2 /**-------------------------------------------------------------------**
4 **-------------------------------------------------------------------**
6 **-------------------------------------------------------------------**
7 ** First version: october 26th 2001 **
8 **-------------------------------------------------------------------**/
11 /******************************************************************************
12 * CLooG : the Chunky Loop Generator (experimental) *
13 ******************************************************************************
15 * Copyright (C) 2001-2005 Cedric Bastoul *
17 * This library is free software; you can redistribute it and/or *
18 * modify it under the terms of the GNU Lesser General Public *
19 * License as published by the Free Software Foundation; either *
20 * version 2.1 of the License, or (at your option) any later version. *
22 * This library is distributed in the hope that it will be useful, *
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of *
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU *
25 * Lesser General Public License for more details. *
27 * You should have received a copy of the GNU Lesser General Public *
28 * License along with this library; if not, write to the Free Software *
29 * Foundation, Inc., 51 Franklin Street, Fifth Floor, *
30 * Boston, MA 02110-1301 USA *
32 * CLooG, the Chunky Loop Generator *
33 * Written by Cedric Bastoul, Cedric.Bastoul@inria.fr *
35 ******************************************************************************/
36 /* CAUTION: the english used for comments is probably the worst you ever read,
37 * please feel free to correct and improve it !
42 # include "../include/cloog/cloog.h"
44 #define ALLOC(type) (type*)malloc(sizeof(type))
47 /******************************************************************************
48 * Memory leaks hunting *
49 ******************************************************************************/
53 * These functions and global variables are devoted to memory leaks hunting: we
54 * want to know at each moment how many CloogLoop structures had been allocated
55 * (cloog_loop_allocated) and how many had been freed (cloog_loop_freed).
56 * Each time a CloogLoog structure is allocated, a call to the function
57 * cloog_loop_leak_up() must be carried out, and respectively
58 * cloog_loop_leak_down() when a CloogLoop structure is freed. The special
59 * variable cloog_loop_max gives the maximal number of CloogLoop structures
60 * simultaneously alive (i.e. allocated and non-freed) in memory.
61 * - July 3rd->11th 2003: first version (memory leaks hunt and correction).
65 static void cloog_loop_leak_up(CloogState
*state
)
67 state
->loop_allocated
++;
68 if ((state
->loop_allocated
- state
->loop_freed
) > state
->loop_max
)
69 state
->loop_max
= state
->loop_allocated
- state
->loop_freed
;
73 static void cloog_loop_leak_down(CloogState
*state
)
79 /******************************************************************************
80 * Structure display function *
81 ******************************************************************************/
85 * cloog_loop_print_structure function:
86 * Displays a loop structure in a way that trends to be understandable without
87 * falling in a deep depression or, for the lucky ones, getting a headache...
88 * Written by Olivier Chorier, Luc Marchaud, Pierre Martin and Romain Tartiere.
89 * - April 24th 2005: Initial version.
90 * - May 21rd 2005: - New parameter `F' for destination file (ie stdout),
92 * - May 26th 2005: Memory leak hunt.
93 * - June 2nd 2005: (Ced) Integration and minor fixes.
94 * -June 22nd 2005: (Ced) Adaptation for GMP.
96 void cloog_loop_print_structure(FILE * file
, CloogLoop
* loop
, int level
)
100 { /* Go to the right level. */
101 for (i
=0; i
<level
; i
++)
102 fprintf(file
,"|\t") ;
104 fprintf(file
,"+-- CloogLoop\n") ;
110 { /* Go to the right level. */
111 for (i
=0; i
<level
; i
++)
112 fprintf(file
,"|\t") ;
114 fprintf(file
,"| CloogLoop\n") ;
120 for(j
=0; j
<=level
+1; j
++)
121 fprintf(file
,"|\t") ;
124 /* Print the domain. */
125 cloog_domain_print_structure(file
, loop
->domain
, level
+1, "CloogDomain");
127 /* Print the stride. */
128 for(j
=0; j
<=level
; j
++)
129 fprintf(file
,"|\t") ;
131 fprintf(file
, "Stride: ");
132 cloog_int_print(file
, loop
->stride
->stride
);
134 fprintf(file
, "Offset: ");
135 cloog_int_print(file
, loop
->stride
->offset
);
140 for(j
=0; j
<=level
+1; j
++)
141 fprintf(file
,"|\t") ;
144 /* Print the block. */
145 cloog_block_print_structure(file
,loop
->block
,level
+1) ;
148 for (i
=0; i
<=level
+1; i
++)
149 fprintf(file
,"|\t") ;
152 /* Print inner if any. */
154 cloog_loop_print_structure(file
,loop
->inner
,level
+1) ;
156 /* And let's go for the next one. */
159 /* One more time something that is here only for a better look. */
161 { /* Two blank lines if this is the end of the linked list. */
163 { for (i
=0; i
<=level
; i
++)
164 fprintf(file
,"|\t") ;
170 { /* A special blank line if the is a next loop. */
171 for (i
=0; i
<=level
; i
++)
172 fprintf(file
,"|\t") ;
173 fprintf(file
,"V\n") ;
180 * cloog_loop_print function:
181 * This function prints the content of a CloogLoop structure (start) into a
182 * file (file, possibly stdout).
183 * - June 2nd 2005: Now this very old function (probably as old as CLooG) is
184 * only a frontend to cloog_loop_print_structure, with a quite
185 * better human-readable representation.
187 void cloog_loop_print(FILE * file
, CloogLoop
* loop
)
188 { cloog_loop_print_structure(file
,loop
,0) ;
192 /******************************************************************************
193 * Memory deallocation function *
194 ******************************************************************************/
198 * cloog_loop_free function:
199 * This function frees the allocated memory for a CloogLoop structure (loop),
200 * and frees its inner loops and its next loops.
201 * - June 22nd 2005: Adaptation for GMP.
203 void cloog_loop_free(CloogLoop
* loop
)
206 while (loop
!= NULL
) {
207 cloog_loop_leak_down(loop
->state
);
210 cloog_domain_free(loop
->domain
) ;
211 cloog_domain_free(loop
->unsimplified
);
212 cloog_block_free(loop
->block
) ;
213 if (loop
->inner
!= NULL
)
214 cloog_loop_free(loop
->inner
) ;
216 cloog_stride_free(loop
->stride
);
224 * cloog_loop_free_parts function:
225 * This function frees the allocated memory for some parts of a CloogLoop
226 * structure (loop), each other argument is a boolean having to be set to 1 if
227 * we want to free the corresponding part, 0 otherwise. This function applies
228 * the same freeing policy to its inner ans next loops recursively.
229 * - July 3rd 2003: first version.
230 * - June 22nd 2005: Adaptation for GMP.
232 void cloog_loop_free_parts(loop
, domain
, block
, inner
, next
)
234 int domain
, block
, inner
, next
;
235 { CloogLoop
* follow
;
237 while (loop
!= NULL
) {
238 cloog_loop_leak_down(loop
->state
);
239 follow
= loop
->next
;
242 cloog_domain_free(loop
->domain
) ;
245 cloog_block_free(loop
->block
) ;
247 if ((inner
) && (loop
->inner
!= NULL
))
248 cloog_loop_free_parts(loop
->inner
,domain
,block
,inner
,1) ;
250 cloog_stride_free(loop
->stride
);
260 /******************************************************************************
261 * Reading functions *
262 ******************************************************************************/
266 * Construct a CloogLoop structure from a given iteration domain
267 * and statement number.
269 CloogLoop
*cloog_loop_from_domain(CloogState
*state
, CloogDomain
*domain
,
274 CloogStatement
* statement
;
276 /* Memory allocation and information reading for the first domain: */
277 loop
= cloog_loop_malloc(state
);
279 loop
->domain
= domain
;
280 if (loop
->domain
!= NULL
)
281 nb_iterators
= cloog_domain_dimension(loop
->domain
);
284 /* included statement block. */
285 statement
= cloog_statement_alloc(state
, number
+ 1);
286 loop
->block
= cloog_block_alloc(statement
, 0, NULL
, nb_iterators
);
293 * cloog_loop_read function:
294 * This function reads loop data from a file (foo, possibly stdin) and
295 * returns a pointer to a CloogLoop structure containing the read information.
296 * This function can be used only for input file reading, when one loop is
297 * associated with one statement.
298 * - number is the statement block number carried by the loop (-1 if none).
299 * - nb_parameters is the number of parameters.
301 * - September 9th 2002: first version.
302 * - April 16th 2005: adaptation to new CloogStatement struct (with number).
303 * - June 11th 2005: adaptation to new CloogBlock structure.
304 * - June 22nd 2005: Adaptation for GMP.
306 CloogLoop
*cloog_loop_read(CloogState
*state
,
307 FILE *foo
, int number
, int nb_parameters
)
313 domain
= cloog_domain_union_read(state
, foo
, nb_parameters
);
315 /* To read that stupid "0 0 0" line. */
316 while (fgets(s
,MAX_STRING
,foo
) == 0) ;
317 while ((*s
=='#' || *s
=='\n') || (sscanf(s
," %d %d %d",&op1
,&op2
,&op3
)<3))
318 fgets(s
,MAX_STRING
,foo
) ;
320 return cloog_loop_from_domain(state
, domain
, number
);
324 /******************************************************************************
325 * Processing functions *
326 ******************************************************************************/
330 * cloog_loop_malloc function:
331 * This function allocates the memory space for a CloogLoop structure and
332 * sets its fields with default values. Then it returns a pointer to the
334 * - November 21th 2005: first version.
336 CloogLoop
*cloog_loop_malloc(CloogState
*state
)
339 /* Memory allocation for the CloogLoop structure. */
340 loop
= (CloogLoop
*)malloc(sizeof(CloogLoop
)) ;
342 cloog_die("memory overflow.\n");
343 cloog_loop_leak_up(state
);
346 /* We set the various fields with default values. */
348 loop
->domain
= NULL
;
349 loop
->unsimplified
= NULL
;
362 * cloog_loop_alloc function:
363 * This function allocates the memory space for a CloogLoop structure and
364 * sets its fields with those given as input. Then it returns a pointer to the
366 * - October 27th 2001: first version.
367 * - June 22nd 2005: Adaptation for GMP.
368 * - November 21th 2005: use of cloog_loop_malloc.
370 CloogLoop
*cloog_loop_alloc(CloogState
*state
,
371 CloogDomain
*domain
, int otl
, CloogStride
*stride
,
372 CloogBlock
*block
, CloogLoop
*inner
, CloogLoop
*next
)
375 loop
= cloog_loop_malloc(state
);
377 loop
->domain
= domain
;
378 loop
->block
= block
;
379 loop
->inner
= inner
;
382 loop
->stride
= cloog_stride_copy(stride
);
389 * cloog_loop_add function:
390 * This function adds a CloogLoop structure (loop) at a given place (now) of a
391 * NULL terminated list of CloogLoop structures. The beginning of this list
392 * is (start). This function updates (now) to (loop), and updates (start) if the
393 * added element is the first one -that is when (start) is NULL-.
394 * - October 28th 2001: first version.
396 void cloog_loop_add(CloogLoop
** start
, CloogLoop
** now
, CloogLoop
* loop
)
397 { if (*start
== NULL
)
402 { (*now
)->next
= loop
;
403 *now
= (*now
)->next
;
409 * cloog_loop_add function:
410 * This function adds a CloogLoop structure (loop) at a given place (now) of a
411 * NULL terminated list of CloogLoop structures. The beginning of this list
412 * is (start). This function updates (now) to the end of the loop list (loop),
413 * and updates (start) if the added element is the first one -that is when
415 * - September 9th 2005: first version.
417 void cloog_loop_add_list(CloogLoop
** start
, CloogLoop
** now
, CloogLoop
* loop
)
418 { if (*start
== NULL
)
423 { (*now
)->next
= loop
;
424 *now
= (*now
)->next
;
427 while ((*now
)->next
!= NULL
)
428 *now
= (*now
)->next
;
433 * cloog_loop_copy function:
434 * This function returns a copy of the CloogLoop structure given as input. In
435 * fact, there is just new allocations for the CloogLoop structures, but their
436 * contents are the same.
437 * - October 28th 2001: first version.
438 * - July 3rd->11th 2003: memory leaks hunt and correction.
440 CloogLoop
* cloog_loop_copy(CloogLoop
* source
)
443 CloogDomain
* domain
;
447 { domain
= cloog_domain_copy(source
->domain
) ;
448 block
= cloog_block_copy(source
->block
) ;
449 loop
= cloog_loop_alloc(source
->state
, domain
, source
->otl
,
450 source
->stride
, block
, NULL
, NULL
);
451 loop
->usr
= source
->usr
;
452 loop
->inner
= cloog_loop_copy(source
->inner
) ;
453 loop
->next
= cloog_loop_copy(source
->next
) ;
460 * cloog_loop_add_disjoint function:
461 * This function adds some CloogLoop structures at a given place (now) of a
462 * NULL terminated list of CloogLoop structures. The beginning of this list
463 * is (start). (loop) can be an union of polyhedra, this function separates the
464 * union into a list of *disjoint* polyhedra then adds the list. This function
465 * updates (now) to the end of the list and updates (start) if first added
466 * element is the first of the principal list -that is when (start) is NULL-.
467 * (loop) can be freed by this function, basically when its domain is actually
468 * a union of polyhedra, but don't worry, all the useful data are now stored
469 * inside the list (start). We do not use PolyLib's Domain_Disjoint function,
470 * since the number of union components is often higher (thus code size too).
471 * - October 28th 2001: first version.
472 * - November 14th 2001: bug correction (this one was hard to find !).
473 * - July 3rd->11th 2003: memory leaks hunt and correction.
474 * - June 22nd 2005: Adaptation for GMP.
475 * - October 27th 2005: (debug) included blocks were not copied for new loops.
477 void cloog_loop_add_disjoint(start
, now
, loop
)
478 CloogLoop
** start
, ** now
, * loop
;
480 CloogLoop
* sep
, * inner
;
481 CloogDomain
*domain
, *seen
, *temp
, *rest
;
484 if (cloog_domain_isconvex(loop
->domain
))
485 cloog_loop_add(start
,now
,loop
) ;
487 domain
= cloog_domain_simplify_union(loop
->domain
);
488 loop
->domain
= NULL
;
490 /* We separate the first element of the rest of the union. */
491 domain
= cloog_domain_cut_first(domain
, &rest
);
493 /* This first element is the first of the list of disjoint polyhedra. */
494 sep
= cloog_loop_alloc(loop
->state
, domain
, 0, NULL
,
495 loop
->block
, loop
->inner
, NULL
);
496 cloog_loop_add(start
,now
,sep
) ;
498 seen
= cloog_domain_copy(domain
);
499 while (!cloog_domain_isempty(domain
= rest
)) {
500 temp
= cloog_domain_cut_first(domain
, &rest
);
501 domain
= cloog_domain_difference(temp
, seen
);
502 cloog_domain_free(temp
);
504 if (cloog_domain_isempty(domain
)) {
505 cloog_domain_free(domain
);
509 /* Each new loop will have its own life, for instance we can free its
510 * inner loop and included block. Then each one must have its own copy
511 * of both 'inner' and 'block'.
513 inner
= cloog_loop_copy(loop
->inner
) ;
514 block
= cloog_block_copy(loop
->block
) ;
516 sep
= cloog_loop_alloc(loop
->state
, cloog_domain_copy(domain
),
517 0, NULL
, block
, inner
, NULL
);
518 /* domain can be an union too. If so: recursion. */
519 if (cloog_domain_isconvex(domain
))
520 cloog_loop_add(start
,now
,sep
) ;
522 cloog_loop_add_disjoint(start
,now
,sep
) ;
524 if (cloog_domain_isempty(rest
)) {
525 cloog_domain_free(domain
);
529 seen
= cloog_domain_union(seen
, domain
);
531 cloog_domain_free(rest
);
532 cloog_domain_free(seen
);
533 cloog_loop_free_parts(loop
,0,0,0,0) ;
539 * cloog_loop_disjoint function:
540 * This function returns a list of loops such that each loop with non-convex
541 * domain in the input list (loop) is separated into several loops where the
542 * domains are the components of the union of *disjoint* polyhedra equivalent
543 * to the original non-convex domain. See cloog_loop_add_disjoint comments
545 * - September 16th 2005: first version.
547 CloogLoop
* cloog_loop_disjoint(CloogLoop
* loop
)
548 { CloogLoop
*res
=NULL
, * now
=NULL
, * next
;
550 /* Because this is often the case, don't waste time ! */
551 if (loop
&& !loop
->next
&& cloog_domain_isconvex(loop
->domain
))
555 { next
= loop
->next
;
557 cloog_loop_add_disjoint(&res
,&now
,loop
) ;
566 * cloog_loop_restrict function:
567 * This function returns the (loop) in the context of (context): it makes the
568 * intersection between the (loop) domain and the (context), then it returns
569 * a pointer to a new loop, with this intersection as domain.
571 * - October 27th 2001: first version.
572 * - June 15th 2005: a memory leak fixed (domain was not freed when empty).
573 * - June 22nd 2005: Adaptation for GMP.
575 CloogLoop
*cloog_loop_restrict(CloogLoop
*loop
, CloogDomain
*context
)
576 { int new_dimension
;
577 CloogDomain
* domain
, * extended_context
, * new_domain
;
578 CloogLoop
* new_loop
;
580 domain
= loop
->domain
;
581 if (cloog_domain_dimension(domain
) > cloog_domain_dimension(context
))
583 new_dimension
= cloog_domain_dimension(domain
);
584 extended_context
= cloog_domain_extend(context
, new_dimension
);
585 new_domain
= cloog_domain_intersection(extended_context
,loop
->domain
) ;
586 cloog_domain_free(extended_context
) ;
589 new_domain
= cloog_domain_intersection(context
,loop
->domain
) ;
591 if (cloog_domain_isempty(new_domain
))
592 { cloog_domain_free(new_domain
) ;
596 new_loop
= cloog_loop_alloc(loop
->state
, new_domain
,
597 0, NULL
, loop
->block
, loop
->inner
, NULL
);
604 * Call cloog_loop_restrict on each loop in the list "loop" and return
605 * the concatenated result.
607 CloogLoop
*cloog_loop_restrict_all(CloogLoop
*loop
, CloogDomain
*context
)
610 CloogLoop
*res
= NULL
;
611 CloogLoop
**res_next
= &res
;
613 for (; loop
; loop
= next
) {
616 *res_next
= cloog_loop_restrict(loop
, context
);
618 res_next
= &(*res_next
)->next
;
619 cloog_loop_free_parts(loop
, 1, 0, 0, 0);
622 cloog_loop_free(loop
);
629 CloogLoop
*cloog_loop_restrict_inner(CloogLoop
*loop
)
633 for (l
= loop
; l
; l
= l
->next
)
634 l
->inner
= cloog_loop_restrict_all(l
->inner
, l
->domain
);
640 * cloog_loop_project function:
641 * This function returns the projection of (loop) on the (level) first
642 * dimensions (outer loops). It makes the projection of the (loop) domain,
643 * then it returns a pointer to a new loop, with this projection as domain.
645 * - October 27th 2001: first version.
646 * - July 3rd->11th 2003: memory leaks hunt and correction.
647 * - June 22nd 2005: Adaptation for GMP.
649 CloogLoop
* cloog_loop_project(CloogLoop
* loop
, int level
)
651 CloogDomain
* new_domain
;
652 CloogLoop
* new_loop
, * copy
;
654 copy
= cloog_loop_alloc(loop
->state
, loop
->domain
, loop
->otl
, loop
->stride
,
655 loop
->block
, loop
->inner
, NULL
);
657 if (cloog_domain_dimension(loop
->domain
) == level
)
658 new_domain
= cloog_domain_copy(loop
->domain
) ;
660 new_domain
= cloog_domain_project(loop
->domain
, level
);
662 new_loop
= cloog_loop_alloc(loop
->state
, new_domain
, 0, NULL
,
670 * Call cloog_loop_project on each loop in the list "loop" and return
671 * the concatenated result.
673 CloogLoop
*cloog_loop_project_all(CloogLoop
*loop
, int level
)
676 CloogLoop
*res
= NULL
;
677 CloogLoop
**res_next
= &res
;
679 for (; loop
; loop
= next
) {
682 *res_next
= cloog_loop_project(loop
, level
);
683 res_next
= &(*res_next
)->next
;
684 cloog_loop_free_parts(loop
, 0, 0, 0, 0);
692 * cloog_loop_concat function:
693 * This function returns a pointer to the concatenation of the
694 * CloogLoop lists given as input.
695 * - October 28th 2001: first version.
697 CloogLoop
* cloog_loop_concat(CloogLoop
* a
, CloogLoop
* b
)
698 { CloogLoop
* loop
, * temp
;
703 { while (temp
->next
!= NULL
)
715 * cloog_loop_combine:
716 * Combine consecutive loops with identical domains into
717 * a single loop with the concatenation of their inner loops
720 CloogLoop
*cloog_loop_combine(CloogLoop
*loop
)
722 CloogLoop
*first
, *second
;
724 for (first
= loop
; first
; first
= first
->next
) {
725 while (first
->next
) {
726 if (!cloog_domain_lazy_equal(first
->domain
, first
->next
->domain
))
728 second
= first
->next
;
729 first
->inner
= cloog_loop_concat(first
->inner
, second
->inner
);
730 first
->next
= second
->next
;
731 cloog_loop_free_parts(second
, 1, 0, 0, 0);
739 * Remove loops from list that have an empty domain.
741 CloogLoop
*cloog_loop_remove_empty_domain_loops(CloogLoop
*loop
)
743 CloogLoop
*l
, *res
, *next
, **res_next
;
747 for (l
= loop
; l
; l
= next
) {
749 if (cloog_domain_isempty(l
->domain
))
750 cloog_loop_free_parts(l
, 1, 1, 1, 0);
753 res_next
= &(*res_next
)->next
;
761 CloogLoop
*cloog_loop_decompose_inner(CloogLoop
*loop
,
762 int level
, int scalar
, int *scaldims
, int nb_scattdims
);
764 /* For each loop with only one inner loop, replace the domain
765 * of the loop with the projection of the domain of the inner
766 * loop. To increase the number of loops with a single inner
767 * we first decompose the inner loops into strongly connected
770 CloogLoop
*cloog_loop_specialize(CloogLoop
*loop
,
771 int level
, int scalar
, int *scaldims
, int nb_scattdims
)
777 loop
= cloog_loop_decompose_inner(loop
, level
, scalar
,
778 scaldims
, nb_scattdims
);
780 for (l
= loop
; l
; l
= l
->next
) {
783 if (!cloog_domain_isconvex(l
->inner
->domain
))
786 dim
= cloog_domain_dimension(l
->domain
);
787 domain
= cloog_domain_project(l
->inner
->domain
, dim
);
788 if (cloog_domain_isconvex(domain
)) {
789 cloog_domain_free(l
->domain
);
792 cloog_domain_free(domain
);
796 return cloog_loop_remove_empty_domain_loops(loop
);
800 * cloog_loop_separate function:
801 * This function implements the Quillere algorithm for separation of multiple
802 * loops: for a given set of polyhedra (loop), it computes a set of disjoint
803 * polyhedra such that the unions of these sets are equal, and returns this set.
804 * - October 28th 2001: first version.
805 * - November 14th 2001: elimination of some unused blocks.
806 * - August 13th 2002: (debug) in the case of union of polyhedra for one
807 * loop, redundant constraints are fired.
808 * - July 3rd->11th 2003: memory leaks hunt and correction.
809 * - June 22nd 2005: Adaptation for GMP.
810 * - October 16th 2005: Removal of the non-shared constraint elimination when
811 * there is only one loop in the list (seems to work
812 * without now, DomainSimplify may have been improved).
813 * The problem was visible with test/iftest2.cloog.
815 CloogLoop
* cloog_loop_separate(CloogLoop
* loop
)
816 { int lazy_equal
=0, disjoint
= 0;
817 CloogLoop
* new_loop
, * new_inner
, * res
, * now
, * temp
, * Q
,
818 * inner
, * old
/*, * previous, * next*/ ;
819 CloogDomain
*UQ
, *domain
;
824 loop
= cloog_loop_combine(loop
);
826 if (loop
->next
== NULL
)
827 return cloog_loop_disjoint(loop
) ;
829 UQ
= cloog_domain_copy(loop
->domain
) ;
830 domain
= cloog_domain_copy(loop
->domain
) ;
831 res
= cloog_loop_alloc(loop
->state
, domain
, 0, NULL
,
832 loop
->block
, loop
->inner
, NULL
);
835 while((loop
= loop
->next
) != NULL
)
838 /* For all Q, add Q-loop associated with the blocks of Q alone,
839 * and Q inter loop associated with the blocks of Q and loop.
841 for (Q
= res
; Q
; Q
= Q
->next
) {
842 /* Add (Q inter loop). */
843 if ((disjoint
= cloog_domain_lazy_disjoint(Q
->domain
,loop
->domain
)))
846 { if ((lazy_equal
= cloog_domain_lazy_equal(Q
->domain
,loop
->domain
)))
847 domain
= cloog_domain_copy(Q
->domain
) ;
849 domain
= cloog_domain_intersection(Q
->domain
,loop
->domain
) ;
851 if (!cloog_domain_isempty(domain
))
852 { new_inner
= cloog_loop_concat(cloog_loop_copy(Q
->inner
),
853 cloog_loop_copy(loop
->inner
)) ;
854 new_loop
= cloog_loop_alloc(loop
->state
, domain
, 0, NULL
,
855 NULL
, new_inner
, NULL
);
856 cloog_loop_add_disjoint(&temp
,&now
,new_loop
) ;
860 cloog_domain_free(domain
);
864 /* Add (Q - loop). */
866 domain
= cloog_domain_copy(Q
->domain
) ;
869 domain
= cloog_domain_empty(Q
->domain
);
871 domain
= cloog_domain_difference(Q
->domain
,loop
->domain
) ;
874 if (!cloog_domain_isempty(domain
)) {
875 new_loop
= cloog_loop_alloc(loop
->state
, domain
, 0, NULL
,
876 NULL
, Q
->inner
, NULL
);
877 cloog_loop_add_disjoint(&temp
,&now
,new_loop
) ;
880 { cloog_domain_free(domain
) ;
881 /* If Q->inner is no more useful, we can free it. */
884 cloog_loop_free(inner
) ;
888 /* Add loop-UQ associated with the blocks of loop alone.*/
889 if (cloog_domain_lazy_disjoint(loop
->domain
,UQ
))
890 domain
= cloog_domain_copy(loop
->domain
) ;
892 { if (cloog_domain_lazy_equal(loop
->domain
,UQ
))
893 domain
= cloog_domain_empty(UQ
);
895 domain
= cloog_domain_difference(loop
->domain
,UQ
) ;
898 if (!cloog_domain_isempty(domain
)) {
899 new_loop
= cloog_loop_alloc(loop
->state
, domain
, 0, NULL
,
900 NULL
, loop
->inner
, NULL
);
901 cloog_loop_add_disjoint(&temp
,&now
,new_loop
) ;
904 { cloog_domain_free(domain
) ;
905 /* If loop->inner is no more useful, we can free it. */
906 cloog_loop_free(loop
->inner
) ;
911 if (loop
->next
!= NULL
)
912 UQ
= cloog_domain_union(UQ
, cloog_domain_copy(loop
->domain
));
914 cloog_domain_free(UQ
);
916 cloog_loop_free_parts(res
,1,0,0,1) ;
920 cloog_loop_free_parts(old
,1,0,0,1) ;
926 static CloogDomain
*bounding_domain(CloogDomain
*dom
, CloogOptions
*options
)
929 return cloog_domain_simple_convex(dom
);
931 return cloog_domain_convex(dom
);
936 * cloog_loop_merge function:
937 * This function is the 'soft' version of loop_separate if we are looking for
938 * a code much simpler (and less efficicient). This function returns the new
940 * - October 29th 2001: first version.
941 * - July 3rd->11th 2003: memory leaks hunt and correction.
942 * - June 22nd 2005: Adaptation for GMP.
944 CloogLoop
*cloog_loop_merge(CloogLoop
*loop
, int level
, CloogOptions
*options
)
946 CloogLoop
*res
, *new_inner
, *old
;
947 CloogDomain
*new_domain
, *temp
;
952 if (loop
->next
== NULL
)
953 return cloog_loop_disjoint(loop
);
958 new_inner
= loop
->inner
;
960 for (loop
= loop
->next
; loop
; loop
= loop
->next
) {
961 temp
= cloog_domain_union(temp
, loop
->domain
);
963 new_inner
= cloog_loop_concat(new_inner
, loop
->inner
);
966 new_domain
= bounding_domain(temp
, options
);
968 if (level
> 0 && !cloog_domain_is_bounded(new_domain
, level
) &&
969 cloog_domain_is_bounded(temp
, level
)) {
970 CloogDomain
*splitter
, *t2
;
972 cloog_domain_free(new_domain
);
973 splitter
= cloog_domain_bound_splitter(temp
, level
);
976 while (!cloog_domain_isconvex(splitter
)) {
977 CloogDomain
*first
, *rest
;
978 first
= cloog_domain_cut_first(splitter
, &rest
);
980 t2
= cloog_domain_intersection(first
, temp
);
981 cloog_domain_free(first
);
983 new_domain
= bounding_domain(t2
, options
);
984 cloog_domain_free(t2
);
986 if (cloog_domain_isempty(new_domain
)) {
987 cloog_domain_free(new_domain
);
990 res
= cloog_loop_alloc(old
->state
, new_domain
, 0, NULL
,
991 NULL
, cloog_loop_copy(new_inner
), res
);
994 t2
= cloog_domain_intersection(splitter
, temp
);
995 cloog_domain_free(splitter
);
997 new_domain
= bounding_domain(t2
, options
);
998 cloog_domain_free(t2
);
1000 if (cloog_domain_isempty(new_domain
)) {
1001 cloog_domain_free(new_domain
);
1002 cloog_loop_free(new_inner
);
1004 res
= cloog_loop_alloc(old
->state
, new_domain
, 0, NULL
,
1005 NULL
, new_inner
, res
);
1007 res
= cloog_loop_alloc(old
->state
, new_domain
, 0, NULL
,
1008 NULL
, new_inner
, NULL
);
1010 cloog_domain_free(temp
);
1012 cloog_loop_free_parts(old
, 0, 0, 0, 1);
1018 static int cloog_loop_count(CloogLoop
*loop
)
1022 for (nb_loops
= 0; loop
; loop
= loop
->next
)
1030 * cloog_loop_sort function:
1031 * Adaptation from LoopGen 0.4 by F. Quillere. This function sorts a list of
1032 * parameterized disjoint polyhedra, in order to not have lexicographic order
1033 * violation (see Quillere paper).
1034 * - September 16th 2005: inclusion of cloog_loop_number (October 29th 2001).
1036 CloogLoop
*cloog_loop_sort(CloogLoop
*loop
, int level
)
1038 CloogLoop
*res
, *now
, **loop_array
;
1040 int i
, nb_loops
=0, * permut
;
1042 /* There is no need to sort the parameter domains. */
1046 /* We will need to know how many loops are in the list. */
1047 nb_loops
= cloog_loop_count(loop
);
1049 /* If there is only one loop, it's the end. */
1053 /* We have to allocate memory for some useful components:
1054 * - loop_array: the loop array,
1055 * - doms: the array of domains to sort,
1056 * - permut: will give us a possible sort (maybe not the only one).
1058 loop_array
= (CloogLoop
**)malloc(nb_loops
*sizeof(CloogLoop
*)) ;
1059 doms
= (CloogDomain
**)malloc(nb_loops
*sizeof(CloogDomain
*));
1060 permut
= (int *)malloc(nb_loops
*sizeof(int)) ;
1062 /* We fill up the loop and domain arrays. */
1063 for (i
=0;i
<nb_loops
;i
++,loop
=loop
->next
)
1064 { loop_array
[i
] = loop
;
1065 doms
[i
] = loop_array
[i
]->domain
;
1068 /* cloog_domain_sort will fill up permut. */
1069 cloog_domain_sort(doms
, nb_loops
, level
, permut
);
1071 /* With permut and loop_array we build the sorted list. */
1073 for (i
=0;i
<nb_loops
;i
++)
1074 { /* To avoid pointer looping... loop_add will rebuild the list. */
1075 loop_array
[permut
[i
]-1]->next
= NULL
;
1076 cloog_loop_add(&res
,&now
,loop_array
[permut
[i
]-1]) ;
1088 * cloog_loop_nest function:
1089 * This function changes the loop list in such a way that we have no more than
1090 * one dimension added by level. It returns an equivalent loop list with
1092 * - October 29th 2001: first version.
1093 * - July 3rd->11th 2003: memory leaks hunt and correction.
1094 * - June 22nd 2005: Adaptation for GMP.
1095 * - November 21th 2005: (debug) now OK when cloog_loop_restrict returns NULL.
1097 CloogLoop
*cloog_loop_nest(CloogLoop
*loop
, CloogDomain
*context
, int level
)
1099 CloogLoop
* p
, * temp
, * res
, * now
, * next
;
1100 CloogDomain
* new_domain
;
1102 loop
= cloog_loop_disjoint(loop
);
1105 /* Each domain is changed by its intersection with the context. */
1106 while (loop
!= NULL
)
1107 { p
= cloog_loop_restrict(loop
, context
);
1111 { cloog_loop_free_parts(loop
,1,0,0,0) ;
1113 temp
= cloog_loop_alloc(p
->state
, p
->domain
, 0, NULL
,
1114 p
->block
, p
->inner
, NULL
);
1116 /* If the intersection dimension is too big, we make projections smaller
1117 * and smaller, and each projection includes the preceding projection
1118 * (thus, in the target list, dimensions are added one by one).
1120 if (cloog_domain_dimension(p
->domain
) >= level
)
1121 for (l
= cloog_domain_dimension(p
->domain
); l
>= level
; l
--) {
1122 new_domain
= cloog_domain_project(p
->domain
, l
);
1123 temp
= cloog_loop_alloc(p
->state
, new_domain
, 0, NULL
,
1127 /* p is no more useful (but its content yes !). */
1128 cloog_loop_free_parts(p
,0,0,0,0) ;
1130 cloog_loop_add(&res
,&now
,temp
) ;
1133 cloog_loop_free_parts(loop
,1,1,1,0) ;
1142 /* Check if the domains of the inner loops impose a stride constraint
1143 * on the given level.
1144 * The core of the search is implemented in cloog_domain_list_stride.
1145 * Here, we simply construct a list of domains to pass to this function
1146 * and if a stride is found, we adjust the lower bounds by calling
1147 * cloog_domain_stride_lower_bound.
1149 static int cloog_loop_variable_offset_stride(CloogLoop
*loop
, int level
)
1151 CloogDomainList
*list
= NULL
;
1153 CloogStride
*stride
;
1155 for (inner
= loop
->inner
; inner
; inner
= inner
->next
) {
1156 CloogDomainList
*entry
= ALLOC(CloogDomainList
);
1157 entry
->domain
= cloog_domain_copy(inner
->domain
);
1162 stride
= cloog_domain_list_stride(list
, level
);
1164 cloog_domain_list_free(list
);
1169 loop
->stride
= stride
;
1170 loop
->domain
= cloog_domain_stride_lower_bound(loop
->domain
, level
, stride
);
1177 * cloog_loop_stride function:
1178 * This function will find the stride of a loop for the iterator at the column
1179 * number 'level' in the constraint matrix. It will update the lower bound of
1180 * the iterator accordingly. Basically, the function will try to find in the
1181 * inner loops a common condition on this iterator for the inner loop iterators
1182 * to be integral. For instance, let us consider a loop with the iterator i,
1183 * the iteration domain -4<=i<=n, and its two inner loops with the iterator j.
1184 * The first inner loop has the constraint 3j=i, and the second one has the
1185 * constraint 6j=i. Then the common constraint on i for j to be integral is
1186 * i%3=0, the stride for i is 3. Lastly, we have to find the new lower bound
1187 * for i: the first value satisfying the common constraint: -3. At the end, the
1188 * iteration domain for i is -3<=i<=n and the stride for i is 3.
1190 * The algorithm implemented in this function only allows for strides
1191 * on loops with a lower bound that has a constant remainder on division
1192 * by the stride. Before initiating this procedure, we first check
1193 * if we can find a stride with a lower bound with a variable offset in
1194 * cloog_loop_variable_offset_stride.
1196 * - loop is the loop including the iteration domain of the considered iterator,
1197 * - level is the column number of the iterator in the matrix of contraints.
1199 * - June 29th 2003: first version (work in progress since June 26th 2003).
1200 * - July 14th 2003: simpler version.
1201 * - June 22nd 2005: Adaptation for GMP (from S. Verdoolaege's 0.12.1 version).
1203 void cloog_loop_stride(CloogLoop
* loop
, int level
)
1204 { int first_search
;
1205 cloog_int_t stride
, ref_offset
, offset
, potential
;
1208 if (!cloog_domain_can_stride(loop
->domain
, level
))
1211 if (cloog_loop_variable_offset_stride(loop
, level
))
1214 cloog_int_init(stride
);
1215 cloog_int_init(ref_offset
);
1216 cloog_int_init(offset
);
1217 cloog_int_init(potential
);
1219 cloog_int_set_si(ref_offset
, 0);
1220 cloog_int_set_si(offset
, 0);
1222 /* Default stride. */
1223 cloog_int_set_si(stride
, 1);
1225 inner
= loop
->inner
;
1227 while (inner
!= NULL
)
1228 { /* If the minimun stride has not been found yet, find the stride. */
1229 if ((first_search
) || (!cloog_int_is_one(stride
)))
1231 cloog_domain_stride(inner
->domain
, level
, &potential
, &offset
);
1232 if (!cloog_int_is_one(potential
) && (!first_search
))
1233 { /* Offsets must be the same for common stride. */
1234 cloog_int_gcd(stride
, potential
, stride
);
1235 if (!cloog_int_is_zero(stride
)) {
1236 cloog_int_fdiv_r(offset
, offset
, stride
);
1237 cloog_int_fdiv_r(ref_offset
, ref_offset
, stride
);
1239 if (cloog_int_ne(offset
,ref_offset
))
1240 cloog_int_set_si(stride
, 1);
1243 cloog_int_set(stride
, potential
);
1244 cloog_int_set(ref_offset
, offset
);
1250 inner
= inner
->next
;
1253 if (cloog_int_is_zero(stride
))
1254 cloog_int_set_si(stride
, 1);
1256 /* Update the values if necessary. */
1257 if (!cloog_int_is_one(stride
))
1258 { /* Update the stride value. */
1259 if (!cloog_int_is_zero(offset
))
1260 cloog_int_sub(offset
, stride
, offset
);
1261 loop
->stride
= cloog_stride_alloc(stride
, offset
);
1262 loop
->domain
= cloog_domain_stride_lower_bound(loop
->domain
, level
,
1266 cloog_int_clear(stride
);
1267 cloog_int_clear(ref_offset
);
1268 cloog_int_clear(offset
);
1269 cloog_int_clear(potential
);
1273 void cloog_loop_otl(CloogLoop
*loop
, int level
)
1275 if (cloog_domain_is_otl(loop
->domain
, level
))
1281 * cloog_loop_stop function:
1282 * This function implements the 'stop' option : each domain of each loop
1283 * in the list 'loop' is replaced by 'context'. 'context' should be the
1284 * domain of the outer loop. By using this method, there are no more dimensions
1285 * to scan and the simplification step will automaticaly remove the domains
1286 * since they are the same as the corresponding contexts. The effect of this
1287 * function is to stop the code generation at the level this function is called,
1288 * the resulting code do not consider the next dimensions.
1289 * - January 11th 2005: first version.
1291 CloogLoop
* cloog_loop_stop(CloogLoop
* loop
, CloogDomain
* context
)
1295 { cloog_domain_free(loop
->domain
) ;
1296 loop
->domain
= cloog_domain_copy(context
) ;
1297 loop
->next
= cloog_loop_stop(loop
->next
, context
) ;
1304 static int level_is_constant(int level
, int scalar
, int *scaldims
, int nb_scattdims
)
1306 return level
&& (level
+scalar
<= nb_scattdims
) && (scaldims
[level
+scalar
-1]);
1311 * Compare the constant dimensions of loops 'l1' and 'l2' starting at 'scalar'
1312 * and return -1 if the vector of constant dimensions of 'l1' is smaller
1313 * than that of 'l2', 0 if they are the same and +1 if that of 'l1' is
1314 * greater than that of 'l2'.
1315 * This function should be called on the innermost loop (the loop
1316 * containing a block).
1317 * \param l1 Loop to be compared with l2.
1318 * \param l2 Loop to be compared with l1.
1319 * \param level Current non-scalar dimension.
1320 * \param scaldims Boolean array saying whether a dimension is scalar or not.
1321 * \param nb_scattdims Size of the scaldims array.
1322 * \param scalar Current scalar dimension.
1323 * \return -1 if (l1 < l2), 0 if (l1 == l2) and +1 if (l1 > l2)
1325 int cloog_loop_constant_cmp(CloogLoop
*l1
, CloogLoop
*l2
, int level
,
1326 int *scaldims
, int nb_scattdims
, int scalar
)
1328 CloogBlock
*b1
, *b2
;
1331 while (level_is_constant(level
, scalar
, scaldims
, nb_scattdims
)) {
1332 int cmp
= cloog_int_cmp(b1
->scaldims
[scalar
], b2
->scaldims
[scalar
]);
1342 * cloog_loop_scalar_gt function:
1343 * This function returns 1 if loop 'l1' is greater than loop 'l2' for the
1344 * scalar dimension vector that begins at dimension 'scalar', 0 otherwise. What
1345 * we want to know is whether a loop is scheduled before another one or not.
1346 * This function solves the problem when the considered dimension for scheduling
1347 * is a scalar dimension. Since there may be a succession of scalar dimensions,
1348 * this function will reason about the vector of scalar dimension that begins
1349 * at dimension 'level+scalar' and finish to the first non-scalar dimension.
1350 * \param l1 Loop to be compared with l2.
1351 * \param l2 Loop to be compared with l1.
1352 * \param level Current non-scalar dimension.
1353 * \param scaldims Boolean array saying whether a dimension is scalar or not.
1354 * \param nb_scattdims Size of the scaldims array.
1355 * \param scalar Current scalar dimension.
1356 * \return 1 if (l1 > l2), 0 otherwise.
1358 * - September 9th 2005: first version.
1359 * - October 15nd 2007: now "greater than" instead of "greater or equal".
1361 int cloog_loop_scalar_gt(l1
, l2
, level
, scaldims
, nb_scattdims
, scalar
)
1362 CloogLoop
* l1
, * l2
;
1363 int level
, * scaldims
, nb_scattdims
, scalar
;
1365 return cloog_loop_constant_cmp(l1
, l2
, level
, scaldims
, nb_scattdims
, scalar
) > 0;
1370 * cloog_loop_scalar_eq function:
1371 * This function returns 1 if loop 'l1' is equal to loop 'l2' for the scalar
1372 * dimension vector that begins at dimension 'scalar', 0 otherwise. What we want
1373 * to know is whether two loops are scheduled for the same time or not.
1374 * This function solves the problem when the considered dimension for scheduling
1375 * is a scalar dimension. Since there may be a succession of scalar dimensions,
1376 * this function will reason about the vector of scalar dimension that begins
1377 * at dimension 'level+scalar' and finish to the first non-scalar dimension.
1378 * - l1 and l2 are the loops to compare,
1379 * - level is the current non-scalar dimension,
1380 * - scaldims is the boolean array saying whether a dimension is scalar or not,
1381 * - nb_scattdims is the size of the scaldims array,
1382 * - scalar is the current scalar dimension.
1384 * - September 9th 2005 : first version.
1386 int cloog_loop_scalar_eq(l1
, l2
, level
, scaldims
, nb_scattdims
, scalar
)
1387 CloogLoop
* l1
, * l2
;
1388 int level
, * scaldims
, nb_scattdims
, scalar
;
1390 return cloog_loop_constant_cmp(l1
, l2
, level
, scaldims
, nb_scattdims
, scalar
) == 0;
1395 * cloog_loop_scalar_sort function:
1396 * This function sorts a linked list of loops (loop) with respect to the
1397 * scalar dimension vector that begins at dimension 'scalar'. Since there may
1398 * be a succession of scalar dimensions, this function will reason about the
1399 * vector of scalar dimension that begins at dimension 'level+scalar' and
1400 * finish to the first non-scalar dimension.
1401 * \param loop Loop list to sort.
1402 * \param level Current non-scalar dimension.
1403 * \param scaldims Boolean array saying whether a dimension is scalar or not.
1404 * \param nb_scattdims Size of the scaldims array.
1405 * \param scalar Current scalar dimension.
1406 * \return A pointer to the sorted list.
1408 * - July 2nd 2005: first developments.
1409 * - September 2nd 2005: first version.
1410 * - October 15nd 2007: complete rewrite to remove bugs, now a bubble sort.
1412 CloogLoop
* cloog_loop_scalar_sort(loop
, level
, scaldims
, nb_scattdims
, scalar
)
1414 int level
, * scaldims
, nb_scattdims
, scalar
;
1416 CloogLoop
**current
;
1420 for (current
= &loop
; (*current
)->next
; current
= &(*current
)->next
) {
1421 CloogLoop
*next
= (*current
)->next
;
1422 if (cloog_loop_scalar_gt(*current
,next
,level
,scaldims
,nb_scattdims
,scalar
)) {
1424 (*current
)->next
= next
->next
;
1425 next
->next
= *current
;
1436 * cloog_loop_generate_backtrack function:
1437 * adaptation from LoopGen 0.4 by F. Quillere. This function implements the
1438 * backtrack of the Quillere et al. algorithm (see the Quillere paper).
1439 * It eliminates unused iterations of the current level for the new one. See the
1440 * example called linearity-1-1 example with and without this part for an idea.
1441 * - October 26th 2001: first version in cloog_loop_generate_general.
1442 * - July 31th 2002: (debug) no more parasite loops (REALLY hard !).
1443 * - October 30th 2005: extraction from cloog_loop_generate_general.
1445 CloogLoop
*cloog_loop_generate_backtrack(CloogLoop
*loop
,
1446 int level
, CloogOptions
*options
)
1448 CloogDomain
* domain
;
1449 CloogLoop
* now
, * now2
, * next
, * next2
, * end
, * temp
, * l
, * inner
,
1455 while (temp
!= NULL
)
1457 inner
= temp
->inner
;
1459 while (inner
!= NULL
)
1460 { next
= inner
->next
;
1461 /* This 'if' and its first part is the debug of july 31th 2002. */
1462 if (inner
->block
!= NULL
) {
1463 end
= cloog_loop_alloc(temp
->state
, inner
->domain
, 0, NULL
,
1464 inner
->block
, NULL
, NULL
);
1465 domain
= cloog_domain_copy(temp
->domain
) ;
1466 new_loop
= cloog_loop_alloc(temp
->state
, domain
, 0, NULL
,
1470 new_loop
= cloog_loop_project(inner
, level
);
1472 cloog_loop_free_parts(inner
,0,0,0,0) ;
1473 cloog_loop_add(&l
,&now2
,new_loop
) ;
1477 temp
->inner
= NULL
;
1480 { l
= cloog_loop_separate(l
) ;
1481 l
= cloog_loop_sort(l
, level
);
1483 l
->stride
= cloog_stride_copy(l
->stride
);
1484 cloog_loop_add(&loop
,&now
,l
) ;
1488 next2
= temp
->next
;
1489 cloog_loop_free_parts(temp
,1,0,0,0) ;
1498 * Return 1 if we need to continue recursing to the specified level.
1500 int cloog_loop_more(CloogLoop
*loop
, int level
, int scalar
, int nb_scattdims
)
1502 return level
+ scalar
<= nb_scattdims
||
1503 cloog_domain_dimension(loop
->domain
) >= level
;
1506 CloogLoop
*cloog_loop_generate_restricted_or_stop(CloogLoop
*loop
,
1507 CloogDomain
*context
,
1508 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
1509 CloogOptions
*options
);
1512 * cloog_loop_generate_general function:
1513 * Adaptation from LoopGen 0.4 by F. Quillere. This function implements the
1514 * Quillere algorithm for polyhedron scanning from step 3 to 5.
1515 * (see the Quillere paper).
1516 * - loop is the loop for which we have to generate a scanning code,
1517 * - level is the current non-scalar dimension,
1518 * - scalar is the current scalar dimension,
1519 * - scaldims is the boolean array saying whether a dimension is scalar or not,
1520 * - nb_scattdims is the size of the scaldims array,
1521 * - options are the general code generation options.
1523 * - October 26th 2001: first version.
1524 * - July 3rd->11th 2003: memory leaks hunt and correction.
1525 * - June 22nd 2005: Adaptation for GMP.
1526 * - September 2nd 2005: The function have been cutted out in two pieces:
1527 * cloog_loop_generate and this one, in order to handle
1528 * the scalar dimension case more efficiently with
1529 * cloog_loop_generate_scalar.
1530 * - November 15th 2005: (debug) the result of the cloog_loop_generate call may
1531 * be a list of polyhedra (especially if stop option is
1532 * used): cloog_loop_add_list instead of cloog_loop_add.
1534 CloogLoop
*cloog_loop_generate_general(CloogLoop
*loop
,
1535 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
1536 CloogOptions
*options
)
1538 CloogLoop
* res
, * now
, * temp
, * l
, * new_loop
, * inner
, * now2
, * end
,
1540 CloogDomain
* domain
;
1543 /* 3. Separate all projections into disjoint polyhedra. */
1544 if ((options
->f
> level
+scalar
) || (options
->f
< 0))
1545 res
= cloog_loop_merge(loop
, level
, options
);
1547 res
= cloog_loop_separate(loop
);
1551 /* 3b. -correction- sort the loops to determine their textual order. */
1552 res
= cloog_loop_sort(res
, level
);
1554 res
= cloog_loop_restrict_inner(res
);
1557 res
= cloog_loop_specialize(res
, level
, scalar
, scaldims
, nb_scattdims
);
1559 /* 4. Recurse for each loop with the current domain as context. */
1562 if (!level
|| (level
+scalar
< options
->l
) || (options
->l
< 0))
1564 { if (level
&& options
->strides
)
1565 cloog_loop_stride(temp
, level
);
1566 if (level
&& options
->otl
)
1567 cloog_loop_otl(temp
, level
);
1568 inner
= temp
->inner
;
1569 domain
= temp
->domain
;
1571 while (inner
!= NULL
)
1572 { /* 4b. -ced- recurse for each sub-list of non terminal loops. */
1573 if (cloog_loop_more(inner
, level
+ 1, scalar
, nb_scattdims
)) {
1575 while ((end
->next
!= NULL
) &&
1576 cloog_loop_more(end
->next
, level
+ 1, scalar
, nb_scattdims
))
1582 l
= cloog_loop_generate_restricted_or_stop(inner
, domain
,
1583 level
+ 1, scalar
, scaldims
, nb_scattdims
, options
);
1586 cloog_loop_add_list(&into
,&now
,l
) ;
1591 { cloog_loop_add(&into
,&now
,inner
) ;
1592 inner
= inner
->next
;
1597 temp
->inner
= into
;
1598 cloog_loop_add(&res
,&now2
,temp
) ;
1602 while (temp
!= NULL
)
1603 { next
= temp
->next
;
1604 l
= cloog_loop_nest(temp
->inner
, temp
->domain
, level
+1);
1605 new_loop
= cloog_loop_alloc(temp
->state
, temp
->domain
, 0, NULL
,
1607 temp
->inner
= NULL
;
1609 cloog_loop_free_parts(temp
,0,0,0,0) ;
1610 cloog_loop_add(&res
,&now
,new_loop
) ;
1614 /* 5. eliminate unused iterations of the current level for the new one. See
1615 * the example called linearity-1-1 example with and without this part
1618 if (options
->backtrack
&& level
&&
1619 ((level
+scalar
< options
->l
) || (options
->l
< 0)) &&
1620 ((options
->f
<= level
+scalar
) && !(options
->f
< 0)))
1621 res
= cloog_loop_generate_backtrack(res
, level
, options
);
1623 /* Pray for my new paper to be accepted somewhere since the following stuff
1624 * is really amazing :-) !
1625 * Far long later: The paper has been accepted to PACT 2004 :-))). But there
1626 * are still some bugs and I have no time to fix them. Thus now you have to
1627 * pray for me to get an academic position for that really amazing stuff :-) !
1628 * Later again: OK, I get my academic position, but still I have not enough
1629 * time to fix and clean this part... Pray again :-) !!!
1631 /* res = cloog_loop_unisolate(res,level) ;*/
1637 CloogLoop
*cloog_loop_generate_restricted(CloogLoop
*loop
,
1638 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
1639 CloogOptions
*options
);
1643 * cloog_loop_generate_scalar function:
1644 * This function applies the simplified code generation scheme in the trivial
1645 * case of scalar dimensions. When dealing with scalar dimensions, there is
1646 * no need of costly polyhedral operations for separation or sorting: sorting
1647 * is a question of comparing scalar vectors and separation amounts to consider
1648 * only loops with the same scalar vector for the next step of the code
1649 * generation process. This function achieves the separation/sorting process
1650 * for the vector of scalar dimension that begins at dimension 'level+scalar'
1651 * and finish to the first non-scalar dimension.
1652 * - loop is the loop for which we have to generate a scanning code,
1653 * - level is the current non-scalar dimension,
1654 * - scalar is the current scalar dimension,
1655 * - scaldims is the boolean array saying whether a dimension is scalar or not,
1656 * - nb_scattdims is the size of the scaldims array,
1657 * - options are the general code generation options.
1659 * - September 2nd 2005: First version.
1661 CloogLoop
*cloog_loop_generate_scalar(CloogLoop
*loop
,
1662 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
1663 CloogOptions
*options
)
1664 { CloogLoop
* res
, * now
, * temp
, * l
, * end
, * next
, * ref
;
1667 /* We sort the loop list with respect to the current scalar vector. */
1668 res
= cloog_loop_scalar_sort(loop
,level
,scaldims
,nb_scattdims
,scalar
) ;
1670 scalar_new
= scalar
+ scaldims
[level
+ scalar
- 1];
1674 while (temp
!= NULL
)
1675 { /* Then we will appy the general code generation process to each sub-list
1676 * of loops with the same scalar vector.
1681 while((end
->next
!= NULL
) &&
1682 cloog_loop_more(end
->next
, level
, scalar_new
, nb_scattdims
) &&
1683 cloog_loop_scalar_eq(ref
,end
->next
,level
,scaldims
,nb_scattdims
,scalar
))
1689 /* For the next dimension, scalar value is updated by adding the scalar
1690 * vector size, which is stored at scaldims[level+scalar-1].
1692 if (cloog_loop_more(temp
, level
, scalar_new
, nb_scattdims
)) {
1693 l
= cloog_loop_generate_restricted(temp
, level
, scalar_new
,
1694 scaldims
, nb_scattdims
, options
);
1697 cloog_loop_add_list(&res
, &now
, l
);
1699 cloog_loop_add(&res
, &now
, temp
);
1708 /* Compare loop with the next loop based on their constant dimensions.
1709 * The result is < 0, == 0 or > 0 depending on whether the constant
1710 * dimensions of loop are lexicographically smaller, equal or greater
1711 * than those of loop->next.
1712 * If loop is the last in the list, then it is assumed to be smaller
1713 * than the "next" one.
1715 static int cloog_loop_next_scal_cmp(CloogLoop
*loop
)
1723 nb_scaldims
= loop
->block
->nb_scaldims
;
1724 if (loop
->next
->block
->nb_scaldims
< nb_scaldims
)
1725 nb_scaldims
= loop
->next
->block
->nb_scaldims
;
1727 for (i
= 0; i
< nb_scaldims
; ++i
) {
1728 int cmp
= cloog_int_cmp(loop
->block
->scaldims
[i
],
1729 loop
->next
->block
->scaldims
[i
]);
1733 return loop
->block
->nb_scaldims
- loop
->next
->block
->nb_scaldims
;
1737 /* Check whether the globally constant dimensions of a and b
1738 * have the same value for all globally constant dimensions
1739 * that are situated before any (locally) non-constant dimension.
1741 static int cloog_loop_equal_prefix(CloogLoop
*a
, CloogLoop
*b
,
1742 int *scaldims
, int nb_scattdims
)
1748 for (i
= 0; i
< nb_scattdims
; ++i
) {
1753 if (!cloog_int_eq(a
->block
->scaldims
[cst
], b
->block
->scaldims
[cst
]))
1757 for (i
= i
+ 1; i
< nb_scattdims
; ++i
) {
1760 if (!cloog_domain_lazy_isconstant(a
->domain
, dim
))
1762 /* No need to check that dim is also constant in b and that the
1763 * constant values are equal. That will happen during the check
1764 * whether the two domains are equal.
1772 /* Try to block adjacent loops in the loop list "loop".
1773 * We only attempt blocking if the constant dimensions of the loops
1774 * in the least are (not necessarily strictly) increasing.
1775 * Then we look for a sublist such that the first (begin) has constant
1776 * dimensions strictly larger than the previous loop in the complete
1777 * list and such that the loop (end) after the last loop in the sublist
1778 * has constant dimensions strictly larger than the last loop in the sublist.
1779 * Furthermore, all loops in the sublist should have the same domain
1780 * (with globally constant dimensions removed) and the difference
1781 * (if any) in constant dimensions may only occur after all the
1782 * (locally) constant dimensions.
1783 * If we find such a sublist, then the blocks of all but the first
1784 * are merged into the block of the first.
1786 * Note that this function can only be called before the global
1787 * blocklist has been created because it may otherwise modify and destroy
1788 * elements on that list.
1790 CloogLoop
*cloog_loop_block(CloogLoop
*loop
, int *scaldims
, int nb_scattdims
)
1792 CloogLoop
*begin
, *end
, *l
;
1793 int begin_after_previous
;
1794 int end_after_previous
;
1798 for (begin
= loop
; begin
; begin
= begin
->next
) {
1799 if (!begin
->block
|| !begin
->block
->scaldims
)
1801 if (cloog_loop_next_scal_cmp(begin
) > 0)
1805 begin_after_previous
= 1;
1806 for (begin
= loop
; begin
; begin
= begin
->next
) {
1807 if (!begin_after_previous
) {
1808 begin_after_previous
= cloog_loop_next_scal_cmp(begin
) < 0;
1812 end_after_previous
= cloog_loop_next_scal_cmp(begin
) < 0;
1813 for (end
= begin
->next
; end
; end
= end
->next
) {
1814 if (!cloog_loop_equal_prefix(begin
, end
, scaldims
, nb_scattdims
))
1816 if (!cloog_domain_lazy_equal(begin
->domain
, end
->domain
))
1818 end_after_previous
= cloog_loop_next_scal_cmp(end
) < 0;
1820 if (end
!= begin
->next
&& end_after_previous
) {
1821 for (l
= begin
->next
; l
!= end
; l
= begin
->next
) {
1822 cloog_block_merge(begin
->block
, l
->block
);
1823 begin
->next
= l
->next
;
1824 cloog_loop_free_parts(l
, 1, 0, 1, 0);
1828 begin_after_previous
= cloog_loop_next_scal_cmp(begin
) < 0;
1836 * Check whether for any fixed iteration of the outer loops,
1837 * there is an iteration of loop1 that is lexicographically greater
1838 * than an iteration of loop2.
1839 * Return 1 if there exists (or may exist) such a pair.
1840 * Return 0 if all iterations of loop1 are lexicographically smaller
1841 * than the iterations of loop2.
1842 * If no iteration is lexicographically greater, but if there are
1843 * iterations that are equal to iterations of loop2, then return "def".
1844 * This is useful for ensuring that such statements are not reordered.
1845 * Some users, including the test_run target in test, expect
1846 * the statements at a given point to be run in the original order.
1847 * Passing the value "0" for "def" would allow such statements to be reordered
1848 * and would allow for the detection of more components.
1850 int cloog_loop_follows(CloogLoop
*loop1
, CloogLoop
*loop2
,
1851 int level
, int scalar
, int *scaldims
, int nb_scattdims
, int def
)
1855 dim1
= cloog_domain_dimension(loop1
->domain
);
1856 dim2
= cloog_domain_dimension(loop2
->domain
);
1857 while ((level
<= dim1
&& level
<= dim2
) ||
1858 level_is_constant(level
, scalar
, scaldims
, nb_scattdims
)) {
1859 if (level_is_constant(level
, scalar
, scaldims
, nb_scattdims
)) {
1860 int cmp
= cloog_loop_constant_cmp(loop1
, loop2
, level
, scaldims
,
1861 nb_scattdims
, scalar
);
1866 scalar
+= scaldims
[level
+ scalar
- 1];
1868 int follows
= cloog_domain_follows(loop1
->domain
, loop2
->domain
,
1882 /* Structure for representing the nodes in the graph being traversed
1883 * using Tarjan's algorithm.
1884 * index represents the order in which nodes are visited.
1885 * min_index is the index of the root of a (sub)component.
1886 * on_stack indicates whether the node is currently on the stack.
1888 struct cloog_loop_sort_node
{
1893 /* Structure for representing the graph being traversed
1894 * using Tarjan's algorithm.
1895 * len is the number of nodes
1896 * node is an array of nodes
1897 * stack contains the nodes on the path from the root to the current node
1898 * sp is the stack pointer
1899 * index is the index of the last node visited
1900 * order contains the elements of the components separated by -1
1901 * op represents the current position in order
1903 struct cloog_loop_sort
{
1905 struct cloog_loop_sort_node
*node
;
1913 /* Allocate and initialize cloog_loop_sort structure.
1915 static struct cloog_loop_sort
*cloog_loop_sort_alloc(int len
)
1917 struct cloog_loop_sort
*s
;
1920 s
= (struct cloog_loop_sort
*)malloc(sizeof(struct cloog_loop_sort
));
1923 s
->node
= (struct cloog_loop_sort_node
*)
1924 malloc(len
* sizeof(struct cloog_loop_sort_node
));
1926 for (i
= 0; i
< len
; ++i
)
1927 s
->node
[i
].index
= -1;
1928 s
->stack
= (int *)malloc(len
* sizeof(int));
1930 s
->order
= (int *)malloc(2 * len
* sizeof(int));
1940 /* Free cloog_loop_sort structure.
1942 static void cloog_loop_sort_free(struct cloog_loop_sort
*s
)
1951 /* Check whether for any fixed iteration of the outer loops,
1952 * there is an iteration of loop1 that is lexicographically greater
1953 * than an iteration of loop2, where the iteration domains are
1954 * available in the inner loops of the arguments.
1956 * By using this functions to detect components, we ensure that
1957 * two CloogLoops appear in the same component if some iterations of
1958 * each loop should be executed before some iterations of the other loop.
1959 * Since we also want two CloogLoops that have exactly the same
1960 * iteration domain at the current level to be placed in the same component,
1961 * we first check if these domains are indeed the same.
1963 static int inner_loop_follows(CloogLoop
*loop1
, CloogLoop
*loop2
,
1964 int level
, int scalar
, int *scaldims
, int nb_scattdims
, int def
)
1968 f
= cloog_domain_lazy_equal(loop1
->domain
, loop2
->domain
);
1970 f
= cloog_loop_follows(loop1
->inner
, loop2
->inner
,
1971 level
, scalar
, scaldims
, nb_scattdims
, def
);
1977 /* Perform Tarjan's algorithm for computing the strongly connected components
1978 * in the graph with the individual CloogLoops as vertices.
1979 * Two CloopLoops appear in the same component if they both (indirectly)
1980 * "follow" each other, where the following relation is determined
1981 * by the follows function.
1983 static void cloog_loop_components_tarjan(struct cloog_loop_sort
*s
,
1984 CloogLoop
**loop_array
, int i
, int level
, int scalar
, int *scaldims
,
1986 int (*follows
)(CloogLoop
*loop1
, CloogLoop
*loop2
,
1987 int level
, int scalar
, int *scaldims
, int nb_scattdims
, int def
))
1991 s
->node
[i
].index
= s
->index
;
1992 s
->node
[i
].min_index
= s
->index
;
1993 s
->node
[i
].on_stack
= 1;
1995 s
->stack
[s
->sp
++] = i
;
1997 for (j
= s
->len
- 1; j
>= 0; --j
) {
2002 if (s
->node
[j
].index
>= 0 &&
2003 (!s
->node
[j
].on_stack
||
2004 s
->node
[j
].index
> s
->node
[i
].min_index
))
2007 f
= follows(loop_array
[i
], loop_array
[j
],
2008 level
, scalar
, scaldims
, nb_scattdims
, i
> j
);
2012 if (s
->node
[j
].index
< 0) {
2013 cloog_loop_components_tarjan(s
, loop_array
, j
, level
, scalar
,
2014 scaldims
, nb_scattdims
, follows
);
2015 if (s
->node
[j
].min_index
< s
->node
[i
].min_index
)
2016 s
->node
[i
].min_index
= s
->node
[j
].min_index
;
2017 } else if (s
->node
[j
].index
< s
->node
[i
].min_index
)
2018 s
->node
[i
].min_index
= s
->node
[j
].index
;
2021 if (s
->node
[i
].index
!= s
->node
[i
].min_index
)
2025 j
= s
->stack
[--s
->sp
];
2026 s
->node
[j
].on_stack
= 0;
2027 s
->order
[s
->op
++] = j
;
2029 s
->order
[s
->op
++] = -1;
2033 static int qsort_index_cmp(const void *p1
, const void *p2
)
2035 return *(int *)p1
- *(int *)p2
;
2038 /* Sort the elements of the component starting at list.
2039 * The list is terminated by a -1.
2041 static void sort_component(int *list
)
2045 for (len
= 0; list
[len
] != -1; ++len
)
2048 qsort(list
, len
, sizeof(int), qsort_index_cmp
);
2051 /* Given an array of indices "list" into the "loop_array" array,
2052 * terminated by -1, construct a linked list of the corresponding
2053 * entries and put the result in *res.
2054 * The value returned is the number of CloogLoops in the (linked) list
2056 static int extract_component(CloogLoop
**loop_array
, int *list
, CloogLoop
**res
)
2060 sort_component(list
);
2061 while (list
[i
] != -1) {
2062 *res
= loop_array
[list
[i
]];
2063 res
= &(*res
)->next
;
2073 * Call cloog_loop_generate_scalar or cloog_loop_generate_general
2074 * on each of the strongly connected components in the list of CloogLoops
2075 * pointed to by "loop".
2077 * We use Tarjan's algorithm to find the strongly connected components.
2078 * Note that this algorithm also topologically sorts the components.
2080 * The components are treated separately to avoid spurious separations.
2081 * The concatentation of the results may contain successive loops
2082 * with the same bounds, so we try to combine such loops.
2084 CloogLoop
*cloog_loop_generate_components(CloogLoop
*loop
,
2085 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
2086 CloogOptions
*options
)
2090 CloogLoop
*res
, **res_next
;
2091 CloogLoop
**loop_array
;
2092 struct cloog_loop_sort
*s
;
2094 if (level
== 0 || !loop
->next
)
2095 return cloog_loop_generate_general(loop
, level
, scalar
,
2096 scaldims
, nb_scattdims
, options
);
2098 nb_loops
= cloog_loop_count(loop
);
2100 loop_array
= (CloogLoop
**)malloc(nb_loops
* sizeof(CloogLoop
*));
2103 for (i
= 0, tmp
= loop
; i
< nb_loops
; i
++, tmp
= tmp
->next
)
2104 loop_array
[i
] = tmp
;
2106 s
= cloog_loop_sort_alloc(nb_loops
);
2107 for (i
= nb_loops
- 1; i
>= 0; --i
) {
2108 if (s
->node
[i
].index
>= 0)
2110 cloog_loop_components_tarjan(s
, loop_array
, i
, level
, scalar
, scaldims
,
2111 nb_scattdims
, &inner_loop_follows
);
2118 int n
= extract_component(loop_array
, &s
->order
[i
], &tmp
);
2121 *res_next
= cloog_loop_generate_general(tmp
, level
, scalar
,
2122 scaldims
, nb_scattdims
, options
);
2124 res_next
= &(*res_next
)->next
;
2127 cloog_loop_sort_free(s
);
2131 res
= cloog_loop_combine(res
);
2137 /* For each loop in the list "loop", decompose the list of
2138 * inner loops into strongly connected components and put
2139 * the components into separate loops at the top level.
2141 CloogLoop
*cloog_loop_decompose_inner(CloogLoop
*loop
,
2142 int level
, int scalar
, int *scaldims
, int nb_scattdims
)
2145 CloogLoop
**loop_array
;
2146 int i
, n_loops
, max_loops
= 0;
2147 struct cloog_loop_sort
*s
;
2149 for (l
= loop
; l
; l
= l
->next
) {
2150 n_loops
= cloog_loop_count(l
->inner
);
2151 if (max_loops
< n_loops
)
2152 max_loops
= n_loops
;
2158 loop_array
= (CloogLoop
**)malloc(max_loops
* sizeof(CloogLoop
*));
2161 for (l
= loop
; l
; l
= l
->next
) {
2164 for (i
= 0, tmp
= l
->inner
; tmp
; i
++, tmp
= tmp
->next
)
2165 loop_array
[i
] = tmp
;
2170 s
= cloog_loop_sort_alloc(n_loops
);
2171 for (i
= n_loops
- 1; i
>= 0; --i
) {
2172 if (s
->node
[i
].index
>= 0)
2174 cloog_loop_components_tarjan(s
, loop_array
, i
, level
, scalar
,
2175 scaldims
, nb_scattdims
, &cloog_loop_follows
);
2178 n
= extract_component(loop_array
, s
->order
, &l
->inner
);
2184 n
= extract_component(loop_array
, &s
->order
[i
], &inner
);
2187 tmp
= cloog_loop_alloc(l
->state
, cloog_domain_copy(l
->domain
),
2188 l
->otl
, l
->stride
, l
->block
, inner
, l
->next
);
2193 cloog_loop_sort_free(s
);
2202 CloogLoop
*cloog_loop_generate_restricted(CloogLoop
*loop
,
2203 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
2204 CloogOptions
*options
)
2206 /* To save both time and memory, we switch here depending on whether the
2207 * current dimension is scalar (simplified processing) or not (general
2210 if (level_is_constant(level
, scalar
, scaldims
, nb_scattdims
))
2211 return cloog_loop_generate_scalar(loop
, level
, scalar
,
2212 scaldims
, nb_scattdims
, options
);
2214 * 2. Compute the projection of each polyhedron onto the outermost
2215 * loop variable and the parameters.
2217 loop
= cloog_loop_project_all(loop
, level
);
2219 return cloog_loop_generate_components(loop
, level
, scalar
, scaldims
,
2220 nb_scattdims
, options
);
2224 CloogLoop
*cloog_loop_generate_restricted_or_stop(CloogLoop
*loop
,
2225 CloogDomain
*context
,
2226 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
2227 CloogOptions
*options
)
2229 /* If the user asked to stop code generation at this level, let's stop. */
2230 if ((options
->stop
>= 0) && (level
+scalar
>= options
->stop
+1))
2231 return cloog_loop_stop(loop
,context
) ;
2233 return cloog_loop_generate_restricted(loop
, level
, scalar
, scaldims
,
2234 nb_scattdims
, options
);
2239 * cloog_loop_generate function:
2240 * Adaptation from LoopGen 0.4 by F. Quillere. This function implements the
2241 * Quillere algorithm for polyhedron scanning from step 1 to 2.
2242 * (see the Quillere paper).
2243 * - loop is the loop for which we have to generate a scanning code,
2244 * - context is the context of the current loop (constraints on parameter and/or
2245 * on outer loop counters),
2246 * - level is the current non-scalar dimension,
2247 * - scalar is the current scalar dimension,
2248 * - scaldims is the boolean array saying whether a dimension is scalar or not,
2249 * - nb_scattdims is the size of the scaldims array,
2250 * - options are the general code generation options.
2252 * - October 26th 2001: first version.
2253 * - July 3rd->11th 2003: memory leaks hunt and correction.
2254 * - June 15th 2005: a memory leak fixed (loop was not entirely freed when
2255 * the result of cloog_loop_restrict was NULL).
2256 * - June 22nd 2005: Adaptation for GMP.
2257 * - September 2nd 2005: The function have been cutted out in two pieces:
2258 * cloog_loop_generate and this one, in order to handle
2259 * the scalar dimension case more efficiently with
2260 * cloog_loop_generate_scalar.
2261 * - November 15th 2005: (debug) Condition for stop option no more take care of
2262 * further scalar dimensions.
2264 CloogLoop
*cloog_loop_generate(CloogLoop
*loop
, CloogDomain
*context
,
2265 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
2266 CloogOptions
*options
)
2268 /* 1. Replace each polyhedron by its intersection with the context.
2270 loop
= cloog_loop_restrict_all(loop
, context
);
2274 return cloog_loop_generate_restricted_or_stop(loop
, context
,
2275 level
, scalar
, scaldims
, nb_scattdims
, options
);
2280 * Internal function for simplifying a single loop in a list of loops.
2281 * See cloog_loop_simplify.
2283 static CloogLoop
*loop_simplify(CloogLoop
*loop
, CloogDomain
*context
,
2284 int level
, CloogOptions
*options
)
2287 CloogBlock
* new_block
;
2288 CloogLoop
*simplified
, *inner
;
2289 CloogDomain
* domain
, * simp
, * inter
, * extended_context
;
2291 if (!cloog_domain_isconvex(loop
->domain
))
2292 loop
->domain
= cloog_domain_simplify_union(loop
->domain
);
2294 domain
= loop
->domain
;
2296 domain_dim
= cloog_domain_dimension(domain
);
2297 extended_context
= cloog_domain_extend(context
, domain_dim
);
2298 inter
= cloog_domain_intersection(domain
,extended_context
) ;
2299 simp
= cloog_domain_simplify(inter
,extended_context
) ;
2300 cloog_domain_free(extended_context
) ;
2302 /* If the constraint system is never true, go to the next one. */
2303 if (cloog_domain_never_integral(simp
)) {
2304 cloog_loop_free(loop
->inner
);
2305 cloog_domain_free(inter
);
2306 cloog_domain_free(simp
);
2310 inner
= cloog_loop_simplify(loop
->inner
, inter
, level
+1, options
);
2312 if ((inner
== NULL
) && (loop
->block
== NULL
)) {
2313 cloog_domain_free(inter
);
2314 cloog_domain_free(simp
);
2318 new_block
= cloog_block_copy(loop
->block
) ;
2320 simplified
= cloog_loop_alloc(loop
->state
, simp
, loop
->otl
, loop
->stride
,
2321 new_block
, inner
, NULL
);
2323 if (loop
->block
&& options
->save_domains
)
2324 simplified
->unsimplified
= inter
;
2326 cloog_domain_free(inter
);
2328 return(simplified
) ;
2333 * cloog_loop_simplify function:
2334 * This function implements the part 6. of the Quillere algorithm, it
2335 * recursively simplifies each loop in the context of the preceding loop domain.
2336 * It returns a pointer to the simplified loop list.
2337 * The cloog_domain_simplify (DomainSimplify) behaviour is really bad with
2338 * polyhedra union and some really awful sidesteppings were written, I plan
2340 * - October 31th 2001: first version.
2341 * - July 3rd->11th 2003: memory leaks hunt and correction.
2342 * - April 16th 2005: a memory leak fixed (extended_context was not freed).
2343 * - June 15th 2005: a memory leak fixed (loop was not conveniently freed
2344 * when the constraint system is never true).
2345 * - October 27th 2005: - this function called before cloog_loop_fast_simplify
2346 * is now the official cloog_loop_simplify function in
2347 * replacement of a slower and more complex one (after
2348 * deep changes in the pretty printer).
2349 * - we use cloog_loop_disjoint to fix the problem when
2350 * simplifying gives a union of polyhedra (before, it
2351 * was under the responsibility of the pretty printer).
2353 CloogLoop
*cloog_loop_simplify(CloogLoop
*loop
, CloogDomain
*context
, int level
,
2354 CloogOptions
*options
)
2357 CloogLoop
*res
= NULL
;
2358 CloogLoop
**next
= &res
;
2360 for (now
= loop
; now
; now
= now
->next
) {
2361 *next
= loop_simplify(now
, context
, level
, options
);
2363 now
->inner
= NULL
; /* For loop integrity. */
2364 cloog_domain_free(now
->domain
);
2368 next
= &(*next
)->next
;
2370 cloog_loop_free(loop
);
2372 /* Examples like test/iftest2.cloog give unions of polyhedra after
2373 * simplifying, thus we have to make them disjoint. Another good reason to
2374 * put the simplifying step in the Quillere backtrack.
2376 res
= cloog_loop_disjoint(res
);
2383 * cloog_loop_scatter function:
2384 * This function add the scattering (scheduling) informations in a loop.
2386 void cloog_loop_scatter(CloogLoop
* loop
, CloogScattering
*scatt
)
2388 loop
->domain
= cloog_domain_scatter(loop
->domain
, scatt
);