nsz Git - libfirm/blob - ir/be/becopyilp2.c

   1 /*
   2  * Copyright (C) 1995-2007 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  * @version     $Id$
  26  *
  27  * ILP formalization using G=(V, E, Q):
  28  *  - 2 class of variables: coloring vars x_ic   and   equal color vars y_ij
  29  *  - Path constraints
  30  *  - Clique-star constraints
  31  *
  32  *
  33  *      \min \sum_{ (i,j) \in Q }  w_ij y_ij
  34  *
  35  *              \sum_c x_nc                     =  1                    n \in N, c \in C
  36  *
  37  *              x_nc                            =  0                    n \in N, c \not\in C(n)
  38  *
  39  *              \sum x_nc                       <= 1                    x_nc \in Clique \in AllCliques,  c \in C
  40  *
  41  *              \sum_{e \in p} y_e      >= 1                    p \in P         path constraints
  42  *
  43  *              \sum_{e \in cs} y_e     >= |cs| - 1             cs \in CP       clique-star constraints
  44  *
  45  *              x_nc, y_ij \in N,   w_ij \in R^+
  46  */
  47 #ifdef HAVE_CONFIG_H
  48 #include "config.h"
  49 #endif /* HAVE_CONFIG_H */
  50
  51 #include "firm_config.h"
  52
  53 #ifdef WITH_ILP
  54
  55 #include "bitset.h"
  56 #include "raw_bitset.h"
  57 #include "pdeq.h"
  58
  59 #include "irtools.h"
  60 #include "irgwalk.h"
  61 #include "becopyilp_t.h"
  62 #include "beifg_t.h"
  63 #include "besched_t.h"
  64 #include "benodesets.h"
  65
  66 #define DEBUG_LVL 1
  67
  68 typedef struct _local_env_t {
  69         double time_limit;
  70         int first_x_var, last_x_var;
  71         int n_colors;
  72         bitset_t *normal_colors;
  73         pmap *nr_2_irn;
  74         DEBUG_ONLY(firm_dbg_module_t *dbg;)
  75 } local_env_t;
  76
  77 static void build_coloring_cstr(ilp_env_t *ienv) {
  78         local_env_t *lenv = ienv->env;
  79         be_ifg_t *ifg     = ienv->co->cenv->ifg;
  80         void *iter        = be_ifg_nodes_iter_alloca(ifg);
  81         bitset_t *colors;
  82         ir_node *irn;
  83         char buf[16];
  84
  85         colors = bitset_alloca(arch_register_class_n_regs(ienv->co->cls));
  86
  87         be_ifg_foreach_node(ifg, iter, irn)
  88                 if (!sr_is_removed(ienv->sr, irn)) {
  89                         bitset_pos_t col;
  90                         int cst_idx;
  91                         const arch_register_req_t *req;
  92                         int curr_node_color = get_irn_col(ienv->co, irn);
  93                         int node_nr = (int)get_irn_node_nr(irn);
  94                         local_env_t *lenv = ienv->env;
  95
  96                         pmap_insert(lenv->nr_2_irn, INT_TO_PTR(node_nr), irn);
  97
  98                         req = arch_get_register_req(ienv->co->aenv, irn, -1);
  99
 100                         bitset_clear_all(colors);
 101
 102                         /* get assignable colors */
 103                         if (arch_register_req_is(req, limited)) {
 104                                 rbitset_copy_to_bitset(req->limited, colors);
 105                         } else {
 106                                 bitset_copy(colors, lenv->normal_colors);
 107                         }
 108
 109                         /* add the coloring constraint */
 110                         cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 1.0);
 111
 112                         bitset_foreach(colors, col) {
 113                                 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, col), lpp_binary, 0.0);
 114                                 lpp_set_start_value(ienv->lp, var_idx, (col == (unsigned) curr_node_color) ? 1.0 : 0.0);
 115                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
 116
 117                                 lenv->last_x_var = var_idx;
 118                                 if (lenv->first_x_var == -1)
 119                                         lenv->first_x_var = var_idx;
 120                         }
 121
 122                         /* add register constraint constraints */
 123                         bitset_foreach_clear(colors, col) {
 124                                 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 0.0);
 125                                 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, col), lpp_binary, 0.0);
 126                                 lpp_set_start_value(ienv->lp, var_idx, 0.0);
 127                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
 128
 129                                 lenv->last_x_var = var_idx;
 130                         }
 131                 }
 132 }
 133
 134 static void build_interference_cstr(ilp_env_t *ienv) {
 135         lpp_t *lpp        = ienv->lp;
 136         local_env_t *lenv = ienv->env;
 137         be_ifg_t *ifg     = ienv->co->cenv->ifg;
 138         int n_colors      = lenv->n_colors;
 139         int i, col;
 140
 141         void *iter = be_ifg_cliques_iter_alloca(ifg);
 142         ir_node **clique = alloca(sizeof(*clique) * n_colors);
 143         int size;
 144
 145         char buf[16];
 146
 147         /* for each maximal clique */
 148         be_ifg_foreach_clique(ifg, iter, clique, &size) {
 149                 int realsize = 0;
 150
 151                 for (i=0; i<size; ++i)
 152                         if (!sr_is_removed(ienv->sr, clique[i]))
 153                                 ++realsize;
 154
 155                 if (realsize < 2)
 156                         continue;
 157
 158                 /* for all colors */
 159                 for (col=0; col<n_colors; ++col) {
 160                         int cst_idx = lpp_add_cst(lpp, NULL, lpp_less, 1.0);
 161
 162                         /* for each member of this clique */
 163                         for (i=0; i<size; ++i) {
 164                                 ir_node *irn = clique[i];
 165
 166                                 if (!sr_is_removed(ienv->sr, irn)) {
 167                                         int var_idx = lpp_get_var_idx(lpp, name_cdd(buf, 'x', (int)get_irn_node_nr(irn), col));
 168                                         lpp_set_factor_fast(lpp, cst_idx, var_idx, 1);
 169                                 }
 170                         }
 171                 }
 172         }
 173 }
 174
 175
 176 /**
 177  * TODO: Remove the dependency of the opt-units data structure
 178  *       by walking over all affinity edges. Graph structure
 179  *       does not provide this walker, yet.
 180  */
 181 static void build_affinity_cstr(ilp_env_t *ienv) {
 182         local_env_t *lenv = ienv->env;
 183         int n_colors      = lenv->n_colors;
 184         unit_t *curr;
 185
 186         /* for all optimization units */
 187         list_for_each_entry(unit_t, curr, &ienv->co->units, units) {
 188                 ir_node *root, *arg;
 189                 int root_nr, arg_nr, i, col, y_idx, root_idx, arg_idx;
 190                 char buf[16];
 191                 int root_col, arg_col;
 192
 193                 root = curr->nodes[0];
 194                 root_nr = (int) get_irn_node_nr(root);
 195                 root_col = get_irn_col(ienv->co, root);
 196
 197                 for (i = 1; i < curr->node_count; ++i) {
 198                         arg = curr->nodes[i];
 199                         arg_nr = (int) get_irn_node_nr(arg);
 200                         arg_col = get_irn_col(ienv->co, arg);
 201
 202                         /* add a new affinity variable */
 203                         y_idx = lpp_add_var(ienv->lp, name_cdd_sorted(buf, 'y', root_nr, arg_nr), lpp_binary, curr->costs[i]);
 204                         lpp_set_start_value(ienv->lp, y_idx, (root_col==arg_col) ? 0.0 : 1.0);
 205
 206                         /* add constraints relating the affinity var to the color vars */
 207                         for (col=0; col<n_colors; ++col) {
 208                                 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_less, 0.0);
 209                                 root_idx = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', root_nr, col));
 210                                 arg_idx  = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', arg_nr,  col));
 211
 212                                 lpp_set_factor_fast(ienv->lp, cst_idx, root_idx,  1.0);
 213                                 lpp_set_factor_fast(ienv->lp, cst_idx, arg_idx,  -1.0);
 214                                 lpp_set_factor_fast(ienv->lp, cst_idx, y_idx, -1.0);
 215                         }
 216                 }
 217         }
 218 }
 219
 220 /**
 221  * Helping stuff for build_clique_star_cstr
 222  */
 223 typedef struct _edge_t {
 224         ir_node *n1, *n2;
 225 } edge_t;
 226
 227 static int compare_edge_t(const void *k1, const void *k2, size_t size) {
 228         const edge_t *e1 = k1;
 229         const edge_t *e2 = k2;
 230         (void) size;
 231
 232         return ! (e1->n1 == e2->n1   &&   e1->n2 == e2->n2);
 233 }
 234
 235 #define HASH_EDGE(e) (nodeset_hash((e)->n1) ^ nodeset_hash((e)->n2))
 236
 237 static INLINE edge_t *add_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
 238         edge_t new_edge;
 239
 240         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 241                 new_edge.n1 = n1;
 242                 new_edge.n2 = n2;
 243         } else {
 244                 new_edge.n1 = n2;
 245                 new_edge.n2 = n1;
 246         }
 247         (*counter)++;
 248         return set_insert(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 249 }
 250
 251 static INLINE edge_t *find_edge(set *edges, ir_node *n1, ir_node *n2) {
 252         edge_t new_edge;
 253
 254         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 255                 new_edge.n1 = n1;
 256                 new_edge.n2 = n2;
 257         } else {
 258                 new_edge.n1 = n2;
 259                 new_edge.n2 = n1;
 260         }
 261         return set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 262 }
 263
 264 static INLINE void remove_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
 265         edge_t new_edge, *e;
 266
 267         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 268                 new_edge.n1 = n1;
 269                 new_edge.n2 = n2;
 270         } else {
 271                 new_edge.n1 = n2;
 272                 new_edge.n2 = n1;
 273         }
 274         e = set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 275         if (e) {
 276                 e->n1 = NULL;
 277                 e->n2 = NULL;
 278                 (*counter)--;
 279         }
 280 }
 281
 282 #define pset_foreach(pset, irn)  for(irn=pset_first(pset); irn; irn=pset_next(pset))
 283
 284 /**
 285  * Search for an interference clique and an external node
 286  * with affinity edges to all nodes of the clique.
 287  * At most 1 node of the clique can be colored equally with the external node.
 288  */
 289 static void build_clique_star_cstr(ilp_env_t *ienv) {
 290         affinity_node_t *aff;
 291
 292         /* for each node with affinity edges */
 293         co_gs_foreach_aff_node(ienv->co, aff) {
 294                 struct obstack ob;
 295                 neighb_t *nbr;
 296                 ir_node *center = aff->irn;
 297                 ir_node **nodes;
 298                 set *edges;
 299                 int i, o, n_nodes, n_edges;
 300
 301                 obstack_init(&ob);
 302                 edges = new_set(compare_edge_t, 8);
 303
 304                 /* get all affinity neighbours */
 305                 n_nodes = 0;
 306                 co_gs_foreach_neighb(aff, nbr) {
 307                         obstack_ptr_grow(&ob, nbr->irn);
 308                         ++n_nodes;
 309                 }
 310                 nodes = obstack_finish(&ob);
 311
 312                 /* get all interference edges between these */
 313                 n_edges = 0;
 314                 for (i=0; i<n_nodes; ++i)
 315                         for (o=0; o<i; ++o)
 316                                 if (be_ifg_connected(ienv->co->cenv->ifg, nodes[i], nodes[o]))
 317                                         add_edge(edges, nodes[i], nodes[o], &n_edges);
 318
 319                 /* cover all these interference edges with maximal cliques */
 320                 while (n_edges) {
 321                         edge_t *e;
 322                         pset *clique = pset_new_ptr(8);
 323                         int growed;
 324
 325                         /* get 2 starting nodes to form a clique */
 326                         for (e=set_first(edges); !e->n1; e=set_next(edges))
 327                                 /*nothing*/ ;
 328
 329                         /* we could be stepped out of the loop before the set iterated to the end */
 330                         set_break(edges);
 331
 332                         pset_insert_ptr(clique, e->n1);
 333                         pset_insert_ptr(clique, e->n2);
 334                         remove_edge(edges, e->n1, e->n2, &n_edges);
 335
 336                         /* while the clique is growing */
 337                         do {
 338                                 growed = 0;
 339
 340                                 /* search for a candidate to extend the clique */
 341                                 for (i=0; i<n_nodes; ++i) {
 342                                         ir_node *member, *cand = nodes[i];
 343                                         int is_cand;
 344
 345                                         /* if its already in the clique try the next */
 346                                         if (pset_find_ptr(clique, cand))
 347                                                 continue;
 348
 349                                         /* are there all necessary interferences? */
 350                                         is_cand = 1;
 351                                         pset_foreach(clique, member) {
 352                                                 if (!find_edge(edges, cand, member)) {
 353                                                         is_cand = 0;
 354                                                         pset_break(clique);
 355                                                         break;
 356                                                 }
 357                                         }
 358
 359                                         /* now we know if we have a clique extender */
 360                                         if (is_cand) {
 361                                                 /* first remove all covered edges */
 362                                                 pset_foreach(clique, member)
 363                                                         remove_edge(edges, cand, member, &n_edges);
 364
 365                                                 /* insert into clique */
 366                                                 pset_insert_ptr(clique, cand);
 367                                                 growed = 1;
 368                                                 break;
 369                                         }
 370                                 }
 371                         } while (growed);
 372
 373                         /* now the clique is maximal. Finally add the constraint */
 374                         {
 375                                 ir_node *member;
 376                                 int var_idx, cst_idx, center_nr, member_nr;
 377                                 char buf[16];
 378
 379                                 cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, pset_count(clique)-1);
 380                                 center_nr = get_irn_node_nr(center);
 381
 382                                 pset_foreach(clique, member) {
 383                                         member_nr = get_irn_node_nr(member);
 384                                         var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', center_nr, member_nr));
 385                                         lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 386                                 }
 387                         }
 388
 389                         del_pset(clique);
 390                 }
 391
 392                 del_set(edges);
 393                 obstack_free(&ob, NULL);
 394         }
 395 }
 396
 397
 398 static void extend_path(ilp_env_t *ienv, pdeq *path, ir_node *irn) {
 399         be_ifg_t *ifg = ienv->co->cenv->ifg;
 400         int i, len;
 401         ir_node **curr_path;
 402         affinity_node_t *aff;
 403         neighb_t *nbr;
 404
 405         /* do not walk backwards or in circles */
 406         if (pdeq_contains(path, irn))
 407                 return;
 408
 409         /* insert the new irn */
 410         pdeq_putr(path, irn);
 411
 412
 413
 414         /* check for forbidden interferences */
 415         len = pdeq_len(path);
 416         curr_path = alloca(len * sizeof(*curr_path));
 417         pdeq_copyl(path, (const void **)curr_path);
 418
 419         for (i=1; i<len; ++i)
 420                 if (be_ifg_connected(ifg, irn, curr_path[i]))
 421                         goto end;
 422
 423
 424
 425         /* check for terminating interference */
 426         if (be_ifg_connected(ifg, irn, curr_path[0])) {
 427
 428                 /* One node is not a path. */
 429                 /* And a path of length 2 is covered by a clique star constraint. */
 430                 if (len > 2) {
 431                         /* finally build the constraint */
 432                         int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, 1.0);
 433                         for (i=1; i<len; ++i) {
 434                                 char buf[16];
 435                                 int nr_1    = get_irn_node_nr(curr_path[i-1]);
 436                                 int nr_2    = get_irn_node_nr(curr_path[i]);
 437                                 int var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', nr_1, nr_2));
 438                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 439                         }
 440                 }
 441
 442                 /* this path cannot be extended anymore */
 443                 goto end;
 444         }
 445
 446
 447
 448         /* recursively extend the path */
 449         aff = get_affinity_info(ienv->co, irn);
 450         co_gs_foreach_neighb(aff, nbr)
 451                 extend_path(ienv, path, nbr->irn);
 452
 453
 454 end:
 455         /* remove the irn */
 456         pdeq_getr(path);
 457
 458 }
 459
 460 /**
 461  *  Search a path of affinity edges, whose ends are connected
 462  *  by an interference edge and there are no other interference
 463  *  edges in between.
 464  *  Then at least one of these affinity edges must break.
 465  */
 466 static void build_path_cstr(ilp_env_t *ienv) {
 467         affinity_node_t *aff_info;
 468
 469         /* for each node with affinity edges */
 470         co_gs_foreach_aff_node(ienv->co, aff_info) {
 471                 pdeq *path = new_pdeq();
 472
 473                 extend_path(ienv, path, aff_info->irn);
 474
 475                 del_pdeq(path);
 476         }
 477 }
 478
 479 static void ilp2_build(ilp_env_t *ienv) {
 480         local_env_t *lenv = ienv->env;
 481         int lower_bound;
 482
 483         ienv->lp = new_lpp(ienv->co->name, lpp_minimize);
 484         build_coloring_cstr(ienv);
 485         build_interference_cstr(ienv);
 486         build_affinity_cstr(ienv);
 487         build_clique_star_cstr(ienv);
 488         build_path_cstr(ienv);
 489
 490         lower_bound = co_get_lower_bound(ienv->co) - co_get_inevit_copy_costs(ienv->co);
 491         lpp_set_bound(ienv->lp, lower_bound);
 492         lpp_set_time_limit(ienv->lp, lenv->time_limit);
 493 }
 494
 495 static void ilp2_apply(ilp_env_t *ienv) {
 496         local_env_t *lenv = ienv->env;
 497         double *sol;
 498         lpp_sol_state_t state;
 499         int i, count;
 500
 501         /* first check if there was sth. to optimize */
 502         if (lenv->first_x_var >= 0) {
 503
 504                 count = lenv->last_x_var - lenv->first_x_var + 1;
 505                 sol = xmalloc(count * sizeof(sol[0]));
 506                 state = lpp_get_solution(ienv->lp, sol, lenv->first_x_var, lenv->last_x_var);
 507                 if (state != lpp_optimal) {
 508                         printf("WARNING %s: Solution state is not 'optimal': %d\n", ienv->co->name, state);
 509                         assert(state >= lpp_feasible && "The solution should at least be feasible!");
 510                 }
 511
 512                 for (i=0; i<count; ++i) {
 513                         int nodenr, color;
 514                         char var_name[16];
 515
 516                         if (sol[i] > 1-EPSILON) { /* split variable name into components */
 517                                 lpp_get_var_name(ienv->lp, lenv->first_x_var+i, var_name, sizeof(var_name));
 518
 519                                 if (sscanf(var_name, "x_%d_%d", &nodenr, &color) == 2) {
 520                                         ir_node *irn = pmap_get(lenv->nr_2_irn, INT_TO_PTR(nodenr));
 521                                         assert(irn && "This node number must be present in the map");
 522
 523                                         set_irn_col(ienv->co, irn, color);
 524                                 } else
 525                                         assert(0 && "This should be a x-var");
 526                         }
 527                 }
 528         }
 529
 530 #ifdef COPYOPT_STAT
 531         /* TODO adapt to multiple possible ILPs */
 532         copystat_add_ilp_time((int)(1000.0*lpp_get_sol_time(pi->curr_lp)));  //now we have ms
 533         copystat_add_ilp_vars(lpp_get_var_count(pi->curr_lp));
 534         copystat_add_ilp_csts(lpp_get_cst_count(pi->curr_lp));
 535         copystat_add_ilp_iter(lpp_get_iter_cnt(pi->curr_lp));
 536 #endif
 537 }
 538
 539 int co_solve_ilp2(copy_opt_t *co) {
 540         lpp_sol_state_t sol_state;
 541         ilp_env_t *ienv;
 542         local_env_t my;
 543
 544         ASSERT_OU_AVAIL(co); //See build_clique_st
 545         ASSERT_GS_AVAIL(co);
 546
 547         my.time_limit  = 0;
 548         my.first_x_var = -1;
 549         my.last_x_var  = -1;
 550         my.nr_2_irn    = pmap_create();
 551         FIRM_DBG_REGISTER(my.dbg, "firm.be.coilp2");
 552
 553         my.normal_colors = bitset_alloca(arch_register_class_n_regs(co->cls));
 554         bitset_clear_all(my.normal_colors);
 555         arch_put_non_ignore_regs(co->aenv, co->cls, my.normal_colors);
 556         my.n_colors = bitset_popcnt(my.normal_colors);
 557
 558         ienv = new_ilp_env(co, ilp2_build, ilp2_apply, &my);
 559
 560         sol_state = ilp_go(ienv);
 561
 562         pmap_destroy(my.nr_2_irn);
 563         free_ilp_env(ienv);
 564
 565         return sol_state == lpp_optimal;
 566 }
 567
 568 #else /* WITH_ILP */
 569
 570 static INLINE void only_that_you_can_compile_without_WITH_ILP_defined(void) {
 571 }
 572
 573 #endif /* WITH_ILP */