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