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