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"
45 /******************************************************************************
46 * Memory leaks hunting *
47 ******************************************************************************/
51 * These functions and global variables are devoted to memory leaks hunting: we
52 * want to know at each moment how many CloogLoop structures had been allocated
53 * (cloog_loop_allocated) and how many had been freed (cloog_loop_freed).
54 * Each time a CloogLoog structure is allocated, a call to the function
55 * cloog_loop_leak_up() must be carried out, and respectively
56 * cloog_loop_leak_down() when a CloogLoop structure is freed. The special
57 * variable cloog_loop_max gives the maximal number of CloogLoop structures
58 * simultaneously alive (i.e. allocated and non-freed) in memory.
59 * - July 3rd->11th 2003: first version (memory leaks hunt and correction).
63 static void cloog_loop_leak_up(CloogState
*state
)
65 state
->loop_allocated
++;
66 if ((state
->loop_allocated
- state
->loop_freed
) > state
->loop_max
)
67 state
->loop_max
= state
->loop_allocated
- state
->loop_freed
;
71 static void cloog_loop_leak_down(CloogState
*state
)
77 /******************************************************************************
78 * Structure display function *
79 ******************************************************************************/
83 * cloog_loop_print_structure function:
84 * Displays a loop structure in a way that trends to be understandable without
85 * falling in a deep depression or, for the lucky ones, getting a headache...
86 * Written by Olivier Chorier, Luc Marchaud, Pierre Martin and Romain Tartiere.
87 * - April 24th 2005: Initial version.
88 * - May 21rd 2005: - New parameter `F' for destination file (ie stdout),
90 * - May 26th 2005: Memory leak hunt.
91 * - June 2nd 2005: (Ced) Integration and minor fixes.
92 * -June 22nd 2005: (Ced) Adaptation for GMP.
94 void cloog_loop_print_structure(FILE * file
, CloogLoop
* loop
, int level
)
98 { /* Go to the right level. */
99 for (i
=0; i
<level
; i
++)
100 fprintf(file
,"|\t") ;
102 fprintf(file
,"+-- CloogLoop\n") ;
108 { /* Go to the right level. */
109 for (i
=0; i
<level
; i
++)
110 fprintf(file
,"|\t") ;
112 fprintf(file
,"| CloogLoop\n") ;
118 for(j
=0; j
<=level
+1; j
++)
119 fprintf(file
,"|\t") ;
122 /* Print the domain. */
123 cloog_domain_print_structure(file
, loop
->domain
, level
+1, "CloogDomain");
125 /* Print the stride. */
126 for(j
=0; j
<=level
; j
++)
127 fprintf(file
,"|\t") ;
128 fprintf(file
, "Stride: ") ;
129 cloog_int_print(file
, loop
->stride
);
130 fprintf(file
, "\n") ;
131 fprintf(file
, "Offset: ") ;
132 cloog_int_print(file
, loop
->offset
);
133 fprintf(file
, "\n") ;
136 for(j
=0; j
<=level
+1; j
++)
137 fprintf(file
,"|\t") ;
140 /* Print the block. */
141 cloog_block_print_structure(file
,loop
->block
,level
+1) ;
144 for (i
=0; i
<=level
+1; i
++)
145 fprintf(file
,"|\t") ;
148 /* Print inner if any. */
150 cloog_loop_print_structure(file
,loop
->inner
,level
+1) ;
152 /* And let's go for the next one. */
155 /* One more time something that is here only for a better look. */
157 { /* Two blank lines if this is the end of the linked list. */
159 { for (i
=0; i
<=level
; i
++)
160 fprintf(file
,"|\t") ;
166 { /* A special blank line if the is a next loop. */
167 for (i
=0; i
<=level
; i
++)
168 fprintf(file
,"|\t") ;
169 fprintf(file
,"V\n") ;
176 * cloog_loop_print function:
177 * This function prints the content of a CloogLoop structure (start) into a
178 * file (file, possibly stdout).
179 * - June 2nd 2005: Now this very old function (probably as old as CLooG) is
180 * only a frontend to cloog_loop_print_structure, with a quite
181 * better human-readable representation.
183 void cloog_loop_print(FILE * file
, CloogLoop
* loop
)
184 { cloog_loop_print_structure(file
,loop
,0) ;
188 /******************************************************************************
189 * Memory deallocation function *
190 ******************************************************************************/
194 * cloog_loop_free function:
195 * This function frees the allocated memory for a CloogLoop structure (loop),
196 * and frees its inner loops and its next loops.
197 * - June 22nd 2005: Adaptation for GMP.
199 void cloog_loop_free(CloogLoop
* loop
)
202 while (loop
!= NULL
) {
203 cloog_loop_leak_down(loop
->state
);
206 cloog_domain_free(loop
->domain
) ;
207 cloog_block_free(loop
->block
) ;
208 if (loop
->inner
!= NULL
)
209 cloog_loop_free(loop
->inner
) ;
211 cloog_int_clear(loop
->stride
);
212 cloog_int_clear(loop
->offset
);
220 * cloog_loop_free_parts function:
221 * This function frees the allocated memory for some parts of a CloogLoop
222 * structure (loop), each other argument is a boolean having to be set to 1 if
223 * we want to free the corresponding part, 0 otherwise. This function applies
224 * the same freeing policy to its inner ans next loops recursively.
225 * - July 3rd 2003: first version.
226 * - June 22nd 2005: Adaptation for GMP.
228 void cloog_loop_free_parts(loop
, domain
, block
, inner
, next
)
230 int domain
, block
, inner
, next
;
231 { CloogLoop
* follow
;
233 while (loop
!= NULL
) {
234 cloog_loop_leak_down(loop
->state
);
235 follow
= loop
->next
;
238 cloog_domain_free(loop
->domain
) ;
241 cloog_block_free(loop
->block
) ;
243 if ((inner
) && (loop
->inner
!= NULL
))
244 cloog_loop_free_parts(loop
->inner
,domain
,block
,inner
,1) ;
246 cloog_int_clear(loop
->stride
);
247 cloog_int_clear(loop
->offset
);
257 /******************************************************************************
258 * Reading functions *
259 ******************************************************************************/
263 * cloog_loop_read function:
264 * This function reads loop data into a file (foo, possibly stdin) and
265 * returns a pointer to a CloogLoop structure containing the read information.
266 * This function can be used only for input file reading, when one loop is
267 * associated with one statement.
268 * - number is the statement block number carried by the loop (-1 if none).
269 * - nb_parameters is the number of parameters.
271 * - September 9th 2002: first version.
272 * - April 16th 2005: adaptation to new CloogStatement struct (with number).
273 * - June 11th 2005: adaptation to new CloogBlock structure.
274 * - June 22nd 2005: Adaptation for GMP.
276 CloogLoop
*cloog_loop_read(CloogState
*state
,
277 FILE * foo
, int number
, int nb_parameters
)
278 { int nb_iterators
, op1
, op2
, op3
;
281 CloogStatement
* statement
;
283 cloog_loop_leak_up(state
);
285 /* Memory allocation and information reading for the first domain: */
286 loop
= (CloogLoop
*)malloc(sizeof(CloogLoop
)) ;
288 cloog_die("memory overflow.\n");
291 loop
->domain
= cloog_domain_union_read(state
, foo
, nb_parameters
);
292 if (loop
->domain
!= NULL
)
293 nb_iterators
= cloog_domain_dimension(loop
->domain
);
296 /* stride is initialized to 1. */
297 cloog_int_init(loop
->stride
);
298 cloog_int_set_si(loop
->stride
, 1);
299 cloog_int_init(loop
->offset
);
300 cloog_int_set_si(loop
->offset
, 0);
301 /* included statement block. */
302 statement
= cloog_statement_alloc(state
, number
+ 1);
303 loop
->block
= cloog_block_alloc(statement
, 0, NULL
, nb_iterators
);
305 /* inner is NULL at beginning. */
310 /* To read that stupid "0 0 0" line. */
311 while (fgets(s
,MAX_STRING
,foo
) == 0) ;
312 while ((*s
=='#' || *s
=='\n') || (sscanf(s
," %d %d %d",&op1
,&op2
,&op3
)<3))
313 fgets(s
,MAX_STRING
,foo
) ;
319 /******************************************************************************
320 * Processing functions *
321 ******************************************************************************/
325 * cloog_loop_malloc function:
326 * This function allocates the memory space for a CloogLoop structure and
327 * sets its fields with default values. Then it returns a pointer to the
329 * - November 21th 2005: first version.
331 CloogLoop
*cloog_loop_malloc(CloogState
*state
)
334 /* Memory allocation for the CloogLoop structure. */
335 loop
= (CloogLoop
*)malloc(sizeof(CloogLoop
)) ;
337 cloog_die("memory overflow.\n");
338 cloog_loop_leak_up(state
);
341 /* We set the various fields with default values. */
343 loop
->domain
= NULL
;
348 cloog_int_init(loop
->stride
);
349 cloog_int_set_si(loop
->stride
, 1);
350 cloog_int_init(loop
->offset
);
351 cloog_int_set_si(loop
->offset
, 0);
358 * cloog_loop_alloc function:
359 * This function allocates the memory space for a CloogLoop structure and
360 * sets its fields with those given as input. Then it returns a pointer to the
362 * - October 27th 2001: first version.
363 * - June 22nd 2005: Adaptation for GMP.
364 * - November 21th 2005: use of cloog_loop_malloc.
366 CloogLoop
*cloog_loop_alloc(CloogState
*state
,
367 CloogDomain
*domain
, cloog_int_t stride
, cloog_int_t offset
,
368 CloogBlock
*block
, CloogLoop
*inner
, CloogLoop
*next
)
371 loop
= cloog_loop_malloc(state
);
373 loop
->domain
= domain
;
374 loop
->block
= block
;
375 loop
->inner
= inner
;
377 cloog_int_set(loop
->stride
, stride
);
378 cloog_int_set(loop
->offset
, offset
);
385 * cloog_loop_add function:
386 * This function adds a CloogLoop structure (loop) at a given place (now) of a
387 * NULL terminated list of CloogLoop structures. The beginning of this list
388 * is (start). This function updates (now) to (loop), and updates (start) if the
389 * added element is the first one -that is when (start) is NULL-.
390 * - October 28th 2001: first version.
392 void cloog_loop_add(CloogLoop
** start
, CloogLoop
** now
, CloogLoop
* loop
)
393 { if (*start
== NULL
)
398 { (*now
)->next
= loop
;
399 *now
= (*now
)->next
;
405 * cloog_loop_add function:
406 * This function adds a CloogLoop structure (loop) at a given place (now) of a
407 * NULL terminated list of CloogLoop structures. The beginning of this list
408 * is (start). This function updates (now) to the end of the loop list (loop),
409 * and updates (start) if the added element is the first one -that is when
411 * - September 9th 2005: first version.
413 void cloog_loop_add_list(CloogLoop
** start
, CloogLoop
** now
, CloogLoop
* loop
)
414 { if (*start
== NULL
)
419 { (*now
)->next
= loop
;
420 *now
= (*now
)->next
;
423 while ((*now
)->next
!= NULL
)
424 *now
= (*now
)->next
;
429 * cloog_loop_copy function:
430 * This function returns a copy of the CloogLoop structure given as input. In
431 * fact, there is just new allocations for the CloogLoop structures, but their
432 * contents are the same.
433 * - October 28th 2001: first version.
434 * - July 3rd->11th 2003: memory leaks hunt and correction.
436 CloogLoop
* cloog_loop_copy(CloogLoop
* source
)
439 CloogDomain
* domain
;
443 { domain
= cloog_domain_copy(source
->domain
) ;
444 block
= cloog_block_copy(source
->block
) ;
445 loop
= cloog_loop_alloc(source
->state
, domain
, source
->stride
,
446 source
->offset
, block
, NULL
, NULL
);
447 loop
->usr
= source
->usr
;
448 loop
->inner
= cloog_loop_copy(source
->inner
) ;
449 loop
->next
= cloog_loop_copy(source
->next
) ;
456 * cloog_loop_add_disjoint function:
457 * This function adds some CloogLoop structures at a given place (now) of a
458 * NULL terminated list of CloogLoop structures. The beginning of this list
459 * is (start). (loop) can be an union of polyhedra, this function separates the
460 * union into a list of *disjoint* polyhedra then adds the list. This function
461 * updates (now) to the end of the list and updates (start) if first added
462 * element is the first of the principal list -that is when (start) is NULL-.
463 * (loop) can be freed by this function, basically when its domain is actually
464 * a union of polyhedra, but don't worry, all the useful data are now stored
465 * inside the list (start). We do not use PolyLib's Domain_Disjoint function,
466 * since the number of union components is often higher (thus code size too).
467 * - October 28th 2001: first version.
468 * - November 14th 2001: bug correction (this one was hard to find !).
469 * - July 3rd->11th 2003: memory leaks hunt and correction.
470 * - June 22nd 2005: Adaptation for GMP.
471 * - October 27th 2005: (debug) included blocks were not copied for new loops.
473 void cloog_loop_add_disjoint(start
, now
, loop
)
474 CloogLoop
** start
, ** now
, * loop
;
476 CloogLoop
* sep
, * inner
;
477 CloogDomain
*domain
, *seen
, *temp
, *rest
;
480 if (cloog_domain_isconvex(loop
->domain
))
481 cloog_loop_add(start
,now
,loop
) ;
483 domain
= cloog_domain_simplify_union(loop
->domain
);
484 loop
->domain
= NULL
;
486 /* We separate the first element of the rest of the union. */
487 domain
= cloog_domain_cut_first(domain
, &rest
);
489 /* This first element is the first of the list of disjoint polyhedra. */
490 sep
= cloog_loop_alloc(loop
->state
, domain
, loop
->state
->one
,
491 loop
->state
->zero
, loop
->block
, loop
->inner
, NULL
);
492 cloog_loop_add(start
,now
,sep
) ;
494 seen
= cloog_domain_copy(domain
);
495 while (!cloog_domain_isempty(domain
= rest
)) {
496 temp
= cloog_domain_cut_first(domain
, &rest
);
497 domain
= cloog_domain_difference(temp
, seen
);
498 cloog_domain_free(temp
);
500 if (cloog_domain_isempty(domain
)) {
501 cloog_domain_free(domain
);
505 /* Each new loop will have its own life, for instance we can free its
506 * inner loop and included block. Then each one must have its own copy
507 * of both 'inner' and 'block'.
509 inner
= cloog_loop_copy(loop
->inner
) ;
510 block
= cloog_block_copy(loop
->block
) ;
512 sep
= cloog_loop_alloc(loop
->state
, cloog_domain_copy(domain
),
513 loop
->state
->one
, loop
->state
->zero
,
515 /* domain can be an union too. If so: recursion. */
516 if (cloog_domain_isconvex(domain
))
517 cloog_loop_add(start
,now
,sep
) ;
519 cloog_loop_add_disjoint(start
,now
,sep
) ;
521 if (cloog_domain_isempty(rest
)) {
522 cloog_domain_free(domain
);
526 seen
= cloog_domain_union(seen
, domain
);
528 cloog_domain_free(rest
);
529 cloog_domain_free(seen
);
530 cloog_loop_free_parts(loop
,0,0,0,0) ;
536 * cloog_loop_disjoint function:
537 * This function returns a list of loops such that each loop with non-convex
538 * domain in the input list (loop) is separated into several loops where the
539 * domains are the components of the union of *disjoint* polyhedra equivalent
540 * to the original non-convex domain. See cloog_loop_add_disjoint comments
542 * - September 16th 2005: first version.
544 CloogLoop
* cloog_loop_disjoint(CloogLoop
* loop
)
545 { CloogLoop
*res
=NULL
, * now
=NULL
, * next
;
547 /* Because this is often the case, don't waste time ! */
548 if (loop
&& !loop
->next
&& cloog_domain_isconvex(loop
->domain
))
552 { next
= loop
->next
;
554 cloog_loop_add_disjoint(&res
,&now
,loop
) ;
563 * cloog_loop_restrict function:
564 * This function returns the (loop) in the context of (context): it makes the
565 * intersection between the (loop) domain and the (context), then it returns
566 * a pointer to a new loop, with this intersection as domain.
568 * - October 27th 2001: first version.
569 * - June 15th 2005: a memory leak fixed (domain was not freed when empty).
570 * - June 22nd 2005: Adaptation for GMP.
572 CloogLoop
*cloog_loop_restrict(CloogLoop
*loop
, CloogDomain
*context
)
573 { int new_dimension
;
574 CloogDomain
* domain
, * extended_context
, * new_domain
;
575 CloogLoop
* new_loop
;
577 domain
= loop
->domain
;
578 if (cloog_domain_dimension(domain
) > cloog_domain_dimension(context
))
580 new_dimension
= cloog_domain_dimension(domain
);
581 extended_context
= cloog_domain_extend(context
, new_dimension
);
582 new_domain
= cloog_domain_intersection(extended_context
,loop
->domain
) ;
583 cloog_domain_free(extended_context
) ;
586 new_domain
= cloog_domain_intersection(context
,loop
->domain
) ;
588 if (cloog_domain_isempty(new_domain
))
589 { cloog_domain_free(new_domain
) ;
593 new_loop
= cloog_loop_alloc(loop
->state
, new_domain
,
594 loop
->state
->one
, loop
->state
->zero
,
595 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
->stride
, loop
->offset
,
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
, loop
->state
->one
,
661 loop
->state
->zero
, NULL
, copy
, 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
)
774 loop
= cloog_loop_decompose_inner(loop
, level
, scalar
,
775 scaldims
, nb_scattdims
);
777 for (l
= loop
; l
; l
= l
->next
) {
780 if (!cloog_domain_isconvex(l
->inner
->domain
))
783 dim
= cloog_domain_dimension(l
->domain
);
784 cloog_domain_free(l
->domain
);
785 l
->domain
= cloog_domain_project(l
->inner
->domain
, dim
);
788 return cloog_loop_remove_empty_domain_loops(loop
);
792 * cloog_loop_separate function:
793 * This function implements the Quillere algorithm for separation of multiple
794 * loops: for a given set of polyhedra (loop), it computes a set of disjoint
795 * polyhedra such that the unions of these sets are equal, and returns this set.
796 * - October 28th 2001: first version.
797 * - November 14th 2001: elimination of some unused blocks.
798 * - August 13th 2002: (debug) in the case of union of polyhedra for one
799 * loop, redundant constraints are fired.
800 * - July 3rd->11th 2003: memory leaks hunt and correction.
801 * - June 22nd 2005: Adaptation for GMP.
802 * - October 16th 2005: Removal of the non-shared constraint elimination when
803 * there is only one loop in the list (seems to work
804 * without now, DomainSimplify may have been improved).
805 * The problem was visible with test/iftest2.cloog.
807 CloogLoop
* cloog_loop_separate(CloogLoop
* loop
)
808 { int lazy_equal
=0, disjoint
= 0;
809 CloogLoop
* new_loop
, * new_inner
, * res
, * now
, * temp
, * Q
,
810 * inner
, * old
/*, * previous, * next*/ ;
811 CloogDomain
*UQ
, *domain
;
816 loop
= cloog_loop_combine(loop
);
818 if (loop
->next
== NULL
)
819 return cloog_loop_disjoint(loop
) ;
821 UQ
= cloog_domain_copy(loop
->domain
) ;
822 domain
= cloog_domain_copy(loop
->domain
) ;
823 res
= cloog_loop_alloc(loop
->state
, domain
, loop
->state
->one
,
824 loop
->state
->zero
, loop
->block
, loop
->inner
, NULL
);
827 while((loop
= loop
->next
) != NULL
)
830 /* For all Q, add Q-loop associated with the blocks of Q alone,
831 * and Q inter loop associated with the blocks of Q and loop.
833 for (Q
= res
; Q
; Q
= Q
->next
) {
834 /* Add (Q inter loop). */
835 if ((disjoint
= cloog_domain_lazy_disjoint(Q
->domain
,loop
->domain
)))
838 { if ((lazy_equal
= cloog_domain_lazy_equal(Q
->domain
,loop
->domain
)))
839 domain
= cloog_domain_copy(Q
->domain
) ;
841 domain
= cloog_domain_intersection(Q
->domain
,loop
->domain
) ;
843 if (!cloog_domain_isempty(domain
))
844 { new_inner
= cloog_loop_concat(cloog_loop_copy(Q
->inner
),
845 cloog_loop_copy(loop
->inner
)) ;
846 new_loop
= cloog_loop_alloc(loop
->state
, domain
, loop
->state
->one
,
847 loop
->state
->zero
, NULL
, new_inner
, NULL
);
848 cloog_loop_add_disjoint(&temp
,&now
,new_loop
) ;
852 cloog_domain_free(domain
);
856 /* Add (Q - loop). */
858 domain
= cloog_domain_copy(Q
->domain
) ;
861 domain
= cloog_domain_empty(Q
->domain
);
863 domain
= cloog_domain_difference(Q
->domain
,loop
->domain
) ;
866 if (!cloog_domain_isempty(domain
)) {
867 new_loop
= cloog_loop_alloc(loop
->state
, domain
, loop
->state
->one
,
868 loop
->state
->zero
, NULL
, Q
->inner
, NULL
);
869 cloog_loop_add_disjoint(&temp
,&now
,new_loop
) ;
872 { cloog_domain_free(domain
) ;
873 /* If Q->inner is no more useful, we can free it. */
876 cloog_loop_free(inner
) ;
880 /* Add loop-UQ associated with the blocks of loop alone.*/
881 if (cloog_domain_lazy_disjoint(loop
->domain
,UQ
))
882 domain
= cloog_domain_copy(loop
->domain
) ;
884 { if (cloog_domain_lazy_equal(loop
->domain
,UQ
))
885 domain
= cloog_domain_empty(UQ
);
887 domain
= cloog_domain_difference(loop
->domain
,UQ
) ;
890 if (!cloog_domain_isempty(domain
)) {
891 new_loop
= cloog_loop_alloc(loop
->state
, domain
, loop
->state
->one
,
892 loop
->state
->zero
, NULL
, loop
->inner
, NULL
);
893 cloog_loop_add_disjoint(&temp
,&now
,new_loop
) ;
896 { cloog_domain_free(domain
) ;
897 /* If loop->inner is no more useful, we can free it. */
898 cloog_loop_free(loop
->inner
) ;
903 if (loop
->next
!= NULL
)
904 UQ
= cloog_domain_union(UQ
, cloog_domain_copy(loop
->domain
));
906 cloog_domain_free(UQ
);
908 cloog_loop_free_parts(res
,1,0,0,1) ;
912 cloog_loop_free_parts(old
,1,0,0,1) ;
918 static CloogDomain
*bounding_domain(CloogDomain
*dom
, CloogOptions
*options
)
921 return cloog_domain_simple_convex(dom
);
923 return cloog_domain_convex(dom
);
928 * cloog_loop_merge function:
929 * This function is the 'soft' version of loop_separate if we are looking for
930 * a code much simpler (and less efficicient). This function returns the new
932 * - October 29th 2001: first version.
933 * - July 3rd->11th 2003: memory leaks hunt and correction.
934 * - June 22nd 2005: Adaptation for GMP.
936 CloogLoop
*cloog_loop_merge(CloogLoop
*loop
, int level
, CloogOptions
*options
)
938 CloogLoop
*res
, *new_inner
, *old
;
939 CloogDomain
*new_domain
, *temp
;
944 if (loop
->next
== NULL
)
945 return cloog_loop_disjoint(loop
);
950 new_inner
= loop
->inner
;
952 for (loop
= loop
->next
; loop
; loop
= loop
->next
) {
953 temp
= cloog_domain_union(temp
, loop
->domain
);
955 new_inner
= cloog_loop_concat(new_inner
, loop
->inner
);
958 new_domain
= bounding_domain(temp
, options
);
960 if (level
> 0 && !cloog_domain_is_bounded(new_domain
, level
) &&
961 cloog_domain_is_bounded(temp
, level
)) {
962 CloogDomain
*splitter
, *t2
;
964 cloog_domain_free(new_domain
);
965 splitter
= cloog_domain_bound_splitter(temp
, level
);
968 while (!cloog_domain_isconvex(splitter
)) {
969 CloogDomain
*first
, *rest
;
970 first
= cloog_domain_cut_first(splitter
, &rest
);
972 t2
= cloog_domain_intersection(first
, temp
);
973 cloog_domain_free(first
);
975 new_domain
= bounding_domain(t2
, options
);
976 cloog_domain_free(t2
);
978 if (cloog_domain_isempty(new_domain
)) {
979 cloog_domain_free(new_domain
);
982 res
= cloog_loop_alloc(old
->state
, new_domain
, old
->state
->one
,
983 old
->state
->zero
, NULL
,
984 cloog_loop_copy(new_inner
), res
);
987 t2
= cloog_domain_intersection(splitter
, temp
);
988 cloog_domain_free(splitter
);
990 new_domain
= bounding_domain(t2
, options
);
991 cloog_domain_free(t2
);
993 if (cloog_domain_isempty(new_domain
)) {
994 cloog_domain_free(new_domain
);
995 cloog_loop_free(new_inner
);
997 res
= cloog_loop_alloc(old
->state
, new_domain
, old
->state
->one
,
998 old
->state
->zero
, NULL
, new_inner
, res
);
1000 res
= cloog_loop_alloc(old
->state
, new_domain
, old
->state
->one
,
1001 old
->state
->zero
, NULL
, new_inner
, NULL
);
1003 cloog_domain_free(temp
);
1005 cloog_loop_free_parts(old
, 0, 0, 0, 1);
1011 static int cloog_loop_count(CloogLoop
*loop
)
1015 for (nb_loops
= 0; loop
; loop
= loop
->next
)
1023 * cloog_loop_sort function:
1024 * Adaptation from LoopGen 0.4 by F. Quillere. This function sorts a list of
1025 * parameterized disjoint polyhedra, in order to not have lexicographic order
1026 * violation (see Quillere paper).
1027 * - September 16th 2005: inclusion of cloog_loop_number (October 29th 2001).
1029 CloogLoop
*cloog_loop_sort(CloogLoop
*loop
, int level
)
1031 CloogLoop
*res
, *now
, **loop_array
;
1033 int i
, nb_loops
=0, * permut
;
1035 /* There is no need to sort the parameter domains. */
1039 /* We will need to know how many loops are in the list. */
1040 nb_loops
= cloog_loop_count(loop
);
1042 /* If there is only one loop, it's the end. */
1046 /* We have to allocate memory for some useful components:
1047 * - loop_array: the loop array,
1048 * - doms: the array of domains to sort,
1049 * - permut: will give us a possible sort (maybe not the only one).
1051 loop_array
= (CloogLoop
**)malloc(nb_loops
*sizeof(CloogLoop
*)) ;
1052 doms
= (CloogDomain
**)malloc(nb_loops
*sizeof(CloogDomain
*));
1053 permut
= (int *)malloc(nb_loops
*sizeof(int)) ;
1055 /* We fill up the loop and domain arrays. */
1056 for (i
=0;i
<nb_loops
;i
++,loop
=loop
->next
)
1057 { loop_array
[i
] = loop
;
1058 doms
[i
] = loop_array
[i
]->domain
;
1061 /* cloog_domain_sort will fill up permut. */
1062 cloog_domain_sort(doms
, nb_loops
, level
, permut
);
1064 /* With permut and loop_array we build the sorted list. */
1066 for (i
=0;i
<nb_loops
;i
++)
1067 { /* To avoid pointer looping... loop_add will rebuild the list. */
1068 loop_array
[permut
[i
]-1]->next
= NULL
;
1069 cloog_loop_add(&res
,&now
,loop_array
[permut
[i
]-1]) ;
1081 * cloog_loop_nest function:
1082 * This function changes the loop list in such a way that we have no more than
1083 * one dimension added by level. It returns an equivalent loop list with
1085 * - October 29th 2001: first version.
1086 * - July 3rd->11th 2003: memory leaks hunt and correction.
1087 * - June 22nd 2005: Adaptation for GMP.
1088 * - November 21th 2005: (debug) now OK when cloog_loop_restrict returns NULL.
1090 CloogLoop
*cloog_loop_nest(CloogLoop
*loop
, CloogDomain
*context
, int level
)
1092 CloogLoop
* p
, * temp
, * res
, * now
, * next
;
1093 CloogDomain
* new_domain
;
1095 loop
= cloog_loop_disjoint(loop
);
1098 /* Each domain is changed by its intersection with the context. */
1099 while (loop
!= NULL
)
1100 { p
= cloog_loop_restrict(loop
, context
);
1104 { cloog_loop_free_parts(loop
,1,0,0,0) ;
1106 temp
= cloog_loop_alloc(p
->state
, p
->domain
, p
->state
->one
,
1107 p
->state
->zero
, p
->block
, p
->inner
, NULL
);
1109 /* If the intersection dimension is too big, we make projections smaller
1110 * and smaller, and each projection includes the preceding projection
1111 * (thus, in the target list, dimensions are added one by one).
1113 if (cloog_domain_dimension(p
->domain
) >= level
)
1114 for (l
= cloog_domain_dimension(p
->domain
); l
>= level
; l
--) {
1115 new_domain
= cloog_domain_project(p
->domain
, l
);
1116 temp
= cloog_loop_alloc(p
->state
, new_domain
, p
->state
->one
,
1117 p
->state
->zero
, NULL
, temp
, NULL
);
1120 /* p is no more useful (but its content yes !). */
1121 cloog_loop_free_parts(p
,0,0,0,0) ;
1123 cloog_loop_add(&res
,&now
,temp
) ;
1126 cloog_loop_free_parts(loop
,1,1,1,0) ;
1136 * cloog_loop_stride function:
1137 * This function will find the stride of a loop for the iterator at the column
1138 * number 'level' in the constraint matrix. It will update the lower bound of
1139 * the iterator accordingly. Basically, the function will try to find in the
1140 * inner loops a common condition on this iterator for the inner loop iterators
1141 * to be integral. For instance, let us consider a loop with the iterator i,
1142 * the iteration domain -4<=i<=n, and its two inner loops with the iterator j.
1143 * The first inner loop has the constraint 3j=i, and the second one has the
1144 * constraint 6j=i. Then the common constraint on i for j to be integral is
1145 * i%3=0, the stride for i is 3. Lastly, we have to find the new lower bound
1146 * for i: the first value satisfying the common constraint: -3. At the end, the
1147 * iteration domain for i is -3<=i<=n and the stride for i is 3.
1148 * - loop is the loop including the iteration domain of the considered iterator,
1149 * - level is the column number of the iterator in the matrix of contraints.
1151 * - June 29th 2003: first version (work in progress since June 26th 2003).
1152 * - July 14th 2003: simpler version.
1153 * - June 22nd 2005: Adaptation for GMP (from S. Verdoolaege's 0.12.1 version).
1155 void cloog_loop_stride(CloogLoop
* loop
, int level
)
1156 { int first_search
;
1157 cloog_int_t stride
, ref_offset
, offset
, potential
, lower
;
1160 cloog_int_init(stride
);
1161 cloog_int_init(ref_offset
);
1162 cloog_int_init(offset
);
1163 cloog_int_init(potential
);
1164 cloog_int_init(lower
);
1166 cloog_int_set_si(ref_offset
, 0);
1167 cloog_int_set_si(offset
, 0);
1168 cloog_int_set_si(lower
, 0);
1170 /* Default stride. */
1171 cloog_int_set_si(stride
, 1);
1173 inner
= loop
->inner
;
1175 if (cloog_domain_integral_lowerbound(loop
->domain
,level
,&lower
))
1176 while (inner
!= NULL
)
1177 { /* If the minimun stride has not been found yet, find the stride. */
1178 if ((first_search
) || (!cloog_int_is_one(stride
)))
1180 cloog_domain_stride(inner
->domain
, level
, &potential
, &offset
);
1181 if (!cloog_int_is_one(potential
) && (!first_search
))
1182 { /* Offsets must be the same for common stride. */
1183 cloog_int_gcd(stride
, potential
, stride
);
1184 if (!cloog_int_is_zero(stride
)) {
1185 cloog_int_fdiv_r(offset
, offset
, stride
);
1186 cloog_int_fdiv_r(ref_offset
, ref_offset
, stride
);
1188 if (cloog_int_ne(offset
,ref_offset
))
1189 cloog_int_set_si(stride
, 1);
1192 cloog_int_set(stride
, potential
);
1193 cloog_int_set(ref_offset
, offset
);
1199 inner
= inner
->next
;
1202 if (cloog_int_is_zero(stride
))
1203 cloog_int_set_si(stride
, 1);
1205 /* Update the values if necessary. */
1206 if (!cloog_int_is_one(stride
))
1207 { /* Update the stride value. */
1208 cloog_int_set(loop
->stride
, stride
);
1209 if (!cloog_int_is_zero(offset
))
1210 cloog_int_sub(loop
->offset
, stride
, offset
);
1213 cloog_int_clear(stride
);
1214 cloog_int_clear(ref_offset
);
1215 cloog_int_clear(offset
);
1216 cloog_int_clear(potential
);
1217 cloog_int_clear(lower
);
1222 * cloog_loop_stop function:
1223 * This function implements the 'stop' option : each domain of each loop
1224 * in the list 'loop' is replaced by 'context'. 'context' should be the
1225 * domain of the outer loop. By using this method, there are no more dimensions
1226 * to scan and the simplification step will automaticaly remove the domains
1227 * since they are the same as the corresponding contexts. The effect of this
1228 * function is to stop the code generation at the level this function is called,
1229 * the resulting code do not consider the next dimensions.
1230 * - January 11th 2005: first version.
1232 CloogLoop
* cloog_loop_stop(CloogLoop
* loop
, CloogDomain
* context
)
1236 { cloog_domain_free(loop
->domain
) ;
1237 loop
->domain
= cloog_domain_copy(context
) ;
1238 loop
->next
= cloog_loop_stop(loop
->next
, context
) ;
1245 static int level_is_constant(int level
, int scalar
, int *scaldims
, int nb_scattdims
)
1247 return level
&& (level
+scalar
<= nb_scattdims
) && (scaldims
[level
+scalar
-1]);
1252 * Compare the constant dimensions of loops 'l1' and 'l2' starting at 'scalar'
1253 * and return -1 if the vector of constant dimensions of 'l1' is smaller
1254 * than that of 'l2', 0 if they are the same and +1 if that of 'l1' is
1255 * greater than that of 'l2'.
1256 * This function should be called on the innermost loop (the loop
1257 * containing a block).
1258 * \param l1 Loop to be compared with l2.
1259 * \param l2 Loop to be compared with l1.
1260 * \param level Current non-scalar dimension.
1261 * \param scaldims Boolean array saying whether a dimension is scalar or not.
1262 * \param nb_scattdims Size of the scaldims array.
1263 * \param scalar Current scalar dimension.
1264 * \return -1 if (l1 < l2), 0 if (l1 == l2) and +1 if (l1 > l2)
1266 int cloog_loop_constant_cmp(CloogLoop
*l1
, CloogLoop
*l2
, int level
,
1267 int *scaldims
, int nb_scattdims
, int scalar
)
1269 CloogBlock
*b1
, *b2
;
1272 while (level_is_constant(level
, scalar
, scaldims
, nb_scattdims
)) {
1273 int cmp
= cloog_int_cmp(b1
->scaldims
[scalar
], b2
->scaldims
[scalar
]);
1283 * cloog_loop_scalar_gt function:
1284 * This function returns 1 if loop 'l1' is greater than loop 'l2' for the
1285 * scalar dimension vector that begins at dimension 'scalar', 0 otherwise. What
1286 * we want to know is whether a loop is scheduled before another one or not.
1287 * This function solves the problem when the considered dimension for scheduling
1288 * is a scalar dimension. Since there may be a succession of scalar dimensions,
1289 * this function will reason about the vector of scalar dimension that begins
1290 * at dimension 'level+scalar' and finish to the first non-scalar dimension.
1291 * \param l1 Loop to be compared with l2.
1292 * \param l2 Loop to be compared with l1.
1293 * \param level Current non-scalar dimension.
1294 * \param scaldims Boolean array saying whether a dimension is scalar or not.
1295 * \param nb_scattdims Size of the scaldims array.
1296 * \param scalar Current scalar dimension.
1297 * \return 1 if (l1 > l2), 0 otherwise.
1299 * - September 9th 2005: first version.
1300 * - October 15nd 2007: now "greater than" instead of "greater or equal".
1302 int cloog_loop_scalar_gt(l1
, l2
, level
, scaldims
, nb_scattdims
, scalar
)
1303 CloogLoop
* l1
, * l2
;
1304 int level
, * scaldims
, nb_scattdims
, scalar
;
1306 return cloog_loop_constant_cmp(l1
, l2
, level
, scaldims
, nb_scattdims
, scalar
) > 0;
1311 * cloog_loop_scalar_eq function:
1312 * This function returns 1 if loop 'l1' is equal to loop 'l2' for the scalar
1313 * dimension vector that begins at dimension 'scalar', 0 otherwise. What we want
1314 * to know is whether two loops are scheduled for the same time or not.
1315 * This function solves the problem when the considered dimension for scheduling
1316 * is a scalar dimension. Since there may be a succession of scalar dimensions,
1317 * this function will reason about the vector of scalar dimension that begins
1318 * at dimension 'level+scalar' and finish to the first non-scalar dimension.
1319 * - l1 and l2 are the loops to compare,
1320 * - level is the current non-scalar dimension,
1321 * - scaldims is the boolean array saying whether a dimension is scalar or not,
1322 * - nb_scattdims is the size of the scaldims array,
1323 * - scalar is the current scalar dimension.
1325 * - September 9th 2005 : first version.
1327 int cloog_loop_scalar_eq(l1
, l2
, level
, scaldims
, nb_scattdims
, scalar
)
1328 CloogLoop
* l1
, * l2
;
1329 int level
, * scaldims
, nb_scattdims
, scalar
;
1331 return cloog_loop_constant_cmp(l1
, l2
, level
, scaldims
, nb_scattdims
, scalar
) == 0;
1336 * cloog_loop_scalar_sort function:
1337 * This function sorts a linked list of loops (loop) with respect to the
1338 * scalar dimension vector that begins at dimension 'scalar'. Since there may
1339 * be a succession of scalar dimensions, this function will reason about the
1340 * vector of scalar dimension that begins at dimension 'level+scalar' and
1341 * finish to the first non-scalar dimension.
1342 * \param loop Loop list to sort.
1343 * \param level Current non-scalar dimension.
1344 * \param scaldims Boolean array saying whether a dimension is scalar or not.
1345 * \param nb_scattdims Size of the scaldims array.
1346 * \param scalar Current scalar dimension.
1347 * \return A pointer to the sorted list.
1349 * - July 2nd 2005: first developments.
1350 * - September 2nd 2005: first version.
1351 * - October 15nd 2007: complete rewrite to remove bugs, now a bubble sort.
1353 CloogLoop
* cloog_loop_scalar_sort(loop
, level
, scaldims
, nb_scattdims
, scalar
)
1355 int level
, * scaldims
, nb_scattdims
, scalar
;
1357 CloogLoop
**current
;
1361 for (current
= &loop
; (*current
)->next
; current
= &(*current
)->next
) {
1362 CloogLoop
*next
= (*current
)->next
;
1363 if (cloog_loop_scalar_gt(*current
,next
,level
,scaldims
,nb_scattdims
,scalar
)) {
1365 (*current
)->next
= next
->next
;
1366 next
->next
= *current
;
1377 * cloog_loop_generate_backtrack function:
1378 * adaptation from LoopGen 0.4 by F. Quillere. This function implements the
1379 * backtrack of the Quillere et al. algorithm (see the Quillere paper).
1380 * It eliminates unused iterations of the current level for the new one. See the
1381 * example called linearity-1-1 example with and without this part for an idea.
1382 * - October 26th 2001: first version in cloog_loop_generate_general.
1383 * - July 31th 2002: (debug) no more parasite loops (REALLY hard !).
1384 * - October 30th 2005: extraction from cloog_loop_generate_general.
1386 CloogLoop
*cloog_loop_generate_backtrack(CloogLoop
*loop
,
1387 int level
, CloogOptions
*options
)
1389 CloogDomain
* domain
;
1390 CloogLoop
* now
, * now2
, * next
, * next2
, * end
, * temp
, * l
, * inner
,
1396 while (temp
!= NULL
)
1398 inner
= temp
->inner
;
1400 while (inner
!= NULL
)
1401 { next
= inner
->next
;
1402 /* This 'if' and its first part is the debug of july 31th 2002. */
1403 if (inner
->block
!= NULL
) {
1404 end
= cloog_loop_alloc(temp
->state
, inner
->domain
, temp
->state
->one
,
1405 temp
->state
->zero
, inner
->block
, NULL
, NULL
);
1406 domain
= cloog_domain_copy(temp
->domain
) ;
1407 new_loop
= cloog_loop_alloc(temp
->state
, domain
, temp
->state
->one
,
1408 temp
->state
->zero
, NULL
, end
, NULL
);
1411 new_loop
= cloog_loop_project(inner
, level
);
1413 cloog_loop_free_parts(inner
,0,0,0,0) ;
1414 cloog_loop_add(&l
,&now2
,new_loop
) ;
1418 temp
->inner
= NULL
;
1421 { l
= cloog_loop_separate(l
) ;
1422 l
= cloog_loop_sort(l
, level
);
1424 cloog_int_set(l
->stride
, temp
->stride
);
1425 cloog_int_set(l
->offset
, temp
->offset
);
1426 cloog_loop_add(&loop
,&now
,l
) ;
1430 next2
= temp
->next
;
1431 cloog_loop_free_parts(temp
,1,0,0,0) ;
1440 * Return 1 if we need to continue recursing to the specified level.
1442 int cloog_loop_more(CloogLoop
*loop
, int level
, int scalar
, int nb_scattdims
)
1444 return level
+ scalar
<= nb_scattdims
||
1445 cloog_domain_dimension(loop
->domain
) >= level
;
1448 CloogLoop
*cloog_loop_generate_restricted_or_stop(CloogLoop
*loop
,
1449 CloogDomain
*context
,
1450 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
1451 CloogOptions
*options
);
1454 * cloog_loop_generate_general function:
1455 * Adaptation from LoopGen 0.4 by F. Quillere. This function implements the
1456 * Quillere algorithm for polyhedron scanning from step 3 to 5.
1457 * (see the Quillere paper).
1458 * - loop is the loop for which we have to generate a scanning code,
1459 * - level is the current non-scalar dimension,
1460 * - scalar is the current scalar dimension,
1461 * - scaldims is the boolean array saying whether a dimension is scalar or not,
1462 * - nb_scattdims is the size of the scaldims array,
1463 * - options are the general code generation options.
1465 * - October 26th 2001: first version.
1466 * - July 3rd->11th 2003: memory leaks hunt and correction.
1467 * - June 22nd 2005: Adaptation for GMP.
1468 * - September 2nd 2005: The function have been cutted out in two pieces:
1469 * cloog_loop_generate and this one, in order to handle
1470 * the scalar dimension case more efficiently with
1471 * cloog_loop_generate_scalar.
1472 * - November 15th 2005: (debug) the result of the cloog_loop_generate call may
1473 * be a list of polyhedra (especially if stop option is
1474 * used): cloog_loop_add_list instead of cloog_loop_add.
1476 CloogLoop
*cloog_loop_generate_general(CloogLoop
*loop
,
1477 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
1478 CloogOptions
*options
)
1480 CloogLoop
* res
, * now
, * temp
, * l
, * new_loop
, * inner
, * now2
, * end
,
1482 CloogDomain
* domain
;
1485 /* 3. Separate all projections into disjoint polyhedra. */
1486 if ((options
->f
> level
+scalar
) || (options
->f
< 0))
1487 res
= cloog_loop_merge(loop
, level
, options
);
1489 res
= cloog_loop_separate(loop
);
1493 /* 3b. -correction- sort the loops to determine their textual order. */
1494 res
= cloog_loop_sort(res
, level
);
1496 res
= cloog_loop_restrict_inner(res
);
1499 res
= cloog_loop_specialize(res
, level
, scalar
, scaldims
, nb_scattdims
);
1501 /* 4. Recurse for each loop with the current domain as context. */
1504 if (!level
|| (level
+scalar
< options
->l
) || (options
->l
< 0))
1506 { if (level
&& options
->strides
)
1507 cloog_loop_stride(temp
, level
);
1508 inner
= temp
->inner
;
1509 domain
= temp
->domain
;
1511 while (inner
!= NULL
)
1512 { /* 4b. -ced- recurse for each sub-list of non terminal loops. */
1513 if (cloog_loop_more(inner
, level
+ 1, scalar
, nb_scattdims
)) {
1515 while ((end
->next
!= NULL
) &&
1516 cloog_loop_more(end
->next
, level
+ 1, scalar
, nb_scattdims
))
1522 l
= cloog_loop_generate_restricted_or_stop(inner
, domain
,
1523 level
+ 1, scalar
, scaldims
, nb_scattdims
, options
);
1526 cloog_loop_add_list(&into
,&now
,l
) ;
1531 { cloog_loop_add(&into
,&now
,inner
) ;
1532 inner
= inner
->next
;
1537 temp
->inner
= into
;
1538 cloog_loop_add(&res
,&now2
,temp
) ;
1542 while (temp
!= NULL
)
1543 { next
= temp
->next
;
1544 l
= cloog_loop_nest(temp
->inner
, temp
->domain
, level
+1);
1545 new_loop
= cloog_loop_alloc(temp
->state
, temp
->domain
, temp
->state
->one
,
1546 temp
->state
->zero
, NULL
, l
, NULL
);
1547 temp
->inner
= NULL
;
1549 cloog_loop_free_parts(temp
,0,0,0,0) ;
1550 cloog_loop_add(&res
,&now
,new_loop
) ;
1554 /* 5. eliminate unused iterations of the current level for the new one. See
1555 * the example called linearity-1-1 example with and without this part
1558 if (options
->backtrack
&& level
&&
1559 ((level
+scalar
< options
->l
) || (options
->l
< 0)) &&
1560 ((options
->f
<= level
+scalar
) && !(options
->f
< 0)))
1561 res
= cloog_loop_generate_backtrack(res
, level
, options
);
1563 /* Pray for my new paper to be accepted somewhere since the following stuff
1564 * is really amazing :-) !
1565 * Far long later: The paper has been accepted to PACT 2004 :-))). But there
1566 * are still some bugs and I have no time to fix them. Thus now you have to
1567 * pray for me to get an academic position for that really amazing stuff :-) !
1568 * Later again: OK, I get my academic position, but still I have not enough
1569 * time to fix and clean this part... Pray again :-) !!!
1571 /* res = cloog_loop_unisolate(res,level) ;*/
1577 CloogLoop
*cloog_loop_generate_restricted(CloogLoop
*loop
,
1578 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
1579 CloogOptions
*options
);
1583 * cloog_loop_generate_scalar function:
1584 * This function applies the simplified code generation scheme in the trivial
1585 * case of scalar dimensions. When dealing with scalar dimensions, there is
1586 * no need of costly polyhedral operations for separation or sorting: sorting
1587 * is a question of comparing scalar vectors and separation amounts to consider
1588 * only loops with the same scalar vector for the next step of the code
1589 * generation process. This function achieves the separation/sorting process
1590 * for the vector of scalar dimension that begins at dimension 'level+scalar'
1591 * and finish to the first non-scalar dimension.
1592 * - loop is the loop for which we have to generate a scanning code,
1593 * - level is the current non-scalar dimension,
1594 * - scalar is the current scalar dimension,
1595 * - scaldims is the boolean array saying whether a dimension is scalar or not,
1596 * - nb_scattdims is the size of the scaldims array,
1597 * - options are the general code generation options.
1599 * - September 2nd 2005: First version.
1601 CloogLoop
*cloog_loop_generate_scalar(CloogLoop
*loop
,
1602 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
1603 CloogOptions
*options
)
1604 { CloogLoop
* res
, * now
, * temp
, * l
, * end
, * next
, * ref
;
1607 /* We sort the loop list with respect to the current scalar vector. */
1608 res
= cloog_loop_scalar_sort(loop
,level
,scaldims
,nb_scattdims
,scalar
) ;
1610 scalar_new
= scalar
+ scaldims
[level
+ scalar
- 1];
1614 while (temp
!= NULL
)
1615 { /* Then we will appy the general code generation process to each sub-list
1616 * of loops with the same scalar vector.
1621 while((end
->next
!= NULL
) &&
1622 cloog_loop_more(end
->next
, level
, scalar_new
, nb_scattdims
) &&
1623 cloog_loop_scalar_eq(ref
,end
->next
,level
,scaldims
,nb_scattdims
,scalar
))
1629 /* For the next dimension, scalar value is updated by adding the scalar
1630 * vector size, which is stored at scaldims[level+scalar-1].
1632 if (cloog_loop_more(temp
, level
, scalar_new
, nb_scattdims
)) {
1633 l
= cloog_loop_generate_restricted(temp
, level
, scalar_new
,
1634 scaldims
, nb_scattdims
, options
);
1637 cloog_loop_add_list(&res
, &now
, l
);
1639 cloog_loop_add(&res
, &now
, temp
);
1648 /* Compare loop with the next loop based on their constant dimensions.
1649 * The result is < 0, == 0 or > 0 depending on whether the constant
1650 * dimensions of loop are lexicographically smaller, equal or greater
1651 * than those of loop->next.
1652 * If loop is the last in the list, then it is assumed to be smaller
1653 * than the "next" one.
1655 static int cloog_loop_next_scal_cmp(CloogLoop
*loop
)
1663 nb_scaldims
= loop
->block
->nb_scaldims
;
1664 if (loop
->next
->block
->nb_scaldims
< nb_scaldims
)
1665 nb_scaldims
= loop
->next
->block
->nb_scaldims
;
1667 for (i
= 0; i
< nb_scaldims
; ++i
) {
1668 int cmp
= cloog_int_cmp(loop
->block
->scaldims
[i
],
1669 loop
->next
->block
->scaldims
[i
]);
1673 return loop
->block
->nb_scaldims
- loop
->next
->block
->nb_scaldims
;
1677 /* Check whether the globally constant dimensions of a and b
1678 * have the same value for all globally constant dimensions
1679 * that are situated before any (locally) non-constant dimension.
1681 static int cloog_loop_equal_prefix(CloogLoop
*a
, CloogLoop
*b
,
1682 int *scaldims
, int nb_scattdims
)
1688 for (i
= 0; i
< nb_scattdims
; ++i
) {
1693 if (!cloog_int_eq(a
->block
->scaldims
[cst
], b
->block
->scaldims
[cst
]))
1697 for (i
= i
+ 1; i
< nb_scattdims
; ++i
) {
1700 if (!cloog_domain_lazy_isconstant(a
->domain
, dim
))
1702 /* No need to check that dim is also constant in b and that the
1703 * constant values are equal. That will happen during the check
1704 * whether the two domains are equal.
1712 /* Try to block adjacent loops in the loop list "loop".
1713 * We only attempt blocking if the constant dimensions of the loops
1714 * in the least are (not necessarily strictly) increasing.
1715 * Then we look for a sublist such that the first (begin) has constant
1716 * dimensions strictly larger than the previous loop in the complete
1717 * list and such that the loop (end) after the last loop in the sublist
1718 * has constant dimensions strictly larger than the last loop in the sublist.
1719 * Furthermore, all loops in the sublist should have the same domain
1720 * (with globally constant dimensions removed) and the difference
1721 * (if any) in constant dimensions may only occur after all the
1722 * (locally) constant dimensions.
1723 * If we find such a sublist, then the blocks of all but the first
1724 * are merged into the block of the first.
1726 * Note that this function can only be called before the global
1727 * blocklist has been created because it may otherwise modify and destroy
1728 * elements on that list.
1730 CloogLoop
*cloog_loop_block(CloogLoop
*loop
, int *scaldims
, int nb_scattdims
)
1732 CloogLoop
*begin
, *end
, *l
;
1733 int begin_after_previous
;
1734 int end_after_previous
;
1738 for (begin
= loop
; begin
; begin
= begin
->next
) {
1739 if (!begin
->block
|| !begin
->block
->scaldims
)
1741 if (cloog_loop_next_scal_cmp(loop
) > 0)
1745 begin_after_previous
= 1;
1746 for (begin
= loop
; begin
; begin
= begin
->next
) {
1747 if (!begin_after_previous
) {
1748 begin_after_previous
= cloog_loop_next_scal_cmp(begin
) < 0;
1752 end_after_previous
= cloog_loop_next_scal_cmp(begin
) < 0;
1753 for (end
= begin
->next
; end
; end
= end
->next
) {
1754 if (!cloog_loop_equal_prefix(begin
, end
, scaldims
, nb_scattdims
))
1756 if (!cloog_domain_lazy_equal(begin
->domain
, end
->domain
))
1758 end_after_previous
= cloog_loop_next_scal_cmp(end
) < 0;
1760 if (end
!= begin
->next
&& end_after_previous
) {
1761 for (l
= begin
->next
; l
!= end
; l
= begin
->next
) {
1762 cloog_block_merge(begin
->block
, l
->block
);
1763 begin
->next
= l
->next
;
1764 cloog_loop_free_parts(l
, 1, 0, 1, 0);
1768 begin_after_previous
= cloog_loop_next_scal_cmp(begin
) < 0;
1776 * Check whether for any fixed iteration of the outer loops,
1777 * there is an iteration of loop1 that is lexicographically greater
1778 * than an iteration of loop2.
1779 * Return 1 if there exists (or may exist) such a pair.
1780 * Return 0 if all iterations of loop1 are lexicographically smaller
1781 * than the iterations of loop2.
1782 * If no iteration is lexicographically greater, but if there are
1783 * iterations that are equal to iterations of loop2, then return "def".
1784 * This is useful for ensuring that such statements are not reordered.
1785 * Some users, including the test_run target in test, expect
1786 * the statements at a given point to be run in the original order.
1787 * Passing the value "0" for "def" would allow such statements to be reordered
1788 * and would allow for the detection of more components.
1790 int cloog_loop_follows(CloogLoop
*loop1
, CloogLoop
*loop2
,
1791 int level
, int scalar
, int *scaldims
, int nb_scattdims
, int def
)
1795 dim1
= cloog_domain_dimension(loop1
->domain
);
1796 dim2
= cloog_domain_dimension(loop2
->domain
);
1797 while ((level
<= dim1
&& level
<= dim2
) ||
1798 level_is_constant(level
, scalar
, scaldims
, nb_scattdims
)) {
1799 if (level_is_constant(level
, scalar
, scaldims
, nb_scattdims
)) {
1800 int cmp
= cloog_loop_constant_cmp(loop1
, loop2
, level
, scaldims
,
1801 nb_scattdims
, scalar
);
1806 scalar
+= scaldims
[level
+ scalar
- 1];
1808 int follows
= cloog_domain_follows(loop1
->domain
, loop2
->domain
,
1822 /* Structure for representing the nodes in the graph being traversed
1823 * using Tarjan's algorithm.
1824 * index represents the order in which nodes are visited.
1825 * min_index is the index of the root of a (sub)component.
1826 * on_stack indicates whether the node is currently on the stack.
1828 struct cloog_loop_sort_node
{
1833 /* Structure for representing the graph being traversed
1834 * using Tarjan's algorithm.
1835 * len is the number of nodes
1836 * node is an array of nodes
1837 * stack contains the nodes on the path from the root to the current node
1838 * sp is the stack pointer
1839 * index is the index of the last node visited
1840 * order contains the elements of the components separated by -1
1841 * op represents the current position in order
1843 struct cloog_loop_sort
{
1845 struct cloog_loop_sort_node
*node
;
1853 /* Allocate and initialize cloog_loop_sort structure.
1855 static struct cloog_loop_sort
*cloog_loop_sort_alloc(int len
)
1857 struct cloog_loop_sort
*s
;
1860 s
= (struct cloog_loop_sort
*)malloc(sizeof(struct cloog_loop_sort
));
1863 s
->node
= (struct cloog_loop_sort_node
*)
1864 malloc(len
* sizeof(struct cloog_loop_sort_node
));
1866 for (i
= 0; i
< len
; ++i
)
1867 s
->node
[i
].index
= -1;
1868 s
->stack
= (int *)malloc(len
* sizeof(int));
1870 s
->order
= (int *)malloc(2 * len
* sizeof(int));
1880 /* Free cloog_loop_sort structure.
1882 static void cloog_loop_sort_free(struct cloog_loop_sort
*s
)
1891 /* Check whether for any fixed iteration of the outer loops,
1892 * there is an iteration of loop1 that is lexicographically greater
1893 * than an iteration of loop2, where the iteration domains are
1894 * available in the inner loops of the arguments.
1896 * By using this functions to detect components, we ensure that
1897 * two CloogLoops appear in the same component if some iterations of
1898 * each loop should be executed before some iterations of the other loop.
1899 * Since we also want two CloogLoops that have exactly the same
1900 * iteration domain at the current level to be placed in the same component,
1901 * we first check if these domains are indeed the same.
1903 static int inner_loop_follows(CloogLoop
*loop1
, CloogLoop
*loop2
,
1904 int level
, int scalar
, int *scaldims
, int nb_scattdims
, int def
)
1908 f
= cloog_domain_lazy_equal(loop1
->domain
, loop2
->domain
);
1910 f
= cloog_loop_follows(loop1
->inner
, loop2
->inner
,
1911 level
, scalar
, scaldims
, nb_scattdims
, def
);
1917 /* Perform Tarjan's algorithm for computing the strongly connected components
1918 * in the graph with the individual CloogLoops as vertices.
1919 * Two CloopLoops appear in the same component if they both (indirectly)
1920 * "follow" each other, where the following relation is determined
1921 * by the follows function.
1923 static void cloog_loop_components_tarjan(struct cloog_loop_sort
*s
,
1924 CloogLoop
**loop_array
, int i
, int level
, int scalar
, int *scaldims
,
1926 int (*follows
)(CloogLoop
*loop1
, CloogLoop
*loop2
,
1927 int level
, int scalar
, int *scaldims
, int nb_scattdims
, int def
))
1931 s
->node
[i
].index
= s
->index
;
1932 s
->node
[i
].min_index
= s
->index
;
1933 s
->node
[i
].on_stack
= 1;
1935 s
->stack
[s
->sp
++] = i
;
1937 for (j
= s
->len
- 1; j
>= 0; --j
) {
1942 if (s
->node
[j
].index
>= 0 &&
1943 (!s
->node
[j
].on_stack
||
1944 s
->node
[j
].index
> s
->node
[i
].min_index
))
1947 f
= follows(loop_array
[i
], loop_array
[j
],
1948 level
, scalar
, scaldims
, nb_scattdims
, i
> j
);
1952 if (s
->node
[j
].index
< 0) {
1953 cloog_loop_components_tarjan(s
, loop_array
, j
, level
, scalar
,
1954 scaldims
, nb_scattdims
, follows
);
1955 if (s
->node
[j
].min_index
< s
->node
[i
].min_index
)
1956 s
->node
[i
].min_index
= s
->node
[j
].min_index
;
1957 } else if (s
->node
[j
].index
< s
->node
[i
].min_index
)
1958 s
->node
[i
].min_index
= s
->node
[j
].index
;
1961 if (s
->node
[i
].index
!= s
->node
[i
].min_index
)
1965 j
= s
->stack
[--s
->sp
];
1966 s
->node
[j
].on_stack
= 0;
1967 s
->order
[s
->op
++] = j
;
1969 s
->order
[s
->op
++] = -1;
1973 static int qsort_index_cmp(const void *p1
, const void *p2
)
1975 return *(int *)p1
- *(int *)p2
;
1978 /* Sort the elements of the component starting at list.
1979 * The list is terminated by a -1.
1981 static void sort_component(int *list
)
1985 for (len
= 0; list
[len
] != -1; ++len
)
1988 qsort(list
, len
, sizeof(int), qsort_index_cmp
);
1991 /* Given an array of indices "list" into the "loop_array" array,
1992 * terminated by -1, construct a linked list of the corresponding
1993 * entries and put the result in *res.
1994 * The value returned is the number of CloogLoops in the (linked) list
1996 static int extract_component(CloogLoop
**loop_array
, int *list
, CloogLoop
**res
)
2000 sort_component(list
);
2001 while (list
[i
] != -1) {
2002 *res
= loop_array
[list
[i
]];
2003 res
= &(*res
)->next
;
2013 * Call cloog_loop_generate_scalar or cloog_loop_generate_general
2014 * on each of the strongly connected components in the list of CloogLoops
2015 * pointed to by "loop".
2017 * We use Tarjan's algorithm to find the strongly connected components.
2018 * Note that this algorithm also topologically sorts the components.
2020 * The components are treated separately to avoid spurious separations.
2021 * The concatentation of the results may contain successive loops
2022 * with the same bounds, so we try to combine such loops.
2024 CloogLoop
*cloog_loop_generate_components(CloogLoop
*loop
,
2025 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
2026 CloogOptions
*options
)
2030 CloogLoop
*res
, **res_next
;
2031 CloogLoop
**loop_array
;
2032 struct cloog_loop_sort
*s
;
2034 if (level
== 0 || !loop
->next
)
2035 return cloog_loop_generate_general(loop
, level
, scalar
,
2036 scaldims
, nb_scattdims
, options
);
2038 nb_loops
= cloog_loop_count(loop
);
2040 loop_array
= (CloogLoop
**)malloc(nb_loops
* sizeof(CloogLoop
*));
2043 for (i
= 0, tmp
= loop
; i
< nb_loops
; i
++, tmp
= tmp
->next
)
2044 loop_array
[i
] = tmp
;
2046 s
= cloog_loop_sort_alloc(nb_loops
);
2047 for (i
= nb_loops
- 1; i
>= 0; --i
) {
2048 if (s
->node
[i
].index
>= 0)
2050 cloog_loop_components_tarjan(s
, loop_array
, i
, level
, scalar
, scaldims
,
2051 nb_scattdims
, &inner_loop_follows
);
2058 int n
= extract_component(loop_array
, &s
->order
[i
], &tmp
);
2061 *res_next
= cloog_loop_generate_general(tmp
, level
, scalar
,
2062 scaldims
, nb_scattdims
, options
);
2064 res_next
= &(*res_next
)->next
;
2067 cloog_loop_sort_free(s
);
2071 res
= cloog_loop_combine(res
);
2077 /* For each loop in the list "loop", decompose the list of
2078 * inner loops into strongly connected components and put
2079 * the components into separate loops at the top level.
2081 CloogLoop
*cloog_loop_decompose_inner(CloogLoop
*loop
,
2082 int level
, int scalar
, int *scaldims
, int nb_scattdims
)
2085 CloogLoop
**loop_array
;
2086 int i
, n_loops
, max_loops
= 0;
2087 struct cloog_loop_sort
*s
;
2089 for (l
= loop
; l
; l
= l
->next
) {
2090 n_loops
= cloog_loop_count(l
->inner
);
2091 if (max_loops
< n_loops
)
2092 max_loops
= n_loops
;
2098 loop_array
= (CloogLoop
**)malloc(max_loops
* sizeof(CloogLoop
*));
2101 for (l
= loop
; l
; l
= l
->next
) {
2104 for (i
= 0, tmp
= l
->inner
; tmp
; i
++, tmp
= tmp
->next
)
2105 loop_array
[i
] = tmp
;
2110 s
= cloog_loop_sort_alloc(n_loops
);
2111 for (i
= n_loops
- 1; i
>= 0; --i
) {
2112 if (s
->node
[i
].index
>= 0)
2114 cloog_loop_components_tarjan(s
, loop_array
, i
, level
, scalar
,
2115 scaldims
, nb_scattdims
, &cloog_loop_follows
);
2118 n
= extract_component(loop_array
, s
->order
, &l
->inner
);
2124 n
= extract_component(loop_array
, &s
->order
[i
], &inner
);
2127 tmp
= cloog_loop_alloc(l
->state
, cloog_domain_copy(l
->domain
),
2128 l
->stride
, l
->offset
, l
->block
, inner
, l
->next
);
2133 cloog_loop_sort_free(s
);
2142 CloogLoop
*cloog_loop_generate_restricted(CloogLoop
*loop
,
2143 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
2144 CloogOptions
*options
)
2146 /* To save both time and memory, we switch here depending on whether the
2147 * current dimension is scalar (simplified processing) or not (general
2150 if (level_is_constant(level
, scalar
, scaldims
, nb_scattdims
))
2151 return cloog_loop_generate_scalar(loop
, level
, scalar
,
2152 scaldims
, nb_scattdims
, options
);
2154 * 2. Compute the projection of each polyhedron onto the outermost
2155 * loop variable and the parameters.
2157 loop
= cloog_loop_project_all(loop
, level
);
2159 return cloog_loop_generate_components(loop
, level
, scalar
, scaldims
,
2160 nb_scattdims
, options
);
2164 CloogLoop
*cloog_loop_generate_restricted_or_stop(CloogLoop
*loop
,
2165 CloogDomain
*context
,
2166 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
2167 CloogOptions
*options
)
2169 /* If the user asked to stop code generation at this level, let's stop. */
2170 if ((options
->stop
>= 0) && (level
+scalar
>= options
->stop
+1))
2171 return cloog_loop_stop(loop
,context
) ;
2173 return cloog_loop_generate_restricted(loop
, level
, scalar
, scaldims
,
2174 nb_scattdims
, options
);
2179 * cloog_loop_generate function:
2180 * Adaptation from LoopGen 0.4 by F. Quillere. This function implements the
2181 * Quillere algorithm for polyhedron scanning from step 1 to 2.
2182 * (see the Quillere paper).
2183 * - loop is the loop for which we have to generate a scanning code,
2184 * - context is the context of the current loop (constraints on parameter and/or
2185 * on outer loop counters),
2186 * - level is the current non-scalar dimension,
2187 * - scalar is the current scalar dimension,
2188 * - scaldims is the boolean array saying whether a dimension is scalar or not,
2189 * - nb_scattdims is the size of the scaldims array,
2190 * - options are the general code generation options.
2192 * - October 26th 2001: first version.
2193 * - July 3rd->11th 2003: memory leaks hunt and correction.
2194 * - June 15th 2005: a memory leak fixed (loop was not entirely freed when
2195 * the result of cloog_loop_restrict was NULL).
2196 * - June 22nd 2005: Adaptation for GMP.
2197 * - September 2nd 2005: The function have been cutted out in two pieces:
2198 * cloog_loop_generate and this one, in order to handle
2199 * the scalar dimension case more efficiently with
2200 * cloog_loop_generate_scalar.
2201 * - November 15th 2005: (debug) Condition for stop option no more take care of
2202 * further scalar dimensions.
2204 CloogLoop
*cloog_loop_generate(CloogLoop
*loop
, CloogDomain
*context
,
2205 int level
, int scalar
, int *scaldims
, int nb_scattdims
,
2206 CloogOptions
*options
)
2208 /* 1. Replace each polyhedron by its intersection with the context.
2210 loop
= cloog_loop_restrict_all(loop
, context
);
2214 return cloog_loop_generate_restricted_or_stop(loop
, context
,
2215 level
, scalar
, scaldims
, nb_scattdims
, options
);
2220 * Internal function for simplifying a single loop in a list of loops.
2221 * See cloog_loop_simplify.
2223 static CloogLoop
*loop_simplify(CloogLoop
*loop
, CloogDomain
*context
,
2227 CloogBlock
* new_block
;
2228 CloogLoop
*simplified
, *inner
;
2229 CloogDomain
* domain
, * simp
, * inter
, * extended_context
;
2231 if (!cloog_domain_isconvex(loop
->domain
))
2232 loop
->domain
= cloog_domain_simplify_union(loop
->domain
);
2234 domain
= loop
->domain
;
2236 domain_dim
= cloog_domain_dimension(domain
);
2237 extended_context
= cloog_domain_extend(context
, domain_dim
);
2238 inter
= cloog_domain_intersection(domain
,extended_context
) ;
2239 simp
= cloog_domain_simplify(inter
,extended_context
) ;
2240 cloog_domain_free(extended_context
) ;
2242 /* If the constraint system is never true, go to the next one. */
2243 if (cloog_domain_never_integral(simp
)) {
2244 cloog_loop_free(loop
->inner
);
2245 cloog_domain_free(inter
);
2246 cloog_domain_free(simp
);
2250 inner
= cloog_loop_simplify(loop
->inner
, inter
, level
+1);
2251 cloog_domain_free(inter
) ;
2253 if ((inner
== NULL
) && (loop
->block
== NULL
)) {
2254 cloog_domain_free(simp
);
2258 new_block
= cloog_block_copy(loop
->block
) ;
2260 simplified
= cloog_loop_alloc(loop
->state
, simp
, loop
->stride
, loop
->offset
,
2261 new_block
, inner
, NULL
);
2263 return(simplified
) ;
2268 * cloog_loop_simplify function:
2269 * This function implements the part 6. of the Quillere algorithm, it
2270 * recursively simplifies each loop in the context of the preceding loop domain.
2271 * It returns a pointer to the simplified loop list.
2272 * The cloog_domain_simplify (DomainSimplify) behaviour is really bad with
2273 * polyhedra union and some really awful sidesteppings were written, I plan
2275 * - October 31th 2001: first version.
2276 * - July 3rd->11th 2003: memory leaks hunt and correction.
2277 * - April 16th 2005: a memory leak fixed (extended_context was not freed).
2278 * - June 15th 2005: a memory leak fixed (loop was not conveniently freed
2279 * when the constraint system is never true).
2280 * - October 27th 2005: - this function called before cloog_loop_fast_simplify
2281 * is now the official cloog_loop_simplify function in
2282 * replacement of a slower and more complex one (after
2283 * deep changes in the pretty printer).
2284 * - we use cloog_loop_disjoint to fix the problem when
2285 * simplifying gives a union of polyhedra (before, it
2286 * was under the responsibility of the pretty printer).
2288 CloogLoop
*cloog_loop_simplify(CloogLoop
*loop
, CloogDomain
*context
, int level
)
2291 CloogLoop
*res
= NULL
;
2292 CloogLoop
**next
= &res
;
2294 for (now
= loop
; now
; now
= now
->next
) {
2295 *next
= loop_simplify(now
, context
, level
);
2297 now
->inner
= NULL
; /* For loop integrity. */
2298 cloog_domain_free(now
->domain
);
2302 next
= &(*next
)->next
;
2304 cloog_loop_free(loop
);
2306 /* Examples like test/iftest2.cloog give unions of polyhedra after
2307 * simplifying, thus we we have to disjoint them. Another good reason to
2308 * put the simplifying step in the Quillere backtrack.
2310 res
= cloog_loop_disjoint(res
);
2317 * cloog_loop_scatter function:
2318 * This function add the scattering (scheduling) informations in a loop.
2320 void cloog_loop_scatter(CloogLoop
* loop
, CloogScattering
*scatt
)
2322 loop
->domain
= cloog_domain_scatter(loop
->domain
, scatt
);