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 ctx
= pet_tree_get_ctx(tree
);
243 scop
= pet_scop_add_seq(ctx
, scop_decl
, scop
);
248 /* Does "tree" represent a kill statement?
249 * That is, is it an expression statement that "calls" __pencil_kill?
251 static int is_pencil_kill(__isl_keep pet_tree
*tree
)
258 if (tree
->type
!= pet_tree_expr
)
260 expr
= tree
->u
.e
.expr
;
261 if (pet_expr_get_type(expr
) != pet_expr_call
)
263 name
= pet_expr_call_get_name(expr
);
266 return !strcmp(name
, "__pencil_kill");
269 /* Add a kill to "scop" that kills what is accessed by
270 * the access expression "expr".
272 * If the access expression has any arguments (after evaluation
273 * in the context of "pc"), then we ignore it, since we cannot
274 * tell which elements are definitely killed.
276 * Otherwise, we extend the index expression to the dimension
277 * of the accessed array and intersect with the extent of the array and
278 * add a kill expression that kills these array elements is added to "scop".
280 static struct pet_scop
*scop_add_kill(struct pet_scop
*scop
,
281 __isl_take pet_expr
*expr
, __isl_take pet_loc
*loc
,
282 __isl_keep pet_context
*pc
, struct pet_state
*state
)
286 isl_multi_pw_aff
*index
;
289 struct pet_array
*array
;
290 struct pet_scop
*scop_i
;
292 expr
= pet_context_evaluate_expr(pc
, expr
);
295 if (expr
->n_arg
!= 0) {
299 array
= extract_array(expr
, pc
, state
);
302 index
= pet_expr_access_get_index(expr
);
304 map
= isl_map_from_multi_pw_aff(isl_multi_pw_aff_copy(index
));
305 id
= isl_map_get_tuple_id(map
, isl_dim_out
);
306 dim1
= isl_set_dim(array
->extent
, isl_dim_set
);
307 dim2
= isl_map_dim(map
, isl_dim_out
);
308 map
= isl_map_add_dims(map
, isl_dim_out
, dim1
- dim2
);
309 map
= isl_map_set_tuple_id(map
, isl_dim_out
, id
);
310 map
= isl_map_intersect_range(map
, isl_set_copy(array
->extent
));
311 pet_array_free(array
);
312 kill
= pet_expr_kill_from_access_and_index(map
, index
);
313 scop_i
= scop_from_evaluated_expr(kill
, state
->n_stmt
++, loc
, pc
);
314 scop
= pet_scop_add_par(state
->ctx
, scop
, scop_i
);
319 return pet_scop_free(scop
);
322 /* For each argument of the __pencil_kill call in "tree" that
323 * represents an access, add a kill statement to "scop" killing the accessed
326 static struct pet_scop
*scop_from_pencil_kill(__isl_keep pet_tree
*tree
,
327 __isl_keep pet_context
*pc
, struct pet_state
*state
)
330 struct pet_scop
*scop
;
333 call
= tree
->u
.e
.expr
;
335 scop
= pet_scop_empty(pet_context_get_space(pc
));
337 n
= pet_expr_get_n_arg(call
);
338 for (i
= 0; i
< n
; ++i
) {
342 arg
= pet_expr_get_arg(call
, i
);
344 return pet_scop_free(scop
);
345 if (pet_expr_get_type(arg
) != pet_expr_access
) {
349 loc
= pet_tree_get_loc(tree
);
350 scop
= scop_add_kill(scop
, arg
, loc
, pc
, state
);
356 /* Construct a pet_scop for an expression statement within the context "pc".
358 * If the expression calls __pencil_kill, then it needs to be converted
359 * into zero or more kill statements.
360 * Otherwise, a scop is extracted directly from the tree.
362 static struct pet_scop
*scop_from_tree_expr(__isl_keep pet_tree
*tree
,
363 __isl_keep pet_context
*pc
, struct pet_state
*state
)
367 is_kill
= is_pencil_kill(tree
);
371 return scop_from_pencil_kill(tree
, pc
, state
);
372 return scop_from_unevaluated_tree(pet_tree_copy(tree
),
373 state
->n_stmt
++, pc
);
376 /* Return those elements in the space of "cond" that come after
377 * (based on "sign") an element in "cond" in the final dimension.
379 static __isl_give isl_set
*after(__isl_take isl_set
*cond
, int sign
)
382 isl_map
*previous_to_this
;
385 dim
= isl_set_dim(cond
, isl_dim_set
);
386 space
= isl_space_map_from_set(isl_set_get_space(cond
));
387 previous_to_this
= isl_map_universe(space
);
388 for (i
= 0; i
+ 1 < dim
; ++i
)
389 previous_to_this
= isl_map_equate(previous_to_this
,
390 isl_dim_in
, i
, isl_dim_out
, i
);
392 previous_to_this
= isl_map_order_lt(previous_to_this
,
393 isl_dim_in
, dim
- 1, isl_dim_out
, dim
- 1);
395 previous_to_this
= isl_map_order_gt(previous_to_this
,
396 isl_dim_in
, dim
- 1, isl_dim_out
, dim
- 1);
398 cond
= isl_set_apply(cond
, previous_to_this
);
403 /* Remove those iterations of "domain" that have an earlier iteration
404 * (based on "sign") in the final dimension where "skip" is satisfied.
405 * If "apply_skip_map" is set, then "skip_map" is first applied
406 * to the embedded skip condition before removing it from the domain.
408 static __isl_give isl_set
*apply_affine_break(__isl_take isl_set
*domain
,
409 __isl_take isl_set
*skip
, int sign
,
410 int apply_skip_map
, __isl_keep isl_map
*skip_map
)
413 skip
= isl_set_apply(skip
, isl_map_copy(skip_map
));
414 skip
= isl_set_intersect(skip
, isl_set_copy(domain
));
415 return isl_set_subtract(domain
, after(skip
, sign
));
418 /* Create a single-dimensional multi-affine expression on the domain space
419 * of "pc" that is equal to the final dimension of this domain.
420 * "loop_nr" is the sequence number of the corresponding loop.
421 * If "id" is not NULL, then it is used as the output tuple name.
422 * Otherwise, the name is constructed as L_<loop_nr>.
424 static __isl_give isl_multi_aff
*map_to_last(__isl_keep pet_context
*pc
,
425 int loop_nr
, __isl_keep isl_id
*id
)
435 space
= pet_context_get_space(pc
);
436 pos
= isl_space_dim(space
, isl_dim_set
) - 1;
437 ls
= isl_local_space_from_space(space
);
438 aff
= isl_aff_var_on_domain(ls
, isl_dim_set
, pos
);
439 ma
= isl_multi_aff_from_aff(aff
);
442 label
= isl_id_copy(id
);
444 snprintf(name
, sizeof(name
), "L_%d", loop_nr
);
445 label
= isl_id_alloc(pet_context_get_ctx(pc
), name
, NULL
);
447 ma
= isl_multi_aff_set_tuple_id(ma
, isl_dim_out
, label
);
452 /* Create an affine expression that maps elements
453 * of an array "id_test" to the previous element in the final dimension
454 * (according to "inc"), provided this element belongs to "domain".
455 * That is, create the affine expression
457 * { id[outer,x] -> id[outer,x - inc] : (outer,x - inc) in domain }
459 static __isl_give isl_multi_pw_aff
*map_to_previous(__isl_take isl_id
*id_test
,
460 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
467 isl_multi_pw_aff
*prev
;
469 pos
= isl_set_dim(domain
, isl_dim_set
) - 1;
470 space
= isl_set_get_space(domain
);
471 space
= isl_space_map_from_set(space
);
472 ma
= isl_multi_aff_identity(space
);
473 aff
= isl_multi_aff_get_aff(ma
, pos
);
474 aff
= isl_aff_add_constant_val(aff
, isl_val_neg(inc
));
475 ma
= isl_multi_aff_set_aff(ma
, pos
, aff
);
476 domain
= isl_set_preimage_multi_aff(domain
, isl_multi_aff_copy(ma
));
477 prev
= isl_multi_pw_aff_from_multi_aff(ma
);
478 pa
= isl_multi_pw_aff_get_pw_aff(prev
, pos
);
479 pa
= isl_pw_aff_intersect_domain(pa
, domain
);
480 prev
= isl_multi_pw_aff_set_pw_aff(prev
, pos
, pa
);
481 prev
= isl_multi_pw_aff_set_tuple_id(prev
, isl_dim_out
, id_test
);
486 /* Add an implication to "scop" expressing that if an element of
487 * virtual array "id_test" has value "satisfied" then all previous elements
488 * of this array (in the final dimension) also have that value.
489 * The set of previous elements is bounded by "domain".
490 * If "sign" is negative then the iterator
491 * is decreasing and we express that all subsequent array elements
492 * (but still defined previously) have the same value.
494 static struct pet_scop
*add_implication(struct pet_scop
*scop
,
495 __isl_take isl_id
*id_test
, __isl_take isl_set
*domain
, int sign
,
502 dim
= isl_set_dim(domain
, isl_dim_set
);
503 domain
= isl_set_set_tuple_id(domain
, id_test
);
504 space
= isl_space_map_from_set(isl_set_get_space(domain
));
505 map
= isl_map_universe(space
);
506 for (i
= 0; i
+ 1 < dim
; ++i
)
507 map
= isl_map_equate(map
, isl_dim_in
, i
, isl_dim_out
, i
);
509 map
= isl_map_order_ge(map
,
510 isl_dim_in
, dim
- 1, isl_dim_out
, dim
- 1);
512 map
= isl_map_order_le(map
,
513 isl_dim_in
, dim
- 1, isl_dim_out
, dim
- 1);
514 map
= isl_map_intersect_range(map
, domain
);
515 scop
= pet_scop_add_implication(scop
, map
, satisfied
);
520 /* Add a filter to "scop" that imposes that it is only executed
521 * when the variable identified by "id_test" has a zero value
522 * for all previous iterations of "domain".
524 * In particular, add a filter that imposes that the array
525 * has a zero value at the previous iteration of domain and
526 * add an implication that implies that it then has that
527 * value for all previous iterations.
529 static struct pet_scop
*scop_add_break(struct pet_scop
*scop
,
530 __isl_take isl_id
*id_test
, __isl_take isl_set
*domain
,
531 __isl_take isl_val
*inc
)
533 isl_multi_pw_aff
*prev
;
534 int sign
= isl_val_sgn(inc
);
536 prev
= map_to_previous(isl_id_copy(id_test
), isl_set_copy(domain
), inc
);
537 scop
= add_implication(scop
, id_test
, domain
, sign
, 0);
538 scop
= pet_scop_filter(scop
, prev
, 0);
543 static struct pet_scop
*scop_from_tree(__isl_keep pet_tree
*tree
,
544 __isl_keep pet_context
*pc
, struct pet_state
*state
);
546 /* Construct a pet_scop for an infinite loop around the given body
547 * within the context "pc".
548 * "loop_id" is the label on the loop or NULL if there is no such label.
550 * The domain of "pc" has already been extended with an infinite loop
554 * We extract a pet_scop for the body and then embed it in a loop with
557 * { [outer,t] -> [t] }
559 * If the body contains any break, then it is taken into
560 * account in apply_affine_break (if the skip condition is affine)
561 * or in scop_add_break (if the skip condition is not affine).
563 * Note that in case of an affine skip condition,
564 * since we are dealing with a loop without loop iterator,
565 * the skip condition cannot refer to the current loop iterator and
566 * so effectively, the effect on the iteration domain is of the form
568 * { [outer,0]; [outer,t] : t >= 1 and not skip }
570 static struct pet_scop
*scop_from_infinite_loop(__isl_keep pet_tree
*body
,
571 __isl_keep isl_id
*loop_id
, __isl_keep pet_context
*pc
,
572 struct pet_state
*state
)
578 isl_multi_aff
*sched
;
579 struct pet_scop
*scop
;
580 int has_affine_break
;
583 ctx
= pet_tree_get_ctx(body
);
584 domain
= pet_context_get_domain(pc
);
585 sched
= map_to_last(pc
, state
->n_loop
++, loop_id
);
587 scop
= scop_from_tree(body
, pc
, state
);
589 has_affine_break
= pet_scop_has_affine_skip(scop
, pet_skip_later
);
590 if (has_affine_break
)
591 skip
= pet_scop_get_affine_skip_domain(scop
, pet_skip_later
);
592 has_var_break
= pet_scop_has_var_skip(scop
, pet_skip_later
);
594 id_test
= pet_scop_get_skip_id(scop
, pet_skip_later
);
596 scop
= pet_scop_reset_skips(scop
);
597 scop
= pet_scop_embed(scop
, isl_set_copy(domain
), sched
);
598 if (has_affine_break
) {
599 domain
= apply_affine_break(domain
, skip
, 1, 0, NULL
);
600 scop
= pet_scop_intersect_domain_prefix(scop
,
601 isl_set_copy(domain
));
604 scop
= scop_add_break(scop
, id_test
, domain
, isl_val_one(ctx
));
606 isl_set_free(domain
);
611 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
616 * within the context "pc".
618 * Extend the domain of "pc" with an extra inner loop
622 * and construct the scop in scop_from_infinite_loop.
624 static struct pet_scop
*scop_from_infinite_for(__isl_keep pet_tree
*tree
,
625 __isl_keep pet_context
*pc
, struct pet_state
*state
)
627 struct pet_scop
*scop
;
629 pc
= pet_context_copy(pc
);
630 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.l
.body
);
632 pc
= pet_context_add_infinite_loop(pc
);
634 scop
= scop_from_infinite_loop(tree
->u
.l
.body
, tree
->label
, pc
, state
);
636 pet_context_free(pc
);
641 /* Construct a pet_scop for a while loop of the form
646 * within the context "pc".
648 * The domain of "pc" has already been extended with an infinite loop
652 * Here, we add the constraints on the outer loop iterators
653 * implied by "pa" and construct the scop in scop_from_infinite_loop.
654 * Note that the intersection with these constraints
655 * may result in an empty loop.
657 static struct pet_scop
*scop_from_affine_while(__isl_keep pet_tree
*tree
,
658 __isl_take isl_pw_aff
*pa
, __isl_take pet_context
*pc
,
659 struct pet_state
*state
)
661 struct pet_scop
*scop
;
662 isl_set
*dom
, *local
;
665 valid
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
666 dom
= isl_pw_aff_non_zero_set(pa
);
667 local
= isl_set_add_dims(isl_set_copy(dom
), isl_dim_set
, 1);
668 pc
= pet_context_intersect_domain(pc
, local
);
669 scop
= scop_from_infinite_loop(tree
->u
.l
.body
, tree
->label
, pc
, state
);
670 scop
= pet_scop_restrict(scop
, dom
);
671 scop
= pet_scop_restrict_context(scop
, valid
);
673 pet_context_free(pc
);
677 /* Construct a scop for a while, given the scops for the condition
678 * and the body, the filter identifier and the iteration domain of
681 * In particular, the scop for the condition is filtered to depend
682 * on "id_test" evaluating to true for all previous iterations
683 * of the loop, while the scop for the body is filtered to depend
684 * on "id_test" evaluating to true for all iterations up to the
686 * The actual filter only imposes that this virtual array has
687 * value one on the previous or the current iteration.
688 * The fact that this condition also applies to the previous
689 * iterations is enforced by an implication.
691 * These filtered scops are then combined into a single scop,
692 * with the condition scop scheduled before the body scop.
694 * "sign" is positive if the iterator increases and negative
697 static struct pet_scop
*scop_add_while(struct pet_scop
*scop_cond
,
698 struct pet_scop
*scop_body
, __isl_take isl_id
*id_test
,
699 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
701 isl_ctx
*ctx
= isl_set_get_ctx(domain
);
703 isl_multi_pw_aff
*test_index
;
704 isl_multi_pw_aff
*prev
;
705 int sign
= isl_val_sgn(inc
);
706 struct pet_scop
*scop
;
708 prev
= map_to_previous(isl_id_copy(id_test
), isl_set_copy(domain
), inc
);
709 scop_cond
= pet_scop_filter(scop_cond
, prev
, 1);
711 space
= isl_space_map_from_set(isl_set_get_space(domain
));
712 test_index
= isl_multi_pw_aff_identity(space
);
713 test_index
= isl_multi_pw_aff_set_tuple_id(test_index
, isl_dim_out
,
714 isl_id_copy(id_test
));
715 scop_body
= pet_scop_filter(scop_body
, test_index
, 1);
717 scop
= pet_scop_add_seq(ctx
, scop_cond
, scop_body
);
718 scop
= add_implication(scop
, id_test
, domain
, sign
, 1);
723 /* Create a pet_scop with a single statement with name S_<stmt_nr>,
724 * evaluating "cond" and writing the result to a virtual scalar,
725 * as expressed by "index".
726 * The expression "cond" has not yet been evaluated in the context of "pc".
727 * Do so within the context "pc".
728 * The location of the statement is set to "loc".
730 static struct pet_scop
*scop_from_non_affine_condition(
731 __isl_take pet_expr
*cond
, int stmt_nr
,
732 __isl_take isl_multi_pw_aff
*index
,
733 __isl_take pet_loc
*loc
, __isl_keep pet_context
*pc
)
735 pet_expr
*expr
, *write
;
737 cond
= pet_context_evaluate_expr(pc
, cond
);
739 write
= pet_expr_from_index(index
);
740 write
= pet_expr_access_set_write(write
, 1);
741 write
= pet_expr_access_set_read(write
, 0);
742 expr
= pet_expr_new_binary(1, pet_op_assign
, write
, cond
);
744 return scop_from_evaluated_expr(expr
, stmt_nr
, loc
, pc
);
747 /* Given that "scop" has an affine skip condition of type pet_skip_now,
748 * apply this skip condition to the domain of "pc".
749 * That is, remove the elements satisfying the skip condition from
750 * the domain of "pc".
752 static __isl_give pet_context
*apply_affine_continue(__isl_take pet_context
*pc
,
753 struct pet_scop
*scop
)
755 isl_set
*domain
, *skip
;
757 skip
= pet_scop_get_affine_skip_domain(scop
, pet_skip_now
);
758 domain
= pet_context_get_domain(pc
);
759 domain
= isl_set_subtract(domain
, skip
);
760 pc
= pet_context_intersect_domain(pc
, domain
);
765 /* Add a scop for evaluating the loop increment "inc" add the end
766 * of a loop body "scop" within the context "pc".
768 * The skip conditions resulting from continue statements inside
769 * the body do not apply to "inc", but those resulting from break
770 * statements do need to get applied.
772 static struct pet_scop
*scop_add_inc(struct pet_scop
*scop
,
773 __isl_take pet_expr
*inc
, __isl_take pet_loc
*loc
,
774 __isl_keep pet_context
*pc
, struct pet_state
*state
)
776 struct pet_scop
*scop_inc
;
778 pc
= pet_context_copy(pc
);
780 if (pet_scop_has_skip(scop
, pet_skip_later
)) {
781 isl_multi_pw_aff
*skip
;
782 skip
= pet_scop_get_skip(scop
, pet_skip_later
);
783 scop
= pet_scop_set_skip(scop
, pet_skip_now
, skip
);
784 if (pet_scop_has_affine_skip(scop
, pet_skip_now
))
785 pc
= apply_affine_continue(pc
, scop
);
787 pet_scop_reset_skip(scop
, pet_skip_now
);
788 scop_inc
= scop_from_expr(inc
, state
->n_stmt
++, loc
, pc
);
789 scop
= pet_scop_add_seq(state
->ctx
, scop
, scop_inc
);
791 pet_context_free(pc
);
796 /* Construct a generic while scop, with iteration domain
797 * { [t] : t >= 0 } around the scop for "tree_body" within the context "pc".
798 * "loop_id" is the label on the loop or NULL if there is no such label.
799 * The domain of "pc" has already been extended with this infinite loop
803 * The scop consists of two parts,
804 * one for evaluating the condition "cond" and one for the body.
805 * If "expr_inc" is not NULL, then a scop for evaluating this expression
806 * is added at the end of the body,
807 * after replacing any skip conditions resulting from continue statements
808 * by the skip conditions resulting from break statements (if any).
810 * The schedules are combined as a sequence to reflect that the condition is
811 * evaluated before the body is executed and the body is filtered to depend
812 * on the result of the condition evaluating to true on all iterations
813 * up to the current iteration, while the evaluation of the condition itself
814 * is filtered to depend on the result of the condition evaluating to true
815 * on all previous iterations.
816 * The context of the scop representing the body is dropped
817 * because we don't know how many times the body will be executed,
820 * If the body contains any break, then it is taken into
821 * account in apply_affine_break (if the skip condition is affine)
822 * or in scop_add_break (if the skip condition is not affine).
824 * Note that in case of an affine skip condition,
825 * since we are dealing with a loop without loop iterator,
826 * the skip condition cannot refer to the current loop iterator and
827 * so effectively, the effect on the iteration domain is of the form
829 * { [outer,0]; [outer,t] : t >= 1 and not skip }
831 static struct pet_scop
*scop_from_non_affine_while(__isl_take pet_expr
*cond
,
832 __isl_take pet_loc
*loc
, __isl_keep pet_tree
*tree_body
,
833 __isl_keep isl_id
*loop_id
, __isl_take pet_expr
*expr_inc
,
834 __isl_take pet_context
*pc
, struct pet_state
*state
)
837 isl_id
*id_test
, *id_break_test
;
839 isl_multi_pw_aff
*test_index
;
842 isl_multi_aff
*sched
;
843 struct pet_scop
*scop
, *scop_body
;
844 int has_affine_break
;
848 space
= pet_context_get_space(pc
);
849 test_index
= pet_create_test_index(space
, state
->n_test
++);
850 scop
= scop_from_non_affine_condition(cond
, state
->n_stmt
++,
851 isl_multi_pw_aff_copy(test_index
),
852 pet_loc_copy(loc
), pc
);
853 id_test
= isl_multi_pw_aff_get_tuple_id(test_index
, isl_dim_out
);
854 domain
= pet_context_get_domain(pc
);
855 scop
= pet_scop_add_boolean_array(scop
, isl_set_copy(domain
),
856 test_index
, state
->int_size
);
858 sched
= map_to_last(pc
, state
->n_loop
++, loop_id
);
860 scop_body
= scop_from_tree(tree_body
, pc
, state
);
862 has_affine_break
= pet_scop_has_affine_skip(scop_body
, pet_skip_later
);
863 if (has_affine_break
)
864 skip
= pet_scop_get_affine_skip_domain(scop_body
,
866 has_var_break
= pet_scop_has_var_skip(scop_body
, pet_skip_later
);
868 id_break_test
= pet_scop_get_skip_id(scop_body
, pet_skip_later
);
870 scop_body
= pet_scop_reset_context(scop_body
);
872 scop_body
= scop_add_inc(scop_body
, expr_inc
, loc
, pc
, state
);
875 scop_body
= pet_scop_reset_skips(scop_body
);
877 if (has_affine_break
) {
878 domain
= apply_affine_break(domain
, skip
, 1, 0, NULL
);
879 scop
= pet_scop_intersect_domain_prefix(scop
,
880 isl_set_copy(domain
));
881 scop_body
= pet_scop_intersect_domain_prefix(scop_body
,
882 isl_set_copy(domain
));
885 scop
= scop_add_break(scop
, isl_id_copy(id_break_test
),
886 isl_set_copy(domain
), isl_val_one(ctx
));
887 scop_body
= scop_add_break(scop_body
, id_break_test
,
888 isl_set_copy(domain
), isl_val_one(ctx
));
890 scop
= scop_add_while(scop
, scop_body
, id_test
, isl_set_copy(domain
),
893 scop
= pet_scop_embed(scop
, domain
, sched
);
895 pet_context_free(pc
);
899 /* Check if the while loop is of the form
901 * while (affine expression)
904 * If so, call scop_from_affine_while to construct a scop.
906 * Otherwise, pass control to scop_from_non_affine_while.
908 * "pc" is the context in which the affine expressions in the scop are created.
909 * The domain of "pc" is extended with an infinite loop
913 * before passing control to scop_from_affine_while or
914 * scop_from_non_affine_while.
916 static struct pet_scop
*scop_from_while(__isl_keep pet_tree
*tree
,
917 __isl_keep pet_context
*pc
, struct pet_state
*state
)
925 pc
= pet_context_copy(pc
);
926 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.l
.body
);
928 cond_expr
= pet_expr_copy(tree
->u
.l
.cond
);
929 cond_expr
= pet_context_evaluate_expr(pc
, cond_expr
);
930 pa
= pet_expr_extract_affine_condition(cond_expr
, pc
);
931 pet_expr_free(cond_expr
);
933 pc
= pet_context_add_infinite_loop(pc
);
938 if (!isl_pw_aff_involves_nan(pa
))
939 return scop_from_affine_while(tree
, pa
, pc
, state
);
941 return scop_from_non_affine_while(pet_expr_copy(tree
->u
.l
.cond
),
942 pet_tree_get_loc(tree
), tree
->u
.l
.body
,
943 tree
->label
, NULL
, pc
, state
);
945 pet_context_free(pc
);
949 /* Check whether "cond" expresses a simple loop bound
950 * on the final set dimension.
951 * In particular, if "up" is set then "cond" should contain only
952 * upper bounds on the final set dimension.
953 * Otherwise, it should contain only lower bounds.
955 static int is_simple_bound(__isl_keep isl_set
*cond
, __isl_keep isl_val
*inc
)
959 pos
= isl_set_dim(cond
, isl_dim_set
) - 1;
960 if (isl_val_is_pos(inc
))
961 return !isl_set_dim_has_any_lower_bound(cond
, isl_dim_set
, pos
);
963 return !isl_set_dim_has_any_upper_bound(cond
, isl_dim_set
, pos
);
966 /* Extend a condition on a given iteration of a loop to one that
967 * imposes the same condition on all previous iterations.
968 * "domain" expresses the lower [upper] bound on the iterations
969 * when inc is positive [negative] in its final dimension.
971 * In particular, we construct the condition (when inc is positive)
973 * forall i' : (domain(i') and i' <= i) => cond(i')
975 * (where "<=" applies to the final dimension)
976 * which is equivalent to
978 * not exists i' : domain(i') and i' <= i and not cond(i')
980 * We construct this set by subtracting the satisfying cond from domain,
983 * { [i'] -> [i] : i' <= i }
985 * and then subtracting the result from domain again.
987 static __isl_give isl_set
*valid_for_each_iteration(__isl_take isl_set
*cond
,
988 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
991 isl_map
*previous_to_this
;
994 dim
= isl_set_dim(cond
, isl_dim_set
);
995 space
= isl_space_map_from_set(isl_set_get_space(cond
));
996 previous_to_this
= isl_map_universe(space
);
997 for (i
= 0; i
+ 1 < dim
; ++i
)
998 previous_to_this
= isl_map_equate(previous_to_this
,
999 isl_dim_in
, i
, isl_dim_out
, i
);
1000 if (isl_val_is_pos(inc
))
1001 previous_to_this
= isl_map_order_le(previous_to_this
,
1002 isl_dim_in
, dim
- 1, isl_dim_out
, dim
- 1);
1004 previous_to_this
= isl_map_order_ge(previous_to_this
,
1005 isl_dim_in
, dim
- 1, isl_dim_out
, dim
- 1);
1007 cond
= isl_set_subtract(isl_set_copy(domain
), cond
);
1008 cond
= isl_set_apply(cond
, previous_to_this
);
1009 cond
= isl_set_subtract(domain
, cond
);
1016 /* Given an initial value of the form
1018 * { [outer,i] -> init(outer) }
1020 * construct a domain of the form
1022 * { [outer,i] : exists a: i = init(outer) + a * inc and a >= 0 }
1024 static __isl_give isl_set
*strided_domain(__isl_take isl_pw_aff
*init
,
1025 __isl_take isl_val
*inc
)
1030 isl_local_space
*ls
;
1033 dim
= isl_pw_aff_dim(init
, isl_dim_in
);
1035 init
= isl_pw_aff_add_dims(init
, isl_dim_in
, 1);
1036 space
= isl_pw_aff_get_domain_space(init
);
1037 ls
= isl_local_space_from_space(space
);
1038 aff
= isl_aff_zero_on_domain(isl_local_space_copy(ls
));
1039 aff
= isl_aff_add_coefficient_val(aff
, isl_dim_in
, dim
, inc
);
1040 init
= isl_pw_aff_add(init
, isl_pw_aff_from_aff(aff
));
1042 aff
= isl_aff_var_on_domain(ls
, isl_dim_set
, dim
- 1);
1043 set
= isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff
), init
);
1045 set
= isl_set_lower_bound_si(set
, isl_dim_set
, dim
, 0);
1046 set
= isl_set_project_out(set
, isl_dim_set
, dim
, 1);
1051 /* Assuming "cond" represents a bound on a loop where the loop
1052 * iterator "iv" is incremented (or decremented) by one, check if wrapping
1055 * Under the given assumptions, wrapping is only possible if "cond" allows
1056 * for the last value before wrapping, i.e., 2^width - 1 in case of an
1057 * increasing iterator and 0 in case of a decreasing iterator.
1059 static int can_wrap(__isl_keep isl_set
*cond
, __isl_keep pet_expr
*iv
,
1060 __isl_keep isl_val
*inc
)
1067 test
= isl_set_copy(cond
);
1069 ctx
= isl_set_get_ctx(test
);
1070 if (isl_val_is_neg(inc
))
1071 limit
= isl_val_zero(ctx
);
1073 limit
= isl_val_int_from_ui(ctx
, pet_expr_get_type_size(iv
));
1074 limit
= isl_val_2exp(limit
);
1075 limit
= isl_val_sub_ui(limit
, 1);
1078 test
= isl_set_fix_val(cond
, isl_dim_set
, 0, limit
);
1079 cw
= !isl_set_is_empty(test
);
1089 * construct the following affine expression on this space
1091 * { [outer, v] -> [outer, v mod 2^width] }
1093 * where width is the number of bits used to represent the values
1094 * of the unsigned variable "iv".
1096 static __isl_give isl_multi_aff
*compute_wrapping(__isl_take isl_space
*space
,
1097 __isl_keep pet_expr
*iv
)
1105 dim
= isl_space_dim(space
, isl_dim_set
);
1107 ctx
= isl_space_get_ctx(space
);
1108 mod
= isl_val_int_from_ui(ctx
, pet_expr_get_type_size(iv
));
1109 mod
= isl_val_2exp(mod
);
1111 space
= isl_space_map_from_set(space
);
1112 ma
= isl_multi_aff_identity(space
);
1114 aff
= isl_multi_aff_get_aff(ma
, dim
- 1);
1115 aff
= isl_aff_mod_val(aff
, mod
);
1116 ma
= isl_multi_aff_set_aff(ma
, dim
- 1, aff
);
1121 /* Given two sets in the space
1125 * where l represents the outer loop iterators, compute the set
1126 * of values of l that ensure that "set1" is a subset of "set2".
1128 * set1 is a subset of set2 if
1130 * forall i: set1(l,i) => set2(l,i)
1134 * not exists i: set1(l,i) and not set2(l,i)
1138 * not exists i: (set1 \ set2)(l,i)
1140 static __isl_give isl_set
*enforce_subset(__isl_take isl_set
*set1
,
1141 __isl_take isl_set
*set2
)
1145 pos
= isl_set_dim(set1
, isl_dim_set
) - 1;
1146 set1
= isl_set_subtract(set1
, set2
);
1147 set1
= isl_set_eliminate(set1
, isl_dim_set
, pos
, 1);
1148 return isl_set_complement(set1
);
1151 /* Compute the set of outer iterator values for which "cond" holds
1152 * on the next iteration of the inner loop for each element of "dom".
1154 * We first construct mapping { [l,i] -> [l,i + inc] } (where l refers
1155 * to the outer loop iterators), plug that into "cond"
1156 * and then compute the set of outer iterators for which "dom" is a subset
1159 static __isl_give isl_set
*valid_on_next(__isl_take isl_set
*cond
,
1160 __isl_take isl_set
*dom
, __isl_take isl_val
*inc
)
1167 pos
= isl_set_dim(dom
, isl_dim_set
) - 1;
1168 space
= isl_set_get_space(dom
);
1169 space
= isl_space_map_from_set(space
);
1170 ma
= isl_multi_aff_identity(space
);
1171 aff
= isl_multi_aff_get_aff(ma
, pos
);
1172 aff
= isl_aff_add_constant_val(aff
, inc
);
1173 ma
= isl_multi_aff_set_aff(ma
, pos
, aff
);
1174 cond
= isl_set_preimage_multi_aff(cond
, ma
);
1176 return enforce_subset(dom
, cond
);
1179 /* Extract the for loop "tree" as a while loop within the context "pc_init".
1180 * In particular, "pc_init" represents the context of the loop,
1181 * whereas "pc" represents the context of the body of the loop and
1182 * has already had its domain extended with an infinite loop
1186 * The for loop has the form
1188 * for (iv = init; cond; iv += inc)
1199 * except that the skips resulting from any continue statements
1200 * in body do not apply to the increment, but are replaced by the skips
1201 * resulting from break statements.
1203 * If the loop iterator is declared in the for loop, then it is killed before
1204 * and after the loop.
1206 static struct pet_scop
*scop_from_non_affine_for(__isl_keep pet_tree
*tree
,
1207 __isl_keep pet_context
*init_pc
, __isl_take pet_context
*pc
,
1208 struct pet_state
*state
)
1212 pet_expr
*expr_iv
, *init
, *inc
;
1213 struct pet_scop
*scop_init
, *scop
;
1215 struct pet_array
*array
;
1216 struct pet_scop
*scop_kill
;
1218 iv
= pet_expr_access_get_id(tree
->u
.l
.iv
);
1219 pc
= pet_context_clear_value(pc
, iv
);
1221 declared
= tree
->u
.l
.declared
;
1223 expr_iv
= pet_expr_copy(tree
->u
.l
.iv
);
1224 type_size
= pet_expr_get_type_size(expr_iv
);
1225 init
= pet_expr_copy(tree
->u
.l
.init
);
1226 init
= pet_expr_new_binary(type_size
, pet_op_assign
, expr_iv
, init
);
1227 scop_init
= scop_from_expr(init
, state
->n_stmt
++,
1228 pet_tree_get_loc(tree
), init_pc
);
1230 expr_iv
= pet_expr_copy(tree
->u
.l
.iv
);
1231 type_size
= pet_expr_get_type_size(expr_iv
);
1232 inc
= pet_expr_copy(tree
->u
.l
.inc
);
1233 inc
= pet_expr_new_binary(type_size
, pet_op_add_assign
, expr_iv
, inc
);
1235 scop
= scop_from_non_affine_while(pet_expr_copy(tree
->u
.l
.cond
),
1236 pet_tree_get_loc(tree
), tree
->u
.l
.body
, tree
->label
,
1237 inc
, pet_context_copy(pc
), state
);
1239 scop
= pet_scop_add_seq(state
->ctx
, scop_init
, scop
);
1241 pet_context_free(pc
);
1246 array
= extract_array(tree
->u
.l
.iv
, init_pc
, state
);
1248 array
->declared
= 1;
1249 scop_kill
= kill(pet_tree_get_loc(tree
), array
, init_pc
, state
);
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
= pet_scop_add_seq(state
->ctx
, scop
, scop_kill
);
1258 /* Given an access expression "expr", is the variable accessed by
1259 * "expr" assigned anywhere inside "tree"?
1261 static int is_assigned(__isl_keep pet_expr
*expr
, __isl_keep pet_tree
*tree
)
1266 id
= pet_expr_access_get_id(expr
);
1267 assigned
= pet_tree_writes(tree
, id
);
1273 /* Are all nested access parameters in "pa" allowed given "tree".
1274 * In particular, is none of them written by anywhere inside "tree".
1276 * If "tree" has any continue or break nodes in the current loop level,
1277 * then no nested access parameters are allowed.
1278 * In particular, if there is any nested access in a guard
1279 * for a piece of code containing a "continue", then we want to introduce
1280 * a separate statement for evaluating this guard so that we can express
1281 * that the result is false for all previous iterations.
1283 static int is_nested_allowed(__isl_keep isl_pw_aff
*pa
,
1284 __isl_keep pet_tree
*tree
)
1291 if (!pet_nested_any_in_pw_aff(pa
))
1294 if (pet_tree_has_continue_or_break(tree
))
1297 nparam
= isl_pw_aff_dim(pa
, isl_dim_param
);
1298 for (i
= 0; i
< nparam
; ++i
) {
1299 isl_id
*id
= isl_pw_aff_get_dim_id(pa
, isl_dim_param
, i
);
1303 if (!pet_nested_in_id(id
)) {
1308 expr
= pet_nested_extract_expr(id
);
1309 allowed
= pet_expr_get_type(expr
) == pet_expr_access
&&
1310 !is_assigned(expr
, tree
);
1312 pet_expr_free(expr
);
1322 /* Internal data structure for collect_local.
1323 * "pc" and "state" are needed to extract pet_arrays for the local variables.
1324 * "local" collects the results.
1326 struct pet_tree_collect_local_data
{
1328 struct pet_state
*state
;
1329 isl_union_set
*local
;
1332 /* Add the variable accessed by "var" to data->local.
1333 * We extract a representation of the variable from
1334 * the pet_array constructed using extract_array
1335 * to ensure consistency with the rest of the scop.
1337 static int add_local(struct pet_tree_collect_local_data
*data
,
1338 __isl_keep pet_expr
*var
)
1340 struct pet_array
*array
;
1343 array
= extract_array(var
, data
->pc
, data
->state
);
1347 universe
= isl_set_universe(isl_set_get_space(array
->extent
));
1348 data
->local
= isl_union_set_add_set(data
->local
, universe
);
1349 pet_array_free(array
);
1354 /* If the node "tree" declares a variable, then add it to
1357 static int extract_local_var(__isl_keep pet_tree
*tree
, void *user
)
1359 enum pet_tree_type type
;
1360 struct pet_tree_collect_local_data
*data
= user
;
1362 type
= pet_tree_get_type(tree
);
1363 if (type
== pet_tree_decl
|| type
== pet_tree_decl_init
)
1364 return add_local(data
, tree
->u
.d
.var
);
1369 /* If the node "tree" is a for loop that declares its induction variable,
1370 * then add it this induction variable to data->local.
1372 static int extract_local_iterator(__isl_keep pet_tree
*tree
, void *user
)
1374 struct pet_tree_collect_local_data
*data
= user
;
1376 if (pet_tree_get_type(tree
) == pet_tree_for
&& tree
->u
.l
.declared
)
1377 return add_local(data
, tree
->u
.l
.iv
);
1382 /* Collect and return all local variables of the for loop represented
1383 * by "tree", with "scop" the corresponding pet_scop.
1384 * "pc" and "state" are needed to extract pet_arrays for the local variables.
1386 * We collect not only the variables that are declared inside "tree",
1387 * but also the loop iterators that are declared anywhere inside
1388 * any possible macro statements in "scop".
1389 * The latter also appear as declared variable in the scop,
1390 * whereas other declared loop iterators only appear implicitly
1391 * in the iteration domains.
1393 static __isl_give isl_union_set
*collect_local(struct pet_scop
*scop
,
1394 __isl_keep pet_tree
*tree
, __isl_keep pet_context
*pc
,
1395 struct pet_state
*state
)
1399 struct pet_tree_collect_local_data data
= { pc
, state
};
1401 ctx
= pet_tree_get_ctx(tree
);
1402 data
.local
= isl_union_set_empty(isl_space_params_alloc(ctx
, 0));
1404 if (pet_tree_foreach_sub_tree(tree
, &extract_local_var
, &data
) < 0)
1405 return isl_union_set_free(data
.local
);
1407 for (i
= 0; i
< scop
->n_stmt
; ++i
) {
1408 pet_tree
*body
= scop
->stmts
[i
]->body
;
1409 if (pet_tree_foreach_sub_tree(body
, &extract_local_iterator
,
1411 return isl_union_set_free(data
.local
);
1417 /* Add an independence to "scop" if the for node "tree" was marked
1419 * "domain" is the set of loop iterators, with the current for loop
1420 * innermost. If "sign" is positive, then the inner iterator increases.
1421 * Otherwise it decreases.
1422 * "pc" and "state" are needed to extract pet_arrays for the local variables.
1424 * If the tree was marked, then collect all local variables and
1425 * add an independence.
1427 static struct pet_scop
*set_independence(struct pet_scop
*scop
,
1428 __isl_keep pet_tree
*tree
, __isl_keep isl_set
*domain
, int sign
,
1429 __isl_keep pet_context
*pc
, struct pet_state
*state
)
1431 isl_union_set
*local
;
1433 if (!tree
->u
.l
.independent
)
1436 local
= collect_local(scop
, tree
, pc
, state
);
1437 scop
= pet_scop_set_independent(scop
, domain
, local
, sign
);
1442 /* Construct a pet_scop for a for tree with static affine initialization
1443 * and constant increment within the context "pc".
1444 * The domain of "pc" has already been extended with an (at this point
1445 * unbounded) inner loop iterator corresponding to the current for loop.
1447 * The condition is allowed to contain nested accesses, provided
1448 * they are not being written to inside the body of the loop.
1449 * Otherwise, or if the condition is otherwise non-affine, the for loop is
1450 * essentially treated as a while loop, with iteration domain
1451 * { [l,i] : i >= init }, where l refers to the outer loop iterators.
1453 * We extract a pet_scop for the body after intersecting the domain of "pc"
1455 * { [l,i] : i >= init and condition' }
1459 * { [l,i] : i <= init and condition' }
1461 * Where condition' is equal to condition if the latter is
1462 * a simple upper [lower] bound and a condition that is extended
1463 * to apply to all previous iterations otherwise.
1464 * Afterwards, the schedule of the pet_scop is extended with
1472 * If the condition is non-affine, then we drop the condition from the
1473 * iteration domain and instead create a separate statement
1474 * for evaluating the condition. The body is then filtered to depend
1475 * on the result of the condition evaluating to true on all iterations
1476 * up to the current iteration, while the evaluation the condition itself
1477 * is filtered to depend on the result of the condition evaluating to true
1478 * on all previous iterations.
1479 * The context of the scop representing the body is dropped
1480 * because we don't know how many times the body will be executed,
1483 * If the stride of the loop is not 1, then "i >= init" is replaced by
1485 * (exists a: i = init + stride * a and a >= 0)
1487 * If the loop iterator i is unsigned, then wrapping may occur.
1488 * We therefore use a virtual iterator instead that does not wrap.
1489 * However, the condition in the code applies
1490 * to the wrapped value, so we need to change condition(l,i)
1491 * into condition([l,i % 2^width]). Similarly, we replace all accesses
1492 * to the original iterator by the wrapping of the virtual iterator.
1493 * Note that there may be no need to perform this final wrapping
1494 * if the loop condition (after wrapping) satisfies certain conditions.
1495 * However, the is_simple_bound condition is not enough since it doesn't
1496 * check if there even is an upper bound.
1498 * Wrapping on unsigned iterators can be avoided entirely if
1499 * loop condition is simple, the loop iterator is incremented
1500 * [decremented] by one and the last value before wrapping cannot
1501 * possibly satisfy the loop condition.
1503 * Valid outer iterators for a for loop are those for which the initial
1504 * value itself, the increment on each domain iteration and
1505 * the condition on both the initial value and
1506 * the result of incrementing the iterator for each iteration of the domain
1508 * If the loop condition is non-affine, then we only consider validity
1509 * of the initial value.
1511 * If the body contains any break, then we keep track of it in "skip"
1512 * (if the skip condition is affine) or it is handled in scop_add_break
1513 * (if the skip condition is not affine).
1514 * Note that the affine break condition needs to be considered with
1515 * respect to previous iterations in the virtual domain (if any).
1517 static struct pet_scop
*scop_from_affine_for(__isl_keep pet_tree
*tree
,
1518 __isl_take isl_pw_aff
*init_val
, __isl_take isl_pw_aff
*pa_inc
,
1519 __isl_take isl_val
*inc
, __isl_take pet_context
*pc
,
1520 struct pet_state
*state
)
1523 isl_multi_aff
*sched
;
1524 isl_set
*cond
= NULL
;
1525 isl_set
*skip
= NULL
;
1526 isl_id
*id_test
= NULL
, *id_break_test
;
1527 struct pet_scop
*scop
, *scop_cond
= NULL
;
1534 int has_affine_break
;
1536 isl_map
*rev_wrap
= NULL
;
1537 isl_map
*init_val_map
;
1539 isl_set
*valid_init
;
1540 isl_set
*valid_cond
;
1541 isl_set
*valid_cond_init
;
1542 isl_set
*valid_cond_next
;
1544 pet_expr
*cond_expr
;
1545 pet_context
*pc_nested
;
1547 pos
= pet_context_dim(pc
) - 1;
1549 domain
= pet_context_get_domain(pc
);
1550 cond_expr
= pet_expr_copy(tree
->u
.l
.cond
);
1551 cond_expr
= pet_context_evaluate_expr(pc
, cond_expr
);
1552 pc_nested
= pet_context_copy(pc
);
1553 pc_nested
= pet_context_set_allow_nested(pc_nested
, 1);
1554 pa
= pet_expr_extract_affine_condition(cond_expr
, pc_nested
);
1555 pet_context_free(pc_nested
);
1556 pet_expr_free(cond_expr
);
1558 valid_inc
= isl_pw_aff_domain(pa_inc
);
1560 is_unsigned
= pet_expr_get_type_size(tree
->u
.l
.iv
) > 0;
1562 is_non_affine
= isl_pw_aff_involves_nan(pa
) ||
1563 !is_nested_allowed(pa
, tree
->u
.l
.body
);
1565 pa
= isl_pw_aff_free(pa
);
1567 valid_cond
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
1568 cond
= isl_pw_aff_non_zero_set(pa
);
1570 cond
= isl_set_universe(isl_set_get_space(domain
));
1572 valid_cond
= isl_set_coalesce(valid_cond
);
1573 is_one
= isl_val_is_one(inc
) || isl_val_is_negone(inc
);
1574 is_virtual
= is_unsigned
&&
1575 (!is_one
|| can_wrap(cond
, tree
->u
.l
.iv
, inc
));
1577 init_val_map
= isl_map_from_pw_aff(isl_pw_aff_copy(init_val
));
1578 init_val_map
= isl_map_equate(init_val_map
, isl_dim_in
, pos
,
1580 valid_cond_init
= enforce_subset(isl_map_domain(init_val_map
),
1581 isl_set_copy(valid_cond
));
1582 if (is_one
&& !is_virtual
) {
1585 isl_pw_aff_free(init_val
);
1586 pa
= pet_expr_extract_comparison(
1587 isl_val_is_pos(inc
) ? pet_op_ge
: pet_op_le
,
1588 tree
->u
.l
.iv
, tree
->u
.l
.init
, pc
);
1589 valid_init
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
1590 valid_init
= isl_set_eliminate(valid_init
, isl_dim_set
,
1591 isl_set_dim(domain
, isl_dim_set
) - 1, 1);
1592 cond
= isl_pw_aff_non_zero_set(pa
);
1593 domain
= isl_set_intersect(domain
, cond
);
1597 valid_init
= isl_pw_aff_domain(isl_pw_aff_copy(init_val
));
1598 strided
= strided_domain(init_val
, isl_val_copy(inc
));
1599 domain
= isl_set_intersect(domain
, strided
);
1603 isl_multi_aff
*wrap
;
1604 wrap
= compute_wrapping(isl_set_get_space(cond
), tree
->u
.l
.iv
);
1605 pc
= pet_context_preimage_domain(pc
, wrap
);
1606 rev_wrap
= isl_map_from_multi_aff(wrap
);
1607 rev_wrap
= isl_map_reverse(rev_wrap
);
1608 cond
= isl_set_apply(cond
, isl_map_copy(rev_wrap
));
1609 valid_cond
= isl_set_apply(valid_cond
, isl_map_copy(rev_wrap
));
1610 valid_inc
= isl_set_apply(valid_inc
, isl_map_copy(rev_wrap
));
1612 is_simple
= is_simple_bound(cond
, inc
);
1614 cond
= isl_set_gist(cond
, isl_set_copy(domain
));
1615 is_simple
= is_simple_bound(cond
, inc
);
1618 cond
= valid_for_each_iteration(cond
,
1619 isl_set_copy(domain
), isl_val_copy(inc
));
1620 cond
= isl_set_align_params(cond
, isl_set_get_space(domain
));
1621 domain
= isl_set_intersect(domain
, cond
);
1622 sched
= map_to_last(pc
, state
->n_loop
++, tree
->label
);
1623 if (isl_val_is_neg(inc
))
1624 sched
= isl_multi_aff_neg(sched
);
1626 valid_cond_next
= valid_on_next(valid_cond
, isl_set_copy(domain
),
1628 valid_inc
= enforce_subset(isl_set_copy(domain
), valid_inc
);
1630 pc
= pet_context_intersect_domain(pc
, isl_set_copy(domain
));
1632 if (is_non_affine
) {
1634 isl_multi_pw_aff
*test_index
;
1635 space
= isl_set_get_space(domain
);
1636 test_index
= pet_create_test_index(space
, state
->n_test
++);
1637 scop_cond
= scop_from_non_affine_condition(
1638 pet_expr_copy(tree
->u
.l
.cond
), state
->n_stmt
++,
1639 isl_multi_pw_aff_copy(test_index
),
1640 pet_tree_get_loc(tree
), pc
);
1641 id_test
= isl_multi_pw_aff_get_tuple_id(test_index
,
1643 scop_cond
= pet_scop_add_boolean_array(scop_cond
,
1644 isl_set_copy(domain
), test_index
,
1648 scop
= scop_from_tree(tree
->u
.l
.body
, pc
, state
);
1649 has_affine_break
= scop
&&
1650 pet_scop_has_affine_skip(scop
, pet_skip_later
);
1651 if (has_affine_break
)
1652 skip
= pet_scop_get_affine_skip_domain(scop
, pet_skip_later
);
1653 has_var_break
= scop
&& pet_scop_has_var_skip(scop
, pet_skip_later
);
1655 id_break_test
= pet_scop_get_skip_id(scop
, pet_skip_later
);
1656 if (is_non_affine
) {
1657 scop
= pet_scop_reset_context(scop
);
1659 scop
= pet_scop_reset_skips(scop
);
1660 scop
= pet_scop_resolve_nested(scop
);
1661 if (has_affine_break
) {
1662 domain
= apply_affine_break(domain
, skip
, isl_val_sgn(inc
),
1663 is_virtual
, rev_wrap
);
1664 scop
= pet_scop_intersect_domain_prefix(scop
,
1665 isl_set_copy(domain
));
1667 isl_map_free(rev_wrap
);
1669 scop
= scop_add_break(scop
, id_break_test
, isl_set_copy(domain
),
1672 scop
= scop_add_while(scop_cond
, scop
, id_test
,
1673 isl_set_copy(domain
),
1676 scop
= set_independence(scop
, tree
, domain
, isl_val_sgn(inc
),
1678 scop
= pet_scop_embed(scop
, domain
, sched
);
1679 if (is_non_affine
) {
1680 isl_set_free(valid_inc
);
1682 valid_inc
= isl_set_intersect(valid_inc
, valid_cond_next
);
1683 valid_inc
= isl_set_intersect(valid_inc
, valid_cond_init
);
1684 valid_inc
= isl_set_project_out(valid_inc
, isl_dim_set
, pos
, 1);
1685 scop
= pet_scop_restrict_context(scop
, valid_inc
);
1690 valid_init
= isl_set_project_out(valid_init
, isl_dim_set
, pos
, 1);
1691 scop
= pet_scop_restrict_context(scop
, valid_init
);
1693 pet_context_free(pc
);
1697 /* Construct a pet_scop for a for statement within the context of "pc".
1699 * We update the context to reflect the writes to the loop variable and
1700 * the writes inside the body.
1702 * Then we check if the initialization of the for loop
1703 * is a static affine value and the increment is a constant.
1704 * If so, we construct the pet_scop using scop_from_affine_for.
1705 * Otherwise, we treat the for loop as a while loop
1706 * in scop_from_non_affine_for.
1708 * Note that the initialization and the increment are extracted
1709 * in a context where the current loop iterator has been added
1710 * to the context. If these turn out not be affine, then we
1711 * have reconstruct the body context without an assignment
1712 * to this loop iterator, as this variable will then not be
1713 * treated as a dimension of the iteration domain, but as any
1716 static struct pet_scop
*scop_from_for(__isl_keep pet_tree
*tree
,
1717 __isl_keep pet_context
*init_pc
, struct pet_state
*state
)
1721 isl_pw_aff
*pa_inc
, *init_val
;
1722 pet_context
*pc
, *pc_init_val
;
1727 iv
= pet_expr_access_get_id(tree
->u
.l
.iv
);
1728 pc
= pet_context_copy(init_pc
);
1729 pc
= pet_context_add_inner_iterator(pc
, iv
);
1730 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.l
.body
);
1732 pc_init_val
= pet_context_copy(pc
);
1733 pc_init_val
= pet_context_clear_value(pc_init_val
, isl_id_copy(iv
));
1734 init_val
= pet_expr_extract_affine(tree
->u
.l
.init
, pc_init_val
);
1735 pet_context_free(pc_init_val
);
1736 pa_inc
= pet_expr_extract_affine(tree
->u
.l
.inc
, pc
);
1737 inc
= pet_extract_cst(pa_inc
);
1738 if (!pa_inc
|| !init_val
|| !inc
)
1740 if (!isl_pw_aff_involves_nan(pa_inc
) &&
1741 !isl_pw_aff_involves_nan(init_val
) && !isl_val_is_nan(inc
))
1742 return scop_from_affine_for(tree
, init_val
, pa_inc
, inc
,
1745 isl_pw_aff_free(pa_inc
);
1746 isl_pw_aff_free(init_val
);
1748 pet_context_free(pc
);
1750 pc
= pet_context_copy(init_pc
);
1751 pc
= pet_context_add_infinite_loop(pc
);
1752 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.l
.body
);
1753 return scop_from_non_affine_for(tree
, init_pc
, pc
, state
);
1755 isl_pw_aff_free(pa_inc
);
1756 isl_pw_aff_free(init_val
);
1758 pet_context_free(pc
);
1762 /* Check whether "expr" is an affine constraint within the context "pc".
1764 static int is_affine_condition(__isl_keep pet_expr
*expr
,
1765 __isl_keep pet_context
*pc
)
1770 pa
= pet_expr_extract_affine_condition(expr
, pc
);
1773 is_affine
= !isl_pw_aff_involves_nan(pa
);
1774 isl_pw_aff_free(pa
);
1779 /* Check if the given if statement is a conditional assignement
1780 * with a non-affine condition.
1782 * In particular we check if "stmt" is of the form
1789 * where the condition is non-affine and a is some array or scalar access.
1791 static int is_conditional_assignment(__isl_keep pet_tree
*tree
,
1792 __isl_keep pet_context
*pc
)
1796 pet_expr
*expr1
, *expr2
;
1798 ctx
= pet_tree_get_ctx(tree
);
1799 if (!pet_options_get_detect_conditional_assignment(ctx
))
1801 if (tree
->type
!= pet_tree_if_else
)
1803 if (tree
->u
.i
.then_body
->type
!= pet_tree_expr
)
1805 if (tree
->u
.i
.else_body
->type
!= pet_tree_expr
)
1807 expr1
= tree
->u
.i
.then_body
->u
.e
.expr
;
1808 expr2
= tree
->u
.i
.else_body
->u
.e
.expr
;
1809 if (pet_expr_get_type(expr1
) != pet_expr_op
)
1811 if (pet_expr_get_type(expr2
) != pet_expr_op
)
1813 if (pet_expr_op_get_type(expr1
) != pet_op_assign
)
1815 if (pet_expr_op_get_type(expr2
) != pet_op_assign
)
1817 expr1
= pet_expr_get_arg(expr1
, 0);
1818 expr2
= pet_expr_get_arg(expr2
, 0);
1819 equal
= pet_expr_is_equal(expr1
, expr2
);
1820 pet_expr_free(expr1
);
1821 pet_expr_free(expr2
);
1822 if (equal
< 0 || !equal
)
1824 if (is_affine_condition(tree
->u
.i
.cond
, pc
))
1830 /* Given that "tree" is of the form
1837 * where a is some array or scalar access, construct a pet_scop
1838 * corresponding to this conditional assignment within the context "pc".
1839 * "cond_pa" is an affine expression with nested accesses representing
1842 * The constructed pet_scop then corresponds to the expression
1844 * a = condition ? f(...) : g(...)
1846 * All access relations in f(...) are intersected with condition
1847 * while all access relation in g(...) are intersected with the complement.
1849 static struct pet_scop
*scop_from_conditional_assignment(
1850 __isl_keep pet_tree
*tree
, __isl_take isl_pw_aff
*cond_pa
,
1851 __isl_take pet_context
*pc
, struct pet_state
*state
)
1854 isl_set
*cond
, *comp
;
1855 isl_multi_pw_aff
*index
;
1856 pet_expr
*expr1
, *expr2
;
1857 pet_expr
*pe_cond
, *pe_then
, *pe_else
, *pe
, *pe_write
;
1858 struct pet_scop
*scop
;
1860 cond
= isl_pw_aff_non_zero_set(isl_pw_aff_copy(cond_pa
));
1861 comp
= isl_pw_aff_zero_set(isl_pw_aff_copy(cond_pa
));
1862 index
= isl_multi_pw_aff_from_pw_aff(cond_pa
);
1864 expr1
= tree
->u
.i
.then_body
->u
.e
.expr
;
1865 expr2
= tree
->u
.i
.else_body
->u
.e
.expr
;
1867 pe_cond
= pet_expr_from_index(index
);
1869 pe_then
= pet_expr_get_arg(expr1
, 1);
1870 pe_then
= pet_context_evaluate_expr(pc
, pe_then
);
1871 pe_then
= pet_expr_restrict(pe_then
, cond
);
1872 pe_else
= pet_expr_get_arg(expr2
, 1);
1873 pe_else
= pet_context_evaluate_expr(pc
, pe_else
);
1874 pe_else
= pet_expr_restrict(pe_else
, comp
);
1875 pe_write
= pet_expr_get_arg(expr1
, 0);
1876 pe_write
= pet_context_evaluate_expr(pc
, pe_write
);
1878 pe
= pet_expr_new_ternary(pe_cond
, pe_then
, pe_else
);
1879 type_size
= pet_expr_get_type_size(pe_write
);
1880 pe
= pet_expr_new_binary(type_size
, pet_op_assign
, pe_write
, pe
);
1882 scop
= scop_from_evaluated_expr(pe
, state
->n_stmt
++,
1883 pet_tree_get_loc(tree
), pc
);
1885 pet_context_free(pc
);
1890 /* Construct a pet_scop for a non-affine if statement within the context "pc".
1892 * We create a separate statement that writes the result
1893 * of the non-affine condition to a virtual scalar.
1894 * A constraint requiring the value of this virtual scalar to be one
1895 * is added to the iteration domains of the then branch.
1896 * Similarly, a constraint requiring the value of this virtual scalar
1897 * to be zero is added to the iteration domains of the else branch, if any.
1898 * We combine the schedules as a sequence to ensure that the virtual scalar
1899 * is written before it is read.
1901 * If there are any breaks or continues in the then and/or else
1902 * branches, then we may have to compute a new skip condition.
1903 * This is handled using a pet_skip_info object.
1904 * On initialization, the object checks if skip conditions need
1905 * to be computed. If so, it does so in pet_skip_info_if_extract_index and
1906 * adds them in pet_skip_info_add.
1908 static struct pet_scop
*scop_from_non_affine_if(__isl_keep pet_tree
*tree
,
1909 __isl_take pet_context
*pc
, struct pet_state
*state
)
1914 isl_multi_pw_aff
*test_index
;
1915 struct pet_skip_info skip
;
1916 struct pet_scop
*scop
, *scop_then
, *scop_else
= NULL
;
1918 has_else
= tree
->type
== pet_tree_if_else
;
1920 space
= pet_context_get_space(pc
);
1921 test_index
= pet_create_test_index(space
, state
->n_test
++);
1922 scop
= scop_from_non_affine_condition(pet_expr_copy(tree
->u
.i
.cond
),
1923 state
->n_stmt
++, isl_multi_pw_aff_copy(test_index
),
1924 pet_tree_get_loc(tree
), pc
);
1925 domain
= pet_context_get_domain(pc
);
1926 scop
= pet_scop_add_boolean_array(scop
, domain
,
1927 isl_multi_pw_aff_copy(test_index
), state
->int_size
);
1929 scop_then
= scop_from_tree(tree
->u
.i
.then_body
, pc
, state
);
1931 scop_else
= scop_from_tree(tree
->u
.i
.else_body
, pc
, state
);
1933 pet_skip_info_if_init(&skip
, state
->ctx
, scop_then
, scop_else
,
1935 pet_skip_info_if_extract_index(&skip
, test_index
, pc
, state
);
1937 scop_then
= pet_scop_filter(scop_then
,
1938 isl_multi_pw_aff_copy(test_index
), 1);
1940 scop_else
= pet_scop_filter(scop_else
, test_index
, 0);
1941 scop_then
= pet_scop_add_par(state
->ctx
, scop_then
, scop_else
);
1943 isl_multi_pw_aff_free(test_index
);
1945 scop
= pet_scop_add_seq(state
->ctx
, scop
, scop_then
);
1947 scop
= pet_skip_info_add(&skip
, scop
);
1949 pet_context_free(pc
);
1953 /* Construct a pet_scop for an affine if statement within the context "pc".
1955 * The condition is added to the iteration domains of the then branch,
1956 * while the opposite of the condition in added to the iteration domains
1957 * of the else branch, if any.
1959 * If there are any breaks or continues in the then and/or else
1960 * branches, then we may have to compute a new skip condition.
1961 * This is handled using a pet_skip_info_if object.
1962 * On initialization, the object checks if skip conditions need
1963 * to be computed. If so, it does so in pet_skip_info_if_extract_cond and
1964 * adds them in pet_skip_info_add.
1966 static struct pet_scop
*scop_from_affine_if(__isl_keep pet_tree
*tree
,
1967 __isl_take isl_pw_aff
*cond
, __isl_take pet_context
*pc
,
1968 struct pet_state
*state
)
1972 isl_set
*set
, *complement
;
1974 struct pet_skip_info skip
;
1975 struct pet_scop
*scop
, *scop_then
, *scop_else
= NULL
;
1976 pet_context
*pc_body
;
1978 ctx
= pet_tree_get_ctx(tree
);
1980 has_else
= tree
->type
== pet_tree_if_else
;
1982 valid
= isl_pw_aff_domain(isl_pw_aff_copy(cond
));
1983 set
= isl_pw_aff_non_zero_set(isl_pw_aff_copy(cond
));
1985 pc_body
= pet_context_copy(pc
);
1986 pc_body
= pet_context_intersect_domain(pc_body
, isl_set_copy(set
));
1987 scop_then
= scop_from_tree(tree
->u
.i
.then_body
, pc_body
, state
);
1988 pet_context_free(pc_body
);
1990 pc_body
= pet_context_copy(pc
);
1991 complement
= isl_set_copy(valid
);
1992 complement
= isl_set_subtract(valid
, isl_set_copy(set
));
1993 pc_body
= pet_context_intersect_domain(pc_body
,
1994 isl_set_copy(complement
));
1995 scop_else
= scop_from_tree(tree
->u
.i
.else_body
, pc_body
, state
);
1996 pet_context_free(pc_body
);
1999 pet_skip_info_if_init(&skip
, ctx
, scop_then
, scop_else
, has_else
, 1);
2000 pet_skip_info_if_extract_cond(&skip
, cond
, pc
, state
);
2001 isl_pw_aff_free(cond
);
2003 scop
= pet_scop_restrict(scop_then
, set
);
2006 scop_else
= pet_scop_restrict(scop_else
, complement
);
2007 scop
= pet_scop_add_par(ctx
, scop
, scop_else
);
2009 scop
= pet_scop_resolve_nested(scop
);
2010 scop
= pet_scop_restrict_context(scop
, valid
);
2012 scop
= pet_skip_info_add(&skip
, scop
);
2014 pet_context_free(pc
);
2018 /* Construct a pet_scop for an if statement within the context "pc".
2020 * If the condition fits the pattern of a conditional assignment,
2021 * then it is handled by scop_from_conditional_assignment.
2022 * Note that the condition is only considered for a conditional assignment
2023 * if it is not static-affine. However, it should still convert
2024 * to an affine expression when nesting is allowed.
2026 * Otherwise, we check if the condition is affine.
2027 * If so, we construct the scop in scop_from_affine_if.
2028 * Otherwise, we construct the scop in scop_from_non_affine_if.
2030 * We allow the condition to be dynamic, i.e., to refer to
2031 * scalars or array elements that may be written to outside
2032 * of the given if statement. These nested accesses are then represented
2033 * as output dimensions in the wrapping iteration domain.
2034 * If it is also written _inside_ the then or else branch, then
2035 * we treat the condition as non-affine.
2036 * As explained in extract_non_affine_if, this will introduce
2037 * an extra statement.
2038 * For aesthetic reasons, we want this statement to have a statement
2039 * number that is lower than those of the then and else branches.
2040 * In order to evaluate if we will need such a statement, however, we
2041 * first construct scops for the then and else branches.
2042 * We therefore reserve a statement number if we might have to
2043 * introduce such an extra statement.
2045 static struct pet_scop
*scop_from_if(__isl_keep pet_tree
*tree
,
2046 __isl_keep pet_context
*pc
, struct pet_state
*state
)
2050 pet_expr
*cond_expr
;
2051 pet_context
*pc_nested
;
2056 has_else
= tree
->type
== pet_tree_if_else
;
2058 pc
= pet_context_copy(pc
);
2059 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.i
.then_body
);
2061 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.i
.else_body
);
2063 cond_expr
= pet_expr_copy(tree
->u
.i
.cond
);
2064 cond_expr
= pet_context_evaluate_expr(pc
, cond_expr
);
2065 pc_nested
= pet_context_copy(pc
);
2066 pc_nested
= pet_context_set_allow_nested(pc_nested
, 1);
2067 cond
= pet_expr_extract_affine_condition(cond_expr
, pc_nested
);
2068 pet_context_free(pc_nested
);
2069 pet_expr_free(cond_expr
);
2072 pet_context_free(pc
);
2076 if (isl_pw_aff_involves_nan(cond
)) {
2077 isl_pw_aff_free(cond
);
2078 return scop_from_non_affine_if(tree
, pc
, state
);
2081 if (is_conditional_assignment(tree
, pc
))
2082 return scop_from_conditional_assignment(tree
, cond
, pc
, state
);
2084 if ((!is_nested_allowed(cond
, tree
->u
.i
.then_body
) ||
2085 (has_else
&& !is_nested_allowed(cond
, tree
->u
.i
.else_body
)))) {
2086 isl_pw_aff_free(cond
);
2087 return scop_from_non_affine_if(tree
, pc
, state
);
2090 return scop_from_affine_if(tree
, cond
, pc
, state
);
2093 /* Return a one-dimensional multi piecewise affine expression that is equal
2094 * to the constant 1 and is defined over the given domain.
2096 static __isl_give isl_multi_pw_aff
*one_mpa(__isl_take isl_space
*space
)
2098 isl_local_space
*ls
;
2101 ls
= isl_local_space_from_space(space
);
2102 aff
= isl_aff_zero_on_domain(ls
);
2103 aff
= isl_aff_set_constant_si(aff
, 1);
2105 return isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff
));
2108 /* Construct a pet_scop for a continue statement with the given domain space.
2110 * We simply create an empty scop with a universal pet_skip_now
2111 * skip condition. This skip condition will then be taken into
2112 * account by the enclosing loop construct, possibly after
2113 * being incorporated into outer skip conditions.
2115 static struct pet_scop
*scop_from_continue(__isl_keep pet_tree
*tree
,
2116 __isl_take isl_space
*space
)
2118 struct pet_scop
*scop
;
2120 scop
= pet_scop_empty(isl_space_copy(space
));
2122 scop
= pet_scop_set_skip(scop
, pet_skip_now
, one_mpa(space
));
2127 /* Construct a pet_scop for a break statement with the given domain space.
2129 * We simply create an empty scop with both a universal pet_skip_now
2130 * skip condition and a universal pet_skip_later skip condition.
2131 * These skip conditions will then be taken into
2132 * account by the enclosing loop construct, possibly after
2133 * being incorporated into outer skip conditions.
2135 static struct pet_scop
*scop_from_break(__isl_keep pet_tree
*tree
,
2136 __isl_take isl_space
*space
)
2138 struct pet_scop
*scop
;
2139 isl_multi_pw_aff
*skip
;
2141 scop
= pet_scop_empty(isl_space_copy(space
));
2143 skip
= one_mpa(space
);
2144 scop
= pet_scop_set_skip(scop
, pet_skip_now
,
2145 isl_multi_pw_aff_copy(skip
));
2146 scop
= pet_scop_set_skip(scop
, pet_skip_later
, skip
);
2151 /* Extract a clone of the kill statement in "scop".
2152 * The domain of the clone is given by "domain".
2153 * "scop" is expected to have been created from a DeclStmt
2154 * and should have the kill as its first statement.
2156 static struct pet_scop
*extract_kill(__isl_keep isl_set
*domain
,
2157 struct pet_scop
*scop
, struct pet_state
*state
)
2160 struct pet_stmt
*stmt
;
2162 isl_multi_pw_aff
*mpa
;
2165 if (!domain
|| !scop
)
2167 if (scop
->n_stmt
< 1)
2168 isl_die(isl_set_get_ctx(domain
), isl_error_internal
,
2169 "expecting at least one statement", return NULL
);
2170 stmt
= scop
->stmts
[0];
2171 if (!pet_stmt_is_kill(stmt
))
2172 isl_die(isl_set_get_ctx(domain
), isl_error_internal
,
2173 "expecting kill statement", return NULL
);
2175 kill
= pet_tree_expr_get_expr(stmt
->body
);
2176 space
= pet_stmt_get_space(stmt
);
2177 space
= isl_space_map_from_set(space
);
2178 mpa
= isl_multi_pw_aff_identity(space
);
2179 mpa
= isl_multi_pw_aff_reset_tuple_id(mpa
, isl_dim_in
);
2180 kill
= pet_expr_update_domain(kill
, mpa
);
2181 tree
= pet_tree_new_expr(kill
);
2182 tree
= pet_tree_set_loc(tree
, pet_loc_copy(stmt
->loc
));
2183 stmt
= pet_stmt_from_pet_tree(isl_set_copy(domain
),
2184 state
->n_stmt
++, tree
);
2185 return pet_scop_from_pet_stmt(isl_set_get_space(domain
), stmt
);
2188 /* Does "tree" represent an assignment to a variable?
2190 * The assignment may be one of
2191 * - a declaration with initialization
2192 * - an expression with a top-level assignment operator
2194 static int is_assignment(__isl_keep pet_tree
*tree
)
2198 if (tree
->type
== pet_tree_decl_init
)
2200 return pet_tree_is_assign(tree
);
2203 /* Update "pc" by taking into account the assignment performed by "tree",
2204 * where "tree" satisfies is_assignment.
2206 * In particular, if the lhs of the assignment is a scalar variable and
2207 * if the rhs is an affine expression, then keep track of this value in "pc"
2208 * so that we can plug it in when we later come across the same variable.
2210 * Any previously assigned value to the variable has already been removed
2211 * by scop_handle_writes.
2213 static __isl_give pet_context
*handle_assignment(__isl_take pet_context
*pc
,
2214 __isl_keep pet_tree
*tree
)
2216 pet_expr
*var
, *val
;
2220 if (pet_tree_get_type(tree
) == pet_tree_decl_init
) {
2221 var
= pet_tree_decl_get_var(tree
);
2222 val
= pet_tree_decl_get_init(tree
);
2225 expr
= pet_tree_expr_get_expr(tree
);
2226 var
= pet_expr_get_arg(expr
, 0);
2227 val
= pet_expr_get_arg(expr
, 1);
2228 pet_expr_free(expr
);
2231 if (!pet_expr_is_scalar_access(var
)) {
2237 pa
= pet_expr_extract_affine(val
, pc
);
2239 pc
= pet_context_free(pc
);
2241 if (!isl_pw_aff_involves_nan(pa
)) {
2242 id
= pet_expr_access_get_id(var
);
2243 pc
= pet_context_set_value(pc
, id
, pa
);
2245 isl_pw_aff_free(pa
);
2253 /* Mark all arrays in "scop" as being exposed.
2255 static struct pet_scop
*mark_exposed(struct pet_scop
*scop
)
2261 for (i
= 0; i
< scop
->n_array
; ++i
)
2262 scop
->arrays
[i
]->exposed
= 1;
2266 /* Try and construct a pet_scop corresponding to (part of)
2267 * a sequence of statements within the context "pc".
2269 * After extracting a statement, we update "pc"
2270 * based on the top-level assignments in the statement
2271 * so that we can exploit them in subsequent statements in the same block.
2273 * If there are any breaks or continues in the individual statements,
2274 * then we may have to compute a new skip condition.
2275 * This is handled using a pet_skip_info object.
2276 * On initialization, the object checks if skip conditions need
2277 * to be computed. If so, it does so in pet_skip_info_seq_extract and
2278 * adds them in pet_skip_info_add.
2280 * If "block" is set, then we need to insert kill statements at
2281 * the end of the block for any array that has been declared by
2282 * one of the statements in the sequence. Each of these declarations
2283 * results in the construction of a kill statement at the place
2284 * of the declaration, so we simply collect duplicates of
2285 * those kill statements and append these duplicates to the constructed scop.
2287 * If "block" is not set, then any array declared by one of the statements
2288 * in the sequence is marked as being exposed.
2290 * If autodetect is set, then we allow the extraction of only a subrange
2291 * of the sequence of statements. However, if there is at least one statement
2292 * for which we could not construct a scop and the final range contains
2293 * either no statements or at least one kill, then we discard the entire
2296 static struct pet_scop
*scop_from_block(__isl_keep pet_tree
*tree
,
2297 __isl_keep pet_context
*pc
, struct pet_state
*state
)
2303 struct pet_scop
*scop
, *kills
;
2305 ctx
= pet_tree_get_ctx(tree
);
2307 space
= pet_context_get_space(pc
);
2308 domain
= pet_context_get_domain(pc
);
2309 pc
= pet_context_copy(pc
);
2310 scop
= pet_scop_empty(isl_space_copy(space
));
2311 kills
= pet_scop_empty(space
);
2312 for (i
= 0; i
< tree
->u
.b
.n
; ++i
) {
2313 struct pet_scop
*scop_i
;
2315 if (pet_scop_has_affine_skip(scop
, pet_skip_now
))
2316 pc
= apply_affine_continue(pc
, scop
);
2317 scop_i
= scop_from_tree(tree
->u
.b
.child
[i
], pc
, state
);
2318 pc
= scop_handle_writes(scop_i
, pc
);
2319 if (is_assignment(tree
->u
.b
.child
[i
]))
2320 pc
= handle_assignment(pc
, tree
->u
.b
.child
[i
]);
2321 struct pet_skip_info skip
;
2322 pet_skip_info_seq_init(&skip
, ctx
, scop
, scop_i
);
2323 pet_skip_info_seq_extract(&skip
, pc
, state
);
2324 if (scop_i
&& pet_tree_is_decl(tree
->u
.b
.child
[i
])) {
2325 if (tree
->u
.b
.block
) {
2326 struct pet_scop
*kill
;
2327 kill
= extract_kill(domain
, scop_i
, state
);
2328 kills
= pet_scop_add_par(ctx
, kills
, kill
);
2330 scop_i
= mark_exposed(scop_i
);
2332 scop
= pet_scop_add_seq(ctx
, scop
, scop_i
);
2334 scop
= pet_skip_info_add(&skip
, scop
);
2339 isl_set_free(domain
);
2341 scop
= pet_scop_add_seq(ctx
, scop
, kills
);
2343 pet_context_free(pc
);
2348 /* Internal data structure for extract_declared_arrays.
2350 * "pc" and "state" are used to create pet_array objects and kill statements.
2351 * "any" is initialized to 0 by the caller and set to 1 as soon as we have
2352 * found any declared array.
2353 * "scop" has been initialized by the caller and is used to attach
2354 * the created pet_array objects.
2355 * "kill_before" and "kill_after" are created and updated by
2356 * extract_declared_arrays to collect the kills of the arrays.
2358 struct pet_tree_extract_declared_arrays_data
{
2360 struct pet_state
*state
;
2365 struct pet_scop
*scop
;
2366 struct pet_scop
*kill_before
;
2367 struct pet_scop
*kill_after
;
2370 /* Check if the node "node" declares any array or scalar.
2371 * If so, create the corresponding pet_array and attach it to data->scop.
2372 * Additionally, create two kill statements for the array and add them
2373 * to data->kill_before and data->kill_after.
2375 static int extract_declared_arrays(__isl_keep pet_tree
*node
, void *user
)
2377 enum pet_tree_type type
;
2378 struct pet_tree_extract_declared_arrays_data
*data
= user
;
2379 struct pet_array
*array
;
2380 struct pet_scop
*scop_kill
;
2383 type
= pet_tree_get_type(node
);
2384 if (type
== pet_tree_decl
|| type
== pet_tree_decl_init
)
2385 var
= node
->u
.d
.var
;
2386 else if (type
== pet_tree_for
&& node
->u
.l
.declared
)
2391 array
= extract_array(var
, data
->pc
, data
->state
);
2393 array
->declared
= 1;
2394 data
->scop
= pet_scop_add_array(data
->scop
, array
);
2396 scop_kill
= kill(pet_tree_get_loc(node
), array
, data
->pc
, data
->state
);
2398 data
->kill_before
= scop_kill
;
2400 data
->kill_before
= pet_scop_add_par(data
->ctx
,
2401 data
->kill_before
, scop_kill
);
2403 scop_kill
= kill(pet_tree_get_loc(node
), array
, data
->pc
, data
->state
);
2405 data
->kill_after
= scop_kill
;
2407 data
->kill_after
= pet_scop_add_par(data
->ctx
,
2408 data
->kill_after
, scop_kill
);
2415 /* Convert a pet_tree that consists of more than a single leaf
2416 * to a pet_scop with a single statement encapsulating the entire pet_tree.
2417 * Do so within the context of "pc".
2419 * After constructing the core scop, we also look for any arrays (or scalars)
2420 * that are declared inside "tree". Each of those arrays is marked as
2421 * having been declared and kill statements for these arrays
2422 * are introduced before and after the core scop.
2423 * Note that the input tree is not a leaf so that the declaration
2424 * cannot occur at the outer level.
2426 static struct pet_scop
*scop_from_tree_macro(__isl_take pet_tree
*tree
,
2427 __isl_take isl_id
*label
, __isl_keep pet_context
*pc
,
2428 struct pet_state
*state
)
2430 struct pet_tree_extract_declared_arrays_data data
= { pc
, state
};
2432 data
.scop
= scop_from_unevaluated_tree(pet_tree_copy(tree
),
2433 state
->n_stmt
++, pc
);
2436 data
.ctx
= pet_context_get_ctx(pc
);
2437 if (pet_tree_foreach_sub_tree(tree
, &extract_declared_arrays
,
2439 data
.scop
= pet_scop_free(data
.scop
);
2440 pet_tree_free(tree
);
2445 data
.scop
= pet_scop_add_seq(data
.ctx
, data
.kill_before
, data
.scop
);
2446 data
.scop
= pet_scop_add_seq(data
.ctx
, data
.scop
, data
.kill_after
);
2451 /* Construct a pet_scop that corresponds to the pet_tree "tree"
2452 * within the context "pc" by calling the appropriate function
2453 * based on the type of "tree".
2455 * If the initially constructed pet_scop turns out to involve
2456 * dynamic control and if the user has requested an encapsulation
2457 * of all dynamic control, then this pet_scop is discarded and
2458 * a new pet_scop is created with a single statement representing
2459 * the entire "tree".
2460 * However, if the scop contains any active continue or break,
2461 * then we need to include the loop containing the continue or break
2462 * in the encapsulation. We therefore postpone the encapsulation
2463 * until we have constructed a pet_scop for this enclosing loop.
2465 static struct pet_scop
*scop_from_tree(__isl_keep pet_tree
*tree
,
2466 __isl_keep pet_context
*pc
, struct pet_state
*state
)
2469 struct pet_scop
*scop
= NULL
;
2474 ctx
= pet_tree_get_ctx(tree
);
2475 switch (tree
->type
) {
2476 case pet_tree_error
:
2478 case pet_tree_block
:
2479 return scop_from_block(tree
, pc
, state
);
2480 case pet_tree_break
:
2481 return scop_from_break(tree
, pet_context_get_space(pc
));
2482 case pet_tree_continue
:
2483 return scop_from_continue(tree
, pet_context_get_space(pc
));
2485 case pet_tree_decl_init
:
2486 return scop_from_decl(tree
, pc
, state
);
2488 return scop_from_tree_expr(tree
, pc
, state
);
2490 case pet_tree_if_else
:
2491 scop
= scop_from_if(tree
, pc
, state
);
2494 scop
= scop_from_for(tree
, pc
, state
);
2496 case pet_tree_while
:
2497 scop
= scop_from_while(tree
, pc
, state
);
2499 case pet_tree_infinite_loop
:
2500 scop
= scop_from_infinite_for(tree
, pc
, state
);
2507 if (!pet_options_get_encapsulate_dynamic_control(ctx
) ||
2508 !pet_scop_has_data_dependent_conditions(scop
) ||
2509 pet_scop_has_var_skip(scop
, pet_skip_now
))
2512 pet_scop_free(scop
);
2513 return scop_from_tree_macro(pet_tree_copy(tree
),
2514 isl_id_copy(tree
->label
), pc
, state
);
2517 /* If "tree" has a label that is of the form S_<nr>, then make
2518 * sure that state->n_stmt is greater than nr to ensure that
2519 * we will not generate S_<nr> ourselves.
2521 static int set_first_stmt(__isl_keep pet_tree
*tree
, void *user
)
2523 struct pet_state
*state
= user
;
2531 name
= isl_id_get_name(tree
->label
);
2532 if (strncmp(name
, "S_", 2) != 0)
2534 nr
= atoi(name
+ 2);
2535 if (nr
>= state
->n_stmt
)
2536 state
->n_stmt
= nr
+ 1;
2541 /* Construct a pet_scop that corresponds to the pet_tree "tree".
2542 * "int_size" is the number of bytes need to represent an integer.
2543 * "extract_array" is a callback that we can use to create a pet_array
2544 * that corresponds to the variable accessed by an expression.
2546 * Initialize the global state, construct a context and then
2547 * construct the pet_scop by recursively visiting the tree.
2549 * state.n_stmt is initialized to point beyond any explicit S_<nr> label.
2551 struct pet_scop
*pet_scop_from_pet_tree(__isl_take pet_tree
*tree
, int int_size
,
2552 struct pet_array
*(*extract_array
)(__isl_keep pet_expr
*access
,
2553 __isl_keep pet_context
*pc
, void *user
), void *user
,
2554 __isl_keep pet_context
*pc
)
2556 struct pet_scop
*scop
;
2557 struct pet_state state
= { 0 };
2562 state
.ctx
= pet_tree_get_ctx(tree
);
2563 state
.int_size
= int_size
;
2564 state
.extract_array
= extract_array
;
2566 if (pet_tree_foreach_sub_tree(tree
, &set_first_stmt
, &state
) < 0)
2567 tree
= pet_tree_free(tree
);
2569 scop
= scop_from_tree(tree
, pc
, &state
);
2570 scop
= pet_scop_set_loc(scop
, pet_tree_get_loc(tree
));
2572 pet_tree_free(tree
);
2575 scop
->context
= isl_set_params(scop
->context
);