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 cloog_loop_leak_up(state
);
277 /* Memory allocation and information reading for the first domain: */
278 loop
= (CloogLoop
*)malloc(sizeof(CloogLoop
)) ;
280 cloog_die("memory overflow.\n");
283 loop
->domain
= domain
;
284 if (loop
->domain
!= NULL
)
285 nb_iterators
= cloog_domain_dimension(loop
->domain
);
289 /* assume no stride */
291 /* included statement block. */
292 statement
= cloog_statement_alloc(state
, number
+ 1);
293 loop
->block
= cloog_block_alloc(statement
, 0, NULL
, nb_iterators
);
295 /* inner is NULL at beginning. */
305 * cloog_loop_read function:
306 * This function reads loop data from a file (foo, possibly stdin) and
307 * returns a pointer to a CloogLoop structure containing the read information.
308 * This function can be used only for input file reading, when one loop is
309 * associated with one statement.
310 * - number is the statement block number carried by the loop (-1 if none).
311 * - nb_parameters is the number of parameters.
313 * - September 9th 2002: first version.
314 * - April 16th 2005: adaptation to new CloogStatement struct (with number).
315 * - June 11th 2005: adaptation to new CloogBlock structure.
316 * - June 22nd 2005: Adaptation for GMP.
318 CloogLoop
*cloog_loop_read(CloogState
*state
,
319 FILE *foo
, int number
, int nb_parameters
)
325 domain
= cloog_domain_union_read(state
, foo
, nb_parameters
);
327 /* To read that stupid "0 0 0" line. */
328 while (fgets(s
,MAX_STRING
,foo
) == 0) ;
329 while ((*s
=='#' || *s
=='\n') || (sscanf(s
," %d %d %d",&op1
,&op2
,&op3
)<3))
330 fgets(s
,MAX_STRING
,foo
) ;
332 return cloog_loop_from_domain(state
, domain
, number
);
336 /******************************************************************************
337 * Processing functions *
338 ******************************************************************************/
342 * cloog_loop_malloc function:
343 * This function allocates the memory space for a CloogLoop structure and
344 * sets its fields with default values. Then it returns a pointer to the
346 * - November 21th 2005: first version.
348 CloogLoop
*cloog_loop_malloc(CloogState
*state
)
351 /* Memory allocation for the CloogLoop structure. */
352 loop
= (CloogLoop
*)malloc(sizeof(CloogLoop
)) ;
354 cloog_die("memory overflow.\n");
355 cloog_loop_leak_up(state
);
358 /* We set the various fields with default values. */
360 loop
->domain
= NULL
;
373 * cloog_loop_alloc function:
374 * This function allocates the memory space for a CloogLoop structure and
375 * sets its fields with those given as input. Then it returns a pointer to the
377 * - October 27th 2001: first version.
378 * - June 22nd 2005: Adaptation for GMP.
379 * - November 21th 2005: use of cloog_loop_malloc.
381 CloogLoop
*cloog_loop_alloc(CloogState
*state
,
382 CloogDomain
*domain
, int otl
, CloogStride
*stride
,
383 CloogBlock
*block
, CloogLoop
*inner
, CloogLoop
*next
)
386 loop
= cloog_loop_malloc(state
);
388 loop
->domain
= domain
;
389 loop
->block
= block
;
390 loop
->inner
= inner
;
393 loop
->stride
= cloog_stride_copy(stride
);
400 * cloog_loop_add function:
401 * This function adds a CloogLoop structure (loop) at a given place (now) of a
402 * NULL terminated list of CloogLoop structures. The beginning of this list
403 * is (start). This function updates (now) to (loop), and updates (start) if the
404 * added element is the first one -that is when (start) is NULL-.
405 * - October 28th 2001: first version.
407 void cloog_loop_add(CloogLoop
** start
, CloogLoop
** now
, CloogLoop
* loop
)
408 { if (*start
== NULL
)
413 { (*now
)->next
= loop
;
414 *now
= (*now
)->next
;
420 * cloog_loop_add function:
421 * This function adds a CloogLoop structure (loop) at a given place (now) of a
422 * NULL terminated list of CloogLoop structures. The beginning of this list
423 * is (start). This function updates (now) to the end of the loop list (loop),
424 * and updates (start) if the added element is the first one -that is when
426 * - September 9th 2005: first version.
428 void cloog_loop_add_list(CloogLoop
** start
, CloogLoop
** now
, CloogLoop
* loop
)
429 { if (*start
== NULL
)
434 { (*now
)->next
= loop
;
435 *now
= (*now
)->next
;
438 while ((*now
)->next
!= NULL
)
439 *now
= (*now
)->next
;
444 * cloog_loop_copy function:
445 * This function returns a copy of the CloogLoop structure given as input. In
446 * fact, there is just new allocations for the CloogLoop structures, but their
447 * contents are the same.
448 * - October 28th 2001: first version.
449 * - July 3rd->11th 2003: memory leaks hunt and correction.
451 CloogLoop
* cloog_loop_copy(CloogLoop
* source
)
454 CloogDomain
* domain
;
458 { domain
= cloog_domain_copy(source
->domain
) ;
459 block
= cloog_block_copy(source
->block
) ;
460 loop
= cloog_loop_alloc(source
->state
, domain
, source
->otl
,
461 source
->stride
, block
, NULL
, NULL
);
462 loop
->usr
= source
->usr
;
463 loop
->inner
= cloog_loop_copy(source
->inner
) ;
464 loop
->next
= cloog_loop_copy(source
->next
) ;
471 * cloog_loop_add_disjoint function:
472 * This function adds some CloogLoop structures at a given place (now) of a
473 * NULL terminated list of CloogLoop structures. The beginning of this list
474 * is (start). (loop) can be an union of polyhedra, this function separates the
475 * union into a list of *disjoint* polyhedra then adds the list. This function
476 * updates (now) to the end of the list and updates (start) if first added
477 * element is the first of the principal list -that is when (start) is NULL-.
478 * (loop) can be freed by this function, basically when its domain is actually
479 * a union of polyhedra, but don't worry, all the useful data are now stored
480 * inside the list (start). We do not use PolyLib's Domain_Disjoint function,
481 * since the number of union components is often higher (thus code size too).
482 * - October 28th 2001: first version.
483 * - November 14th 2001: bug correction (this one was hard to find !).
484 * - July 3rd->11th 2003: memory leaks hunt and correction.
485 * - June 22nd 2005: Adaptation for GMP.
486 * - October 27th 2005: (debug) included blocks were not copied for new loops.
488 void cloog_loop_add_disjoint(start
, now
, loop
)
489 CloogLoop
** start
, ** now
, * loop
;
491 CloogLoop
* sep
, * inner
;
492 CloogDomain
*domain
, *seen
, *temp
, *rest
;
495 if (cloog_domain_isconvex(loop
->domain
))
496 cloog_loop_add(start
,now
,loop
) ;
498 domain
= cloog_domain_simplify_union(loop
->domain
);
499 loop
->domain
= NULL
;
501 /* We separate the first element of the rest of the union. */
502 domain
= cloog_domain_cut_first(domain
, &rest
);
504 /* This first element is the first of the list of disjoint polyhedra. */
505 sep
= cloog_loop_alloc(loop
->state
, domain
, 0, NULL
,
506 loop
->block
, loop
->inner
, NULL
);
507 cloog_loop_add(start
,now
,sep
) ;
509 seen
= cloog_domain_copy(domain
);
510 while (!cloog_domain_isempty(domain
= rest
)) {
511 temp
= cloog_domain_cut_first(domain
, &rest
);
512 domain
= cloog_domain_difference(temp
, seen
);
513 cloog_domain_free(temp
);
515 if (cloog_domain_isempty(domain
)) {
516 cloog_domain_free(domain
);
520 /* Each new loop will have its own life, for instance we can free its
521 * inner loop and included block. Then each one must have its own copy
522 * of both 'inner' and 'block'.
524 inner
= cloog_loop_copy(loop
->inner
) ;
525 block
= cloog_block_copy(loop
->block
) ;
527 sep
= cloog_loop_alloc(loop
->state
, cloog_domain_copy(domain
),
528 0, NULL
, block
, inner
, NULL
);
529 /* domain can be an union too. If so: recursion. */
530 if (cloog_domain_isconvex(domain
))
531 cloog_loop_add(start
,now
,sep
) ;
533 cloog_loop_add_disjoint(start
,now
,sep
) ;
535 if (cloog_domain_isempty(rest
)) {
536 cloog_domain_free(domain
);
540 seen
= cloog_domain_union(seen
, domain
);
542 cloog_domain_free(rest
);
543 cloog_domain_free(seen
);
544 cloog_loop_free_parts(loop
,0,0,0,0) ;
550 * cloog_loop_disjoint function:
551 * This function returns a list of loops such that each loop with non-convex
552 * domain in the input list (loop) is separated into several loops where the
553 * domains are the components of the union of *disjoint* polyhedra equivalent
554 * to the original non-convex domain. See cloog_loop_add_disjoint comments
556 * - September 16th 2005: first version.
558 CloogLoop
* cloog_loop_disjoint(CloogLoop
* loop
)
559 { CloogLoop
*res
=NULL
, * now
=NULL
, * next
;
561 /* Because this is often the case, don't waste time ! */
562 if (loop
&& !loop
->next
&& cloog_domain_isconvex(loop
->domain
))
566 { next
= loop
->next
;
568 cloog_loop_add_disjoint(&res
,&now
,loop
) ;
577 * cloog_loop_restrict function:
578 * This function returns the (loop) in the context of (context): it makes the
579 * intersection between the (loop) domain and the (context), then it returns
580 * a pointer to a new loop, with this intersection as domain.
582 * - October 27th 2001: first version.
583 * - June 15th 2005: a memory leak fixed (domain was not freed when empty).
584 * - June 22nd 2005: Adaptation for GMP.
586 CloogLoop
*cloog_loop_restrict(CloogLoop
*loop
, CloogDomain
*context
)
587 { int new_dimension
;
588 CloogDomain
* domain
, * extended_context
, * new_domain
;
589 CloogLoop
* new_loop
;
591 domain
= loop
->domain
;
592 if (cloog_domain_dimension(domain
) > cloog_domain_dimension(context
))
594 new_dimension
= cloog_domain_dimension(domain
);
595 extended_context
= cloog_domain_extend(context
, new_dimension
);
596 new_domain
= cloog_domain_intersection(extended_context
,loop
->domain
) ;
597 cloog_domain_free(extended_context
) ;
600 new_domain
= cloog_domain_intersection(context
,loop
->domain
) ;
602 if (cloog_domain_isempty(new_domain
))
603 { cloog_domain_free(new_domain
) ;
607 new_loop
= cloog_loop_alloc(loop
->state
, new_domain
,
608 0, NULL
, loop
->block
, loop
->inner
, NULL
);
615 * Call cloog_loop_restrict on each loop in the list "loop" and return
616 * the concatenated result.
618 CloogLoop
*cloog_loop_restrict_all(CloogLoop
*loop
, CloogDomain
*context
)
621 CloogLoop
*res
= NULL
;
622 CloogLoop
**res_next
= &res
;
624 for (; loop
; loop
= next
) {
627 *res_next
= cloog_loop_restrict(loop
, context
);
629 res_next
= &(*res_next
)->next
;
630 cloog_loop_free_parts(loop
, 1, 0, 0, 0);
633 cloog_loop_free(loop
);
640 CloogLoop
*cloog_loop_restrict_inner(CloogLoop
*loop
)
644 for (l
= loop
; l
; l
= l
->next
)
645 l
->inner
= cloog_loop_restrict_all(l
->inner
, l
->domain
);
651 * cloog_loop_project function:
652 * This function returns the projection of (loop) on the (level) first
653 * dimensions (outer loops). It makes the projection of the (loop) domain,
654 * then it returns a pointer to a new loop, with this projection as domain.
656 * - October 27th 2001: first version.
657 * - July 3rd->11th 2003: memory leaks hunt and correction.
658 * - June 22nd 2005: Adaptation for GMP.
660 CloogLoop
* cloog_loop_project(CloogLoop
* loop
, int level
)
662 CloogDomain
* new_domain
;
663 CloogLoop
* new_loop
, * copy
;
665 copy
= cloog_loop_alloc(loop
->state
, loop
->domain
, loop
->otl
, loop
->stride
,
666 loop
->block
, loop
->inner
, NULL
);
668 if (cloog_domain_dimension(loop
->domain
) == level
)
669 new_domain
= cloog_domain_copy(loop
->domain
) ;
671 new_domain
= cloog_domain_project(loop
->domain
, level
);
673 new_loop
= cloog_loop_alloc(loop
->state
, new_domain
, 0, NULL
,
681 * Call cloog_loop_project on each loop in the list "loop" and return
682 * the concatenated result.
684 CloogLoop
*cloog_loop_project_all(CloogLoop
*loop
, int level
)
687 CloogLoop
*res
= NULL
;
688 CloogLoop
**res_next
= &res
;
690 for (; loop
; loop
= next
) {
693 *res_next
= cloog_loop_project(loop
, level
);
694 res_next
= &(*res_next
)->next
;
695 cloog_loop_free_parts(loop
, 0, 0, 0, 0);
703 * cloog_loop_concat function:
704 * This function returns a pointer to the concatenation of the
705 * CloogLoop lists given as input.
706 * - October 28th 2001: first version.
708 CloogLoop
* cloog_loop_concat(CloogLoop
* a
, CloogLoop
* b
)
709 { CloogLoop
* loop
, * temp
;
714 { while (temp
->next
!= NULL
)
726 * cloog_loop_combine:
727 * Combine consecutive loops with identical domains into
728 * a single loop with the concatenation of their inner loops
731 CloogLoop
*cloog_loop_combine(CloogLoop
*loop
)
733 CloogLoop
*first
, *second
;
735 for (first
= loop
; first
; first
= first
->next
) {
736 while (first
->next
) {
737 if (!cloog_domain_lazy_equal(first
->domain
, first
->next
->domain
))
739 second
= first
->next
;
740 first
->inner
= cloog_loop_concat(first
->inner
, second
->inner
);
741 first
->next
= second
->next
;
742 cloog_loop_free_parts(second
, 1, 0, 0, 0);
750 * Remove loops from list that have an empty domain.
752 CloogLoop
*cloog_loop_remove_empty_domain_loops(CloogLoop
*loop
)
754 CloogLoop
*l
, *res
, *next
, **res_next
;
758 for (l
= loop
; l
; l
= next
) {
760 if (cloog_domain_isempty(l
->domain
))
761 cloog_loop_free_parts(l
, 1, 1, 1, 0);
764 res_next
= &(*res_next
)->next
;
772 CloogLoop
*cloog_loop_decompose_inner(CloogLoop
*loop
,
773 int level
, int scalar
, int *scaldims
, int nb_scattdims
);
775 /* For each loop with only one inner loop, replace the domain
776 * of the loop with the projection of the domain of the inner
777 * loop. To increase the number of loops with a single inner
778 * we first decompose the inner loops into strongly connected
781 CloogLoop
*cloog_loop_specialize(CloogLoop
*loop
,
782 int level
, int scalar
, int *scaldims
, int nb_scattdims
)
787 loop
= cloog_loop_decompose_inner(loop
, level
, scalar
,
788 scaldims
, nb_scattdims
);
790 for (l
= loop
; l
; l
= l
->next
) {
793 if (!cloog_domain_isconvex(l
->inner
->domain
))
796 dim
= cloog_domain_dimension(l
->domain
);
797 cloog_domain_free(l
->domain
);
798 l
->domain
= cloog_domain_project(l
->inner
->domain
, dim
);
801 return cloog_loop_remove_empty_domain_loops(loop
);
805 * cloog_loop_separate function:
806 * This function implements the Quillere algorithm for separation of multiple
807 * loops: for a given set of polyhedra (loop), it computes a set of disjoint
808 * polyhedra such that the unions of these sets are equal, and returns this set.
809 * - October 28th 2001: first version.
810 * - November 14th 2001: elimination of some unused blocks.
811 * - August 13th 2002: (debug) in the case of union of polyhedra for one
812 * loop, redundant constraints are fired.
813 * - July 3rd->11th 2003: memory leaks hunt and correction.
814 * - June 22nd 2005: Adaptation for GMP.
815 * - October 16th 2005: Removal of the non-shared constraint elimination when
816 * there is only one loop in the list (seems to work
817 * without now, DomainSimplify may have been improved).
818 * The problem was visible with test/iftest2.cloog.
820 CloogLoop
* cloog_loop_separate(CloogLoop
* loop
)
821 { int lazy_equal
=0, disjoint
= 0;
822 CloogLoop
* new_loop
, * new_inner
, * res
, * now
, * temp
, * Q
,
823 * inner
, * old
/*, * previous, * next*/ ;
824 CloogDomain
*UQ
, *domain
;
829 loop
= cloog_loop_combine(loop
);
831 if (loop
->next
== NULL
)
832 return cloog_loop_disjoint(loop
) ;
834 UQ
= cloog_domain_copy(loop
->domain
) ;
835 domain
= cloog_domain_copy(loop
->domain
) ;
836 res
= cloog_loop_alloc(loop
->state
, domain
, 0, NULL
,
837 loop
->block
, loop
->inner
, NULL
);
840 while((loop
= loop
->next
) != NULL
)
843 /* For all Q, add Q-loop associated with the blocks of Q alone,
844 * and Q inter loop associated with the blocks of Q and loop.
846 for (Q
= res
; Q
; Q
= Q
->next
) {
847 /* Add (Q inter loop). */
848 if ((disjoint
= cloog_domain_lazy_disjoint(Q
->domain
,loop
->domain
)))
851 { if ((lazy_equal
= cloog_domain_lazy_equal(Q
->domain
,loop
->domain
)))
852 domain
= cloog_domain_copy(Q
->domain
) ;
854 domain
= cloog_domain_intersection(Q
->domain
,loop
->domain
) ;
856 if (!cloog_domain_isempty(domain
))
857 { new_inner
= cloog_loop_concat(cloog_loop_copy(Q
->inner
),
858 cloog_loop_copy(loop
->inner
)) ;
859 new_loop
= cloog_loop_alloc(loop
->state
, domain
, 0, NULL
,
860 NULL
, new_inner
, NULL
);
861 cloog_loop_add_disjoint(&temp
,&now
,new_loop
) ;
865 cloog_domain_free(domain
);
869 /* Add (Q - loop). */
871 domain
= cloog_domain_copy(Q
->domain
) ;
874 domain
= cloog_domain_empty(Q
->domain
);
876 domain
= cloog_domain_difference(Q
->domain
,loop
->domain
) ;
879 if (!cloog_domain_isempty(domain
)) {
880 new_loop
= cloog_loop_alloc(loop
->state
, domain
, 0, NULL
,
881 NULL
, Q
->inner
, NULL
);
882 cloog_loop_add_disjoint(&temp
,&now
,new_loop
) ;
885 { cloog_domain_free(domain
) ;
886 /* If Q->inner is no more useful, we can free it. */
889 cloog_loop_free(inner
) ;
893 /* Add loop-UQ associated with the blocks of loop alone.*/
894 if (cloog_domain_lazy_disjoint(loop
->domain
,UQ
))
895 domain
= cloog_domain_copy(loop
->domain
) ;
897 { if (cloog_domain_lazy_equal(loop
->domain
,UQ
))
898 domain
= cloog_domain_empty(UQ
);
900 domain
= cloog_domain_difference(loop
->domain
,UQ
) ;
903 if (!cloog_domain_isempty(domain
)) {
904 new_loop
= cloog_loop_alloc(loop
->state
, domain
, 0, NULL
,
905 NULL
, loop
->inner
, NULL
);
906 cloog_loop_add_disjoint(&temp
,&now
,new_loop
) ;
909 { cloog_domain_free(domain
) ;
910 /* If loop->inner is no more useful, we can free it. */
911 cloog_loop_free(loop
->inner
) ;
916 if (loop
->next
!= NULL
)
917 UQ
= cloog_domain_union(UQ
, cloog_domain_copy(loop
->domain
));
919 cloog_domain_free(UQ
);
921 cloog_loop_free_parts(res
,1,0,0,1) ;
925 cloog_loop_free_parts(old
,1,0,0,1) ;
931 static CloogDomain
*bounding_domain(CloogDomain
*dom
, CloogOptions
*options
)
934 return cloog_domain_simple_convex(dom
);
936 return cloog_domain_convex(dom
);
941 * cloog_loop_merge function:
942 * This function is the 'soft' version of loop_separate if we are looking for
943 * a code much simpler (and less efficicient). This function returns the new
945 * - October 29th 2001: first version.
946 * - July 3rd->11th 2003: memory leaks hunt and correction.
947 * - June 22nd 2005: Adaptation for GMP.
949 CloogLoop
*cloog_loop_merge(CloogLoop
*loop
, int level
, CloogOptions
*options
)
951 CloogLoop
*res
, *new_inner
, *old
;
952 CloogDomain
*new_domain
, *temp
;
957 if (loop
->next
== NULL
)
958 return cloog_loop_disjoint(loop
);
963 new_inner
= loop
->inner
;
965 for (loop
= loop
->next
; loop
; loop
= loop
->next
) {
966 temp
= cloog_domain_union(temp
, loop
->domain
);
968 new_inner
= cloog_loop_concat(new_inner
, loop
->inner
);
971 new_domain
= bounding_domain(temp
, options
);
973 if (level
> 0 && !cloog_domain_is_bounded(new_domain
, level
) &&
974 cloog_domain_is_bounded(temp
, level
)) {
975 CloogDomain
*splitter
, *t2
;
977 cloog_domain_free(new_domain
);
978 splitter
= cloog_domain_bound_splitter(temp
, level
);
981 while (!cloog_domain_isconvex(splitter
)) {
982 CloogDomain
*first
, *rest
;
983 first
= cloog_domain_cut_first(splitter
, &rest
);
985 t2
= cloog_domain_intersection(first
, temp
);
986 cloog_domain_free(first
);
988 new_domain
= bounding_domain(t2
, options
);
989 cloog_domain_free(t2
);
991 if (cloog_domain_isempty(new_domain
)) {
992 cloog_domain_free(new_domain
);
995 res
= cloog_loop_alloc(old
->state
, new_domain
, 0, NULL
,
996 NULL
, cloog_loop_copy(new_inner
), res
);
999 t2
= cloog_domain_intersection(splitter
, temp
);
1000 cloog_domain_free(splitter
);
1002 new_domain
= bounding_domain(t2
, options
);
1003 cloog_domain_free(t2
);
1005 if (cloog_domain_isempty(new_domain
)) {
1006 cloog_domain_free(new_domain
);
1007 cloog_loop_free(new_inner
);
1009 res
= cloog_loop_alloc(old
->state
, new_domain
, 0, NULL
,
1010 NULL
, new_inner
, res
);
1012 res
= cloog_loop_alloc(old
->state
, new_domain
, 0, NULL
,
1013 NULL
, new_inner
, NULL
);
1015 cloog_domain_free(temp
);
1017 cloog_loop_free_parts(old
, 0, 0, 0, 1);
1023 static int cloog_loop_count(CloogLoop
*loop
)
1027 for (nb_loops
= 0; loop
; loop
= loop
->next
)
1035 * cloog_loop_sort function:
1036 * Adaptation from LoopGen 0.4 by F. Quillere. This function sorts a list of
1037 * parameterized disjoint polyhedra, in order to not have lexicographic order
1038 * violation (see Quillere paper).
1039 * - September 16th 2005: inclusion of cloog_loop_number (October 29th 2001).
1041 CloogLoop
*cloog_loop_sort(CloogLoop
*loop
, int level
)
1043 CloogLoop
*res
, *now
, **loop_array
;
1045 int i
, nb_loops
=0, * permut
;
1047 /* There is no need to sort the parameter domains. */
1051 /* We will need to know how many loops are in the list. */
1052 nb_loops
= cloog_loop_count(loop
);
1054 /* If there is only one loop, it's the end. */
1058 /* We have to allocate memory for some useful components:
1059 * - loop_array: the loop array,
1060 * - doms: the array of domains to sort,
1061 * - permut: will give us a possible sort (maybe not the only one).
1063 loop_array
= (CloogLoop
**)malloc(nb_loops
*sizeof(CloogLoop
*)) ;
1064 doms
= (CloogDomain
**)malloc(nb_loops
*sizeof(CloogDomain
*));
1065 permut
= (int *)malloc(nb_loops
*sizeof(int)) ;
1067 /* We fill up the loop and domain arrays. */
1068 for (i
=0;i
<nb_loops
;i
++,loop
=loop
->next
)
1069 { loop_array
[i
] = loop
;
1070 doms
[i
] = loop_array
[i
]->domain
;
1073 /* cloog_domain_sort will fill up permut. */
1074 cloog_domain_sort(doms
, nb_loops
, level
, permut
);
1076 /* With permut and loop_array we build the sorted list. */
1078 for (i
=0;i
<nb_loops
;i
++)
1079 { /* To avoid pointer looping... loop_add will rebuild the list. */
1080 loop_array
[permut
[i
]-1]->next
= NULL
;
1081 cloog_loop_add(&res
,&now
,loop_array
[permut
[i
]-1]) ;
1093 * cloog_loop_nest function:
1094 * This function changes the loop list in such a way that we have no more than
1095 * one dimension added by level. It returns an equivalent loop list with
1097 * - October 29th 2001: first version.
1098 * - July 3rd->11th 2003: memory leaks hunt and correction.
1099 * - June 22nd 2005: Adaptation for GMP.
1100 * - November 21th 2005: (debug) now OK when cloog_loop_restrict returns NULL.
1102 CloogLoop
*cloog_loop_nest(CloogLoop
*loop
, CloogDomain
*context
, int level
)
1104 CloogLoop
* p
, * temp
, * res
, * now
, * next
;
1105 CloogDomain
* new_domain
;
1107 loop
= cloog_loop_disjoint(loop
);
1110 /* Each domain is changed by its intersection with the context. */
1111 while (loop
!= NULL
)
1112 { p
= cloog_loop_restrict(loop
, context
);
1116 { cloog_loop_free_parts(loop
,1,0,0,0) ;
1118 temp
= cloog_loop_alloc(p
->state
, p
->domain
, 0, NULL
,
1119 p
->block
, p
->inner
, NULL
);
1121 /* If the intersection dimension is too big, we make projections smaller
1122 * and smaller, and each projection includes the preceding projection
1123 * (thus, in the target list, dimensions are added one by one).
1125 if (cloog_domain_dimension(p
->domain
) >= level
)
1126 for (l
= cloog_domain_dimension(p
->domain
); l
>= level
; l
--) {
1127 new_domain
= cloog_domain_project(p
->domain
, l
);
1128 temp
= cloog_loop_alloc(p
->state
, new_domain
, 0, NULL
,
1132 /* p is no more useful (but its content yes !). */
1133 cloog_loop_free_parts(p
,0,0,0,0) ;
1135 cloog_loop_add(&res
,&now
,temp
) ;
1138 cloog_loop_free_parts(loop
,1,1,1,0) ;
1147 /* Check if the domains of the inner loops impose a stride constraint
1148 * on the given level.
1149 * The core of the search is implemented in cloog_domain_list_stride.
1150 * Here, we simply construct a list of domains to pass to this function
1151 * and if a stride is found, we adjust the lower bounds by calling
1152 * cloog_domain_stride_lower_bound.
1154 static int cloog_loop_variable_offset_stride(CloogLoop
*loop
, int level
)
1156 CloogDomainList
*list
= NULL
;
1158 CloogStride
*stride
;
1160 for (inner
= loop
->inner
; inner
; inner
= inner
->next
) {
1161 CloogDomainList
*entry
= ALLOC(CloogDomainList
);
1162 entry
->domain
= cloog_domain_copy(inner
->domain
);
1167 stride
= cloog_domain_list_stride(list
, level
);
1169 cloog_domain_list_free(list
);
1174 loop
->stride
= stride
;
1175 loop
->domain
= cloog_domain_stride_lower_bound(loop
->domain
, level
, stride
);
1182 * cloog_loop_stride function:
1183 * This function will find the stride of a loop for the iterator at the column
1184 * number 'level' in the constraint matrix. It will update the lower bound of
1185 * the iterator accordingly. Basically, the function will try to find in the
1186 * inner loops a common condition on this iterator for the inner loop iterators
1187 * to be integral. For instance, let us consider a loop with the iterator i,
1188 * the iteration domain -4<=i<=n, and its two inner loops with the iterator j.
1189 * The first inner loop has the constraint 3j=i, and the second one has the
1190 * constraint 6j=i. Then the common constraint on i for j to be integral is
1191 * i%3=0, the stride for i is 3. Lastly, we have to find the new lower bound
1192 * for i: the first value satisfying the common constraint: -3. At the end, the
1193 * iteration domain for i is -3<=i<=n and the stride for i is 3.
1195 * The algorithm implemented in this function only allows for strides
1196 * on loops with a lower bound that has a constant remainder on division
1197 * by the stride. Before initiating this procedure, we first check
1198 * if we can find a stride with a lower bound with a variable offset in
1199 * cloog_loop_variable_offset_stride.
1201 * - loop is the loop including the iteration domain of the considered iterator,
1202 * - level is the column number of the iterator in the matrix of contraints.
1204 * - June 29th 2003: first version (work in progress since June 26th 2003).
1205 * - July 14th 2003: simpler version.
1206 * - June 22nd 2005: Adaptation for GMP (from S. Verdoolaege's 0.12.1 version).
1208 void cloog_loop_stride(CloogLoop
* loop
, int level
)
1209 { int first_search
;
1210 cloog_int_t stride
, ref_offset
, offset
, potential
;
1213 if (!cloog_domain_can_stride(loop
->domain
, level
))
1216 if (cloog_loop_variable_offset_stride(loop
, level
))
1219 cloog_int_init(stride
);
1220 cloog_int_init(ref_offset
);
1221 cloog_int_init(offset
);
1222 cloog_int_init(potential
);
1224 cloog_int_set_si(ref_offset
, 0);
1225 cloog_int_set_si(offset
, 0);
1227 /* Default stride. */
1228 cloog_int_set_si(stride
, 1);
1230 inner
= loop
->inner
;
1232 while (inner
!= NULL
)
1233 { /* If the minimun stride has not been found yet, find the stride. */
1234 if ((first_search
) || (!cloog_int_is_one(stride
)))
1236 cloog_domain_stride(inner
->domain
, level
, &potential
, &offset
);
1237 if (!cloog_int_is_one(potential
) && (!first_search
))
1238 { /* Offsets must be the same for common stride. */
1239 cloog_int_gcd(stride
, potential
, stride
);
1240 if (!cloog_int_is_zero(stride
)) {
1241 cloog_int_fdiv_r(offset
, offset
, stride
);
1242 cloog_int_fdiv_r(ref_offset
, ref_offset
, stride
);
1244 if (cloog_int_ne(offset
,ref_offset
))
1245 cloog_int_set_si(stride
, 1);
1248 cloog_int_set(stride
, potential
);
1249 cloog_int_set(ref_offset
, offset
);
1255 inner
= inner
->next
;
1258 if (cloog_int_is_zero(stride
))
1259 cloog_int_set_si(stride
, 1);
1261 /* Update the values if necessary. */
1262 if (!cloog_int_is_one(stride
))
1263 { /* Update the stride value. */
1264 if (!cloog_int_is_zero(offset
))
1265 cloog_int_sub(offset
, stride
, offset
);
1266 loop
->stride
= cloog_stride_alloc(stride
, offset
);
1267 loop
->domain
= cloog_domain_stride_lower_bound(loop
->domain
, level
,
1271 cloog_int_clear(stride
);
1272 cloog_int_clear(ref_offset
);
1273 cloog_int_clear(offset
);
1274 cloog_int_clear(potential
);
1278 void cloog_loop_otl(CloogLoop
*loop
, int level
)
1280 if (cloog_domain_is_otl(loop
->domain
, level
))
1286 * cloog_loop_stop function:
1287 * This function implements the 'stop' option : each domain of each loop
1288 * in the list 'loop' is replaced by 'context'. 'context' should be the
1289 * domain of the outer loop. By using this method, there are no more dimensions
1290 * to scan and the simplification step will automaticaly remove the domains
1291 * since they are the same as the corresponding contexts. The effect of this
1292 * function is to stop the code generation at the level this function is called,
1293 * the resulting code do not consider the next dimensions.
1294 * - January 11th 2005: first version.
1296 CloogLoop
* cloog_loop_stop(CloogLoop
* loop
, CloogDomain
* context
)
1300 { cloog_domain_free(loop
->domain
) ;
1301 loop
->domain
= cloog_domain_copy(context
) ;
1302 loop
->next
= cloog_loop_stop(loop
->next
, context
) ;
1309 static int level_is_constant(int level
, int scalar
, int *scaldims
, int nb_scattdims
)
1311 return level
&& (level
+scalar
<= nb_scattdims
) && (scaldims
[level
+scalar
-1]);
1316 * Compare the constant dimensions of loops 'l1' and 'l2' starting at 'scalar'
1317 * and return -1 if the vector of constant dimensions of 'l1' is smaller
1318 * than that of 'l2', 0 if they are the same and +1 if that of 'l1' is
1319 * greater than that of 'l2'.
1320 * This function should be called on the innermost loop (the loop
1321 * containing a block).
1322 * \param l1 Loop to be compared with l2.
1323 * \param l2 Loop to be compared with l1.
1324 * \param level Current non-scalar dimension.
1325 * \param scaldims Boolean array saying whether a dimension is scalar or not.
1326 * \param nb_scattdims Size of the scaldims array.
1327 * \param scalar Current scalar dimension.
1328 * \return -1 if (l1 < l2), 0 if (l1 == l2) and +1 if (l1 > l2)
1330 int cloog_loop_constant_cmp(CloogLoop
*l1
, CloogLoop
*l2
, int level
,
1331 int *scaldims
, int nb_scattdims
, int scalar
)
1333 CloogBlock
*b1
, *b2
;
1336 while (level_is_constant(level
, scalar
, scaldims
, nb_scattdims
)) {
1337 int cmp
= cloog_int_cmp(b1
->scaldims
[scalar
], b2
->scaldims
[scalar
]);
1347 * cloog_loop_scalar_gt function:
1348 * This function returns 1 if loop 'l1' is greater than loop 'l2' for the
1349 * scalar dimension vector that begins at dimension 'scalar', 0 otherwise. What
1350 * we want to know is whether a loop is scheduled before another one or not.
1351 * This function solves the problem when the considered dimension for scheduling
1352 * is a scalar dimension. Since there may be a succession of scalar dimensions,
1353 * this function will reason about the vector of scalar dimension that begins
1354 * at dimension 'level+scalar' and finish to the first non-scalar dimension.
1355 * \param l1 Loop to be compared with l2.
1356 * \param l2 Loop to be compared with l1.
1357 * \param level Current non-scalar dimension.
1358 * \param scaldims Boolean array saying whether a dimension is scalar or not.
1359 * \param nb_scattdims Size of the scaldims array.
1360 * \param scalar Current scalar dimension.
1361 * \return 1 if (l1 > l2), 0 otherwise.
1363 * - September 9th 2005: first version.
1364 * - October 15nd 2007: now "greater than" instead of "greater or equal".
1366 int cloog_loop_scalar_gt(l1
, l2
, level
, scaldims
, nb_scattdims
, scalar
)
1367 CloogLoop
* l1
, * l2
;
1368 int level
, * scaldims
, nb_scattdims
, scalar
;
1370 return cloog_loop_constant_cmp(l1
, l2
, level
, scaldims
, nb_scattdims
, scalar
) > 0;
1375 * cloog_loop_scalar_eq function:
1376 * This function returns 1 if loop 'l1' is equal to loop 'l2' for the scalar
1377 * dimension vector that begins at dimension 'scalar', 0 otherwise. What we want
1378 * to know is whether two loops are scheduled for the same time or not.
1379 * This function solves the problem when the considered dimension for scheduling
1380 * is a scalar dimension. Since there may be a succession of scalar dimensions,
1381 * this function will reason about the vector of scalar dimension that begins
1382 * at dimension 'level+scalar' and finish to the first non-scalar dimension.
1383 * - l1 and l2 are the loops to compare,
1384 * - level is the current non-scalar dimension,
1385 * - scaldims is the boolean array saying whether a dimension is scalar or not,
1386 * - nb_scattdims is the size of the scaldims array,
1387 * - scalar is the current scalar dimension.
1389 * - September 9th 2005 : first version.
1391 int cloog_loop_scalar_eq(l1
, l2
, level
, scaldims
, nb_scattdims
, scalar
)
1392 CloogLoop
* l1
, * l2
;
1393 int level
, * scaldims
, nb_scattdims
, scalar
;
1395 return cloog_loop_constant_cmp(l1
, l2
, level
, scaldims
, nb_scattdims
, scalar
) == 0;
1400 * cloog_loop_scalar_sort function:
1401 * This function sorts a linked list of loops (loop) with respect to the
1402 * scalar dimension vector that begins at dimension 'scalar'. Since there may
1403 * be a succession of scalar dimensions, this function will reason about the
1404 * vector of scalar dimension that begins at dimension 'level+scalar' and
1405 * finish to the first non-scalar dimension.
1406 * \param loop Loop list to sort.
1407 * \param level Current non-scalar dimension.
1408 * \param scaldims Boolean array saying whether a dimension is scalar or not.
1409 * \param nb_scattdims Size of the scaldims array.
1410 * \param scalar Current scalar dimension.
1411 * \return A pointer to the sorted list.
1413 * - July 2nd 2005: first developments.
1414 * - September 2nd 2005: first version.
1415 * - October 15nd 2007: complete rewrite to remove bugs, now a bubble sort.
1417 CloogLoop
* cloog_loop_scalar_sort(loop
, level
, scaldims
, nb_scattdims
, scalar
)
1419 int level
, * scaldims
, nb_scattdims
, scalar
;
1421 CloogLoop
**current
;
1425 for (current
= &loop
; (*current
)->next
; current
= &(*current
)->next
) {
1426 CloogLoop
*next
= (*current
)->next
;
1427 if (cloog_loop_scalar_gt(*current
,next
,level
,scaldims
,nb_scattdims
,scalar
)) {
1429 (*current
)->next
= next
->next
;
1430 next
->next
= *current
;
1441 * cloog_loop_generate_backtrack function:
1442 * adaptation from LoopGen 0.4 by F. Quillere. This function implements the
1443 * backtrack of the Quillere et al. algorithm (see the Quillere paper).
1444 * It eliminates unused iterations of the current level for the new one. See the
1445 * example called linearity-1-1 example with and without this part for an idea.
1446 * - October 26th 2001: first version in cloog_loop_generate_general.
1447 * - July 31th 2002: (debug) no more parasite loops (REALLY hard !).
1448 * - October 30th 2005: extraction from cloog_loop_generate_general.
1450 CloogLoop
*cloog_loop_generate_backtrack(CloogLoop
*loop
,
1451 int level
, CloogOptions
*options
)
1453 CloogDomain
* domain
;
1454 CloogLoop
* now
, * now2
, * next
, * next2
, * end
, * temp
, * l
, * inner
,
1460 while (temp
!= NULL
)
1462 inner
= temp
->inner
;
1464 while (inner
!= NULL
)
1465 { next
= inner
->next
;
1466 /* This 'if' and its first part is the debug of july 31th 2002. */
1467 if (inner
->block
!= NULL
) {
1468 end
= cloog_loop_alloc(temp
->state
, inner
->domain
, 0, NULL
,
1469 inner
->block
, NULL
, NULL
);
1470 domain
= cloog_domain_copy(temp
->domain
) ;
1471 new_loop
= cloog_loop_alloc(temp
->state
, domain
, 0, NULL
,
1475 new_loop
= cloog_loop_project(inner
, level
);
1477 cloog_loop_free_parts(inner
,0,0,0,0) ;
1478 cloog_loop_add(&l
,&now2
,new_loop
) ;
1482 temp
->inner
= NULL
;
1485 { l
= cloog_loop_separate(l
) ;
1486 l
= cloog_loop_sort(l
, level
);
1488 l
->stride
= cloog_stride_copy(l
->stride
);
1489 cloog_loop_add(&loop
,&now
,l
) ;
1493 next2
= temp
->next
;
1494 cloog_loop_free_parts(temp
,1,0,0,0) ;
1503 * Return 1 if we need to continue recursing to the specified level.
1505 int cloog_loop_more(CloogLoop
*loop
, int level
, int scalar
, int nb_scattdims
)
1507 return level
+ scalar
<= nb_scattdims
||
1508 cloog_domain_dimension(loop
->domain
) >= level
;
1511 CloogLoop
*cloog_loop_generate_restricted_or_stop(CloogLoop
*loop
,
1512 CloogDomain
*context
,
1513 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
1514 CloogOptions
*options
);
1517 * cloog_loop_generate_general function:
1518 * Adaptation from LoopGen 0.4 by F. Quillere. This function implements the
1519 * Quillere algorithm for polyhedron scanning from step 3 to 5.
1520 * (see the Quillere paper).
1521 * - loop is the loop for which we have to generate a scanning code,
1522 * - level is the current non-scalar dimension,
1523 * - scalar is the current scalar dimension,
1524 * - scaldims is the boolean array saying whether a dimension is scalar or not,
1525 * - nb_scattdims is the size of the scaldims array,
1526 * - options are the general code generation options.
1528 * - October 26th 2001: first version.
1529 * - July 3rd->11th 2003: memory leaks hunt and correction.
1530 * - June 22nd 2005: Adaptation for GMP.
1531 * - September 2nd 2005: The function have been cutted out in two pieces:
1532 * cloog_loop_generate and this one, in order to handle
1533 * the scalar dimension case more efficiently with
1534 * cloog_loop_generate_scalar.
1535 * - November 15th 2005: (debug) the result of the cloog_loop_generate call may
1536 * be a list of polyhedra (especially if stop option is
1537 * used): cloog_loop_add_list instead of cloog_loop_add.
1539 CloogLoop
*cloog_loop_generate_general(CloogLoop
*loop
,
1540 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
1541 CloogOptions
*options
)
1543 CloogLoop
* res
, * now
, * temp
, * l
, * new_loop
, * inner
, * now2
, * end
,
1545 CloogDomain
* domain
;
1548 /* 3. Separate all projections into disjoint polyhedra. */
1549 if ((options
->f
> level
+scalar
) || (options
->f
< 0))
1550 res
= cloog_loop_merge(loop
, level
, options
);
1552 res
= cloog_loop_separate(loop
);
1556 /* 3b. -correction- sort the loops to determine their textual order. */
1557 res
= cloog_loop_sort(res
, level
);
1559 res
= cloog_loop_restrict_inner(res
);
1562 res
= cloog_loop_specialize(res
, level
, scalar
, scaldims
, nb_scattdims
);
1564 /* 4. Recurse for each loop with the current domain as context. */
1567 if (!level
|| (level
+scalar
< options
->l
) || (options
->l
< 0))
1569 { if (level
&& options
->strides
)
1570 cloog_loop_stride(temp
, level
);
1571 if (level
&& options
->otl
)
1572 cloog_loop_otl(temp
, level
);
1573 inner
= temp
->inner
;
1574 domain
= temp
->domain
;
1576 while (inner
!= NULL
)
1577 { /* 4b. -ced- recurse for each sub-list of non terminal loops. */
1578 if (cloog_loop_more(inner
, level
+ 1, scalar
, nb_scattdims
)) {
1580 while ((end
->next
!= NULL
) &&
1581 cloog_loop_more(end
->next
, level
+ 1, scalar
, nb_scattdims
))
1587 l
= cloog_loop_generate_restricted_or_stop(inner
, domain
,
1588 level
+ 1, scalar
, scaldims
, nb_scattdims
, options
);
1591 cloog_loop_add_list(&into
,&now
,l
) ;
1596 { cloog_loop_add(&into
,&now
,inner
) ;
1597 inner
= inner
->next
;
1602 temp
->inner
= into
;
1603 cloog_loop_add(&res
,&now2
,temp
) ;
1607 while (temp
!= NULL
)
1608 { next
= temp
->next
;
1609 l
= cloog_loop_nest(temp
->inner
, temp
->domain
, level
+1);
1610 new_loop
= cloog_loop_alloc(temp
->state
, temp
->domain
, 0, NULL
,
1612 temp
->inner
= NULL
;
1614 cloog_loop_free_parts(temp
,0,0,0,0) ;
1615 cloog_loop_add(&res
,&now
,new_loop
) ;
1619 /* 5. eliminate unused iterations of the current level for the new one. See
1620 * the example called linearity-1-1 example with and without this part
1623 if (options
->backtrack
&& level
&&
1624 ((level
+scalar
< options
->l
) || (options
->l
< 0)) &&
1625 ((options
->f
<= level
+scalar
) && !(options
->f
< 0)))
1626 res
= cloog_loop_generate_backtrack(res
, level
, options
);
1628 /* Pray for my new paper to be accepted somewhere since the following stuff
1629 * is really amazing :-) !
1630 * Far long later: The paper has been accepted to PACT 2004 :-))). But there
1631 * are still some bugs and I have no time to fix them. Thus now you have to
1632 * pray for me to get an academic position for that really amazing stuff :-) !
1633 * Later again: OK, I get my academic position, but still I have not enough
1634 * time to fix and clean this part... Pray again :-) !!!
1636 /* res = cloog_loop_unisolate(res,level) ;*/
1642 CloogLoop
*cloog_loop_generate_restricted(CloogLoop
*loop
,
1643 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
1644 CloogOptions
*options
);
1648 * cloog_loop_generate_scalar function:
1649 * This function applies the simplified code generation scheme in the trivial
1650 * case of scalar dimensions. When dealing with scalar dimensions, there is
1651 * no need of costly polyhedral operations for separation or sorting: sorting
1652 * is a question of comparing scalar vectors and separation amounts to consider
1653 * only loops with the same scalar vector for the next step of the code
1654 * generation process. This function achieves the separation/sorting process
1655 * for the vector of scalar dimension that begins at dimension 'level+scalar'
1656 * and finish to the first non-scalar dimension.
1657 * - loop is the loop for which we have to generate a scanning code,
1658 * - level is the current non-scalar dimension,
1659 * - scalar is the current scalar dimension,
1660 * - scaldims is the boolean array saying whether a dimension is scalar or not,
1661 * - nb_scattdims is the size of the scaldims array,
1662 * - options are the general code generation options.
1664 * - September 2nd 2005: First version.
1666 CloogLoop
*cloog_loop_generate_scalar(CloogLoop
*loop
,
1667 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
1668 CloogOptions
*options
)
1669 { CloogLoop
* res
, * now
, * temp
, * l
, * end
, * next
, * ref
;
1672 /* We sort the loop list with respect to the current scalar vector. */
1673 res
= cloog_loop_scalar_sort(loop
,level
,scaldims
,nb_scattdims
,scalar
) ;
1675 scalar_new
= scalar
+ scaldims
[level
+ scalar
- 1];
1679 while (temp
!= NULL
)
1680 { /* Then we will appy the general code generation process to each sub-list
1681 * of loops with the same scalar vector.
1686 while((end
->next
!= NULL
) &&
1687 cloog_loop_more(end
->next
, level
, scalar_new
, nb_scattdims
) &&
1688 cloog_loop_scalar_eq(ref
,end
->next
,level
,scaldims
,nb_scattdims
,scalar
))
1694 /* For the next dimension, scalar value is updated by adding the scalar
1695 * vector size, which is stored at scaldims[level+scalar-1].
1697 if (cloog_loop_more(temp
, level
, scalar_new
, nb_scattdims
)) {
1698 l
= cloog_loop_generate_restricted(temp
, level
, scalar_new
,
1699 scaldims
, nb_scattdims
, options
);
1702 cloog_loop_add_list(&res
, &now
, l
);
1704 cloog_loop_add(&res
, &now
, temp
);
1713 /* Compare loop with the next loop based on their constant dimensions.
1714 * The result is < 0, == 0 or > 0 depending on whether the constant
1715 * dimensions of loop are lexicographically smaller, equal or greater
1716 * than those of loop->next.
1717 * If loop is the last in the list, then it is assumed to be smaller
1718 * than the "next" one.
1720 static int cloog_loop_next_scal_cmp(CloogLoop
*loop
)
1728 nb_scaldims
= loop
->block
->nb_scaldims
;
1729 if (loop
->next
->block
->nb_scaldims
< nb_scaldims
)
1730 nb_scaldims
= loop
->next
->block
->nb_scaldims
;
1732 for (i
= 0; i
< nb_scaldims
; ++i
) {
1733 int cmp
= cloog_int_cmp(loop
->block
->scaldims
[i
],
1734 loop
->next
->block
->scaldims
[i
]);
1738 return loop
->block
->nb_scaldims
- loop
->next
->block
->nb_scaldims
;
1742 /* Check whether the globally constant dimensions of a and b
1743 * have the same value for all globally constant dimensions
1744 * that are situated before any (locally) non-constant dimension.
1746 static int cloog_loop_equal_prefix(CloogLoop
*a
, CloogLoop
*b
,
1747 int *scaldims
, int nb_scattdims
)
1753 for (i
= 0; i
< nb_scattdims
; ++i
) {
1758 if (!cloog_int_eq(a
->block
->scaldims
[cst
], b
->block
->scaldims
[cst
]))
1762 for (i
= i
+ 1; i
< nb_scattdims
; ++i
) {
1765 if (!cloog_domain_lazy_isconstant(a
->domain
, dim
))
1767 /* No need to check that dim is also constant in b and that the
1768 * constant values are equal. That will happen during the check
1769 * whether the two domains are equal.
1777 /* Try to block adjacent loops in the loop list "loop".
1778 * We only attempt blocking if the constant dimensions of the loops
1779 * in the least are (not necessarily strictly) increasing.
1780 * Then we look for a sublist such that the first (begin) has constant
1781 * dimensions strictly larger than the previous loop in the complete
1782 * list and such that the loop (end) after the last loop in the sublist
1783 * has constant dimensions strictly larger than the last loop in the sublist.
1784 * Furthermore, all loops in the sublist should have the same domain
1785 * (with globally constant dimensions removed) and the difference
1786 * (if any) in constant dimensions may only occur after all the
1787 * (locally) constant dimensions.
1788 * If we find such a sublist, then the blocks of all but the first
1789 * are merged into the block of the first.
1791 * Note that this function can only be called before the global
1792 * blocklist has been created because it may otherwise modify and destroy
1793 * elements on that list.
1795 CloogLoop
*cloog_loop_block(CloogLoop
*loop
, int *scaldims
, int nb_scattdims
)
1797 CloogLoop
*begin
, *end
, *l
;
1798 int begin_after_previous
;
1799 int end_after_previous
;
1803 for (begin
= loop
; begin
; begin
= begin
->next
) {
1804 if (!begin
->block
|| !begin
->block
->scaldims
)
1806 if (cloog_loop_next_scal_cmp(begin
) > 0)
1810 begin_after_previous
= 1;
1811 for (begin
= loop
; begin
; begin
= begin
->next
) {
1812 if (!begin_after_previous
) {
1813 begin_after_previous
= cloog_loop_next_scal_cmp(begin
) < 0;
1817 end_after_previous
= cloog_loop_next_scal_cmp(begin
) < 0;
1818 for (end
= begin
->next
; end
; end
= end
->next
) {
1819 if (!cloog_loop_equal_prefix(begin
, end
, scaldims
, nb_scattdims
))
1821 if (!cloog_domain_lazy_equal(begin
->domain
, end
->domain
))
1823 end_after_previous
= cloog_loop_next_scal_cmp(end
) < 0;
1825 if (end
!= begin
->next
&& end_after_previous
) {
1826 for (l
= begin
->next
; l
!= end
; l
= begin
->next
) {
1827 cloog_block_merge(begin
->block
, l
->block
);
1828 begin
->next
= l
->next
;
1829 cloog_loop_free_parts(l
, 1, 0, 1, 0);
1833 begin_after_previous
= cloog_loop_next_scal_cmp(begin
) < 0;
1841 * Check whether for any fixed iteration of the outer loops,
1842 * there is an iteration of loop1 that is lexicographically greater
1843 * than an iteration of loop2.
1844 * Return 1 if there exists (or may exist) such a pair.
1845 * Return 0 if all iterations of loop1 are lexicographically smaller
1846 * than the iterations of loop2.
1847 * If no iteration is lexicographically greater, but if there are
1848 * iterations that are equal to iterations of loop2, then return "def".
1849 * This is useful for ensuring that such statements are not reordered.
1850 * Some users, including the test_run target in test, expect
1851 * the statements at a given point to be run in the original order.
1852 * Passing the value "0" for "def" would allow such statements to be reordered
1853 * and would allow for the detection of more components.
1855 int cloog_loop_follows(CloogLoop
*loop1
, CloogLoop
*loop2
,
1856 int level
, int scalar
, int *scaldims
, int nb_scattdims
, int def
)
1860 dim1
= cloog_domain_dimension(loop1
->domain
);
1861 dim2
= cloog_domain_dimension(loop2
->domain
);
1862 while ((level
<= dim1
&& level
<= dim2
) ||
1863 level_is_constant(level
, scalar
, scaldims
, nb_scattdims
)) {
1864 if (level_is_constant(level
, scalar
, scaldims
, nb_scattdims
)) {
1865 int cmp
= cloog_loop_constant_cmp(loop1
, loop2
, level
, scaldims
,
1866 nb_scattdims
, scalar
);
1871 scalar
+= scaldims
[level
+ scalar
- 1];
1873 int follows
= cloog_domain_follows(loop1
->domain
, loop2
->domain
,
1887 /* Structure for representing the nodes in the graph being traversed
1888 * using Tarjan's algorithm.
1889 * index represents the order in which nodes are visited.
1890 * min_index is the index of the root of a (sub)component.
1891 * on_stack indicates whether the node is currently on the stack.
1893 struct cloog_loop_sort_node
{
1898 /* Structure for representing the graph being traversed
1899 * using Tarjan's algorithm.
1900 * len is the number of nodes
1901 * node is an array of nodes
1902 * stack contains the nodes on the path from the root to the current node
1903 * sp is the stack pointer
1904 * index is the index of the last node visited
1905 * order contains the elements of the components separated by -1
1906 * op represents the current position in order
1908 struct cloog_loop_sort
{
1910 struct cloog_loop_sort_node
*node
;
1918 /* Allocate and initialize cloog_loop_sort structure.
1920 static struct cloog_loop_sort
*cloog_loop_sort_alloc(int len
)
1922 struct cloog_loop_sort
*s
;
1925 s
= (struct cloog_loop_sort
*)malloc(sizeof(struct cloog_loop_sort
));
1928 s
->node
= (struct cloog_loop_sort_node
*)
1929 malloc(len
* sizeof(struct cloog_loop_sort_node
));
1931 for (i
= 0; i
< len
; ++i
)
1932 s
->node
[i
].index
= -1;
1933 s
->stack
= (int *)malloc(len
* sizeof(int));
1935 s
->order
= (int *)malloc(2 * len
* sizeof(int));
1945 /* Free cloog_loop_sort structure.
1947 static void cloog_loop_sort_free(struct cloog_loop_sort
*s
)
1956 /* Check whether for any fixed iteration of the outer loops,
1957 * there is an iteration of loop1 that is lexicographically greater
1958 * than an iteration of loop2, where the iteration domains are
1959 * available in the inner loops of the arguments.
1961 * By using this functions to detect components, we ensure that
1962 * two CloogLoops appear in the same component if some iterations of
1963 * each loop should be executed before some iterations of the other loop.
1964 * Since we also want two CloogLoops that have exactly the same
1965 * iteration domain at the current level to be placed in the same component,
1966 * we first check if these domains are indeed the same.
1968 static int inner_loop_follows(CloogLoop
*loop1
, CloogLoop
*loop2
,
1969 int level
, int scalar
, int *scaldims
, int nb_scattdims
, int def
)
1973 f
= cloog_domain_lazy_equal(loop1
->domain
, loop2
->domain
);
1975 f
= cloog_loop_follows(loop1
->inner
, loop2
->inner
,
1976 level
, scalar
, scaldims
, nb_scattdims
, def
);
1982 /* Perform Tarjan's algorithm for computing the strongly connected components
1983 * in the graph with the individual CloogLoops as vertices.
1984 * Two CloopLoops appear in the same component if they both (indirectly)
1985 * "follow" each other, where the following relation is determined
1986 * by the follows function.
1988 static void cloog_loop_components_tarjan(struct cloog_loop_sort
*s
,
1989 CloogLoop
**loop_array
, int i
, int level
, int scalar
, int *scaldims
,
1991 int (*follows
)(CloogLoop
*loop1
, CloogLoop
*loop2
,
1992 int level
, int scalar
, int *scaldims
, int nb_scattdims
, int def
))
1996 s
->node
[i
].index
= s
->index
;
1997 s
->node
[i
].min_index
= s
->index
;
1998 s
->node
[i
].on_stack
= 1;
2000 s
->stack
[s
->sp
++] = i
;
2002 for (j
= s
->len
- 1; j
>= 0; --j
) {
2007 if (s
->node
[j
].index
>= 0 &&
2008 (!s
->node
[j
].on_stack
||
2009 s
->node
[j
].index
> s
->node
[i
].min_index
))
2012 f
= follows(loop_array
[i
], loop_array
[j
],
2013 level
, scalar
, scaldims
, nb_scattdims
, i
> j
);
2017 if (s
->node
[j
].index
< 0) {
2018 cloog_loop_components_tarjan(s
, loop_array
, j
, level
, scalar
,
2019 scaldims
, nb_scattdims
, follows
);
2020 if (s
->node
[j
].min_index
< s
->node
[i
].min_index
)
2021 s
->node
[i
].min_index
= s
->node
[j
].min_index
;
2022 } else if (s
->node
[j
].index
< s
->node
[i
].min_index
)
2023 s
->node
[i
].min_index
= s
->node
[j
].index
;
2026 if (s
->node
[i
].index
!= s
->node
[i
].min_index
)
2030 j
= s
->stack
[--s
->sp
];
2031 s
->node
[j
].on_stack
= 0;
2032 s
->order
[s
->op
++] = j
;
2034 s
->order
[s
->op
++] = -1;
2038 static int qsort_index_cmp(const void *p1
, const void *p2
)
2040 return *(int *)p1
- *(int *)p2
;
2043 /* Sort the elements of the component starting at list.
2044 * The list is terminated by a -1.
2046 static void sort_component(int *list
)
2050 for (len
= 0; list
[len
] != -1; ++len
)
2053 qsort(list
, len
, sizeof(int), qsort_index_cmp
);
2056 /* Given an array of indices "list" into the "loop_array" array,
2057 * terminated by -1, construct a linked list of the corresponding
2058 * entries and put the result in *res.
2059 * The value returned is the number of CloogLoops in the (linked) list
2061 static int extract_component(CloogLoop
**loop_array
, int *list
, CloogLoop
**res
)
2065 sort_component(list
);
2066 while (list
[i
] != -1) {
2067 *res
= loop_array
[list
[i
]];
2068 res
= &(*res
)->next
;
2078 * Call cloog_loop_generate_scalar or cloog_loop_generate_general
2079 * on each of the strongly connected components in the list of CloogLoops
2080 * pointed to by "loop".
2082 * We use Tarjan's algorithm to find the strongly connected components.
2083 * Note that this algorithm also topologically sorts the components.
2085 * The components are treated separately to avoid spurious separations.
2086 * The concatentation of the results may contain successive loops
2087 * with the same bounds, so we try to combine such loops.
2089 CloogLoop
*cloog_loop_generate_components(CloogLoop
*loop
,
2090 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
2091 CloogOptions
*options
)
2095 CloogLoop
*res
, **res_next
;
2096 CloogLoop
**loop_array
;
2097 struct cloog_loop_sort
*s
;
2099 if (level
== 0 || !loop
->next
)
2100 return cloog_loop_generate_general(loop
, level
, scalar
,
2101 scaldims
, nb_scattdims
, options
);
2103 nb_loops
= cloog_loop_count(loop
);
2105 loop_array
= (CloogLoop
**)malloc(nb_loops
* sizeof(CloogLoop
*));
2108 for (i
= 0, tmp
= loop
; i
< nb_loops
; i
++, tmp
= tmp
->next
)
2109 loop_array
[i
] = tmp
;
2111 s
= cloog_loop_sort_alloc(nb_loops
);
2112 for (i
= nb_loops
- 1; i
>= 0; --i
) {
2113 if (s
->node
[i
].index
>= 0)
2115 cloog_loop_components_tarjan(s
, loop_array
, i
, level
, scalar
, scaldims
,
2116 nb_scattdims
, &inner_loop_follows
);
2123 int n
= extract_component(loop_array
, &s
->order
[i
], &tmp
);
2126 *res_next
= cloog_loop_generate_general(tmp
, level
, scalar
,
2127 scaldims
, nb_scattdims
, options
);
2129 res_next
= &(*res_next
)->next
;
2132 cloog_loop_sort_free(s
);
2136 res
= cloog_loop_combine(res
);
2142 /* For each loop in the list "loop", decompose the list of
2143 * inner loops into strongly connected components and put
2144 * the components into separate loops at the top level.
2146 CloogLoop
*cloog_loop_decompose_inner(CloogLoop
*loop
,
2147 int level
, int scalar
, int *scaldims
, int nb_scattdims
)
2150 CloogLoop
**loop_array
;
2151 int i
, n_loops
, max_loops
= 0;
2152 struct cloog_loop_sort
*s
;
2154 for (l
= loop
; l
; l
= l
->next
) {
2155 n_loops
= cloog_loop_count(l
->inner
);
2156 if (max_loops
< n_loops
)
2157 max_loops
= n_loops
;
2163 loop_array
= (CloogLoop
**)malloc(max_loops
* sizeof(CloogLoop
*));
2166 for (l
= loop
; l
; l
= l
->next
) {
2169 for (i
= 0, tmp
= l
->inner
; tmp
; i
++, tmp
= tmp
->next
)
2170 loop_array
[i
] = tmp
;
2175 s
= cloog_loop_sort_alloc(n_loops
);
2176 for (i
= n_loops
- 1; i
>= 0; --i
) {
2177 if (s
->node
[i
].index
>= 0)
2179 cloog_loop_components_tarjan(s
, loop_array
, i
, level
, scalar
,
2180 scaldims
, nb_scattdims
, &cloog_loop_follows
);
2183 n
= extract_component(loop_array
, s
->order
, &l
->inner
);
2189 n
= extract_component(loop_array
, &s
->order
[i
], &inner
);
2192 tmp
= cloog_loop_alloc(l
->state
, cloog_domain_copy(l
->domain
),
2193 l
->otl
, l
->stride
, l
->block
, inner
, l
->next
);
2198 cloog_loop_sort_free(s
);
2207 CloogLoop
*cloog_loop_generate_restricted(CloogLoop
*loop
,
2208 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
2209 CloogOptions
*options
)
2211 /* To save both time and memory, we switch here depending on whether the
2212 * current dimension is scalar (simplified processing) or not (general
2215 if (level_is_constant(level
, scalar
, scaldims
, nb_scattdims
))
2216 return cloog_loop_generate_scalar(loop
, level
, scalar
,
2217 scaldims
, nb_scattdims
, options
);
2219 * 2. Compute the projection of each polyhedron onto the outermost
2220 * loop variable and the parameters.
2222 loop
= cloog_loop_project_all(loop
, level
);
2224 return cloog_loop_generate_components(loop
, level
, scalar
, scaldims
,
2225 nb_scattdims
, options
);
2229 CloogLoop
*cloog_loop_generate_restricted_or_stop(CloogLoop
*loop
,
2230 CloogDomain
*context
,
2231 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
2232 CloogOptions
*options
)
2234 /* If the user asked to stop code generation at this level, let's stop. */
2235 if ((options
->stop
>= 0) && (level
+scalar
>= options
->stop
+1))
2236 return cloog_loop_stop(loop
,context
) ;
2238 return cloog_loop_generate_restricted(loop
, level
, scalar
, scaldims
,
2239 nb_scattdims
, options
);
2244 * cloog_loop_generate function:
2245 * Adaptation from LoopGen 0.4 by F. Quillere. This function implements the
2246 * Quillere algorithm for polyhedron scanning from step 1 to 2.
2247 * (see the Quillere paper).
2248 * - loop is the loop for which we have to generate a scanning code,
2249 * - context is the context of the current loop (constraints on parameter and/or
2250 * on outer loop counters),
2251 * - level is the current non-scalar dimension,
2252 * - scalar is the current scalar dimension,
2253 * - scaldims is the boolean array saying whether a dimension is scalar or not,
2254 * - nb_scattdims is the size of the scaldims array,
2255 * - options are the general code generation options.
2257 * - October 26th 2001: first version.
2258 * - July 3rd->11th 2003: memory leaks hunt and correction.
2259 * - June 15th 2005: a memory leak fixed (loop was not entirely freed when
2260 * the result of cloog_loop_restrict was NULL).
2261 * - June 22nd 2005: Adaptation for GMP.
2262 * - September 2nd 2005: The function have been cutted out in two pieces:
2263 * cloog_loop_generate and this one, in order to handle
2264 * the scalar dimension case more efficiently with
2265 * cloog_loop_generate_scalar.
2266 * - November 15th 2005: (debug) Condition for stop option no more take care of
2267 * further scalar dimensions.
2269 CloogLoop
*cloog_loop_generate(CloogLoop
*loop
, CloogDomain
*context
,
2270 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
2271 CloogOptions
*options
)
2273 /* 1. Replace each polyhedron by its intersection with the context.
2275 loop
= cloog_loop_restrict_all(loop
, context
);
2279 return cloog_loop_generate_restricted_or_stop(loop
, context
,
2280 level
, scalar
, scaldims
, nb_scattdims
, options
);
2285 * Internal function for simplifying a single loop in a list of loops.
2286 * See cloog_loop_simplify.
2288 static CloogLoop
*loop_simplify(CloogLoop
*loop
, CloogDomain
*context
,
2292 CloogBlock
* new_block
;
2293 CloogLoop
*simplified
, *inner
;
2294 CloogDomain
* domain
, * simp
, * inter
, * extended_context
;
2296 if (!cloog_domain_isconvex(loop
->domain
))
2297 loop
->domain
= cloog_domain_simplify_union(loop
->domain
);
2299 domain
= loop
->domain
;
2301 domain_dim
= cloog_domain_dimension(domain
);
2302 extended_context
= cloog_domain_extend(context
, domain_dim
);
2303 inter
= cloog_domain_intersection(domain
,extended_context
) ;
2304 simp
= cloog_domain_simplify(inter
,extended_context
) ;
2305 cloog_domain_free(extended_context
) ;
2307 /* If the constraint system is never true, go to the next one. */
2308 if (cloog_domain_never_integral(simp
)) {
2309 cloog_loop_free(loop
->inner
);
2310 cloog_domain_free(inter
);
2311 cloog_domain_free(simp
);
2315 inner
= cloog_loop_simplify(loop
->inner
, inter
, level
+1);
2316 cloog_domain_free(inter
) ;
2318 if ((inner
== NULL
) && (loop
->block
== NULL
)) {
2319 cloog_domain_free(simp
);
2323 new_block
= cloog_block_copy(loop
->block
) ;
2325 simplified
= cloog_loop_alloc(loop
->state
, simp
, loop
->otl
, loop
->stride
,
2326 new_block
, inner
, NULL
);
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
)
2356 CloogLoop
*res
= NULL
;
2357 CloogLoop
**next
= &res
;
2359 for (now
= loop
; now
; now
= now
->next
) {
2360 *next
= loop_simplify(now
, context
, level
);
2362 now
->inner
= NULL
; /* For loop integrity. */
2363 cloog_domain_free(now
->domain
);
2367 next
= &(*next
)->next
;
2369 cloog_loop_free(loop
);
2371 /* Examples like test/iftest2.cloog give unions of polyhedra after
2372 * simplifying, thus we we have to disjoint them. Another good reason to
2373 * put the simplifying step in the Quillere backtrack.
2375 res
= cloog_loop_disjoint(res
);
2382 * cloog_loop_scatter function:
2383 * This function add the scattering (scheduling) informations in a loop.
2385 void cloog_loop_scatter(CloogLoop
* loop
, CloogScattering
*scatt
)
2387 loop
->domain
= cloog_domain_scatter(loop
->domain
, scatt
);