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