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