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 #ifdef WITH_ILP
  52
  53 #include "bitset.h"
  54 #include "raw_bitset.h"
  55 #include "pdeq.h"
  56
  57 #include "irtools.h"
  58 #include "irgwalk.h"
  59 #include "becopyilp_t.h"
  60 #include "beifg_t.h"
  61 #include "besched_t.h"
  62 #include "benodesets.h"
  63
  64 #define DEBUG_LVL 1
  65
  66 typedef struct _local_env_t {
  67         double time_limit;
  68         int first_x_var, last_x_var;
  69         pmap *nr_2_irn;
  70         DEBUG_ONLY(firm_dbg_module_t *dbg;)
  71 } local_env_t;
  72
  73 static void build_coloring_cstr(ilp_env_t *ienv) {
  74         be_ifg_t *ifg = ienv->co->cenv->ifg;
  75         void *iter = be_ifg_nodes_iter_alloca(ifg);
  76         bitset_t *colors;
  77         ir_node *irn;
  78         char buf[16];
  79
  80         colors = bitset_alloca(arch_register_class_n_regs(ienv->co->cls));
  81
  82         be_ifg_foreach_node(ifg, iter, irn)
  83                 if (!sr_is_removed(ienv->sr, irn)) {
  84                         int col, cst_idx;
  85                         const arch_register_req_t *req;
  86                         int curr_node_color = get_irn_col(ienv->co, irn);
  87                         int node_nr = (int)get_irn_node_nr(irn);
  88                         local_env_t *lenv = ienv->env;
  89
  90                         pmap_insert(lenv->nr_2_irn, INT_TO_PTR(node_nr), irn);
  91
  92                         req = arch_get_register_req(ienv->co->aenv, irn, -1);
  93
  94                         /* get assignable colors */
  95                         if (arch_register_req_is(req, limited)) {
  96                                 rbitset_copy_to_bitset(req->limited, colors);
  97                         } else {
  98                                 arch_register_class_put(req->cls, colors);
  99                                 // bitset_andnot(colors, ienv->co->cenv->ignore_colors);
 100                         }
 101
 102                         /* add the coloring constraint */
 103                         cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 1.0);
 104
 105                         bitset_foreach(colors, col) {
 106                                 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, col), lpp_binary, 0.0);
 107                                 lpp_set_start_value(ienv->lp, var_idx, (col == curr_node_color) ? 1.0 : 0.0);
 108                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
 109
 110                                 lenv->last_x_var = var_idx;
 111                                 if (lenv->first_x_var == -1)
 112                                         lenv->first_x_var = var_idx;
 113                         }
 114
 115                         /* add register constraint constraints */
 116                         bitset_foreach_clear(colors, col) {
 117                                 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 0.0);
 118                                 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, col), lpp_binary, 0.0);
 119                                 lpp_set_start_value(ienv->lp, var_idx, 0.0);
 120                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
 121
 122                                 lenv->last_x_var = var_idx;
 123                         }
 124                 }
 125 }
 126
 127 static void build_interference_cstr(ilp_env_t *ienv) {
 128         lpp_t *lpp = ienv->lp;
 129         be_ifg_t *ifg = ienv->co->cenv->ifg;
 130         int n_colors = arch_register_class_n_regs(ienv->co->cls);
 131         int i, col;
 132
 133         void *iter = be_ifg_cliques_iter_alloca(ifg);
 134         ir_node **clique = alloca(sizeof(*clique) * n_colors);
 135         int size;
 136
 137         char buf[16];
 138
 139         /* for each maximal clique */
 140         be_ifg_foreach_clique(ifg, iter, clique, &size) {
 141                 int realsize = 0;
 142
 143                 for (i=0; i<size; ++i)
 144                         if (!sr_is_removed(ienv->sr, clique[i]))
 145                                 ++realsize;
 146
 147                 if (realsize < 2)
 148                         continue;
 149
 150                 /* for all colors */
 151                 for (col=0; col<n_colors; ++col) {
 152                         int cst_idx = lpp_add_cst(lpp, NULL, lpp_less, 1.0);
 153
 154                         /* for each member of this clique */
 155                         for (i=0; i<size; ++i) {
 156                                 ir_node *irn = clique[i];
 157
 158                                 if (!sr_is_removed(ienv->sr, irn)) {
 159                                         int var_idx = lpp_get_var_idx(lpp, name_cdd(buf, 'x', (int)get_irn_node_nr(irn), col));
 160                                         lpp_set_factor_fast(lpp, cst_idx, var_idx, 1);
 161                                 }
 162                         }
 163                 }
 164         }
 165 }
 166
 167
 168 /**
 169  * TODO: Remove the dependency of the opt-units data structure
 170  *       by walking over all affinity edges. Graph structure
 171  *       does not provide this walker, yet.
 172  */
 173 static void build_affinity_cstr(ilp_env_t *ienv) {
 174         unit_t *curr;
 175         int n_colors = arch_register_class_n_regs(ienv->co->cls);
 176
 177         /* for all optimization units */
 178         list_for_each_entry(unit_t, curr, &ienv->co->units, units) {
 179                 ir_node *root, *arg;
 180                 int root_nr, arg_nr, i, col, y_idx, root_idx, arg_idx;
 181                 char buf[16];
 182                 int root_col, arg_col;
 183
 184                 root = curr->nodes[0];
 185                 root_nr = (int) get_irn_node_nr(root);
 186                 root_col = get_irn_col(ienv->co, root);
 187
 188                 for (i = 1; i < curr->node_count; ++i) {
 189                         arg = curr->nodes[i];
 190                         arg_nr = (int) get_irn_node_nr(arg);
 191                         arg_col = get_irn_col(ienv->co, arg);
 192
 193                         /* add a new affinity variable */
 194                         y_idx = lpp_add_var(ienv->lp, name_cdd_sorted(buf, 'y', root_nr, arg_nr), lpp_binary, curr->costs[i]);
 195                         lpp_set_start_value(ienv->lp, y_idx, (root_col==arg_col) ? 0.0 : 1.0);
 196
 197                         /* add constraints relating the affinity var to the color vars */
 198                         for (col=0; col<n_colors; ++col) {
 199                                 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_less, 0.0);
 200                                 root_idx = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', root_nr, col));
 201                                 arg_idx  = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', arg_nr,  col));
 202
 203                                 lpp_set_factor_fast(ienv->lp, cst_idx, root_idx,  1.0);
 204                                 lpp_set_factor_fast(ienv->lp, cst_idx, arg_idx,  -1.0);
 205                                 lpp_set_factor_fast(ienv->lp, cst_idx, y_idx, -1.0);
 206                         }
 207                 }
 208         }
 209 }
 210
 211 /**
 212  * Helping stuff for build_clique_star_cstr
 213  */
 214 typedef struct _edge_t {
 215         ir_node *n1, *n2;
 216 } edge_t;
 217
 218 static int compare_edge_t(const void *k1, const void *k2, size_t size) {
 219         const edge_t *e1 = k1;
 220         const edge_t *e2 = k2;
 221
 222         return ! (e1->n1 == e2->n1   &&   e1->n2 == e2->n2);
 223 }
 224
 225 #define HASH_EDGE(e) (nodeset_hash((e)->n1) ^ nodeset_hash((e)->n2))
 226
 227 static INLINE edge_t *add_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
 228         edge_t new_edge;
 229
 230         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 231                 new_edge.n1 = n1;
 232                 new_edge.n2 = n2;
 233         } else {
 234                 new_edge.n1 = n2;
 235                 new_edge.n2 = n1;
 236         }
 237         (*counter)++;
 238         return set_insert(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 239 }
 240
 241 static INLINE edge_t *find_edge(set *edges, ir_node *n1, ir_node *n2) {
 242         edge_t new_edge;
 243
 244         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 245                 new_edge.n1 = n1;
 246                 new_edge.n2 = n2;
 247         } else {
 248                 new_edge.n1 = n2;
 249                 new_edge.n2 = n1;
 250         }
 251         return set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 252 }
 253
 254 static INLINE void remove_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
 255         edge_t new_edge, *e;
 256
 257         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 258                 new_edge.n1 = n1;
 259                 new_edge.n2 = n2;
 260         } else {
 261                 new_edge.n1 = n2;
 262                 new_edge.n2 = n1;
 263         }
 264         e = set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 265         if (e) {
 266                 e->n1 = NULL;
 267                 e->n2 = NULL;
 268                 (*counter)--;
 269         }
 270 }
 271
 272 #define pset_foreach(pset, irn)  for(irn=pset_first(pset); irn; irn=pset_next(pset))
 273
 274 /**
 275  * Search for an interference clique and an external node
 276  * with affinity edges to all nodes of the clique.
 277  * At most 1 node of the clique can be colored equally with the external node.
 278  */
 279 static void build_clique_star_cstr(ilp_env_t *ienv) {
 280         affinity_node_t *aff;
 281
 282         /* for each node with affinity edges */
 283         co_gs_foreach_aff_node(ienv->co, aff) {
 284                 struct obstack ob;
 285                 neighb_t *nbr;
 286                 ir_node *center = aff->irn;
 287                 ir_node **nodes;
 288                 set *edges;
 289                 int i, o, n_nodes, n_edges;
 290
 291                 obstack_init(&ob);
 292                 edges = new_set(compare_edge_t, 8);
 293
 294                 /* get all affinity neighbours */
 295                 n_nodes = 0;
 296                 co_gs_foreach_neighb(aff, nbr) {
 297                         obstack_ptr_grow(&ob, nbr->irn);
 298                         ++n_nodes;
 299                 }
 300                 nodes = obstack_finish(&ob);
 301
 302                 /* get all interference edges between these */
 303                 n_edges = 0;
 304                 for (i=0; i<n_nodes; ++i)
 305                         for (o=0; o<i; ++o)
 306                                 if (be_ifg_connected(ienv->co->cenv->ifg, nodes[i], nodes[o]))
 307                                         add_edge(edges, nodes[i], nodes[o], &n_edges);
 308
 309                 /* cover all these interference edges with maximal cliques */
 310                 while (n_edges) {
 311                         edge_t *e;
 312                         pset *clique = pset_new_ptr(8);
 313                         int growed;
 314
 315                         /* get 2 starting nodes to form a clique */
 316                         for (e=set_first(edges); !e->n1; e=set_next(edges))
 317                                 /*nothing*/ ;
 318
 319                         /* we could be stepped out of the loop before the set iterated to the end */
 320                         set_break(edges);
 321
 322                         pset_insert_ptr(clique, e->n1);
 323                         pset_insert_ptr(clique, e->n2);
 324                         remove_edge(edges, e->n1, e->n2, &n_edges);
 325
 326                         /* while the clique is growing */
 327                         do {
 328                                 growed = 0;
 329
 330                                 /* search for a candidate to extend the clique */
 331                                 for (i=0; i<n_nodes; ++i) {
 332                                         ir_node *member, *cand = nodes[i];
 333                                         int is_cand;
 334
 335                                         /* if its already in the clique try the next */
 336                                         if (pset_find_ptr(clique, cand))
 337                                                 continue;
 338
 339                                         /* are there all necessary interferences? */
 340                                         is_cand = 1;
 341                                         pset_foreach(clique, member) {
 342                                                 if (!find_edge(edges, cand, member)) {
 343                                                         is_cand = 0;
 344                                                         pset_break(clique);
 345                                                         break;
 346                                                 }
 347                                         }
 348
 349                                         /* now we know if we have a clique extender */
 350                                         if (is_cand) {
 351                                                 /* first remove all covered edges */
 352                                                 pset_foreach(clique, member)
 353                                                         remove_edge(edges, cand, member, &n_edges);
 354
 355                                                 /* insert into clique */
 356                                                 pset_insert_ptr(clique, cand);
 357                                                 growed = 1;
 358                                                 break;
 359                                         }
 360                                 }
 361                         } while (growed);
 362
 363                         /* now the clique is maximal. Finally add the constraint */
 364                         {
 365                                 ir_node *member;
 366                                 int var_idx, cst_idx, center_nr, member_nr;
 367                                 char buf[16];
 368
 369                                 cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, pset_count(clique)-1);
 370                                 center_nr = get_irn_node_nr(center);
 371
 372                                 pset_foreach(clique, member) {
 373                                         member_nr = get_irn_node_nr(member);
 374                                         var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', center_nr, member_nr));
 375                                         lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 376                                 }
 377                         }
 378
 379                         del_pset(clique);
 380                 }
 381
 382                 del_set(edges);
 383                 obstack_free(&ob, NULL);
 384         }
 385 }
 386
 387
 388 static void extend_path(ilp_env_t *ienv, pdeq *path, ir_node *irn) {
 389         be_ifg_t *ifg = ienv->co->cenv->ifg;
 390         int i, len;
 391         ir_node **curr_path;
 392         affinity_node_t *aff;
 393         neighb_t *nbr;
 394
 395         /* do not walk backwards or in circles */
 396         if (pdeq_contains(path, irn))
 397                 return;
 398
 399         /* insert the new irn */
 400         pdeq_putr(path, irn);
 401
 402
 403
 404         /* check for forbidden interferences */
 405         len = pdeq_len(path);
 406         curr_path = alloca(len * sizeof(*curr_path));
 407         pdeq_copyl(path, (const void **)curr_path);
 408
 409         for (i=1; i<len; ++i)
 410                 if (be_ifg_connected(ifg, irn, curr_path[i]))
 411                         goto end;
 412
 413
 414
 415         /* check for terminating interference */
 416         if (be_ifg_connected(ifg, irn, curr_path[0])) {
 417
 418                 /* One node is not a path. */
 419                 /* And a path of length 2 is covered by a clique star constraint. */
 420                 if (len > 2) {
 421                         /* finally build the constraint */
 422                         int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, 1.0);
 423                         for (i=1; i<len; ++i) {
 424                                 char buf[16];
 425                                 int nr_1    = get_irn_node_nr(curr_path[i-1]);
 426                                 int nr_2    = get_irn_node_nr(curr_path[i]);
 427                                 int var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', nr_1, nr_2));
 428                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 429                         }
 430                 }
 431
 432                 /* this path cannot be extended anymore */
 433                 goto end;
 434         }
 435
 436
 437
 438         /* recursively extend the path */
 439         aff = get_affinity_info(ienv->co, irn);
 440         co_gs_foreach_neighb(aff, nbr)
 441                 extend_path(ienv, path, nbr->irn);
 442
 443
 444 end:
 445         /* remove the irn */
 446         pdeq_getr(path);
 447
 448 }
 449
 450 /**
 451  *  Search a path of affinity edges, whose ends are connected
 452  *  by an interference edge and there are no other interference
 453  *  edges in between.
 454  *  Then at least one of these affinity edges must break.
 455  */
 456 static void build_path_cstr(ilp_env_t *ienv) {
 457         affinity_node_t *aff_info;
 458
 459         /* for each node with affinity edges */
 460         co_gs_foreach_aff_node(ienv->co, aff_info) {
 461                 pdeq *path = new_pdeq();
 462
 463                 extend_path(ienv, path, aff_info->irn);
 464
 465                 del_pdeq(path);
 466         }
 467 }
 468
 469 static void ilp2_build(ilp_env_t *ienv) {
 470         local_env_t *lenv = ienv->env;
 471         int lower_bound;
 472
 473         ienv->lp = new_lpp(ienv->co->name, lpp_minimize);
 474         build_coloring_cstr(ienv);
 475         build_interference_cstr(ienv);
 476         build_affinity_cstr(ienv);
 477         build_clique_star_cstr(ienv);
 478         build_path_cstr(ienv);
 479
 480         lower_bound = co_get_lower_bound(ienv->co) - co_get_inevit_copy_costs(ienv->co);
 481         lpp_set_bound(ienv->lp, lower_bound);
 482         lpp_set_time_limit(ienv->lp, lenv->time_limit);
 483 }
 484
 485 static void ilp2_apply(ilp_env_t *ienv) {
 486         local_env_t *lenv = ienv->env;
 487         double *sol;
 488         lpp_sol_state_t state;
 489         int i, count;
 490
 491         /* first check if there was sth. to optimize */
 492         if (lenv->first_x_var >= 0) {
 493
 494                 count = lenv->last_x_var - lenv->first_x_var + 1;
 495                 sol = xmalloc(count * sizeof(sol[0]));
 496                 state = lpp_get_solution(ienv->lp, sol, lenv->first_x_var, lenv->last_x_var);
 497                 if (state != lpp_optimal) {
 498                         printf("WARNING %s: Solution state is not 'optimal': %d\n", ienv->co->name, state);
 499                         assert(state >= lpp_feasible && "The solution should at least be feasible!");
 500                 }
 501
 502                 for (i=0; i<count; ++i) {
 503                         int nodenr, color;
 504                         char var_name[16];
 505
 506                         if (sol[i] > 1-EPSILON) { /* split variable name into components */
 507                                 lpp_get_var_name(ienv->lp, lenv->first_x_var+i, var_name, sizeof(var_name));
 508
 509                                 if (sscanf(var_name, "x_%d_%d", &nodenr, &color) == 2) {
 510                                         ir_node *irn = pmap_get(lenv->nr_2_irn, INT_TO_PTR(nodenr));
 511                                         assert(irn && "This node number must be present in the map");
 512
 513                                         set_irn_col(ienv->co, irn, color);
 514                                 } else
 515                                         assert(0 && "This should be a x-var");
 516                         }
 517                 }
 518         }
 519
 520 #ifdef COPYOPT_STAT
 521         /* TODO adapt to multiple possible ILPs */
 522         copystat_add_ilp_time((int)(1000.0*lpp_get_sol_time(pi->curr_lp)));  //now we have ms
 523         copystat_add_ilp_vars(lpp_get_var_count(pi->curr_lp));
 524         copystat_add_ilp_csts(lpp_get_cst_count(pi->curr_lp));
 525         copystat_add_ilp_iter(lpp_get_iter_cnt(pi->curr_lp));
 526 #endif
 527 }
 528
 529 int co_solve_ilp2(copy_opt_t *co) {
 530         lpp_sol_state_t sol_state;
 531         ilp_env_t *ienv;
 532         local_env_t my;
 533
 534         ASSERT_OU_AVAIL(co); //See build_clique_st
 535         ASSERT_GS_AVAIL(co);
 536
 537         my.time_limit  = 0;
 538         my.first_x_var = -1;
 539         my.last_x_var  = -1;
 540         my.nr_2_irn    = pmap_create();
 541         FIRM_DBG_REGISTER(my.dbg, "firm.be.coilp2");
 542
 543         ienv = new_ilp_env(co, ilp2_build, ilp2_apply, &my);
 544
 545         sol_state = ilp_go(ienv);
 546
 547         pmap_destroy(my.nr_2_irn);
 548         free_ilp_env(ienv);
 549
 550         return sol_state == lpp_optimal;
 551 }
 552
 553 #else /* WITH_ILP */
 554
 555 static INLINE void only_that_you_can_compile_without_WITH_ILP_defined(void) {
 556 }
 557
 558 #endif /* WITH_ILP */