Clarify interpretation of limited for wide values
[libfirm] / ir / be / becopyilp2.c
1 /*
2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
3  *
4  * This file is part of libFirm.
5  *
6  * This file may be distributed and/or modified under the terms of the
7  * GNU General Public License version 2 as published by the Free Software
8  * Foundation and appearing in the file LICENSE.GPL included in the
9  * packaging of this file.
10  *
11  * Licensees holding valid libFirm Professional Edition licenses may use
12  * this file in accordance with the libFirm Commercial License.
13  * Agreement provided with the Software.
14  *
15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
17  * PURPOSE.
18  */
19
20 /**
21  * @file
22  * @brief       ILP based copy minimization.
23  * @author      Daniel Grund
24  * @date        28.02.2006
25  *
26  * ILP formalization using G=(V, E, Q):
27  *  - 2 class of variables: coloring vars x_ic   and   equal color vars y_ij
28  *  - Path constraints
29  *  - Clique-star constraints
30  *
31  *
32  * \min \sum_{ (i,j) \in Q }  w_ij y_ij
33  *
34  *     \sum_c x_nc           =  1           n \in N, c \in C
35  *
36  *     x_nc                  =  0           n \in N, c \not\in C(n)
37  *
38  *     \sum x_nc            <=  1           x_nc \in Clique \in AllCliques,  c \in C
39  *
40  *     \sum_{e \in p} y_e   >=  1           p \in P      path constraints
41  *
42  *     \sum_{e \in cs} y_e  >= |cs| - 1     cs \in CP    clique-star constraints
43  *
44  *     x_nc, y_ij \in N,   w_ij \in R^+
45  */
46 #include "config.h"
47
48 #include "bitset.h"
49 #include "raw_bitset.h"
50 #include "pdeq.h"
51
52 #include "util.h"
53 #include "irgwalk.h"
54 #include "becopyilp_t.h"
55 #include "beifg.h"
56 #include "besched.h"
57 #include "bemodule.h"
58
59 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
60
61 typedef struct local_env_t {
62         int             first_x_var;
63         int             last_x_var;
64         const unsigned *allocatable_colors;
65 } local_env_t;
66
67 static unsigned check_alignment_constraints(ir_node *node)
68 {
69         const arch_register_req_t *req = arch_get_irn_register_req(node);
70         // For larger than 1 variables, support only aligned constraints
71         assert(((!(req->type & arch_register_req_type_aligned)
72                          && req->width == 1)
73                         || (req->type & arch_register_req_type_aligned))
74                    && "Unaligned large (width > 1) variables not supported");
75         return (req->type & arch_register_req_type_aligned) && req->width > 1;
76 }
77
78 static void make_color_var_name(char *buf, size_t buf_size,
79                                 const ir_node *node, unsigned color)
80 {
81         unsigned node_idx = get_irn_idx(node);
82         snprintf(buf, buf_size, "x_%u_%u", node_idx, color);
83 }
84
85 static void build_coloring_cstr(ilp_env_t *ienv)
86 {
87         local_env_t    *lenv   = ienv->env;
88         be_ifg_t       *ifg    = ienv->co->cenv->ifg;
89         unsigned        n_regs = arch_register_class_n_regs(ienv->co->cls);
90         const unsigned *allocatable_colors = lenv->allocatable_colors;
91         nodes_iter_t    iter;
92         unsigned       *colors;
93         ir_node        *irn;
94         char            buf[32];
95
96         rbitset_alloca(colors, n_regs);
97
98         be_ifg_foreach_node(ifg, &iter, irn) {
99                 const arch_register_req_t *req;
100                 unsigned                   col;
101                 int                        cst_idx;
102                 unsigned                   curr_node_color;
103                 unsigned                   has_alignment_cstr;
104
105                 if (sr_is_removed(ienv->sr, irn))
106                         continue;
107
108                 has_alignment_cstr = check_alignment_constraints(irn);
109
110                 req = arch_get_irn_register_req(irn);
111
112                 /* get assignable colors */
113                 if (arch_register_req_is(req, limited)) {
114                         rbitset_copy(colors, req->limited, n_regs);
115                 } else {
116                         rbitset_copy(colors, allocatable_colors, n_regs);
117                 }
118
119                 /* add the coloring constraint */
120                 cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 1.0);
121
122                 curr_node_color = get_irn_col(irn);
123                 for (col = 0; col < n_regs; ++col) {
124                         int    var_idx;
125                         double val;
126                         if (!rbitset_is_set(colors, col)
127                                 || (has_alignment_cstr && ((col % req->width) != 0)))
128                                 continue;
129
130                         make_color_var_name(buf, sizeof(buf), irn, col);
131                         var_idx = lpp_add_var(ienv->lp, buf, lpp_binary, 0.0);
132                         lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
133
134                         val = (col == curr_node_color) ? 1.0 : 0.0;
135                         lpp_set_start_value(ienv->lp, var_idx, val);
136
137                         lenv->last_x_var = var_idx;
138                         if (lenv->first_x_var == -1)
139                                 lenv->first_x_var = var_idx;
140                 }
141
142                 /* add register constraint constraints */
143                 for (col = 0; col < n_regs; ++col) {
144                         int cst_idx;
145                         int var_idx;
146                         if (rbitset_is_set(colors, col)
147                                 // for aligned variable, we set the unaligned part to 0
148                                 && (!has_alignment_cstr || ((col % req->width) == 0)))
149                                 continue;
150
151                         make_color_var_name(buf, sizeof(buf), irn, col);
152                         cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 0.0);
153                         var_idx = lpp_add_var(ienv->lp, buf, lpp_binary, 0.0);
154                         lpp_set_start_value(ienv->lp, var_idx, 0.0);
155                         lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
156
157                         lenv->last_x_var = var_idx;
158                 }
159         }
160 }
161
162 static void build_interference_cstr(ilp_env_t *ienv)
163 {
164         lpp_t          *lpp      = ienv->lp;
165         local_env_t    *lenv     = ienv->env;
166         be_ifg_t       *ifg      = ienv->co->cenv->ifg;
167         unsigned        n_colors = arch_register_class_n_regs(ienv->co->cls);
168         ir_node       **clique   = ALLOCAN(ir_node*, n_colors);
169         const unsigned *allocatable_colors = lenv->allocatable_colors;
170         cliques_iter_t iter;
171         int            size;
172         unsigned       col;
173         int            i;
174
175         /* for each maximal clique */
176         be_ifg_foreach_clique(ifg, &iter, clique, &size) {
177                 unsigned realsize = 0;
178
179                 for (i=0; i<size; ++i) {
180                         if (!sr_is_removed(ienv->sr, clique[i]))
181                                 ++realsize;
182                 }
183
184                 if (realsize < 2)
185                         continue;
186
187                 /* for all colors */
188                 for (col=0; col<n_colors; ++col) {
189                         int cst_idx;
190                         if (!rbitset_is_set(allocatable_colors, col))
191                                 continue;
192
193                         cst_idx = lpp_add_cst(lpp, NULL, lpp_less_equal, 1.0);
194
195                         /* for each member of this clique */
196                         for (i=0; i<size; ++i) {
197                                 ir_node *irn = clique[i];
198                                 char     buf[32];
199                                 int      var_idx;
200                                 unsigned aligment_offset = 0;
201
202                                 if (sr_is_removed(ienv->sr, irn))
203                                         continue;
204
205                                 // Use the first part of the large registers for all
206                                 // interference, since it is the only colorable one.
207                                 if (check_alignment_constraints(irn)) {
208                                         const arch_register_req_t *req
209                                                 = arch_get_irn_register_req(irn);
210                                         aligment_offset = col % req->width;
211                                 }
212
213                                 make_color_var_name(buf, sizeof(buf), irn,
214                                                                         col - aligment_offset);
215                                 var_idx = lpp_get_var_idx(lpp, buf);
216                                 lpp_set_factor_fast(lpp, cst_idx, var_idx, 1);
217                         }
218                 }
219         }
220 }
221
222 static void make_affinity_var_name(char *buf, size_t buf_size,
223                                    const ir_node *node1, const ir_node *node2)
224 {
225         unsigned n1 = get_irn_idx(node1);
226         unsigned n2 = get_irn_idx(node2);
227         if (n1 < n2) {
228                 snprintf(buf, buf_size, "y_%u_%u", n1, n2);
229         } else {
230                 snprintf(buf, buf_size, "y_%u_%u", n2, n1);
231         }
232 }
233
234
235 /**
236  * TODO: Remove the dependency of the opt-units data structure
237  *       by walking over all affinity edges. Graph structure
238  *       does not provide this walker, yet.
239  */
240 static void build_affinity_cstr(ilp_env_t *ienv)
241 {
242         unsigned  n_colors = arch_register_class_n_regs(ienv->co->cls);
243         unit_t   *curr;
244
245         /* for all optimization units */
246         list_for_each_entry(unit_t, curr, &ienv->co->units, units) {
247                 ir_node *root     = curr->nodes[0];
248                 unsigned root_col = get_irn_col(root);
249                 int      i;
250
251                 for (i = 1; i < curr->node_count; ++i) {
252                         ir_node *arg     = curr->nodes[i];
253                         unsigned arg_col = get_irn_col(arg);
254                         double   val;
255                         char     buf[32];
256                         unsigned col;
257                         int      y_idx;
258
259                         /* add a new affinity variable */
260                         make_affinity_var_name(buf, sizeof(buf), arg, root);
261                         y_idx = lpp_add_var(ienv->lp, buf, lpp_binary, curr->costs[i]);
262                         val   = (root_col == arg_col) ? 0.0 : 1.0;
263                         lpp_set_start_value(ienv->lp, y_idx, val);
264
265                         /* add constraints relating the affinity var to the color vars */
266                         for (col=0; col<n_colors; ++col) {
267                                 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_less_equal, 0.0);
268                                 int root_idx;
269                                 int arg_idx;
270
271                                 make_color_var_name(buf, sizeof(buf), root, col);
272                                 root_idx = lpp_get_var_idx(ienv->lp, buf);
273                                 make_color_var_name(buf, sizeof(buf), arg, col);
274                                 arg_idx  = lpp_get_var_idx(ienv->lp, buf);
275
276                                 lpp_set_factor_fast(ienv->lp, cst_idx, root_idx,  1.0);
277                                 lpp_set_factor_fast(ienv->lp, cst_idx, arg_idx,  -1.0);
278                                 lpp_set_factor_fast(ienv->lp, cst_idx, y_idx, -1.0);
279                         }
280                 }
281         }
282 }
283
284 /**
285  * Helping stuff for build_clique_star_cstr
286  */
287 typedef struct edge_t {
288         ir_node *n1;
289         ir_node *n2;
290 } edge_t;
291
292 static int compare_edge_t(const void *k1, const void *k2, size_t size)
293 {
294         const edge_t *e1 = k1;
295         const edge_t *e2 = k2;
296         (void) size;
297
298         return ! (e1->n1 == e2->n1 && e1->n2 == e2->n2);
299 }
300
301 #define HASH_EDGE(e) (hash_irn((e)->n1) ^ hash_irn((e)->n2))
302
303 static inline edge_t *add_edge(set *edges, ir_node *n1, ir_node *n2, size_t *counter)
304 {
305         edge_t new_edge;
306
307         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
308                 new_edge.n1 = n1;
309                 new_edge.n2 = n2;
310         } else {
311                 new_edge.n1 = n2;
312                 new_edge.n2 = n1;
313         }
314         (*counter)++;
315         return set_insert(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
316 }
317
318 static inline edge_t *find_edge(set *edges, ir_node *n1, ir_node *n2)
319 {
320         edge_t new_edge;
321
322         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
323                 new_edge.n1 = n1;
324                 new_edge.n2 = n2;
325         } else {
326                 new_edge.n1 = n2;
327                 new_edge.n2 = n1;
328         }
329         return set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
330 }
331
332 static inline void remove_edge(set *edges, ir_node *n1, ir_node *n2, size_t *counter)
333 {
334         edge_t new_edge, *e;
335
336         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
337                 new_edge.n1 = n1;
338                 new_edge.n2 = n2;
339         } else {
340                 new_edge.n1 = n2;
341                 new_edge.n2 = n1;
342         }
343         e = set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
344         if (e) {
345                 e->n1 = NULL;
346                 e->n2 = NULL;
347                 (*counter)--;
348         }
349 }
350
351 #define pset_foreach(pset, irn)  for (irn=pset_first(pset); irn; irn=pset_next(pset))
352
353 /**
354  * Search for an interference clique and an external node
355  * with affinity edges to all nodes of the clique.
356  * At most 1 node of the clique can be colored equally with the external node.
357  */
358 static void build_clique_star_cstr(ilp_env_t *ienv)
359 {
360         affinity_node_t *aff;
361
362         /* for each node with affinity edges */
363         co_gs_foreach_aff_node(ienv->co, aff) {
364                 struct obstack ob;
365                 neighb_t *nbr;
366                 const ir_node *center = aff->irn;
367                 ir_node **nodes;
368                 set *edges;
369                 int i, o, n_nodes;
370                 size_t n_edges;
371
372                 if (arch_irn_is_ignore(aff->irn))
373                         continue;
374
375                 obstack_init(&ob);
376                 edges = new_set(compare_edge_t, 8);
377
378                 /* get all affinity neighbours */
379                 n_nodes = 0;
380                 co_gs_foreach_neighb(aff, nbr) {
381                         if (!arch_irn_is_ignore(nbr->irn)) {
382                                 obstack_ptr_grow(&ob, nbr->irn);
383                                 ++n_nodes;
384                         }
385                 }
386                 nodes = obstack_finish(&ob);
387
388                 /* get all interference edges between these */
389                 n_edges = 0;
390                 for (i=0; i<n_nodes; ++i) {
391                         for (o=0; o<i; ++o) {
392                                 if (be_ifg_connected(ienv->co->cenv->ifg, nodes[i], nodes[o]))
393                                         add_edge(edges, nodes[i], nodes[o], &n_edges);
394                         }
395                 }
396
397                 /* cover all these interference edges with maximal cliques */
398                 while (n_edges) {
399                         edge_t *e;
400                         pset   *clique = pset_new_ptr(8);
401                         bool    growed;
402
403                         /* get 2 starting nodes to form a clique */
404                         for (e=set_first(edges); !e->n1; e=set_next(edges)) {
405                         }
406
407                         /* we could be stepped out of the loop before the set iterated to the end */
408                         set_break(edges);
409
410                         pset_insert_ptr(clique, e->n1);
411                         pset_insert_ptr(clique, e->n2);
412                         remove_edge(edges, e->n1, e->n2, &n_edges);
413
414                         /* while the clique is growing */
415                         do {
416                                 growed = false;
417
418                                 /* search for a candidate to extend the clique */
419                                 for (i=0; i<n_nodes; ++i) {
420                                         ir_node *cand = nodes[i];
421                                         ir_node *member;
422                                         bool     is_cand;
423
424                                         /* if its already in the clique try the next */
425                                         if (pset_find_ptr(clique, cand))
426                                                 continue;
427
428                                         /* are there all necessary interferences? */
429                                         is_cand = true;
430                                         pset_foreach(clique, member) {
431                                                 if (!find_edge(edges, cand, member)) {
432                                                         is_cand = false;
433                                                         pset_break(clique);
434                                                         break;
435                                                 }
436                                         }
437
438                                         /* now we know if we have a clique extender */
439                                         if (is_cand) {
440                                                 /* first remove all covered edges */
441                                                 pset_foreach(clique, member)
442                                                         remove_edge(edges, cand, member, &n_edges);
443
444                                                 /* insert into clique */
445                                                 pset_insert_ptr(clique, cand);
446                                                 growed = true;
447                                                 break;
448                                         }
449                                 }
450                         } while (growed);
451
452                         /* now the clique is maximal. Finally add the constraint */
453                         {
454                                 ir_node *member;
455                                 int      var_idx;
456                                 int      cst_idx;
457                                 char     buf[32];
458
459                                 cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater_equal, pset_count(clique)-1);
460
461                                 pset_foreach(clique, member) {
462                                         make_affinity_var_name(buf, sizeof(buf), center, member);
463                                         var_idx = lpp_get_var_idx(ienv->lp, buf);
464                                         lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
465                                 }
466                         }
467
468                         del_pset(clique);
469                 }
470
471                 del_set(edges);
472                 obstack_free(&ob, NULL);
473         }
474 }
475
476 static void extend_path(ilp_env_t *ienv, pdeq *path, const ir_node *irn)
477 {
478         be_ifg_t *ifg = ienv->co->cenv->ifg;
479         int i, len;
480         ir_node **curr_path;
481         affinity_node_t *aff;
482         neighb_t *nbr;
483
484         /* do not walk backwards or in circles */
485         if (pdeq_contains(path, irn))
486                 return;
487
488         if (arch_irn_is_ignore(irn))
489                 return;
490
491         /* insert the new irn */
492         pdeq_putr(path, irn);
493
494         /* check for forbidden interferences */
495         len       = pdeq_len(path);
496         curr_path = ALLOCAN(ir_node*, len);
497         pdeq_copyl(path, (const void **)curr_path);
498
499         for (i=1; i<len; ++i) {
500                 if (be_ifg_connected(ifg, irn, curr_path[i]))
501                         goto end;
502         }
503
504         /* check for terminating interference */
505         if (be_ifg_connected(ifg, irn, curr_path[0])) {
506                 /* One node is not a path. */
507                 /* And a path of length 2 is covered by a clique star constraint. */
508                 if (len > 2) {
509                         /* finally build the constraint */
510                         int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater_equal, 1.0);
511                         for (i=1; i<len; ++i) {
512                                 char buf[32];
513                                 int  var_idx;
514
515                                 make_affinity_var_name(buf, sizeof(buf), curr_path[i-1], curr_path[i]);
516                                 var_idx = lpp_get_var_idx(ienv->lp, buf);
517                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
518                         }
519                 }
520
521                 /* this path cannot be extended anymore */
522                 goto end;
523         }
524
525         /* recursively extend the path */
526         aff = get_affinity_info(ienv->co, irn);
527         co_gs_foreach_neighb(aff, nbr) {
528                 extend_path(ienv, path, nbr->irn);
529         }
530
531 end:
532         /* remove the irn */
533         pdeq_getr(path);
534 }
535
536 /**
537  * Search a path of affinity edges, whose ends are connected
538  * by an interference edge and there are no other interference
539  * edges in between.
540  * Then at least one of these affinity edges must break.
541  */
542 static void build_path_cstr(ilp_env_t *ienv)
543 {
544         affinity_node_t *aff_info;
545
546         /* for each node with affinity edges */
547         co_gs_foreach_aff_node(ienv->co, aff_info) {
548                 pdeq *path = new_pdeq();
549
550                 extend_path(ienv, path, aff_info->irn);
551
552                 del_pdeq(path);
553         }
554 }
555
556 static void ilp2_build(ilp_env_t *ienv)
557 {
558         int lower_bound;
559
560         ienv->lp = lpp_new(ienv->co->name, lpp_minimize);
561         build_coloring_cstr(ienv);
562         build_interference_cstr(ienv);
563         build_affinity_cstr(ienv);
564         build_clique_star_cstr(ienv);
565         build_path_cstr(ienv);
566
567         lower_bound = co_get_lower_bound(ienv->co) - co_get_inevit_copy_costs(ienv->co);
568         lpp_set_bound(ienv->lp, lower_bound);
569 }
570
571 static void ilp2_apply(ilp_env_t *ienv)
572 {
573         local_env_t *lenv = ienv->env;
574         ir_graph    *irg  = ienv->co->irg;
575
576         /* first check if there was sth. to optimize */
577         if (lenv->first_x_var >= 0) {
578                 int              count = lenv->last_x_var - lenv->first_x_var + 1;
579                 double          *sol   = XMALLOCN(double, count);
580                 lpp_sol_state_t  state = lpp_get_solution(ienv->lp, sol, lenv->first_x_var, lenv->last_x_var);
581                 int              i;
582
583                 if (state != lpp_optimal) {
584                         printf("WARNING %s: Solution state is not 'optimal': %d\n",
585                                ienv->co->name, (int)state);
586                         if (state < lpp_feasible) {
587                                 panic("Copy coalescing solution not feasible!");
588                         }
589                 }
590
591                 for (i=0; i<count; ++i) {
592                         unsigned nodenr;
593                         unsigned color;
594                         char     var_name[32];
595                         if (sol[i] <= 1-EPSILON)
596                                 continue;
597                         /* split variable name into components */
598                         lpp_get_var_name(ienv->lp, lenv->first_x_var+i, var_name, sizeof(var_name));
599
600                         if (sscanf(var_name, "x_%u_%u", &nodenr, &color) == 2) {
601                                 ir_node *irn = get_idx_irn(irg, nodenr);
602                                 set_irn_col(ienv->co->cls, irn, color);
603                         } else {
604                                 panic("This should be a x-var");
605                         }
606                 }
607
608                 xfree(sol);
609         }
610
611 #ifdef COPYOPT_STAT
612         /* TODO adapt to multiple possible ILPs */
613         copystat_add_ilp_time((int)(1000.0*lpp_get_sol_time(pi->curr_lp)));  //now we have ms
614         copystat_add_ilp_vars(lpp_get_var_count(pi->curr_lp));
615         copystat_add_ilp_csts(lpp_get_cst_count(pi->curr_lp));
616         copystat_add_ilp_iter(lpp_get_iter_cnt(pi->curr_lp));
617 #endif
618 }
619
620 /**
621  * Solves the problem using mixed integer programming
622  * @returns 1 iff solution state was optimal
623  * Uses the OU and the GRAPH data structure
624  * Dependency of the OU structure can be removed
625  */
626 static int co_solve_ilp2(copy_opt_t *co)
627 {
628         unsigned       *allocatable_colors;
629         unsigned        n_regs = arch_register_class_n_regs(co->cls);
630         lpp_sol_state_t sol_state;
631         ilp_env_t      *ienv;
632         local_env_t     my;
633
634         ASSERT_OU_AVAIL(co); //See build_clique_st
635         ASSERT_GS_AVAIL(co);
636
637         my.first_x_var = -1;
638         my.last_x_var  = -1;
639         FIRM_DBG_REGISTER(dbg, "firm.be.coilp2");
640
641         rbitset_alloca(allocatable_colors, n_regs);
642         be_set_allocatable_regs(co->irg, co->cls, allocatable_colors);
643         my.allocatable_colors = allocatable_colors;
644
645         ienv = new_ilp_env(co, ilp2_build, ilp2_apply, &my);
646
647         sol_state = ilp_go(ienv);
648
649         free_ilp_env(ienv);
650
651         return sol_state == lpp_optimal;
652 }
653
654 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyilp2)
655 void be_init_copyilp2(void)
656 {
657         static co_algo_info copyheur = {
658                 co_solve_ilp2, 1
659         };
660
661         be_register_copyopt("ilp", &copyheur);
662 }