2 * Copyright 2011 Leiden University. All rights reserved.
3 * Copyright 2012-2014 Ecole Normale Superieure. All rights reserved.
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above
13 * copyright notice, this list of conditions and the following
14 * disclaimer in the documentation and/or other materials provided
15 * with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY LEIDEN UNIVERSITY ''AS IS'' AND ANY
18 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
20 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL LEIDEN UNIVERSITY OR
21 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
22 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
23 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA,
24 * OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
27 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29 * The views and conclusions contained in the software and documentation
30 * are those of the authors and should not be interpreted as
31 * representing official policies, either expressed or implied, of
35 #include <isl/id_to_pw_aff.h>
44 #include "tree2scop.h"
46 /* Update "pc" by taking into account the writes in "stmt".
47 * That is, first mark all scalar variables that are written by "stmt"
48 * as having an unknown value. Afterwards,
49 * if "stmt" is a top-level (i.e., unconditional) assignment
50 * to a scalar variable, then update "pc" accordingly.
52 * In particular, if the lhs of the assignment is a scalar variable, then mark
53 * the variable as having been assigned. If, furthermore, the rhs
54 * is an affine expression, then keep track of this value in "pc"
55 * so that we can plug it in when we later come across the same variable.
57 * We skip assignments to virtual arrays (those with NULL user pointer).
59 static __isl_give pet_context
*handle_writes(struct pet_stmt
*stmt
,
60 __isl_take pet_context
*pc
)
62 pet_expr
*body
= stmt
->body
;
67 pc
= pet_context_clear_writes_in_expr(pc
, body
);
71 if (pet_expr_get_type(body
) != pet_expr_op
)
73 if (pet_expr_op_get_type(body
) != pet_op_assign
)
75 if (!isl_set_plain_is_universe(stmt
->domain
))
77 arg
= pet_expr_get_arg(body
, 0);
78 if (!pet_expr_is_scalar_access(arg
)) {
83 id
= pet_expr_access_get_id(arg
);
86 if (!isl_id_get_user(id
)) {
91 arg
= pet_expr_get_arg(body
, 1);
92 pa
= pet_expr_extract_affine(arg
, pc
);
93 pc
= pet_context_mark_assigned(pc
, isl_id_copy(id
));
96 if (pa
&& isl_pw_aff_involves_nan(pa
)) {
102 pc
= pet_context_set_value(pc
, id
, pa
);
107 /* Update "pc" based on the write accesses (and, in particular,
108 * assignments) in "scop".
110 static __isl_give pet_context
*scop_handle_writes(struct pet_scop
*scop
,
111 __isl_take pet_context
*pc
)
116 return pet_context_free(pc
);
117 for (i
= 0; i
< scop
->n_stmt
; ++i
)
118 pc
= handle_writes(scop
->stmts
[i
], pc
);
123 /* Convert a top-level pet_expr to a pet_scop with one statement
124 * within the context "pc".
125 * This mainly involves resolving nested expression parameters
126 * and setting the name of the iteration space.
127 * The name is given by "label" if it is non-NULL. Otherwise,
128 * it is of the form S_<stmt_nr>.
129 * The location of the statement is set to "loc".
131 static struct pet_scop
*scop_from_expr(__isl_take pet_expr
*expr
,
132 __isl_take isl_id
*label
, int stmt_nr
, __isl_take pet_loc
*loc
,
133 __isl_keep pet_context
*pc
)
138 ctx
= pet_expr_get_ctx(expr
);
140 expr
= pet_expr_plug_in_args(expr
, pc
);
141 expr
= pet_expr_resolve_nested(expr
);
142 expr
= pet_expr_resolve_assume(expr
, pc
);
143 ps
= pet_stmt_from_pet_expr(loc
, label
, stmt_nr
, expr
);
144 return pet_scop_from_pet_stmt(ctx
, ps
);
147 /* Construct a pet_scop with a single statement killing the entire
149 * The location of the statement is set to "loc".
151 static struct pet_scop
*kill(__isl_take pet_loc
*loc
, struct pet_array
*array
,
152 __isl_keep pet_context
*pc
, struct pet_state
*state
)
157 isl_multi_pw_aff
*index
;
160 struct pet_scop
*scop
;
164 ctx
= isl_set_get_ctx(array
->extent
);
165 access
= isl_map_from_range(isl_set_copy(array
->extent
));
166 id
= isl_set_get_tuple_id(array
->extent
);
167 space
= isl_space_alloc(ctx
, 0, 0, 0);
168 space
= isl_space_set_tuple_id(space
, isl_dim_out
, id
);
169 index
= isl_multi_pw_aff_zero(space
);
170 expr
= pet_expr_kill_from_access_and_index(access
, index
);
171 return scop_from_expr(expr
, NULL
, state
->n_stmt
++, loc
, pc
);
177 /* Construct and return a pet_array corresponding to the variable
178 * accessed by "access" by calling the extract_array callback.
180 static struct pet_array
*extract_array(__isl_keep pet_expr
*access
,
181 __isl_keep pet_context
*pc
, struct pet_state
*state
)
183 return state
->extract_array(access
, pc
, state
->user
);
186 /* Construct a pet_scop for a (single) variable declaration
187 * within the context "pc".
189 * The scop contains the variable being declared (as an array)
190 * and a statement killing the array.
192 * If the declaration comes with an initialization, then the scop
193 * also contains an assignment to the variable.
195 static struct pet_scop
*scop_from_decl(__isl_keep pet_tree
*tree
,
196 __isl_keep pet_context
*pc
, struct pet_state
*state
)
200 struct pet_array
*array
;
201 struct pet_scop
*scop_decl
, *scop
;
202 pet_expr
*lhs
, *rhs
, *pe
;
204 array
= extract_array(tree
->u
.d
.var
, pc
, state
);
207 scop_decl
= kill(pet_tree_get_loc(tree
), array
, pc
, state
);
208 scop_decl
= pet_scop_add_array(scop_decl
, array
);
210 if (tree
->type
!= pet_tree_decl_init
)
213 lhs
= pet_expr_copy(tree
->u
.d
.var
);
214 rhs
= pet_expr_copy(tree
->u
.d
.init
);
215 type_size
= pet_expr_get_type_size(lhs
);
216 pe
= pet_expr_new_binary(type_size
, pet_op_assign
, lhs
, rhs
);
217 scop
= scop_from_expr(pe
, NULL
, state
->n_stmt
++,
218 pet_tree_get_loc(tree
), pc
);
220 scop_decl
= pet_scop_prefix(scop_decl
, 0);
221 scop
= pet_scop_prefix(scop
, 1);
223 ctx
= pet_tree_get_ctx(tree
);
224 scop
= pet_scop_add_seq(ctx
, scop_decl
, scop
);
229 /* Embed the given iteration domain in an extra outer loop
230 * with induction variable "var".
231 * If this variable appeared as a parameter in the constraints,
232 * it is replaced by the new outermost dimension.
234 static __isl_give isl_set
*embed(__isl_take isl_set
*set
,
235 __isl_take isl_id
*var
)
239 set
= isl_set_insert_dims(set
, isl_dim_set
, 0, 1);
240 pos
= isl_set_find_dim_by_id(set
, isl_dim_param
, var
);
242 set
= isl_set_equate(set
, isl_dim_param
, pos
, isl_dim_set
, 0);
243 set
= isl_set_project_out(set
, isl_dim_param
, pos
, 1);
250 /* Return those elements in the space of "cond" that come after
251 * (based on "sign") an element in "cond".
253 static __isl_give isl_set
*after(__isl_take isl_set
*cond
, int sign
)
255 isl_map
*previous_to_this
;
258 previous_to_this
= isl_map_lex_lt(isl_set_get_space(cond
));
260 previous_to_this
= isl_map_lex_gt(isl_set_get_space(cond
));
262 cond
= isl_set_apply(cond
, previous_to_this
);
267 /* Remove those iterations of "domain" that have an earlier iteration
268 * (based on "sign") where "skip" is satisfied.
269 * "domain" has an extra outer loop compared to "skip".
270 * The skip condition is first embedded in the same space as "domain".
271 * If "apply_skip_map" is set, then "skip_map" is first applied
272 * to the embedded skip condition before removing it from the domain.
274 static __isl_give isl_set
*apply_affine_break(__isl_take isl_set
*domain
,
275 __isl_take isl_set
*skip
, int sign
,
276 int apply_skip_map
, __isl_keep isl_map
*skip_map
)
278 skip
= embed(skip
, isl_set_get_dim_id(domain
, isl_dim_set
, 0));
280 skip
= isl_set_apply(skip
, isl_map_copy(skip_map
));
281 skip
= isl_set_intersect(skip
, isl_set_copy(domain
));
282 return isl_set_subtract(domain
, after(skip
, sign
));
285 /* Create the infinite iteration domain
289 static __isl_give isl_set
*infinite_domain(__isl_take isl_id
*id
)
291 isl_ctx
*ctx
= isl_id_get_ctx(id
);
294 domain
= isl_set_nat_universe(isl_space_set_alloc(ctx
, 0, 1));
295 domain
= isl_set_set_dim_id(domain
, isl_dim_set
, 0, id
);
300 /* Create an identity affine expression on the space containing "domain",
301 * which is assumed to be one-dimensional.
303 static __isl_give isl_aff
*identity_aff(__isl_keep isl_set
*domain
)
307 ls
= isl_local_space_from_space(isl_set_get_space(domain
));
308 return isl_aff_var_on_domain(ls
, isl_dim_set
, 0);
311 /* Create an affine expression that maps elements
312 * of a single-dimensional array "id_test" to the previous element
313 * (according to "inc"), provided this element belongs to "domain".
314 * That is, create the affine expression
316 * { id[x] -> id[x - inc] : x - inc in domain }
318 static __isl_give isl_multi_pw_aff
*map_to_previous(__isl_take isl_id
*id_test
,
319 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
324 isl_multi_pw_aff
*prev
;
326 space
= isl_set_get_space(domain
);
327 ls
= isl_local_space_from_space(space
);
328 aff
= isl_aff_var_on_domain(ls
, isl_dim_set
, 0);
329 aff
= isl_aff_add_constant_val(aff
, isl_val_neg(inc
));
330 prev
= isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff
));
331 domain
= isl_set_preimage_multi_pw_aff(domain
,
332 isl_multi_pw_aff_copy(prev
));
333 prev
= isl_multi_pw_aff_intersect_domain(prev
, domain
);
334 prev
= isl_multi_pw_aff_set_tuple_id(prev
, isl_dim_out
, id_test
);
339 /* Add an implication to "scop" expressing that if an element of
340 * virtual array "id_test" has value "satisfied" then all previous elements
341 * of this array also have that value. The set of previous elements
342 * is bounded by "domain". If "sign" is negative then the iterator
343 * is decreasing and we express that all subsequent array elements
344 * (but still defined previously) have the same value.
346 static struct pet_scop
*add_implication(struct pet_scop
*scop
,
347 __isl_take isl_id
*id_test
, __isl_take isl_set
*domain
, int sign
,
353 domain
= isl_set_set_tuple_id(domain
, id_test
);
354 space
= isl_set_get_space(domain
);
356 map
= isl_map_lex_ge(space
);
358 map
= isl_map_lex_le(space
);
359 map
= isl_map_intersect_range(map
, domain
);
360 scop
= pet_scop_add_implication(scop
, map
, satisfied
);
365 /* Add a filter to "scop" that imposes that it is only executed
366 * when the variable identified by "id_test" has a zero value
367 * for all previous iterations of "domain".
369 * In particular, add a filter that imposes that the array
370 * has a zero value at the previous iteration of domain and
371 * add an implication that implies that it then has that
372 * value for all previous iterations.
374 static struct pet_scop
*scop_add_break(struct pet_scop
*scop
,
375 __isl_take isl_id
*id_test
, __isl_take isl_set
*domain
,
376 __isl_take isl_val
*inc
)
378 isl_multi_pw_aff
*prev
;
379 int sign
= isl_val_sgn(inc
);
381 prev
= map_to_previous(isl_id_copy(id_test
), isl_set_copy(domain
), inc
);
382 scop
= add_implication(scop
, id_test
, domain
, sign
, 0);
383 scop
= pet_scop_filter(scop
, prev
, 0);
388 static struct pet_scop
*scop_from_tree(__isl_keep pet_tree
*tree
,
389 __isl_keep pet_context
*pc
, struct pet_state
*state
);
391 /* Construct a pet_scop for an infinite loop around the given body
392 * within the context "pc".
394 * We extract a pet_scop for the body and then embed it in a loop with
403 * If the body contains any break, then it is taken into
404 * account in apply_affine_break (if the skip condition is affine)
405 * or in scop_add_break (if the skip condition is not affine).
407 * Note that in case of an affine skip condition,
408 * since we are dealing with a loop without loop iterator,
409 * the skip condition cannot refer to the current loop iterator and
410 * so effectively, the iteration domain is of the form
412 * { [0]; [t] : t >= 1 and not skip }
414 static struct pet_scop
*scop_from_infinite_loop(__isl_keep pet_tree
*body
,
415 __isl_keep pet_context
*pc
, struct pet_state
*state
)
418 isl_id
*id
, *id_test
;
422 struct pet_scop
*scop
;
423 int has_affine_break
;
426 ctx
= pet_tree_get_ctx(body
);
427 id
= isl_id_alloc(ctx
, "t", NULL
);
428 domain
= infinite_domain(isl_id_copy(id
));
429 ident
= identity_aff(domain
);
431 scop
= scop_from_tree(body
, pc
, state
);
433 has_affine_break
= pet_scop_has_affine_skip(scop
, pet_skip_later
);
434 if (has_affine_break
)
435 skip
= pet_scop_get_affine_skip_domain(scop
, pet_skip_later
);
436 has_var_break
= pet_scop_has_var_skip(scop
, pet_skip_later
);
438 id_test
= pet_scop_get_skip_id(scop
, pet_skip_later
);
440 scop
= pet_scop_embed(scop
, isl_set_copy(domain
),
441 isl_aff_copy(ident
), ident
, id
);
442 if (has_affine_break
) {
443 domain
= apply_affine_break(domain
, skip
, 1, 0, NULL
);
444 scop
= pet_scop_intersect_domain_prefix(scop
,
445 isl_set_copy(domain
));
448 scop
= scop_add_break(scop
, id_test
, domain
, isl_val_one(ctx
));
450 isl_set_free(domain
);
455 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
460 * within the context "pc".
462 static struct pet_scop
*scop_from_infinite_for(__isl_keep pet_tree
*tree
,
463 __isl_keep pet_context
*pc
, struct pet_state
*state
)
465 struct pet_scop
*scop
;
467 pc
= pet_context_copy(pc
);
468 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.l
.body
);
470 scop
= scop_from_infinite_loop(tree
->u
.l
.body
, pc
, state
);
472 pet_context_free(pc
);
477 /* Construct a pet_scop for a while loop of the form
482 * within the context "pc".
483 * In particular, construct a scop for an infinite loop around body and
484 * intersect the domain with the affine expression.
485 * Note that this intersection may result in an empty loop.
487 static struct pet_scop
*scop_from_affine_while(__isl_keep pet_tree
*tree
,
488 __isl_take isl_pw_aff
*pa
, __isl_take pet_context
*pc
,
489 struct pet_state
*state
)
491 struct pet_scop
*scop
;
495 valid
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
496 dom
= isl_pw_aff_non_zero_set(pa
);
497 scop
= scop_from_infinite_loop(tree
->u
.l
.body
, pc
, state
);
498 scop
= pet_scop_restrict(scop
, isl_set_params(dom
));
499 scop
= pet_scop_restrict_context(scop
, isl_set_params(valid
));
501 pet_context_free(pc
);
505 /* Construct a scop for a while, given the scops for the condition
506 * and the body, the filter identifier and the iteration domain of
509 * In particular, the scop for the condition is filtered to depend
510 * on "id_test" evaluating to true for all previous iterations
511 * of the loop, while the scop for the body is filtered to depend
512 * on "id_test" evaluating to true for all iterations up to the
514 * The actual filter only imposes that this virtual array has
515 * value one on the previous or the current iteration.
516 * The fact that this condition also applies to the previous
517 * iterations is enforced by an implication.
519 * These filtered scops are then combined into a single scop.
521 * "sign" is positive if the iterator increases and negative
524 static struct pet_scop
*scop_add_while(struct pet_scop
*scop_cond
,
525 struct pet_scop
*scop_body
, __isl_take isl_id
*id_test
,
526 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
528 isl_ctx
*ctx
= isl_set_get_ctx(domain
);
530 isl_multi_pw_aff
*test_index
;
531 isl_multi_pw_aff
*prev
;
532 int sign
= isl_val_sgn(inc
);
533 struct pet_scop
*scop
;
535 prev
= map_to_previous(isl_id_copy(id_test
), isl_set_copy(domain
), inc
);
536 scop_cond
= pet_scop_filter(scop_cond
, prev
, 1);
538 space
= isl_space_map_from_set(isl_set_get_space(domain
));
539 test_index
= isl_multi_pw_aff_identity(space
);
540 test_index
= isl_multi_pw_aff_set_tuple_id(test_index
, isl_dim_out
,
541 isl_id_copy(id_test
));
542 scop_body
= pet_scop_filter(scop_body
, test_index
, 1);
544 scop
= pet_scop_add_seq(ctx
, scop_cond
, scop_body
);
545 scop
= add_implication(scop
, id_test
, domain
, sign
, 1);
550 /* Create a pet_scop with a single statement with name S_<stmt_nr>,
551 * evaluating "cond" and writing the result to a virtual scalar,
552 * as expressed by "index".
553 * Do so within the context "pc".
554 * The location of the statement is set to "loc".
556 static struct pet_scop
*scop_from_non_affine_condition(
557 __isl_take pet_expr
*cond
, int stmt_nr
,
558 __isl_take isl_multi_pw_aff
*index
,
559 __isl_take pet_loc
*loc
, __isl_keep pet_context
*pc
)
561 pet_expr
*expr
, *write
;
563 write
= pet_expr_from_index(index
);
564 write
= pet_expr_access_set_write(write
, 1);
565 write
= pet_expr_access_set_read(write
, 0);
566 expr
= pet_expr_new_binary(1, pet_op_assign
, write
, cond
);
568 return scop_from_expr(expr
, NULL
, stmt_nr
, loc
, pc
);
571 /* Construct a generic while scop, with iteration domain
572 * { [t] : t >= 0 } around the scop for "tree_body" within the context "pc".
573 * The scop consists of two parts,
574 * one for evaluating the condition "cond" and one for the body.
575 * "test_nr" is the sequence number of the virtual test variable that contains
576 * the result of the condition and "stmt_nr" is the sequence number
577 * of the statement that evaluates the condition.
578 * If "expr_inc" is not NULL, then a scop for evaluating this expression
579 * is added at the end of the body,
580 * after replacing any skip conditions resulting from continue statements
581 * by the skip conditions resulting from break statements (if any).
583 * The schedule is adjusted to reflect that the condition is evaluated
584 * before the body is executed and the body is filtered to depend
585 * on the result of the condition evaluating to true on all iterations
586 * up to the current iteration, while the evaluation of the condition itself
587 * is filtered to depend on the result of the condition evaluating to true
588 * on all previous iterations.
589 * The context of the scop representing the body is dropped
590 * because we don't know how many times the body will be executed,
593 * If the body contains any break, then it is taken into
594 * account in apply_affine_break (if the skip condition is affine)
595 * or in scop_add_break (if the skip condition is not affine).
597 * Note that in case of an affine skip condition,
598 * since we are dealing with a loop without loop iterator,
599 * the skip condition cannot refer to the current loop iterator and
600 * so effectively, the iteration domain is of the form
602 * { [0]; [t] : t >= 1 and not skip }
604 static struct pet_scop
*scop_from_non_affine_while(__isl_take pet_expr
*cond
,
605 int test_nr
, int stmt_nr
, __isl_take pet_loc
*loc
,
606 __isl_keep pet_tree
*tree_body
, __isl_take pet_expr
*expr_inc
,
607 __isl_take pet_context
*pc
, struct pet_state
*state
)
610 isl_id
*id
, *id_test
, *id_break_test
;
611 isl_multi_pw_aff
*test_index
;
615 struct pet_scop
*scop
, *scop_body
;
616 int has_affine_break
;
620 test_index
= pet_create_test_index(ctx
, test_nr
);
621 scop
= scop_from_non_affine_condition(cond
, stmt_nr
,
622 isl_multi_pw_aff_copy(test_index
),
623 pet_loc_copy(loc
), pc
);
624 id_test
= isl_multi_pw_aff_get_tuple_id(test_index
, isl_dim_out
);
625 scop
= pet_scop_add_boolean_array(scop
, test_index
, state
->int_size
);
627 id
= isl_id_alloc(ctx
, "t", NULL
);
628 domain
= infinite_domain(isl_id_copy(id
));
629 ident
= identity_aff(domain
);
631 scop_body
= scop_from_tree(tree_body
, pc
, state
);
633 has_affine_break
= pet_scop_has_affine_skip(scop_body
, pet_skip_later
);
634 if (has_affine_break
)
635 skip
= pet_scop_get_affine_skip_domain(scop_body
,
637 has_var_break
= pet_scop_has_var_skip(scop_body
, pet_skip_later
);
639 id_break_test
= pet_scop_get_skip_id(scop_body
, pet_skip_later
);
641 scop
= pet_scop_prefix(scop
, 0);
642 scop
= pet_scop_embed(scop
, isl_set_copy(domain
), isl_aff_copy(ident
),
643 isl_aff_copy(ident
), isl_id_copy(id
));
644 scop_body
= pet_scop_reset_context(scop_body
);
645 scop_body
= pet_scop_prefix(scop_body
, 1);
647 struct pet_scop
*scop_inc
;
648 scop_inc
= scop_from_expr(expr_inc
, NULL
, state
->n_stmt
++,
650 scop_inc
= pet_scop_prefix(scop_inc
, 2);
651 if (pet_scop_has_skip(scop_body
, pet_skip_later
)) {
652 isl_multi_pw_aff
*skip
;
653 skip
= pet_scop_get_skip(scop_body
, pet_skip_later
);
654 scop_body
= pet_scop_set_skip(scop_body
,
657 pet_scop_reset_skip(scop_body
, pet_skip_now
);
658 scop_body
= pet_scop_add_seq(ctx
, scop_body
, scop_inc
);
661 scop_body
= pet_scop_embed(scop_body
, isl_set_copy(domain
),
662 isl_aff_copy(ident
), ident
, id
);
664 if (has_affine_break
) {
665 domain
= apply_affine_break(domain
, skip
, 1, 0, NULL
);
666 scop
= pet_scop_intersect_domain_prefix(scop
,
667 isl_set_copy(domain
));
668 scop_body
= pet_scop_intersect_domain_prefix(scop_body
,
669 isl_set_copy(domain
));
672 scop
= scop_add_break(scop
, isl_id_copy(id_break_test
),
673 isl_set_copy(domain
), isl_val_one(ctx
));
674 scop_body
= scop_add_break(scop_body
, id_break_test
,
675 isl_set_copy(domain
), isl_val_one(ctx
));
677 scop
= scop_add_while(scop
, scop_body
, id_test
, domain
,
680 pet_context_free(pc
);
684 /* Check if the while loop is of the form
686 * while (affine expression)
689 * If so, call scop_from_affine_while to construct a scop.
691 * Otherwise, pass control to scop_from_non_affine_while.
693 * "pc" is the context in which the affine expressions in the scop are created.
695 static struct pet_scop
*scop_from_while(__isl_keep pet_tree
*tree
,
696 __isl_keep pet_context
*pc
, struct pet_state
*state
)
699 int test_nr
, stmt_nr
;
705 pc
= pet_context_copy(pc
);
706 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.l
.body
);
708 cond_expr
= pet_expr_copy(tree
->u
.l
.cond
);
709 cond_expr
= pet_expr_plug_in_args(cond_expr
, pc
);
710 pa
= pet_expr_extract_affine_condition(cond_expr
, pc
);
711 pet_expr_free(cond_expr
);
716 if (!isl_pw_aff_involves_nan(pa
))
717 return scop_from_affine_while(tree
, pa
, pc
, state
);
719 test_nr
= state
->n_test
++;
720 stmt_nr
= state
->n_stmt
++;
721 return scop_from_non_affine_while(pet_expr_copy(tree
->u
.l
.cond
),
722 test_nr
, stmt_nr
, pet_tree_get_loc(tree
),
723 tree
->u
.l
.body
, NULL
, pc
, state
);
725 pet_context_free(pc
);
729 /* Check whether "cond" expresses a simple loop bound
730 * on the only set dimension.
731 * In particular, if "up" is set then "cond" should contain only
732 * upper bounds on the set dimension.
733 * Otherwise, it should contain only lower bounds.
735 static int is_simple_bound(__isl_keep isl_set
*cond
, __isl_keep isl_val
*inc
)
737 if (isl_val_is_pos(inc
))
738 return !isl_set_dim_has_any_lower_bound(cond
, isl_dim_set
, 0);
740 return !isl_set_dim_has_any_upper_bound(cond
, isl_dim_set
, 0);
743 /* Extend a condition on a given iteration of a loop to one that
744 * imposes the same condition on all previous iterations.
745 * "domain" expresses the lower [upper] bound on the iterations
746 * when inc is positive [negative].
748 * In particular, we construct the condition (when inc is positive)
750 * forall i' : (domain(i') and i' <= i) => cond(i')
752 * which is equivalent to
754 * not exists i' : domain(i') and i' <= i and not cond(i')
756 * We construct this set by negating cond, applying a map
758 * { [i'] -> [i] : domain(i') and i' <= i }
760 * and then negating the result again.
762 static __isl_give isl_set
*valid_for_each_iteration(__isl_take isl_set
*cond
,
763 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
765 isl_map
*previous_to_this
;
767 if (isl_val_is_pos(inc
))
768 previous_to_this
= isl_map_lex_le(isl_set_get_space(domain
));
770 previous_to_this
= isl_map_lex_ge(isl_set_get_space(domain
));
772 previous_to_this
= isl_map_intersect_domain(previous_to_this
, domain
);
774 cond
= isl_set_complement(cond
);
775 cond
= isl_set_apply(cond
, previous_to_this
);
776 cond
= isl_set_complement(cond
);
783 /* Construct a domain of the form
785 * [id] -> { : exists a: id = init + a * inc and a >= 0 }
787 static __isl_give isl_set
*strided_domain(__isl_take isl_id
*id
,
788 __isl_take isl_pw_aff
*init
, __isl_take isl_val
*inc
)
794 init
= isl_pw_aff_insert_dims(init
, isl_dim_in
, 0, 1);
795 dim
= isl_pw_aff_get_domain_space(init
);
796 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
797 aff
= isl_aff_add_coefficient_val(aff
, isl_dim_in
, 0, inc
);
798 init
= isl_pw_aff_add(init
, isl_pw_aff_from_aff(aff
));
800 dim
= isl_space_set_alloc(isl_pw_aff_get_ctx(init
), 1, 1);
801 dim
= isl_space_set_dim_id(dim
, isl_dim_param
, 0, id
);
802 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
803 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_param
, 0, 1);
805 set
= isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff
), init
);
807 set
= isl_set_lower_bound_si(set
, isl_dim_set
, 0, 0);
809 return isl_set_params(set
);
812 /* Assuming "cond" represents a bound on a loop where the loop
813 * iterator "iv" is incremented (or decremented) by one, check if wrapping
816 * Under the given assumptions, wrapping is only possible if "cond" allows
817 * for the last value before wrapping, i.e., 2^width - 1 in case of an
818 * increasing iterator and 0 in case of a decreasing iterator.
820 static int can_wrap(__isl_keep isl_set
*cond
, __isl_keep pet_expr
*iv
,
821 __isl_keep isl_val
*inc
)
828 test
= isl_set_copy(cond
);
830 ctx
= isl_set_get_ctx(test
);
831 if (isl_val_is_neg(inc
))
832 limit
= isl_val_zero(ctx
);
834 limit
= isl_val_int_from_ui(ctx
, pet_expr_get_type_size(iv
));
835 limit
= isl_val_2exp(limit
);
836 limit
= isl_val_sub_ui(limit
, 1);
839 test
= isl_set_fix_val(cond
, isl_dim_set
, 0, limit
);
840 cw
= !isl_set_is_empty(test
);
846 /* Given a one-dimensional space, construct the following affine expression
849 * { [v] -> [v mod 2^width] }
851 * where width is the number of bits used to represent the values
852 * of the unsigned variable "iv".
854 static __isl_give isl_aff
*compute_wrapping(__isl_take isl_space
*dim
,
855 __isl_keep pet_expr
*iv
)
861 ctx
= isl_space_get_ctx(dim
);
862 mod
= isl_val_int_from_ui(ctx
, pet_expr_get_type_size(iv
));
863 mod
= isl_val_2exp(mod
);
865 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
866 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, 0, 1);
867 aff
= isl_aff_mod_val(aff
, mod
);
872 /* Project out the parameter "id" from "set".
874 static __isl_give isl_set
*set_project_out_by_id(__isl_take isl_set
*set
,
875 __isl_keep isl_id
*id
)
879 pos
= isl_set_find_dim_by_id(set
, isl_dim_param
, id
);
881 set
= isl_set_project_out(set
, isl_dim_param
, pos
, 1);
886 /* Compute the set of parameters for which "set1" is a subset of "set2".
888 * set1 is a subset of set2 if
890 * forall i in set1 : i in set2
894 * not exists i in set1 and i not in set2
898 * not exists i in set1 \ set2
900 static __isl_give isl_set
*enforce_subset(__isl_take isl_set
*set1
,
901 __isl_take isl_set
*set2
)
903 return isl_set_complement(isl_set_params(isl_set_subtract(set1
, set2
)));
906 /* Compute the set of parameter values for which "cond" holds
907 * on the next iteration for each element of "dom".
909 * We first construct mapping { [i] -> [i + inc] }, apply that to "dom"
910 * and then compute the set of parameters for which the result is a subset
913 static __isl_give isl_set
*valid_on_next(__isl_take isl_set
*cond
,
914 __isl_take isl_set
*dom
, __isl_take isl_val
*inc
)
920 space
= isl_set_get_space(dom
);
921 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
922 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, 0, 1);
923 aff
= isl_aff_add_constant_val(aff
, inc
);
924 next
= isl_map_from_basic_map(isl_basic_map_from_aff(aff
));
926 dom
= isl_set_apply(dom
, next
);
928 return enforce_subset(dom
, cond
);
931 /* Extract the for loop "tree" as a while loop within the context "pc".
933 * That is, the for loop has the form
935 * for (iv = init; cond; iv += inc)
946 * except that the skips resulting from any continue statements
947 * in body do not apply to the increment, but are replaced by the skips
948 * resulting from break statements.
950 * If the loop iterator is declared in the for loop, then it is killed before
951 * and after the loop.
953 static struct pet_scop
*scop_from_non_affine_for(__isl_keep pet_tree
*tree
,
954 __isl_take pet_context
*pc
, struct pet_state
*state
)
957 int test_nr
, stmt_nr
;
959 pet_expr
*expr_iv
, *init
, *inc
;
960 struct pet_scop
*scop_init
, *scop
;
962 struct pet_array
*array
;
963 struct pet_scop
*scop_kill
;
965 iv
= pet_expr_access_get_id(tree
->u
.l
.iv
);
966 pc
= pet_context_mark_assigned(pc
, iv
);
968 declared
= tree
->u
.l
.declared
;
970 expr_iv
= pet_expr_copy(tree
->u
.l
.iv
);
971 type_size
= pet_expr_get_type_size(expr_iv
);
972 init
= pet_expr_copy(tree
->u
.l
.init
);
973 init
= pet_expr_new_binary(type_size
, pet_op_assign
, expr_iv
, init
);
974 scop_init
= scop_from_expr(init
, NULL
, state
->n_stmt
++,
975 pet_tree_get_loc(tree
), pc
);
976 scop_init
= pet_scop_prefix(scop_init
, declared
);
978 test_nr
= state
->n_test
++;
979 stmt_nr
= state
->n_stmt
++;
981 expr_iv
= pet_expr_copy(tree
->u
.l
.iv
);
982 type_size
= pet_expr_get_type_size(expr_iv
);
983 inc
= pet_expr_copy(tree
->u
.l
.inc
);
984 inc
= pet_expr_new_binary(type_size
, pet_op_add_assign
, expr_iv
, inc
);
986 scop
= scop_from_non_affine_while(pet_expr_copy(tree
->u
.l
.cond
),
987 test_nr
, stmt_nr
, pet_tree_get_loc(tree
),
988 tree
->u
.l
.body
, inc
, pet_context_copy(pc
), state
);
990 scop
= pet_scop_prefix(scop
, declared
+ 1);
991 scop
= pet_scop_add_seq(state
->ctx
, scop_init
, scop
);
994 pet_context_free(pc
);
998 array
= extract_array(tree
->u
.l
.iv
, pc
, state
);
1000 array
->declared
= 1;
1001 scop_kill
= kill(pet_tree_get_loc(tree
), array
, pc
, state
);
1002 scop_kill
= pet_scop_prefix(scop_kill
, 0);
1003 scop
= pet_scop_add_seq(state
->ctx
, scop_kill
, scop
);
1004 scop_kill
= kill(pet_tree_get_loc(tree
), array
, pc
, state
);
1005 scop_kill
= pet_scop_add_array(scop_kill
, array
);
1006 scop_kill
= pet_scop_prefix(scop_kill
, 3);
1007 scop
= pet_scop_add_seq(state
->ctx
, scop
, scop_kill
);
1009 pet_context_free(pc
);
1013 /* Given an access expression "expr", is the variable accessed by
1014 * "expr" assigned anywhere inside "tree"?
1016 static int is_assigned(__isl_keep pet_expr
*expr
, __isl_keep pet_tree
*tree
)
1021 id
= pet_expr_access_get_id(expr
);
1022 assigned
= pet_tree_writes(tree
, id
);
1028 /* Are all nested access parameters in "pa" allowed given "tree".
1029 * In particular, is none of them written by anywhere inside "tree".
1031 * If "tree" has any continue nodes in the current loop level,
1032 * then no nested access parameters are allowed.
1033 * In particular, if there is any nested access in a guard
1034 * for a piece of code containing a "continue", then we want to introduce
1035 * a separate statement for evaluating this guard so that we can express
1036 * that the result is false for all previous iterations.
1038 static int is_nested_allowed(__isl_keep isl_pw_aff
*pa
,
1039 __isl_keep pet_tree
*tree
)
1046 if (!pet_nested_any_in_pw_aff(pa
))
1049 if (pet_tree_has_continue(tree
))
1052 nparam
= isl_pw_aff_dim(pa
, isl_dim_param
);
1053 for (i
= 0; i
< nparam
; ++i
) {
1054 isl_id
*id
= isl_pw_aff_get_dim_id(pa
, isl_dim_param
, i
);
1058 if (!pet_nested_in_id(id
)) {
1063 expr
= pet_nested_extract_expr(id
);
1064 allowed
= pet_expr_get_type(expr
) == pet_expr_access
&&
1065 !is_assigned(expr
, tree
);
1067 pet_expr_free(expr
);
1077 /* Construct a pet_scop for a for tree with static affine initialization
1078 * and constant increment within the context "pc".
1080 * The condition is allowed to contain nested accesses, provided
1081 * they are not being written to inside the body of the loop.
1082 * Otherwise, or if the condition is otherwise non-affine, the for loop is
1083 * essentially treated as a while loop, with iteration domain
1084 * { [i] : i >= init }.
1086 * We extract a pet_scop for the body and then embed it in a loop with
1087 * iteration domain and schedule
1089 * { [i] : i >= init and condition' }
1094 * { [i] : i <= init and condition' }
1097 * Where condition' is equal to condition if the latter is
1098 * a simple upper [lower] bound and a condition that is extended
1099 * to apply to all previous iterations otherwise.
1101 * If the condition is non-affine, then we drop the condition from the
1102 * iteration domain and instead create a separate statement
1103 * for evaluating the condition. The body is then filtered to depend
1104 * on the result of the condition evaluating to true on all iterations
1105 * up to the current iteration, while the evaluation the condition itself
1106 * is filtered to depend on the result of the condition evaluating to true
1107 * on all previous iterations.
1108 * The context of the scop representing the body is dropped
1109 * because we don't know how many times the body will be executed,
1112 * If the stride of the loop is not 1, then "i >= init" is replaced by
1114 * (exists a: i = init + stride * a and a >= 0)
1116 * If the loop iterator i is unsigned, then wrapping may occur.
1117 * We therefore use a virtual iterator instead that does not wrap.
1118 * However, the condition in the code applies
1119 * to the wrapped value, so we need to change condition(i)
1120 * into condition([i % 2^width]). Similarly, we replace all accesses
1121 * to the original iterator by the wrapping of the virtual iterator.
1122 * Note that there may be no need to perform this final wrapping
1123 * if the loop condition (after wrapping) satisfies certain conditions.
1124 * However, the is_simple_bound condition is not enough since it doesn't
1125 * check if there even is an upper bound.
1127 * Wrapping on unsigned iterators can be avoided entirely if
1128 * loop condition is simple, the loop iterator is incremented
1129 * [decremented] by one and the last value before wrapping cannot
1130 * possibly satisfy the loop condition.
1132 * Valid parameters for a for loop are those for which the initial
1133 * value itself, the increment on each domain iteration and
1134 * the condition on both the initial value and
1135 * the result of incrementing the iterator for each iteration of the domain
1137 * If the loop condition is non-affine, then we only consider validity
1138 * of the initial value.
1140 * If the body contains any break, then we keep track of it in "skip"
1141 * (if the skip condition is affine) or it is handled in scop_add_break
1142 * (if the skip condition is not affine).
1143 * Note that the affine break condition needs to be considered with
1144 * respect to previous iterations in the virtual domain (if any).
1146 static struct pet_scop
*scop_from_affine_for(__isl_keep pet_tree
*tree
,
1147 __isl_take isl_pw_aff
*init_val
, __isl_take isl_pw_aff
*pa_inc
,
1148 __isl_take isl_val
*inc
, __isl_take pet_context
*pc
,
1149 struct pet_state
*state
)
1151 isl_local_space
*ls
;
1154 isl_set
*cond
= NULL
;
1155 isl_set
*skip
= NULL
;
1156 isl_id
*id
, *id_test
= NULL
, *id_break_test
;
1157 struct pet_scop
*scop
, *scop_cond
= NULL
;
1163 int has_affine_break
;
1165 isl_map
*rev_wrap
= NULL
;
1166 isl_aff
*wrap
= NULL
;
1168 isl_set
*valid_init
;
1169 isl_set
*valid_cond
;
1170 isl_set
*valid_cond_init
;
1171 isl_set
*valid_cond_next
;
1173 pet_expr
*cond_expr
;
1174 pet_context
*pc_nested
;
1176 id
= pet_expr_access_get_id(tree
->u
.l
.iv
);
1178 cond_expr
= pet_expr_copy(tree
->u
.l
.cond
);
1179 cond_expr
= pet_expr_plug_in_args(cond_expr
, pc
);
1180 pc_nested
= pet_context_copy(pc
);
1181 pc_nested
= pet_context_set_allow_nested(pc_nested
, 1);
1182 pa
= pet_expr_extract_affine_condition(cond_expr
, pc_nested
);
1183 pet_context_free(pc_nested
);
1184 pet_expr_free(cond_expr
);
1186 valid_inc
= isl_pw_aff_domain(pa_inc
);
1188 is_unsigned
= pet_expr_get_type_size(tree
->u
.l
.iv
) > 0;
1190 is_non_affine
= isl_pw_aff_involves_nan(pa
) ||
1191 !is_nested_allowed(pa
, tree
->u
.l
.body
);
1193 pa
= isl_pw_aff_free(pa
);
1195 valid_cond
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
1196 cond
= isl_pw_aff_non_zero_set(pa
);
1197 if (is_non_affine
) {
1198 isl_multi_pw_aff
*test_index
;
1199 test_index
= pet_create_test_index(state
->ctx
, state
->n_test
++);
1200 scop_cond
= scop_from_non_affine_condition(
1201 pet_expr_copy(tree
->u
.l
.cond
), state
->n_stmt
++,
1202 isl_multi_pw_aff_copy(test_index
),
1203 pet_tree_get_loc(tree
), pc
);
1204 id_test
= isl_multi_pw_aff_get_tuple_id(test_index
,
1206 scop_cond
= pet_scop_add_boolean_array(scop_cond
, test_index
,
1208 scop_cond
= pet_scop_prefix(scop_cond
, 0);
1209 cond
= isl_set_universe(isl_space_set_alloc(state
->ctx
, 0, 0));
1212 cond
= embed(cond
, isl_id_copy(id
));
1213 valid_cond
= isl_set_coalesce(valid_cond
);
1214 valid_cond
= embed(valid_cond
, isl_id_copy(id
));
1215 valid_inc
= embed(valid_inc
, isl_id_copy(id
));
1216 is_one
= isl_val_is_one(inc
) || isl_val_is_negone(inc
);
1217 is_virtual
= is_unsigned
&&
1218 (!is_one
|| can_wrap(cond
, tree
->u
.l
.iv
, inc
));
1220 valid_cond_init
= enforce_subset(
1221 isl_map_range(isl_map_from_pw_aff(isl_pw_aff_copy(init_val
))),
1222 isl_set_copy(valid_cond
));
1223 if (is_one
&& !is_virtual
) {
1224 isl_pw_aff_free(init_val
);
1225 pa
= pet_expr_extract_comparison(
1226 isl_val_is_pos(inc
) ? pet_op_ge
: pet_op_le
,
1227 tree
->u
.l
.iv
, tree
->u
.l
.init
, pc
);
1228 valid_init
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
1229 valid_init
= set_project_out_by_id(valid_init
, id
);
1230 domain
= isl_pw_aff_non_zero_set(pa
);
1232 valid_init
= isl_pw_aff_domain(isl_pw_aff_copy(init_val
));
1233 domain
= strided_domain(isl_id_copy(id
), init_val
,
1237 domain
= embed(domain
, isl_id_copy(id
));
1239 wrap
= compute_wrapping(isl_set_get_space(cond
), tree
->u
.l
.iv
);
1240 rev_wrap
= isl_map_from_aff(isl_aff_copy(wrap
));
1241 rev_wrap
= isl_map_reverse(rev_wrap
);
1242 cond
= isl_set_apply(cond
, isl_map_copy(rev_wrap
));
1243 valid_cond
= isl_set_apply(valid_cond
, isl_map_copy(rev_wrap
));
1244 valid_inc
= isl_set_apply(valid_inc
, isl_map_copy(rev_wrap
));
1246 is_simple
= is_simple_bound(cond
, inc
);
1248 cond
= isl_set_gist(cond
, isl_set_copy(domain
));
1249 is_simple
= is_simple_bound(cond
, inc
);
1252 cond
= valid_for_each_iteration(cond
,
1253 isl_set_copy(domain
), isl_val_copy(inc
));
1254 domain
= isl_set_intersect(domain
, cond
);
1255 domain
= isl_set_set_dim_id(domain
, isl_dim_set
, 0, isl_id_copy(id
));
1256 ls
= isl_local_space_from_space(isl_set_get_space(domain
));
1257 sched
= isl_aff_var_on_domain(ls
, isl_dim_set
, 0);
1258 if (isl_val_is_neg(inc
))
1259 sched
= isl_aff_neg(sched
);
1261 valid_cond_next
= valid_on_next(valid_cond
, isl_set_copy(domain
),
1263 valid_inc
= enforce_subset(isl_set_copy(domain
), valid_inc
);
1266 wrap
= identity_aff(domain
);
1268 scop
= scop_from_tree(tree
->u
.l
.body
, pc
, state
);
1270 scop_cond
= pet_scop_embed(scop_cond
, isl_set_copy(domain
),
1271 isl_aff_copy(sched
), isl_aff_copy(wrap
), isl_id_copy(id
));
1272 has_affine_break
= scop
&&
1273 pet_scop_has_affine_skip(scop
, pet_skip_later
);
1274 if (has_affine_break
)
1275 skip
= pet_scop_get_affine_skip_domain(scop
, pet_skip_later
);
1276 has_var_break
= scop
&& pet_scop_has_var_skip(scop
, pet_skip_later
);
1278 id_break_test
= pet_scop_get_skip_id(scop
, pet_skip_later
);
1279 if (is_non_affine
) {
1280 scop
= pet_scop_reset_context(scop
);
1281 scop
= pet_scop_prefix(scop
, 1);
1283 scop
= pet_scop_embed(scop
, isl_set_copy(domain
), sched
, wrap
, id
);
1284 scop
= pet_scop_resolve_nested(scop
);
1285 if (has_affine_break
) {
1286 domain
= apply_affine_break(domain
, skip
, isl_val_sgn(inc
),
1287 is_virtual
, rev_wrap
);
1288 scop
= pet_scop_intersect_domain_prefix(scop
,
1289 isl_set_copy(domain
));
1291 isl_map_free(rev_wrap
);
1293 scop
= scop_add_break(scop
, id_break_test
, isl_set_copy(domain
),
1295 if (is_non_affine
) {
1296 scop
= scop_add_while(scop_cond
, scop
, id_test
, domain
,
1298 isl_set_free(valid_inc
);
1300 scop
= pet_scop_restrict_context(scop
, valid_inc
);
1301 scop
= pet_scop_restrict_context(scop
, valid_cond_next
);
1302 scop
= pet_scop_restrict_context(scop
, valid_cond_init
);
1303 isl_set_free(domain
);
1308 scop
= pet_scop_restrict_context(scop
, isl_set_params(valid_init
));
1310 pet_context_free(pc
);
1314 /* Construct a pet_scop for a for statement within the context of "pc".
1316 * We update the context to reflect the writes to the loop variable and
1317 * the writes inside the body.
1319 * Then we check if the initialization of the for loop
1320 * is a static affine value and the increment is a constant.
1321 * If so, we construct the pet_scop using scop_from_affine_for.
1322 * Otherwise, we treat the for loop as a while loop
1323 * in scop_from_non_affine_for.
1325 static struct pet_scop
*scop_from_for(__isl_keep pet_tree
*tree
,
1326 __isl_keep pet_context
*pc
, struct pet_state
*state
)
1330 isl_pw_aff
*pa_inc
, *init_val
;
1331 pet_context
*pc_init_val
;
1336 iv
= pet_expr_access_get_id(tree
->u
.l
.iv
);
1337 pc
= pet_context_copy(pc
);
1338 pc
= pet_context_clear_value(pc
, iv
);
1339 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.l
.body
);
1341 pc_init_val
= pet_context_copy(pc
);
1342 pc_init_val
= pet_context_mark_unknown(pc_init_val
, isl_id_copy(iv
));
1343 init_val
= pet_expr_extract_affine(tree
->u
.l
.init
, pc_init_val
);
1344 pet_context_free(pc_init_val
);
1345 pa_inc
= pet_expr_extract_affine(tree
->u
.l
.inc
, pc
);
1346 inc
= pet_extract_cst(pa_inc
);
1347 if (!pa_inc
|| !init_val
|| !inc
)
1349 if (!isl_pw_aff_involves_nan(pa_inc
) &&
1350 !isl_pw_aff_involves_nan(init_val
) && !isl_val_is_nan(inc
))
1351 return scop_from_affine_for(tree
, init_val
, pa_inc
, inc
,
1354 isl_pw_aff_free(pa_inc
);
1355 isl_pw_aff_free(init_val
);
1357 return scop_from_non_affine_for(tree
, pc
, state
);
1359 isl_pw_aff_free(pa_inc
);
1360 isl_pw_aff_free(init_val
);
1362 pet_context_free(pc
);
1366 /* Check whether "expr" is an affine constraint within the context "pc".
1368 static int is_affine_condition(__isl_keep pet_expr
*expr
,
1369 __isl_keep pet_context
*pc
)
1374 pa
= pet_expr_extract_affine_condition(expr
, pc
);
1377 is_affine
= !isl_pw_aff_involves_nan(pa
);
1378 isl_pw_aff_free(pa
);
1383 /* Check if the given if statement is a conditional assignement
1384 * with a non-affine condition.
1386 * In particular we check if "stmt" is of the form
1393 * where the condition is non-affine and a is some array or scalar access.
1395 static int is_conditional_assignment(__isl_keep pet_tree
*tree
,
1396 __isl_keep pet_context
*pc
)
1400 pet_expr
*expr1
, *expr2
;
1402 ctx
= pet_tree_get_ctx(tree
);
1403 if (!pet_options_get_detect_conditional_assignment(ctx
))
1405 if (tree
->type
!= pet_tree_if_else
)
1407 if (tree
->u
.i
.then_body
->type
!= pet_tree_expr
)
1409 if (tree
->u
.i
.else_body
->type
!= pet_tree_expr
)
1411 expr1
= tree
->u
.i
.then_body
->u
.e
.expr
;
1412 expr2
= tree
->u
.i
.else_body
->u
.e
.expr
;
1413 if (pet_expr_get_type(expr1
) != pet_expr_op
)
1415 if (pet_expr_get_type(expr2
) != pet_expr_op
)
1417 if (pet_expr_op_get_type(expr1
) != pet_op_assign
)
1419 if (pet_expr_op_get_type(expr2
) != pet_op_assign
)
1421 expr1
= pet_expr_get_arg(expr1
, 0);
1422 expr2
= pet_expr_get_arg(expr2
, 0);
1423 equal
= pet_expr_is_equal(expr1
, expr2
);
1424 pet_expr_free(expr1
);
1425 pet_expr_free(expr2
);
1426 if (equal
< 0 || !equal
)
1428 if (is_affine_condition(tree
->u
.i
.cond
, pc
))
1434 /* Given that "tree" is of the form
1441 * where a is some array or scalar access, construct a pet_scop
1442 * corresponding to this conditional assignment within the context "pc".
1444 * The constructed pet_scop then corresponds to the expression
1446 * a = condition ? f(...) : g(...)
1448 * All access relations in f(...) are intersected with condition
1449 * while all access relation in g(...) are intersected with the complement.
1451 static struct pet_scop
*scop_from_conditional_assignment(
1452 __isl_keep pet_tree
*tree
, __isl_take pet_context
*pc
,
1453 struct pet_state
*state
)
1457 isl_set
*cond
, *comp
;
1458 isl_multi_pw_aff
*index
;
1459 pet_expr
*expr1
, *expr2
;
1460 pet_expr
*pe_cond
, *pe_then
, *pe_else
, *pe
, *pe_write
;
1461 pet_context
*pc_nested
;
1462 struct pet_scop
*scop
;
1464 pe_cond
= pet_expr_copy(tree
->u
.i
.cond
);
1465 pe_cond
= pet_expr_plug_in_args(pe_cond
, pc
);
1466 pc_nested
= pet_context_copy(pc
);
1467 pc_nested
= pet_context_set_allow_nested(pc_nested
, 1);
1468 pa
= pet_expr_extract_affine_condition(pe_cond
, pc_nested
);
1469 pet_context_free(pc_nested
);
1470 pet_expr_free(pe_cond
);
1471 cond
= isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa
));
1472 comp
= isl_pw_aff_zero_set(isl_pw_aff_copy(pa
));
1473 index
= isl_multi_pw_aff_from_pw_aff(pa
);
1475 expr1
= tree
->u
.i
.then_body
->u
.e
.expr
;
1476 expr2
= tree
->u
.i
.else_body
->u
.e
.expr
;
1478 pe_cond
= pet_expr_from_index(index
);
1480 pe_then
= pet_expr_get_arg(expr1
, 1);
1481 pe_then
= pet_expr_restrict(pe_then
, cond
);
1482 pe_else
= pet_expr_get_arg(expr2
, 1);
1483 pe_else
= pet_expr_restrict(pe_else
, comp
);
1484 pe_write
= pet_expr_get_arg(expr1
, 0);
1486 pe
= pet_expr_new_ternary(pe_cond
, pe_then
, pe_else
);
1487 type_size
= pet_expr_get_type_size(pe_write
);
1488 pe
= pet_expr_new_binary(type_size
, pet_op_assign
, pe_write
, pe
);
1490 scop
= scop_from_expr(pe
, NULL
, state
->n_stmt
++,
1491 pet_tree_get_loc(tree
), pc
);
1493 pet_context_free(pc
);
1498 /* Construct a pet_scop for a non-affine if statement within the context "pc".
1500 * We create a separate statement that writes the result
1501 * of the non-affine condition to a virtual scalar.
1502 * A constraint requiring the value of this virtual scalar to be one
1503 * is added to the iteration domains of the then branch.
1504 * Similarly, a constraint requiring the value of this virtual scalar
1505 * to be zero is added to the iteration domains of the else branch, if any.
1506 * We adjust the schedules to ensure that the virtual scalar is written
1507 * before it is read.
1509 * If there are any breaks or continues in the then and/or else
1510 * branches, then we may have to compute a new skip condition.
1511 * This is handled using a pet_skip_info object.
1512 * On initialization, the object checks if skip conditions need
1513 * to be computed. If so, it does so in pet_skip_info_if_extract_index and
1514 * adds them in pet_skip_info_if_add.
1516 static struct pet_scop
*scop_from_non_affine_if(__isl_keep pet_tree
*tree
,
1517 struct pet_scop
*scop_then
, struct pet_scop
*scop_else
, int stmt_id
,
1518 __isl_take pet_context
*pc
, struct pet_state
*state
)
1521 int save_n_stmt
= state
->n_stmt
;
1522 isl_multi_pw_aff
*test_index
;
1523 struct pet_skip_info skip
;
1524 struct pet_scop
*scop
;
1526 has_else
= tree
->type
== pet_tree_if_else
;
1528 test_index
= pet_create_test_index(state
->ctx
, state
->n_test
++);
1529 state
->n_stmt
= stmt_id
;
1530 scop
= scop_from_non_affine_condition(pet_expr_copy(tree
->u
.i
.cond
),
1531 state
->n_stmt
++, isl_multi_pw_aff_copy(test_index
),
1532 pet_tree_get_loc(tree
), pc
);
1533 state
->n_stmt
= save_n_stmt
;
1534 scop
= pet_scop_add_boolean_array(scop
,
1535 isl_multi_pw_aff_copy(test_index
), state
->int_size
);
1537 pet_skip_info_if_init(&skip
, state
->ctx
, scop_then
, scop_else
,
1539 pet_skip_info_if_extract_index(&skip
, test_index
, state
);
1541 scop
= pet_scop_prefix(scop
, 0);
1542 scop_then
= pet_scop_prefix(scop_then
, 1);
1543 scop_then
= pet_scop_filter(scop_then
,
1544 isl_multi_pw_aff_copy(test_index
), 1);
1546 scop_else
= pet_scop_prefix(scop_else
, 1);
1547 scop_else
= pet_scop_filter(scop_else
, test_index
, 0);
1548 scop_then
= pet_scop_add_par(state
->ctx
, scop_then
, scop_else
);
1550 isl_multi_pw_aff_free(test_index
);
1552 scop
= pet_scop_add_seq(state
->ctx
, scop
, scop_then
);
1554 scop
= pet_skip_info_if_add(&skip
, scop
, 2);
1556 pet_context_free(pc
);
1560 /* Construct a pet_scop for an affine if statement within the context "pc".
1562 * The condition is added to the iteration domains of the then branch,
1563 * while the opposite of the condition in added to the iteration domains
1564 * of the else branch, if any.
1566 * If there are any breaks or continues in the then and/or else
1567 * branches, then we may have to compute a new skip condition.
1568 * This is handled using a pet_skip_info_if object.
1569 * On initialization, the object checks if skip conditions need
1570 * to be computed. If so, it does so in pet_skip_info_if_extract_cond and
1571 * adds them in pet_skip_info_if_add.
1573 static struct pet_scop
*scop_from_affine_if(__isl_keep pet_tree
*tree
,
1574 __isl_take isl_pw_aff
*cond
,
1575 struct pet_scop
*scop_then
, struct pet_scop
*scop_else
,
1576 struct pet_state
*state
)
1582 struct pet_skip_info skip
;
1583 struct pet_scop
*scop
;
1585 ctx
= pet_tree_get_ctx(tree
);
1587 has_else
= tree
->type
== pet_tree_if_else
;
1589 pet_skip_info_if_init(&skip
, ctx
, scop_then
, scop_else
, has_else
, 1);
1590 pet_skip_info_if_extract_cond(&skip
, cond
, state
);
1592 valid
= isl_pw_aff_domain(isl_pw_aff_copy(cond
));
1593 set
= isl_pw_aff_non_zero_set(cond
);
1594 scop
= pet_scop_restrict(scop_then
, isl_set_params(isl_set_copy(set
)));
1597 set
= isl_set_subtract(isl_set_copy(valid
), set
);
1598 scop_else
= pet_scop_restrict(scop_else
, isl_set_params(set
));
1599 scop
= pet_scop_add_par(ctx
, scop
, scop_else
);
1602 scop
= pet_scop_resolve_nested(scop
);
1603 scop
= pet_scop_restrict_context(scop
, isl_set_params(valid
));
1605 if (pet_skip_info_has_skip(&skip
))
1606 scop
= pet_scop_prefix(scop
, 0);
1607 scop
= pet_skip_info_if_add(&skip
, scop
, 1);
1612 /* Construct a pet_scop for an if statement within the context "pc".
1614 * If the condition fits the pattern of a conditional assignment,
1615 * then it is handled by scop_from_conditional_assignment.
1617 * Otherwise, we check if the condition is affine.
1618 * If so, we construct the scop in scop_from_affine_if.
1619 * Otherwise, we construct the scop in scop_from_non_affine_if.
1621 * We allow the condition to be dynamic, i.e., to refer to
1622 * scalars or array elements that may be written to outside
1623 * of the given if statement. These nested accesses are then represented
1624 * as output dimensions in the wrapping iteration domain.
1625 * If it is also written _inside_ the then or else branch, then
1626 * we treat the condition as non-affine.
1627 * As explained in extract_non_affine_if, this will introduce
1628 * an extra statement.
1629 * For aesthetic reasons, we want this statement to have a statement
1630 * number that is lower than those of the then and else branches.
1631 * In order to evaluate if we will need such a statement, however, we
1632 * first construct scops for the then and else branches.
1633 * We therefore reserve a statement number if we might have to
1634 * introduce such an extra statement.
1636 static struct pet_scop
*scop_from_if(__isl_keep pet_tree
*tree
,
1637 __isl_keep pet_context
*pc
, struct pet_state
*state
)
1642 pet_expr
*cond_expr
;
1643 struct pet_scop
*scop_then
, *scop_else
= NULL
;
1644 pet_context
*pc_nested
;
1649 has_else
= tree
->type
== pet_tree_if_else
;
1651 pc
= pet_context_copy(pc
);
1652 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.i
.then_body
);
1654 pc
= pet_context_clear_writes_in_tree(pc
, tree
->u
.i
.else_body
);
1656 if (is_conditional_assignment(tree
, pc
))
1657 return scop_from_conditional_assignment(tree
, pc
, state
);
1659 cond_expr
= pet_expr_copy(tree
->u
.i
.cond
);
1660 cond_expr
= pet_expr_plug_in_args(cond_expr
, pc
);
1661 pc_nested
= pet_context_copy(pc
);
1662 pc_nested
= pet_context_set_allow_nested(pc_nested
, 1);
1663 cond
= pet_expr_extract_affine_condition(cond_expr
, pc_nested
);
1664 pet_context_free(pc_nested
);
1665 pet_expr_free(cond_expr
);
1668 pet_context_free(pc
);
1672 if (isl_pw_aff_involves_nan(cond
) || pet_nested_any_in_pw_aff(cond
))
1673 stmt_id
= state
->n_stmt
++;
1675 scop_then
= scop_from_tree(tree
->u
.i
.then_body
, pc
, state
);
1677 scop_else
= scop_from_tree(tree
->u
.i
.else_body
, pc
, state
);
1679 if (isl_pw_aff_involves_nan(cond
)) {
1680 isl_pw_aff_free(cond
);
1681 return scop_from_non_affine_if(tree
, scop_then
, scop_else
,
1682 stmt_id
, pc
, state
);
1685 if ((!is_nested_allowed(cond
, tree
->u
.i
.then_body
) ||
1686 (has_else
&& !is_nested_allowed(cond
, tree
->u
.i
.else_body
)))) {
1687 isl_pw_aff_free(cond
);
1688 return scop_from_non_affine_if(tree
, scop_then
, scop_else
,
1689 stmt_id
, pc
, state
);
1692 pet_context_free(pc
);
1693 return scop_from_affine_if(tree
, cond
, scop_then
, scop_else
, state
);
1696 /* Return a one-dimensional multi piecewise affine expression that is equal
1697 * to the constant 1 and is defined over a zero-dimensional domain.
1699 static __isl_give isl_multi_pw_aff
*one_mpa(isl_ctx
*ctx
)
1702 isl_local_space
*ls
;
1705 space
= isl_space_set_alloc(ctx
, 0, 0);
1706 ls
= isl_local_space_from_space(space
);
1707 aff
= isl_aff_zero_on_domain(ls
);
1708 aff
= isl_aff_set_constant_si(aff
, 1);
1710 return isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff
));
1713 /* Construct a pet_scop for a continue statement.
1715 * We simply create an empty scop with a universal pet_skip_now
1716 * skip condition. This skip condition will then be taken into
1717 * account by the enclosing loop construct, possibly after
1718 * being incorporated into outer skip conditions.
1720 static struct pet_scop
*scop_from_continue(__isl_keep pet_tree
*tree
)
1722 struct pet_scop
*scop
;
1725 ctx
= pet_tree_get_ctx(tree
);
1726 scop
= pet_scop_empty(ctx
);
1730 scop
= pet_scop_set_skip(scop
, pet_skip_now
, one_mpa(ctx
));
1735 /* Construct a pet_scop for a break statement.
1737 * We simply create an empty scop with both a universal pet_skip_now
1738 * skip condition and a universal pet_skip_later skip condition.
1739 * These skip conditions will then be taken into
1740 * account by the enclosing loop construct, possibly after
1741 * being incorporated into outer skip conditions.
1743 static struct pet_scop
*scop_from_break(__isl_keep pet_tree
*tree
)
1745 struct pet_scop
*scop
;
1747 isl_multi_pw_aff
*skip
;
1749 ctx
= pet_tree_get_ctx(tree
);
1750 scop
= pet_scop_empty(ctx
);
1754 skip
= one_mpa(ctx
);
1755 scop
= pet_scop_set_skip(scop
, pet_skip_now
,
1756 isl_multi_pw_aff_copy(skip
));
1757 scop
= pet_scop_set_skip(scop
, pet_skip_later
, skip
);
1762 /* Extract a clone of the kill statement in "scop".
1763 * "scop" is expected to have been created from a DeclStmt
1764 * and should have the kill as its first statement.
1766 static struct pet_scop
*extract_kill(isl_ctx
*ctx
, struct pet_scop
*scop
,
1767 struct pet_state
*state
)
1770 struct pet_stmt
*stmt
;
1771 isl_multi_pw_aff
*index
;
1777 if (scop
->n_stmt
< 1)
1778 isl_die(ctx
, isl_error_internal
,
1779 "expecting at least one statement", return NULL
);
1780 stmt
= scop
->stmts
[0];
1781 if (!pet_stmt_is_kill(stmt
))
1782 isl_die(ctx
, isl_error_internal
,
1783 "expecting kill statement", return NULL
);
1785 arg
= pet_expr_get_arg(stmt
->body
, 0);
1786 index
= pet_expr_access_get_index(arg
);
1787 access
= pet_expr_access_get_access(arg
);
1789 index
= isl_multi_pw_aff_reset_tuple_id(index
, isl_dim_in
);
1790 access
= isl_map_reset_tuple_id(access
, isl_dim_in
);
1791 kill
= pet_expr_kill_from_access_and_index(access
, index
);
1792 stmt
= pet_stmt_from_pet_expr(pet_loc_copy(stmt
->loc
),
1793 NULL
, state
->n_stmt
++, kill
);
1794 return pet_scop_from_pet_stmt(ctx
, stmt
);
1797 /* Mark all arrays in "scop" as being exposed.
1799 static struct pet_scop
*mark_exposed(struct pet_scop
*scop
)
1805 for (i
= 0; i
< scop
->n_array
; ++i
)
1806 scop
->arrays
[i
]->exposed
= 1;
1810 /* Try and construct a pet_scop corresponding to (part of)
1811 * a sequence of statements within the context "pc".
1813 * After extracting a statement, we update "pc"
1814 * based on the top-level assignments in the statement
1815 * so that we can exploit them in subsequent statements in the same block.
1817 * If there are any breaks or continues in the individual statements,
1818 * then we may have to compute a new skip condition.
1819 * This is handled using a pet_skip_info object.
1820 * On initialization, the object checks if skip conditions need
1821 * to be computed. If so, it does so in pet_skip_info_seq_extract and
1822 * adds them in pet_skip_info_seq_add.
1824 * If "block" is set, then we need to insert kill statements at
1825 * the end of the block for any array that has been declared by
1826 * one of the statements in the sequence. Each of these declarations
1827 * results in the construction of a kill statement at the place
1828 * of the declaration, so we simply collect duplicates of
1829 * those kill statements and append these duplicates to the constructed scop.
1831 * If "block" is not set, then any array declared by one of the statements
1832 * in the sequence is marked as being exposed.
1834 * If autodetect is set, then we allow the extraction of only a subrange
1835 * of the sequence of statements. However, if there is at least one statement
1836 * for which we could not construct a scop and the final range contains
1837 * either no statements or at least one kill, then we discard the entire
1840 static struct pet_scop
*scop_from_block(__isl_keep pet_tree
*tree
,
1841 __isl_keep pet_context
*pc
, struct pet_state
*state
)
1845 struct pet_scop
*scop
, *kills
;
1847 ctx
= pet_tree_get_ctx(tree
);
1849 pc
= pet_context_copy(pc
);
1850 scop
= pet_scop_empty(ctx
);
1851 kills
= pet_scop_empty(ctx
);
1852 for (i
= 0; i
< tree
->u
.b
.n
; ++i
) {
1853 struct pet_scop
*scop_i
;
1855 scop_i
= scop_from_tree(tree
->u
.b
.child
[i
], pc
, state
);
1856 pc
= scop_handle_writes(scop_i
, pc
);
1857 struct pet_skip_info skip
;
1858 pet_skip_info_seq_init(&skip
, ctx
, scop
, scop_i
);
1859 pet_skip_info_seq_extract(&skip
, state
);
1860 if (pet_skip_info_has_skip(&skip
))
1861 scop_i
= pet_scop_prefix(scop_i
, 0);
1862 if (scop_i
&& pet_tree_is_decl(tree
->u
.b
.child
[i
])) {
1863 if (tree
->u
.b
.block
) {
1864 struct pet_scop
*kill
;
1865 kill
= extract_kill(ctx
, scop_i
, state
);
1866 kills
= pet_scop_add_par(ctx
, kills
, kill
);
1868 scop_i
= mark_exposed(scop_i
);
1870 scop_i
= pet_scop_prefix(scop_i
, i
);
1871 scop
= pet_scop_add_seq(ctx
, scop
, scop_i
);
1873 scop
= pet_skip_info_seq_add(&skip
, scop
, i
);
1879 kills
= pet_scop_prefix(kills
, tree
->u
.b
.n
);
1880 scop
= pet_scop_add_seq(ctx
, scop
, kills
);
1882 pet_context_free(pc
);
1887 /* Construct a pet_scop that corresponds to the pet_tree "tree"
1888 * within the context "pc" by calling the appropriate function
1889 * based on the type of "tree".
1891 static struct pet_scop
*scop_from_tree(__isl_keep pet_tree
*tree
,
1892 __isl_keep pet_context
*pc
, struct pet_state
*state
)
1897 switch (tree
->type
) {
1898 case pet_tree_error
:
1900 case pet_tree_block
:
1901 return scop_from_block(tree
, pc
, state
);
1902 case pet_tree_break
:
1903 return scop_from_break(tree
);
1904 case pet_tree_continue
:
1905 return scop_from_continue(tree
);
1907 case pet_tree_decl_init
:
1908 return scop_from_decl(tree
, pc
, state
);
1910 return scop_from_expr(pet_expr_copy(tree
->u
.e
.expr
),
1911 isl_id_copy(tree
->label
),
1913 pet_tree_get_loc(tree
), pc
);
1915 case pet_tree_if_else
:
1916 return scop_from_if(tree
, pc
, state
);
1918 return scop_from_for(tree
, pc
, state
);
1919 case pet_tree_while
:
1920 return scop_from_while(tree
, pc
, state
);
1921 case pet_tree_infinite_loop
:
1922 return scop_from_infinite_for(tree
, pc
, state
);
1925 isl_die(tree
->ctx
, isl_error_internal
, "unhandled type",
1929 /* Construct a pet_scop that corresponds to the pet_tree "tree".
1930 * "int_size" is the number of bytes need to represent an integer.
1931 * "extract_array" is a callback that we can use to create a pet_array
1932 * that corresponds to the variable accessed by an expression.
1934 * Initialize the global state, construct a context and then
1935 * construct the pet_scop by recursively visiting the tree.
1937 struct pet_scop
*pet_scop_from_pet_tree(__isl_take pet_tree
*tree
, int int_size
,
1938 struct pet_array
*(*extract_array
)(__isl_keep pet_expr
*access
,
1939 __isl_keep pet_context
*pc
, void *user
), void *user
,
1940 __isl_keep pet_context
*pc
)
1942 struct pet_scop
*scop
;
1943 struct pet_state state
= { 0 };
1948 state
.ctx
= pet_tree_get_ctx(tree
);
1949 state
.int_size
= int_size
;
1950 state
.extract_array
= extract_array
;
1953 scop
= scop_from_tree(tree
, pc
, &state
);
1954 scop
= pet_scop_set_loc(scop
, pet_tree_get_loc(tree
));
1956 pet_tree_free(tree
);