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