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
38 #include <isl/id_to_pw_aff.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
)
67 return pet_context_free(pc
);
68 for (i
= 0; i
< scop
->n_stmt
; ++i
)
69 pc
= handle_writes(scop
->stmts
[i
], 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
,
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
)
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
)
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
)
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
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
)
180 isl_multi_pw_aff
*index
;
183 struct pet_scop
*scop
;
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
);
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
)
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
);
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
)
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
);
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
)
261 if (tree
->type
!= pet_tree_expr
)
263 expr
= tree
->u
.e
.expr
;
264 if (pet_expr_get_type(expr
) != pet_expr_call
)
266 name
= pet_expr_call_get_name(expr
);
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
)
289 isl_multi_pw_aff
*index
;
292 struct pet_array
*array
;
293 struct pet_scop
*scop_i
;
295 expr
= pet_context_evaluate_expr(pc
, expr
);
298 if (expr
->n_arg
!= 0) {
302 array
= extract_array(expr
, pc
, state
);
305 index
= pet_expr_access_get_index(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
);
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
329 static struct pet_scop
*scop_from_pencil_kill(__isl_keep pet_tree
*tree
,
330 __isl_keep pet_context
*pc
, struct pet_state
*state
)
333 struct pet_scop
*scop
;
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
) {
345 arg
= pet_expr_get_arg(call
, i
);
347 return pet_scop_free(scop
);
348 if (pet_expr_get_type(arg
) != pet_expr_access
) {
352 loc
= pet_tree_get_loc(tree
);
353 scop
= scop_add_kill(scop
, arg
, loc
, pc
, state
);
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
)
370 is_kill
= is_pencil_kill(tree
);
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
)
385 isl_map
*previous_to_this
;
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
);
395 previous_to_this
= isl_map_order_lt(previous_to_this
,
396 isl_dim_in
, dim
- 1, isl_dim_out
, dim
- 1);
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
);
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
)
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 a single-dimensional multi-affine expression on the domain space
422 * of "pc" that is equal to the final dimension of this domain.
423 * "loop_nr" is the sequence number of the corresponding loop.
425 static __isl_give isl_multi_aff
*map_to_last(__isl_keep pet_context
*pc
,
436 space
= pet_context_get_space(pc
);
437 pos
= isl_space_dim(space
, isl_dim_set
) - 1;
438 ls
= isl_local_space_from_space(space
);
439 aff
= isl_aff_var_on_domain(ls
, isl_dim_set
, pos
);
440 ma
= isl_multi_aff_from_aff(aff
);
442 snprintf(name
, sizeof(name
), "L_%d", loop_nr
);
443 label
= isl_id_alloc(pet_context_get_ctx(pc
), name
, NULL
);
444 ma
= isl_multi_aff_set_tuple_id(ma
, isl_dim_out
, label
);
449 /* Create an affine expression that maps elements
450 * of an array "id_test" to the previous element in the final dimension
451 * (according to "inc"), provided this element belongs to "domain".
452 * That is, create the affine expression
454 * { id[outer,x] -> id[outer,x - inc] : (outer,x - inc) in domain }
456 static __isl_give isl_multi_pw_aff
*map_to_previous(__isl_take isl_id
*id_test
,
457 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
464 isl_multi_pw_aff
*prev
;
466 pos
= isl_set_dim(domain
, isl_dim_set
) - 1;
467 space
= isl_set_get_space(domain
);
468 space
= isl_space_map_from_set(space
);
469 ma
= isl_multi_aff_identity(space
);
470 aff
= isl_multi_aff_get_aff(ma
, pos
);
471 aff
= isl_aff_add_constant_val(aff
, isl_val_neg(inc
));
472 ma
= isl_multi_aff_set_aff(ma
, pos
, aff
);
473 domain
= isl_set_preimage_multi_aff(domain
, isl_multi_aff_copy(ma
));
474 prev
= isl_multi_pw_aff_from_multi_aff(ma
);
475 pa
= isl_multi_pw_aff_get_pw_aff(prev
, pos
);
476 pa
= isl_pw_aff_intersect_domain(pa
, domain
);
477 prev
= isl_multi_pw_aff_set_pw_aff(prev
, pos
, pa
);
478 prev
= isl_multi_pw_aff_set_tuple_id(prev
, isl_dim_out
, id_test
);
483 /* Add an implication to "scop" expressing that if an element of
484 * virtual array "id_test" has value "satisfied" then all previous elements
485 * of this array (in the final dimension) also have that value.
486 * The set of previous elements is bounded by "domain".
487 * If "sign" is negative then the iterator
488 * is decreasing and we express that all subsequent array elements
489 * (but still defined previously) have the same value.
491 static struct pet_scop
*add_implication(struct pet_scop
*scop
,
492 __isl_take isl_id
*id_test
, __isl_take isl_set
*domain
, int sign
,
499 dim
= isl_set_dim(domain
, isl_dim_set
);
500 domain
= isl_set_set_tuple_id(domain
, id_test
);
501 space
= isl_space_map_from_set(isl_set_get_space(domain
));
502 map
= isl_map_universe(space
);
503 for (i
= 0; i
+ 1 < dim
; ++i
)
504 map
= isl_map_equate(map
, isl_dim_in
, i
, isl_dim_out
, i
);
506 map
= isl_map_order_ge(map
,
507 isl_dim_in
, dim
- 1, isl_dim_out
, dim
- 1);
509 map
= isl_map_order_le(map
,
510 isl_dim_in
, dim
- 1, isl_dim_out
, dim
- 1);
511 map
= isl_map_intersect_range(map
, domain
);
512 scop
= pet_scop_add_implication(scop
, map
, satisfied
);
517 /* Add a filter to "scop" that imposes that it is only executed
518 * when the variable identified by "id_test" has a zero value
519 * for all previous iterations of "domain".
521 * In particular, add a filter that imposes that the array
522 * has a zero value at the previous iteration of domain and
523 * add an implication that implies that it then has that
524 * value for all previous iterations.
526 static struct pet_scop
*scop_add_break(struct pet_scop
*scop
,
527 __isl_take isl_id
*id_test
, __isl_take isl_set
*domain
,
528 __isl_take isl_val
*inc
)
530 isl_multi_pw_aff
*prev
;
531 int sign
= isl_val_sgn(inc
);
533 prev
= map_to_previous(isl_id_copy(id_test
), isl_set_copy(domain
), inc
);
534 scop
= add_implication(scop
, id_test
, domain
, sign
, 0);
535 scop
= pet_scop_filter(scop
, prev
, 0);
540 static struct pet_scop
*scop_from_tree(__isl_keep pet_tree
*tree
,
541 __isl_keep pet_context
*pc
, struct pet_state
*state
);
543 /* Construct a pet_scop for an infinite loop around the given body
544 * within the context "pc".
546 * The domain of "pc" has already been extended with an infinite loop
550 * We extract a pet_scop for the body and then embed it in a loop with
553 * { [outer,t] -> [t] }
555 * If the body contains any break, then it is taken into
556 * account in apply_affine_break (if the skip condition is affine)
557 * or in scop_add_break (if the skip condition is not affine).
559 * Note that in case of an affine skip condition,
560 * since we are dealing with a loop without loop iterator,
561 * the skip condition cannot refer to the current loop iterator and
562 * so effectively, the effect on the iteration domain is of the form
564 * { [outer,0]; [outer,t] : t >= 1 and not skip }
566 static struct pet_scop
*scop_from_infinite_loop(__isl_keep pet_tree
*body
,
567 __isl_keep pet_context
*pc
, struct pet_state
*state
)
573 isl_multi_aff
*sched
;
574 struct pet_scop
*scop
;
575 int has_affine_break
;
578 ctx
= pet_tree_get_ctx(body
);
579 domain
= pet_context_get_domain(pc
);
580 sched
= map_to_last(pc
, state
->n_loop
++);
582 scop
= scop_from_tree(body
, pc
, state
);
584 has_affine_break
= pet_scop_has_affine_skip(scop
, pet_skip_later
);
585 if (has_affine_break
)
586 skip
= pet_scop_get_affine_skip_domain(scop
, pet_skip_later
);
587 has_var_break
= pet_scop_has_var_skip(scop
, pet_skip_later
);
589 id_test
= pet_scop_get_skip_id(scop
, pet_skip_later
);
591 scop
= pet_scop_reset_skips(scop
);
592 scop
= pet_scop_embed(scop
, isl_set_copy(domain
), sched
);
593 if (has_affine_break
) {
594 domain
= apply_affine_break(domain
, skip
, 1, 0, NULL
);
595 scop
= pet_scop_intersect_domain_prefix(scop
,
596 isl_set_copy(domain
));
599 scop
= scop_add_break(scop
, id_test
, domain
, isl_val_one(ctx
));
601 isl_set_free(domain
);
606 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
611 * within the context "pc".
613 * Extend the domain of "pc" with an extra inner loop
617 * and construct the scop in scop_from_infinite_loop.
619 static struct pet_scop
*scop_from_infinite_for(__isl_keep pet_tree
*tree
,
620 __isl_keep pet_context
*pc
, struct pet_state
*state
)
622 struct pet_scop
*scop
;
624 pc
= pet_context_copy(pc
);
625 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.l
.body
);
627 pc
= pet_context_add_infinite_loop(pc
);
629 scop
= scop_from_infinite_loop(tree
->u
.l
.body
, pc
, state
);
631 pet_context_free(pc
);
636 /* Construct a pet_scop for a while loop of the form
641 * within the context "pc".
643 * The domain of "pc" has already been extended with an infinite loop
647 * Here, we add the constraints on the outer loop iterators
648 * implied by "pa" and construct the scop in scop_from_infinite_loop.
649 * Note that the intersection with these constraints
650 * may result in an empty loop.
652 static struct pet_scop
*scop_from_affine_while(__isl_keep pet_tree
*tree
,
653 __isl_take isl_pw_aff
*pa
, __isl_take pet_context
*pc
,
654 struct pet_state
*state
)
656 struct pet_scop
*scop
;
657 isl_set
*dom
, *local
;
660 valid
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
661 dom
= isl_pw_aff_non_zero_set(pa
);
662 local
= isl_set_add_dims(isl_set_copy(dom
), isl_dim_set
, 1);
663 pc
= pet_context_intersect_domain(pc
, local
);
664 scop
= scop_from_infinite_loop(tree
->u
.l
.body
, pc
, state
);
665 scop
= pet_scop_restrict(scop
, dom
);
666 scop
= pet_scop_restrict_context(scop
, valid
);
668 pet_context_free(pc
);
672 /* Construct a scop for a while, given the scops for the condition
673 * and the body, the filter identifier and the iteration domain of
676 * In particular, the scop for the condition is filtered to depend
677 * on "id_test" evaluating to true for all previous iterations
678 * of the loop, while the scop for the body is filtered to depend
679 * on "id_test" evaluating to true for all iterations up to the
681 * The actual filter only imposes that this virtual array has
682 * value one on the previous or the current iteration.
683 * The fact that this condition also applies to the previous
684 * iterations is enforced by an implication.
686 * These filtered scops are then combined into a single scop,
687 * with the condition scop scheduled before the body scop.
689 * "sign" is positive if the iterator increases and negative
692 static struct pet_scop
*scop_add_while(struct pet_scop
*scop_cond
,
693 struct pet_scop
*scop_body
, __isl_take isl_id
*id_test
,
694 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
696 isl_ctx
*ctx
= isl_set_get_ctx(domain
);
698 isl_multi_pw_aff
*test_index
;
699 isl_multi_pw_aff
*prev
;
700 int sign
= isl_val_sgn(inc
);
701 struct pet_scop
*scop
;
703 prev
= map_to_previous(isl_id_copy(id_test
), isl_set_copy(domain
), inc
);
704 scop_cond
= pet_scop_filter(scop_cond
, prev
, 1);
706 space
= isl_space_map_from_set(isl_set_get_space(domain
));
707 test_index
= isl_multi_pw_aff_identity(space
);
708 test_index
= isl_multi_pw_aff_set_tuple_id(test_index
, isl_dim_out
,
709 isl_id_copy(id_test
));
710 scop_body
= pet_scop_filter(scop_body
, test_index
, 1);
712 scop
= pet_scop_add_seq(ctx
, scop_cond
, scop_body
);
713 scop
= add_implication(scop
, id_test
, domain
, sign
, 1);
718 /* Create a pet_scop with a single statement with name S_<stmt_nr>,
719 * evaluating "cond" and writing the result to a virtual scalar,
720 * as expressed by "index".
721 * The expression "cond" has not yet been evaluated in the context of "pc".
722 * Do so within the context "pc".
723 * The location of the statement is set to "loc".
725 static struct pet_scop
*scop_from_non_affine_condition(
726 __isl_take pet_expr
*cond
, int stmt_nr
,
727 __isl_take isl_multi_pw_aff
*index
,
728 __isl_take pet_loc
*loc
, __isl_keep pet_context
*pc
)
730 pet_expr
*expr
, *write
;
732 cond
= pet_context_evaluate_expr(pc
, cond
);
734 write
= pet_expr_from_index(index
);
735 write
= pet_expr_access_set_write(write
, 1);
736 write
= pet_expr_access_set_read(write
, 0);
737 expr
= pet_expr_new_binary(1, pet_op_assign
, write
, cond
);
739 return scop_from_evaluated_expr(expr
, stmt_nr
, loc
, pc
);
742 /* Given that "scop" has an affine skip condition of type pet_skip_now,
743 * apply this skip condition to the domain of "pc".
744 * That is, remove the elements satisfying the skip condition from
745 * the domain of "pc".
747 static __isl_give pet_context
*apply_affine_continue(__isl_take pet_context
*pc
,
748 struct pet_scop
*scop
)
750 isl_set
*domain
, *skip
;
752 skip
= pet_scop_get_affine_skip_domain(scop
, pet_skip_now
);
753 domain
= pet_context_get_domain(pc
);
754 domain
= isl_set_subtract(domain
, skip
);
755 pc
= pet_context_intersect_domain(pc
, domain
);
760 /* Add a scop for evaluating the loop increment "inc" add the end
761 * of a loop body "scop" within the context "pc".
763 * The skip conditions resulting from continue statements inside
764 * the body do not apply to "inc", but those resulting from break
765 * statements do need to get applied.
767 static struct pet_scop
*scop_add_inc(struct pet_scop
*scop
,
768 __isl_take pet_expr
*inc
, __isl_take pet_loc
*loc
,
769 __isl_keep pet_context
*pc
, struct pet_state
*state
)
771 struct pet_scop
*scop_inc
;
773 pc
= pet_context_copy(pc
);
775 if (pet_scop_has_skip(scop
, pet_skip_later
)) {
776 isl_multi_pw_aff
*skip
;
777 skip
= pet_scop_get_skip(scop
, pet_skip_later
);
778 scop
= pet_scop_set_skip(scop
, pet_skip_now
, skip
);
779 if (pet_scop_has_affine_skip(scop
, pet_skip_now
))
780 pc
= apply_affine_continue(pc
, scop
);
782 pet_scop_reset_skip(scop
, pet_skip_now
);
783 scop_inc
= scop_from_expr(inc
, state
->n_stmt
++, loc
, pc
);
784 scop_inc
= pet_scop_prefix(scop_inc
, 2);
785 scop
= pet_scop_add_seq(state
->ctx
, scop
, scop_inc
);
787 pet_context_free(pc
);
792 /* Construct a generic while scop, with iteration domain
793 * { [t] : t >= 0 } around the scop for "tree_body" within the context "pc".
794 * The domain of "pc" has already been extended with this infinite loop
798 * The scop consists of two parts,
799 * one for evaluating the condition "cond" and one for the body.
800 * If "expr_inc" is not NULL, then a scop for evaluating this expression
801 * is added at the end of the body,
802 * after replacing any skip conditions resulting from continue statements
803 * by the skip conditions resulting from break statements (if any).
805 * The schedules are combined as a sequence to reflect that the condition is
806 * evaluated before the body is executed and the body is filtered to depend
807 * on the result of the condition evaluating to true on all iterations
808 * up to the current iteration, while the evaluation of the condition itself
809 * is filtered to depend on the result of the condition evaluating to true
810 * on all previous iterations.
811 * The context of the scop representing the body is dropped
812 * because we don't know how many times the body will be executed,
815 * If the body contains any break, then it is taken into
816 * account in apply_affine_break (if the skip condition is affine)
817 * or in scop_add_break (if the skip condition is not affine).
819 * Note that in case of an affine skip condition,
820 * since we are dealing with a loop without loop iterator,
821 * the skip condition cannot refer to the current loop iterator and
822 * so effectively, the effect on the iteration domain is of the form
824 * { [outer,0]; [outer,t] : t >= 1 and not skip }
826 static struct pet_scop
*scop_from_non_affine_while(__isl_take pet_expr
*cond
,
827 __isl_take pet_loc
*loc
, __isl_keep pet_tree
*tree_body
,
828 __isl_take pet_expr
*expr_inc
, __isl_take pet_context
*pc
,
829 struct pet_state
*state
)
832 isl_id
*id_test
, *id_break_test
;
834 isl_multi_pw_aff
*test_index
;
837 isl_multi_aff
*sched
;
838 struct pet_scop
*scop
, *scop_body
;
839 int has_affine_break
;
843 space
= pet_context_get_space(pc
);
844 test_index
= pet_create_test_index(space
, state
->n_test
++);
845 scop
= scop_from_non_affine_condition(cond
, state
->n_stmt
++,
846 isl_multi_pw_aff_copy(test_index
),
847 pet_loc_copy(loc
), pc
);
848 id_test
= isl_multi_pw_aff_get_tuple_id(test_index
, isl_dim_out
);
849 domain
= pet_context_get_domain(pc
);
850 scop
= pet_scop_add_boolean_array(scop
, isl_set_copy(domain
),
851 test_index
, state
->int_size
);
853 sched
= map_to_last(pc
, state
->n_loop
++);
855 scop_body
= scop_from_tree(tree_body
, pc
, state
);
857 has_affine_break
= pet_scop_has_affine_skip(scop_body
, pet_skip_later
);
858 if (has_affine_break
)
859 skip
= pet_scop_get_affine_skip_domain(scop_body
,
861 has_var_break
= pet_scop_has_var_skip(scop_body
, pet_skip_later
);
863 id_break_test
= pet_scop_get_skip_id(scop_body
, pet_skip_later
);
865 scop
= pet_scop_prefix(scop
, 0);
866 scop_body
= pet_scop_reset_context(scop_body
);
867 scop_body
= pet_scop_prefix(scop_body
, 1);
869 scop_body
= scop_add_inc(scop_body
, expr_inc
, loc
, pc
, state
);
872 scop_body
= pet_scop_reset_skips(scop_body
);
874 if (has_affine_break
) {
875 domain
= apply_affine_break(domain
, skip
, 1, 0, NULL
);
876 scop
= pet_scop_intersect_domain_prefix(scop
,
877 isl_set_copy(domain
));
878 scop_body
= pet_scop_intersect_domain_prefix(scop_body
,
879 isl_set_copy(domain
));
882 scop
= scop_add_break(scop
, isl_id_copy(id_break_test
),
883 isl_set_copy(domain
), isl_val_one(ctx
));
884 scop_body
= scop_add_break(scop_body
, id_break_test
,
885 isl_set_copy(domain
), isl_val_one(ctx
));
887 scop
= scop_add_while(scop
, scop_body
, id_test
, isl_set_copy(domain
),
890 scop
= pet_scop_embed(scop
, domain
, sched
);
892 pet_context_free(pc
);
896 /* Check if the while loop is of the form
898 * while (affine expression)
901 * If so, call scop_from_affine_while to construct a scop.
903 * Otherwise, pass control to scop_from_non_affine_while.
905 * "pc" is the context in which the affine expressions in the scop are created.
906 * The domain of "pc" is extended with an infinite loop
910 * before passing control to scop_from_affine_while or
911 * scop_from_non_affine_while.
913 static struct pet_scop
*scop_from_while(__isl_keep pet_tree
*tree
,
914 __isl_keep pet_context
*pc
, struct pet_state
*state
)
922 pc
= pet_context_copy(pc
);
923 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.l
.body
);
925 cond_expr
= pet_expr_copy(tree
->u
.l
.cond
);
926 cond_expr
= pet_context_evaluate_expr(pc
, cond_expr
);
927 pa
= pet_expr_extract_affine_condition(cond_expr
, pc
);
928 pet_expr_free(cond_expr
);
930 pc
= pet_context_add_infinite_loop(pc
);
935 if (!isl_pw_aff_involves_nan(pa
))
936 return scop_from_affine_while(tree
, pa
, pc
, state
);
938 return scop_from_non_affine_while(pet_expr_copy(tree
->u
.l
.cond
),
939 pet_tree_get_loc(tree
), tree
->u
.l
.body
, NULL
,
942 pet_context_free(pc
);
946 /* Check whether "cond" expresses a simple loop bound
947 * on the final set dimension.
948 * In particular, if "up" is set then "cond" should contain only
949 * upper bounds on the final set dimension.
950 * Otherwise, it should contain only lower bounds.
952 static int is_simple_bound(__isl_keep isl_set
*cond
, __isl_keep isl_val
*inc
)
956 pos
= isl_set_dim(cond
, isl_dim_set
) - 1;
957 if (isl_val_is_pos(inc
))
958 return !isl_set_dim_has_any_lower_bound(cond
, isl_dim_set
, pos
);
960 return !isl_set_dim_has_any_upper_bound(cond
, isl_dim_set
, pos
);
963 /* Extend a condition on a given iteration of a loop to one that
964 * imposes the same condition on all previous iterations.
965 * "domain" expresses the lower [upper] bound on the iterations
966 * when inc is positive [negative] in its final dimension.
968 * In particular, we construct the condition (when inc is positive)
970 * forall i' : (domain(i') and i' <= i) => cond(i')
972 * (where "<=" applies to the final dimension)
973 * which is equivalent to
975 * not exists i' : domain(i') and i' <= i and not cond(i')
977 * We construct this set by subtracting the satisfying cond from domain,
980 * { [i'] -> [i] : i' <= i }
982 * and then subtracting the result from domain again.
984 static __isl_give isl_set
*valid_for_each_iteration(__isl_take isl_set
*cond
,
985 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
988 isl_map
*previous_to_this
;
991 dim
= isl_set_dim(cond
, isl_dim_set
);
992 space
= isl_space_map_from_set(isl_set_get_space(cond
));
993 previous_to_this
= isl_map_universe(space
);
994 for (i
= 0; i
+ 1 < dim
; ++i
)
995 previous_to_this
= isl_map_equate(previous_to_this
,
996 isl_dim_in
, i
, isl_dim_out
, i
);
997 if (isl_val_is_pos(inc
))
998 previous_to_this
= isl_map_order_le(previous_to_this
,
999 isl_dim_in
, dim
- 1, isl_dim_out
, dim
- 1);
1001 previous_to_this
= isl_map_order_ge(previous_to_this
,
1002 isl_dim_in
, dim
- 1, isl_dim_out
, dim
- 1);
1004 cond
= isl_set_subtract(isl_set_copy(domain
), cond
);
1005 cond
= isl_set_apply(cond
, previous_to_this
);
1006 cond
= isl_set_subtract(domain
, cond
);
1013 /* Given an initial value of the form
1015 * { [outer,i] -> init(outer) }
1017 * construct a domain of the form
1019 * { [outer,i] : exists a: i = init(outer) + a * inc and a >= 0 }
1021 static __isl_give isl_set
*strided_domain(__isl_take isl_pw_aff
*init
,
1022 __isl_take isl_val
*inc
)
1027 isl_local_space
*ls
;
1030 dim
= isl_pw_aff_dim(init
, isl_dim_in
);
1032 init
= isl_pw_aff_add_dims(init
, isl_dim_in
, 1);
1033 space
= isl_pw_aff_get_domain_space(init
);
1034 ls
= isl_local_space_from_space(space
);
1035 aff
= isl_aff_zero_on_domain(isl_local_space_copy(ls
));
1036 aff
= isl_aff_add_coefficient_val(aff
, isl_dim_in
, dim
, inc
);
1037 init
= isl_pw_aff_add(init
, isl_pw_aff_from_aff(aff
));
1039 aff
= isl_aff_var_on_domain(ls
, isl_dim_set
, dim
- 1);
1040 set
= isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff
), init
);
1042 set
= isl_set_lower_bound_si(set
, isl_dim_set
, dim
, 0);
1043 set
= isl_set_project_out(set
, isl_dim_set
, dim
, 1);
1048 /* Assuming "cond" represents a bound on a loop where the loop
1049 * iterator "iv" is incremented (or decremented) by one, check if wrapping
1052 * Under the given assumptions, wrapping is only possible if "cond" allows
1053 * for the last value before wrapping, i.e., 2^width - 1 in case of an
1054 * increasing iterator and 0 in case of a decreasing iterator.
1056 static int can_wrap(__isl_keep isl_set
*cond
, __isl_keep pet_expr
*iv
,
1057 __isl_keep isl_val
*inc
)
1064 test
= isl_set_copy(cond
);
1066 ctx
= isl_set_get_ctx(test
);
1067 if (isl_val_is_neg(inc
))
1068 limit
= isl_val_zero(ctx
);
1070 limit
= isl_val_int_from_ui(ctx
, pet_expr_get_type_size(iv
));
1071 limit
= isl_val_2exp(limit
);
1072 limit
= isl_val_sub_ui(limit
, 1);
1075 test
= isl_set_fix_val(cond
, isl_dim_set
, 0, limit
);
1076 cw
= !isl_set_is_empty(test
);
1086 * construct the following affine expression on this space
1088 * { [outer, v] -> [outer, v mod 2^width] }
1090 * where width is the number of bits used to represent the values
1091 * of the unsigned variable "iv".
1093 static __isl_give isl_multi_aff
*compute_wrapping(__isl_take isl_space
*space
,
1094 __isl_keep pet_expr
*iv
)
1102 dim
= isl_space_dim(space
, isl_dim_set
);
1104 ctx
= isl_space_get_ctx(space
);
1105 mod
= isl_val_int_from_ui(ctx
, pet_expr_get_type_size(iv
));
1106 mod
= isl_val_2exp(mod
);
1108 space
= isl_space_map_from_set(space
);
1109 ma
= isl_multi_aff_identity(space
);
1111 aff
= isl_multi_aff_get_aff(ma
, dim
- 1);
1112 aff
= isl_aff_mod_val(aff
, mod
);
1113 ma
= isl_multi_aff_set_aff(ma
, dim
- 1, aff
);
1118 /* Given two sets in the space
1122 * where l represents the outer loop iterators, compute the set
1123 * of values of l that ensure that "set1" is a subset of "set2".
1125 * set1 is a subset of set2 if
1127 * forall i: set1(l,i) => set2(l,i)
1131 * not exists i: set1(l,i) and not set2(l,i)
1135 * not exists i: (set1 \ set2)(l,i)
1137 static __isl_give isl_set
*enforce_subset(__isl_take isl_set
*set1
,
1138 __isl_take isl_set
*set2
)
1142 pos
= isl_set_dim(set1
, isl_dim_set
) - 1;
1143 set1
= isl_set_subtract(set1
, set2
);
1144 set1
= isl_set_eliminate(set1
, isl_dim_set
, pos
, 1);
1145 return isl_set_complement(set1
);
1148 /* Compute the set of outer iterator values for which "cond" holds
1149 * on the next iteration of the inner loop for each element of "dom".
1151 * We first construct mapping { [l,i] -> [l,i + inc] } (where l refers
1152 * to the outer loop iterators), plug that into "cond"
1153 * and then compute the set of outer iterators for which "dom" is a subset
1156 static __isl_give isl_set
*valid_on_next(__isl_take isl_set
*cond
,
1157 __isl_take isl_set
*dom
, __isl_take isl_val
*inc
)
1164 pos
= isl_set_dim(dom
, isl_dim_set
) - 1;
1165 space
= isl_set_get_space(dom
);
1166 space
= isl_space_map_from_set(space
);
1167 ma
= isl_multi_aff_identity(space
);
1168 aff
= isl_multi_aff_get_aff(ma
, pos
);
1169 aff
= isl_aff_add_constant_val(aff
, inc
);
1170 ma
= isl_multi_aff_set_aff(ma
, pos
, aff
);
1171 cond
= isl_set_preimage_multi_aff(cond
, ma
);
1173 return enforce_subset(dom
, cond
);
1176 /* Extract the for loop "tree" as a while loop within the context "pc_init".
1177 * In particular, "pc_init" represents the context of the loop,
1178 * whereas "pc" represents the context of the body of the loop and
1179 * has already had its domain extended with an infinite loop
1183 * The for loop has the form
1185 * for (iv = init; cond; iv += inc)
1196 * except that the skips resulting from any continue statements
1197 * in body do not apply to the increment, but are replaced by the skips
1198 * resulting from break statements.
1200 * If the loop iterator is declared in the for loop, then it is killed before
1201 * and after the loop.
1203 static struct pet_scop
*scop_from_non_affine_for(__isl_keep pet_tree
*tree
,
1204 __isl_keep pet_context
*init_pc
, __isl_take pet_context
*pc
,
1205 struct pet_state
*state
)
1209 pet_expr
*expr_iv
, *init
, *inc
;
1210 struct pet_scop
*scop_init
, *scop
;
1212 struct pet_array
*array
;
1213 struct pet_scop
*scop_kill
;
1215 iv
= pet_expr_access_get_id(tree
->u
.l
.iv
);
1216 pc
= pet_context_clear_value(pc
, iv
);
1218 declared
= tree
->u
.l
.declared
;
1220 expr_iv
= pet_expr_copy(tree
->u
.l
.iv
);
1221 type_size
= pet_expr_get_type_size(expr_iv
);
1222 init
= pet_expr_copy(tree
->u
.l
.init
);
1223 init
= pet_expr_new_binary(type_size
, pet_op_assign
, expr_iv
, init
);
1224 scop_init
= scop_from_expr(init
, state
->n_stmt
++,
1225 pet_tree_get_loc(tree
), init_pc
);
1226 scop_init
= pet_scop_prefix(scop_init
, declared
);
1228 expr_iv
= pet_expr_copy(tree
->u
.l
.iv
);
1229 type_size
= pet_expr_get_type_size(expr_iv
);
1230 inc
= pet_expr_copy(tree
->u
.l
.inc
);
1231 inc
= pet_expr_new_binary(type_size
, pet_op_add_assign
, expr_iv
, inc
);
1233 scop
= scop_from_non_affine_while(pet_expr_copy(tree
->u
.l
.cond
),
1234 pet_tree_get_loc(tree
), tree
->u
.l
.body
, inc
,
1235 pet_context_copy(pc
), state
);
1237 scop
= pet_scop_prefix(scop
, declared
+ 1);
1238 scop
= pet_scop_add_seq(state
->ctx
, scop_init
, scop
);
1240 pet_context_free(pc
);
1245 array
= extract_array(tree
->u
.l
.iv
, init_pc
, state
);
1247 array
->declared
= 1;
1248 scop_kill
= kill(pet_tree_get_loc(tree
), array
, init_pc
, state
);
1249 scop_kill
= pet_scop_prefix(scop_kill
, 0);
1250 scop
= pet_scop_add_seq(state
->ctx
, scop_kill
, scop
);
1251 scop_kill
= kill(pet_tree_get_loc(tree
), array
, init_pc
, state
);
1252 scop_kill
= pet_scop_add_array(scop_kill
, array
);
1253 scop_kill
= pet_scop_prefix(scop_kill
, 3);
1254 scop
= pet_scop_add_seq(state
->ctx
, scop
, scop_kill
);
1259 /* Given an access expression "expr", is the variable accessed by
1260 * "expr" assigned anywhere inside "tree"?
1262 static int is_assigned(__isl_keep pet_expr
*expr
, __isl_keep pet_tree
*tree
)
1267 id
= pet_expr_access_get_id(expr
);
1268 assigned
= pet_tree_writes(tree
, id
);
1274 /* Are all nested access parameters in "pa" allowed given "tree".
1275 * In particular, is none of them written by anywhere inside "tree".
1277 * If "tree" has any continue or break nodes in the current loop level,
1278 * then no nested access parameters are allowed.
1279 * In particular, if there is any nested access in a guard
1280 * for a piece of code containing a "continue", then we want to introduce
1281 * a separate statement for evaluating this guard so that we can express
1282 * that the result is false for all previous iterations.
1284 static int is_nested_allowed(__isl_keep isl_pw_aff
*pa
,
1285 __isl_keep pet_tree
*tree
)
1292 if (!pet_nested_any_in_pw_aff(pa
))
1295 if (pet_tree_has_continue_or_break(tree
))
1298 nparam
= isl_pw_aff_dim(pa
, isl_dim_param
);
1299 for (i
= 0; i
< nparam
; ++i
) {
1300 isl_id
*id
= isl_pw_aff_get_dim_id(pa
, isl_dim_param
, i
);
1304 if (!pet_nested_in_id(id
)) {
1309 expr
= pet_nested_extract_expr(id
);
1310 allowed
= pet_expr_get_type(expr
) == pet_expr_access
&&
1311 !is_assigned(expr
, tree
);
1313 pet_expr_free(expr
);
1323 /* Internal data structure for collect_local.
1324 * "pc" and "state" are needed to extract pet_arrays for the local variables.
1325 * "local" collects the results.
1327 struct pet_tree_collect_local_data
{
1329 struct pet_state
*state
;
1330 isl_union_set
*local
;
1333 /* Add the variable accessed by "var" to data->local.
1334 * We extract a representation of the variable from
1335 * the pet_array constructed using extract_array
1336 * to ensure consistency with the rest of the scop.
1338 static int add_local(struct pet_tree_collect_local_data
*data
,
1339 __isl_keep pet_expr
*var
)
1341 struct pet_array
*array
;
1344 array
= extract_array(var
, data
->pc
, data
->state
);
1348 universe
= isl_set_universe(isl_set_get_space(array
->extent
));
1349 data
->local
= isl_union_set_add_set(data
->local
, universe
);
1350 pet_array_free(array
);
1355 /* If the node "tree" declares a variable, then add it to
1358 static int extract_local_var(__isl_keep pet_tree
*tree
, void *user
)
1360 enum pet_tree_type type
;
1361 struct pet_tree_collect_local_data
*data
= user
;
1363 type
= pet_tree_get_type(tree
);
1364 if (type
== pet_tree_decl
|| type
== pet_tree_decl_init
)
1365 return add_local(data
, tree
->u
.d
.var
);
1370 /* If the node "tree" is a for loop that declares its induction variable,
1371 * then add it this induction variable to data->local.
1373 static int extract_local_iterator(__isl_keep pet_tree
*tree
, void *user
)
1375 struct pet_tree_collect_local_data
*data
= user
;
1377 if (pet_tree_get_type(tree
) == pet_tree_for
&& tree
->u
.l
.declared
)
1378 return add_local(data
, tree
->u
.l
.iv
);
1383 /* Collect and return all local variables of the for loop represented
1384 * by "tree", with "scop" the corresponding pet_scop.
1385 * "pc" and "state" are needed to extract pet_arrays for the local variables.
1387 * We collect not only the variables that are declared inside "tree",
1388 * but also the loop iterators that are declared anywhere inside
1389 * any possible macro statements in "scop".
1390 * The latter also appear as declared variable in the scop,
1391 * whereas other declared loop iterators only appear implicitly
1392 * in the iteration domains.
1394 static __isl_give isl_union_set
*collect_local(struct pet_scop
*scop
,
1395 __isl_keep pet_tree
*tree
, __isl_keep pet_context
*pc
,
1396 struct pet_state
*state
)
1400 struct pet_tree_collect_local_data data
= { pc
, state
};
1402 ctx
= pet_tree_get_ctx(tree
);
1403 data
.local
= isl_union_set_empty(isl_space_params_alloc(ctx
, 0));
1405 if (pet_tree_foreach_sub_tree(tree
, &extract_local_var
, &data
) < 0)
1406 return isl_union_set_free(data
.local
);
1408 for (i
= 0; i
< scop
->n_stmt
; ++i
) {
1409 pet_tree
*body
= scop
->stmts
[i
]->body
;
1410 if (pet_tree_foreach_sub_tree(body
, &extract_local_iterator
,
1412 return isl_union_set_free(data
.local
);
1418 /* Add an independence to "scop" if the for node "tree" was marked
1420 * "domain" is the set of loop iterators, with the current for loop
1421 * innermost. If "sign" is positive, then the inner iterator increases.
1422 * Otherwise it decreases.
1423 * "pc" and "state" are needed to extract pet_arrays for the local variables.
1425 * If the tree was marked, then collect all local variables and
1426 * add an independence.
1428 static struct pet_scop
*set_independence(struct pet_scop
*scop
,
1429 __isl_keep pet_tree
*tree
, __isl_keep isl_set
*domain
, int sign
,
1430 __isl_keep pet_context
*pc
, struct pet_state
*state
)
1432 isl_union_set
*local
;
1434 if (!tree
->u
.l
.independent
)
1437 local
= collect_local(scop
, tree
, pc
, state
);
1438 scop
= pet_scop_set_independent(scop
, domain
, local
, sign
);
1443 /* Construct a pet_scop for a for tree with static affine initialization
1444 * and constant increment within the context "pc".
1445 * The domain of "pc" has already been extended with an (at this point
1446 * unbounded) inner loop iterator corresponding to the current for loop.
1448 * The condition is allowed to contain nested accesses, provided
1449 * they are not being written to inside the body of the loop.
1450 * Otherwise, or if the condition is otherwise non-affine, the for loop is
1451 * essentially treated as a while loop, with iteration domain
1452 * { [l,i] : i >= init }, where l refers to the outer loop iterators.
1454 * We extract a pet_scop for the body after intersecting the domain of "pc"
1456 * { [l,i] : i >= init and condition' }
1460 * { [l,i] : i <= init and condition' }
1462 * Where condition' is equal to condition if the latter is
1463 * a simple upper [lower] bound and a condition that is extended
1464 * to apply to all previous iterations otherwise.
1465 * Afterwards, the schedule of the pet_scop is extended with
1473 * If the condition is non-affine, then we drop the condition from the
1474 * iteration domain and instead create a separate statement
1475 * for evaluating the condition. The body is then filtered to depend
1476 * on the result of the condition evaluating to true on all iterations
1477 * up to the current iteration, while the evaluation the condition itself
1478 * is filtered to depend on the result of the condition evaluating to true
1479 * on all previous iterations.
1480 * The context of the scop representing the body is dropped
1481 * because we don't know how many times the body will be executed,
1484 * If the stride of the loop is not 1, then "i >= init" is replaced by
1486 * (exists a: i = init + stride * a and a >= 0)
1488 * If the loop iterator i is unsigned, then wrapping may occur.
1489 * We therefore use a virtual iterator instead that does not wrap.
1490 * However, the condition in the code applies
1491 * to the wrapped value, so we need to change condition(l,i)
1492 * into condition([l,i % 2^width]). Similarly, we replace all accesses
1493 * to the original iterator by the wrapping of the virtual iterator.
1494 * Note that there may be no need to perform this final wrapping
1495 * if the loop condition (after wrapping) satisfies certain conditions.
1496 * However, the is_simple_bound condition is not enough since it doesn't
1497 * check if there even is an upper bound.
1499 * Wrapping on unsigned iterators can be avoided entirely if
1500 * loop condition is simple, the loop iterator is incremented
1501 * [decremented] by one and the last value before wrapping cannot
1502 * possibly satisfy the loop condition.
1504 * Valid outer iterators for a for loop are those for which the initial
1505 * value itself, the increment on each domain iteration and
1506 * the condition on both the initial value and
1507 * the result of incrementing the iterator for each iteration of the domain
1509 * If the loop condition is non-affine, then we only consider validity
1510 * of the initial value.
1512 * If the body contains any break, then we keep track of it in "skip"
1513 * (if the skip condition is affine) or it is handled in scop_add_break
1514 * (if the skip condition is not affine).
1515 * Note that the affine break condition needs to be considered with
1516 * respect to previous iterations in the virtual domain (if any).
1518 static struct pet_scop
*scop_from_affine_for(__isl_keep pet_tree
*tree
,
1519 __isl_take isl_pw_aff
*init_val
, __isl_take isl_pw_aff
*pa_inc
,
1520 __isl_take isl_val
*inc
, __isl_take pet_context
*pc
,
1521 struct pet_state
*state
)
1524 isl_multi_aff
*sched
;
1525 isl_set
*cond
= NULL
;
1526 isl_set
*skip
= NULL
;
1527 isl_id
*id_test
= NULL
, *id_break_test
;
1528 struct pet_scop
*scop
, *scop_cond
= NULL
;
1535 int has_affine_break
;
1537 isl_map
*rev_wrap
= NULL
;
1538 isl_map
*init_val_map
;
1540 isl_set
*valid_init
;
1541 isl_set
*valid_cond
;
1542 isl_set
*valid_cond_init
;
1543 isl_set
*valid_cond_next
;
1545 pet_expr
*cond_expr
;
1546 pet_context
*pc_nested
;
1548 pos
= pet_context_dim(pc
) - 1;
1550 domain
= pet_context_get_domain(pc
);
1551 cond_expr
= pet_expr_copy(tree
->u
.l
.cond
);
1552 cond_expr
= pet_context_evaluate_expr(pc
, cond_expr
);
1553 pc_nested
= pet_context_copy(pc
);
1554 pc_nested
= pet_context_set_allow_nested(pc_nested
, 1);
1555 pa
= pet_expr_extract_affine_condition(cond_expr
, pc_nested
);
1556 pet_context_free(pc_nested
);
1557 pet_expr_free(cond_expr
);
1559 valid_inc
= isl_pw_aff_domain(pa_inc
);
1561 is_unsigned
= pet_expr_get_type_size(tree
->u
.l
.iv
) > 0;
1563 is_non_affine
= isl_pw_aff_involves_nan(pa
) ||
1564 !is_nested_allowed(pa
, tree
->u
.l
.body
);
1566 pa
= isl_pw_aff_free(pa
);
1568 valid_cond
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
1569 cond
= isl_pw_aff_non_zero_set(pa
);
1571 cond
= isl_set_universe(isl_set_get_space(domain
));
1573 valid_cond
= isl_set_coalesce(valid_cond
);
1574 is_one
= isl_val_is_one(inc
) || isl_val_is_negone(inc
);
1575 is_virtual
= is_unsigned
&&
1576 (!is_one
|| can_wrap(cond
, tree
->u
.l
.iv
, inc
));
1578 init_val_map
= isl_map_from_pw_aff(isl_pw_aff_copy(init_val
));
1579 init_val_map
= isl_map_equate(init_val_map
, isl_dim_in
, pos
,
1581 valid_cond_init
= enforce_subset(isl_map_domain(init_val_map
),
1582 isl_set_copy(valid_cond
));
1583 if (is_one
&& !is_virtual
) {
1586 isl_pw_aff_free(init_val
);
1587 pa
= pet_expr_extract_comparison(
1588 isl_val_is_pos(inc
) ? pet_op_ge
: pet_op_le
,
1589 tree
->u
.l
.iv
, tree
->u
.l
.init
, pc
);
1590 valid_init
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
1591 valid_init
= isl_set_eliminate(valid_init
, isl_dim_set
,
1592 isl_set_dim(domain
, isl_dim_set
) - 1, 1);
1593 cond
= isl_pw_aff_non_zero_set(pa
);
1594 domain
= isl_set_intersect(domain
, cond
);
1598 valid_init
= isl_pw_aff_domain(isl_pw_aff_copy(init_val
));
1599 strided
= strided_domain(init_val
, isl_val_copy(inc
));
1600 domain
= isl_set_intersect(domain
, strided
);
1604 isl_multi_aff
*wrap
;
1605 wrap
= compute_wrapping(isl_set_get_space(cond
), tree
->u
.l
.iv
);
1606 pc
= pet_context_preimage_domain(pc
, wrap
);
1607 rev_wrap
= isl_map_from_multi_aff(wrap
);
1608 rev_wrap
= isl_map_reverse(rev_wrap
);
1609 cond
= isl_set_apply(cond
, isl_map_copy(rev_wrap
));
1610 valid_cond
= isl_set_apply(valid_cond
, isl_map_copy(rev_wrap
));
1611 valid_inc
= isl_set_apply(valid_inc
, isl_map_copy(rev_wrap
));
1613 is_simple
= is_simple_bound(cond
, inc
);
1615 cond
= isl_set_gist(cond
, isl_set_copy(domain
));
1616 is_simple
= is_simple_bound(cond
, inc
);
1619 cond
= valid_for_each_iteration(cond
,
1620 isl_set_copy(domain
), isl_val_copy(inc
));
1621 cond
= isl_set_align_params(cond
, isl_set_get_space(domain
));
1622 domain
= isl_set_intersect(domain
, cond
);
1623 sched
= map_to_last(pc
, state
->n_loop
++);
1624 if (isl_val_is_neg(inc
))
1625 sched
= isl_multi_aff_neg(sched
);
1627 valid_cond_next
= valid_on_next(valid_cond
, isl_set_copy(domain
),
1629 valid_inc
= enforce_subset(isl_set_copy(domain
), valid_inc
);
1631 pc
= pet_context_intersect_domain(pc
, isl_set_copy(domain
));
1633 if (is_non_affine
) {
1635 isl_multi_pw_aff
*test_index
;
1636 space
= isl_set_get_space(domain
);
1637 test_index
= pet_create_test_index(space
, state
->n_test
++);
1638 scop_cond
= scop_from_non_affine_condition(
1639 pet_expr_copy(tree
->u
.l
.cond
), state
->n_stmt
++,
1640 isl_multi_pw_aff_copy(test_index
),
1641 pet_tree_get_loc(tree
), pc
);
1642 id_test
= isl_multi_pw_aff_get_tuple_id(test_index
,
1644 scop_cond
= pet_scop_add_boolean_array(scop_cond
,
1645 isl_set_copy(domain
), test_index
,
1647 scop_cond
= pet_scop_prefix(scop_cond
, 0);
1650 scop
= scop_from_tree(tree
->u
.l
.body
, pc
, state
);
1651 has_affine_break
= scop
&&
1652 pet_scop_has_affine_skip(scop
, pet_skip_later
);
1653 if (has_affine_break
)
1654 skip
= pet_scop_get_affine_skip_domain(scop
, pet_skip_later
);
1655 has_var_break
= scop
&& pet_scop_has_var_skip(scop
, pet_skip_later
);
1657 id_break_test
= pet_scop_get_skip_id(scop
, pet_skip_later
);
1658 if (is_non_affine
) {
1659 scop
= pet_scop_reset_context(scop
);
1660 scop
= pet_scop_prefix(scop
, 1);
1662 scop
= pet_scop_reset_skips(scop
);
1663 scop
= pet_scop_resolve_nested(scop
);
1664 if (has_affine_break
) {
1665 domain
= apply_affine_break(domain
, skip
, isl_val_sgn(inc
),
1666 is_virtual
, rev_wrap
);
1667 scop
= pet_scop_intersect_domain_prefix(scop
,
1668 isl_set_copy(domain
));
1670 isl_map_free(rev_wrap
);
1672 scop
= scop_add_break(scop
, id_break_test
, isl_set_copy(domain
),
1675 scop
= scop_add_while(scop_cond
, scop
, id_test
,
1676 isl_set_copy(domain
),
1679 scop
= set_independence(scop
, tree
, domain
, isl_val_sgn(inc
),
1681 scop
= pet_scop_embed(scop
, domain
, sched
);
1682 if (is_non_affine
) {
1683 isl_set_free(valid_inc
);
1685 valid_inc
= isl_set_intersect(valid_inc
, valid_cond_next
);
1686 valid_inc
= isl_set_intersect(valid_inc
, valid_cond_init
);
1687 valid_inc
= isl_set_project_out(valid_inc
, isl_dim_set
, pos
, 1);
1688 scop
= pet_scop_restrict_context(scop
, valid_inc
);
1693 valid_init
= isl_set_project_out(valid_init
, isl_dim_set
, pos
, 1);
1694 scop
= pet_scop_restrict_context(scop
, valid_init
);
1696 pet_context_free(pc
);
1700 /* Construct a pet_scop for a for statement within the context of "pc".
1702 * We update the context to reflect the writes to the loop variable and
1703 * the writes inside the body.
1705 * Then we check if the initialization of the for loop
1706 * is a static affine value and the increment is a constant.
1707 * If so, we construct the pet_scop using scop_from_affine_for.
1708 * Otherwise, we treat the for loop as a while loop
1709 * in scop_from_non_affine_for.
1711 * Note that the initialization and the increment are extracted
1712 * in a context where the current loop iterator has been added
1713 * to the context. If these turn out not be affine, then we
1714 * have reconstruct the body context without an assignment
1715 * to this loop iterator, as this variable will then not be
1716 * treated as a dimension of the iteration domain, but as any
1719 static struct pet_scop
*scop_from_for(__isl_keep pet_tree
*tree
,
1720 __isl_keep pet_context
*init_pc
, struct pet_state
*state
)
1724 isl_pw_aff
*pa_inc
, *init_val
;
1725 pet_context
*pc
, *pc_init_val
;
1730 iv
= pet_expr_access_get_id(tree
->u
.l
.iv
);
1731 pc
= pet_context_copy(init_pc
);
1732 pc
= pet_context_add_inner_iterator(pc
, iv
);
1733 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.l
.body
);
1735 pc_init_val
= pet_context_copy(pc
);
1736 pc_init_val
= pet_context_clear_value(pc_init_val
, isl_id_copy(iv
));
1737 init_val
= pet_expr_extract_affine(tree
->u
.l
.init
, pc_init_val
);
1738 pet_context_free(pc_init_val
);
1739 pa_inc
= pet_expr_extract_affine(tree
->u
.l
.inc
, pc
);
1740 inc
= pet_extract_cst(pa_inc
);
1741 if (!pa_inc
|| !init_val
|| !inc
)
1743 if (!isl_pw_aff_involves_nan(pa_inc
) &&
1744 !isl_pw_aff_involves_nan(init_val
) && !isl_val_is_nan(inc
))
1745 return scop_from_affine_for(tree
, init_val
, pa_inc
, inc
,
1748 isl_pw_aff_free(pa_inc
);
1749 isl_pw_aff_free(init_val
);
1751 pet_context_free(pc
);
1753 pc
= pet_context_copy(init_pc
);
1754 pc
= pet_context_add_infinite_loop(pc
);
1755 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.l
.body
);
1756 return scop_from_non_affine_for(tree
, init_pc
, pc
, state
);
1758 isl_pw_aff_free(pa_inc
);
1759 isl_pw_aff_free(init_val
);
1761 pet_context_free(pc
);
1765 /* Check whether "expr" is an affine constraint within the context "pc".
1767 static int is_affine_condition(__isl_keep pet_expr
*expr
,
1768 __isl_keep pet_context
*pc
)
1773 pa
= pet_expr_extract_affine_condition(expr
, pc
);
1776 is_affine
= !isl_pw_aff_involves_nan(pa
);
1777 isl_pw_aff_free(pa
);
1782 /* Check if the given if statement is a conditional assignement
1783 * with a non-affine condition.
1785 * In particular we check if "stmt" is of the form
1792 * where the condition is non-affine and a is some array or scalar access.
1794 static int is_conditional_assignment(__isl_keep pet_tree
*tree
,
1795 __isl_keep pet_context
*pc
)
1799 pet_expr
*expr1
, *expr2
;
1801 ctx
= pet_tree_get_ctx(tree
);
1802 if (!pet_options_get_detect_conditional_assignment(ctx
))
1804 if (tree
->type
!= pet_tree_if_else
)
1806 if (tree
->u
.i
.then_body
->type
!= pet_tree_expr
)
1808 if (tree
->u
.i
.else_body
->type
!= pet_tree_expr
)
1810 expr1
= tree
->u
.i
.then_body
->u
.e
.expr
;
1811 expr2
= tree
->u
.i
.else_body
->u
.e
.expr
;
1812 if (pet_expr_get_type(expr1
) != pet_expr_op
)
1814 if (pet_expr_get_type(expr2
) != pet_expr_op
)
1816 if (pet_expr_op_get_type(expr1
) != pet_op_assign
)
1818 if (pet_expr_op_get_type(expr2
) != pet_op_assign
)
1820 expr1
= pet_expr_get_arg(expr1
, 0);
1821 expr2
= pet_expr_get_arg(expr2
, 0);
1822 equal
= pet_expr_is_equal(expr1
, expr2
);
1823 pet_expr_free(expr1
);
1824 pet_expr_free(expr2
);
1825 if (equal
< 0 || !equal
)
1827 if (is_affine_condition(tree
->u
.i
.cond
, pc
))
1833 /* Given that "tree" is of the form
1840 * where a is some array or scalar access, construct a pet_scop
1841 * corresponding to this conditional assignment within the context "pc".
1842 * "cond_pa" is an affine expression with nested accesses representing
1845 * The constructed pet_scop then corresponds to the expression
1847 * a = condition ? f(...) : g(...)
1849 * All access relations in f(...) are intersected with condition
1850 * while all access relation in g(...) are intersected with the complement.
1852 static struct pet_scop
*scop_from_conditional_assignment(
1853 __isl_keep pet_tree
*tree
, __isl_take isl_pw_aff
*cond_pa
,
1854 __isl_take pet_context
*pc
, struct pet_state
*state
)
1857 isl_set
*cond
, *comp
;
1858 isl_multi_pw_aff
*index
;
1859 pet_expr
*expr1
, *expr2
;
1860 pet_expr
*pe_cond
, *pe_then
, *pe_else
, *pe
, *pe_write
;
1861 struct pet_scop
*scop
;
1863 cond
= isl_pw_aff_non_zero_set(isl_pw_aff_copy(cond_pa
));
1864 comp
= isl_pw_aff_zero_set(isl_pw_aff_copy(cond_pa
));
1865 index
= isl_multi_pw_aff_from_pw_aff(cond_pa
);
1867 expr1
= tree
->u
.i
.then_body
->u
.e
.expr
;
1868 expr2
= tree
->u
.i
.else_body
->u
.e
.expr
;
1870 pe_cond
= pet_expr_from_index(index
);
1872 pe_then
= pet_expr_get_arg(expr1
, 1);
1873 pe_then
= pet_context_evaluate_expr(pc
, pe_then
);
1874 pe_then
= pet_expr_restrict(pe_then
, cond
);
1875 pe_else
= pet_expr_get_arg(expr2
, 1);
1876 pe_else
= pet_context_evaluate_expr(pc
, pe_else
);
1877 pe_else
= pet_expr_restrict(pe_else
, comp
);
1878 pe_write
= pet_expr_get_arg(expr1
, 0);
1879 pe_write
= pet_context_evaluate_expr(pc
, pe_write
);
1881 pe
= pet_expr_new_ternary(pe_cond
, pe_then
, pe_else
);
1882 type_size
= pet_expr_get_type_size(pe_write
);
1883 pe
= pet_expr_new_binary(type_size
, pet_op_assign
, pe_write
, pe
);
1885 scop
= scop_from_evaluated_expr(pe
, state
->n_stmt
++,
1886 pet_tree_get_loc(tree
), pc
);
1888 pet_context_free(pc
);
1893 /* Construct a pet_scop for a non-affine if statement within the context "pc".
1895 * We create a separate statement that writes the result
1896 * of the non-affine condition to a virtual scalar.
1897 * A constraint requiring the value of this virtual scalar to be one
1898 * is added to the iteration domains of the then branch.
1899 * Similarly, a constraint requiring the value of this virtual scalar
1900 * to be zero is added to the iteration domains of the else branch, if any.
1901 * We combine the schedules as a sequence to ensure that the virtual scalar
1902 * is written before it is read.
1904 * If there are any breaks or continues in the then and/or else
1905 * branches, then we may have to compute a new skip condition.
1906 * This is handled using a pet_skip_info object.
1907 * On initialization, the object checks if skip conditions need
1908 * to be computed. If so, it does so in pet_skip_info_if_extract_index and
1909 * adds them in pet_skip_info_if_add.
1911 static struct pet_scop
*scop_from_non_affine_if(__isl_keep pet_tree
*tree
,
1912 __isl_take pet_context
*pc
, struct pet_state
*state
)
1917 isl_multi_pw_aff
*test_index
;
1918 struct pet_skip_info skip
;
1919 struct pet_scop
*scop
, *scop_then
, *scop_else
= NULL
;
1921 has_else
= tree
->type
== pet_tree_if_else
;
1923 space
= pet_context_get_space(pc
);
1924 test_index
= pet_create_test_index(space
, state
->n_test
++);
1925 scop
= scop_from_non_affine_condition(pet_expr_copy(tree
->u
.i
.cond
),
1926 state
->n_stmt
++, isl_multi_pw_aff_copy(test_index
),
1927 pet_tree_get_loc(tree
), pc
);
1928 domain
= pet_context_get_domain(pc
);
1929 scop
= pet_scop_add_boolean_array(scop
, domain
,
1930 isl_multi_pw_aff_copy(test_index
), state
->int_size
);
1932 scop_then
= scop_from_tree(tree
->u
.i
.then_body
, pc
, state
);
1934 scop_else
= scop_from_tree(tree
->u
.i
.else_body
, pc
, state
);
1936 pet_skip_info_if_init(&skip
, state
->ctx
, scop_then
, scop_else
,
1938 pet_skip_info_if_extract_index(&skip
, test_index
, pc
, state
);
1940 scop
= pet_scop_prefix(scop
, 0);
1941 scop_then
= pet_scop_prefix(scop_then
, 1);
1942 scop_then
= pet_scop_filter(scop_then
,
1943 isl_multi_pw_aff_copy(test_index
), 1);
1945 scop_else
= pet_scop_prefix(scop_else
, 1);
1946 scop_else
= pet_scop_filter(scop_else
, test_index
, 0);
1947 scop_then
= pet_scop_add_par(state
->ctx
, scop_then
, scop_else
);
1949 isl_multi_pw_aff_free(test_index
);
1951 scop
= pet_scop_add_seq(state
->ctx
, scop
, scop_then
);
1953 scop
= pet_skip_info_if_add(&skip
, scop
, 2);
1955 pet_context_free(pc
);
1959 /* Construct a pet_scop for an affine if statement within the context "pc".
1961 * The condition is added to the iteration domains of the then branch,
1962 * while the opposite of the condition in added to the iteration domains
1963 * of the else branch, if any.
1965 * If there are any breaks or continues in the then and/or else
1966 * branches, then we may have to compute a new skip condition.
1967 * This is handled using a pet_skip_info_if object.
1968 * On initialization, the object checks if skip conditions need
1969 * to be computed. If so, it does so in pet_skip_info_if_extract_cond and
1970 * adds them in pet_skip_info_if_add.
1972 static struct pet_scop
*scop_from_affine_if(__isl_keep pet_tree
*tree
,
1973 __isl_take isl_pw_aff
*cond
, __isl_take pet_context
*pc
,
1974 struct pet_state
*state
)
1978 isl_set
*set
, *complement
;
1980 struct pet_skip_info skip
;
1981 struct pet_scop
*scop
, *scop_then
, *scop_else
= NULL
;
1982 pet_context
*pc_body
;
1984 ctx
= pet_tree_get_ctx(tree
);
1986 has_else
= tree
->type
== pet_tree_if_else
;
1988 valid
= isl_pw_aff_domain(isl_pw_aff_copy(cond
));
1989 set
= isl_pw_aff_non_zero_set(isl_pw_aff_copy(cond
));
1991 pc_body
= pet_context_copy(pc
);
1992 pc_body
= pet_context_intersect_domain(pc_body
, isl_set_copy(set
));
1993 scop_then
= scop_from_tree(tree
->u
.i
.then_body
, pc_body
, state
);
1994 pet_context_free(pc_body
);
1996 pc_body
= pet_context_copy(pc
);
1997 complement
= isl_set_copy(valid
);
1998 complement
= isl_set_subtract(valid
, isl_set_copy(set
));
1999 pc_body
= pet_context_intersect_domain(pc_body
,
2000 isl_set_copy(complement
));
2001 scop_else
= scop_from_tree(tree
->u
.i
.else_body
, pc_body
, state
);
2002 pet_context_free(pc_body
);
2005 pet_skip_info_if_init(&skip
, ctx
, scop_then
, scop_else
, has_else
, 1);
2006 pet_skip_info_if_extract_cond(&skip
, cond
, pc
, state
);
2007 isl_pw_aff_free(cond
);
2009 scop
= pet_scop_restrict(scop_then
, set
);
2012 scop_else
= pet_scop_restrict(scop_else
, complement
);
2013 scop
= pet_scop_add_par(ctx
, scop
, scop_else
);
2015 scop
= pet_scop_resolve_nested(scop
);
2016 scop
= pet_scop_restrict_context(scop
, valid
);
2018 if (pet_skip_info_has_skip(&skip
))
2019 scop
= pet_scop_prefix(scop
, 0);
2020 scop
= pet_skip_info_if_add(&skip
, scop
, 1);
2022 pet_context_free(pc
);
2026 /* Construct a pet_scop for an if statement within the context "pc".
2028 * If the condition fits the pattern of a conditional assignment,
2029 * then it is handled by scop_from_conditional_assignment.
2030 * Note that the condition is only considered for a conditional assignment
2031 * if it is not static-affine. However, it should still convert
2032 * to an affine expression when nesting is allowed.
2034 * Otherwise, we check if the condition is affine.
2035 * If so, we construct the scop in scop_from_affine_if.
2036 * Otherwise, we construct the scop in scop_from_non_affine_if.
2038 * We allow the condition to be dynamic, i.e., to refer to
2039 * scalars or array elements that may be written to outside
2040 * of the given if statement. These nested accesses are then represented
2041 * as output dimensions in the wrapping iteration domain.
2042 * If it is also written _inside_ the then or else branch, then
2043 * we treat the condition as non-affine.
2044 * As explained in extract_non_affine_if, this will introduce
2045 * an extra statement.
2046 * For aesthetic reasons, we want this statement to have a statement
2047 * number that is lower than those of the then and else branches.
2048 * In order to evaluate if we will need such a statement, however, we
2049 * first construct scops for the then and else branches.
2050 * We therefore reserve a statement number if we might have to
2051 * introduce such an extra statement.
2053 static struct pet_scop
*scop_from_if(__isl_keep pet_tree
*tree
,
2054 __isl_keep pet_context
*pc
, struct pet_state
*state
)
2058 pet_expr
*cond_expr
;
2059 pet_context
*pc_nested
;
2064 has_else
= tree
->type
== pet_tree_if_else
;
2066 pc
= pet_context_copy(pc
);
2067 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.i
.then_body
);
2069 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.i
.else_body
);
2071 cond_expr
= pet_expr_copy(tree
->u
.i
.cond
);
2072 cond_expr
= pet_context_evaluate_expr(pc
, cond_expr
);
2073 pc_nested
= pet_context_copy(pc
);
2074 pc_nested
= pet_context_set_allow_nested(pc_nested
, 1);
2075 cond
= pet_expr_extract_affine_condition(cond_expr
, pc_nested
);
2076 pet_context_free(pc_nested
);
2077 pet_expr_free(cond_expr
);
2080 pet_context_free(pc
);
2084 if (isl_pw_aff_involves_nan(cond
)) {
2085 isl_pw_aff_free(cond
);
2086 return scop_from_non_affine_if(tree
, pc
, state
);
2089 if (is_conditional_assignment(tree
, pc
))
2090 return scop_from_conditional_assignment(tree
, cond
, pc
, state
);
2092 if ((!is_nested_allowed(cond
, tree
->u
.i
.then_body
) ||
2093 (has_else
&& !is_nested_allowed(cond
, tree
->u
.i
.else_body
)))) {
2094 isl_pw_aff_free(cond
);
2095 return scop_from_non_affine_if(tree
, pc
, state
);
2098 return scop_from_affine_if(tree
, cond
, pc
, state
);
2101 /* Return a one-dimensional multi piecewise affine expression that is equal
2102 * to the constant 1 and is defined over the given domain.
2104 static __isl_give isl_multi_pw_aff
*one_mpa(__isl_take isl_space
*space
)
2106 isl_local_space
*ls
;
2109 ls
= isl_local_space_from_space(space
);
2110 aff
= isl_aff_zero_on_domain(ls
);
2111 aff
= isl_aff_set_constant_si(aff
, 1);
2113 return isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff
));
2116 /* Construct a pet_scop for a continue statement with the given domain space.
2118 * We simply create an empty scop with a universal pet_skip_now
2119 * skip condition. This skip condition will then be taken into
2120 * account by the enclosing loop construct, possibly after
2121 * being incorporated into outer skip conditions.
2123 static struct pet_scop
*scop_from_continue(__isl_keep pet_tree
*tree
,
2124 __isl_take isl_space
*space
)
2126 struct pet_scop
*scop
;
2128 scop
= pet_scop_empty(isl_space_copy(space
));
2130 scop
= pet_scop_set_skip(scop
, pet_skip_now
, one_mpa(space
));
2135 /* Construct a pet_scop for a break statement with the given domain space.
2137 * We simply create an empty scop with both a universal pet_skip_now
2138 * skip condition and a universal pet_skip_later skip condition.
2139 * These skip conditions will then be taken into
2140 * account by the enclosing loop construct, possibly after
2141 * being incorporated into outer skip conditions.
2143 static struct pet_scop
*scop_from_break(__isl_keep pet_tree
*tree
,
2144 __isl_take isl_space
*space
)
2146 struct pet_scop
*scop
;
2147 isl_multi_pw_aff
*skip
;
2149 scop
= pet_scop_empty(isl_space_copy(space
));
2151 skip
= one_mpa(space
);
2152 scop
= pet_scop_set_skip(scop
, pet_skip_now
,
2153 isl_multi_pw_aff_copy(skip
));
2154 scop
= pet_scop_set_skip(scop
, pet_skip_later
, skip
);
2159 /* Extract a clone of the kill statement in "scop".
2160 * The domain of the clone is given by "domain".
2161 * "scop" is expected to have been created from a DeclStmt
2162 * and should have the kill as its first statement.
2164 static struct pet_scop
*extract_kill(__isl_keep isl_set
*domain
,
2165 struct pet_scop
*scop
, struct pet_state
*state
)
2168 struct pet_stmt
*stmt
;
2170 isl_multi_pw_aff
*mpa
;
2173 if (!domain
|| !scop
)
2175 if (scop
->n_stmt
< 1)
2176 isl_die(isl_set_get_ctx(domain
), isl_error_internal
,
2177 "expecting at least one statement", return NULL
);
2178 stmt
= scop
->stmts
[0];
2179 if (!pet_stmt_is_kill(stmt
))
2180 isl_die(isl_set_get_ctx(domain
), isl_error_internal
,
2181 "expecting kill statement", return NULL
);
2183 kill
= pet_tree_expr_get_expr(stmt
->body
);
2184 space
= pet_stmt_get_space(stmt
);
2185 space
= isl_space_map_from_set(space
);
2186 mpa
= isl_multi_pw_aff_identity(space
);
2187 mpa
= isl_multi_pw_aff_reset_tuple_id(mpa
, isl_dim_in
);
2188 kill
= pet_expr_update_domain(kill
, mpa
);
2189 tree
= pet_tree_new_expr(kill
);
2190 tree
= pet_tree_set_loc(tree
, pet_loc_copy(stmt
->loc
));
2191 stmt
= pet_stmt_from_pet_tree(isl_set_copy(domain
),
2192 state
->n_stmt
++, tree
);
2193 return pet_scop_from_pet_stmt(isl_set_get_space(domain
), stmt
);
2196 /* Does "tree" represent an assignment to a variable?
2198 * The assignment may be one of
2199 * - a declaration with initialization
2200 * - an expression with a top-level assignment operator
2202 static int is_assignment(__isl_keep pet_tree
*tree
)
2206 if (tree
->type
== pet_tree_decl_init
)
2208 return pet_tree_is_assign(tree
);
2211 /* Update "pc" by taking into account the assignment performed by "tree",
2212 * where "tree" satisfies is_assignment.
2214 * In particular, if the lhs of the assignment is a scalar variable and
2215 * if the rhs is an affine expression, then keep track of this value in "pc"
2216 * so that we can plug it in when we later come across the same variable.
2218 * Any previously assigned value to the variable has already been removed
2219 * by scop_handle_writes.
2221 static __isl_give pet_context
*handle_assignment(__isl_take pet_context
*pc
,
2222 __isl_keep pet_tree
*tree
)
2224 pet_expr
*var
, *val
;
2228 if (pet_tree_get_type(tree
) == pet_tree_decl_init
) {
2229 var
= pet_tree_decl_get_var(tree
);
2230 val
= pet_tree_decl_get_init(tree
);
2233 expr
= pet_tree_expr_get_expr(tree
);
2234 var
= pet_expr_get_arg(expr
, 0);
2235 val
= pet_expr_get_arg(expr
, 1);
2236 pet_expr_free(expr
);
2239 if (!pet_expr_is_scalar_access(var
)) {
2245 pa
= pet_expr_extract_affine(val
, pc
);
2247 pc
= pet_context_free(pc
);
2249 if (!isl_pw_aff_involves_nan(pa
)) {
2250 id
= pet_expr_access_get_id(var
);
2251 pc
= pet_context_set_value(pc
, id
, pa
);
2253 isl_pw_aff_free(pa
);
2261 /* Mark all arrays in "scop" as being exposed.
2263 static struct pet_scop
*mark_exposed(struct pet_scop
*scop
)
2269 for (i
= 0; i
< scop
->n_array
; ++i
)
2270 scop
->arrays
[i
]->exposed
= 1;
2274 /* Try and construct a pet_scop corresponding to (part of)
2275 * a sequence of statements within the context "pc".
2277 * After extracting a statement, we update "pc"
2278 * based on the top-level assignments in the statement
2279 * so that we can exploit them in subsequent statements in the same block.
2281 * If there are any breaks or continues in the individual statements,
2282 * then we may have to compute a new skip condition.
2283 * This is handled using a pet_skip_info object.
2284 * On initialization, the object checks if skip conditions need
2285 * to be computed. If so, it does so in pet_skip_info_seq_extract and
2286 * adds them in pet_skip_info_seq_add.
2288 * If "block" is set, then we need to insert kill statements at
2289 * the end of the block for any array that has been declared by
2290 * one of the statements in the sequence. Each of these declarations
2291 * results in the construction of a kill statement at the place
2292 * of the declaration, so we simply collect duplicates of
2293 * those kill statements and append these duplicates to the constructed scop.
2295 * If "block" is not set, then any array declared by one of the statements
2296 * in the sequence is marked as being exposed.
2298 * If autodetect is set, then we allow the extraction of only a subrange
2299 * of the sequence of statements. However, if there is at least one statement
2300 * for which we could not construct a scop and the final range contains
2301 * either no statements or at least one kill, then we discard the entire
2304 static struct pet_scop
*scop_from_block(__isl_keep pet_tree
*tree
,
2305 __isl_keep pet_context
*pc
, struct pet_state
*state
)
2311 struct pet_scop
*scop
, *kills
;
2313 ctx
= pet_tree_get_ctx(tree
);
2315 space
= pet_context_get_space(pc
);
2316 domain
= pet_context_get_domain(pc
);
2317 pc
= pet_context_copy(pc
);
2318 scop
= pet_scop_empty(isl_space_copy(space
));
2319 kills
= pet_scop_empty(space
);
2320 for (i
= 0; i
< tree
->u
.b
.n
; ++i
) {
2321 struct pet_scop
*scop_i
;
2323 if (pet_scop_has_affine_skip(scop
, pet_skip_now
))
2324 pc
= apply_affine_continue(pc
, scop
);
2325 scop_i
= scop_from_tree(tree
->u
.b
.child
[i
], pc
, state
);
2326 pc
= scop_handle_writes(scop_i
, pc
);
2327 if (is_assignment(tree
->u
.b
.child
[i
]))
2328 pc
= handle_assignment(pc
, tree
->u
.b
.child
[i
]);
2329 struct pet_skip_info skip
;
2330 pet_skip_info_seq_init(&skip
, ctx
, scop
, scop_i
);
2331 pet_skip_info_seq_extract(&skip
, pc
, state
);
2332 if (pet_skip_info_has_skip(&skip
))
2333 scop_i
= pet_scop_prefix(scop_i
, 0);
2334 if (scop_i
&& pet_tree_is_decl(tree
->u
.b
.child
[i
])) {
2335 if (tree
->u
.b
.block
) {
2336 struct pet_scop
*kill
;
2337 kill
= extract_kill(domain
, scop_i
, state
);
2338 kills
= pet_scop_add_par(ctx
, kills
, kill
);
2340 scop_i
= mark_exposed(scop_i
);
2342 scop_i
= pet_scop_prefix(scop_i
, i
);
2343 scop
= pet_scop_add_seq(ctx
, scop
, scop_i
);
2345 scop
= pet_skip_info_seq_add(&skip
, scop
, i
);
2350 isl_set_free(domain
);
2352 kills
= pet_scop_prefix(kills
, tree
->u
.b
.n
);
2353 scop
= pet_scop_add_seq(ctx
, scop
, kills
);
2355 pet_context_free(pc
);
2360 /* Internal data structure for extract_declared_arrays.
2362 * "pc" and "state" are used to create pet_array objects and kill statements.
2363 * "any" is initialized to 0 by the caller and set to 1 as soon as we have
2364 * found any declared array.
2365 * "scop" has been initialized by the caller and is used to attach
2366 * the created pet_array objects.
2367 * "kill_before" and "kill_after" are created and updated by
2368 * extract_declared_arrays to collect the kills of the arrays.
2370 struct pet_tree_extract_declared_arrays_data
{
2372 struct pet_state
*state
;
2377 struct pet_scop
*scop
;
2378 struct pet_scop
*kill_before
;
2379 struct pet_scop
*kill_after
;
2382 /* Check if the node "node" declares any array or scalar.
2383 * If so, create the corresponding pet_array and attach it to data->scop.
2384 * Additionally, create two kill statements for the array and add them
2385 * to data->kill_before and data->kill_after.
2387 static int extract_declared_arrays(__isl_keep pet_tree
*node
, void *user
)
2389 enum pet_tree_type type
;
2390 struct pet_tree_extract_declared_arrays_data
*data
= user
;
2391 struct pet_array
*array
;
2392 struct pet_scop
*scop_kill
;
2395 type
= pet_tree_get_type(node
);
2396 if (type
== pet_tree_decl
|| type
== pet_tree_decl_init
)
2397 var
= node
->u
.d
.var
;
2398 else if (type
== pet_tree_for
&& node
->u
.l
.declared
)
2403 array
= extract_array(var
, data
->pc
, data
->state
);
2405 array
->declared
= 1;
2406 data
->scop
= pet_scop_add_array(data
->scop
, array
);
2408 scop_kill
= kill(pet_tree_get_loc(node
), array
, data
->pc
, data
->state
);
2410 data
->kill_before
= scop_kill
;
2412 data
->kill_before
= pet_scop_add_par(data
->ctx
,
2413 data
->kill_before
, scop_kill
);
2415 scop_kill
= kill(pet_tree_get_loc(node
), array
, data
->pc
, data
->state
);
2417 data
->kill_after
= scop_kill
;
2419 data
->kill_after
= pet_scop_add_par(data
->ctx
,
2420 data
->kill_after
, scop_kill
);
2427 /* Convert a pet_tree that consists of more than a single leaf
2428 * to a pet_scop with a single statement encapsulating the entire pet_tree.
2429 * Do so within the context of "pc".
2431 * After constructing the core scop, we also look for any arrays (or scalars)
2432 * that are declared inside "tree". Each of those arrays is marked as
2433 * having been declared and kill statements for these arrays
2434 * are introduced before and after the core scop.
2435 * Note that the input tree is not a leaf so that the declaration
2436 * cannot occur at the outer level.
2438 static struct pet_scop
*scop_from_tree_macro(__isl_take pet_tree
*tree
,
2439 __isl_take isl_id
*label
, __isl_keep pet_context
*pc
,
2440 struct pet_state
*state
)
2442 struct pet_tree_extract_declared_arrays_data data
= { pc
, state
};
2444 data
.scop
= scop_from_unevaluated_tree(pet_tree_copy(tree
),
2445 state
->n_stmt
++, pc
);
2448 data
.ctx
= pet_context_get_ctx(pc
);
2449 if (pet_tree_foreach_sub_tree(tree
, &extract_declared_arrays
,
2451 data
.scop
= pet_scop_free(data
.scop
);
2452 pet_tree_free(tree
);
2457 data
.kill_before
= pet_scop_prefix(data
.kill_before
, 0);
2458 data
.scop
= pet_scop_prefix(data
.scop
, 1);
2459 data
.kill_after
= pet_scop_prefix(data
.kill_after
, 2);
2461 data
.scop
= pet_scop_add_seq(data
.ctx
, data
.kill_before
, data
.scop
);
2462 data
.scop
= pet_scop_add_seq(data
.ctx
, data
.scop
, data
.kill_after
);
2467 /* Construct a pet_scop that corresponds to the pet_tree "tree"
2468 * within the context "pc" by calling the appropriate function
2469 * based on the type of "tree".
2471 * If the initially constructed pet_scop turns out to involve
2472 * dynamic control and if the user has requested an encapsulation
2473 * of all dynamic control, then this pet_scop is discarded and
2474 * a new pet_scop is created with a single statement representing
2475 * the entire "tree".
2476 * However, if the scop contains any active continue or break,
2477 * then we need to include the loop containing the continue or break
2478 * in the encapsulation. We therefore postpone the encapsulation
2479 * until we have constructed a pet_scop for this enclosing loop.
2481 static struct pet_scop
*scop_from_tree(__isl_keep pet_tree
*tree
,
2482 __isl_keep pet_context
*pc
, struct pet_state
*state
)
2485 struct pet_scop
*scop
= NULL
;
2490 ctx
= pet_tree_get_ctx(tree
);
2491 switch (tree
->type
) {
2492 case pet_tree_error
:
2494 case pet_tree_block
:
2495 return scop_from_block(tree
, pc
, state
);
2496 case pet_tree_break
:
2497 return scop_from_break(tree
, pet_context_get_space(pc
));
2498 case pet_tree_continue
:
2499 return scop_from_continue(tree
, pet_context_get_space(pc
));
2501 case pet_tree_decl_init
:
2502 return scop_from_decl(tree
, pc
, state
);
2504 return scop_from_tree_expr(tree
, pc
, state
);
2506 case pet_tree_if_else
:
2507 scop
= scop_from_if(tree
, pc
, state
);
2510 scop
= scop_from_for(tree
, pc
, state
);
2512 case pet_tree_while
:
2513 scop
= scop_from_while(tree
, pc
, state
);
2515 case pet_tree_infinite_loop
:
2516 scop
= scop_from_infinite_for(tree
, pc
, state
);
2523 if (!pet_options_get_encapsulate_dynamic_control(ctx
) ||
2524 !pet_scop_has_data_dependent_conditions(scop
) ||
2525 pet_scop_has_var_skip(scop
, pet_skip_now
))
2528 pet_scop_free(scop
);
2529 return scop_from_tree_macro(pet_tree_copy(tree
),
2530 isl_id_copy(tree
->label
), pc
, state
);
2533 /* If "tree" has a label that is of the form S_<nr>, then make
2534 * sure that state->n_stmt is greater than nr to ensure that
2535 * we will not generate S_<nr> ourselves.
2537 static int set_first_stmt(__isl_keep pet_tree
*tree
, void *user
)
2539 struct pet_state
*state
= user
;
2547 name
= isl_id_get_name(tree
->label
);
2548 if (strncmp(name
, "S_", 2) != 0)
2550 nr
= atoi(name
+ 2);
2551 if (nr
>= state
->n_stmt
)
2552 state
->n_stmt
= nr
+ 1;
2557 /* Construct a pet_scop that corresponds to the pet_tree "tree".
2558 * "int_size" is the number of bytes need to represent an integer.
2559 * "extract_array" is a callback that we can use to create a pet_array
2560 * that corresponds to the variable accessed by an expression.
2562 * Initialize the global state, construct a context and then
2563 * construct the pet_scop by recursively visiting the tree.
2565 * state.n_stmt is initialized to point beyond any explicit S_<nr> label.
2567 struct pet_scop
*pet_scop_from_pet_tree(__isl_take pet_tree
*tree
, int int_size
,
2568 struct pet_array
*(*extract_array
)(__isl_keep pet_expr
*access
,
2569 __isl_keep pet_context
*pc
, void *user
), void *user
,
2570 __isl_keep pet_context
*pc
)
2572 struct pet_scop
*scop
;
2573 struct pet_state state
= { 0 };
2578 state
.ctx
= pet_tree_get_ctx(tree
);
2579 state
.int_size
= int_size
;
2580 state
.extract_array
= extract_array
;
2582 if (pet_tree_foreach_sub_tree(tree
, &set_first_stmt
, &state
) < 0)
2583 tree
= pet_tree_free(tree
);
2585 scop
= scop_from_tree(tree
, pc
, &state
);
2586 scop
= pet_scop_set_loc(scop
, pet_tree_get_loc(tree
));
2588 pet_tree_free(tree
);
2591 scop
->context
= isl_set_params(scop
->context
);