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

   1 /**
   2  * Author:      Daniel Grund
   3  * Date:                28.02.2006
   4  * Copyright:   (c) Universitaet Karlsruhe
   5  * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
   6  *
   7  *
   8  * ILP formalization using G=(V, E, Q):
   9  *  - 2 class of variables: coloring vars x_ic   and   equal color vars y_ij
  10  *  - Path constraints
  11  *  - Clique-star constraints
  12  *
  13  *
  14  *      \min \sum_{ (i,j) \in Q }  w_ij y_ij
  15  *
  16  *              \sum_c x_nc                     =  1                    n \in N, c \in C
  17  *
  18  *              x_nc                            =  0                    n \in N, c \not\in C(n)
  19  *
  20  *              \sum x_nc                       <= 1                    x_nc \in Clique \in AllCliques,  c \in C
  21  *
  22  *              \sum_{e \in p} y_e      >= 1                    p \in P         path constraints
  23  *
  24  *              \sum_{e \in cs} y_e     >= |cs| - 1             cs \in CP       clique-star constraints
  25  *
  26  *              x_nc, y_ij \in N,   w_ij \in R^+
  27  */
  28
  29 #ifdef HAVE_CONFIG_H
  30 #include "config.h"
  31 #endif /* HAVE_CONFIG_H */
  32
  33 #ifdef WITH_ILP
  34
  35 #include <bitset.h>
  36 #include "pdeq.h"
  37
  38 #include "irtools.h"
  39 #include "irgwalk.h"
  40 #include "becopyilp_t.h"
  41 #include "beifg_t.h"
  42 #include "besched_t.h"
  43 #include "benodesets.h"
  44
  45 #define DEBUG_LVL 1
  46
  47 typedef struct _local_env_t {
  48         double time_limit;
  49         int first_x_var, last_x_var;
  50         pmap *nr_2_irn;
  51         DEBUG_ONLY(firm_dbg_module_t *dbg;)
  52 } local_env_t;
  53
  54 static void build_coloring_cstr(ilp_env_t *ienv) {
  55         be_ifg_t *ifg = ienv->co->cenv->ifg;
  56         void *iter = be_ifg_nodes_iter_alloca(ifg);
  57         bitset_t *colors;
  58         ir_node *irn;
  59         char buf[16];
  60
  61         colors = bitset_alloca(arch_register_class_n_regs(ienv->co->cls));
  62
  63         be_ifg_foreach_node(ifg, iter, irn)
  64                 if (!sr_is_removed(ienv->sr, irn)) {
  65                         int col, cst_idx;
  66                         arch_register_req_t req;
  67                         int curr_node_color = get_irn_col(ienv->co, irn);
  68                         int node_nr = (int)get_irn_node_nr(irn);
  69                         local_env_t *lenv = ienv->env;
  70
  71                         pmap_insert(lenv->nr_2_irn, INT_TO_PTR(node_nr), irn);
  72
  73                         arch_get_register_req(ienv->co->aenv, &req, irn, -1);
  74
  75                         /* get assignable colors */
  76                         if (arch_register_req_is(&req, limited))
  77                                 req.limited(req.limited_env, colors);
  78                         else {
  79                                 arch_register_class_put(req.cls, colors);
  80                                 // bitset_andnot(colors, ienv->co->cenv->ignore_colors);
  81                         }
  82
  83                         /* add the coloring constraint */
  84                         cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 1.0);
  85
  86                         bitset_foreach(colors, col) {
  87                                 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, col), lpp_binary, 0.0);
  88                                 lpp_set_start_value(ienv->lp, var_idx, (col == curr_node_color) ? 1.0 : 0.0);
  89                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
  90
  91                                 lenv->last_x_var = var_idx;
  92                                 if (lenv->first_x_var == -1)
  93                                         lenv->first_x_var = var_idx;
  94                         }
  95
  96                         /* add register constraint constraints */
  97                         bitset_foreach_clear(colors, col) {
  98                                 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 0.0);
  99                                 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, col), lpp_binary, 0.0);
 100                                 lpp_set_start_value(ienv->lp, var_idx, 0.0);
 101                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
 102
 103                                 lenv->last_x_var = var_idx;
 104                         }
 105                 }
 106 }
 107
 108 static void build_interference_cstr(ilp_env_t *ienv) {
 109         lpp_t *lpp = ienv->lp;
 110         be_ifg_t *ifg = ienv->co->cenv->ifg;
 111         int n_colors = arch_register_class_n_regs(ienv->co->cls);
 112         int i, col;
 113
 114         void *iter = be_ifg_cliques_iter_alloca(ifg);
 115         ir_node **clique = alloca(sizeof(*clique) * n_colors);
 116         int size;
 117
 118         char buf[16];
 119
 120         /* for each maximal clique */
 121         be_ifg_foreach_clique(ifg, iter, clique, &size) {
 122                 int realsize = 0;
 123
 124                 for (i=0; i<size; ++i)
 125                         if (!sr_is_removed(ienv->sr, clique[i]))
 126                                 ++realsize;
 127
 128                 if (realsize < 2)
 129                         continue;
 130
 131                 /* for all colors */
 132                 for (col=0; col<n_colors; ++col) {
 133                         int cst_idx = lpp_add_cst(lpp, NULL, lpp_less, 1.0);
 134
 135                         /* for each member of this clique */
 136                         for (i=0; i<size; ++i) {
 137                                 ir_node *irn = clique[i];
 138
 139                                 if (!sr_is_removed(ienv->sr, irn)) {
 140                                         int var_idx = lpp_get_var_idx(lpp, name_cdd(buf, 'x', (int)get_irn_node_nr(irn), col));
 141                                         lpp_set_factor_fast(lpp, cst_idx, var_idx, 1);
 142                                 }
 143                         }
 144                 }
 145         }
 146 }
 147
 148
 149 /**
 150  * TODO: Remove the dependency of the opt-units data structure
 151  *       by walking over all affinity edges. Graph structure
 152  *       does not provide this walker, yet.
 153  */
 154 static void build_affinity_cstr(ilp_env_t *ienv) {
 155         unit_t *curr;
 156         int n_colors = arch_register_class_n_regs(ienv->co->cls);
 157
 158         /* for all optimization units */
 159         list_for_each_entry(unit_t, curr, &ienv->co->units, units) {
 160                 ir_node *root, *arg;
 161                 int root_nr, arg_nr, i, col, y_idx, root_idx, arg_idx;
 162                 char buf[16];
 163                 int root_col, arg_col;
 164
 165                 root = curr->nodes[0];
 166                 root_nr = (int) get_irn_node_nr(root);
 167                 root_col = get_irn_col(ienv->co, root);
 168
 169                 for (i = 1; i < curr->node_count; ++i) {
 170                         arg = curr->nodes[i];
 171                         arg_nr = (int) get_irn_node_nr(arg);
 172                         arg_col = get_irn_col(ienv->co, arg);
 173
 174                         /* add a new affinity variable */
 175                         y_idx = lpp_add_var(ienv->lp, name_cdd_sorted(buf, 'y', root_nr, arg_nr), lpp_binary, curr->costs[i]);
 176                         lpp_set_start_value(ienv->lp, y_idx, (root_col==arg_col) ? 0.0 : 1.0);
 177
 178                         /* add constraints relating the affinity var to the color vars */
 179                         for (col=0; col<n_colors; ++col) {
 180                                 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_less, 0.0);
 181                                 root_idx = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', root_nr, col));
 182                                 arg_idx  = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', arg_nr,  col));
 183
 184                                 lpp_set_factor_fast(ienv->lp, cst_idx, root_idx,  1.0);
 185                                 lpp_set_factor_fast(ienv->lp, cst_idx, arg_idx,  -1.0);
 186                                 lpp_set_factor_fast(ienv->lp, cst_idx, y_idx, -1.0);
 187                         }
 188                 }
 189         }
 190 }
 191
 192 /**
 193  * Helping stuff for build_clique_star_cstr
 194  */
 195 typedef struct _edge_t {
 196         ir_node *n1, *n2;
 197 } edge_t;
 198
 199 static int compare_edge_t(const void *k1, const void *k2, size_t size) {
 200         const edge_t *e1 = k1;
 201         const edge_t *e2 = k2;
 202
 203         return ! (e1->n1 == e2->n1   &&   e1->n2 == e2->n2);
 204 }
 205
 206 #define HASH_EDGE(e) (nodeset_hash((e)->n1) ^ nodeset_hash((e)->n2))
 207
 208 static INLINE edge_t *add_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
 209         edge_t new_edge;
 210
 211         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 212                 new_edge.n1 = n1;
 213                 new_edge.n2 = n2;
 214         } else {
 215                 new_edge.n1 = n2;
 216                 new_edge.n2 = n1;
 217         }
 218         (*counter)++;
 219         return set_insert(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 220 }
 221
 222 static INLINE edge_t *find_edge(set *edges, ir_node *n1, ir_node *n2) {
 223         edge_t new_edge;
 224
 225         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 226                 new_edge.n1 = n1;
 227                 new_edge.n2 = n2;
 228         } else {
 229                 new_edge.n1 = n2;
 230                 new_edge.n2 = n1;
 231         }
 232         return set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 233 }
 234
 235 static INLINE void remove_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
 236         edge_t new_edge, *e;
 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         e = set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 246         if (e) {
 247                 e->n1 = NULL;
 248                 e->n2 = NULL;
 249                 (*counter)--;
 250         }
 251 }
 252
 253 #define pset_foreach(pset, irn)  for(irn=pset_first(pset); irn; irn=pset_next(pset))
 254
 255 /**
 256  * Search for an interference clique and an external node
 257  * with affinity edges to all nodes of the clique.
 258  * At most 1 node of the clique can be colored equally with the external node.
 259  */
 260 static void build_clique_star_cstr(ilp_env_t *ienv) {
 261         affinity_node_t *aff;
 262
 263         /* for each node with affinity edges */
 264         co_gs_foreach_aff_node(ienv->co, aff) {
 265                 struct obstack ob;
 266                 neighb_t *nbr;
 267                 ir_node *center = aff->irn;
 268                 ir_node **nodes;
 269                 set *edges;
 270                 int i, o, n_nodes, n_edges;
 271
 272                 obstack_init(&ob);
 273                 edges = new_set(compare_edge_t, 8);
 274
 275                 /* get all affinity neighbours */
 276                 n_nodes = 0;
 277                 co_gs_foreach_neighb(aff, nbr) {
 278                         obstack_ptr_grow(&ob, nbr->irn);
 279                         ++n_nodes;
 280                 }
 281                 nodes = obstack_finish(&ob);
 282
 283                 /* get all interference edges between these */
 284                 n_edges = 0;
 285                 for (i=0; i<n_nodes; ++i)
 286                         for (o=0; o<i; ++o)
 287                                 if (be_ifg_connected(ienv->co->cenv->ifg, nodes[i], nodes[o]))
 288                                         add_edge(edges, nodes[i], nodes[o], &n_edges);
 289
 290                 /* cover all these interference edges with maximal cliques */
 291                 while (n_edges) {
 292                         edge_t *e;
 293                         pset *clique = pset_new_ptr(8);
 294                         int growed;
 295
 296                         /* get 2 starting nodes to form a clique */
 297                         for (e=set_first(edges); !e->n1; e=set_next(edges))
 298                                 /*nothing*/ ;
 299
 300                         pset_insert_ptr(clique, e->n1);
 301                         pset_insert_ptr(clique, e->n2);
 302                         remove_edge(edges, e->n1, e->n2, &n_edges);
 303
 304                         /* while the clique is growing */
 305                         do {
 306                                 growed = 0;
 307
 308                                 /* search for a candidate to extend the clique */
 309                                 for (i=0; i<n_nodes; ++i) {
 310                                         ir_node *member, *cand = nodes[i];
 311                                         int is_cand;
 312
 313                                         /* if its already in the clique try the next */
 314                                         if (pset_find_ptr(clique, cand))
 315                                                 continue;
 316
 317                                         /* are there all necessary interferences? */
 318                                         is_cand = 1;
 319                                         pset_foreach(clique, member) {
 320                                                 if (!find_edge(edges, cand, member)) {
 321                                                         is_cand = 0;
 322                                                         pset_break(clique);
 323                                                         break;
 324                                                 }
 325                                         }
 326
 327                                         /* now we know if we have a clique extender */
 328                                         if (is_cand) {
 329                                                 /* first remove all covered edges */
 330                                                 pset_foreach(clique, member)
 331                                                         remove_edge(edges, cand, member, &n_edges);
 332
 333                                                 /* insert into clique */
 334                                                 pset_insert_ptr(clique, cand);
 335                                                 growed = 1;
 336                                                 break;
 337                                         }
 338                                 }
 339                         } while (growed);
 340
 341                         /* now the clique is maximal. Finally add the constraint */
 342                         {
 343                                 ir_node *member;
 344                                 int var_idx, cst_idx, center_nr, member_nr;
 345                                 char buf[16];
 346
 347                                 cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, pset_count(clique)-1);
 348                                 center_nr = get_irn_node_nr(center);
 349
 350                                 pset_foreach(clique, member) {
 351                                         member_nr = get_irn_node_nr(member);
 352                                         var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', center_nr, member_nr));
 353                                         lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 354                                 }
 355                         }
 356
 357                         del_pset(clique);
 358                 }
 359
 360                 del_set(edges);
 361                 obstack_free(&ob, NULL);
 362         }
 363 }
 364
 365
 366 static void extend_path(ilp_env_t *ienv, pdeq *path, ir_node *irn) {
 367         be_ifg_t *ifg = ienv->co->cenv->ifg;
 368         int i, len;
 369         ir_node **curr_path;
 370         affinity_node_t *aff;
 371         neighb_t *nbr;
 372
 373         /* do not walk backwards or in circles */
 374         if (pdeq_contains(path, irn))
 375                 return;
 376
 377         /* insert the new irn */
 378         pdeq_putr(path, irn);
 379
 380
 381
 382         /* check for forbidden interferences */
 383         len = pdeq_len(path);
 384         curr_path = alloca(len * sizeof(*curr_path));
 385         pdeq_copyl(path, (const void **)curr_path);
 386
 387         for (i=1; i<len; ++i)
 388                 if (be_ifg_connected(ifg, irn, curr_path[i]))
 389                         goto end;
 390
 391
 392
 393         /* check for terminating interference */
 394         if (be_ifg_connected(ifg, irn, curr_path[0])) {
 395
 396                 /* One node is not a path. */
 397                 /* And a path of length 2 is covered by a clique star constraint. */
 398                 if (len > 2) {
 399                         /* finally build the constraint */
 400                         int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, 1.0);
 401                         for (i=1; i<len; ++i) {
 402                                 char buf[16];
 403                                 int nr_1    = get_irn_node_nr(curr_path[i-1]);
 404                                 int nr_2    = get_irn_node_nr(curr_path[i]);
 405                                 int var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', nr_1, nr_2));
 406                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 407                         }
 408                 }
 409
 410                 /* this path cannot be extended anymore */
 411                 goto end;
 412         }
 413
 414
 415
 416         /* recursively extend the path */
 417         aff = get_affinity_info(ienv->co, irn);
 418         co_gs_foreach_neighb(aff, nbr)
 419                 extend_path(ienv, path, nbr->irn);
 420
 421
 422 end:
 423         /* remove the irn */
 424         pdeq_getr(path);
 425
 426 }
 427
 428 /**
 429  *  Search a path of affinity edges, whose ends are connected
 430  *  by an interference edge and there are no other interference
 431  *  edges in between.
 432  *  Then at least one of these affinity edges must break.
 433  */
 434 static void build_path_cstr(ilp_env_t *ienv) {
 435         affinity_node_t *aff_info;
 436
 437         /* for each node with affinity edges */
 438         co_gs_foreach_aff_node(ienv->co, aff_info) {
 439                 pdeq *path = new_pdeq();
 440
 441                 extend_path(ienv, path, aff_info->irn);
 442
 443                 del_pdeq(path);
 444         }
 445 }
 446
 447 static void ilp2_build(ilp_env_t *ienv) {
 448         local_env_t *lenv = ienv->env;
 449         int lower_bound;
 450
 451         ienv->lp = new_lpp(ienv->co->name, lpp_minimize);
 452         build_coloring_cstr(ienv);
 453         build_interference_cstr(ienv);
 454         build_affinity_cstr(ienv);
 455         build_clique_star_cstr(ienv);
 456         build_path_cstr(ienv);
 457
 458         lower_bound = co_get_lower_bound(ienv->co) - co_get_inevit_copy_costs(ienv->co);
 459         lpp_set_bound(ienv->lp, lower_bound);
 460         lpp_set_time_limit(ienv->lp, lenv->time_limit);
 461 }
 462
 463 static void ilp2_apply(ilp_env_t *ienv) {
 464         local_env_t *lenv = ienv->env;
 465         double *sol;
 466         lpp_sol_state_t state;
 467         int i, count;
 468
 469         /* first check if there was sth. to optimize */
 470         if (lenv->first_x_var >= 0) {
 471
 472                 count = lenv->last_x_var - lenv->first_x_var + 1;
 473                 sol = xmalloc(count * sizeof(sol[0]));
 474                 state = lpp_get_solution(ienv->lp, sol, lenv->first_x_var, lenv->last_x_var);
 475                 if (state != lpp_optimal) {
 476                         printf("WARNING %s: Solution state is not 'optimal': %d\n", ienv->co->name, state);
 477                         assert(state >= lpp_feasible && "The solution should at least be feasible!");
 478                 }
 479
 480                 for (i=0; i<count; ++i) {
 481                         int nodenr, color;
 482                         char var_name[16];
 483
 484                         if (sol[i] > 1-EPSILON) { /* split variable name into components */
 485                                 lpp_get_var_name(ienv->lp, lenv->first_x_var+i, var_name, sizeof(var_name));
 486
 487                                 if (sscanf(var_name, "x_%d_%d", &nodenr, &color) == 2) {
 488                                         ir_node *irn = pmap_get(lenv->nr_2_irn, INT_TO_PTR(nodenr));
 489                                         assert(irn && "This node number must be present in the map");
 490
 491                                         set_irn_col(ienv->co, irn, color);
 492                                 } else
 493                                         assert(0 && "This should be a x-var");
 494                         }
 495                 }
 496         }
 497
 498 #ifdef COPYOPT_STAT
 499         /* TODO adapt to multiple possible ILPs */
 500         copystat_add_ilp_time((int)(1000.0*lpp_get_sol_time(pi->curr_lp)));  //now we have ms
 501         copystat_add_ilp_vars(lpp_get_var_count(pi->curr_lp));
 502         copystat_add_ilp_csts(lpp_get_cst_count(pi->curr_lp));
 503         copystat_add_ilp_iter(lpp_get_iter_cnt(pi->curr_lp));
 504 #endif
 505 }
 506
 507 int co_solve_ilp2(copy_opt_t *co) {
 508         lpp_sol_state_t sol_state;
 509         ilp_env_t *ienv;
 510         local_env_t my;
 511
 512         ASSERT_OU_AVAIL(co); //See build_clique_st
 513         ASSERT_GS_AVAIL(co);
 514
 515         my.time_limit  = 0;
 516         my.first_x_var = -1;
 517         my.last_x_var  = -1;
 518         my.nr_2_irn    = pmap_create();
 519         FIRM_DBG_REGISTER(my.dbg, "firm.be.coilp2");
 520
 521         ienv = new_ilp_env(co, ilp2_build, ilp2_apply, &my);
 522
 523         sol_state = ilp_go(ienv);
 524
 525         pmap_destroy(my.nr_2_irn);
 526         free_ilp_env(ienv);
 527
 528         return sol_state == lpp_optimal;
 529 }
 530
 531 #else /* WITH_ILP */
 532
 533 static void only_that_you_can_compile_without_WITH_ILP_defined(void) {
 534 }
 535
 536 #endif /* WITH_ILP */