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