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_block_free(loop
->block
) ;
212 if (loop
->inner
!= NULL
)
213 cloog_loop_free(loop
->inner
) ;
215 cloog_stride_free(loop
->stride
);
223 * cloog_loop_free_parts function:
224 * This function frees the allocated memory for some parts of a CloogLoop
225 * structure (loop), each other argument is a boolean having to be set to 1 if
226 * we want to free the corresponding part, 0 otherwise. This function applies
227 * the same freeing policy to its inner ans next loops recursively.
228 * - July 3rd 2003: first version.
229 * - June 22nd 2005: Adaptation for GMP.
231 void cloog_loop_free_parts(loop
, domain
, block
, inner
, next
)
233 int domain
, block
, inner
, next
;
234 { CloogLoop
* follow
;
236 while (loop
!= NULL
) {
237 cloog_loop_leak_down(loop
->state
);
238 follow
= loop
->next
;
241 cloog_domain_free(loop
->domain
) ;
244 cloog_block_free(loop
->block
) ;
246 if ((inner
) && (loop
->inner
!= NULL
))
247 cloog_loop_free_parts(loop
->inner
,domain
,block
,inner
,1) ;
249 cloog_stride_free(loop
->stride
);
259 /******************************************************************************
260 * Reading functions *
261 ******************************************************************************/
265 * Construct a CloogLoop structure from a given iteration domain
266 * and statement number.
268 CloogLoop
*cloog_loop_from_domain(CloogState
*state
, CloogDomain
*domain
,
273 CloogStatement
* statement
;
275 /* Memory allocation and information reading for the first domain: */
276 loop
= cloog_loop_malloc(state
);
278 loop
->domain
= domain
;
279 if (loop
->domain
!= NULL
)
280 nb_iterators
= cloog_domain_dimension(loop
->domain
);
283 /* included statement block. */
284 statement
= cloog_statement_alloc(state
, number
+ 1);
285 loop
->block
= cloog_block_alloc(statement
, 0, NULL
, nb_iterators
);
292 * cloog_loop_read function:
293 * This function reads loop data from a file (foo, possibly stdin) and
294 * returns a pointer to a CloogLoop structure containing the read information.
295 * This function can be used only for input file reading, when one loop is
296 * associated with one statement.
297 * - number is the statement block number carried by the loop (-1 if none).
298 * - nb_parameters is the number of parameters.
300 * - September 9th 2002: first version.
301 * - April 16th 2005: adaptation to new CloogStatement struct (with number).
302 * - June 11th 2005: adaptation to new CloogBlock structure.
303 * - June 22nd 2005: Adaptation for GMP.
305 CloogLoop
*cloog_loop_read(CloogState
*state
,
306 FILE *foo
, int number
, int nb_parameters
)
312 domain
= cloog_domain_union_read(state
, foo
, nb_parameters
);
314 /* To read that stupid "0 0 0" line. */
315 while (fgets(s
,MAX_STRING
,foo
) == 0) ;
316 while ((*s
=='#' || *s
=='\n') || (sscanf(s
," %d %d %d",&op1
,&op2
,&op3
)<3))
317 fgets(s
,MAX_STRING
,foo
) ;
319 return cloog_loop_from_domain(state
, domain
, number
);
323 /******************************************************************************
324 * Processing functions *
325 ******************************************************************************/
329 * cloog_loop_malloc function:
330 * This function allocates the memory space for a CloogLoop structure and
331 * sets its fields with default values. Then it returns a pointer to the
333 * - November 21th 2005: first version.
335 CloogLoop
*cloog_loop_malloc(CloogState
*state
)
338 /* Memory allocation for the CloogLoop structure. */
339 loop
= (CloogLoop
*)malloc(sizeof(CloogLoop
)) ;
341 cloog_die("memory overflow.\n");
342 cloog_loop_leak_up(state
);
345 /* We set the various fields with default values. */
347 loop
->domain
= NULL
;
360 * cloog_loop_alloc function:
361 * This function allocates the memory space for a CloogLoop structure and
362 * sets its fields with those given as input. Then it returns a pointer to the
364 * - October 27th 2001: first version.
365 * - June 22nd 2005: Adaptation for GMP.
366 * - November 21th 2005: use of cloog_loop_malloc.
368 CloogLoop
*cloog_loop_alloc(CloogState
*state
,
369 CloogDomain
*domain
, int otl
, CloogStride
*stride
,
370 CloogBlock
*block
, CloogLoop
*inner
, CloogLoop
*next
)
373 loop
= cloog_loop_malloc(state
);
375 loop
->domain
= domain
;
376 loop
->block
= block
;
377 loop
->inner
= inner
;
380 loop
->stride
= cloog_stride_copy(stride
);
387 * cloog_loop_add function:
388 * This function adds a CloogLoop structure (loop) at a given place (now) of a
389 * NULL terminated list of CloogLoop structures. The beginning of this list
390 * is (start). This function updates (now) to (loop), and updates (start) if the
391 * added element is the first one -that is when (start) is NULL-.
392 * - October 28th 2001: first version.
394 void cloog_loop_add(CloogLoop
** start
, CloogLoop
** now
, CloogLoop
* loop
)
395 { if (*start
== NULL
)
400 { (*now
)->next
= loop
;
401 *now
= (*now
)->next
;
407 * cloog_loop_add function:
408 * This function adds a CloogLoop structure (loop) at a given place (now) of a
409 * NULL terminated list of CloogLoop structures. The beginning of this list
410 * is (start). This function updates (now) to the end of the loop list (loop),
411 * and updates (start) if the added element is the first one -that is when
413 * - September 9th 2005: first version.
415 void cloog_loop_add_list(CloogLoop
** start
, CloogLoop
** now
, CloogLoop
* loop
)
416 { if (*start
== NULL
)
421 { (*now
)->next
= loop
;
422 *now
= (*now
)->next
;
425 while ((*now
)->next
!= NULL
)
426 *now
= (*now
)->next
;
431 * cloog_loop_copy function:
432 * This function returns a copy of the CloogLoop structure given as input. In
433 * fact, there is just new allocations for the CloogLoop structures, but their
434 * contents are the same.
435 * - October 28th 2001: first version.
436 * - July 3rd->11th 2003: memory leaks hunt and correction.
438 CloogLoop
* cloog_loop_copy(CloogLoop
* source
)
441 CloogDomain
* domain
;
445 { domain
= cloog_domain_copy(source
->domain
) ;
446 block
= cloog_block_copy(source
->block
) ;
447 loop
= cloog_loop_alloc(source
->state
, domain
, source
->otl
,
448 source
->stride
, block
, NULL
, NULL
);
449 loop
->usr
= source
->usr
;
450 loop
->inner
= cloog_loop_copy(source
->inner
) ;
451 loop
->next
= cloog_loop_copy(source
->next
) ;
458 * cloog_loop_add_disjoint function:
459 * This function adds some CloogLoop structures at a given place (now) of a
460 * NULL terminated list of CloogLoop structures. The beginning of this list
461 * is (start). (loop) can be an union of polyhedra, this function separates the
462 * union into a list of *disjoint* polyhedra then adds the list. This function
463 * updates (now) to the end of the list and updates (start) if first added
464 * element is the first of the principal list -that is when (start) is NULL-.
465 * (loop) can be freed by this function, basically when its domain is actually
466 * a union of polyhedra, but don't worry, all the useful data are now stored
467 * inside the list (start). We do not use PolyLib's Domain_Disjoint function,
468 * since the number of union components is often higher (thus code size too).
469 * - October 28th 2001: first version.
470 * - November 14th 2001: bug correction (this one was hard to find !).
471 * - July 3rd->11th 2003: memory leaks hunt and correction.
472 * - June 22nd 2005: Adaptation for GMP.
473 * - October 27th 2005: (debug) included blocks were not copied for new loops.
475 void cloog_loop_add_disjoint(start
, now
, loop
)
476 CloogLoop
** start
, ** now
, * loop
;
478 CloogLoop
* sep
, * inner
;
479 CloogDomain
*domain
, *seen
, *temp
, *rest
;
482 if (cloog_domain_isconvex(loop
->domain
))
483 cloog_loop_add(start
,now
,loop
) ;
485 domain
= cloog_domain_simplify_union(loop
->domain
);
486 loop
->domain
= NULL
;
488 /* We separate the first element of the rest of the union. */
489 domain
= cloog_domain_cut_first(domain
, &rest
);
491 /* This first element is the first of the list of disjoint polyhedra. */
492 sep
= cloog_loop_alloc(loop
->state
, domain
, 0, NULL
,
493 loop
->block
, loop
->inner
, NULL
);
494 cloog_loop_add(start
,now
,sep
) ;
496 seen
= cloog_domain_copy(domain
);
497 while (!cloog_domain_isempty(domain
= rest
)) {
498 temp
= cloog_domain_cut_first(domain
, &rest
);
499 domain
= cloog_domain_difference(temp
, seen
);
500 cloog_domain_free(temp
);
502 if (cloog_domain_isempty(domain
)) {
503 cloog_domain_free(domain
);
507 /* Each new loop will have its own life, for instance we can free its
508 * inner loop and included block. Then each one must have its own copy
509 * of both 'inner' and 'block'.
511 inner
= cloog_loop_copy(loop
->inner
) ;
512 block
= cloog_block_copy(loop
->block
) ;
514 sep
= cloog_loop_alloc(loop
->state
, cloog_domain_copy(domain
),
515 0, NULL
, block
, inner
, NULL
);
516 /* domain can be an union too. If so: recursion. */
517 if (cloog_domain_isconvex(domain
))
518 cloog_loop_add(start
,now
,sep
) ;
520 cloog_loop_add_disjoint(start
,now
,sep
) ;
522 if (cloog_domain_isempty(rest
)) {
523 cloog_domain_free(domain
);
527 seen
= cloog_domain_union(seen
, domain
);
529 cloog_domain_free(rest
);
530 cloog_domain_free(seen
);
531 cloog_loop_free_parts(loop
,0,0,0,0) ;
537 * cloog_loop_disjoint function:
538 * This function returns a list of loops such that each loop with non-convex
539 * domain in the input list (loop) is separated into several loops where the
540 * domains are the components of the union of *disjoint* polyhedra equivalent
541 * to the original non-convex domain. See cloog_loop_add_disjoint comments
543 * - September 16th 2005: first version.
545 CloogLoop
* cloog_loop_disjoint(CloogLoop
* loop
)
546 { CloogLoop
*res
=NULL
, * now
=NULL
, * next
;
548 /* Because this is often the case, don't waste time ! */
549 if (loop
&& !loop
->next
&& cloog_domain_isconvex(loop
->domain
))
553 { next
= loop
->next
;
555 cloog_loop_add_disjoint(&res
,&now
,loop
) ;
564 * cloog_loop_restrict function:
565 * This function returns the (loop) in the context of (context): it makes the
566 * intersection between the (loop) domain and the (context), then it returns
567 * a pointer to a new loop, with this intersection as domain.
569 * - October 27th 2001: first version.
570 * - June 15th 2005: a memory leak fixed (domain was not freed when empty).
571 * - June 22nd 2005: Adaptation for GMP.
573 CloogLoop
*cloog_loop_restrict(CloogLoop
*loop
, CloogDomain
*context
)
574 { int new_dimension
;
575 CloogDomain
* domain
, * extended_context
, * new_domain
;
576 CloogLoop
* new_loop
;
578 domain
= loop
->domain
;
579 if (cloog_domain_dimension(domain
) > cloog_domain_dimension(context
))
581 new_dimension
= cloog_domain_dimension(domain
);
582 extended_context
= cloog_domain_extend(context
, new_dimension
);
583 new_domain
= cloog_domain_intersection(extended_context
,loop
->domain
) ;
584 cloog_domain_free(extended_context
) ;
587 new_domain
= cloog_domain_intersection(context
,loop
->domain
) ;
589 if (cloog_domain_isempty(new_domain
))
590 { cloog_domain_free(new_domain
) ;
594 new_loop
= cloog_loop_alloc(loop
->state
, new_domain
,
595 0, NULL
, loop
->block
, loop
->inner
, NULL
);
602 * Call cloog_loop_restrict on each loop in the list "loop" and return
603 * the concatenated result.
605 CloogLoop
*cloog_loop_restrict_all(CloogLoop
*loop
, CloogDomain
*context
)
608 CloogLoop
*res
= NULL
;
609 CloogLoop
**res_next
= &res
;
611 for (; loop
; loop
= next
) {
614 *res_next
= cloog_loop_restrict(loop
, context
);
616 res_next
= &(*res_next
)->next
;
617 cloog_loop_free_parts(loop
, 1, 0, 0, 0);
620 cloog_loop_free(loop
);
627 CloogLoop
*cloog_loop_restrict_inner(CloogLoop
*loop
)
631 for (l
= loop
; l
; l
= l
->next
)
632 l
->inner
= cloog_loop_restrict_all(l
->inner
, l
->domain
);
638 * cloog_loop_project function:
639 * This function returns the projection of (loop) on the (level) first
640 * dimensions (outer loops). It makes the projection of the (loop) domain,
641 * then it returns a pointer to a new loop, with this projection as domain.
643 * - October 27th 2001: first version.
644 * - July 3rd->11th 2003: memory leaks hunt and correction.
645 * - June 22nd 2005: Adaptation for GMP.
647 CloogLoop
* cloog_loop_project(CloogLoop
* loop
, int level
)
649 CloogDomain
* new_domain
;
650 CloogLoop
* new_loop
, * copy
;
652 copy
= cloog_loop_alloc(loop
->state
, loop
->domain
, loop
->otl
, loop
->stride
,
653 loop
->block
, loop
->inner
, NULL
);
655 if (cloog_domain_dimension(loop
->domain
) == level
)
656 new_domain
= cloog_domain_copy(loop
->domain
) ;
658 new_domain
= cloog_domain_project(loop
->domain
, level
);
660 new_loop
= cloog_loop_alloc(loop
->state
, new_domain
, 0, NULL
,
668 * Call cloog_loop_project on each loop in the list "loop" and return
669 * the concatenated result.
671 CloogLoop
*cloog_loop_project_all(CloogLoop
*loop
, int level
)
674 CloogLoop
*res
= NULL
;
675 CloogLoop
**res_next
= &res
;
677 for (; loop
; loop
= next
) {
680 *res_next
= cloog_loop_project(loop
, level
);
681 res_next
= &(*res_next
)->next
;
682 cloog_loop_free_parts(loop
, 0, 0, 0, 0);
690 * cloog_loop_concat function:
691 * This function returns a pointer to the concatenation of the
692 * CloogLoop lists given as input.
693 * - October 28th 2001: first version.
695 CloogLoop
* cloog_loop_concat(CloogLoop
* a
, CloogLoop
* b
)
696 { CloogLoop
* loop
, * temp
;
701 { while (temp
->next
!= NULL
)
713 * cloog_loop_combine:
714 * Combine consecutive loops with identical domains into
715 * a single loop with the concatenation of their inner loops
718 CloogLoop
*cloog_loop_combine(CloogLoop
*loop
)
720 CloogLoop
*first
, *second
;
722 for (first
= loop
; first
; first
= first
->next
) {
723 while (first
->next
) {
724 if (!cloog_domain_lazy_equal(first
->domain
, first
->next
->domain
))
726 second
= first
->next
;
727 first
->inner
= cloog_loop_concat(first
->inner
, second
->inner
);
728 first
->next
= second
->next
;
729 cloog_loop_free_parts(second
, 1, 0, 0, 0);
737 * Remove loops from list that have an empty domain.
739 CloogLoop
*cloog_loop_remove_empty_domain_loops(CloogLoop
*loop
)
741 CloogLoop
*l
, *res
, *next
, **res_next
;
745 for (l
= loop
; l
; l
= next
) {
747 if (cloog_domain_isempty(l
->domain
))
748 cloog_loop_free_parts(l
, 1, 1, 1, 0);
751 res_next
= &(*res_next
)->next
;
759 CloogLoop
*cloog_loop_decompose_inner(CloogLoop
*loop
,
760 int level
, int scalar
, int *scaldims
, int nb_scattdims
);
762 /* For each loop with only one inner loop, replace the domain
763 * of the loop with the projection of the domain of the inner
764 * loop. To increase the number of loops with a single inner
765 * we first decompose the inner loops into strongly connected
768 CloogLoop
*cloog_loop_specialize(CloogLoop
*loop
,
769 int level
, int scalar
, int *scaldims
, int nb_scattdims
)
775 loop
= cloog_loop_decompose_inner(loop
, level
, scalar
,
776 scaldims
, nb_scattdims
);
778 for (l
= loop
; l
; l
= l
->next
) {
781 if (!cloog_domain_isconvex(l
->inner
->domain
))
784 dim
= cloog_domain_dimension(l
->domain
);
785 domain
= cloog_domain_project(l
->inner
->domain
, dim
);
786 if (cloog_domain_isconvex(domain
)) {
787 cloog_domain_free(l
->domain
);
790 cloog_domain_free(domain
);
794 return cloog_loop_remove_empty_domain_loops(loop
);
798 * cloog_loop_separate function:
799 * This function implements the Quillere algorithm for separation of multiple
800 * loops: for a given set of polyhedra (loop), it computes a set of disjoint
801 * polyhedra such that the unions of these sets are equal, and returns this set.
802 * - October 28th 2001: first version.
803 * - November 14th 2001: elimination of some unused blocks.
804 * - August 13th 2002: (debug) in the case of union of polyhedra for one
805 * loop, redundant constraints are fired.
806 * - July 3rd->11th 2003: memory leaks hunt and correction.
807 * - June 22nd 2005: Adaptation for GMP.
808 * - October 16th 2005: Removal of the non-shared constraint elimination when
809 * there is only one loop in the list (seems to work
810 * without now, DomainSimplify may have been improved).
811 * The problem was visible with test/iftest2.cloog.
813 CloogLoop
* cloog_loop_separate(CloogLoop
* loop
)
814 { int lazy_equal
=0, disjoint
= 0;
815 CloogLoop
* new_loop
, * new_inner
, * res
, * now
, * temp
, * Q
,
816 * inner
, * old
/*, * previous, * next*/ ;
817 CloogDomain
*UQ
, *domain
;
822 loop
= cloog_loop_combine(loop
);
824 if (loop
->next
== NULL
)
825 return cloog_loop_disjoint(loop
) ;
827 UQ
= cloog_domain_copy(loop
->domain
) ;
828 domain
= cloog_domain_copy(loop
->domain
) ;
829 res
= cloog_loop_alloc(loop
->state
, domain
, 0, NULL
,
830 loop
->block
, loop
->inner
, NULL
);
833 while((loop
= loop
->next
) != NULL
)
836 /* For all Q, add Q-loop associated with the blocks of Q alone,
837 * and Q inter loop associated with the blocks of Q and loop.
839 for (Q
= res
; Q
; Q
= Q
->next
) {
840 /* Add (Q inter loop). */
841 if ((disjoint
= cloog_domain_lazy_disjoint(Q
->domain
,loop
->domain
)))
844 { if ((lazy_equal
= cloog_domain_lazy_equal(Q
->domain
,loop
->domain
)))
845 domain
= cloog_domain_copy(Q
->domain
) ;
847 domain
= cloog_domain_intersection(Q
->domain
,loop
->domain
) ;
849 if (!cloog_domain_isempty(domain
))
850 { new_inner
= cloog_loop_concat(cloog_loop_copy(Q
->inner
),
851 cloog_loop_copy(loop
->inner
)) ;
852 new_loop
= cloog_loop_alloc(loop
->state
, domain
, 0, NULL
,
853 NULL
, new_inner
, NULL
);
854 cloog_loop_add_disjoint(&temp
,&now
,new_loop
) ;
858 cloog_domain_free(domain
);
862 /* Add (Q - loop). */
864 domain
= cloog_domain_copy(Q
->domain
) ;
867 domain
= cloog_domain_empty(Q
->domain
);
869 domain
= cloog_domain_difference(Q
->domain
,loop
->domain
) ;
872 if (!cloog_domain_isempty(domain
)) {
873 new_loop
= cloog_loop_alloc(loop
->state
, domain
, 0, NULL
,
874 NULL
, Q
->inner
, NULL
);
875 cloog_loop_add_disjoint(&temp
,&now
,new_loop
) ;
878 { cloog_domain_free(domain
) ;
879 /* If Q->inner is no more useful, we can free it. */
882 cloog_loop_free(inner
) ;
886 /* Add loop-UQ associated with the blocks of loop alone.*/
887 if (cloog_domain_lazy_disjoint(loop
->domain
,UQ
))
888 domain
= cloog_domain_copy(loop
->domain
) ;
890 { if (cloog_domain_lazy_equal(loop
->domain
,UQ
))
891 domain
= cloog_domain_empty(UQ
);
893 domain
= cloog_domain_difference(loop
->domain
,UQ
) ;
896 if (!cloog_domain_isempty(domain
)) {
897 new_loop
= cloog_loop_alloc(loop
->state
, domain
, 0, NULL
,
898 NULL
, loop
->inner
, NULL
);
899 cloog_loop_add_disjoint(&temp
,&now
,new_loop
) ;
902 { cloog_domain_free(domain
) ;
903 /* If loop->inner is no more useful, we can free it. */
904 cloog_loop_free(loop
->inner
) ;
909 if (loop
->next
!= NULL
)
910 UQ
= cloog_domain_union(UQ
, cloog_domain_copy(loop
->domain
));
912 cloog_domain_free(UQ
);
914 cloog_loop_free_parts(res
,1,0,0,1) ;
918 cloog_loop_free_parts(old
,1,0,0,1) ;
924 static CloogDomain
*bounding_domain(CloogDomain
*dom
, CloogOptions
*options
)
927 return cloog_domain_simple_convex(dom
);
929 return cloog_domain_convex(dom
);
934 * cloog_loop_merge function:
935 * This function is the 'soft' version of loop_separate if we are looking for
936 * a code much simpler (and less efficicient). This function returns the new
938 * - October 29th 2001: first version.
939 * - July 3rd->11th 2003: memory leaks hunt and correction.
940 * - June 22nd 2005: Adaptation for GMP.
942 CloogLoop
*cloog_loop_merge(CloogLoop
*loop
, int level
, CloogOptions
*options
)
944 CloogLoop
*res
, *new_inner
, *old
;
945 CloogDomain
*new_domain
, *temp
;
950 if (loop
->next
== NULL
)
951 return cloog_loop_disjoint(loop
);
956 new_inner
= loop
->inner
;
958 for (loop
= loop
->next
; loop
; loop
= loop
->next
) {
959 temp
= cloog_domain_union(temp
, loop
->domain
);
961 new_inner
= cloog_loop_concat(new_inner
, loop
->inner
);
964 new_domain
= bounding_domain(temp
, options
);
966 if (level
> 0 && !cloog_domain_is_bounded(new_domain
, level
) &&
967 cloog_domain_is_bounded(temp
, level
)) {
968 CloogDomain
*splitter
, *t2
;
970 cloog_domain_free(new_domain
);
971 splitter
= cloog_domain_bound_splitter(temp
, level
);
974 while (!cloog_domain_isconvex(splitter
)) {
975 CloogDomain
*first
, *rest
;
976 first
= cloog_domain_cut_first(splitter
, &rest
);
978 t2
= cloog_domain_intersection(first
, temp
);
979 cloog_domain_free(first
);
981 new_domain
= bounding_domain(t2
, options
);
982 cloog_domain_free(t2
);
984 if (cloog_domain_isempty(new_domain
)) {
985 cloog_domain_free(new_domain
);
988 res
= cloog_loop_alloc(old
->state
, new_domain
, 0, NULL
,
989 NULL
, cloog_loop_copy(new_inner
), res
);
992 t2
= cloog_domain_intersection(splitter
, temp
);
993 cloog_domain_free(splitter
);
995 new_domain
= bounding_domain(t2
, options
);
996 cloog_domain_free(t2
);
998 if (cloog_domain_isempty(new_domain
)) {
999 cloog_domain_free(new_domain
);
1000 cloog_loop_free(new_inner
);
1002 res
= cloog_loop_alloc(old
->state
, new_domain
, 0, NULL
,
1003 NULL
, new_inner
, res
);
1005 res
= cloog_loop_alloc(old
->state
, new_domain
, 0, NULL
,
1006 NULL
, new_inner
, NULL
);
1008 cloog_domain_free(temp
);
1010 cloog_loop_free_parts(old
, 0, 0, 0, 1);
1016 static int cloog_loop_count(CloogLoop
*loop
)
1020 for (nb_loops
= 0; loop
; loop
= loop
->next
)
1028 * cloog_loop_sort function:
1029 * Adaptation from LoopGen 0.4 by F. Quillere. This function sorts a list of
1030 * parameterized disjoint polyhedra, in order to not have lexicographic order
1031 * violation (see Quillere paper).
1032 * - September 16th 2005: inclusion of cloog_loop_number (October 29th 2001).
1034 CloogLoop
*cloog_loop_sort(CloogLoop
*loop
, int level
)
1036 CloogLoop
*res
, *now
, **loop_array
;
1038 int i
, nb_loops
=0, * permut
;
1040 /* There is no need to sort the parameter domains. */
1044 /* We will need to know how many loops are in the list. */
1045 nb_loops
= cloog_loop_count(loop
);
1047 /* If there is only one loop, it's the end. */
1051 /* We have to allocate memory for some useful components:
1052 * - loop_array: the loop array,
1053 * - doms: the array of domains to sort,
1054 * - permut: will give us a possible sort (maybe not the only one).
1056 loop_array
= (CloogLoop
**)malloc(nb_loops
*sizeof(CloogLoop
*)) ;
1057 doms
= (CloogDomain
**)malloc(nb_loops
*sizeof(CloogDomain
*));
1058 permut
= (int *)malloc(nb_loops
*sizeof(int)) ;
1060 /* We fill up the loop and domain arrays. */
1061 for (i
=0;i
<nb_loops
;i
++,loop
=loop
->next
)
1062 { loop_array
[i
] = loop
;
1063 doms
[i
] = loop_array
[i
]->domain
;
1066 /* cloog_domain_sort will fill up permut. */
1067 cloog_domain_sort(doms
, nb_loops
, level
, permut
);
1069 /* With permut and loop_array we build the sorted list. */
1071 for (i
=0;i
<nb_loops
;i
++)
1072 { /* To avoid pointer looping... loop_add will rebuild the list. */
1073 loop_array
[permut
[i
]-1]->next
= NULL
;
1074 cloog_loop_add(&res
,&now
,loop_array
[permut
[i
]-1]) ;
1086 * cloog_loop_nest function:
1087 * This function changes the loop list in such a way that we have no more than
1088 * one dimension added by level. It returns an equivalent loop list with
1090 * - October 29th 2001: first version.
1091 * - July 3rd->11th 2003: memory leaks hunt and correction.
1092 * - June 22nd 2005: Adaptation for GMP.
1093 * - November 21th 2005: (debug) now OK when cloog_loop_restrict returns NULL.
1095 CloogLoop
*cloog_loop_nest(CloogLoop
*loop
, CloogDomain
*context
, int level
)
1097 CloogLoop
* p
, * temp
, * res
, * now
, * next
;
1098 CloogDomain
* new_domain
;
1100 loop
= cloog_loop_disjoint(loop
);
1103 /* Each domain is changed by its intersection with the context. */
1104 while (loop
!= NULL
)
1105 { p
= cloog_loop_restrict(loop
, context
);
1109 { cloog_loop_free_parts(loop
,1,0,0,0) ;
1111 temp
= cloog_loop_alloc(p
->state
, p
->domain
, 0, NULL
,
1112 p
->block
, p
->inner
, NULL
);
1114 /* If the intersection dimension is too big, we make projections smaller
1115 * and smaller, and each projection includes the preceding projection
1116 * (thus, in the target list, dimensions are added one by one).
1118 if (cloog_domain_dimension(p
->domain
) >= level
)
1119 for (l
= cloog_domain_dimension(p
->domain
); l
>= level
; l
--) {
1120 new_domain
= cloog_domain_project(p
->domain
, l
);
1121 temp
= cloog_loop_alloc(p
->state
, new_domain
, 0, NULL
,
1125 /* p is no more useful (but its content yes !). */
1126 cloog_loop_free_parts(p
,0,0,0,0) ;
1128 cloog_loop_add(&res
,&now
,temp
) ;
1131 cloog_loop_free_parts(loop
,1,1,1,0) ;
1140 /* Check if the domains of the inner loops impose a stride constraint
1141 * on the given level.
1142 * The core of the search is implemented in cloog_domain_list_stride.
1143 * Here, we simply construct a list of domains to pass to this function
1144 * and if a stride is found, we adjust the lower bounds by calling
1145 * cloog_domain_stride_lower_bound.
1147 static int cloog_loop_variable_offset_stride(CloogLoop
*loop
, int level
)
1149 CloogDomainList
*list
= NULL
;
1151 CloogStride
*stride
;
1153 for (inner
= loop
->inner
; inner
; inner
= inner
->next
) {
1154 CloogDomainList
*entry
= ALLOC(CloogDomainList
);
1155 entry
->domain
= cloog_domain_copy(inner
->domain
);
1160 stride
= cloog_domain_list_stride(list
, level
);
1162 cloog_domain_list_free(list
);
1167 loop
->stride
= stride
;
1168 loop
->domain
= cloog_domain_stride_lower_bound(loop
->domain
, level
, stride
);
1175 * cloog_loop_stride function:
1176 * This function will find the stride of a loop for the iterator at the column
1177 * number 'level' in the constraint matrix. It will update the lower bound of
1178 * the iterator accordingly. Basically, the function will try to find in the
1179 * inner loops a common condition on this iterator for the inner loop iterators
1180 * to be integral. For instance, let us consider a loop with the iterator i,
1181 * the iteration domain -4<=i<=n, and its two inner loops with the iterator j.
1182 * The first inner loop has the constraint 3j=i, and the second one has the
1183 * constraint 6j=i. Then the common constraint on i for j to be integral is
1184 * i%3=0, the stride for i is 3. Lastly, we have to find the new lower bound
1185 * for i: the first value satisfying the common constraint: -3. At the end, the
1186 * iteration domain for i is -3<=i<=n and the stride for i is 3.
1188 * The algorithm implemented in this function only allows for strides
1189 * on loops with a lower bound that has a constant remainder on division
1190 * by the stride. Before initiating this procedure, we first check
1191 * if we can find a stride with a lower bound with a variable offset in
1192 * cloog_loop_variable_offset_stride.
1194 * - loop is the loop including the iteration domain of the considered iterator,
1195 * - level is the column number of the iterator in the matrix of contraints.
1197 * - June 29th 2003: first version (work in progress since June 26th 2003).
1198 * - July 14th 2003: simpler version.
1199 * - June 22nd 2005: Adaptation for GMP (from S. Verdoolaege's 0.12.1 version).
1201 void cloog_loop_stride(CloogLoop
* loop
, int level
)
1202 { int first_search
;
1203 cloog_int_t stride
, ref_offset
, offset
, potential
;
1206 if (!cloog_domain_can_stride(loop
->domain
, level
))
1209 if (cloog_loop_variable_offset_stride(loop
, level
))
1212 cloog_int_init(stride
);
1213 cloog_int_init(ref_offset
);
1214 cloog_int_init(offset
);
1215 cloog_int_init(potential
);
1217 cloog_int_set_si(ref_offset
, 0);
1218 cloog_int_set_si(offset
, 0);
1220 /* Default stride. */
1221 cloog_int_set_si(stride
, 1);
1223 inner
= loop
->inner
;
1225 while (inner
!= NULL
)
1226 { /* If the minimun stride has not been found yet, find the stride. */
1227 if ((first_search
) || (!cloog_int_is_one(stride
)))
1229 cloog_domain_stride(inner
->domain
, level
, &potential
, &offset
);
1230 if (!cloog_int_is_one(potential
) && (!first_search
))
1231 { /* Offsets must be the same for common stride. */
1232 cloog_int_gcd(stride
, potential
, stride
);
1233 if (!cloog_int_is_zero(stride
)) {
1234 cloog_int_fdiv_r(offset
, offset
, stride
);
1235 cloog_int_fdiv_r(ref_offset
, ref_offset
, stride
);
1237 if (cloog_int_ne(offset
,ref_offset
))
1238 cloog_int_set_si(stride
, 1);
1241 cloog_int_set(stride
, potential
);
1242 cloog_int_set(ref_offset
, offset
);
1248 inner
= inner
->next
;
1251 if (cloog_int_is_zero(stride
))
1252 cloog_int_set_si(stride
, 1);
1254 /* Update the values if necessary. */
1255 if (!cloog_int_is_one(stride
))
1256 { /* Update the stride value. */
1257 if (!cloog_int_is_zero(offset
))
1258 cloog_int_sub(offset
, stride
, offset
);
1259 loop
->stride
= cloog_stride_alloc(stride
, offset
);
1260 loop
->domain
= cloog_domain_stride_lower_bound(loop
->domain
, level
,
1264 cloog_int_clear(stride
);
1265 cloog_int_clear(ref_offset
);
1266 cloog_int_clear(offset
);
1267 cloog_int_clear(potential
);
1271 void cloog_loop_otl(CloogLoop
*loop
, int level
)
1273 if (cloog_domain_is_otl(loop
->domain
, level
))
1279 * cloog_loop_stop function:
1280 * This function implements the 'stop' option : each domain of each loop
1281 * in the list 'loop' is replaced by 'context'. 'context' should be the
1282 * domain of the outer loop. By using this method, there are no more dimensions
1283 * to scan and the simplification step will automaticaly remove the domains
1284 * since they are the same as the corresponding contexts. The effect of this
1285 * function is to stop the code generation at the level this function is called,
1286 * the resulting code do not consider the next dimensions.
1287 * - January 11th 2005: first version.
1289 CloogLoop
* cloog_loop_stop(CloogLoop
* loop
, CloogDomain
* context
)
1293 { cloog_domain_free(loop
->domain
) ;
1294 loop
->domain
= cloog_domain_copy(context
) ;
1295 loop
->next
= cloog_loop_stop(loop
->next
, context
) ;
1302 static int level_is_constant(int level
, int scalar
, int *scaldims
, int nb_scattdims
)
1304 return level
&& (level
+scalar
<= nb_scattdims
) && (scaldims
[level
+scalar
-1]);
1309 * Compare the constant dimensions of loops 'l1' and 'l2' starting at 'scalar'
1310 * and return -1 if the vector of constant dimensions of 'l1' is smaller
1311 * than that of 'l2', 0 if they are the same and +1 if that of 'l1' is
1312 * greater than that of 'l2'.
1313 * This function should be called on the innermost loop (the loop
1314 * containing a block).
1315 * \param l1 Loop to be compared with l2.
1316 * \param l2 Loop to be compared with l1.
1317 * \param level Current non-scalar dimension.
1318 * \param scaldims Boolean array saying whether a dimension is scalar or not.
1319 * \param nb_scattdims Size of the scaldims array.
1320 * \param scalar Current scalar dimension.
1321 * \return -1 if (l1 < l2), 0 if (l1 == l2) and +1 if (l1 > l2)
1323 int cloog_loop_constant_cmp(CloogLoop
*l1
, CloogLoop
*l2
, int level
,
1324 int *scaldims
, int nb_scattdims
, int scalar
)
1326 CloogBlock
*b1
, *b2
;
1329 while (level_is_constant(level
, scalar
, scaldims
, nb_scattdims
)) {
1330 int cmp
= cloog_int_cmp(b1
->scaldims
[scalar
], b2
->scaldims
[scalar
]);
1340 * cloog_loop_scalar_gt function:
1341 * This function returns 1 if loop 'l1' is greater than loop 'l2' for the
1342 * scalar dimension vector that begins at dimension 'scalar', 0 otherwise. What
1343 * we want to know is whether a loop is scheduled before another one or not.
1344 * This function solves the problem when the considered dimension for scheduling
1345 * is a scalar dimension. Since there may be a succession of scalar dimensions,
1346 * this function will reason about the vector of scalar dimension that begins
1347 * at dimension 'level+scalar' and finish to the first non-scalar dimension.
1348 * \param l1 Loop to be compared with l2.
1349 * \param l2 Loop to be compared with l1.
1350 * \param level Current non-scalar dimension.
1351 * \param scaldims Boolean array saying whether a dimension is scalar or not.
1352 * \param nb_scattdims Size of the scaldims array.
1353 * \param scalar Current scalar dimension.
1354 * \return 1 if (l1 > l2), 0 otherwise.
1356 * - September 9th 2005: first version.
1357 * - October 15nd 2007: now "greater than" instead of "greater or equal".
1359 int cloog_loop_scalar_gt(l1
, l2
, level
, scaldims
, nb_scattdims
, scalar
)
1360 CloogLoop
* l1
, * l2
;
1361 int level
, * scaldims
, nb_scattdims
, scalar
;
1363 return cloog_loop_constant_cmp(l1
, l2
, level
, scaldims
, nb_scattdims
, scalar
) > 0;
1368 * cloog_loop_scalar_eq function:
1369 * This function returns 1 if loop 'l1' is equal to loop 'l2' for the scalar
1370 * dimension vector that begins at dimension 'scalar', 0 otherwise. What we want
1371 * to know is whether two loops are scheduled for the same time or not.
1372 * This function solves the problem when the considered dimension for scheduling
1373 * is a scalar dimension. Since there may be a succession of scalar dimensions,
1374 * this function will reason about the vector of scalar dimension that begins
1375 * at dimension 'level+scalar' and finish to the first non-scalar dimension.
1376 * - l1 and l2 are the loops to compare,
1377 * - level is the current non-scalar dimension,
1378 * - scaldims is the boolean array saying whether a dimension is scalar or not,
1379 * - nb_scattdims is the size of the scaldims array,
1380 * - scalar is the current scalar dimension.
1382 * - September 9th 2005 : first version.
1384 int cloog_loop_scalar_eq(l1
, l2
, level
, scaldims
, nb_scattdims
, scalar
)
1385 CloogLoop
* l1
, * l2
;
1386 int level
, * scaldims
, nb_scattdims
, scalar
;
1388 return cloog_loop_constant_cmp(l1
, l2
, level
, scaldims
, nb_scattdims
, scalar
) == 0;
1393 * cloog_loop_scalar_sort function:
1394 * This function sorts a linked list of loops (loop) with respect to the
1395 * scalar dimension vector that begins at dimension 'scalar'. Since there may
1396 * be a succession of scalar dimensions, this function will reason about the
1397 * vector of scalar dimension that begins at dimension 'level+scalar' and
1398 * finish to the first non-scalar dimension.
1399 * \param loop Loop list to sort.
1400 * \param level Current non-scalar dimension.
1401 * \param scaldims Boolean array saying whether a dimension is scalar or not.
1402 * \param nb_scattdims Size of the scaldims array.
1403 * \param scalar Current scalar dimension.
1404 * \return A pointer to the sorted list.
1406 * - July 2nd 2005: first developments.
1407 * - September 2nd 2005: first version.
1408 * - October 15nd 2007: complete rewrite to remove bugs, now a bubble sort.
1410 CloogLoop
* cloog_loop_scalar_sort(loop
, level
, scaldims
, nb_scattdims
, scalar
)
1412 int level
, * scaldims
, nb_scattdims
, scalar
;
1414 CloogLoop
**current
;
1418 for (current
= &loop
; (*current
)->next
; current
= &(*current
)->next
) {
1419 CloogLoop
*next
= (*current
)->next
;
1420 if (cloog_loop_scalar_gt(*current
,next
,level
,scaldims
,nb_scattdims
,scalar
)) {
1422 (*current
)->next
= next
->next
;
1423 next
->next
= *current
;
1434 * cloog_loop_generate_backtrack function:
1435 * adaptation from LoopGen 0.4 by F. Quillere. This function implements the
1436 * backtrack of the Quillere et al. algorithm (see the Quillere paper).
1437 * It eliminates unused iterations of the current level for the new one. See the
1438 * example called linearity-1-1 example with and without this part for an idea.
1439 * - October 26th 2001: first version in cloog_loop_generate_general.
1440 * - July 31th 2002: (debug) no more parasite loops (REALLY hard !).
1441 * - October 30th 2005: extraction from cloog_loop_generate_general.
1443 CloogLoop
*cloog_loop_generate_backtrack(CloogLoop
*loop
,
1444 int level
, CloogOptions
*options
)
1446 CloogDomain
* domain
;
1447 CloogLoop
* now
, * now2
, * next
, * next2
, * end
, * temp
, * l
, * inner
,
1453 while (temp
!= NULL
)
1455 inner
= temp
->inner
;
1457 while (inner
!= NULL
)
1458 { next
= inner
->next
;
1459 /* This 'if' and its first part is the debug of july 31th 2002. */
1460 if (inner
->block
!= NULL
) {
1461 end
= cloog_loop_alloc(temp
->state
, inner
->domain
, 0, NULL
,
1462 inner
->block
, NULL
, NULL
);
1463 domain
= cloog_domain_copy(temp
->domain
) ;
1464 new_loop
= cloog_loop_alloc(temp
->state
, domain
, 0, NULL
,
1468 new_loop
= cloog_loop_project(inner
, level
);
1470 cloog_loop_free_parts(inner
,0,0,0,0) ;
1471 cloog_loop_add(&l
,&now2
,new_loop
) ;
1475 temp
->inner
= NULL
;
1478 { l
= cloog_loop_separate(l
) ;
1479 l
= cloog_loop_sort(l
, level
);
1481 l
->stride
= cloog_stride_copy(l
->stride
);
1482 cloog_loop_add(&loop
,&now
,l
) ;
1486 next2
= temp
->next
;
1487 cloog_loop_free_parts(temp
,1,0,0,0) ;
1496 * Return 1 if we need to continue recursing to the specified level.
1498 int cloog_loop_more(CloogLoop
*loop
, int level
, int scalar
, int nb_scattdims
)
1500 return level
+ scalar
<= nb_scattdims
||
1501 cloog_domain_dimension(loop
->domain
) >= level
;
1504 CloogLoop
*cloog_loop_generate_restricted_or_stop(CloogLoop
*loop
,
1505 CloogDomain
*context
,
1506 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
1507 CloogOptions
*options
);
1510 * cloog_loop_generate_general function:
1511 * Adaptation from LoopGen 0.4 by F. Quillere. This function implements the
1512 * Quillere algorithm for polyhedron scanning from step 3 to 5.
1513 * (see the Quillere paper).
1514 * - loop is the loop for which we have to generate a scanning code,
1515 * - level is the current non-scalar dimension,
1516 * - scalar is the current scalar dimension,
1517 * - scaldims is the boolean array saying whether a dimension is scalar or not,
1518 * - nb_scattdims is the size of the scaldims array,
1519 * - options are the general code generation options.
1521 * - October 26th 2001: first version.
1522 * - July 3rd->11th 2003: memory leaks hunt and correction.
1523 * - June 22nd 2005: Adaptation for GMP.
1524 * - September 2nd 2005: The function have been cutted out in two pieces:
1525 * cloog_loop_generate and this one, in order to handle
1526 * the scalar dimension case more efficiently with
1527 * cloog_loop_generate_scalar.
1528 * - November 15th 2005: (debug) the result of the cloog_loop_generate call may
1529 * be a list of polyhedra (especially if stop option is
1530 * used): cloog_loop_add_list instead of cloog_loop_add.
1532 CloogLoop
*cloog_loop_generate_general(CloogLoop
*loop
,
1533 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
1534 CloogOptions
*options
)
1536 CloogLoop
* res
, * now
, * temp
, * l
, * new_loop
, * inner
, * now2
, * end
,
1538 CloogDomain
* domain
;
1541 /* 3. Separate all projections into disjoint polyhedra. */
1542 if ((options
->f
> level
+scalar
) || (options
->f
< 0))
1543 res
= cloog_loop_merge(loop
, level
, options
);
1545 res
= cloog_loop_separate(loop
);
1549 /* 3b. -correction- sort the loops to determine their textual order. */
1550 res
= cloog_loop_sort(res
, level
);
1552 res
= cloog_loop_restrict_inner(res
);
1555 res
= cloog_loop_specialize(res
, level
, scalar
, scaldims
, nb_scattdims
);
1557 /* 4. Recurse for each loop with the current domain as context. */
1560 if (!level
|| (level
+scalar
< options
->l
) || (options
->l
< 0))
1562 { if (level
&& options
->strides
)
1563 cloog_loop_stride(temp
, level
);
1564 if (level
&& options
->otl
)
1565 cloog_loop_otl(temp
, level
);
1566 inner
= temp
->inner
;
1567 domain
= temp
->domain
;
1569 while (inner
!= NULL
)
1570 { /* 4b. -ced- recurse for each sub-list of non terminal loops. */
1571 if (cloog_loop_more(inner
, level
+ 1, scalar
, nb_scattdims
)) {
1573 while ((end
->next
!= NULL
) &&
1574 cloog_loop_more(end
->next
, level
+ 1, scalar
, nb_scattdims
))
1580 l
= cloog_loop_generate_restricted_or_stop(inner
, domain
,
1581 level
+ 1, scalar
, scaldims
, nb_scattdims
, options
);
1584 cloog_loop_add_list(&into
,&now
,l
) ;
1589 { cloog_loop_add(&into
,&now
,inner
) ;
1590 inner
= inner
->next
;
1595 temp
->inner
= into
;
1596 cloog_loop_add(&res
,&now2
,temp
) ;
1600 while (temp
!= NULL
)
1601 { next
= temp
->next
;
1602 l
= cloog_loop_nest(temp
->inner
, temp
->domain
, level
+1);
1603 new_loop
= cloog_loop_alloc(temp
->state
, temp
->domain
, 0, NULL
,
1605 temp
->inner
= NULL
;
1607 cloog_loop_free_parts(temp
,0,0,0,0) ;
1608 cloog_loop_add(&res
,&now
,new_loop
) ;
1612 /* 5. eliminate unused iterations of the current level for the new one. See
1613 * the example called linearity-1-1 example with and without this part
1616 if (options
->backtrack
&& level
&&
1617 ((level
+scalar
< options
->l
) || (options
->l
< 0)) &&
1618 ((options
->f
<= level
+scalar
) && !(options
->f
< 0)))
1619 res
= cloog_loop_generate_backtrack(res
, level
, options
);
1621 /* Pray for my new paper to be accepted somewhere since the following stuff
1622 * is really amazing :-) !
1623 * Far long later: The paper has been accepted to PACT 2004 :-))). But there
1624 * are still some bugs and I have no time to fix them. Thus now you have to
1625 * pray for me to get an academic position for that really amazing stuff :-) !
1626 * Later again: OK, I get my academic position, but still I have not enough
1627 * time to fix and clean this part... Pray again :-) !!!
1629 /* res = cloog_loop_unisolate(res,level) ;*/
1635 CloogLoop
*cloog_loop_generate_restricted(CloogLoop
*loop
,
1636 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
1637 CloogOptions
*options
);
1641 * cloog_loop_generate_scalar function:
1642 * This function applies the simplified code generation scheme in the trivial
1643 * case of scalar dimensions. When dealing with scalar dimensions, there is
1644 * no need of costly polyhedral operations for separation or sorting: sorting
1645 * is a question of comparing scalar vectors and separation amounts to consider
1646 * only loops with the same scalar vector for the next step of the code
1647 * generation process. This function achieves the separation/sorting process
1648 * for the vector of scalar dimension that begins at dimension 'level+scalar'
1649 * and finish to the first non-scalar dimension.
1650 * - loop is the loop for which we have to generate a scanning code,
1651 * - level is the current non-scalar dimension,
1652 * - scalar is the current scalar dimension,
1653 * - scaldims is the boolean array saying whether a dimension is scalar or not,
1654 * - nb_scattdims is the size of the scaldims array,
1655 * - options are the general code generation options.
1657 * - September 2nd 2005: First version.
1659 CloogLoop
*cloog_loop_generate_scalar(CloogLoop
*loop
,
1660 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
1661 CloogOptions
*options
)
1662 { CloogLoop
* res
, * now
, * temp
, * l
, * end
, * next
, * ref
;
1665 /* We sort the loop list with respect to the current scalar vector. */
1666 res
= cloog_loop_scalar_sort(loop
,level
,scaldims
,nb_scattdims
,scalar
) ;
1668 scalar_new
= scalar
+ scaldims
[level
+ scalar
- 1];
1672 while (temp
!= NULL
)
1673 { /* Then we will appy the general code generation process to each sub-list
1674 * of loops with the same scalar vector.
1679 while((end
->next
!= NULL
) &&
1680 cloog_loop_more(end
->next
, level
, scalar_new
, nb_scattdims
) &&
1681 cloog_loop_scalar_eq(ref
,end
->next
,level
,scaldims
,nb_scattdims
,scalar
))
1687 /* For the next dimension, scalar value is updated by adding the scalar
1688 * vector size, which is stored at scaldims[level+scalar-1].
1690 if (cloog_loop_more(temp
, level
, scalar_new
, nb_scattdims
)) {
1691 l
= cloog_loop_generate_restricted(temp
, level
, scalar_new
,
1692 scaldims
, nb_scattdims
, options
);
1695 cloog_loop_add_list(&res
, &now
, l
);
1697 cloog_loop_add(&res
, &now
, temp
);
1706 /* Compare loop with the next loop based on their constant dimensions.
1707 * The result is < 0, == 0 or > 0 depending on whether the constant
1708 * dimensions of loop are lexicographically smaller, equal or greater
1709 * than those of loop->next.
1710 * If loop is the last in the list, then it is assumed to be smaller
1711 * than the "next" one.
1713 static int cloog_loop_next_scal_cmp(CloogLoop
*loop
)
1721 nb_scaldims
= loop
->block
->nb_scaldims
;
1722 if (loop
->next
->block
->nb_scaldims
< nb_scaldims
)
1723 nb_scaldims
= loop
->next
->block
->nb_scaldims
;
1725 for (i
= 0; i
< nb_scaldims
; ++i
) {
1726 int cmp
= cloog_int_cmp(loop
->block
->scaldims
[i
],
1727 loop
->next
->block
->scaldims
[i
]);
1731 return loop
->block
->nb_scaldims
- loop
->next
->block
->nb_scaldims
;
1735 /* Check whether the globally constant dimensions of a and b
1736 * have the same value for all globally constant dimensions
1737 * that are situated before any (locally) non-constant dimension.
1739 static int cloog_loop_equal_prefix(CloogLoop
*a
, CloogLoop
*b
,
1740 int *scaldims
, int nb_scattdims
)
1746 for (i
= 0; i
< nb_scattdims
; ++i
) {
1751 if (!cloog_int_eq(a
->block
->scaldims
[cst
], b
->block
->scaldims
[cst
]))
1755 for (i
= i
+ 1; i
< nb_scattdims
; ++i
) {
1758 if (!cloog_domain_lazy_isconstant(a
->domain
, dim
))
1760 /* No need to check that dim is also constant in b and that the
1761 * constant values are equal. That will happen during the check
1762 * whether the two domains are equal.
1770 /* Try to block adjacent loops in the loop list "loop".
1771 * We only attempt blocking if the constant dimensions of the loops
1772 * in the least are (not necessarily strictly) increasing.
1773 * Then we look for a sublist such that the first (begin) has constant
1774 * dimensions strictly larger than the previous loop in the complete
1775 * list and such that the loop (end) after the last loop in the sublist
1776 * has constant dimensions strictly larger than the last loop in the sublist.
1777 * Furthermore, all loops in the sublist should have the same domain
1778 * (with globally constant dimensions removed) and the difference
1779 * (if any) in constant dimensions may only occur after all the
1780 * (locally) constant dimensions.
1781 * If we find such a sublist, then the blocks of all but the first
1782 * are merged into the block of the first.
1784 * Note that this function can only be called before the global
1785 * blocklist has been created because it may otherwise modify and destroy
1786 * elements on that list.
1788 CloogLoop
*cloog_loop_block(CloogLoop
*loop
, int *scaldims
, int nb_scattdims
)
1790 CloogLoop
*begin
, *end
, *l
;
1791 int begin_after_previous
;
1792 int end_after_previous
;
1796 for (begin
= loop
; begin
; begin
= begin
->next
) {
1797 if (!begin
->block
|| !begin
->block
->scaldims
)
1799 if (cloog_loop_next_scal_cmp(begin
) > 0)
1803 begin_after_previous
= 1;
1804 for (begin
= loop
; begin
; begin
= begin
->next
) {
1805 if (!begin_after_previous
) {
1806 begin_after_previous
= cloog_loop_next_scal_cmp(begin
) < 0;
1810 end_after_previous
= cloog_loop_next_scal_cmp(begin
) < 0;
1811 for (end
= begin
->next
; end
; end
= end
->next
) {
1812 if (!cloog_loop_equal_prefix(begin
, end
, scaldims
, nb_scattdims
))
1814 if (!cloog_domain_lazy_equal(begin
->domain
, end
->domain
))
1816 end_after_previous
= cloog_loop_next_scal_cmp(end
) < 0;
1818 if (end
!= begin
->next
&& end_after_previous
) {
1819 for (l
= begin
->next
; l
!= end
; l
= begin
->next
) {
1820 cloog_block_merge(begin
->block
, l
->block
);
1821 begin
->next
= l
->next
;
1822 cloog_loop_free_parts(l
, 1, 0, 1, 0);
1826 begin_after_previous
= cloog_loop_next_scal_cmp(begin
) < 0;
1834 * Check whether for any fixed iteration of the outer loops,
1835 * there is an iteration of loop1 that is lexicographically greater
1836 * than an iteration of loop2.
1837 * Return 1 if there exists (or may exist) such a pair.
1838 * Return 0 if all iterations of loop1 are lexicographically smaller
1839 * than the iterations of loop2.
1840 * If no iteration is lexicographically greater, but if there are
1841 * iterations that are equal to iterations of loop2, then return "def".
1842 * This is useful for ensuring that such statements are not reordered.
1843 * Some users, including the test_run target in test, expect
1844 * the statements at a given point to be run in the original order.
1845 * Passing the value "0" for "def" would allow such statements to be reordered
1846 * and would allow for the detection of more components.
1848 int cloog_loop_follows(CloogLoop
*loop1
, CloogLoop
*loop2
,
1849 int level
, int scalar
, int *scaldims
, int nb_scattdims
, int def
)
1853 dim1
= cloog_domain_dimension(loop1
->domain
);
1854 dim2
= cloog_domain_dimension(loop2
->domain
);
1855 while ((level
<= dim1
&& level
<= dim2
) ||
1856 level_is_constant(level
, scalar
, scaldims
, nb_scattdims
)) {
1857 if (level_is_constant(level
, scalar
, scaldims
, nb_scattdims
)) {
1858 int cmp
= cloog_loop_constant_cmp(loop1
, loop2
, level
, scaldims
,
1859 nb_scattdims
, scalar
);
1864 scalar
+= scaldims
[level
+ scalar
- 1];
1866 int follows
= cloog_domain_follows(loop1
->domain
, loop2
->domain
,
1880 /* Structure for representing the nodes in the graph being traversed
1881 * using Tarjan's algorithm.
1882 * index represents the order in which nodes are visited.
1883 * min_index is the index of the root of a (sub)component.
1884 * on_stack indicates whether the node is currently on the stack.
1886 struct cloog_loop_sort_node
{
1891 /* Structure for representing the graph being traversed
1892 * using Tarjan's algorithm.
1893 * len is the number of nodes
1894 * node is an array of nodes
1895 * stack contains the nodes on the path from the root to the current node
1896 * sp is the stack pointer
1897 * index is the index of the last node visited
1898 * order contains the elements of the components separated by -1
1899 * op represents the current position in order
1901 struct cloog_loop_sort
{
1903 struct cloog_loop_sort_node
*node
;
1911 /* Allocate and initialize cloog_loop_sort structure.
1913 static struct cloog_loop_sort
*cloog_loop_sort_alloc(int len
)
1915 struct cloog_loop_sort
*s
;
1918 s
= (struct cloog_loop_sort
*)malloc(sizeof(struct cloog_loop_sort
));
1921 s
->node
= (struct cloog_loop_sort_node
*)
1922 malloc(len
* sizeof(struct cloog_loop_sort_node
));
1924 for (i
= 0; i
< len
; ++i
)
1925 s
->node
[i
].index
= -1;
1926 s
->stack
= (int *)malloc(len
* sizeof(int));
1928 s
->order
= (int *)malloc(2 * len
* sizeof(int));
1938 /* Free cloog_loop_sort structure.
1940 static void cloog_loop_sort_free(struct cloog_loop_sort
*s
)
1949 /* Check whether for any fixed iteration of the outer loops,
1950 * there is an iteration of loop1 that is lexicographically greater
1951 * than an iteration of loop2, where the iteration domains are
1952 * available in the inner loops of the arguments.
1954 * By using this functions to detect components, we ensure that
1955 * two CloogLoops appear in the same component if some iterations of
1956 * each loop should be executed before some iterations of the other loop.
1957 * Since we also want two CloogLoops that have exactly the same
1958 * iteration domain at the current level to be placed in the same component,
1959 * we first check if these domains are indeed the same.
1961 static int inner_loop_follows(CloogLoop
*loop1
, CloogLoop
*loop2
,
1962 int level
, int scalar
, int *scaldims
, int nb_scattdims
, int def
)
1966 f
= cloog_domain_lazy_equal(loop1
->domain
, loop2
->domain
);
1968 f
= cloog_loop_follows(loop1
->inner
, loop2
->inner
,
1969 level
, scalar
, scaldims
, nb_scattdims
, def
);
1975 /* Perform Tarjan's algorithm for computing the strongly connected components
1976 * in the graph with the individual CloogLoops as vertices.
1977 * Two CloopLoops appear in the same component if they both (indirectly)
1978 * "follow" each other, where the following relation is determined
1979 * by the follows function.
1981 static void cloog_loop_components_tarjan(struct cloog_loop_sort
*s
,
1982 CloogLoop
**loop_array
, int i
, int level
, int scalar
, int *scaldims
,
1984 int (*follows
)(CloogLoop
*loop1
, CloogLoop
*loop2
,
1985 int level
, int scalar
, int *scaldims
, int nb_scattdims
, int def
))
1989 s
->node
[i
].index
= s
->index
;
1990 s
->node
[i
].min_index
= s
->index
;
1991 s
->node
[i
].on_stack
= 1;
1993 s
->stack
[s
->sp
++] = i
;
1995 for (j
= s
->len
- 1; j
>= 0; --j
) {
2000 if (s
->node
[j
].index
>= 0 &&
2001 (!s
->node
[j
].on_stack
||
2002 s
->node
[j
].index
> s
->node
[i
].min_index
))
2005 f
= follows(loop_array
[i
], loop_array
[j
],
2006 level
, scalar
, scaldims
, nb_scattdims
, i
> j
);
2010 if (s
->node
[j
].index
< 0) {
2011 cloog_loop_components_tarjan(s
, loop_array
, j
, level
, scalar
,
2012 scaldims
, nb_scattdims
, follows
);
2013 if (s
->node
[j
].min_index
< s
->node
[i
].min_index
)
2014 s
->node
[i
].min_index
= s
->node
[j
].min_index
;
2015 } else if (s
->node
[j
].index
< s
->node
[i
].min_index
)
2016 s
->node
[i
].min_index
= s
->node
[j
].index
;
2019 if (s
->node
[i
].index
!= s
->node
[i
].min_index
)
2023 j
= s
->stack
[--s
->sp
];
2024 s
->node
[j
].on_stack
= 0;
2025 s
->order
[s
->op
++] = j
;
2027 s
->order
[s
->op
++] = -1;
2031 static int qsort_index_cmp(const void *p1
, const void *p2
)
2033 return *(int *)p1
- *(int *)p2
;
2036 /* Sort the elements of the component starting at list.
2037 * The list is terminated by a -1.
2039 static void sort_component(int *list
)
2043 for (len
= 0; list
[len
] != -1; ++len
)
2046 qsort(list
, len
, sizeof(int), qsort_index_cmp
);
2049 /* Given an array of indices "list" into the "loop_array" array,
2050 * terminated by -1, construct a linked list of the corresponding
2051 * entries and put the result in *res.
2052 * The value returned is the number of CloogLoops in the (linked) list
2054 static int extract_component(CloogLoop
**loop_array
, int *list
, CloogLoop
**res
)
2058 sort_component(list
);
2059 while (list
[i
] != -1) {
2060 *res
= loop_array
[list
[i
]];
2061 res
= &(*res
)->next
;
2071 * Call cloog_loop_generate_scalar or cloog_loop_generate_general
2072 * on each of the strongly connected components in the list of CloogLoops
2073 * pointed to by "loop".
2075 * We use Tarjan's algorithm to find the strongly connected components.
2076 * Note that this algorithm also topologically sorts the components.
2078 * The components are treated separately to avoid spurious separations.
2079 * The concatentation of the results may contain successive loops
2080 * with the same bounds, so we try to combine such loops.
2082 CloogLoop
*cloog_loop_generate_components(CloogLoop
*loop
,
2083 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
2084 CloogOptions
*options
)
2088 CloogLoop
*res
, **res_next
;
2089 CloogLoop
**loop_array
;
2090 struct cloog_loop_sort
*s
;
2092 if (level
== 0 || !loop
->next
)
2093 return cloog_loop_generate_general(loop
, level
, scalar
,
2094 scaldims
, nb_scattdims
, options
);
2096 nb_loops
= cloog_loop_count(loop
);
2098 loop_array
= (CloogLoop
**)malloc(nb_loops
* sizeof(CloogLoop
*));
2101 for (i
= 0, tmp
= loop
; i
< nb_loops
; i
++, tmp
= tmp
->next
)
2102 loop_array
[i
] = tmp
;
2104 s
= cloog_loop_sort_alloc(nb_loops
);
2105 for (i
= nb_loops
- 1; i
>= 0; --i
) {
2106 if (s
->node
[i
].index
>= 0)
2108 cloog_loop_components_tarjan(s
, loop_array
, i
, level
, scalar
, scaldims
,
2109 nb_scattdims
, &inner_loop_follows
);
2116 int n
= extract_component(loop_array
, &s
->order
[i
], &tmp
);
2119 *res_next
= cloog_loop_generate_general(tmp
, level
, scalar
,
2120 scaldims
, nb_scattdims
, options
);
2122 res_next
= &(*res_next
)->next
;
2125 cloog_loop_sort_free(s
);
2129 res
= cloog_loop_combine(res
);
2135 /* For each loop in the list "loop", decompose the list of
2136 * inner loops into strongly connected components and put
2137 * the components into separate loops at the top level.
2139 CloogLoop
*cloog_loop_decompose_inner(CloogLoop
*loop
,
2140 int level
, int scalar
, int *scaldims
, int nb_scattdims
)
2143 CloogLoop
**loop_array
;
2144 int i
, n_loops
, max_loops
= 0;
2145 struct cloog_loop_sort
*s
;
2147 for (l
= loop
; l
; l
= l
->next
) {
2148 n_loops
= cloog_loop_count(l
->inner
);
2149 if (max_loops
< n_loops
)
2150 max_loops
= n_loops
;
2156 loop_array
= (CloogLoop
**)malloc(max_loops
* sizeof(CloogLoop
*));
2159 for (l
= loop
; l
; l
= l
->next
) {
2162 for (i
= 0, tmp
= l
->inner
; tmp
; i
++, tmp
= tmp
->next
)
2163 loop_array
[i
] = tmp
;
2168 s
= cloog_loop_sort_alloc(n_loops
);
2169 for (i
= n_loops
- 1; i
>= 0; --i
) {
2170 if (s
->node
[i
].index
>= 0)
2172 cloog_loop_components_tarjan(s
, loop_array
, i
, level
, scalar
,
2173 scaldims
, nb_scattdims
, &cloog_loop_follows
);
2176 n
= extract_component(loop_array
, s
->order
, &l
->inner
);
2182 n
= extract_component(loop_array
, &s
->order
[i
], &inner
);
2185 tmp
= cloog_loop_alloc(l
->state
, cloog_domain_copy(l
->domain
),
2186 l
->otl
, l
->stride
, l
->block
, inner
, l
->next
);
2191 cloog_loop_sort_free(s
);
2200 CloogLoop
*cloog_loop_generate_restricted(CloogLoop
*loop
,
2201 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
2202 CloogOptions
*options
)
2204 /* To save both time and memory, we switch here depending on whether the
2205 * current dimension is scalar (simplified processing) or not (general
2208 if (level_is_constant(level
, scalar
, scaldims
, nb_scattdims
))
2209 return cloog_loop_generate_scalar(loop
, level
, scalar
,
2210 scaldims
, nb_scattdims
, options
);
2212 * 2. Compute the projection of each polyhedron onto the outermost
2213 * loop variable and the parameters.
2215 loop
= cloog_loop_project_all(loop
, level
);
2217 return cloog_loop_generate_components(loop
, level
, scalar
, scaldims
,
2218 nb_scattdims
, options
);
2222 CloogLoop
*cloog_loop_generate_restricted_or_stop(CloogLoop
*loop
,
2223 CloogDomain
*context
,
2224 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
2225 CloogOptions
*options
)
2227 /* If the user asked to stop code generation at this level, let's stop. */
2228 if ((options
->stop
>= 0) && (level
+scalar
>= options
->stop
+1))
2229 return cloog_loop_stop(loop
,context
) ;
2231 return cloog_loop_generate_restricted(loop
, level
, scalar
, scaldims
,
2232 nb_scattdims
, options
);
2237 * cloog_loop_generate function:
2238 * Adaptation from LoopGen 0.4 by F. Quillere. This function implements the
2239 * Quillere algorithm for polyhedron scanning from step 1 to 2.
2240 * (see the Quillere paper).
2241 * - loop is the loop for which we have to generate a scanning code,
2242 * - context is the context of the current loop (constraints on parameter and/or
2243 * on outer loop counters),
2244 * - level is the current non-scalar dimension,
2245 * - scalar is the current scalar dimension,
2246 * - scaldims is the boolean array saying whether a dimension is scalar or not,
2247 * - nb_scattdims is the size of the scaldims array,
2248 * - options are the general code generation options.
2250 * - October 26th 2001: first version.
2251 * - July 3rd->11th 2003: memory leaks hunt and correction.
2252 * - June 15th 2005: a memory leak fixed (loop was not entirely freed when
2253 * the result of cloog_loop_restrict was NULL).
2254 * - June 22nd 2005: Adaptation for GMP.
2255 * - September 2nd 2005: The function have been cutted out in two pieces:
2256 * cloog_loop_generate and this one, in order to handle
2257 * the scalar dimension case more efficiently with
2258 * cloog_loop_generate_scalar.
2259 * - November 15th 2005: (debug) Condition for stop option no more take care of
2260 * further scalar dimensions.
2262 CloogLoop
*cloog_loop_generate(CloogLoop
*loop
, CloogDomain
*context
,
2263 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
2264 CloogOptions
*options
)
2266 /* 1. Replace each polyhedron by its intersection with the context.
2268 loop
= cloog_loop_restrict_all(loop
, context
);
2272 return cloog_loop_generate_restricted_or_stop(loop
, context
,
2273 level
, scalar
, scaldims
, nb_scattdims
, options
);
2278 * Internal function for simplifying a single loop in a list of loops.
2279 * See cloog_loop_simplify.
2281 static CloogLoop
*loop_simplify(CloogLoop
*loop
, CloogDomain
*context
,
2285 CloogBlock
* new_block
;
2286 CloogLoop
*simplified
, *inner
;
2287 CloogDomain
* domain
, * simp
, * inter
, * extended_context
;
2289 if (!cloog_domain_isconvex(loop
->domain
))
2290 loop
->domain
= cloog_domain_simplify_union(loop
->domain
);
2292 domain
= loop
->domain
;
2294 domain_dim
= cloog_domain_dimension(domain
);
2295 extended_context
= cloog_domain_extend(context
, domain_dim
);
2296 inter
= cloog_domain_intersection(domain
,extended_context
) ;
2297 simp
= cloog_domain_simplify(inter
,extended_context
) ;
2298 cloog_domain_free(extended_context
) ;
2300 /* If the constraint system is never true, go to the next one. */
2301 if (cloog_domain_never_integral(simp
)) {
2302 cloog_loop_free(loop
->inner
);
2303 cloog_domain_free(inter
);
2304 cloog_domain_free(simp
);
2308 inner
= cloog_loop_simplify(loop
->inner
, inter
, level
+1);
2309 cloog_domain_free(inter
) ;
2311 if ((inner
== NULL
) && (loop
->block
== NULL
)) {
2312 cloog_domain_free(simp
);
2316 new_block
= cloog_block_copy(loop
->block
) ;
2318 simplified
= cloog_loop_alloc(loop
->state
, simp
, loop
->otl
, loop
->stride
,
2319 new_block
, inner
, NULL
);
2321 return(simplified
) ;
2326 * cloog_loop_simplify function:
2327 * This function implements the part 6. of the Quillere algorithm, it
2328 * recursively simplifies each loop in the context of the preceding loop domain.
2329 * It returns a pointer to the simplified loop list.
2330 * The cloog_domain_simplify (DomainSimplify) behaviour is really bad with
2331 * polyhedra union and some really awful sidesteppings were written, I plan
2333 * - October 31th 2001: first version.
2334 * - July 3rd->11th 2003: memory leaks hunt and correction.
2335 * - April 16th 2005: a memory leak fixed (extended_context was not freed).
2336 * - June 15th 2005: a memory leak fixed (loop was not conveniently freed
2337 * when the constraint system is never true).
2338 * - October 27th 2005: - this function called before cloog_loop_fast_simplify
2339 * is now the official cloog_loop_simplify function in
2340 * replacement of a slower and more complex one (after
2341 * deep changes in the pretty printer).
2342 * - we use cloog_loop_disjoint to fix the problem when
2343 * simplifying gives a union of polyhedra (before, it
2344 * was under the responsibility of the pretty printer).
2346 CloogLoop
*cloog_loop_simplify(CloogLoop
*loop
, CloogDomain
*context
, int level
)
2349 CloogLoop
*res
= NULL
;
2350 CloogLoop
**next
= &res
;
2352 for (now
= loop
; now
; now
= now
->next
) {
2353 *next
= loop_simplify(now
, context
, level
);
2355 now
->inner
= NULL
; /* For loop integrity. */
2356 cloog_domain_free(now
->domain
);
2360 next
= &(*next
)->next
;
2362 cloog_loop_free(loop
);
2364 /* Examples like test/iftest2.cloog give unions of polyhedra after
2365 * simplifying, thus we have to make them disjoint. Another good reason to
2366 * put the simplifying step in the Quillere backtrack.
2368 res
= cloog_loop_disjoint(res
);
2375 * cloog_loop_scatter function:
2376 * This function add the scattering (scheduling) informations in a loop.
2378 void cloog_loop_scatter(CloogLoop
* loop
, CloogScattering
*scatt
)
2380 loop
->domain
= cloog_domain_scatter(loop
->domain
, scatt
);