export pet_expr_new_cast
[pet.git] / tree2scop.c
blobdf8dee73d1a13e161065463d02d8d2977a8577c1
1 /*
2 * Copyright 2011 Leiden University. All rights reserved.
3 * Copyright 2012-2014 Ecole Normale Superieure. All rights reserved.
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above
13 * copyright notice, this list of conditions and the following
14 * disclaimer in the documentation and/or other materials provided
15 * with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY LEIDEN UNIVERSITY ''AS IS'' AND ANY
18 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
20 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL LEIDEN UNIVERSITY OR
21 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
22 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
23 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA,
24 * OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
27 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29 * The views and conclusions contained in the software and documentation
30 * are those of the authors and should not be interpreted as
31 * representing official policies, either expressed or implied, of
32 * Leiden University.
35 #include <stdlib.h>
36 #include <string.h>
38 #include <isl/id_to_pw_aff.h>
40 #include "aff.h"
41 #include "expr.h"
42 #include "expr_arg.h"
43 #include "nest.h"
44 #include "scop.h"
45 #include "skip.h"
46 #include "state.h"
47 #include "tree2scop.h"
49 /* Update "pc" by taking into account the writes in "stmt".
50 * That is, clear any previously assigned values to variables
51 * that are written by "stmt".
53 static __isl_give pet_context *handle_writes(struct pet_stmt *stmt,
54 __isl_take pet_context *pc)
56 return pet_context_clear_writes_in_tree(pc, stmt->body);
59 /* Update "pc" based on the write accesses in "scop".
61 static __isl_give pet_context *scop_handle_writes(struct pet_scop *scop,
62 __isl_take pet_context *pc)
64 int i;
66 if (!scop)
67 return pet_context_free(pc);
68 for (i = 0; i < scop->n_stmt; ++i)
69 pc = handle_writes(scop->stmts[i], pc);
71 return pc;
74 /* Wrapper around pet_expr_resolve_assume
75 * for use as a callback to pet_tree_map_expr.
77 static __isl_give pet_expr *resolve_assume(__isl_take pet_expr *expr,
78 void *user)
80 pet_context *pc = user;
82 return pet_expr_resolve_assume(expr, pc);
85 /* Check if any expression inside "tree" is an assume expression and
86 * if its single argument can be converted to an affine expression
87 * in the context of "pc".
88 * If so, replace the argument by the affine expression.
90 __isl_give pet_tree *pet_tree_resolve_assume(__isl_take pet_tree *tree,
91 __isl_keep pet_context *pc)
93 return pet_tree_map_expr(tree, &resolve_assume, pc);
96 /* Convert a pet_tree to a pet_scop with one statement within the context "pc".
97 * "tree" has already been evaluated in the context of "pc".
98 * This mainly involves resolving nested expression parameters
99 * and setting the name of the iteration space.
100 * The name is given by tree->label if it is non-NULL. Otherwise,
101 * it is of the form S_<stmt_nr>.
103 static struct pet_scop *scop_from_evaluated_tree(__isl_take pet_tree *tree,
104 int stmt_nr, __isl_keep pet_context *pc)
106 isl_space *space;
107 isl_set *domain;
108 struct pet_stmt *ps;
110 space = pet_context_get_space(pc);
112 tree = pet_tree_resolve_nested(tree, space);
113 tree = pet_tree_resolve_assume(tree, pc);
115 domain = pet_context_get_domain(pc);
116 ps = pet_stmt_from_pet_tree(domain, stmt_nr, tree);
117 return pet_scop_from_pet_stmt(space, ps);
120 /* Convert a top-level pet_expr to a pet_scop with one statement
121 * within the context "pc".
122 * "expr" has already been evaluated in the context of "pc".
123 * We construct a pet_tree from "expr" and continue with
124 * scop_from_evaluated_tree.
125 * The name is of the form S_<stmt_nr>.
126 * The location of the statement is set to "loc".
128 static struct pet_scop *scop_from_evaluated_expr(__isl_take pet_expr *expr,
129 int stmt_nr, __isl_take pet_loc *loc, __isl_keep pet_context *pc)
131 pet_tree *tree;
133 tree = pet_tree_new_expr(expr);
134 tree = pet_tree_set_loc(tree, loc);
135 return scop_from_evaluated_tree(tree, stmt_nr, pc);
138 /* Convert a pet_tree to a pet_scop with one statement within the context "pc".
139 * "tree" has not yet been evaluated in the context of "pc".
140 * We evaluate "tree" in the context of "pc" and continue with
141 * scop_from_evaluated_tree.
142 * The statement name is given by tree->label if it is non-NULL. Otherwise,
143 * it is of the form S_<stmt_nr>.
145 static struct pet_scop *scop_from_unevaluated_tree(__isl_take pet_tree *tree,
146 int stmt_nr, __isl_keep pet_context *pc)
148 tree = pet_context_evaluate_tree(pc, tree);
149 return scop_from_evaluated_tree(tree, stmt_nr, pc);
152 /* Convert a top-level pet_expr to a pet_scop with one statement
153 * within the context "pc", where "expr" has not yet been evaluated
154 * in the context of "pc".
155 * We construct a pet_tree from "expr" and continue with
156 * scop_from_unevaluated_tree.
157 * The statement name is of the form S_<stmt_nr>.
158 * The location of the statement is set to "loc".
160 static struct pet_scop *scop_from_expr(__isl_take pet_expr *expr,
161 int stmt_nr, __isl_take pet_loc *loc, __isl_keep pet_context *pc)
163 pet_tree *tree;
165 tree = pet_tree_new_expr(expr);
166 tree = pet_tree_set_loc(tree, loc);
167 return scop_from_unevaluated_tree(tree, stmt_nr, pc);
170 /* Construct a pet_scop with a single statement killing the entire
171 * array "array".
172 * The location of the statement is set to "loc".
174 static struct pet_scop *kill(__isl_take pet_loc *loc, struct pet_array *array,
175 __isl_keep pet_context *pc, struct pet_state *state)
177 isl_ctx *ctx;
178 isl_id *id;
179 isl_space *space;
180 isl_multi_pw_aff *index;
181 isl_map *access;
182 pet_expr *expr;
183 struct pet_scop *scop;
185 if (!array)
186 goto error;
187 ctx = isl_set_get_ctx(array->extent);
188 access = isl_map_from_range(isl_set_copy(array->extent));
189 id = isl_set_get_tuple_id(array->extent);
190 space = isl_space_alloc(ctx, 0, 0, 0);
191 space = isl_space_set_tuple_id(space, isl_dim_out, id);
192 index = isl_multi_pw_aff_zero(space);
193 expr = pet_expr_kill_from_access_and_index(access, index);
194 return scop_from_expr(expr, state->n_stmt++, loc, pc);
195 error:
196 pet_loc_free(loc);
197 return NULL;
200 /* Construct and return a pet_array corresponding to the variable
201 * accessed by "access" by calling the extract_array callback.
203 static struct pet_array *extract_array(__isl_keep pet_expr *access,
204 __isl_keep pet_context *pc, struct pet_state *state)
206 return state->extract_array(access, pc, state->user);
209 /* Construct a pet_scop for a (single) variable declaration
210 * within the context "pc".
212 * The scop contains the variable being declared (as an array)
213 * and a statement killing the array.
215 * If the declaration comes with an initialization, then the scop
216 * also contains an assignment to the variable.
218 static struct pet_scop *scop_from_decl(__isl_keep pet_tree *tree,
219 __isl_keep pet_context *pc, struct pet_state *state)
221 int type_size;
222 isl_ctx *ctx;
223 struct pet_array *array;
224 struct pet_scop *scop_decl, *scop;
225 pet_expr *lhs, *rhs, *pe;
227 array = extract_array(tree->u.d.var, pc, state);
228 if (array)
229 array->declared = 1;
230 scop_decl = kill(pet_tree_get_loc(tree), array, pc, state);
231 scop_decl = pet_scop_add_array(scop_decl, array);
233 if (tree->type != pet_tree_decl_init)
234 return scop_decl;
236 lhs = pet_expr_copy(tree->u.d.var);
237 rhs = pet_expr_copy(tree->u.d.init);
238 type_size = pet_expr_get_type_size(lhs);
239 pe = pet_expr_new_binary(type_size, pet_op_assign, lhs, rhs);
240 scop = scop_from_expr(pe, state->n_stmt++, pet_tree_get_loc(tree), pc);
242 scop_decl = pet_scop_prefix(scop_decl, 0);
243 scop = pet_scop_prefix(scop, 1);
245 ctx = pet_tree_get_ctx(tree);
246 scop = pet_scop_add_seq(ctx, scop_decl, scop);
248 return scop;
251 /* Does "tree" represent a kill statement?
252 * That is, is it an expression statement that "calls" __pencil_kill?
254 static int is_pencil_kill(__isl_keep pet_tree *tree)
256 pet_expr *expr;
257 const char *name;
259 if (!tree)
260 return -1;
261 if (tree->type != pet_tree_expr)
262 return 0;
263 expr = tree->u.e.expr;
264 if (pet_expr_get_type(expr) != pet_expr_call)
265 return 0;
266 name = pet_expr_call_get_name(expr);
267 if (!name)
268 return -1;
269 return !strcmp(name, "__pencil_kill");
272 /* Add a kill to "scop" that kills what is accessed by
273 * the access expression "expr".
275 * If the access expression has any arguments (after evaluation
276 * in the context of "pc"), then we ignore it, since we cannot
277 * tell which elements are definitely killed.
279 * Otherwise, we extend the index expression to the dimension
280 * of the accessed array and intersect with the extent of the array and
281 * add a kill expression that kills these array elements is added to "scop".
283 static struct pet_scop *scop_add_kill(struct pet_scop *scop,
284 __isl_take pet_expr *expr, __isl_take pet_loc *loc,
285 __isl_keep pet_context *pc, struct pet_state *state)
287 int dim1, dim2;
288 isl_id *id;
289 isl_multi_pw_aff *index;
290 isl_map *map;
291 pet_expr *kill;
292 struct pet_array *array;
293 struct pet_scop *scop_i;
295 expr = pet_context_evaluate_expr(pc, expr);
296 if (!expr)
297 goto error;
298 if (expr->n_arg != 0) {
299 pet_expr_free(expr);
300 return scop;
302 array = extract_array(expr, pc, state);
303 if (!array)
304 goto error;
305 index = pet_expr_access_get_index(expr);
306 pet_expr_free(expr);
307 map = isl_map_from_multi_pw_aff(isl_multi_pw_aff_copy(index));
308 id = isl_map_get_tuple_id(map, isl_dim_out);
309 dim1 = isl_set_dim(array->extent, isl_dim_set);
310 dim2 = isl_map_dim(map, isl_dim_out);
311 map = isl_map_add_dims(map, isl_dim_out, dim1 - dim2);
312 map = isl_map_set_tuple_id(map, isl_dim_out, id);
313 map = isl_map_intersect_range(map, isl_set_copy(array->extent));
314 pet_array_free(array);
315 kill = pet_expr_kill_from_access_and_index(map, index);
316 scop_i = scop_from_evaluated_expr(kill, state->n_stmt++, loc, pc);
317 scop = pet_scop_add_par(state->ctx, scop, scop_i);
319 return scop;
320 error:
321 pet_expr_free(expr);
322 return pet_scop_free(scop);
325 /* For each argument of the __pencil_kill call in "tree" that
326 * represents an access, add a kill statement to "scop" killing the accessed
327 * elements.
329 static struct pet_scop *scop_from_pencil_kill(__isl_keep pet_tree *tree,
330 __isl_keep pet_context *pc, struct pet_state *state)
332 pet_expr *call;
333 struct pet_scop *scop;
334 int i, n;
336 call = tree->u.e.expr;
338 scop = pet_scop_empty(pet_context_get_space(pc));
340 n = pet_expr_get_n_arg(call);
341 for (i = 0; i < n; ++i) {
342 pet_expr *arg;
343 pet_loc *loc;
345 arg = pet_expr_get_arg(call, i);
346 if (!arg)
347 return pet_scop_free(scop);
348 if (pet_expr_get_type(arg) != pet_expr_access) {
349 pet_expr_free(arg);
350 continue;
352 loc = pet_tree_get_loc(tree);
353 scop = scop_add_kill(scop, arg, loc, pc, state);
356 return scop;
359 /* Construct a pet_scop for an expression statement within the context "pc".
361 * If the expression calls __pencil_kill, then it needs to be converted
362 * into zero or more kill statements.
363 * Otherwise, a scop is extracted directly from the tree.
365 static struct pet_scop *scop_from_tree_expr(__isl_keep pet_tree *tree,
366 __isl_keep pet_context *pc, struct pet_state *state)
368 int is_kill;
370 is_kill = is_pencil_kill(tree);
371 if (is_kill < 0)
372 return NULL;
373 if (is_kill)
374 return scop_from_pencil_kill(tree, pc, state);
375 return scop_from_unevaluated_tree(pet_tree_copy(tree),
376 state->n_stmt++, pc);
379 /* Return those elements in the space of "cond" that come after
380 * (based on "sign") an element in "cond" in the final dimension.
382 static __isl_give isl_set *after(__isl_take isl_set *cond, int sign)
384 isl_space *space;
385 isl_map *previous_to_this;
386 int i, dim;
388 dim = isl_set_dim(cond, isl_dim_set);
389 space = isl_space_map_from_set(isl_set_get_space(cond));
390 previous_to_this = isl_map_universe(space);
391 for (i = 0; i + 1 < dim; ++i)
392 previous_to_this = isl_map_equate(previous_to_this,
393 isl_dim_in, i, isl_dim_out, i);
394 if (sign > 0)
395 previous_to_this = isl_map_order_lt(previous_to_this,
396 isl_dim_in, dim - 1, isl_dim_out, dim - 1);
397 else
398 previous_to_this = isl_map_order_gt(previous_to_this,
399 isl_dim_in, dim - 1, isl_dim_out, dim - 1);
401 cond = isl_set_apply(cond, previous_to_this);
403 return cond;
406 /* Remove those iterations of "domain" that have an earlier iteration
407 * (based on "sign") in the final dimension where "skip" is satisfied.
408 * If "apply_skip_map" is set, then "skip_map" is first applied
409 * to the embedded skip condition before removing it from the domain.
411 static __isl_give isl_set *apply_affine_break(__isl_take isl_set *domain,
412 __isl_take isl_set *skip, int sign,
413 int apply_skip_map, __isl_keep isl_map *skip_map)
415 if (apply_skip_map)
416 skip = isl_set_apply(skip, isl_map_copy(skip_map));
417 skip = isl_set_intersect(skip , isl_set_copy(domain));
418 return isl_set_subtract(domain, after(skip, sign));
421 /* Create an affine expression on the domain space of "pc" that
422 * is equal to the final dimension of this domain.
424 static __isl_give isl_aff *map_to_last(__isl_keep pet_context *pc)
426 int pos;
427 isl_space *space;
428 isl_local_space *ls;
430 space = pet_context_get_space(pc);
431 pos = isl_space_dim(space, isl_dim_set) - 1;
432 ls = isl_local_space_from_space(space);
433 return isl_aff_var_on_domain(ls, isl_dim_set, pos);
436 /* Create an affine expression that maps elements
437 * of an array "id_test" to the previous element in the final dimension
438 * (according to "inc"), provided this element belongs to "domain".
439 * That is, create the affine expression
441 * { id[outer,x] -> id[outer,x - inc] : (outer,x - inc) in domain }
443 static __isl_give isl_multi_pw_aff *map_to_previous(__isl_take isl_id *id_test,
444 __isl_take isl_set *domain, __isl_take isl_val *inc)
446 int pos;
447 isl_space *space;
448 isl_aff *aff;
449 isl_pw_aff *pa;
450 isl_multi_aff *ma;
451 isl_multi_pw_aff *prev;
453 pos = isl_set_dim(domain, isl_dim_set) - 1;
454 space = isl_set_get_space(domain);
455 space = isl_space_map_from_set(space);
456 ma = isl_multi_aff_identity(space);
457 aff = isl_multi_aff_get_aff(ma, pos);
458 aff = isl_aff_add_constant_val(aff, isl_val_neg(inc));
459 ma = isl_multi_aff_set_aff(ma, pos, aff);
460 domain = isl_set_preimage_multi_aff(domain, isl_multi_aff_copy(ma));
461 prev = isl_multi_pw_aff_from_multi_aff(ma);
462 pa = isl_multi_pw_aff_get_pw_aff(prev, pos);
463 pa = isl_pw_aff_intersect_domain(pa, domain);
464 prev = isl_multi_pw_aff_set_pw_aff(prev, pos, pa);
465 prev = isl_multi_pw_aff_set_tuple_id(prev, isl_dim_out, id_test);
467 return prev;
470 /* Add an implication to "scop" expressing that if an element of
471 * virtual array "id_test" has value "satisfied" then all previous elements
472 * of this array (in the final dimension) also have that value.
473 * The set of previous elements is bounded by "domain".
474 * If "sign" is negative then the iterator
475 * is decreasing and we express that all subsequent array elements
476 * (but still defined previously) have the same value.
478 static struct pet_scop *add_implication(struct pet_scop *scop,
479 __isl_take isl_id *id_test, __isl_take isl_set *domain, int sign,
480 int satisfied)
482 int i, dim;
483 isl_space *space;
484 isl_map *map;
486 dim = isl_set_dim(domain, isl_dim_set);
487 domain = isl_set_set_tuple_id(domain, id_test);
488 space = isl_space_map_from_set(isl_set_get_space(domain));
489 map = isl_map_universe(space);
490 for (i = 0; i + 1 < dim; ++i)
491 map = isl_map_equate(map, isl_dim_in, i, isl_dim_out, i);
492 if (sign > 0)
493 map = isl_map_order_ge(map,
494 isl_dim_in, dim - 1, isl_dim_out, dim - 1);
495 else
496 map = isl_map_order_le(map,
497 isl_dim_in, dim - 1, isl_dim_out, dim - 1);
498 map = isl_map_intersect_range(map, domain);
499 scop = pet_scop_add_implication(scop, map, satisfied);
501 return scop;
504 /* Add a filter to "scop" that imposes that it is only executed
505 * when the variable identified by "id_test" has a zero value
506 * for all previous iterations of "domain".
508 * In particular, add a filter that imposes that the array
509 * has a zero value at the previous iteration of domain and
510 * add an implication that implies that it then has that
511 * value for all previous iterations.
513 static struct pet_scop *scop_add_break(struct pet_scop *scop,
514 __isl_take isl_id *id_test, __isl_take isl_set *domain,
515 __isl_take isl_val *inc)
517 isl_multi_pw_aff *prev;
518 int sign = isl_val_sgn(inc);
520 prev = map_to_previous(isl_id_copy(id_test), isl_set_copy(domain), inc);
521 scop = add_implication(scop, id_test, domain, sign, 0);
522 scop = pet_scop_filter(scop, prev, 0);
524 return scop;
527 static struct pet_scop *scop_from_tree(__isl_keep pet_tree *tree,
528 __isl_keep pet_context *pc, struct pet_state *state);
530 /* Construct a pet_scop for an infinite loop around the given body
531 * within the context "pc".
533 * The domain of "pc" has already been extended with an infinite loop
535 * { [t] : t >= 0 }
537 * We extract a pet_scop for the body and then embed it in a loop with
538 * schedule
540 * { [outer,t] -> [t] }
542 * If the body contains any break, then it is taken into
543 * account in apply_affine_break (if the skip condition is affine)
544 * or in scop_add_break (if the skip condition is not affine).
546 * Note that in case of an affine skip condition,
547 * since we are dealing with a loop without loop iterator,
548 * the skip condition cannot refer to the current loop iterator and
549 * so effectively, the effect on the iteration domain is of the form
551 * { [outer,0]; [outer,t] : t >= 1 and not skip }
553 static struct pet_scop *scop_from_infinite_loop(__isl_keep pet_tree *body,
554 __isl_keep pet_context *pc, struct pet_state *state)
556 isl_ctx *ctx;
557 isl_id *id_test;
558 isl_set *domain;
559 isl_set *skip;
560 isl_aff *sched;
561 struct pet_scop *scop;
562 int has_affine_break;
563 int has_var_break;
565 ctx = pet_tree_get_ctx(body);
566 domain = pet_context_get_domain(pc);
567 sched = map_to_last(pc);
569 scop = scop_from_tree(body, pc, state);
571 has_affine_break = pet_scop_has_affine_skip(scop, pet_skip_later);
572 if (has_affine_break)
573 skip = pet_scop_get_affine_skip_domain(scop, pet_skip_later);
574 has_var_break = pet_scop_has_var_skip(scop, pet_skip_later);
575 if (has_var_break)
576 id_test = pet_scop_get_skip_id(scop, pet_skip_later);
578 scop = pet_scop_embed(scop, isl_set_copy(domain), sched);
579 if (has_affine_break) {
580 domain = apply_affine_break(domain, skip, 1, 0, NULL);
581 scop = pet_scop_intersect_domain_prefix(scop,
582 isl_set_copy(domain));
584 if (has_var_break)
585 scop = scop_add_break(scop, id_test, domain, isl_val_one(ctx));
586 else
587 isl_set_free(domain);
589 return scop;
592 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
594 * for (;;)
595 * body
597 * within the context "pc".
599 * Extend the domain of "pc" with an extra inner loop
601 * { [t] : t >= 0 }
603 * and construct the scop in scop_from_infinite_loop.
605 static struct pet_scop *scop_from_infinite_for(__isl_keep pet_tree *tree,
606 __isl_keep pet_context *pc, struct pet_state *state)
608 struct pet_scop *scop;
610 pc = pet_context_copy(pc);
611 pc = pet_context_clear_writes_in_tree(pc, tree->u.l.body);
613 pc = pet_context_add_infinite_loop(pc);
615 scop = scop_from_infinite_loop(tree->u.l.body, pc, state);
617 pet_context_free(pc);
619 return scop;
622 /* Construct a pet_scop for a while loop of the form
624 * while (pa)
625 * body
627 * within the context "pc".
629 * The domain of "pc" has already been extended with an infinite loop
631 * { [t] : t >= 0 }
633 * Here, we add the constraints on the outer loop iterators
634 * implied by "pa" and construct the scop in scop_from_infinite_loop.
635 * Note that the intersection with these constraints
636 * may result in an empty loop.
638 static struct pet_scop *scop_from_affine_while(__isl_keep pet_tree *tree,
639 __isl_take isl_pw_aff *pa, __isl_take pet_context *pc,
640 struct pet_state *state)
642 struct pet_scop *scop;
643 isl_set *dom, *local;
644 isl_set *valid;
646 valid = isl_pw_aff_domain(isl_pw_aff_copy(pa));
647 dom = isl_pw_aff_non_zero_set(pa);
648 local = isl_set_add_dims(isl_set_copy(dom), isl_dim_set, 1);
649 pc = pet_context_intersect_domain(pc, local);
650 scop = scop_from_infinite_loop(tree->u.l.body, pc, state);
651 scop = pet_scop_restrict(scop, dom);
652 scop = pet_scop_restrict_context(scop, valid);
654 pet_context_free(pc);
655 return scop;
658 /* Construct a scop for a while, given the scops for the condition
659 * and the body, the filter identifier and the iteration domain of
660 * the while loop.
662 * In particular, the scop for the condition is filtered to depend
663 * on "id_test" evaluating to true for all previous iterations
664 * of the loop, while the scop for the body is filtered to depend
665 * on "id_test" evaluating to true for all iterations up to the
666 * current iteration.
667 * The actual filter only imposes that this virtual array has
668 * value one on the previous or the current iteration.
669 * The fact that this condition also applies to the previous
670 * iterations is enforced by an implication.
672 * These filtered scops are then combined into a single scop.
674 * "sign" is positive if the iterator increases and negative
675 * if it decreases.
677 static struct pet_scop *scop_add_while(struct pet_scop *scop_cond,
678 struct pet_scop *scop_body, __isl_take isl_id *id_test,
679 __isl_take isl_set *domain, __isl_take isl_val *inc)
681 isl_ctx *ctx = isl_set_get_ctx(domain);
682 isl_space *space;
683 isl_multi_pw_aff *test_index;
684 isl_multi_pw_aff *prev;
685 int sign = isl_val_sgn(inc);
686 struct pet_scop *scop;
688 prev = map_to_previous(isl_id_copy(id_test), isl_set_copy(domain), inc);
689 scop_cond = pet_scop_filter(scop_cond, prev, 1);
691 space = isl_space_map_from_set(isl_set_get_space(domain));
692 test_index = isl_multi_pw_aff_identity(space);
693 test_index = isl_multi_pw_aff_set_tuple_id(test_index, isl_dim_out,
694 isl_id_copy(id_test));
695 scop_body = pet_scop_filter(scop_body, test_index, 1);
697 scop = pet_scop_add_seq(ctx, scop_cond, scop_body);
698 scop = add_implication(scop, id_test, domain, sign, 1);
700 return scop;
703 /* Create a pet_scop with a single statement with name S_<stmt_nr>,
704 * evaluating "cond" and writing the result to a virtual scalar,
705 * as expressed by "index".
706 * The expression "cond" has not yet been evaluated in the context of "pc".
707 * Do so within the context "pc".
708 * The location of the statement is set to "loc".
710 static struct pet_scop *scop_from_non_affine_condition(
711 __isl_take pet_expr *cond, int stmt_nr,
712 __isl_take isl_multi_pw_aff *index,
713 __isl_take pet_loc *loc, __isl_keep pet_context *pc)
715 pet_expr *expr, *write;
717 cond = pet_context_evaluate_expr(pc, cond);
719 write = pet_expr_from_index(index);
720 write = pet_expr_access_set_write(write, 1);
721 write = pet_expr_access_set_read(write, 0);
722 expr = pet_expr_new_binary(1, pet_op_assign, write, cond);
724 return scop_from_evaluated_expr(expr, stmt_nr, loc, pc);
727 /* Given that "scop" has an affine skip condition of type pet_skip_now,
728 * apply this skip condition to the domain of "pc".
729 * That is, remove the elements satisfying the skip condition from
730 * the domain of "pc".
732 static __isl_give pet_context *apply_affine_continue(__isl_take pet_context *pc,
733 struct pet_scop *scop)
735 isl_set *domain, *skip;
737 skip = pet_scop_get_affine_skip_domain(scop, pet_skip_now);
738 domain = pet_context_get_domain(pc);
739 domain = isl_set_subtract(domain, skip);
740 pc = pet_context_intersect_domain(pc, domain);
742 return pc;
745 /* Add a scop for evaluating the loop increment "inc" add the end
746 * of a loop body "scop" within the context "pc".
748 * The skip conditions resulting from continue statements inside
749 * the body do not apply to "inc", but those resulting from break
750 * statements do need to get applied.
752 static struct pet_scop *scop_add_inc(struct pet_scop *scop,
753 __isl_take pet_expr *inc, __isl_take pet_loc *loc,
754 __isl_keep pet_context *pc, struct pet_state *state)
756 struct pet_scop *scop_inc;
758 pc = pet_context_copy(pc);
760 if (pet_scop_has_skip(scop, pet_skip_later)) {
761 isl_multi_pw_aff *skip;
762 skip = pet_scop_get_skip(scop, pet_skip_later);
763 scop = pet_scop_set_skip(scop, pet_skip_now, skip);
764 if (pet_scop_has_affine_skip(scop, pet_skip_now))
765 pc = apply_affine_continue(pc, scop);
766 } else
767 pet_scop_reset_skip(scop, pet_skip_now);
768 scop_inc = scop_from_expr(inc, state->n_stmt++, loc, pc);
769 scop_inc = pet_scop_prefix(scop_inc, 2);
770 scop = pet_scop_add_seq(state->ctx, scop, scop_inc);
772 pet_context_free(pc);
774 return scop;
777 /* Construct a generic while scop, with iteration domain
778 * { [t] : t >= 0 } around the scop for "tree_body" within the context "pc".
779 * The domain of "pc" has already been extended with this infinite loop
781 * { [t] : t >= 0 }
783 * The scop consists of two parts,
784 * one for evaluating the condition "cond" and one for the body.
785 * If "expr_inc" is not NULL, then a scop for evaluating this expression
786 * is added at the end of the body,
787 * after replacing any skip conditions resulting from continue statements
788 * by the skip conditions resulting from break statements (if any).
790 * The schedule is adjusted to reflect that the condition is evaluated
791 * before the body is executed and the body is filtered to depend
792 * on the result of the condition evaluating to true on all iterations
793 * up to the current iteration, while the evaluation of the condition itself
794 * is filtered to depend on the result of the condition evaluating to true
795 * on all previous iterations.
796 * The context of the scop representing the body is dropped
797 * because we don't know how many times the body will be executed,
798 * if at all.
800 * If the body contains any break, then it is taken into
801 * account in apply_affine_break (if the skip condition is affine)
802 * or in scop_add_break (if the skip condition is not affine).
804 * Note that in case of an affine skip condition,
805 * since we are dealing with a loop without loop iterator,
806 * the skip condition cannot refer to the current loop iterator and
807 * so effectively, the effect on the iteration domain is of the form
809 * { [outer,0]; [outer,t] : t >= 1 and not skip }
811 static struct pet_scop *scop_from_non_affine_while(__isl_take pet_expr *cond,
812 __isl_take pet_loc *loc, __isl_keep pet_tree *tree_body,
813 __isl_take pet_expr *expr_inc, __isl_take pet_context *pc,
814 struct pet_state *state)
816 isl_ctx *ctx;
817 isl_id *id_test, *id_break_test;
818 isl_space *space;
819 isl_multi_pw_aff *test_index;
820 isl_set *domain;
821 isl_set *skip;
822 isl_aff *sched;
823 struct pet_scop *scop, *scop_body;
824 int has_affine_break;
825 int has_var_break;
827 ctx = state->ctx;
828 space = pet_context_get_space(pc);
829 test_index = pet_create_test_index(space, state->n_test++);
830 scop = scop_from_non_affine_condition(cond, state->n_stmt++,
831 isl_multi_pw_aff_copy(test_index),
832 pet_loc_copy(loc), pc);
833 id_test = isl_multi_pw_aff_get_tuple_id(test_index, isl_dim_out);
834 domain = pet_context_get_domain(pc);
835 scop = pet_scop_add_boolean_array(scop, isl_set_copy(domain),
836 test_index, state->int_size);
838 sched = map_to_last(pc);
840 scop_body = scop_from_tree(tree_body, pc, state);
842 has_affine_break = pet_scop_has_affine_skip(scop_body, pet_skip_later);
843 if (has_affine_break)
844 skip = pet_scop_get_affine_skip_domain(scop_body,
845 pet_skip_later);
846 has_var_break = pet_scop_has_var_skip(scop_body, pet_skip_later);
847 if (has_var_break)
848 id_break_test = pet_scop_get_skip_id(scop_body, pet_skip_later);
850 scop = pet_scop_prefix(scop, 0);
851 scop = pet_scop_embed(scop, isl_set_copy(domain), isl_aff_copy(sched));
852 scop_body = pet_scop_reset_context(scop_body);
853 scop_body = pet_scop_prefix(scop_body, 1);
854 if (expr_inc) {
855 scop_body = scop_add_inc(scop_body, expr_inc, loc, pc, state);
856 } else
857 pet_loc_free(loc);
858 scop_body = pet_scop_embed(scop_body, isl_set_copy(domain), sched);
860 if (has_affine_break) {
861 domain = apply_affine_break(domain, skip, 1, 0, NULL);
862 scop = pet_scop_intersect_domain_prefix(scop,
863 isl_set_copy(domain));
864 scop_body = pet_scop_intersect_domain_prefix(scop_body,
865 isl_set_copy(domain));
867 if (has_var_break) {
868 scop = scop_add_break(scop, isl_id_copy(id_break_test),
869 isl_set_copy(domain), isl_val_one(ctx));
870 scop_body = scop_add_break(scop_body, id_break_test,
871 isl_set_copy(domain), isl_val_one(ctx));
873 scop = scop_add_while(scop, scop_body, id_test, domain,
874 isl_val_one(ctx));
876 pet_context_free(pc);
877 return scop;
880 /* Check if the while loop is of the form
882 * while (affine expression)
883 * body
885 * If so, call scop_from_affine_while to construct a scop.
887 * Otherwise, pass control to scop_from_non_affine_while.
889 * "pc" is the context in which the affine expressions in the scop are created.
890 * The domain of "pc" is extended with an infinite loop
892 * { [t] : t >= 0 }
894 * before passing control to scop_from_affine_while or
895 * scop_from_non_affine_while.
897 static struct pet_scop *scop_from_while(__isl_keep pet_tree *tree,
898 __isl_keep pet_context *pc, struct pet_state *state)
900 pet_expr *cond_expr;
901 isl_pw_aff *pa;
903 if (!tree)
904 return NULL;
906 pc = pet_context_copy(pc);
907 pc = pet_context_clear_writes_in_tree(pc, tree->u.l.body);
909 cond_expr = pet_expr_copy(tree->u.l.cond);
910 cond_expr = pet_context_evaluate_expr(pc, cond_expr);
911 pa = pet_expr_extract_affine_condition(cond_expr, pc);
912 pet_expr_free(cond_expr);
914 pc = pet_context_add_infinite_loop(pc);
916 if (!pa)
917 goto error;
919 if (!isl_pw_aff_involves_nan(pa))
920 return scop_from_affine_while(tree, pa, pc, state);
921 isl_pw_aff_free(pa);
922 return scop_from_non_affine_while(pet_expr_copy(tree->u.l.cond),
923 pet_tree_get_loc(tree), tree->u.l.body, NULL,
924 pc, state);
925 error:
926 pet_context_free(pc);
927 return NULL;
930 /* Check whether "cond" expresses a simple loop bound
931 * on the final set dimension.
932 * In particular, if "up" is set then "cond" should contain only
933 * upper bounds on the final set dimension.
934 * Otherwise, it should contain only lower bounds.
936 static int is_simple_bound(__isl_keep isl_set *cond, __isl_keep isl_val *inc)
938 int pos;
940 pos = isl_set_dim(cond, isl_dim_set) - 1;
941 if (isl_val_is_pos(inc))
942 return !isl_set_dim_has_any_lower_bound(cond, isl_dim_set, pos);
943 else
944 return !isl_set_dim_has_any_upper_bound(cond, isl_dim_set, pos);
947 /* Extend a condition on a given iteration of a loop to one that
948 * imposes the same condition on all previous iterations.
949 * "domain" expresses the lower [upper] bound on the iterations
950 * when inc is positive [negative] in its final dimension.
952 * In particular, we construct the condition (when inc is positive)
954 * forall i' : (domain(i') and i' <= i) => cond(i')
956 * (where "<=" applies to the final dimension)
957 * which is equivalent to
959 * not exists i' : domain(i') and i' <= i and not cond(i')
961 * We construct this set by subtracting the satisfying cond from domain,
962 * applying a map
964 * { [i'] -> [i] : i' <= i }
966 * and then subtracting the result from domain again.
968 static __isl_give isl_set *valid_for_each_iteration(__isl_take isl_set *cond,
969 __isl_take isl_set *domain, __isl_take isl_val *inc)
971 isl_space *space;
972 isl_map *previous_to_this;
973 int i, dim;
975 dim = isl_set_dim(cond, isl_dim_set);
976 space = isl_space_map_from_set(isl_set_get_space(cond));
977 previous_to_this = isl_map_universe(space);
978 for (i = 0; i + 1 < dim; ++i)
979 previous_to_this = isl_map_equate(previous_to_this,
980 isl_dim_in, i, isl_dim_out, i);
981 if (isl_val_is_pos(inc))
982 previous_to_this = isl_map_order_le(previous_to_this,
983 isl_dim_in, dim - 1, isl_dim_out, dim - 1);
984 else
985 previous_to_this = isl_map_order_ge(previous_to_this,
986 isl_dim_in, dim - 1, isl_dim_out, dim - 1);
988 cond = isl_set_subtract(isl_set_copy(domain), cond);
989 cond = isl_set_apply(cond, previous_to_this);
990 cond = isl_set_subtract(domain, cond);
992 isl_val_free(inc);
994 return cond;
997 /* Given an initial value of the form
999 * { [outer,i] -> init(outer) }
1001 * construct a domain of the form
1003 * { [outer,i] : exists a: i = init(outer) + a * inc and a >= 0 }
1005 static __isl_give isl_set *strided_domain(__isl_take isl_pw_aff *init,
1006 __isl_take isl_val *inc)
1008 int dim;
1009 isl_aff *aff;
1010 isl_space *space;
1011 isl_local_space *ls;
1012 isl_set *set;
1014 dim = isl_pw_aff_dim(init, isl_dim_in);
1016 init = isl_pw_aff_add_dims(init, isl_dim_in, 1);
1017 space = isl_pw_aff_get_domain_space(init);
1018 ls = isl_local_space_from_space(space);
1019 aff = isl_aff_zero_on_domain(isl_local_space_copy(ls));
1020 aff = isl_aff_add_coefficient_val(aff, isl_dim_in, dim, inc);
1021 init = isl_pw_aff_add(init, isl_pw_aff_from_aff(aff));
1023 aff = isl_aff_var_on_domain(ls, isl_dim_set, dim - 1);
1024 set = isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff), init);
1026 set = isl_set_lower_bound_si(set, isl_dim_set, dim, 0);
1027 set = isl_set_project_out(set, isl_dim_set, dim, 1);
1029 return set;
1032 /* Assuming "cond" represents a bound on a loop where the loop
1033 * iterator "iv" is incremented (or decremented) by one, check if wrapping
1034 * is possible.
1036 * Under the given assumptions, wrapping is only possible if "cond" allows
1037 * for the last value before wrapping, i.e., 2^width - 1 in case of an
1038 * increasing iterator and 0 in case of a decreasing iterator.
1040 static int can_wrap(__isl_keep isl_set *cond, __isl_keep pet_expr *iv,
1041 __isl_keep isl_val *inc)
1043 int cw;
1044 isl_ctx *ctx;
1045 isl_val *limit;
1046 isl_set *test;
1048 test = isl_set_copy(cond);
1050 ctx = isl_set_get_ctx(test);
1051 if (isl_val_is_neg(inc))
1052 limit = isl_val_zero(ctx);
1053 else {
1054 limit = isl_val_int_from_ui(ctx, pet_expr_get_type_size(iv));
1055 limit = isl_val_2exp(limit);
1056 limit = isl_val_sub_ui(limit, 1);
1059 test = isl_set_fix_val(cond, isl_dim_set, 0, limit);
1060 cw = !isl_set_is_empty(test);
1061 isl_set_free(test);
1063 return cw;
1066 /* Given a space
1068 * { [outer, v] },
1070 * construct the following affine expression on this space
1072 * { [outer, v] -> [outer, v mod 2^width] }
1074 * where width is the number of bits used to represent the values
1075 * of the unsigned variable "iv".
1077 static __isl_give isl_multi_aff *compute_wrapping(__isl_take isl_space *space,
1078 __isl_keep pet_expr *iv)
1080 int dim;
1081 isl_ctx *ctx;
1082 isl_val *mod;
1083 isl_aff *aff;
1084 isl_multi_aff *ma;
1086 dim = isl_space_dim(space, isl_dim_set);
1088 ctx = isl_space_get_ctx(space);
1089 mod = isl_val_int_from_ui(ctx, pet_expr_get_type_size(iv));
1090 mod = isl_val_2exp(mod);
1092 space = isl_space_map_from_set(space);
1093 ma = isl_multi_aff_identity(space);
1095 aff = isl_multi_aff_get_aff(ma, dim - 1);
1096 aff = isl_aff_mod_val(aff, mod);
1097 ma = isl_multi_aff_set_aff(ma, dim - 1, aff);
1099 return ma;
1102 /* Given two sets in the space
1104 * { [l,i] },
1106 * where l represents the outer loop iterators, compute the set
1107 * of values of l that ensure that "set1" is a subset of "set2".
1109 * set1 is a subset of set2 if
1111 * forall i: set1(l,i) => set2(l,i)
1113 * or
1115 * not exists i: set1(l,i) and not set2(l,i)
1117 * i.e.,
1119 * not exists i: (set1 \ set2)(l,i)
1121 static __isl_give isl_set *enforce_subset(__isl_take isl_set *set1,
1122 __isl_take isl_set *set2)
1124 int pos;
1126 pos = isl_set_dim(set1, isl_dim_set) - 1;
1127 set1 = isl_set_subtract(set1, set2);
1128 set1 = isl_set_eliminate(set1, isl_dim_set, pos, 1);
1129 return isl_set_complement(set1);
1132 /* Compute the set of outer iterator values for which "cond" holds
1133 * on the next iteration of the inner loop for each element of "dom".
1135 * We first construct mapping { [l,i] -> [l,i + inc] } (where l refers
1136 * to the outer loop iterators), plug that into "cond"
1137 * and then compute the set of outer iterators for which "dom" is a subset
1138 * of the result.
1140 static __isl_give isl_set *valid_on_next(__isl_take isl_set *cond,
1141 __isl_take isl_set *dom, __isl_take isl_val *inc)
1143 int pos;
1144 isl_space *space;
1145 isl_aff *aff;
1146 isl_multi_aff *ma;
1148 pos = isl_set_dim(dom, isl_dim_set) - 1;
1149 space = isl_set_get_space(dom);
1150 space = isl_space_map_from_set(space);
1151 ma = isl_multi_aff_identity(space);
1152 aff = isl_multi_aff_get_aff(ma, pos);
1153 aff = isl_aff_add_constant_val(aff, inc);
1154 ma = isl_multi_aff_set_aff(ma, pos, aff);
1155 cond = isl_set_preimage_multi_aff(cond, ma);
1157 return enforce_subset(dom, cond);
1160 /* Extract the for loop "tree" as a while loop within the context "pc_init".
1161 * In particular, "pc_init" represents the context of the loop,
1162 * whereas "pc" represents the context of the body of the loop and
1163 * has already had its domain extended with an infinite loop
1165 * { [t] : t >= 0 }
1167 * The for loop has the form
1169 * for (iv = init; cond; iv += inc)
1170 * body;
1172 * and is treated as
1174 * iv = init;
1175 * while (cond) {
1176 * body;
1177 * iv += inc;
1180 * except that the skips resulting from any continue statements
1181 * in body do not apply to the increment, but are replaced by the skips
1182 * resulting from break statements.
1184 * If the loop iterator is declared in the for loop, then it is killed before
1185 * and after the loop.
1187 static struct pet_scop *scop_from_non_affine_for(__isl_keep pet_tree *tree,
1188 __isl_keep pet_context *init_pc, __isl_take pet_context *pc,
1189 struct pet_state *state)
1191 int declared;
1192 isl_id *iv;
1193 pet_expr *expr_iv, *init, *inc;
1194 struct pet_scop *scop_init, *scop;
1195 int type_size;
1196 struct pet_array *array;
1197 struct pet_scop *scop_kill;
1199 iv = pet_expr_access_get_id(tree->u.l.iv);
1200 pc = pet_context_clear_value(pc, iv);
1202 declared = tree->u.l.declared;
1204 expr_iv = pet_expr_copy(tree->u.l.iv);
1205 type_size = pet_expr_get_type_size(expr_iv);
1206 init = pet_expr_copy(tree->u.l.init);
1207 init = pet_expr_new_binary(type_size, pet_op_assign, expr_iv, init);
1208 scop_init = scop_from_expr(init, state->n_stmt++,
1209 pet_tree_get_loc(tree), init_pc);
1210 scop_init = pet_scop_prefix(scop_init, declared);
1212 expr_iv = pet_expr_copy(tree->u.l.iv);
1213 type_size = pet_expr_get_type_size(expr_iv);
1214 inc = pet_expr_copy(tree->u.l.inc);
1215 inc = pet_expr_new_binary(type_size, pet_op_add_assign, expr_iv, inc);
1217 scop = scop_from_non_affine_while(pet_expr_copy(tree->u.l.cond),
1218 pet_tree_get_loc(tree), tree->u.l.body, inc,
1219 pet_context_copy(pc), state);
1221 scop = pet_scop_prefix(scop, declared + 1);
1222 scop = pet_scop_add_seq(state->ctx, scop_init, scop);
1224 pet_context_free(pc);
1226 if (!declared)
1227 return scop;
1229 array = extract_array(tree->u.l.iv, init_pc, state);
1230 if (array)
1231 array->declared = 1;
1232 scop_kill = kill(pet_tree_get_loc(tree), array, init_pc, state);
1233 scop_kill = pet_scop_prefix(scop_kill, 0);
1234 scop = pet_scop_add_seq(state->ctx, scop_kill, scop);
1235 scop_kill = kill(pet_tree_get_loc(tree), array, init_pc, state);
1236 scop_kill = pet_scop_add_array(scop_kill, array);
1237 scop_kill = pet_scop_prefix(scop_kill, 3);
1238 scop = pet_scop_add_seq(state->ctx, scop, scop_kill);
1240 return scop;
1243 /* Given an access expression "expr", is the variable accessed by
1244 * "expr" assigned anywhere inside "tree"?
1246 static int is_assigned(__isl_keep pet_expr *expr, __isl_keep pet_tree *tree)
1248 int assigned = 0;
1249 isl_id *id;
1251 id = pet_expr_access_get_id(expr);
1252 assigned = pet_tree_writes(tree, id);
1253 isl_id_free(id);
1255 return assigned;
1258 /* Are all nested access parameters in "pa" allowed given "tree".
1259 * In particular, is none of them written by anywhere inside "tree".
1261 * If "tree" has any continue or break nodes in the current loop level,
1262 * then no nested access parameters are allowed.
1263 * In particular, if there is any nested access in a guard
1264 * for a piece of code containing a "continue", then we want to introduce
1265 * a separate statement for evaluating this guard so that we can express
1266 * that the result is false for all previous iterations.
1268 static int is_nested_allowed(__isl_keep isl_pw_aff *pa,
1269 __isl_keep pet_tree *tree)
1271 int i, nparam;
1273 if (!tree)
1274 return -1;
1276 if (!pet_nested_any_in_pw_aff(pa))
1277 return 1;
1279 if (pet_tree_has_continue_or_break(tree))
1280 return 0;
1282 nparam = isl_pw_aff_dim(pa, isl_dim_param);
1283 for (i = 0; i < nparam; ++i) {
1284 isl_id *id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
1285 pet_expr *expr;
1286 int allowed;
1288 if (!pet_nested_in_id(id)) {
1289 isl_id_free(id);
1290 continue;
1293 expr = pet_nested_extract_expr(id);
1294 allowed = pet_expr_get_type(expr) == pet_expr_access &&
1295 !is_assigned(expr, tree);
1297 pet_expr_free(expr);
1298 isl_id_free(id);
1300 if (!allowed)
1301 return 0;
1304 return 1;
1307 /* Internal data structure for collect_local.
1308 * "pc" and "state" are needed to extract pet_arrays for the local variables.
1309 * "local" collects the results.
1311 struct pet_tree_collect_local_data {
1312 pet_context *pc;
1313 struct pet_state *state;
1314 isl_union_set *local;
1317 /* Add the variable accessed by "var" to data->local.
1318 * We extract a representation of the variable from
1319 * the pet_array constructed using extract_array
1320 * to ensure consistency with the rest of the scop.
1322 static int add_local(struct pet_tree_collect_local_data *data,
1323 __isl_keep pet_expr *var)
1325 struct pet_array *array;
1326 isl_set *universe;
1328 array = extract_array(var, data->pc, data->state);
1329 if (!array)
1330 return -1;
1332 universe = isl_set_universe(isl_set_get_space(array->extent));
1333 data->local = isl_union_set_add_set(data->local, universe);
1334 pet_array_free(array);
1336 return 0;
1339 /* If the node "tree" declares a variable, then add it to
1340 * data->local.
1342 static int extract_local_var(__isl_keep pet_tree *tree, void *user)
1344 enum pet_tree_type type;
1345 struct pet_tree_collect_local_data *data = user;
1347 type = pet_tree_get_type(tree);
1348 if (type == pet_tree_decl || type == pet_tree_decl_init)
1349 return add_local(data, tree->u.d.var);
1351 return 0;
1354 /* If the node "tree" is a for loop that declares its induction variable,
1355 * then add it this induction variable to data->local.
1357 static int extract_local_iterator(__isl_keep pet_tree *tree, void *user)
1359 struct pet_tree_collect_local_data *data = user;
1361 if (pet_tree_get_type(tree) == pet_tree_for && tree->u.l.declared)
1362 return add_local(data, tree->u.l.iv);
1364 return 0;
1367 /* Collect and return all local variables of the for loop represented
1368 * by "tree", with "scop" the corresponding pet_scop.
1369 * "pc" and "state" are needed to extract pet_arrays for the local variables.
1371 * We collect not only the variables that are declared inside "tree",
1372 * but also the loop iterators that are declared anywhere inside
1373 * any possible macro statements in "scop".
1374 * The latter also appear as declared variable in the scop,
1375 * whereas other declared loop iterators only appear implicitly
1376 * in the iteration domains.
1378 static __isl_give isl_union_set *collect_local(struct pet_scop *scop,
1379 __isl_keep pet_tree *tree, __isl_keep pet_context *pc,
1380 struct pet_state *state)
1382 int i;
1383 isl_ctx *ctx;
1384 struct pet_tree_collect_local_data data = { pc, state };
1386 ctx = pet_tree_get_ctx(tree);
1387 data.local = isl_union_set_empty(isl_space_params_alloc(ctx, 0));
1389 if (pet_tree_foreach_sub_tree(tree, &extract_local_var, &data) < 0)
1390 return isl_union_set_free(data.local);
1392 for (i = 0; i < scop->n_stmt; ++i) {
1393 pet_tree *body = scop->stmts[i]->body;
1394 if (pet_tree_foreach_sub_tree(body, &extract_local_iterator,
1395 &data) < 0)
1396 return isl_union_set_free(data.local);
1399 return data.local;
1402 /* Add an independence to "scop" if the for node "tree" was marked
1403 * independent.
1404 * "domain" is the set of loop iterators, with the current for loop
1405 * innermost. If "sign" is positive, then the inner iterator increases.
1406 * Otherwise it decreases.
1407 * "pc" and "state" are needed to extract pet_arrays for the local variables.
1409 * If the tree was marked, then collect all local variables and
1410 * add an independence.
1412 static struct pet_scop *set_independence(struct pet_scop *scop,
1413 __isl_keep pet_tree *tree, __isl_keep isl_set *domain, int sign,
1414 __isl_keep pet_context *pc, struct pet_state *state)
1416 isl_union_set *local;
1418 if (!tree->u.l.independent)
1419 return scop;
1421 local = collect_local(scop, tree, pc, state);
1422 scop = pet_scop_set_independent(scop, domain, local, sign);
1424 return scop;
1427 /* Construct a pet_scop for a for tree with static affine initialization
1428 * and constant increment within the context "pc".
1429 * The domain of "pc" has already been extended with an (at this point
1430 * unbounded) inner loop iterator corresponding to the current for loop.
1432 * The condition is allowed to contain nested accesses, provided
1433 * they are not being written to inside the body of the loop.
1434 * Otherwise, or if the condition is otherwise non-affine, the for loop is
1435 * essentially treated as a while loop, with iteration domain
1436 * { [l,i] : i >= init }, where l refers to the outer loop iterators.
1438 * We extract a pet_scop for the body after intersecting the domain of "pc"
1440 * { [l,i] : i >= init and condition' }
1442 * or
1444 * { [l,i] : i <= init and condition' }
1446 * Where condition' is equal to condition if the latter is
1447 * a simple upper [lower] bound and a condition that is extended
1448 * to apply to all previous iterations otherwise.
1449 * Afterwards, the schedule of the pet_scop is extended with
1451 * { [l,i] -> [i] }
1453 * or
1455 * { [l,i] -> [-i] }
1457 * If the condition is non-affine, then we drop the condition from the
1458 * iteration domain and instead create a separate statement
1459 * for evaluating the condition. The body is then filtered to depend
1460 * on the result of the condition evaluating to true on all iterations
1461 * up to the current iteration, while the evaluation the condition itself
1462 * is filtered to depend on the result of the condition evaluating to true
1463 * on all previous iterations.
1464 * The context of the scop representing the body is dropped
1465 * because we don't know how many times the body will be executed,
1466 * if at all.
1468 * If the stride of the loop is not 1, then "i >= init" is replaced by
1470 * (exists a: i = init + stride * a and a >= 0)
1472 * If the loop iterator i is unsigned, then wrapping may occur.
1473 * We therefore use a virtual iterator instead that does not wrap.
1474 * However, the condition in the code applies
1475 * to the wrapped value, so we need to change condition(l,i)
1476 * into condition([l,i % 2^width]). Similarly, we replace all accesses
1477 * to the original iterator by the wrapping of the virtual iterator.
1478 * Note that there may be no need to perform this final wrapping
1479 * if the loop condition (after wrapping) satisfies certain conditions.
1480 * However, the is_simple_bound condition is not enough since it doesn't
1481 * check if there even is an upper bound.
1483 * Wrapping on unsigned iterators can be avoided entirely if
1484 * loop condition is simple, the loop iterator is incremented
1485 * [decremented] by one and the last value before wrapping cannot
1486 * possibly satisfy the loop condition.
1488 * Valid outer iterators for a for loop are those for which the initial
1489 * value itself, the increment on each domain iteration and
1490 * the condition on both the initial value and
1491 * the result of incrementing the iterator for each iteration of the domain
1492 * can be evaluated.
1493 * If the loop condition is non-affine, then we only consider validity
1494 * of the initial value.
1496 * If the body contains any break, then we keep track of it in "skip"
1497 * (if the skip condition is affine) or it is handled in scop_add_break
1498 * (if the skip condition is not affine).
1499 * Note that the affine break condition needs to be considered with
1500 * respect to previous iterations in the virtual domain (if any).
1502 static struct pet_scop *scop_from_affine_for(__isl_keep pet_tree *tree,
1503 __isl_take isl_pw_aff *init_val, __isl_take isl_pw_aff *pa_inc,
1504 __isl_take isl_val *inc, __isl_take pet_context *pc,
1505 struct pet_state *state)
1507 isl_set *domain;
1508 isl_aff *sched;
1509 isl_set *cond = NULL;
1510 isl_set *skip = NULL;
1511 isl_id *id_test = NULL, *id_break_test;
1512 struct pet_scop *scop, *scop_cond = NULL;
1513 int pos;
1514 int is_one;
1515 int is_unsigned;
1516 int is_simple;
1517 int is_virtual;
1518 int is_non_affine;
1519 int has_affine_break;
1520 int has_var_break;
1521 isl_map *rev_wrap = NULL;
1522 isl_map *init_val_map;
1523 isl_pw_aff *pa;
1524 isl_set *valid_init;
1525 isl_set *valid_cond;
1526 isl_set *valid_cond_init;
1527 isl_set *valid_cond_next;
1528 isl_set *valid_inc;
1529 pet_expr *cond_expr;
1530 pet_context *pc_nested;
1532 pos = pet_context_dim(pc) - 1;
1534 domain = pet_context_get_domain(pc);
1535 cond_expr = pet_expr_copy(tree->u.l.cond);
1536 cond_expr = pet_context_evaluate_expr(pc, cond_expr);
1537 pc_nested = pet_context_copy(pc);
1538 pc_nested = pet_context_set_allow_nested(pc_nested, 1);
1539 pa = pet_expr_extract_affine_condition(cond_expr, pc_nested);
1540 pet_context_free(pc_nested);
1541 pet_expr_free(cond_expr);
1543 valid_inc = isl_pw_aff_domain(pa_inc);
1545 is_unsigned = pet_expr_get_type_size(tree->u.l.iv) > 0;
1547 is_non_affine = isl_pw_aff_involves_nan(pa) ||
1548 !is_nested_allowed(pa, tree->u.l.body);
1549 if (is_non_affine)
1550 pa = isl_pw_aff_free(pa);
1552 valid_cond = isl_pw_aff_domain(isl_pw_aff_copy(pa));
1553 cond = isl_pw_aff_non_zero_set(pa);
1554 if (is_non_affine)
1555 cond = isl_set_universe(isl_set_get_space(domain));
1557 valid_cond = isl_set_coalesce(valid_cond);
1558 is_one = isl_val_is_one(inc) || isl_val_is_negone(inc);
1559 is_virtual = is_unsigned &&
1560 (!is_one || can_wrap(cond, tree->u.l.iv, inc));
1562 init_val_map = isl_map_from_pw_aff(isl_pw_aff_copy(init_val));
1563 init_val_map = isl_map_equate(init_val_map, isl_dim_in, pos,
1564 isl_dim_out, 0);
1565 valid_cond_init = enforce_subset(isl_map_domain(init_val_map),
1566 isl_set_copy(valid_cond));
1567 if (is_one && !is_virtual) {
1568 isl_set *cond;
1570 isl_pw_aff_free(init_val);
1571 pa = pet_expr_extract_comparison(
1572 isl_val_is_pos(inc) ? pet_op_ge : pet_op_le,
1573 tree->u.l.iv, tree->u.l.init, pc);
1574 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(pa));
1575 valid_init = isl_set_eliminate(valid_init, isl_dim_set,
1576 isl_set_dim(domain, isl_dim_set) - 1, 1);
1577 cond = isl_pw_aff_non_zero_set(pa);
1578 domain = isl_set_intersect(domain, cond);
1579 } else {
1580 isl_set *strided;
1582 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(init_val));
1583 strided = strided_domain(init_val, isl_val_copy(inc));
1584 domain = isl_set_intersect(domain, strided);
1587 if (is_virtual) {
1588 isl_multi_aff *wrap;
1589 wrap = compute_wrapping(isl_set_get_space(cond), tree->u.l.iv);
1590 pc = pet_context_preimage_domain(pc, wrap);
1591 rev_wrap = isl_map_from_multi_aff(wrap);
1592 rev_wrap = isl_map_reverse(rev_wrap);
1593 cond = isl_set_apply(cond, isl_map_copy(rev_wrap));
1594 valid_cond = isl_set_apply(valid_cond, isl_map_copy(rev_wrap));
1595 valid_inc = isl_set_apply(valid_inc, isl_map_copy(rev_wrap));
1597 is_simple = is_simple_bound(cond, inc);
1598 if (!is_simple) {
1599 cond = isl_set_gist(cond, isl_set_copy(domain));
1600 is_simple = is_simple_bound(cond, inc);
1602 if (!is_simple)
1603 cond = valid_for_each_iteration(cond,
1604 isl_set_copy(domain), isl_val_copy(inc));
1605 cond = isl_set_align_params(cond, isl_set_get_space(domain));
1606 domain = isl_set_intersect(domain, cond);
1607 sched = map_to_last(pc);
1608 if (isl_val_is_neg(inc))
1609 sched = isl_aff_neg(sched);
1611 valid_cond_next = valid_on_next(valid_cond, isl_set_copy(domain),
1612 isl_val_copy(inc));
1613 valid_inc = enforce_subset(isl_set_copy(domain), valid_inc);
1615 pc = pet_context_intersect_domain(pc, isl_set_copy(domain));
1617 if (is_non_affine) {
1618 isl_space *space;
1619 isl_multi_pw_aff *test_index;
1620 space = isl_set_get_space(domain);
1621 test_index = pet_create_test_index(space, state->n_test++);
1622 scop_cond = scop_from_non_affine_condition(
1623 pet_expr_copy(tree->u.l.cond), state->n_stmt++,
1624 isl_multi_pw_aff_copy(test_index),
1625 pet_tree_get_loc(tree), pc);
1626 id_test = isl_multi_pw_aff_get_tuple_id(test_index,
1627 isl_dim_out);
1628 scop_cond = pet_scop_add_boolean_array(scop_cond,
1629 isl_set_copy(domain), test_index,
1630 state->int_size);
1631 scop_cond = pet_scop_prefix(scop_cond, 0);
1632 scop_cond = pet_scop_embed(scop_cond, isl_set_copy(domain),
1633 isl_aff_copy(sched));
1636 scop = scop_from_tree(tree->u.l.body, pc, state);
1637 has_affine_break = scop &&
1638 pet_scop_has_affine_skip(scop, pet_skip_later);
1639 if (has_affine_break)
1640 skip = pet_scop_get_affine_skip_domain(scop, pet_skip_later);
1641 has_var_break = scop && pet_scop_has_var_skip(scop, pet_skip_later);
1642 if (has_var_break)
1643 id_break_test = pet_scop_get_skip_id(scop, pet_skip_later);
1644 if (is_non_affine) {
1645 scop = pet_scop_reset_context(scop);
1646 scop = pet_scop_prefix(scop, 1);
1648 scop = pet_scop_embed(scop, isl_set_copy(domain), sched);
1649 scop = pet_scop_resolve_nested(scop);
1650 if (has_affine_break) {
1651 domain = apply_affine_break(domain, skip, isl_val_sgn(inc),
1652 is_virtual, rev_wrap);
1653 scop = pet_scop_intersect_domain_prefix(scop,
1654 isl_set_copy(domain));
1656 isl_map_free(rev_wrap);
1657 if (has_var_break)
1658 scop = scop_add_break(scop, id_break_test, isl_set_copy(domain),
1659 isl_val_copy(inc));
1660 if (is_non_affine) {
1661 scop = scop_add_while(scop_cond, scop, id_test, domain,
1662 isl_val_copy(inc));
1663 isl_set_free(valid_inc);
1664 } else {
1665 valid_inc = isl_set_intersect(valid_inc, valid_cond_next);
1666 valid_inc = isl_set_intersect(valid_inc, valid_cond_init);
1667 valid_inc = isl_set_project_out(valid_inc, isl_dim_set, pos, 1);
1668 scop = pet_scop_restrict_context(scop, valid_inc);
1669 scop = set_independence(scop, tree, domain, isl_val_sgn(inc),
1670 pc, state);
1671 isl_set_free(domain);
1674 isl_val_free(inc);
1676 valid_init = isl_set_project_out(valid_init, isl_dim_set, pos, 1);
1677 scop = pet_scop_restrict_context(scop, valid_init);
1679 pet_context_free(pc);
1680 return scop;
1683 /* Construct a pet_scop for a for statement within the context of "pc".
1685 * We update the context to reflect the writes to the loop variable and
1686 * the writes inside the body.
1688 * Then we check if the initialization of the for loop
1689 * is a static affine value and the increment is a constant.
1690 * If so, we construct the pet_scop using scop_from_affine_for.
1691 * Otherwise, we treat the for loop as a while loop
1692 * in scop_from_non_affine_for.
1694 * Note that the initialization and the increment are extracted
1695 * in a context where the current loop iterator has been added
1696 * to the context. If these turn out not be affine, then we
1697 * have reconstruct the body context without an assignment
1698 * to this loop iterator, as this variable will then not be
1699 * treated as a dimension of the iteration domain, but as any
1700 * other variable.
1702 static struct pet_scop *scop_from_for(__isl_keep pet_tree *tree,
1703 __isl_keep pet_context *init_pc, struct pet_state *state)
1705 isl_id *iv;
1706 isl_val *inc;
1707 isl_pw_aff *pa_inc, *init_val;
1708 pet_context *pc, *pc_init_val;
1710 if (!tree)
1711 return NULL;
1713 iv = pet_expr_access_get_id(tree->u.l.iv);
1714 pc = pet_context_copy(init_pc);
1715 pc = pet_context_add_inner_iterator(pc, iv);
1716 pc = pet_context_clear_writes_in_tree(pc, tree->u.l.body);
1718 pc_init_val = pet_context_copy(pc);
1719 pc_init_val = pet_context_clear_value(pc_init_val, isl_id_copy(iv));
1720 init_val = pet_expr_extract_affine(tree->u.l.init, pc_init_val);
1721 pet_context_free(pc_init_val);
1722 pa_inc = pet_expr_extract_affine(tree->u.l.inc, pc);
1723 inc = pet_extract_cst(pa_inc);
1724 if (!pa_inc || !init_val || !inc)
1725 goto error;
1726 if (!isl_pw_aff_involves_nan(pa_inc) &&
1727 !isl_pw_aff_involves_nan(init_val) && !isl_val_is_nan(inc))
1728 return scop_from_affine_for(tree, init_val, pa_inc, inc,
1729 pc, state);
1731 isl_pw_aff_free(pa_inc);
1732 isl_pw_aff_free(init_val);
1733 isl_val_free(inc);
1734 pet_context_free(pc);
1736 pc = pet_context_copy(init_pc);
1737 pc = pet_context_add_infinite_loop(pc);
1738 pc = pet_context_clear_writes_in_tree(pc, tree->u.l.body);
1739 return scop_from_non_affine_for(tree, init_pc, pc, state);
1740 error:
1741 isl_pw_aff_free(pa_inc);
1742 isl_pw_aff_free(init_val);
1743 isl_val_free(inc);
1744 pet_context_free(pc);
1745 return NULL;
1748 /* Check whether "expr" is an affine constraint within the context "pc".
1750 static int is_affine_condition(__isl_keep pet_expr *expr,
1751 __isl_keep pet_context *pc)
1753 isl_pw_aff *pa;
1754 int is_affine;
1756 pa = pet_expr_extract_affine_condition(expr, pc);
1757 if (!pa)
1758 return -1;
1759 is_affine = !isl_pw_aff_involves_nan(pa);
1760 isl_pw_aff_free(pa);
1762 return is_affine;
1765 /* Check if the given if statement is a conditional assignement
1766 * with a non-affine condition.
1768 * In particular we check if "stmt" is of the form
1770 * if (condition)
1771 * a = f(...);
1772 * else
1773 * a = g(...);
1775 * where the condition is non-affine and a is some array or scalar access.
1777 static int is_conditional_assignment(__isl_keep pet_tree *tree,
1778 __isl_keep pet_context *pc)
1780 int equal;
1781 isl_ctx *ctx;
1782 pet_expr *expr1, *expr2;
1784 ctx = pet_tree_get_ctx(tree);
1785 if (!pet_options_get_detect_conditional_assignment(ctx))
1786 return 0;
1787 if (tree->type != pet_tree_if_else)
1788 return 0;
1789 if (tree->u.i.then_body->type != pet_tree_expr)
1790 return 0;
1791 if (tree->u.i.else_body->type != pet_tree_expr)
1792 return 0;
1793 expr1 = tree->u.i.then_body->u.e.expr;
1794 expr2 = tree->u.i.else_body->u.e.expr;
1795 if (pet_expr_get_type(expr1) != pet_expr_op)
1796 return 0;
1797 if (pet_expr_get_type(expr2) != pet_expr_op)
1798 return 0;
1799 if (pet_expr_op_get_type(expr1) != pet_op_assign)
1800 return 0;
1801 if (pet_expr_op_get_type(expr2) != pet_op_assign)
1802 return 0;
1803 expr1 = pet_expr_get_arg(expr1, 0);
1804 expr2 = pet_expr_get_arg(expr2, 0);
1805 equal = pet_expr_is_equal(expr1, expr2);
1806 pet_expr_free(expr1);
1807 pet_expr_free(expr2);
1808 if (equal < 0 || !equal)
1809 return 0;
1810 if (is_affine_condition(tree->u.i.cond, pc))
1811 return 0;
1813 return 1;
1816 /* Given that "tree" is of the form
1818 * if (condition)
1819 * a = f(...);
1820 * else
1821 * a = g(...);
1823 * where a is some array or scalar access, construct a pet_scop
1824 * corresponding to this conditional assignment within the context "pc".
1825 * "cond_pa" is an affine expression with nested accesses representing
1826 * the condition.
1828 * The constructed pet_scop then corresponds to the expression
1830 * a = condition ? f(...) : g(...)
1832 * All access relations in f(...) are intersected with condition
1833 * while all access relation in g(...) are intersected with the complement.
1835 static struct pet_scop *scop_from_conditional_assignment(
1836 __isl_keep pet_tree *tree, __isl_take isl_pw_aff *cond_pa,
1837 __isl_take pet_context *pc, struct pet_state *state)
1839 int type_size;
1840 isl_set *cond, *comp;
1841 isl_multi_pw_aff *index;
1842 pet_expr *expr1, *expr2;
1843 pet_expr *pe_cond, *pe_then, *pe_else, *pe, *pe_write;
1844 struct pet_scop *scop;
1846 cond = isl_pw_aff_non_zero_set(isl_pw_aff_copy(cond_pa));
1847 comp = isl_pw_aff_zero_set(isl_pw_aff_copy(cond_pa));
1848 index = isl_multi_pw_aff_from_pw_aff(cond_pa);
1850 expr1 = tree->u.i.then_body->u.e.expr;
1851 expr2 = tree->u.i.else_body->u.e.expr;
1853 pe_cond = pet_expr_from_index(index);
1855 pe_then = pet_expr_get_arg(expr1, 1);
1856 pe_then = pet_context_evaluate_expr(pc, pe_then);
1857 pe_then = pet_expr_restrict(pe_then, cond);
1858 pe_else = pet_expr_get_arg(expr2, 1);
1859 pe_else = pet_context_evaluate_expr(pc, pe_else);
1860 pe_else = pet_expr_restrict(pe_else, comp);
1861 pe_write = pet_expr_get_arg(expr1, 0);
1862 pe_write = pet_context_evaluate_expr(pc, pe_write);
1864 pe = pet_expr_new_ternary(pe_cond, pe_then, pe_else);
1865 type_size = pet_expr_get_type_size(pe_write);
1866 pe = pet_expr_new_binary(type_size, pet_op_assign, pe_write, pe);
1868 scop = scop_from_evaluated_expr(pe, state->n_stmt++,
1869 pet_tree_get_loc(tree), pc);
1871 pet_context_free(pc);
1873 return scop;
1876 /* Construct a pet_scop for a non-affine if statement within the context "pc".
1878 * We create a separate statement that writes the result
1879 * of the non-affine condition to a virtual scalar.
1880 * A constraint requiring the value of this virtual scalar to be one
1881 * is added to the iteration domains of the then branch.
1882 * Similarly, a constraint requiring the value of this virtual scalar
1883 * to be zero is added to the iteration domains of the else branch, if any.
1884 * We adjust the schedules to ensure that the virtual scalar is written
1885 * before it is read.
1887 * If there are any breaks or continues in the then and/or else
1888 * branches, then we may have to compute a new skip condition.
1889 * This is handled using a pet_skip_info object.
1890 * On initialization, the object checks if skip conditions need
1891 * to be computed. If so, it does so in pet_skip_info_if_extract_index and
1892 * adds them in pet_skip_info_if_add.
1894 static struct pet_scop *scop_from_non_affine_if(__isl_keep pet_tree *tree,
1895 __isl_take pet_context *pc, struct pet_state *state)
1897 int has_else;
1898 isl_space *space;
1899 isl_set *domain;
1900 isl_multi_pw_aff *test_index;
1901 struct pet_skip_info skip;
1902 struct pet_scop *scop, *scop_then, *scop_else = NULL;
1904 has_else = tree->type == pet_tree_if_else;
1906 space = pet_context_get_space(pc);
1907 test_index = pet_create_test_index(space, state->n_test++);
1908 scop = scop_from_non_affine_condition(pet_expr_copy(tree->u.i.cond),
1909 state->n_stmt++, isl_multi_pw_aff_copy(test_index),
1910 pet_tree_get_loc(tree), pc);
1911 domain = pet_context_get_domain(pc);
1912 scop = pet_scop_add_boolean_array(scop, domain,
1913 isl_multi_pw_aff_copy(test_index), state->int_size);
1915 scop_then = scop_from_tree(tree->u.i.then_body, pc, state);
1916 if (has_else)
1917 scop_else = scop_from_tree(tree->u.i.else_body, pc, state);
1919 pet_skip_info_if_init(&skip, state->ctx, scop_then, scop_else,
1920 has_else, 0);
1921 pet_skip_info_if_extract_index(&skip, test_index, pc, state);
1923 scop = pet_scop_prefix(scop, 0);
1924 scop_then = pet_scop_prefix(scop_then, 1);
1925 scop_then = pet_scop_filter(scop_then,
1926 isl_multi_pw_aff_copy(test_index), 1);
1927 if (has_else) {
1928 scop_else = pet_scop_prefix(scop_else, 1);
1929 scop_else = pet_scop_filter(scop_else, test_index, 0);
1930 scop_then = pet_scop_add_par(state->ctx, scop_then, scop_else);
1931 } else
1932 isl_multi_pw_aff_free(test_index);
1934 scop = pet_scop_add_seq(state->ctx, scop, scop_then);
1936 scop = pet_skip_info_if_add(&skip, scop, 2);
1938 pet_context_free(pc);
1939 return scop;
1942 /* Construct a pet_scop for an affine if statement within the context "pc".
1944 * The condition is added to the iteration domains of the then branch,
1945 * while the opposite of the condition in added to the iteration domains
1946 * of the else branch, if any.
1948 * If there are any breaks or continues in the then and/or else
1949 * branches, then we may have to compute a new skip condition.
1950 * This is handled using a pet_skip_info_if object.
1951 * On initialization, the object checks if skip conditions need
1952 * to be computed. If so, it does so in pet_skip_info_if_extract_cond and
1953 * adds them in pet_skip_info_if_add.
1955 static struct pet_scop *scop_from_affine_if(__isl_keep pet_tree *tree,
1956 __isl_take isl_pw_aff *cond, __isl_take pet_context *pc,
1957 struct pet_state *state)
1959 int has_else;
1960 isl_ctx *ctx;
1961 isl_set *set, *complement;
1962 isl_set *valid;
1963 struct pet_skip_info skip;
1964 struct pet_scop *scop, *scop_then, *scop_else = NULL;
1965 pet_context *pc_body;
1967 ctx = pet_tree_get_ctx(tree);
1969 has_else = tree->type == pet_tree_if_else;
1971 valid = isl_pw_aff_domain(isl_pw_aff_copy(cond));
1972 set = isl_pw_aff_non_zero_set(isl_pw_aff_copy(cond));
1974 pc_body = pet_context_copy(pc);
1975 pc_body = pet_context_intersect_domain(pc_body, isl_set_copy(set));
1976 scop_then = scop_from_tree(tree->u.i.then_body, pc_body, state);
1977 pet_context_free(pc_body);
1978 if (has_else) {
1979 pc_body = pet_context_copy(pc);
1980 complement = isl_set_copy(valid);
1981 complement = isl_set_subtract(valid, isl_set_copy(set));
1982 pc_body = pet_context_intersect_domain(pc_body,
1983 isl_set_copy(complement));
1984 scop_else = scop_from_tree(tree->u.i.else_body, pc_body, state);
1985 pet_context_free(pc_body);
1988 pet_skip_info_if_init(&skip, ctx, scop_then, scop_else, has_else, 1);
1989 pet_skip_info_if_extract_cond(&skip, cond, pc, state);
1990 isl_pw_aff_free(cond);
1992 scop = pet_scop_restrict(scop_then, set);
1994 if (has_else) {
1995 scop_else = pet_scop_restrict(scop_else, complement);
1996 scop = pet_scop_add_par(ctx, scop, scop_else);
1998 scop = pet_scop_resolve_nested(scop);
1999 scop = pet_scop_restrict_context(scop, valid);
2001 if (pet_skip_info_has_skip(&skip))
2002 scop = pet_scop_prefix(scop, 0);
2003 scop = pet_skip_info_if_add(&skip, scop, 1);
2005 pet_context_free(pc);
2006 return scop;
2009 /* Construct a pet_scop for an if statement within the context "pc".
2011 * If the condition fits the pattern of a conditional assignment,
2012 * then it is handled by scop_from_conditional_assignment.
2013 * Note that the condition is only considered for a conditional assignment
2014 * if it is not static-affine. However, it should still convert
2015 * to an affine expression when nesting is allowed.
2017 * Otherwise, we check if the condition is affine.
2018 * If so, we construct the scop in scop_from_affine_if.
2019 * Otherwise, we construct the scop in scop_from_non_affine_if.
2021 * We allow the condition to be dynamic, i.e., to refer to
2022 * scalars or array elements that may be written to outside
2023 * of the given if statement. These nested accesses are then represented
2024 * as output dimensions in the wrapping iteration domain.
2025 * If it is also written _inside_ the then or else branch, then
2026 * we treat the condition as non-affine.
2027 * As explained in extract_non_affine_if, this will introduce
2028 * an extra statement.
2029 * For aesthetic reasons, we want this statement to have a statement
2030 * number that is lower than those of the then and else branches.
2031 * In order to evaluate if we will need such a statement, however, we
2032 * first construct scops for the then and else branches.
2033 * We therefore reserve a statement number if we might have to
2034 * introduce such an extra statement.
2036 static struct pet_scop *scop_from_if(__isl_keep pet_tree *tree,
2037 __isl_keep pet_context *pc, struct pet_state *state)
2039 int has_else;
2040 isl_pw_aff *cond;
2041 pet_expr *cond_expr;
2042 pet_context *pc_nested;
2044 if (!tree)
2045 return NULL;
2047 has_else = tree->type == pet_tree_if_else;
2049 pc = pet_context_copy(pc);
2050 pc = pet_context_clear_writes_in_tree(pc, tree->u.i.then_body);
2051 if (has_else)
2052 pc = pet_context_clear_writes_in_tree(pc, tree->u.i.else_body);
2054 cond_expr = pet_expr_copy(tree->u.i.cond);
2055 cond_expr = pet_context_evaluate_expr(pc, cond_expr);
2056 pc_nested = pet_context_copy(pc);
2057 pc_nested = pet_context_set_allow_nested(pc_nested, 1);
2058 cond = pet_expr_extract_affine_condition(cond_expr, pc_nested);
2059 pet_context_free(pc_nested);
2060 pet_expr_free(cond_expr);
2062 if (!cond) {
2063 pet_context_free(pc);
2064 return NULL;
2067 if (isl_pw_aff_involves_nan(cond)) {
2068 isl_pw_aff_free(cond);
2069 return scop_from_non_affine_if(tree, pc, state);
2072 if (is_conditional_assignment(tree, pc))
2073 return scop_from_conditional_assignment(tree, cond, pc, state);
2075 if ((!is_nested_allowed(cond, tree->u.i.then_body) ||
2076 (has_else && !is_nested_allowed(cond, tree->u.i.else_body)))) {
2077 isl_pw_aff_free(cond);
2078 return scop_from_non_affine_if(tree, pc, state);
2081 return scop_from_affine_if(tree, cond, pc, state);
2084 /* Return a one-dimensional multi piecewise affine expression that is equal
2085 * to the constant 1 and is defined over the given domain.
2087 static __isl_give isl_multi_pw_aff *one_mpa(__isl_take isl_space *space)
2089 isl_local_space *ls;
2090 isl_aff *aff;
2092 ls = isl_local_space_from_space(space);
2093 aff = isl_aff_zero_on_domain(ls);
2094 aff = isl_aff_set_constant_si(aff, 1);
2096 return isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff));
2099 /* Construct a pet_scop for a continue statement with the given domain space.
2101 * We simply create an empty scop with a universal pet_skip_now
2102 * skip condition. This skip condition will then be taken into
2103 * account by the enclosing loop construct, possibly after
2104 * being incorporated into outer skip conditions.
2106 static struct pet_scop *scop_from_continue(__isl_keep pet_tree *tree,
2107 __isl_take isl_space *space)
2109 struct pet_scop *scop;
2111 scop = pet_scop_empty(isl_space_copy(space));
2113 scop = pet_scop_set_skip(scop, pet_skip_now, one_mpa(space));
2115 return scop;
2118 /* Construct a pet_scop for a break statement with the given domain space.
2120 * We simply create an empty scop with both a universal pet_skip_now
2121 * skip condition and a universal pet_skip_later skip condition.
2122 * These skip conditions will then be taken into
2123 * account by the enclosing loop construct, possibly after
2124 * being incorporated into outer skip conditions.
2126 static struct pet_scop *scop_from_break(__isl_keep pet_tree *tree,
2127 __isl_take isl_space *space)
2129 struct pet_scop *scop;
2130 isl_multi_pw_aff *skip;
2132 scop = pet_scop_empty(isl_space_copy(space));
2134 skip = one_mpa(space);
2135 scop = pet_scop_set_skip(scop, pet_skip_now,
2136 isl_multi_pw_aff_copy(skip));
2137 scop = pet_scop_set_skip(scop, pet_skip_later, skip);
2139 return scop;
2142 /* Extract a clone of the kill statement in "scop".
2143 * The domain of the clone is given by "domain".
2144 * "scop" is expected to have been created from a DeclStmt
2145 * and should have the kill as its first statement.
2147 static struct pet_scop *extract_kill(__isl_keep isl_set *domain,
2148 struct pet_scop *scop, struct pet_state *state)
2150 pet_expr *kill;
2151 struct pet_stmt *stmt;
2152 isl_space *space;
2153 isl_multi_pw_aff *mpa;
2154 pet_tree *tree;
2156 if (!domain || !scop)
2157 return NULL;
2158 if (scop->n_stmt < 1)
2159 isl_die(isl_set_get_ctx(domain), isl_error_internal,
2160 "expecting at least one statement", return NULL);
2161 stmt = scop->stmts[0];
2162 if (!pet_stmt_is_kill(stmt))
2163 isl_die(isl_set_get_ctx(domain), isl_error_internal,
2164 "expecting kill statement", return NULL);
2166 kill = pet_tree_expr_get_expr(stmt->body);
2167 space = pet_stmt_get_space(stmt);
2168 space = isl_space_map_from_set(space);
2169 mpa = isl_multi_pw_aff_identity(space);
2170 mpa = isl_multi_pw_aff_reset_tuple_id(mpa, isl_dim_in);
2171 kill = pet_expr_update_domain(kill, mpa);
2172 tree = pet_tree_new_expr(kill);
2173 tree = pet_tree_set_loc(tree, pet_loc_copy(stmt->loc));
2174 stmt = pet_stmt_from_pet_tree(isl_set_copy(domain),
2175 state->n_stmt++, tree);
2176 return pet_scop_from_pet_stmt(isl_set_get_space(domain), stmt);
2179 /* Does "tree" represent an assignment to a variable?
2181 * The assignment may be one of
2182 * - a declaration with initialization
2183 * - an expression with a top-level assignment operator
2185 static int is_assignment(__isl_keep pet_tree *tree)
2187 if (!tree)
2188 return 0;
2189 if (tree->type == pet_tree_decl_init)
2190 return 1;
2191 return pet_tree_is_assign(tree);
2194 /* Update "pc" by taking into account the assignment performed by "tree",
2195 * where "tree" satisfies is_assignment.
2197 * In particular, if the lhs of the assignment is a scalar variable and
2198 * if the rhs is an affine expression, then keep track of this value in "pc"
2199 * so that we can plug it in when we later come across the same variable.
2201 * Any previously assigned value to the variable has already been removed
2202 * by scop_handle_writes.
2204 static __isl_give pet_context *handle_assignment(__isl_take pet_context *pc,
2205 __isl_keep pet_tree *tree)
2207 pet_expr *var, *val;
2208 isl_id *id;
2209 isl_pw_aff *pa;
2211 if (pet_tree_get_type(tree) == pet_tree_decl_init) {
2212 var = pet_tree_decl_get_var(tree);
2213 val = pet_tree_decl_get_init(tree);
2214 } else {
2215 pet_expr *expr;
2216 expr = pet_tree_expr_get_expr(tree);
2217 var = pet_expr_get_arg(expr, 0);
2218 val = pet_expr_get_arg(expr, 1);
2219 pet_expr_free(expr);
2222 if (!pet_expr_is_scalar_access(var)) {
2223 pet_expr_free(var);
2224 pet_expr_free(val);
2225 return pc;
2228 pa = pet_expr_extract_affine(val, pc);
2229 if (!pa)
2230 pc = pet_context_free(pc);
2232 if (!isl_pw_aff_involves_nan(pa)) {
2233 id = pet_expr_access_get_id(var);
2234 pc = pet_context_set_value(pc, id, pa);
2235 } else {
2236 isl_pw_aff_free(pa);
2238 pet_expr_free(var);
2239 pet_expr_free(val);
2241 return pc;
2244 /* Mark all arrays in "scop" as being exposed.
2246 static struct pet_scop *mark_exposed(struct pet_scop *scop)
2248 int i;
2250 if (!scop)
2251 return NULL;
2252 for (i = 0; i < scop->n_array; ++i)
2253 scop->arrays[i]->exposed = 1;
2254 return scop;
2257 /* Try and construct a pet_scop corresponding to (part of)
2258 * a sequence of statements within the context "pc".
2260 * After extracting a statement, we update "pc"
2261 * based on the top-level assignments in the statement
2262 * so that we can exploit them in subsequent statements in the same block.
2264 * If there are any breaks or continues in the individual statements,
2265 * then we may have to compute a new skip condition.
2266 * This is handled using a pet_skip_info object.
2267 * On initialization, the object checks if skip conditions need
2268 * to be computed. If so, it does so in pet_skip_info_seq_extract and
2269 * adds them in pet_skip_info_seq_add.
2271 * If "block" is set, then we need to insert kill statements at
2272 * the end of the block for any array that has been declared by
2273 * one of the statements in the sequence. Each of these declarations
2274 * results in the construction of a kill statement at the place
2275 * of the declaration, so we simply collect duplicates of
2276 * those kill statements and append these duplicates to the constructed scop.
2278 * If "block" is not set, then any array declared by one of the statements
2279 * in the sequence is marked as being exposed.
2281 * If autodetect is set, then we allow the extraction of only a subrange
2282 * of the sequence of statements. However, if there is at least one statement
2283 * for which we could not construct a scop and the final range contains
2284 * either no statements or at least one kill, then we discard the entire
2285 * range.
2287 static struct pet_scop *scop_from_block(__isl_keep pet_tree *tree,
2288 __isl_keep pet_context *pc, struct pet_state *state)
2290 int i;
2291 isl_ctx *ctx;
2292 isl_space *space;
2293 isl_set *domain;
2294 struct pet_scop *scop, *kills;
2296 ctx = pet_tree_get_ctx(tree);
2298 space = pet_context_get_space(pc);
2299 domain = pet_context_get_domain(pc);
2300 pc = pet_context_copy(pc);
2301 scop = pet_scop_empty(isl_space_copy(space));
2302 kills = pet_scop_empty(space);
2303 for (i = 0; i < tree->u.b.n; ++i) {
2304 struct pet_scop *scop_i;
2306 if (pet_scop_has_affine_skip(scop, pet_skip_now))
2307 pc = apply_affine_continue(pc, scop);
2308 scop_i = scop_from_tree(tree->u.b.child[i], pc, state);
2309 pc = scop_handle_writes(scop_i, pc);
2310 if (is_assignment(tree->u.b.child[i]))
2311 pc = handle_assignment(pc, tree->u.b.child[i]);
2312 struct pet_skip_info skip;
2313 pet_skip_info_seq_init(&skip, ctx, scop, scop_i);
2314 pet_skip_info_seq_extract(&skip, pc, state);
2315 if (pet_skip_info_has_skip(&skip))
2316 scop_i = pet_scop_prefix(scop_i, 0);
2317 if (scop_i && pet_tree_is_decl(tree->u.b.child[i])) {
2318 if (tree->u.b.block) {
2319 struct pet_scop *kill;
2320 kill = extract_kill(domain, scop_i, state);
2321 kills = pet_scop_add_par(ctx, kills, kill);
2322 } else
2323 scop_i = mark_exposed(scop_i);
2325 scop_i = pet_scop_prefix(scop_i, i);
2326 scop = pet_scop_add_seq(ctx, scop, scop_i);
2328 scop = pet_skip_info_seq_add(&skip, scop, i);
2330 if (!scop)
2331 break;
2333 isl_set_free(domain);
2335 kills = pet_scop_prefix(kills, tree->u.b.n);
2336 scop = pet_scop_add_seq(ctx, scop, kills);
2338 pet_context_free(pc);
2340 return scop;
2343 /* Internal data structure for extract_declared_arrays.
2345 * "pc" and "state" are used to create pet_array objects and kill statements.
2346 * "any" is initialized to 0 by the caller and set to 1 as soon as we have
2347 * found any declared array.
2348 * "scop" has been initialized by the caller and is used to attach
2349 * the created pet_array objects.
2350 * "kill_before" and "kill_after" are created and updated by
2351 * extract_declared_arrays to collect the kills of the arrays.
2353 struct pet_tree_extract_declared_arrays_data {
2354 pet_context *pc;
2355 struct pet_state *state;
2357 isl_ctx *ctx;
2359 int any;
2360 struct pet_scop *scop;
2361 struct pet_scop *kill_before;
2362 struct pet_scop *kill_after;
2365 /* Check if the node "node" declares any array or scalar.
2366 * If so, create the corresponding pet_array and attach it to data->scop.
2367 * Additionally, create two kill statements for the array and add them
2368 * to data->kill_before and data->kill_after.
2370 static int extract_declared_arrays(__isl_keep pet_tree *node, void *user)
2372 enum pet_tree_type type;
2373 struct pet_tree_extract_declared_arrays_data *data = user;
2374 struct pet_array *array;
2375 struct pet_scop *scop_kill;
2376 pet_expr *var;
2378 type = pet_tree_get_type(node);
2379 if (type == pet_tree_decl || type == pet_tree_decl_init)
2380 var = node->u.d.var;
2381 else if (type == pet_tree_for && node->u.l.declared)
2382 var = node->u.l.iv;
2383 else
2384 return 0;
2386 array = extract_array(var, data->pc, data->state);
2387 if (array)
2388 array->declared = 1;
2389 data->scop = pet_scop_add_array(data->scop, array);
2391 scop_kill = kill(pet_tree_get_loc(node), array, data->pc, data->state);
2392 if (!data->any)
2393 data->kill_before = scop_kill;
2394 else
2395 data->kill_before = pet_scop_add_par(data->ctx,
2396 data->kill_before, scop_kill);
2398 scop_kill = kill(pet_tree_get_loc(node), array, data->pc, data->state);
2399 if (!data->any)
2400 data->kill_after = scop_kill;
2401 else
2402 data->kill_after = pet_scop_add_par(data->ctx,
2403 data->kill_after, scop_kill);
2405 data->any = 1;
2407 return 0;
2410 /* Convert a pet_tree that consists of more than a single leaf
2411 * to a pet_scop with a single statement encapsulating the entire pet_tree.
2412 * Do so within the context of "pc".
2414 * After constructing the core scop, we also look for any arrays (or scalars)
2415 * that are declared inside "tree". Each of those arrays is marked as
2416 * having been declared and kill statements for these arrays
2417 * are introduced before and after the core scop.
2418 * Note that the input tree is not a leaf so that the declaration
2419 * cannot occur at the outer level.
2421 static struct pet_scop *scop_from_tree_macro(__isl_take pet_tree *tree,
2422 __isl_take isl_id *label, __isl_keep pet_context *pc,
2423 struct pet_state *state)
2425 struct pet_tree_extract_declared_arrays_data data = { pc, state };
2427 data.scop = scop_from_unevaluated_tree(pet_tree_copy(tree),
2428 state->n_stmt++, pc);
2430 data.any = 0;
2431 data.ctx = pet_context_get_ctx(pc);
2432 if (pet_tree_foreach_sub_tree(tree, &extract_declared_arrays,
2433 &data) < 0)
2434 data.scop = pet_scop_free(data.scop);
2435 pet_tree_free(tree);
2437 if (!data.any)
2438 return data.scop;
2440 data.kill_before = pet_scop_prefix(data.kill_before, 0);
2441 data.scop = pet_scop_prefix(data.scop, 1);
2442 data.kill_after = pet_scop_prefix(data.kill_after, 2);
2444 data.scop = pet_scop_add_seq(data.ctx, data.kill_before, data.scop);
2445 data.scop = pet_scop_add_seq(data.ctx, data.scop, data.kill_after);
2447 return data.scop;
2450 /* Construct a pet_scop that corresponds to the pet_tree "tree"
2451 * within the context "pc" by calling the appropriate function
2452 * based on the type of "tree".
2454 * If the initially constructed pet_scop turns out to involve
2455 * dynamic control and if the user has requested an encapsulation
2456 * of all dynamic control, then this pet_scop is discarded and
2457 * a new pet_scop is created with a single statement representing
2458 * the entire "tree".
2459 * However, if the scop contains any active continue or break,
2460 * then we need to include the loop containing the continue or break
2461 * in the encapsulation. We therefore postpone the encapsulation
2462 * until we have constructed a pet_scop for this enclosing loop.
2464 static struct pet_scop *scop_from_tree(__isl_keep pet_tree *tree,
2465 __isl_keep pet_context *pc, struct pet_state *state)
2467 isl_ctx *ctx;
2468 struct pet_scop *scop = NULL;
2470 if (!tree)
2471 return NULL;
2473 ctx = pet_tree_get_ctx(tree);
2474 switch (tree->type) {
2475 case pet_tree_error:
2476 return NULL;
2477 case pet_tree_block:
2478 return scop_from_block(tree, pc, state);
2479 case pet_tree_break:
2480 return scop_from_break(tree, pet_context_get_space(pc));
2481 case pet_tree_continue:
2482 return scop_from_continue(tree, pet_context_get_space(pc));
2483 case pet_tree_decl:
2484 case pet_tree_decl_init:
2485 return scop_from_decl(tree, pc, state);
2486 case pet_tree_expr:
2487 return scop_from_tree_expr(tree, pc, state);
2488 case pet_tree_if:
2489 case pet_tree_if_else:
2490 scop = scop_from_if(tree, pc, state);
2491 break;
2492 case pet_tree_for:
2493 scop = scop_from_for(tree, pc, state);
2494 break;
2495 case pet_tree_while:
2496 scop = scop_from_while(tree, pc, state);
2497 break;
2498 case pet_tree_infinite_loop:
2499 scop = scop_from_infinite_for(tree, pc, state);
2500 break;
2503 if (!scop)
2504 return NULL;
2506 if (!pet_options_get_encapsulate_dynamic_control(ctx) ||
2507 !pet_scop_has_data_dependent_conditions(scop) ||
2508 pet_scop_has_var_skip(scop, pet_skip_now))
2509 return scop;
2511 pet_scop_free(scop);
2512 return scop_from_tree_macro(pet_tree_copy(tree),
2513 isl_id_copy(tree->label), pc, state);
2516 /* If "tree" has a label that is of the form S_<nr>, then make
2517 * sure that state->n_stmt is greater than nr to ensure that
2518 * we will not generate S_<nr> ourselves.
2520 static int set_first_stmt(__isl_keep pet_tree *tree, void *user)
2522 struct pet_state *state = user;
2523 const char *name;
2524 int nr;
2526 if (!tree)
2527 return -1;
2528 if (!tree->label)
2529 return 0;
2530 name = isl_id_get_name(tree->label);
2531 if (strncmp(name, "S_", 2) != 0)
2532 return 0;
2533 nr = atoi(name + 2);
2534 if (nr >= state->n_stmt)
2535 state->n_stmt = nr + 1;
2537 return 0;
2540 /* Construct a pet_scop that corresponds to the pet_tree "tree".
2541 * "int_size" is the number of bytes need to represent an integer.
2542 * "extract_array" is a callback that we can use to create a pet_array
2543 * that corresponds to the variable accessed by an expression.
2545 * Initialize the global state, construct a context and then
2546 * construct the pet_scop by recursively visiting the tree.
2548 * state.n_stmt is initialized to point beyond any explicit S_<nr> label.
2550 struct pet_scop *pet_scop_from_pet_tree(__isl_take pet_tree *tree, int int_size,
2551 struct pet_array *(*extract_array)(__isl_keep pet_expr *access,
2552 __isl_keep pet_context *pc, void *user), void *user,
2553 __isl_keep pet_context *pc)
2555 struct pet_scop *scop;
2556 struct pet_state state = { 0 };
2558 if (!tree)
2559 return NULL;
2561 state.ctx = pet_tree_get_ctx(tree);
2562 state.int_size = int_size;
2563 state.extract_array = extract_array;
2564 state.user = user;
2565 if (pet_tree_foreach_sub_tree(tree, &set_first_stmt, &state) < 0)
2566 tree = pet_tree_free(tree);
2568 scop = scop_from_tree(tree, pc, &state);
2569 scop = pet_scop_set_loc(scop, pet_tree_get_loc(tree));
2571 pet_tree_free(tree);
2573 if (scop)
2574 scop->context = isl_set_params(scop->context);
2576 return scop;