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
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
35 #include <isl/id_to_pw_aff.h>
44 #include "tree2scop.h"
46 /* Update "pc" by taking into account the writes in "stmt".
47 * That is, clear any previously assigned values to variables
48 * that are written by "stmt".
50 static __isl_give pet_context
*handle_writes(struct pet_stmt
*stmt
,
51 __isl_take pet_context
*pc
)
53 return pet_context_clear_writes_in_tree(pc
, stmt
->body
);
56 /* Update "pc" based on the write accesses in "scop".
58 static __isl_give pet_context
*scop_handle_writes(struct pet_scop
*scop
,
59 __isl_take pet_context
*pc
)
64 return pet_context_free(pc
);
65 for (i
= 0; i
< scop
->n_stmt
; ++i
)
66 pc
= handle_writes(scop
->stmts
[i
], pc
);
71 /* Wrapper around pet_expr_resolve_assume
72 * for use as a callback to pet_tree_map_expr.
74 static __isl_give pet_expr
*resolve_assume(__isl_take pet_expr
*expr
,
77 pet_context
*pc
= user
;
79 return pet_expr_resolve_assume(expr
, pc
);
82 /* Check if any expression inside "tree" is an assume expression and
83 * if its single argument can be converted to an affine expression
84 * in the context of "pc".
85 * If so, replace the argument by the affine expression.
87 __isl_give pet_tree
*pet_tree_resolve_assume(__isl_take pet_tree
*tree
,
88 __isl_keep pet_context
*pc
)
90 return pet_tree_map_expr(tree
, &resolve_assume
, pc
);
93 /* Convert a pet_tree to a pet_scop with one statement within the context "pc".
94 * "tree" has already been evaluated in the context of "pc".
95 * This mainly involves resolving nested expression parameters
96 * and setting the name of the iteration space.
97 * The name is given by tree->label if it is non-NULL. Otherwise,
98 * it is of the form S_<stmt_nr>.
100 static struct pet_scop
*scop_from_evaluated_tree(__isl_take pet_tree
*tree
,
101 int stmt_nr
, __isl_keep pet_context
*pc
)
107 space
= pet_context_get_space(pc
);
109 tree
= pet_tree_resolve_nested(tree
, space
);
110 tree
= pet_tree_resolve_assume(tree
, pc
);
112 domain
= pet_context_get_domain(pc
);
113 ps
= pet_stmt_from_pet_tree(domain
, stmt_nr
, tree
);
114 return pet_scop_from_pet_stmt(space
, ps
);
117 /* Convert a top-level pet_expr to a pet_scop with one statement
118 * within the context "pc".
119 * "expr" has already been evaluated in the context of "pc".
120 * We construct a pet_tree from "expr" and continue with
121 * scop_from_evaluated_tree.
122 * The name is of the form S_<stmt_nr>.
123 * The location of the statement is set to "loc".
125 static struct pet_scop
*scop_from_evaluated_expr(__isl_take pet_expr
*expr
,
126 int stmt_nr
, __isl_take pet_loc
*loc
, __isl_keep pet_context
*pc
)
130 tree
= pet_tree_new_expr(expr
);
131 tree
= pet_tree_set_loc(tree
, loc
);
132 return scop_from_evaluated_tree(tree
, stmt_nr
, pc
);
135 /* Convert a pet_tree to a pet_scop with one statement within the context "pc".
136 * "tree" has not yet been evaluated in the context of "pc".
137 * We evaluate "tree" in the context of "pc" and continue with
138 * scop_from_evaluated_tree.
139 * The statement name is given by tree->label if it is non-NULL. Otherwise,
140 * it is of the form S_<stmt_nr>.
142 static struct pet_scop
*scop_from_unevaluated_tree(__isl_take pet_tree
*tree
,
143 int stmt_nr
, __isl_keep pet_context
*pc
)
145 tree
= pet_context_evaluate_tree(pc
, tree
);
146 return scop_from_evaluated_tree(tree
, stmt_nr
, pc
);
149 /* Convert a top-level pet_expr to a pet_scop with one statement
150 * within the context "pc", where "expr" has not yet been evaluated
151 * in the context of "pc".
152 * We construct a pet_tree from "expr" and continue with
153 * scop_from_unevaluated_tree.
154 * The statement name is of the form S_<stmt_nr>.
155 * The location of the statement is set to "loc".
157 static struct pet_scop
*scop_from_expr(__isl_take pet_expr
*expr
,
158 int stmt_nr
, __isl_take pet_loc
*loc
, __isl_keep pet_context
*pc
)
162 tree
= pet_tree_new_expr(expr
);
163 tree
= pet_tree_set_loc(tree
, loc
);
164 return scop_from_unevaluated_tree(tree
, stmt_nr
, pc
);
167 /* Construct a pet_scop with a single statement killing the entire
169 * The location of the statement is set to "loc".
171 static struct pet_scop
*kill(__isl_take pet_loc
*loc
, struct pet_array
*array
,
172 __isl_keep pet_context
*pc
, struct pet_state
*state
)
177 isl_multi_pw_aff
*index
;
180 struct pet_scop
*scop
;
184 ctx
= isl_set_get_ctx(array
->extent
);
185 access
= isl_map_from_range(isl_set_copy(array
->extent
));
186 id
= isl_set_get_tuple_id(array
->extent
);
187 space
= isl_space_alloc(ctx
, 0, 0, 0);
188 space
= isl_space_set_tuple_id(space
, isl_dim_out
, id
);
189 index
= isl_multi_pw_aff_zero(space
);
190 expr
= pet_expr_kill_from_access_and_index(access
, index
);
191 return scop_from_expr(expr
, state
->n_stmt
++, loc
, pc
);
197 /* Construct and return a pet_array corresponding to the variable
198 * accessed by "access" by calling the extract_array callback.
200 static struct pet_array
*extract_array(__isl_keep pet_expr
*access
,
201 __isl_keep pet_context
*pc
, struct pet_state
*state
)
203 return state
->extract_array(access
, pc
, state
->user
);
206 /* Construct a pet_scop for a (single) variable declaration
207 * within the context "pc".
209 * The scop contains the variable being declared (as an array)
210 * and a statement killing the array.
212 * If the declaration comes with an initialization, then the scop
213 * also contains an assignment to the variable.
215 static struct pet_scop
*scop_from_decl(__isl_keep pet_tree
*tree
,
216 __isl_keep pet_context
*pc
, struct pet_state
*state
)
220 struct pet_array
*array
;
221 struct pet_scop
*scop_decl
, *scop
;
222 pet_expr
*lhs
, *rhs
, *pe
;
224 array
= extract_array(tree
->u
.d
.var
, pc
, state
);
227 scop_decl
= kill(pet_tree_get_loc(tree
), array
, pc
, state
);
228 scop_decl
= pet_scop_add_array(scop_decl
, array
);
230 if (tree
->type
!= pet_tree_decl_init
)
233 lhs
= pet_expr_copy(tree
->u
.d
.var
);
234 rhs
= pet_expr_copy(tree
->u
.d
.init
);
235 type_size
= pet_expr_get_type_size(lhs
);
236 pe
= pet_expr_new_binary(type_size
, pet_op_assign
, lhs
, rhs
);
237 scop
= scop_from_expr(pe
, state
->n_stmt
++, pet_tree_get_loc(tree
), pc
);
239 scop_decl
= pet_scop_prefix(scop_decl
, 0);
240 scop
= pet_scop_prefix(scop
, 1);
242 ctx
= pet_tree_get_ctx(tree
);
243 scop
= pet_scop_add_seq(ctx
, scop_decl
, scop
);
248 /* Return those elements in the space of "cond" that come after
249 * (based on "sign") an element in "cond" in the final dimension.
251 static __isl_give isl_set
*after(__isl_take isl_set
*cond
, int sign
)
254 isl_map
*previous_to_this
;
257 dim
= isl_set_dim(cond
, isl_dim_set
);
258 space
= isl_space_map_from_set(isl_set_get_space(cond
));
259 previous_to_this
= isl_map_universe(space
);
260 for (i
= 0; i
+ 1 < dim
; ++i
)
261 previous_to_this
= isl_map_equate(previous_to_this
,
262 isl_dim_in
, i
, isl_dim_out
, i
);
264 previous_to_this
= isl_map_order_lt(previous_to_this
,
265 isl_dim_in
, dim
- 1, isl_dim_out
, dim
- 1);
267 previous_to_this
= isl_map_order_gt(previous_to_this
,
268 isl_dim_in
, dim
- 1, isl_dim_out
, dim
- 1);
270 cond
= isl_set_apply(cond
, previous_to_this
);
275 /* Remove those iterations of "domain" that have an earlier iteration
276 * (based on "sign") in the final dimension where "skip" is satisfied.
277 * If "apply_skip_map" is set, then "skip_map" is first applied
278 * to the embedded skip condition before removing it from the domain.
280 static __isl_give isl_set
*apply_affine_break(__isl_take isl_set
*domain
,
281 __isl_take isl_set
*skip
, int sign
,
282 int apply_skip_map
, __isl_keep isl_map
*skip_map
)
285 skip
= isl_set_apply(skip
, isl_map_copy(skip_map
));
286 skip
= isl_set_intersect(skip
, isl_set_copy(domain
));
287 return isl_set_subtract(domain
, after(skip
, sign
));
290 /* Create an affine expression on the domain space of "pc" that
291 * is equal to the final dimension of this domain.
293 static __isl_give isl_aff
*map_to_last(__isl_keep pet_context
*pc
)
299 space
= pet_context_get_space(pc
);
300 pos
= isl_space_dim(space
, isl_dim_set
) - 1;
301 ls
= isl_local_space_from_space(space
);
302 return isl_aff_var_on_domain(ls
, isl_dim_set
, pos
);
305 /* Create an affine expression that maps elements
306 * of an array "id_test" to the previous element in the final dimension
307 * (according to "inc"), provided this element belongs to "domain".
308 * That is, create the affine expression
310 * { id[outer,x] -> id[outer,x - inc] : (outer,x - inc) in domain }
312 static __isl_give isl_multi_pw_aff
*map_to_previous(__isl_take isl_id
*id_test
,
313 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
320 isl_multi_pw_aff
*prev
;
322 pos
= isl_set_dim(domain
, isl_dim_set
) - 1;
323 space
= isl_set_get_space(domain
);
324 space
= isl_space_map_from_set(space
);
325 ma
= isl_multi_aff_identity(space
);
326 aff
= isl_multi_aff_get_aff(ma
, pos
);
327 aff
= isl_aff_add_constant_val(aff
, isl_val_neg(inc
));
328 ma
= isl_multi_aff_set_aff(ma
, pos
, aff
);
329 domain
= isl_set_preimage_multi_aff(domain
, isl_multi_aff_copy(ma
));
330 prev
= isl_multi_pw_aff_from_multi_aff(ma
);
331 pa
= isl_multi_pw_aff_get_pw_aff(prev
, pos
);
332 pa
= isl_pw_aff_intersect_domain(pa
, domain
);
333 prev
= isl_multi_pw_aff_set_pw_aff(prev
, pos
, pa
);
334 prev
= isl_multi_pw_aff_set_tuple_id(prev
, isl_dim_out
, id_test
);
339 /* Add an implication to "scop" expressing that if an element of
340 * virtual array "id_test" has value "satisfied" then all previous elements
341 * of this array (in the final dimension) also have that value.
342 * The set of previous elements is bounded by "domain".
343 * If "sign" is negative then the iterator
344 * is decreasing and we express that all subsequent array elements
345 * (but still defined previously) have the same value.
347 static struct pet_scop
*add_implication(struct pet_scop
*scop
,
348 __isl_take isl_id
*id_test
, __isl_take isl_set
*domain
, int sign
,
355 dim
= isl_set_dim(domain
, isl_dim_set
);
356 domain
= isl_set_set_tuple_id(domain
, id_test
);
357 space
= isl_space_map_from_set(isl_set_get_space(domain
));
358 map
= isl_map_universe(space
);
359 for (i
= 0; i
+ 1 < dim
; ++i
)
360 map
= isl_map_equate(map
, isl_dim_in
, i
, isl_dim_out
, i
);
362 map
= isl_map_order_ge(map
,
363 isl_dim_in
, dim
- 1, isl_dim_out
, dim
- 1);
365 map
= isl_map_order_le(map
,
366 isl_dim_in
, dim
- 1, isl_dim_out
, dim
- 1);
367 map
= isl_map_intersect_range(map
, domain
);
368 scop
= pet_scop_add_implication(scop
, map
, satisfied
);
373 /* Add a filter to "scop" that imposes that it is only executed
374 * when the variable identified by "id_test" has a zero value
375 * for all previous iterations of "domain".
377 * In particular, add a filter that imposes that the array
378 * has a zero value at the previous iteration of domain and
379 * add an implication that implies that it then has that
380 * value for all previous iterations.
382 static struct pet_scop
*scop_add_break(struct pet_scop
*scop
,
383 __isl_take isl_id
*id_test
, __isl_take isl_set
*domain
,
384 __isl_take isl_val
*inc
)
386 isl_multi_pw_aff
*prev
;
387 int sign
= isl_val_sgn(inc
);
389 prev
= map_to_previous(isl_id_copy(id_test
), isl_set_copy(domain
), inc
);
390 scop
= add_implication(scop
, id_test
, domain
, sign
, 0);
391 scop
= pet_scop_filter(scop
, prev
, 0);
396 static struct pet_scop
*scop_from_tree(__isl_keep pet_tree
*tree
,
397 __isl_keep pet_context
*pc
, struct pet_state
*state
);
399 /* Construct a pet_scop for an infinite loop around the given body
400 * within the context "pc".
402 * The domain of "pc" has already been extended with an infinite loop
406 * We extract a pet_scop for the body and then embed it in a loop with
409 * { [outer,t] -> [t] }
411 * If the body contains any break, then it is taken into
412 * account in apply_affine_break (if the skip condition is affine)
413 * or in scop_add_break (if the skip condition is not affine).
415 * Note that in case of an affine skip condition,
416 * since we are dealing with a loop without loop iterator,
417 * the skip condition cannot refer to the current loop iterator and
418 * so effectively, the effect on the iteration domain is of the form
420 * { [outer,0]; [outer,t] : t >= 1 and not skip }
422 static struct pet_scop
*scop_from_infinite_loop(__isl_keep pet_tree
*body
,
423 __isl_keep pet_context
*pc
, struct pet_state
*state
)
430 struct pet_scop
*scop
;
431 int has_affine_break
;
434 ctx
= pet_tree_get_ctx(body
);
435 domain
= pet_context_get_domain(pc
);
436 sched
= map_to_last(pc
);
438 scop
= scop_from_tree(body
, pc
, state
);
440 has_affine_break
= pet_scop_has_affine_skip(scop
, pet_skip_later
);
441 if (has_affine_break
)
442 skip
= pet_scop_get_affine_skip_domain(scop
, pet_skip_later
);
443 has_var_break
= pet_scop_has_var_skip(scop
, pet_skip_later
);
445 id_test
= pet_scop_get_skip_id(scop
, pet_skip_later
);
447 scop
= pet_scop_embed(scop
, isl_set_copy(domain
), sched
);
448 if (has_affine_break
) {
449 domain
= apply_affine_break(domain
, skip
, 1, 0, NULL
);
450 scop
= pet_scop_intersect_domain_prefix(scop
,
451 isl_set_copy(domain
));
454 scop
= scop_add_break(scop
, id_test
, domain
, isl_val_one(ctx
));
456 isl_set_free(domain
);
461 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
466 * within the context "pc".
468 * Extend the domain of "pc" with an extra inner loop
472 * and construct the scop in scop_from_infinite_loop.
474 static struct pet_scop
*scop_from_infinite_for(__isl_keep pet_tree
*tree
,
475 __isl_keep pet_context
*pc
, struct pet_state
*state
)
477 struct pet_scop
*scop
;
479 pc
= pet_context_copy(pc
);
480 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.l
.body
);
482 pc
= pet_context_add_infinite_loop(pc
);
484 scop
= scop_from_infinite_loop(tree
->u
.l
.body
, pc
, state
);
486 pet_context_free(pc
);
491 /* Construct a pet_scop for a while loop of the form
496 * within the context "pc".
498 * The domain of "pc" has already been extended with an infinite loop
502 * Here, we add the constraints on the outer loop iterators
503 * implied by "pa" and construct the scop in scop_from_infinite_loop.
504 * Note that the intersection with these constraints
505 * may result in an empty loop.
507 static struct pet_scop
*scop_from_affine_while(__isl_keep pet_tree
*tree
,
508 __isl_take isl_pw_aff
*pa
, __isl_take pet_context
*pc
,
509 struct pet_state
*state
)
511 struct pet_scop
*scop
;
512 isl_set
*dom
, *local
;
515 valid
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
516 dom
= isl_pw_aff_non_zero_set(pa
);
517 local
= isl_set_add_dims(isl_set_copy(dom
), isl_dim_set
, 1);
518 pc
= pet_context_intersect_domain(pc
, local
);
519 scop
= scop_from_infinite_loop(tree
->u
.l
.body
, pc
, state
);
520 scop
= pet_scop_restrict(scop
, dom
);
521 scop
= pet_scop_restrict_context(scop
, valid
);
523 pet_context_free(pc
);
527 /* Construct a scop for a while, given the scops for the condition
528 * and the body, the filter identifier and the iteration domain of
531 * In particular, the scop for the condition is filtered to depend
532 * on "id_test" evaluating to true for all previous iterations
533 * of the loop, while the scop for the body is filtered to depend
534 * on "id_test" evaluating to true for all iterations up to the
536 * The actual filter only imposes that this virtual array has
537 * value one on the previous or the current iteration.
538 * The fact that this condition also applies to the previous
539 * iterations is enforced by an implication.
541 * These filtered scops are then combined into a single scop.
543 * "sign" is positive if the iterator increases and negative
546 static struct pet_scop
*scop_add_while(struct pet_scop
*scop_cond
,
547 struct pet_scop
*scop_body
, __isl_take isl_id
*id_test
,
548 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
550 isl_ctx
*ctx
= isl_set_get_ctx(domain
);
552 isl_multi_pw_aff
*test_index
;
553 isl_multi_pw_aff
*prev
;
554 int sign
= isl_val_sgn(inc
);
555 struct pet_scop
*scop
;
557 prev
= map_to_previous(isl_id_copy(id_test
), isl_set_copy(domain
), inc
);
558 scop_cond
= pet_scop_filter(scop_cond
, prev
, 1);
560 space
= isl_space_map_from_set(isl_set_get_space(domain
));
561 test_index
= isl_multi_pw_aff_identity(space
);
562 test_index
= isl_multi_pw_aff_set_tuple_id(test_index
, isl_dim_out
,
563 isl_id_copy(id_test
));
564 scop_body
= pet_scop_filter(scop_body
, test_index
, 1);
566 scop
= pet_scop_add_seq(ctx
, scop_cond
, scop_body
);
567 scop
= add_implication(scop
, id_test
, domain
, sign
, 1);
572 /* Create a pet_scop with a single statement with name S_<stmt_nr>,
573 * evaluating "cond" and writing the result to a virtual scalar,
574 * as expressed by "index".
575 * The expression "cond" has not yet been evaluated in the context of "pc".
576 * Do so within the context "pc".
577 * The location of the statement is set to "loc".
579 static struct pet_scop
*scop_from_non_affine_condition(
580 __isl_take pet_expr
*cond
, int stmt_nr
,
581 __isl_take isl_multi_pw_aff
*index
,
582 __isl_take pet_loc
*loc
, __isl_keep pet_context
*pc
)
584 pet_expr
*expr
, *write
;
586 cond
= pet_context_evaluate_expr(pc
, cond
);
588 write
= pet_expr_from_index(index
);
589 write
= pet_expr_access_set_write(write
, 1);
590 write
= pet_expr_access_set_read(write
, 0);
591 expr
= pet_expr_new_binary(1, pet_op_assign
, write
, cond
);
593 return scop_from_evaluated_expr(expr
, stmt_nr
, loc
, pc
);
596 /* Construct a generic while scop, with iteration domain
597 * { [t] : t >= 0 } around the scop for "tree_body" within the context "pc".
598 * The domain of "pc" has already been extended with this infinite loop
602 * The scop consists of two parts,
603 * one for evaluating the condition "cond" and one for the body.
604 * If "expr_inc" is not NULL, then a scop for evaluating this expression
605 * is added at the end of the body,
606 * after replacing any skip conditions resulting from continue statements
607 * by the skip conditions resulting from break statements (if any).
609 * The schedule is adjusted to reflect that the condition is evaluated
610 * before the body is executed and the body is filtered to depend
611 * on the result of the condition evaluating to true on all iterations
612 * up to the current iteration, while the evaluation of the condition itself
613 * is filtered to depend on the result of the condition evaluating to true
614 * on all previous iterations.
615 * The context of the scop representing the body is dropped
616 * because we don't know how many times the body will be executed,
619 * If the body contains any break, then it is taken into
620 * account in apply_affine_break (if the skip condition is affine)
621 * or in scop_add_break (if the skip condition is not affine).
623 * Note that in case of an affine skip condition,
624 * since we are dealing with a loop without loop iterator,
625 * the skip condition cannot refer to the current loop iterator and
626 * so effectively, the effect on the iteration domain is of the form
628 * { [outer,0]; [outer,t] : t >= 1 and not skip }
630 static struct pet_scop
*scop_from_non_affine_while(__isl_take pet_expr
*cond
,
631 __isl_take pet_loc
*loc
, __isl_keep pet_tree
*tree_body
,
632 __isl_take pet_expr
*expr_inc
, __isl_take pet_context
*pc
,
633 struct pet_state
*state
)
636 isl_id
*id_test
, *id_break_test
;
638 isl_multi_pw_aff
*test_index
;
642 struct pet_scop
*scop
, *scop_body
;
643 int has_affine_break
;
647 space
= pet_context_get_space(pc
);
648 test_index
= pet_create_test_index(space
, state
->n_test
++);
649 scop
= scop_from_non_affine_condition(cond
, state
->n_stmt
++,
650 isl_multi_pw_aff_copy(test_index
),
651 pet_loc_copy(loc
), pc
);
652 id_test
= isl_multi_pw_aff_get_tuple_id(test_index
, isl_dim_out
);
653 domain
= pet_context_get_domain(pc
);
654 scop
= pet_scop_add_boolean_array(scop
, isl_set_copy(domain
),
655 test_index
, state
->int_size
);
657 sched
= map_to_last(pc
);
659 scop_body
= scop_from_tree(tree_body
, pc
, state
);
661 has_affine_break
= pet_scop_has_affine_skip(scop_body
, pet_skip_later
);
662 if (has_affine_break
)
663 skip
= pet_scop_get_affine_skip_domain(scop_body
,
665 has_var_break
= pet_scop_has_var_skip(scop_body
, pet_skip_later
);
667 id_break_test
= pet_scop_get_skip_id(scop_body
, pet_skip_later
);
669 scop
= pet_scop_prefix(scop
, 0);
670 scop
= pet_scop_embed(scop
, isl_set_copy(domain
), isl_aff_copy(sched
));
671 scop_body
= pet_scop_reset_context(scop_body
);
672 scop_body
= pet_scop_prefix(scop_body
, 1);
674 struct pet_scop
*scop_inc
;
675 scop_inc
= scop_from_expr(expr_inc
, state
->n_stmt
++, loc
, pc
);
676 scop_inc
= pet_scop_prefix(scop_inc
, 2);
677 if (pet_scop_has_skip(scop_body
, pet_skip_later
)) {
678 isl_multi_pw_aff
*skip
;
679 skip
= pet_scop_get_skip(scop_body
, pet_skip_later
);
680 scop_body
= pet_scop_set_skip(scop_body
,
683 pet_scop_reset_skip(scop_body
, pet_skip_now
);
684 scop_body
= pet_scop_add_seq(ctx
, scop_body
, scop_inc
);
687 scop_body
= pet_scop_embed(scop_body
, isl_set_copy(domain
), sched
);
689 if (has_affine_break
) {
690 domain
= apply_affine_break(domain
, skip
, 1, 0, NULL
);
691 scop
= pet_scop_intersect_domain_prefix(scop
,
692 isl_set_copy(domain
));
693 scop_body
= pet_scop_intersect_domain_prefix(scop_body
,
694 isl_set_copy(domain
));
697 scop
= scop_add_break(scop
, isl_id_copy(id_break_test
),
698 isl_set_copy(domain
), isl_val_one(ctx
));
699 scop_body
= scop_add_break(scop_body
, id_break_test
,
700 isl_set_copy(domain
), isl_val_one(ctx
));
702 scop
= scop_add_while(scop
, scop_body
, id_test
, domain
,
705 pet_context_free(pc
);
709 /* Check if the while loop is of the form
711 * while (affine expression)
714 * If so, call scop_from_affine_while to construct a scop.
716 * Otherwise, pass control to scop_from_non_affine_while.
718 * "pc" is the context in which the affine expressions in the scop are created.
719 * The domain of "pc" is extended with an infinite loop
723 * before passing control to scop_from_affine_while or
724 * scop_from_non_affine_while.
726 static struct pet_scop
*scop_from_while(__isl_keep pet_tree
*tree
,
727 __isl_keep pet_context
*pc
, struct pet_state
*state
)
735 pc
= pet_context_copy(pc
);
736 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.l
.body
);
738 cond_expr
= pet_expr_copy(tree
->u
.l
.cond
);
739 cond_expr
= pet_context_evaluate_expr(pc
, cond_expr
);
740 pa
= pet_expr_extract_affine_condition(cond_expr
, pc
);
741 pet_expr_free(cond_expr
);
743 pc
= pet_context_add_infinite_loop(pc
);
748 if (!isl_pw_aff_involves_nan(pa
))
749 return scop_from_affine_while(tree
, pa
, pc
, state
);
751 return scop_from_non_affine_while(pet_expr_copy(tree
->u
.l
.cond
),
752 pet_tree_get_loc(tree
), tree
->u
.l
.body
, NULL
,
755 pet_context_free(pc
);
759 /* Check whether "cond" expresses a simple loop bound
760 * on the final set dimension.
761 * In particular, if "up" is set then "cond" should contain only
762 * upper bounds on the final set dimension.
763 * Otherwise, it should contain only lower bounds.
765 static int is_simple_bound(__isl_keep isl_set
*cond
, __isl_keep isl_val
*inc
)
769 pos
= isl_set_dim(cond
, isl_dim_set
) - 1;
770 if (isl_val_is_pos(inc
))
771 return !isl_set_dim_has_any_lower_bound(cond
, isl_dim_set
, pos
);
773 return !isl_set_dim_has_any_upper_bound(cond
, isl_dim_set
, pos
);
776 /* Extend a condition on a given iteration of a loop to one that
777 * imposes the same condition on all previous iterations.
778 * "domain" expresses the lower [upper] bound on the iterations
779 * when inc is positive [negative] in its final dimension.
781 * In particular, we construct the condition (when inc is positive)
783 * forall i' : (domain(i') and i' <= i) => cond(i')
785 * (where "<=" applies to the final dimension)
786 * which is equivalent to
788 * not exists i' : domain(i') and i' <= i and not cond(i')
790 * We construct this set by subtracting the satisfying cond from domain,
793 * { [i'] -> [i] : i' <= i }
795 * and then subtracting the result from domain again.
797 static __isl_give isl_set
*valid_for_each_iteration(__isl_take isl_set
*cond
,
798 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
801 isl_map
*previous_to_this
;
804 dim
= isl_set_dim(cond
, isl_dim_set
);
805 space
= isl_space_map_from_set(isl_set_get_space(cond
));
806 previous_to_this
= isl_map_universe(space
);
807 for (i
= 0; i
+ 1 < dim
; ++i
)
808 previous_to_this
= isl_map_equate(previous_to_this
,
809 isl_dim_in
, i
, isl_dim_out
, i
);
810 if (isl_val_is_pos(inc
))
811 previous_to_this
= isl_map_order_le(previous_to_this
,
812 isl_dim_in
, dim
- 1, isl_dim_out
, dim
- 1);
814 previous_to_this
= isl_map_order_ge(previous_to_this
,
815 isl_dim_in
, dim
- 1, isl_dim_out
, dim
- 1);
817 cond
= isl_set_subtract(isl_set_copy(domain
), cond
);
818 cond
= isl_set_apply(cond
, previous_to_this
);
819 cond
= isl_set_subtract(domain
, cond
);
826 /* Given an initial value of the form
828 * { [outer,i] -> init(outer) }
830 * construct a domain of the form
832 * { [outer,i] : exists a: i = init(outer) + a * inc and a >= 0 }
834 static __isl_give isl_set
*strided_domain(__isl_take isl_pw_aff
*init
,
835 __isl_take isl_val
*inc
)
843 dim
= isl_pw_aff_dim(init
, isl_dim_in
);
845 init
= isl_pw_aff_add_dims(init
, isl_dim_in
, 1);
846 space
= isl_pw_aff_get_domain_space(init
);
847 ls
= isl_local_space_from_space(space
);
848 aff
= isl_aff_zero_on_domain(isl_local_space_copy(ls
));
849 aff
= isl_aff_add_coefficient_val(aff
, isl_dim_in
, dim
, inc
);
850 init
= isl_pw_aff_add(init
, isl_pw_aff_from_aff(aff
));
852 aff
= isl_aff_var_on_domain(ls
, isl_dim_set
, dim
- 1);
853 set
= isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff
), init
);
855 set
= isl_set_lower_bound_si(set
, isl_dim_set
, dim
, 0);
856 set
= isl_set_project_out(set
, isl_dim_set
, dim
, 1);
861 /* Assuming "cond" represents a bound on a loop where the loop
862 * iterator "iv" is incremented (or decremented) by one, check if wrapping
865 * Under the given assumptions, wrapping is only possible if "cond" allows
866 * for the last value before wrapping, i.e., 2^width - 1 in case of an
867 * increasing iterator and 0 in case of a decreasing iterator.
869 static int can_wrap(__isl_keep isl_set
*cond
, __isl_keep pet_expr
*iv
,
870 __isl_keep isl_val
*inc
)
877 test
= isl_set_copy(cond
);
879 ctx
= isl_set_get_ctx(test
);
880 if (isl_val_is_neg(inc
))
881 limit
= isl_val_zero(ctx
);
883 limit
= isl_val_int_from_ui(ctx
, pet_expr_get_type_size(iv
));
884 limit
= isl_val_2exp(limit
);
885 limit
= isl_val_sub_ui(limit
, 1);
888 test
= isl_set_fix_val(cond
, isl_dim_set
, 0, limit
);
889 cw
= !isl_set_is_empty(test
);
899 * construct the following affine expression on this space
901 * { [outer, v] -> [outer, v mod 2^width] }
903 * where width is the number of bits used to represent the values
904 * of the unsigned variable "iv".
906 static __isl_give isl_multi_aff
*compute_wrapping(__isl_take isl_space
*space
,
907 __isl_keep pet_expr
*iv
)
915 dim
= isl_space_dim(space
, isl_dim_set
);
917 ctx
= isl_space_get_ctx(space
);
918 mod
= isl_val_int_from_ui(ctx
, pet_expr_get_type_size(iv
));
919 mod
= isl_val_2exp(mod
);
921 space
= isl_space_map_from_set(space
);
922 ma
= isl_multi_aff_identity(space
);
924 aff
= isl_multi_aff_get_aff(ma
, dim
- 1);
925 aff
= isl_aff_mod_val(aff
, mod
);
926 ma
= isl_multi_aff_set_aff(ma
, dim
- 1, aff
);
931 /* Given two sets in the space
935 * where l represents the outer loop iterators, compute the set
936 * of values of l that ensure that "set1" is a subset of "set2".
938 * set1 is a subset of set2 if
940 * forall i: set1(l,i) => set2(l,i)
944 * not exists i: set1(l,i) and not set2(l,i)
948 * not exists i: (set1 \ set2)(l,i)
950 static __isl_give isl_set
*enforce_subset(__isl_take isl_set
*set1
,
951 __isl_take isl_set
*set2
)
955 pos
= isl_set_dim(set1
, isl_dim_set
) - 1;
956 set1
= isl_set_subtract(set1
, set2
);
957 set1
= isl_set_eliminate(set1
, isl_dim_set
, pos
, 1);
958 return isl_set_complement(set1
);
961 /* Compute the set of outer iterator values for which "cond" holds
962 * on the next iteration of the inner loop for each element of "dom".
964 * We first construct mapping { [l,i] -> [l,i + inc] } (where l refers
965 * to the outer loop iterators), plug that into "cond"
966 * and then compute the set of outer iterators for which "dom" is a subset
969 static __isl_give isl_set
*valid_on_next(__isl_take isl_set
*cond
,
970 __isl_take isl_set
*dom
, __isl_take isl_val
*inc
)
977 pos
= isl_set_dim(dom
, isl_dim_set
) - 1;
978 space
= isl_set_get_space(dom
);
979 space
= isl_space_map_from_set(space
);
980 ma
= isl_multi_aff_identity(space
);
981 aff
= isl_multi_aff_get_aff(ma
, pos
);
982 aff
= isl_aff_add_constant_val(aff
, inc
);
983 ma
= isl_multi_aff_set_aff(ma
, pos
, aff
);
984 cond
= isl_set_preimage_multi_aff(cond
, ma
);
986 return enforce_subset(dom
, cond
);
989 /* Extract the for loop "tree" as a while loop within the context "pc_init".
990 * In particular, "pc_init" represents the context of the loop,
991 * whereas "pc" represents the context of the body of the loop and
992 * has already had its domain extended with an infinite loop
996 * The for loop has the form
998 * for (iv = init; cond; iv += inc)
1009 * except that the skips resulting from any continue statements
1010 * in body do not apply to the increment, but are replaced by the skips
1011 * resulting from break statements.
1013 * If the loop iterator is declared in the for loop, then it is killed before
1014 * and after the loop.
1016 static struct pet_scop
*scop_from_non_affine_for(__isl_keep pet_tree
*tree
,
1017 __isl_keep pet_context
*init_pc
, __isl_take pet_context
*pc
,
1018 struct pet_state
*state
)
1022 pet_expr
*expr_iv
, *init
, *inc
;
1023 struct pet_scop
*scop_init
, *scop
;
1025 struct pet_array
*array
;
1026 struct pet_scop
*scop_kill
;
1028 iv
= pet_expr_access_get_id(tree
->u
.l
.iv
);
1029 pc
= pet_context_clear_value(pc
, iv
);
1031 declared
= tree
->u
.l
.declared
;
1033 expr_iv
= pet_expr_copy(tree
->u
.l
.iv
);
1034 type_size
= pet_expr_get_type_size(expr_iv
);
1035 init
= pet_expr_copy(tree
->u
.l
.init
);
1036 init
= pet_expr_new_binary(type_size
, pet_op_assign
, expr_iv
, init
);
1037 scop_init
= scop_from_expr(init
, state
->n_stmt
++,
1038 pet_tree_get_loc(tree
), init_pc
);
1039 scop_init
= pet_scop_prefix(scop_init
, declared
);
1041 expr_iv
= pet_expr_copy(tree
->u
.l
.iv
);
1042 type_size
= pet_expr_get_type_size(expr_iv
);
1043 inc
= pet_expr_copy(tree
->u
.l
.inc
);
1044 inc
= pet_expr_new_binary(type_size
, pet_op_add_assign
, expr_iv
, inc
);
1046 scop
= scop_from_non_affine_while(pet_expr_copy(tree
->u
.l
.cond
),
1047 pet_tree_get_loc(tree
), tree
->u
.l
.body
, inc
,
1048 pet_context_copy(pc
), state
);
1050 scop
= pet_scop_prefix(scop
, declared
+ 1);
1051 scop
= pet_scop_add_seq(state
->ctx
, scop_init
, scop
);
1053 pet_context_free(pc
);
1058 array
= extract_array(tree
->u
.l
.iv
, init_pc
, state
);
1060 array
->declared
= 1;
1061 scop_kill
= kill(pet_tree_get_loc(tree
), array
, init_pc
, state
);
1062 scop_kill
= pet_scop_prefix(scop_kill
, 0);
1063 scop
= pet_scop_add_seq(state
->ctx
, scop_kill
, scop
);
1064 scop_kill
= kill(pet_tree_get_loc(tree
), array
, init_pc
, state
);
1065 scop_kill
= pet_scop_add_array(scop_kill
, array
);
1066 scop_kill
= pet_scop_prefix(scop_kill
, 3);
1067 scop
= pet_scop_add_seq(state
->ctx
, scop
, scop_kill
);
1072 /* Given an access expression "expr", is the variable accessed by
1073 * "expr" assigned anywhere inside "tree"?
1075 static int is_assigned(__isl_keep pet_expr
*expr
, __isl_keep pet_tree
*tree
)
1080 id
= pet_expr_access_get_id(expr
);
1081 assigned
= pet_tree_writes(tree
, id
);
1087 /* Are all nested access parameters in "pa" allowed given "tree".
1088 * In particular, is none of them written by anywhere inside "tree".
1090 * If "tree" has any continue nodes in the current loop level,
1091 * then no nested access parameters are allowed.
1092 * In particular, if there is any nested access in a guard
1093 * for a piece of code containing a "continue", then we want to introduce
1094 * a separate statement for evaluating this guard so that we can express
1095 * that the result is false for all previous iterations.
1097 static int is_nested_allowed(__isl_keep isl_pw_aff
*pa
,
1098 __isl_keep pet_tree
*tree
)
1105 if (!pet_nested_any_in_pw_aff(pa
))
1108 if (pet_tree_has_continue(tree
))
1111 nparam
= isl_pw_aff_dim(pa
, isl_dim_param
);
1112 for (i
= 0; i
< nparam
; ++i
) {
1113 isl_id
*id
= isl_pw_aff_get_dim_id(pa
, isl_dim_param
, i
);
1117 if (!pet_nested_in_id(id
)) {
1122 expr
= pet_nested_extract_expr(id
);
1123 allowed
= pet_expr_get_type(expr
) == pet_expr_access
&&
1124 !is_assigned(expr
, tree
);
1126 pet_expr_free(expr
);
1136 /* Internal data structure for collect_local.
1137 * "pc" and "state" are needed to extract pet_arrays for the local variables.
1138 * "local" collects the results.
1140 struct pet_tree_collect_local_data
{
1142 struct pet_state
*state
;
1143 isl_union_set
*local
;
1146 /* Add the variable accessed by "var" to data->local.
1147 * We extract a representation of the variable from
1148 * the pet_array constructed using extract_array
1149 * to ensure consistency with the rest of the scop.
1151 static int add_local(struct pet_tree_collect_local_data
*data
,
1152 __isl_keep pet_expr
*var
)
1154 struct pet_array
*array
;
1157 array
= extract_array(var
, data
->pc
, data
->state
);
1161 universe
= isl_set_universe(isl_set_get_space(array
->extent
));
1162 data
->local
= isl_union_set_add_set(data
->local
, universe
);
1163 pet_array_free(array
);
1168 /* If the node "tree" declares a variable, then add it to
1171 static int extract_local_var(__isl_keep pet_tree
*tree
, void *user
)
1173 enum pet_tree_type type
;
1174 struct pet_tree_collect_local_data
*data
= user
;
1176 type
= pet_tree_get_type(tree
);
1177 if (type
== pet_tree_decl
|| type
== pet_tree_decl_init
)
1178 return add_local(data
, tree
->u
.d
.var
);
1183 /* If the node "tree" is a for loop that declares its induction variable,
1184 * then add it this induction variable to data->local.
1186 static int extract_local_iterator(__isl_keep pet_tree
*tree
, void *user
)
1188 struct pet_tree_collect_local_data
*data
= user
;
1190 if (pet_tree_get_type(tree
) == pet_tree_for
&& tree
->u
.l
.declared
)
1191 return add_local(data
, tree
->u
.l
.iv
);
1196 /* Collect and return all local variables of the for loop represented
1197 * by "tree", with "scop" the corresponding pet_scop.
1198 * "pc" and "state" are needed to extract pet_arrays for the local variables.
1200 * We collect not only the variables that are declared inside "tree",
1201 * but also the loop iterators that are declared anywhere inside
1202 * any possible macro statements in "scop".
1203 * The latter also appear as declared variable in the scop,
1204 * whereas other declared loop iterators only appear implicitly
1205 * in the iteration domains.
1207 static __isl_give isl_union_set
*collect_local(struct pet_scop
*scop
,
1208 __isl_keep pet_tree
*tree
, __isl_keep pet_context
*pc
,
1209 struct pet_state
*state
)
1213 struct pet_tree_collect_local_data data
= { pc
, state
};
1215 ctx
= pet_tree_get_ctx(tree
);
1216 data
.local
= isl_union_set_empty(isl_space_params_alloc(ctx
, 0));
1218 if (pet_tree_foreach_sub_tree(tree
, &extract_local_var
, &data
) < 0)
1219 return isl_union_set_free(data
.local
);
1221 for (i
= 0; i
< scop
->n_stmt
; ++i
) {
1222 pet_tree
*body
= scop
->stmts
[i
]->body
;
1223 if (pet_tree_foreach_sub_tree(body
, &extract_local_iterator
,
1225 return isl_union_set_free(data
.local
);
1231 /* Add an independence to "scop" if the for node "tree" was marked
1233 * "domain" is the set of loop iterators, with the current for loop
1234 * innermost. If "sign" is positive, then the inner iterator increases.
1235 * Otherwise it decreases.
1236 * "pc" and "state" are needed to extract pet_arrays for the local variables.
1238 * If the tree was marked, then collect all local variables and
1239 * add an independence.
1241 static struct pet_scop
*set_independence(struct pet_scop
*scop
,
1242 __isl_keep pet_tree
*tree
, __isl_keep isl_set
*domain
, int sign
,
1243 __isl_keep pet_context
*pc
, struct pet_state
*state
)
1245 isl_union_set
*local
;
1247 if (!tree
->u
.l
.independent
)
1250 local
= collect_local(scop
, tree
, pc
, state
);
1251 scop
= pet_scop_set_independent(scop
, domain
, local
, sign
);
1256 /* Construct a pet_scop for a for tree with static affine initialization
1257 * and constant increment within the context "pc".
1258 * The domain of "pc" has already been extended with an (at this point
1259 * unbounded) inner loop iterator corresponding to the current for loop.
1261 * The condition is allowed to contain nested accesses, provided
1262 * they are not being written to inside the body of the loop.
1263 * Otherwise, or if the condition is otherwise non-affine, the for loop is
1264 * essentially treated as a while loop, with iteration domain
1265 * { [l,i] : i >= init }, where l refers to the outer loop iterators.
1267 * We extract a pet_scop for the body after intersecting the domain of "pc"
1269 * { [l,i] : i >= init and condition' }
1273 * { [l,i] : i <= init and condition' }
1275 * Where condition' is equal to condition if the latter is
1276 * a simple upper [lower] bound and a condition that is extended
1277 * to apply to all previous iterations otherwise.
1278 * Afterwards, the schedule of the pet_scop is extended with
1286 * If the condition is non-affine, then we drop the condition from the
1287 * iteration domain and instead create a separate statement
1288 * for evaluating the condition. The body is then filtered to depend
1289 * on the result of the condition evaluating to true on all iterations
1290 * up to the current iteration, while the evaluation the condition itself
1291 * is filtered to depend on the result of the condition evaluating to true
1292 * on all previous iterations.
1293 * The context of the scop representing the body is dropped
1294 * because we don't know how many times the body will be executed,
1297 * If the stride of the loop is not 1, then "i >= init" is replaced by
1299 * (exists a: i = init + stride * a and a >= 0)
1301 * If the loop iterator i is unsigned, then wrapping may occur.
1302 * We therefore use a virtual iterator instead that does not wrap.
1303 * However, the condition in the code applies
1304 * to the wrapped value, so we need to change condition(l,i)
1305 * into condition([l,i % 2^width]). Similarly, we replace all accesses
1306 * to the original iterator by the wrapping of the virtual iterator.
1307 * Note that there may be no need to perform this final wrapping
1308 * if the loop condition (after wrapping) satisfies certain conditions.
1309 * However, the is_simple_bound condition is not enough since it doesn't
1310 * check if there even is an upper bound.
1312 * Wrapping on unsigned iterators can be avoided entirely if
1313 * loop condition is simple, the loop iterator is incremented
1314 * [decremented] by one and the last value before wrapping cannot
1315 * possibly satisfy the loop condition.
1317 * Valid outer iterators for a for loop are those for which the initial
1318 * value itself, the increment on each domain iteration and
1319 * the condition on both the initial value and
1320 * the result of incrementing the iterator for each iteration of the domain
1322 * If the loop condition is non-affine, then we only consider validity
1323 * of the initial value.
1325 * If the body contains any break, then we keep track of it in "skip"
1326 * (if the skip condition is affine) or it is handled in scop_add_break
1327 * (if the skip condition is not affine).
1328 * Note that the affine break condition needs to be considered with
1329 * respect to previous iterations in the virtual domain (if any).
1331 static struct pet_scop
*scop_from_affine_for(__isl_keep pet_tree
*tree
,
1332 __isl_take isl_pw_aff
*init_val
, __isl_take isl_pw_aff
*pa_inc
,
1333 __isl_take isl_val
*inc
, __isl_take pet_context
*pc
,
1334 struct pet_state
*state
)
1338 isl_set
*cond
= NULL
;
1339 isl_set
*skip
= NULL
;
1340 isl_id
*id_test
= NULL
, *id_break_test
;
1341 struct pet_scop
*scop
, *scop_cond
= NULL
;
1348 int has_affine_break
;
1350 isl_map
*rev_wrap
= NULL
;
1351 isl_map
*init_val_map
;
1353 isl_set
*valid_init
;
1354 isl_set
*valid_cond
;
1355 isl_set
*valid_cond_init
;
1356 isl_set
*valid_cond_next
;
1358 pet_expr
*cond_expr
;
1359 pet_context
*pc_nested
;
1361 pos
= pet_context_dim(pc
) - 1;
1363 domain
= pet_context_get_domain(pc
);
1364 cond_expr
= pet_expr_copy(tree
->u
.l
.cond
);
1365 cond_expr
= pet_context_evaluate_expr(pc
, cond_expr
);
1366 pc_nested
= pet_context_copy(pc
);
1367 pc_nested
= pet_context_set_allow_nested(pc_nested
, 1);
1368 pa
= pet_expr_extract_affine_condition(cond_expr
, pc_nested
);
1369 pet_context_free(pc_nested
);
1370 pet_expr_free(cond_expr
);
1372 valid_inc
= isl_pw_aff_domain(pa_inc
);
1374 is_unsigned
= pet_expr_get_type_size(tree
->u
.l
.iv
) > 0;
1376 is_non_affine
= isl_pw_aff_involves_nan(pa
) ||
1377 !is_nested_allowed(pa
, tree
->u
.l
.body
);
1379 pa
= isl_pw_aff_free(pa
);
1381 valid_cond
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
1382 cond
= isl_pw_aff_non_zero_set(pa
);
1384 cond
= isl_set_universe(isl_set_get_space(domain
));
1386 valid_cond
= isl_set_coalesce(valid_cond
);
1387 is_one
= isl_val_is_one(inc
) || isl_val_is_negone(inc
);
1388 is_virtual
= is_unsigned
&&
1389 (!is_one
|| can_wrap(cond
, tree
->u
.l
.iv
, inc
));
1391 init_val_map
= isl_map_from_pw_aff(isl_pw_aff_copy(init_val
));
1392 init_val_map
= isl_map_equate(init_val_map
, isl_dim_in
, pos
,
1394 valid_cond_init
= enforce_subset(isl_map_domain(init_val_map
),
1395 isl_set_copy(valid_cond
));
1396 if (is_one
&& !is_virtual
) {
1399 isl_pw_aff_free(init_val
);
1400 pa
= pet_expr_extract_comparison(
1401 isl_val_is_pos(inc
) ? pet_op_ge
: pet_op_le
,
1402 tree
->u
.l
.iv
, tree
->u
.l
.init
, pc
);
1403 valid_init
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
1404 valid_init
= isl_set_eliminate(valid_init
, isl_dim_set
,
1405 isl_set_dim(domain
, isl_dim_set
) - 1, 1);
1406 cond
= isl_pw_aff_non_zero_set(pa
);
1407 domain
= isl_set_intersect(domain
, cond
);
1411 valid_init
= isl_pw_aff_domain(isl_pw_aff_copy(init_val
));
1412 strided
= strided_domain(init_val
, isl_val_copy(inc
));
1413 domain
= isl_set_intersect(domain
, strided
);
1417 isl_multi_aff
*wrap
;
1418 wrap
= compute_wrapping(isl_set_get_space(cond
), tree
->u
.l
.iv
);
1419 pc
= pet_context_preimage_domain(pc
, wrap
);
1420 rev_wrap
= isl_map_from_multi_aff(wrap
);
1421 rev_wrap
= isl_map_reverse(rev_wrap
);
1422 cond
= isl_set_apply(cond
, isl_map_copy(rev_wrap
));
1423 valid_cond
= isl_set_apply(valid_cond
, isl_map_copy(rev_wrap
));
1424 valid_inc
= isl_set_apply(valid_inc
, isl_map_copy(rev_wrap
));
1426 is_simple
= is_simple_bound(cond
, inc
);
1428 cond
= isl_set_gist(cond
, isl_set_copy(domain
));
1429 is_simple
= is_simple_bound(cond
, inc
);
1432 cond
= valid_for_each_iteration(cond
,
1433 isl_set_copy(domain
), isl_val_copy(inc
));
1434 cond
= isl_set_align_params(cond
, isl_set_get_space(domain
));
1435 domain
= isl_set_intersect(domain
, cond
);
1436 sched
= map_to_last(pc
);
1437 if (isl_val_is_neg(inc
))
1438 sched
= isl_aff_neg(sched
);
1440 valid_cond_next
= valid_on_next(valid_cond
, isl_set_copy(domain
),
1442 valid_inc
= enforce_subset(isl_set_copy(domain
), valid_inc
);
1444 pc
= pet_context_intersect_domain(pc
, isl_set_copy(domain
));
1446 if (is_non_affine
) {
1448 isl_multi_pw_aff
*test_index
;
1449 space
= isl_set_get_space(domain
);
1450 test_index
= pet_create_test_index(space
, state
->n_test
++);
1451 scop_cond
= scop_from_non_affine_condition(
1452 pet_expr_copy(tree
->u
.l
.cond
), state
->n_stmt
++,
1453 isl_multi_pw_aff_copy(test_index
),
1454 pet_tree_get_loc(tree
), pc
);
1455 id_test
= isl_multi_pw_aff_get_tuple_id(test_index
,
1457 scop_cond
= pet_scop_add_boolean_array(scop_cond
,
1458 isl_set_copy(domain
), test_index
,
1460 scop_cond
= pet_scop_prefix(scop_cond
, 0);
1461 scop_cond
= pet_scop_embed(scop_cond
, isl_set_copy(domain
),
1462 isl_aff_copy(sched
));
1465 scop
= scop_from_tree(tree
->u
.l
.body
, pc
, state
);
1466 has_affine_break
= scop
&&
1467 pet_scop_has_affine_skip(scop
, pet_skip_later
);
1468 if (has_affine_break
)
1469 skip
= pet_scop_get_affine_skip_domain(scop
, pet_skip_later
);
1470 has_var_break
= scop
&& pet_scop_has_var_skip(scop
, pet_skip_later
);
1472 id_break_test
= pet_scop_get_skip_id(scop
, pet_skip_later
);
1473 if (is_non_affine
) {
1474 scop
= pet_scop_reset_context(scop
);
1475 scop
= pet_scop_prefix(scop
, 1);
1477 scop
= pet_scop_embed(scop
, isl_set_copy(domain
), sched
);
1478 scop
= pet_scop_resolve_nested(scop
);
1479 if (has_affine_break
) {
1480 domain
= apply_affine_break(domain
, skip
, isl_val_sgn(inc
),
1481 is_virtual
, rev_wrap
);
1482 scop
= pet_scop_intersect_domain_prefix(scop
,
1483 isl_set_copy(domain
));
1485 isl_map_free(rev_wrap
);
1487 scop
= scop_add_break(scop
, id_break_test
, isl_set_copy(domain
),
1489 if (is_non_affine
) {
1490 scop
= scop_add_while(scop_cond
, scop
, id_test
, domain
,
1492 isl_set_free(valid_inc
);
1494 valid_inc
= isl_set_intersect(valid_inc
, valid_cond_next
);
1495 valid_inc
= isl_set_intersect(valid_inc
, valid_cond_init
);
1496 valid_inc
= isl_set_project_out(valid_inc
, isl_dim_set
, pos
, 1);
1497 scop
= pet_scop_restrict_context(scop
, valid_inc
);
1498 scop
= set_independence(scop
, tree
, domain
, isl_val_sgn(inc
),
1500 isl_set_free(domain
);
1505 valid_init
= isl_set_project_out(valid_init
, isl_dim_set
, pos
, 1);
1506 scop
= pet_scop_restrict_context(scop
, valid_init
);
1508 pet_context_free(pc
);
1512 /* Construct a pet_scop for a for statement within the context of "pc".
1514 * We update the context to reflect the writes to the loop variable and
1515 * the writes inside the body.
1517 * Then we check if the initialization of the for loop
1518 * is a static affine value and the increment is a constant.
1519 * If so, we construct the pet_scop using scop_from_affine_for.
1520 * Otherwise, we treat the for loop as a while loop
1521 * in scop_from_non_affine_for.
1523 * Note that the initialization and the increment are extracted
1524 * in a context where the current loop iterator has been added
1525 * to the context. If these turn out not be affine, then we
1526 * have reconstruct the body context without an assignment
1527 * to this loop iterator, as this variable will then not be
1528 * treated as a dimension of the iteration domain, but as any
1531 static struct pet_scop
*scop_from_for(__isl_keep pet_tree
*tree
,
1532 __isl_keep pet_context
*init_pc
, struct pet_state
*state
)
1536 isl_pw_aff
*pa_inc
, *init_val
;
1537 pet_context
*pc
, *pc_init_val
;
1542 iv
= pet_expr_access_get_id(tree
->u
.l
.iv
);
1543 pc
= pet_context_copy(init_pc
);
1544 pc
= pet_context_add_inner_iterator(pc
, iv
);
1545 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.l
.body
);
1547 pc_init_val
= pet_context_copy(pc
);
1548 pc_init_val
= pet_context_clear_value(pc_init_val
, isl_id_copy(iv
));
1549 init_val
= pet_expr_extract_affine(tree
->u
.l
.init
, pc_init_val
);
1550 pet_context_free(pc_init_val
);
1551 pa_inc
= pet_expr_extract_affine(tree
->u
.l
.inc
, pc
);
1552 inc
= pet_extract_cst(pa_inc
);
1553 if (!pa_inc
|| !init_val
|| !inc
)
1555 if (!isl_pw_aff_involves_nan(pa_inc
) &&
1556 !isl_pw_aff_involves_nan(init_val
) && !isl_val_is_nan(inc
))
1557 return scop_from_affine_for(tree
, init_val
, pa_inc
, inc
,
1560 isl_pw_aff_free(pa_inc
);
1561 isl_pw_aff_free(init_val
);
1563 pet_context_free(pc
);
1565 pc
= pet_context_copy(init_pc
);
1566 pc
= pet_context_add_infinite_loop(pc
);
1567 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.l
.body
);
1568 return scop_from_non_affine_for(tree
, init_pc
, pc
, state
);
1570 isl_pw_aff_free(pa_inc
);
1571 isl_pw_aff_free(init_val
);
1573 pet_context_free(pc
);
1577 /* Check whether "expr" is an affine constraint within the context "pc".
1579 static int is_affine_condition(__isl_keep pet_expr
*expr
,
1580 __isl_keep pet_context
*pc
)
1585 pa
= pet_expr_extract_affine_condition(expr
, pc
);
1588 is_affine
= !isl_pw_aff_involves_nan(pa
);
1589 isl_pw_aff_free(pa
);
1594 /* Check if the given if statement is a conditional assignement
1595 * with a non-affine condition.
1597 * In particular we check if "stmt" is of the form
1604 * where the condition is non-affine and a is some array or scalar access.
1606 static int is_conditional_assignment(__isl_keep pet_tree
*tree
,
1607 __isl_keep pet_context
*pc
)
1611 pet_expr
*expr1
, *expr2
;
1613 ctx
= pet_tree_get_ctx(tree
);
1614 if (!pet_options_get_detect_conditional_assignment(ctx
))
1616 if (tree
->type
!= pet_tree_if_else
)
1618 if (tree
->u
.i
.then_body
->type
!= pet_tree_expr
)
1620 if (tree
->u
.i
.else_body
->type
!= pet_tree_expr
)
1622 expr1
= tree
->u
.i
.then_body
->u
.e
.expr
;
1623 expr2
= tree
->u
.i
.else_body
->u
.e
.expr
;
1624 if (pet_expr_get_type(expr1
) != pet_expr_op
)
1626 if (pet_expr_get_type(expr2
) != pet_expr_op
)
1628 if (pet_expr_op_get_type(expr1
) != pet_op_assign
)
1630 if (pet_expr_op_get_type(expr2
) != pet_op_assign
)
1632 expr1
= pet_expr_get_arg(expr1
, 0);
1633 expr2
= pet_expr_get_arg(expr2
, 0);
1634 equal
= pet_expr_is_equal(expr1
, expr2
);
1635 pet_expr_free(expr1
);
1636 pet_expr_free(expr2
);
1637 if (equal
< 0 || !equal
)
1639 if (is_affine_condition(tree
->u
.i
.cond
, pc
))
1645 /* Given that "tree" is of the form
1652 * where a is some array or scalar access, construct a pet_scop
1653 * corresponding to this conditional assignment within the context "pc".
1655 * The constructed pet_scop then corresponds to the expression
1657 * a = condition ? f(...) : g(...)
1659 * All access relations in f(...) are intersected with condition
1660 * while all access relation in g(...) are intersected with the complement.
1662 static struct pet_scop
*scop_from_conditional_assignment(
1663 __isl_keep pet_tree
*tree
, __isl_take pet_context
*pc
,
1664 struct pet_state
*state
)
1668 isl_set
*cond
, *comp
;
1669 isl_multi_pw_aff
*index
;
1670 pet_expr
*expr1
, *expr2
;
1671 pet_expr
*pe_cond
, *pe_then
, *pe_else
, *pe
, *pe_write
;
1672 pet_context
*pc_nested
;
1673 struct pet_scop
*scop
;
1675 pe_cond
= pet_expr_copy(tree
->u
.i
.cond
);
1676 pe_cond
= pet_context_evaluate_expr(pc
, pe_cond
);
1677 pc_nested
= pet_context_copy(pc
);
1678 pc_nested
= pet_context_set_allow_nested(pc_nested
, 1);
1679 pa
= pet_expr_extract_affine_condition(pe_cond
, pc_nested
);
1680 pet_context_free(pc_nested
);
1681 pet_expr_free(pe_cond
);
1682 cond
= isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa
));
1683 comp
= isl_pw_aff_zero_set(isl_pw_aff_copy(pa
));
1684 index
= isl_multi_pw_aff_from_pw_aff(pa
);
1686 expr1
= tree
->u
.i
.then_body
->u
.e
.expr
;
1687 expr2
= tree
->u
.i
.else_body
->u
.e
.expr
;
1689 pe_cond
= pet_expr_from_index(index
);
1691 pe_then
= pet_expr_get_arg(expr1
, 1);
1692 pe_then
= pet_context_evaluate_expr(pc
, pe_then
);
1693 pe_then
= pet_expr_restrict(pe_then
, cond
);
1694 pe_else
= pet_expr_get_arg(expr2
, 1);
1695 pe_else
= pet_context_evaluate_expr(pc
, pe_else
);
1696 pe_else
= pet_expr_restrict(pe_else
, comp
);
1697 pe_write
= pet_expr_get_arg(expr1
, 0);
1698 pe_write
= pet_context_evaluate_expr(pc
, pe_write
);
1700 pe
= pet_expr_new_ternary(pe_cond
, pe_then
, pe_else
);
1701 type_size
= pet_expr_get_type_size(pe_write
);
1702 pe
= pet_expr_new_binary(type_size
, pet_op_assign
, pe_write
, pe
);
1704 scop
= scop_from_evaluated_expr(pe
, state
->n_stmt
++,
1705 pet_tree_get_loc(tree
), pc
);
1707 pet_context_free(pc
);
1712 /* Construct a pet_scop for a non-affine if statement within the context "pc".
1714 * We create a separate statement that writes the result
1715 * of the non-affine condition to a virtual scalar.
1716 * A constraint requiring the value of this virtual scalar to be one
1717 * is added to the iteration domains of the then branch.
1718 * Similarly, a constraint requiring the value of this virtual scalar
1719 * to be zero is added to the iteration domains of the else branch, if any.
1720 * We adjust the schedules to ensure that the virtual scalar is written
1721 * before it is read.
1723 * If there are any breaks or continues in the then and/or else
1724 * branches, then we may have to compute a new skip condition.
1725 * This is handled using a pet_skip_info object.
1726 * On initialization, the object checks if skip conditions need
1727 * to be computed. If so, it does so in pet_skip_info_if_extract_index and
1728 * adds them in pet_skip_info_if_add.
1730 static struct pet_scop
*scop_from_non_affine_if(__isl_keep pet_tree
*tree
,
1731 __isl_take pet_context
*pc
, struct pet_state
*state
)
1736 isl_multi_pw_aff
*test_index
;
1737 struct pet_skip_info skip
;
1738 struct pet_scop
*scop
, *scop_then
, *scop_else
= NULL
;
1740 has_else
= tree
->type
== pet_tree_if_else
;
1742 space
= pet_context_get_space(pc
);
1743 test_index
= pet_create_test_index(space
, state
->n_test
++);
1744 scop
= scop_from_non_affine_condition(pet_expr_copy(tree
->u
.i
.cond
),
1745 state
->n_stmt
++, isl_multi_pw_aff_copy(test_index
),
1746 pet_tree_get_loc(tree
), pc
);
1747 domain
= pet_context_get_domain(pc
);
1748 scop
= pet_scop_add_boolean_array(scop
, domain
,
1749 isl_multi_pw_aff_copy(test_index
), state
->int_size
);
1751 scop_then
= scop_from_tree(tree
->u
.i
.then_body
, pc
, state
);
1753 scop_else
= scop_from_tree(tree
->u
.i
.else_body
, pc
, state
);
1755 pet_skip_info_if_init(&skip
, state
->ctx
, scop_then
, scop_else
,
1757 pet_skip_info_if_extract_index(&skip
, test_index
, pc
, state
);
1759 scop
= pet_scop_prefix(scop
, 0);
1760 scop_then
= pet_scop_prefix(scop_then
, 1);
1761 scop_then
= pet_scop_filter(scop_then
,
1762 isl_multi_pw_aff_copy(test_index
), 1);
1764 scop_else
= pet_scop_prefix(scop_else
, 1);
1765 scop_else
= pet_scop_filter(scop_else
, test_index
, 0);
1766 scop_then
= pet_scop_add_par(state
->ctx
, scop_then
, scop_else
);
1768 isl_multi_pw_aff_free(test_index
);
1770 scop
= pet_scop_add_seq(state
->ctx
, scop
, scop_then
);
1772 scop
= pet_skip_info_if_add(&skip
, scop
, 2);
1774 pet_context_free(pc
);
1778 /* Construct a pet_scop for an affine if statement within the context "pc".
1780 * The condition is added to the iteration domains of the then branch,
1781 * while the opposite of the condition in added to the iteration domains
1782 * of the else branch, if any.
1784 * If there are any breaks or continues in the then and/or else
1785 * branches, then we may have to compute a new skip condition.
1786 * This is handled using a pet_skip_info_if object.
1787 * On initialization, the object checks if skip conditions need
1788 * to be computed. If so, it does so in pet_skip_info_if_extract_cond and
1789 * adds them in pet_skip_info_if_add.
1791 static struct pet_scop
*scop_from_affine_if(__isl_keep pet_tree
*tree
,
1792 __isl_take isl_pw_aff
*cond
, __isl_take pet_context
*pc
,
1793 struct pet_state
*state
)
1797 isl_set
*set
, *complement
;
1799 struct pet_skip_info skip
;
1800 struct pet_scop
*scop
, *scop_then
, *scop_else
= NULL
;
1801 pet_context
*pc_body
;
1803 ctx
= pet_tree_get_ctx(tree
);
1805 has_else
= tree
->type
== pet_tree_if_else
;
1807 valid
= isl_pw_aff_domain(isl_pw_aff_copy(cond
));
1808 set
= isl_pw_aff_non_zero_set(isl_pw_aff_copy(cond
));
1810 pc_body
= pet_context_copy(pc
);
1811 pc_body
= pet_context_intersect_domain(pc_body
, isl_set_copy(set
));
1812 scop_then
= scop_from_tree(tree
->u
.i
.then_body
, pc_body
, state
);
1813 pet_context_free(pc_body
);
1815 pc_body
= pet_context_copy(pc
);
1816 complement
= isl_set_copy(valid
);
1817 complement
= isl_set_subtract(valid
, isl_set_copy(set
));
1818 pc_body
= pet_context_intersect_domain(pc_body
,
1819 isl_set_copy(complement
));
1820 scop_else
= scop_from_tree(tree
->u
.i
.else_body
, pc_body
, state
);
1821 pet_context_free(pc_body
);
1824 pet_skip_info_if_init(&skip
, ctx
, scop_then
, scop_else
, has_else
, 1);
1825 pet_skip_info_if_extract_cond(&skip
, cond
, pc
, state
);
1826 isl_pw_aff_free(cond
);
1828 scop
= pet_scop_restrict(scop_then
, set
);
1831 scop_else
= pet_scop_restrict(scop_else
, complement
);
1832 scop
= pet_scop_add_par(ctx
, scop
, scop_else
);
1834 scop
= pet_scop_resolve_nested(scop
);
1835 scop
= pet_scop_restrict_context(scop
, valid
);
1837 if (pet_skip_info_has_skip(&skip
))
1838 scop
= pet_scop_prefix(scop
, 0);
1839 scop
= pet_skip_info_if_add(&skip
, scop
, 1);
1841 pet_context_free(pc
);
1845 /* Construct a pet_scop for an if statement within the context "pc".
1847 * If the condition fits the pattern of a conditional assignment,
1848 * then it is handled by scop_from_conditional_assignment.
1850 * Otherwise, we check if the condition is affine.
1851 * If so, we construct the scop in scop_from_affine_if.
1852 * Otherwise, we construct the scop in scop_from_non_affine_if.
1854 * We allow the condition to be dynamic, i.e., to refer to
1855 * scalars or array elements that may be written to outside
1856 * of the given if statement. These nested accesses are then represented
1857 * as output dimensions in the wrapping iteration domain.
1858 * If it is also written _inside_ the then or else branch, then
1859 * we treat the condition as non-affine.
1860 * As explained in extract_non_affine_if, this will introduce
1861 * an extra statement.
1862 * For aesthetic reasons, we want this statement to have a statement
1863 * number that is lower than those of the then and else branches.
1864 * In order to evaluate if we will need such a statement, however, we
1865 * first construct scops for the then and else branches.
1866 * We therefore reserve a statement number if we might have to
1867 * introduce such an extra statement.
1869 static struct pet_scop
*scop_from_if(__isl_keep pet_tree
*tree
,
1870 __isl_keep pet_context
*pc
, struct pet_state
*state
)
1874 pet_expr
*cond_expr
;
1875 pet_context
*pc_nested
;
1880 has_else
= tree
->type
== pet_tree_if_else
;
1882 pc
= pet_context_copy(pc
);
1883 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.i
.then_body
);
1885 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.i
.else_body
);
1887 if (is_conditional_assignment(tree
, pc
))
1888 return scop_from_conditional_assignment(tree
, pc
, state
);
1890 cond_expr
= pet_expr_copy(tree
->u
.i
.cond
);
1891 cond_expr
= pet_context_evaluate_expr(pc
, cond_expr
);
1892 pc_nested
= pet_context_copy(pc
);
1893 pc_nested
= pet_context_set_allow_nested(pc_nested
, 1);
1894 cond
= pet_expr_extract_affine_condition(cond_expr
, pc_nested
);
1895 pet_context_free(pc_nested
);
1896 pet_expr_free(cond_expr
);
1899 pet_context_free(pc
);
1903 if (isl_pw_aff_involves_nan(cond
)) {
1904 isl_pw_aff_free(cond
);
1905 return scop_from_non_affine_if(tree
, pc
, state
);
1908 if ((!is_nested_allowed(cond
, tree
->u
.i
.then_body
) ||
1909 (has_else
&& !is_nested_allowed(cond
, tree
->u
.i
.else_body
)))) {
1910 isl_pw_aff_free(cond
);
1911 return scop_from_non_affine_if(tree
, pc
, state
);
1914 return scop_from_affine_if(tree
, cond
, pc
, state
);
1917 /* Return a one-dimensional multi piecewise affine expression that is equal
1918 * to the constant 1 and is defined over the given domain.
1920 static __isl_give isl_multi_pw_aff
*one_mpa(__isl_take isl_space
*space
)
1922 isl_local_space
*ls
;
1925 ls
= isl_local_space_from_space(space
);
1926 aff
= isl_aff_zero_on_domain(ls
);
1927 aff
= isl_aff_set_constant_si(aff
, 1);
1929 return isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff
));
1932 /* Construct a pet_scop for a continue statement with the given domain space.
1934 * We simply create an empty scop with a universal pet_skip_now
1935 * skip condition. This skip condition will then be taken into
1936 * account by the enclosing loop construct, possibly after
1937 * being incorporated into outer skip conditions.
1939 static struct pet_scop
*scop_from_continue(__isl_keep pet_tree
*tree
,
1940 __isl_take isl_space
*space
)
1942 struct pet_scop
*scop
;
1944 scop
= pet_scop_empty(isl_space_copy(space
));
1946 scop
= pet_scop_set_skip(scop
, pet_skip_now
, one_mpa(space
));
1951 /* Construct a pet_scop for a break statement with the given domain space.
1953 * We simply create an empty scop with both a universal pet_skip_now
1954 * skip condition and a universal pet_skip_later skip condition.
1955 * These skip conditions will then be taken into
1956 * account by the enclosing loop construct, possibly after
1957 * being incorporated into outer skip conditions.
1959 static struct pet_scop
*scop_from_break(__isl_keep pet_tree
*tree
,
1960 __isl_take isl_space
*space
)
1962 struct pet_scop
*scop
;
1963 isl_multi_pw_aff
*skip
;
1965 scop
= pet_scop_empty(isl_space_copy(space
));
1967 skip
= one_mpa(space
);
1968 scop
= pet_scop_set_skip(scop
, pet_skip_now
,
1969 isl_multi_pw_aff_copy(skip
));
1970 scop
= pet_scop_set_skip(scop
, pet_skip_later
, skip
);
1975 /* Extract a clone of the kill statement in "scop".
1976 * The domain of the clone is given by "domain".
1977 * "scop" is expected to have been created from a DeclStmt
1978 * and should have the kill as its first statement.
1980 static struct pet_scop
*extract_kill(__isl_keep isl_set
*domain
,
1981 struct pet_scop
*scop
, struct pet_state
*state
)
1984 struct pet_stmt
*stmt
;
1985 isl_multi_pw_aff
*index
;
1987 pet_expr
*expr
, *arg
;
1990 if (!domain
|| !scop
)
1992 if (scop
->n_stmt
< 1)
1993 isl_die(isl_set_get_ctx(domain
), isl_error_internal
,
1994 "expecting at least one statement", return NULL
);
1995 stmt
= scop
->stmts
[0];
1996 if (!pet_stmt_is_kill(stmt
))
1997 isl_die(isl_set_get_ctx(domain
), isl_error_internal
,
1998 "expecting kill statement", return NULL
);
2000 expr
= pet_tree_expr_get_expr(stmt
->body
);
2001 arg
= pet_expr_get_arg(expr
, 0);
2002 pet_expr_free(expr
);
2003 index
= pet_expr_access_get_index(arg
);
2004 access
= pet_expr_access_get_access(arg
);
2006 index
= isl_multi_pw_aff_reset_tuple_id(index
, isl_dim_in
);
2007 access
= isl_map_reset_tuple_id(access
, isl_dim_in
);
2008 kill
= pet_expr_kill_from_access_and_index(access
, index
);
2009 tree
= pet_tree_new_expr(kill
);
2010 tree
= pet_tree_set_loc(tree
, pet_loc_copy(stmt
->loc
));
2011 stmt
= pet_stmt_from_pet_tree(isl_set_copy(domain
),
2012 state
->n_stmt
++, tree
);
2013 return pet_scop_from_pet_stmt(isl_set_get_space(domain
), stmt
);
2016 /* Does "tree" represent an assignment to a variable?
2018 * The assignment may be one of
2019 * - a declaration with initialization
2020 * - an expression with a top-level assignment operator
2022 static int is_assignment(__isl_keep pet_tree
*tree
)
2026 if (tree
->type
== pet_tree_decl_init
)
2028 return pet_tree_is_assign(tree
);
2031 /* Update "pc" by taking into account the assignment performed by "tree",
2032 * where "tree" satisfies is_assignment.
2034 * In particular, if the lhs of the assignment is a scalar variable and
2035 * if the rhs is an affine expression, then keep track of this value in "pc"
2036 * so that we can plug it in when we later come across the same variable.
2038 * Any previously assigned value to the variable has already been removed
2039 * by scop_handle_writes.
2041 static __isl_give pet_context
*handle_assignment(__isl_take pet_context
*pc
,
2042 __isl_keep pet_tree
*tree
)
2044 pet_expr
*var
, *val
;
2048 if (pet_tree_get_type(tree
) == pet_tree_decl_init
) {
2049 var
= pet_tree_decl_get_var(tree
);
2050 val
= pet_tree_decl_get_init(tree
);
2053 expr
= pet_tree_expr_get_expr(tree
);
2054 var
= pet_expr_get_arg(expr
, 0);
2055 val
= pet_expr_get_arg(expr
, 1);
2056 pet_expr_free(expr
);
2059 if (!pet_expr_is_scalar_access(var
)) {
2065 pa
= pet_expr_extract_affine(val
, pc
);
2067 pc
= pet_context_free(pc
);
2069 if (!isl_pw_aff_involves_nan(pa
)) {
2070 id
= pet_expr_access_get_id(var
);
2071 pc
= pet_context_set_value(pc
, id
, pa
);
2073 isl_pw_aff_free(pa
);
2081 /* Mark all arrays in "scop" as being exposed.
2083 static struct pet_scop
*mark_exposed(struct pet_scop
*scop
)
2089 for (i
= 0; i
< scop
->n_array
; ++i
)
2090 scop
->arrays
[i
]->exposed
= 1;
2094 /* Given that "scop" has an affine skip condition of type pet_skip_now,
2095 * apply this skip condition to the domain of "pc".
2096 * That is, remove the elements satisfying the skip condition from
2097 * the domain of "pc".
2099 static __isl_give pet_context
*apply_affine_continue(__isl_take pet_context
*pc
,
2100 struct pet_scop
*scop
)
2102 isl_set
*domain
, *skip
;
2104 skip
= pet_scop_get_affine_skip_domain(scop
, pet_skip_now
);
2105 domain
= pet_context_get_domain(pc
);
2106 domain
= isl_set_subtract(domain
, skip
);
2107 pc
= pet_context_intersect_domain(pc
, domain
);
2112 /* Try and construct a pet_scop corresponding to (part of)
2113 * a sequence of statements within the context "pc".
2115 * After extracting a statement, we update "pc"
2116 * based on the top-level assignments in the statement
2117 * so that we can exploit them in subsequent statements in the same block.
2119 * If there are any breaks or continues in the individual statements,
2120 * then we may have to compute a new skip condition.
2121 * This is handled using a pet_skip_info object.
2122 * On initialization, the object checks if skip conditions need
2123 * to be computed. If so, it does so in pet_skip_info_seq_extract and
2124 * adds them in pet_skip_info_seq_add.
2126 * If "block" is set, then we need to insert kill statements at
2127 * the end of the block for any array that has been declared by
2128 * one of the statements in the sequence. Each of these declarations
2129 * results in the construction of a kill statement at the place
2130 * of the declaration, so we simply collect duplicates of
2131 * those kill statements and append these duplicates to the constructed scop.
2133 * If "block" is not set, then any array declared by one of the statements
2134 * in the sequence is marked as being exposed.
2136 * If autodetect is set, then we allow the extraction of only a subrange
2137 * of the sequence of statements. However, if there is at least one statement
2138 * for which we could not construct a scop and the final range contains
2139 * either no statements or at least one kill, then we discard the entire
2142 static struct pet_scop
*scop_from_block(__isl_keep pet_tree
*tree
,
2143 __isl_keep pet_context
*pc
, struct pet_state
*state
)
2149 struct pet_scop
*scop
, *kills
;
2151 ctx
= pet_tree_get_ctx(tree
);
2153 space
= pet_context_get_space(pc
);
2154 domain
= pet_context_get_domain(pc
);
2155 pc
= pet_context_copy(pc
);
2156 scop
= pet_scop_empty(isl_space_copy(space
));
2157 kills
= pet_scop_empty(space
);
2158 for (i
= 0; i
< tree
->u
.b
.n
; ++i
) {
2159 struct pet_scop
*scop_i
;
2161 if (pet_scop_has_affine_skip(scop
, pet_skip_now
))
2162 pc
= apply_affine_continue(pc
, scop
);
2163 scop_i
= scop_from_tree(tree
->u
.b
.child
[i
], pc
, state
);
2164 pc
= scop_handle_writes(scop_i
, pc
);
2165 if (is_assignment(tree
->u
.b
.child
[i
]))
2166 pc
= handle_assignment(pc
, tree
->u
.b
.child
[i
]);
2167 struct pet_skip_info skip
;
2168 pet_skip_info_seq_init(&skip
, ctx
, scop
, scop_i
);
2169 pet_skip_info_seq_extract(&skip
, pc
, state
);
2170 if (pet_skip_info_has_skip(&skip
))
2171 scop_i
= pet_scop_prefix(scop_i
, 0);
2172 if (scop_i
&& pet_tree_is_decl(tree
->u
.b
.child
[i
])) {
2173 if (tree
->u
.b
.block
) {
2174 struct pet_scop
*kill
;
2175 kill
= extract_kill(domain
, scop_i
, state
);
2176 kills
= pet_scop_add_par(ctx
, kills
, kill
);
2178 scop_i
= mark_exposed(scop_i
);
2180 scop_i
= pet_scop_prefix(scop_i
, i
);
2181 scop
= pet_scop_add_seq(ctx
, scop
, scop_i
);
2183 scop
= pet_skip_info_seq_add(&skip
, scop
, i
);
2188 isl_set_free(domain
);
2190 kills
= pet_scop_prefix(kills
, tree
->u
.b
.n
);
2191 scop
= pet_scop_add_seq(ctx
, scop
, kills
);
2193 pet_context_free(pc
);
2198 /* Internal data structure for extract_declared_arrays.
2200 * "pc" and "state" are used to create pet_array objects and kill statements.
2201 * "any" is initialized to 0 by the caller and set to 1 as soon as we have
2202 * found any declared array.
2203 * "scop" has been initialized by the caller and is used to attach
2204 * the created pet_array objects.
2205 * "kill_before" and "kill_after" are created and updated by
2206 * extract_declared_arrays to collect the kills of the arrays.
2208 struct pet_tree_extract_declared_arrays_data
{
2210 struct pet_state
*state
;
2215 struct pet_scop
*scop
;
2216 struct pet_scop
*kill_before
;
2217 struct pet_scop
*kill_after
;
2220 /* Check if the node "node" declares any array or scalar.
2221 * If so, create the corresponding pet_array and attach it to data->scop.
2222 * Additionally, create two kill statements for the array and add them
2223 * to data->kill_before and data->kill_after.
2225 static int extract_declared_arrays(__isl_keep pet_tree
*node
, void *user
)
2227 enum pet_tree_type type
;
2228 struct pet_tree_extract_declared_arrays_data
*data
= user
;
2229 struct pet_array
*array
;
2230 struct pet_scop
*scop_kill
;
2233 type
= pet_tree_get_type(node
);
2234 if (type
== pet_tree_decl
|| type
== pet_tree_decl_init
)
2235 var
= node
->u
.d
.var
;
2236 else if (type
== pet_tree_for
&& node
->u
.l
.declared
)
2241 array
= extract_array(var
, data
->pc
, data
->state
);
2243 array
->declared
= 1;
2244 data
->scop
= pet_scop_add_array(data
->scop
, array
);
2246 scop_kill
= kill(pet_tree_get_loc(node
), array
, data
->pc
, data
->state
);
2248 data
->kill_before
= scop_kill
;
2250 data
->kill_before
= pet_scop_add_par(data
->ctx
,
2251 data
->kill_before
, scop_kill
);
2253 scop_kill
= kill(pet_tree_get_loc(node
), array
, data
->pc
, data
->state
);
2255 data
->kill_after
= scop_kill
;
2257 data
->kill_after
= pet_scop_add_par(data
->ctx
,
2258 data
->kill_after
, scop_kill
);
2265 /* Convert a pet_tree that consists of more than a single leaf
2266 * to a pet_scop with a single statement encapsulating the entire pet_tree.
2267 * Do so within the context of "pc".
2269 * After constructing the core scop, we also look for any arrays (or scalars)
2270 * that are declared inside "tree". Each of those arrays is marked as
2271 * having been declared and kill statements for these arrays
2272 * are introduced before and after the core scop.
2273 * Note that the input tree is not a leaf so that the declaration
2274 * cannot occur at the outer level.
2276 static struct pet_scop
*scop_from_tree_macro(__isl_take pet_tree
*tree
,
2277 __isl_take isl_id
*label
, __isl_keep pet_context
*pc
,
2278 struct pet_state
*state
)
2280 struct pet_tree_extract_declared_arrays_data data
= { pc
, state
};
2282 data
.scop
= scop_from_unevaluated_tree(pet_tree_copy(tree
),
2283 state
->n_stmt
++, pc
);
2286 data
.ctx
= pet_context_get_ctx(pc
);
2287 if (pet_tree_foreach_sub_tree(tree
, &extract_declared_arrays
,
2289 data
.scop
= pet_scop_free(data
.scop
);
2290 pet_tree_free(tree
);
2295 data
.kill_before
= pet_scop_prefix(data
.kill_before
, 0);
2296 data
.scop
= pet_scop_prefix(data
.scop
, 1);
2297 data
.kill_after
= pet_scop_prefix(data
.kill_after
, 2);
2299 data
.scop
= pet_scop_add_seq(data
.ctx
, data
.kill_before
, data
.scop
);
2300 data
.scop
= pet_scop_add_seq(data
.ctx
, data
.scop
, data
.kill_after
);
2305 /* Construct a pet_scop that corresponds to the pet_tree "tree"
2306 * within the context "pc" by calling the appropriate function
2307 * based on the type of "tree".
2309 * If the initially constructed pet_scop turns out to involve
2310 * dynamic control and if the user has requested an encapsulation
2311 * of all dynamic control, then this pet_scop is discarded and
2312 * a new pet_scop is created with a single statement representing
2313 * the entire "tree".
2315 static struct pet_scop
*scop_from_tree(__isl_keep pet_tree
*tree
,
2316 __isl_keep pet_context
*pc
, struct pet_state
*state
)
2319 struct pet_scop
*scop
= NULL
;
2324 ctx
= pet_tree_get_ctx(tree
);
2325 switch (tree
->type
) {
2326 case pet_tree_error
:
2328 case pet_tree_block
:
2329 return scop_from_block(tree
, pc
, state
);
2330 case pet_tree_break
:
2331 return scop_from_break(tree
, pet_context_get_space(pc
));
2332 case pet_tree_continue
:
2333 return scop_from_continue(tree
, pet_context_get_space(pc
));
2335 case pet_tree_decl_init
:
2336 return scop_from_decl(tree
, pc
, state
);
2338 return scop_from_unevaluated_tree(pet_tree_copy(tree
),
2339 state
->n_stmt
++, pc
);
2341 case pet_tree_if_else
:
2342 scop
= scop_from_if(tree
, pc
, state
);
2345 scop
= scop_from_for(tree
, pc
, state
);
2347 case pet_tree_while
:
2348 scop
= scop_from_while(tree
, pc
, state
);
2350 case pet_tree_infinite_loop
:
2351 scop
= scop_from_infinite_for(tree
, pc
, state
);
2358 if (!pet_options_get_encapsulate_dynamic_control(ctx
) ||
2359 !pet_scop_has_data_dependent_conditions(scop
))
2362 pet_scop_free(scop
);
2363 return scop_from_tree_macro(pet_tree_copy(tree
),
2364 isl_id_copy(tree
->label
), pc
, state
);
2367 /* Construct a pet_scop that corresponds to the pet_tree "tree".
2368 * "int_size" is the number of bytes need to represent an integer.
2369 * "extract_array" is a callback that we can use to create a pet_array
2370 * that corresponds to the variable accessed by an expression.
2372 * Initialize the global state, construct a context and then
2373 * construct the pet_scop by recursively visiting the tree.
2375 struct pet_scop
*pet_scop_from_pet_tree(__isl_take pet_tree
*tree
, int int_size
,
2376 struct pet_array
*(*extract_array
)(__isl_keep pet_expr
*access
,
2377 __isl_keep pet_context
*pc
, void *user
), void *user
,
2378 __isl_keep pet_context
*pc
)
2380 struct pet_scop
*scop
;
2381 struct pet_state state
= { 0 };
2386 state
.ctx
= pet_tree_get_ctx(tree
);
2387 state
.int_size
= int_size
;
2388 state
.extract_array
= extract_array
;
2391 scop
= scop_from_tree(tree
, pc
, &state
);
2392 scop
= pet_scop_set_loc(scop
, pet_tree_get_loc(tree
));
2394 pet_tree_free(tree
);
2397 scop
->context
= isl_set_params(scop
->context
);