 compute_order_dependences(ps
, isl_union_map_copy(before
));
551 compute_external_false_dependences(ps
, before
);
554 /* Compute the potential flow dependences and the potential live in
557 static void compute_flow_dep(struct ppcg_scop
*ps
)
559 isl_union_map
*may_flow
;
560 isl_union_map
*may_live_in
;
562 isl_union_map_compute_flow(isl_union_map_copy(ps
->reads
),
563 isl_union_map_copy(ps
->must_writes
),
564 isl_union_map_copy(ps
->may_writes
),
565 isl_union_map_copy(ps
->schedule
),
566 &ps
->dep_flow
, &may_flow
,
567 &ps
->live_in
, &may_live_in
);
569 ps
->dep_flow
= isl_union_map_union(ps
->dep_flow
, may_flow
);
570 ps
->live_in
= isl_union_map_union(ps
->live_in
, may_live_in
);
573 /* Compute the dependences of the program represented by "scop".
574 * Store the computed potential flow dependences
575 * in scop->dep_flow and the reads with potentially no corresponding writes in
577 * Store the potential live out accesses in scop->live_out.
578 * Store the potential false (anti and output) dependences in scop->dep_false.
580 * If live range reordering is allowed, then we compute a separate
581 * set of order dependences and a set of external false dependences
582 * in compute_live_range_reordering_dependences.
584 static void compute_dependences(struct ppcg_scop
*scop
)
586 isl_union_map
*dep1
, *dep2
;
587 isl_union_map
*may_source
;
592 compute_live_out(scop
);
594 if (scop
->options
->live_range_reordering
)
595 compute_live_range_reordering_dependences(scop
);
596 else if (scop
->options
->target
!= PPCG_TARGET_C
)
597 compute_tagged_flow_dep(scop
);
599 compute_flow_dep(scop
);
601 may_source
= isl_union_map_union(isl_union_map_copy(scop
->may_writes
),
602 isl_union_map_copy(scop
->reads
));
603 isl_union_map_compute_flow(isl_union_map_copy(scop
->may_writes
),
604 isl_union_map_copy(scop
->must_writes
),
605 may_source
, isl_union_map_copy(scop
->schedule
),
606 &dep1
, &dep2
, NULL
, NULL
);
608 scop
->dep_false
= isl_union_map_union(dep1
, dep2
);
609 scop
->dep_false
= isl_union_map_coalesce(scop
->dep_false
);
612 /* Eliminate dead code from ps->domain.
614 * In particular, intersect ps->domain with the (parts of) iteration
615 * domains that are needed to produce the output or for statement
616 * iterations that call functions.
617 * Also intersect the range of the dataflow dependences with
618 * this domain such that the removed instances will no longer
619 * be considered as targets of dataflow.
621 * We start with the iteration domains that call functions
622 * and the set of iterations that last write to an array
623 * (except those that are later killed).
625 * Then we add those statement iterations that produce
626 * something needed by the "live" statements iterations.
627 * We keep doing this until no more statement iterations can be added.
628 * To ensure that the procedure terminates, we compute the affine
629 * hull of the live iterations (bounded to the original iteration
630 * domains) each time we have added extra iterations.
632 static void eliminate_dead_code(struct ppcg_scop
*ps
)
636 isl_union_map
*tagger
;
638 live
= isl_union_map_domain(isl_union_map_copy(ps
->live_out
));
639 if (!isl_union_set_is_empty(ps
->call
)) {
640 live
= isl_union_set_union(live
, isl_union_set_copy(ps
->call
));
641 live
= isl_union_set_coalesce(live
);
644 dep
= isl_union_map_copy(ps
->dep_flow
);
645 dep
= isl_union_map_reverse(dep
);
648 isl_union_set
*extra
;
650 extra
= isl_union_set_apply(isl_union_set_copy(live
),
651 isl_union_map_copy(dep
));
652 if (isl_union_set_is_subset(extra
, live
)) {
653 isl_union_set_free(extra
);
657 live
= isl_union_set_union(live
, extra
);
658 live
= isl_union_set_affine_hull(live
);
659 live
= isl_union_set_intersect(live
,
660 isl_union_set_copy(ps
->domain
));
663 isl_union_map_free(dep
);
665 ps
->domain
= isl_union_set_intersect(ps
->domain
,
666 isl_union_set_copy(live
));
667 ps
->dep_flow
= isl_union_map_intersect_range(ps
->dep_flow
,
668 isl_union_set_copy(live
));
669 tagger
= isl_union_map_copy(ps
->tagger
);
670 live
= isl_union_set_apply(live
, tagger
);
671 ps
->tagged_dep_flow
= isl_union_map_intersect_range(ps
->tagged_dep_flow
,
675 /* Intersect "set" with the set described by "str", taking the NULL
676 * string to represent the universal set.
678 static __isl_give isl_set
*set_intersect_str(__isl_take isl_set
*set
,
687 ctx
= isl_set_get_ctx(set
);
688 set2
= isl_set_read_from_str(ctx
, str
);
689 set
= isl_set_intersect(set
, set2
);
694 static void *ppcg_scop_free(struct ppcg_scop
*ps
)
699 isl_set_free(ps
->context
);
700 isl_union_set_free(ps
->domain
);
701 isl_union_set_free(ps
->call
);
702 isl_union_map_free(ps
->tagged_reads
);
703 isl_union_map_free(ps
->reads
);
704 isl_union_map_free(ps
->live_in
);
705 isl_union_map_free(ps
->tagged_may_writes
);
706 isl_union_map_free(ps
->tagged_must_writes
);
707 isl_union_map_free(ps
->may_writes
);
708 isl_union_map_free(ps
->must_writes
);
709 isl_union_map_free(ps
->live_out
);
710 isl_union_map_free(ps
->tagged_must_kills
);
711 isl_union_map_free(ps
->tagged_dep_flow
);
712 isl_union_map_free(ps
->dep_flow
);
713 isl_union_map_free(ps
->dep_false
);
714 isl_union_map_free(ps
->dep_external
);
715 isl_union_map_free(ps
->tagged_dep_order
);
716 isl_union_map_free(ps
->dep_order
);
717 isl_union_map_free(ps
->schedule
);
718 isl_union_map_free(ps
->tagger
);
719 isl_union_map_free(ps
->independence
);
720 isl_id_to_ast_expr_free(ps
->names
);
727 /* Extract a ppcg_scop from a pet_scop.
729 * The constructed ppcg_scop refers to elements from the pet_scop
730 * so the pet_scop should not be freed before the ppcg_scop.
732 static struct ppcg_scop
*ppcg_scop_from_pet_scop(struct pet_scop
*scop
,
733 struct ppcg_options
*options
)
737 struct ppcg_scop
*ps
;
742 ctx
= isl_set_get_ctx(scop
->context
);
744 ps
= isl_calloc_type(ctx
, struct ppcg_scop
);
748 ps
->names
= collect_names(scop
);
749 ps
->options
= options
;
750 ps
->start
= pet_loc_get_start(scop
->loc
);
751 ps
->end
= pet_loc_get_end(scop
->loc
);
752 ps
->context
= isl_set_copy(scop
->context
);
753 ps
->context
= set_intersect_str(ps
->context
, options
->ctx
);
754 ps
->domain
= collect_non_kill_domains(scop
);
755 ps
->call
= collect_call_domains(scop
);
756 ps
->tagged_reads
= pet_scop_collect_tagged_may_reads(scop
);
757 ps
->reads
= pet_scop_collect_may_reads(scop
);
758 ps
->tagged_may_writes
= pet_scop_collect_tagged_may_writes(scop
);
759 ps
->may_writes
= pet_scop_collect_may_writes(scop
);
760 ps
->tagged_must_writes
= pet_scop_collect_tagged_must_writes(scop
);
761 ps
->must_writes
= pet_scop_collect_must_writes(scop
);
762 ps
->tagged_must_kills
= pet_scop_collect_tagged_must_kills(scop
);
763 ps
->schedule
= pet_scop_collect_schedule(scop
);
765 ps
->independence
= isl_union_map_empty(isl_set_get_space(ps
->context
));
766 for (i
= 0; i
< scop
->n_independence
; ++i
)
767 ps
->independence
= isl_union_map_union(ps
->independence
,
768 isl_union_map_copy(scop
->independences
[i
]->filter
));
771 compute_dependences(ps
);
772 eliminate_dead_code(ps
);
774 if (!ps
->context
|| !ps
->domain
|| !ps
->call
|| !ps
->reads
||
775 !ps
->may_writes
|| !ps
->must_writes
|| !ps
->tagged_must_kills
||
776 !ps
->schedule
|| !ps
->independence
|| !ps
->names
)
777 return ppcg_scop_free(ps
);
782 /* Internal data structure for ppcg_transform.
784 struct ppcg_transform_data
{
785 struct ppcg_options
*options
;
786 __isl_give isl_printer
*(*transform
)(__isl_take isl_printer
*p
,
787 struct ppcg_scop
*scop
, void *user
);
791 /* Should we print the original code?
792 * That is, does "scop" involve any data dependent conditions or
793 * nested expressions that cannot be handled by pet_stmt_build_ast_exprs?
795 static int print_original(struct pet_scop
*scop
, struct ppcg_options
*options
)
797 if (!pet_scop_can_build_ast_exprs(scop
)) {
798 if (options
->debug
->verbose
)
799 fprintf(stdout
, "Printing original code because "
800 "some index expressions cannot currently "
805 if (pet_scop_has_data_dependent_conditions(scop
)) {
806 if (options
->debug
->verbose
)
807 fprintf(stdout
, "Printing original code because "
808 "input involves data dependent conditions\n");
815 /* Callback for pet_transform_C_source that transforms
816 * the given pet_scop to a ppcg_scop before calling the
817 * ppcg_transform callback.
819 * If "scop" contains any data dependent conditions or if we may
820 * not be able to print the transformed program, then just print
823 static __isl_give isl_printer
*transform(__isl_take isl_printer
*p
,
824 struct pet_scop
*scop
, void *user
)
826 struct ppcg_transform_data
*data
= user
;
827 struct ppcg_scop
*ps
;
829 if (print_original(scop
, data
->options
)) {
830 p
= pet_scop_print_original(scop
, p
);
835 scop
= pet_scop_align_params(scop
);
836 ps
= ppcg_scop_from_pet_scop(scop
, data
->options
);
838 p
= data
->transform(p
, ps
, data
->user
);
846 /* Transform the C source file "input" by rewriting each scop
847 * through a call to "transform".
848 * The transformed C code is written to "out".
850 * This is a wrapper around pet_transform_C_source that transforms
851 * the pet_scop to a ppcg_scop before calling "fn".
853 int ppcg_transform(isl_ctx
*ctx
, const char *input
, FILE *out
,
854 struct ppcg_options
*options
,
855 __isl_give isl_printer
*(*fn
)(__isl_take isl_printer
*p
,
856 struct ppcg_scop
*scop
, void *user
), void *user
)
858 struct ppcg_transform_data data
= { options
, fn
, user
};
859 return pet_transform_C_source(ctx
, input
, out
, &transform
, &data
);
862 /* Check consistency of options.
864 * Return -1 on error.
866 static int check_options(isl_ctx
*ctx
)
868 struct options
*options
;
870 options
= isl_ctx_peek_options(ctx
, &options_args
);
872 isl_die(ctx
, isl_error_internal
,
873 "unable to find options", return -1);
875 if (options
->ppcg
->openmp
&&
876 !isl_options_get_ast_build_atomic_upper_bound(ctx
))
877 isl_die(ctx
, isl_error_invalid
,
878 "OpenMP requires atomic bounds", return -1);
883 int main(int argc
, char **argv
)
887 struct options
*options
;
889 options
= options_new_with_defaults();
892 ctx
= isl_ctx_alloc_with_options(&options_args
, options
);
893 isl_options_set_schedule_outer_coincidence(ctx
, 1);
894 isl_options_set_schedule_maximize_band_depth(ctx
, 1);
895 pet_options_set_encapsulate_dynamic_control(ctx
, 1);
896 argc
= options_parse(options
, argc
, argv
, ISL_ARG_ALL
);
898 if (check_options(ctx
) < 0)
900 else if (options
->ppcg
->target
== PPCG_TARGET_CUDA
)
901 r
= generate_cuda(ctx
, options
->ppcg
, options
->input
);
902 else if (options
->ppcg
->target
== PPCG_TARGET_OPENCL
)
903 r
= generate_opencl(ctx
, options
->ppcg
, options
->input
,
906 r
= generate_cpu(ctx
, options
->ppcg
, options
->input
,