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                         bitset_pos_t col;
  85                         int cst_idx;
  86                         const arch_register_req_t *req;
  87                         int curr_node_color = get_irn_col(ienv->co, irn);
  88                         int node_nr = (int)get_irn_node_nr(irn);
  89                         local_env_t *lenv = ienv->env;
  90
  91                         pmap_insert(lenv->nr_2_irn, INT_TO_PTR(node_nr), irn);
  92
  93                         req = arch_get_register_req(ienv->co->aenv, irn, -1);
  94
  95                         /* get assignable colors */
  96                         if (arch_register_req_is(req, limited)) {
  97                                 rbitset_copy_to_bitset(req->limited, colors);
  98                         } else {
  99                                 arch_register_class_put(req->cls, colors);
 100                                 // bitset_andnot(colors, ienv->co->cenv->ignore_colors);
 101                         }
 102
 103                         /* add the coloring constraint */
 104                         cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 1.0);
 105
 106                         bitset_foreach(colors, col) {
 107                                 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, col), lpp_binary, 0.0);
 108                                 lpp_set_start_value(ienv->lp, var_idx, (col == (unsigned) curr_node_color) ? 1.0 : 0.0);
 109                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
 110
 111                                 lenv->last_x_var = var_idx;
 112                                 if (lenv->first_x_var == -1)
 113                                         lenv->first_x_var = var_idx;
 114                         }
 115
 116                         /* add register constraint constraints */
 117                         bitset_foreach_clear(colors, col) {
 118                                 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 0.0);
 119                                 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, col), lpp_binary, 0.0);
 120                                 lpp_set_start_value(ienv->lp, var_idx, 0.0);
 121                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
 122
 123                                 lenv->last_x_var = var_idx;
 124                         }
 125                 }
 126 }
 127
 128 static void build_interference_cstr(ilp_env_t *ienv) {
 129         lpp_t *lpp = ienv->lp;
 130         be_ifg_t *ifg = ienv->co->cenv->ifg;
 131         int n_colors = arch_register_class_n_regs(ienv->co->cls);
 132         int i, col;
 133
 134         void *iter = be_ifg_cliques_iter_alloca(ifg);
 135         ir_node **clique = alloca(sizeof(*clique) * n_colors);
 136         int size;
 137
 138         char buf[16];
 139
 140         /* for each maximal clique */
 141         be_ifg_foreach_clique(ifg, iter, clique, &size) {
 142                 int realsize = 0;
 143
 144                 for (i=0; i<size; ++i)
 145                         if (!sr_is_removed(ienv->sr, clique[i]))
 146                                 ++realsize;
 147
 148                 if (realsize < 2)
 149                         continue;
 150
 151                 /* for all colors */
 152                 for (col=0; col<n_colors; ++col) {
 153                         int cst_idx = lpp_add_cst(lpp, NULL, lpp_less, 1.0);
 154
 155                         /* for each member of this clique */
 156                         for (i=0; i<size; ++i) {
 157                                 ir_node *irn = clique[i];
 158
 159                                 if (!sr_is_removed(ienv->sr, irn)) {
 160                                         int var_idx = lpp_get_var_idx(lpp, name_cdd(buf, 'x', (int)get_irn_node_nr(irn), col));
 161                                         lpp_set_factor_fast(lpp, cst_idx, var_idx, 1);
 162                                 }
 163                         }
 164                 }
 165         }
 166 }
 167
 168
 169 /**
 170  * TODO: Remove the dependency of the opt-units data structure
 171  *       by walking over all affinity edges. Graph structure
 172  *       does not provide this walker, yet.
 173  */
 174 static void build_affinity_cstr(ilp_env_t *ienv) {
 175         unit_t *curr;
 176         int n_colors = arch_register_class_n_regs(ienv->co->cls);
 177
 178         /* for all optimization units */
 179         list_for_each_entry(unit_t, curr, &ienv->co->units, units) {
 180                 ir_node *root, *arg;
 181                 int root_nr, arg_nr, i, col, y_idx, root_idx, arg_idx;
 182                 char buf[16];
 183                 int root_col, arg_col;
 184
 185                 root = curr->nodes[0];
 186                 root_nr = (int) get_irn_node_nr(root);
 187                 root_col = get_irn_col(ienv->co, root);
 188
 189                 for (i = 1; i < curr->node_count; ++i) {
 190                         arg = curr->nodes[i];
 191                         arg_nr = (int) get_irn_node_nr(arg);
 192                         arg_col = get_irn_col(ienv->co, arg);
 193
 194                         /* add a new affinity variable */
 195                         y_idx = lpp_add_var(ienv->lp, name_cdd_sorted(buf, 'y', root_nr, arg_nr), lpp_binary, curr->costs[i]);
 196                         lpp_set_start_value(ienv->lp, y_idx, (root_col==arg_col) ? 0.0 : 1.0);
 197
 198                         /* add constraints relating the affinity var to the color vars */
 199                         for (col=0; col<n_colors; ++col) {
 200                                 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_less, 0.0);
 201                                 root_idx = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', root_nr, col));
 202                                 arg_idx  = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', arg_nr,  col));
 203
 204                                 lpp_set_factor_fast(ienv->lp, cst_idx, root_idx,  1.0);
 205                                 lpp_set_factor_fast(ienv->lp, cst_idx, arg_idx,  -1.0);
 206                                 lpp_set_factor_fast(ienv->lp, cst_idx, y_idx, -1.0);
 207                         }
 208                 }
 209         }
 210 }
 211
 212 /**
 213  * Helping stuff for build_clique_star_cstr
 214  */
 215 typedef struct _edge_t {
 216         ir_node *n1, *n2;
 217 } edge_t;
 218
 219 static int compare_edge_t(const void *k1, const void *k2, size_t size) {
 220         const edge_t *e1 = k1;
 221         const edge_t *e2 = k2;
 222         (void) size;
 223
 224         return ! (e1->n1 == e2->n1   &&   e1->n2 == e2->n2);
 225 }
 226
 227 #define HASH_EDGE(e) (nodeset_hash((e)->n1) ^ nodeset_hash((e)->n2))
 228
 229 static INLINE edge_t *add_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
 230         edge_t new_edge;
 231
 232         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 233                 new_edge.n1 = n1;
 234                 new_edge.n2 = n2;
 235         } else {
 236                 new_edge.n1 = n2;
 237                 new_edge.n2 = n1;
 238         }
 239         (*counter)++;
 240         return set_insert(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 241 }
 242
 243 static INLINE edge_t *find_edge(set *edges, ir_node *n1, ir_node *n2) {
 244         edge_t new_edge;
 245
 246         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 247                 new_edge.n1 = n1;
 248                 new_edge.n2 = n2;
 249         } else {
 250                 new_edge.n1 = n2;
 251                 new_edge.n2 = n1;
 252         }
 253         return set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 254 }
 255
 256 static INLINE void remove_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
 257         edge_t new_edge, *e;
 258
 259         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 260                 new_edge.n1 = n1;
 261                 new_edge.n2 = n2;
 262         } else {
 263                 new_edge.n1 = n2;
 264                 new_edge.n2 = n1;
 265         }
 266         e = set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 267         if (e) {
 268                 e->n1 = NULL;
 269                 e->n2 = NULL;
 270                 (*counter)--;
 271         }
 272 }
 273
 274 #define pset_foreach(pset, irn)  for(irn=pset_first(pset); irn; irn=pset_next(pset))
 275
 276 /**
 277  * Search for an interference clique and an external node
 278  * with affinity edges to all nodes of the clique.
 279  * At most 1 node of the clique can be colored equally with the external node.
 280  */
 281 static void build_clique_star_cstr(ilp_env_t *ienv) {
 282         affinity_node_t *aff;
 283
 284         /* for each node with affinity edges */
 285         co_gs_foreach_aff_node(ienv->co, aff) {
 286                 struct obstack ob;
 287                 neighb_t *nbr;
 288                 ir_node *center = aff->irn;
 289                 ir_node **nodes;
 290                 set *edges;
 291                 int i, o, n_nodes, n_edges;
 292
 293                 obstack_init(&ob);
 294                 edges = new_set(compare_edge_t, 8);
 295
 296                 /* get all affinity neighbours */
 297                 n_nodes = 0;
 298                 co_gs_foreach_neighb(aff, nbr) {
 299                         obstack_ptr_grow(&ob, nbr->irn);
 300                         ++n_nodes;
 301                 }
 302                 nodes = obstack_finish(&ob);
 303
 304                 /* get all interference edges between these */
 305                 n_edges = 0;
 306                 for (i=0; i<n_nodes; ++i)
 307                         for (o=0; o<i; ++o)
 308                                 if (be_ifg_connected(ienv->co->cenv->ifg, nodes[i], nodes[o]))
 309                                         add_edge(edges, nodes[i], nodes[o], &n_edges);
 310
 311                 /* cover all these interference edges with maximal cliques */
 312                 while (n_edges) {
 313                         edge_t *e;
 314                         pset *clique = pset_new_ptr(8);
 315                         int growed;
 316
 317                         /* get 2 starting nodes to form a clique */
 318                         for (e=set_first(edges); !e->n1; e=set_next(edges))
 319                                 /*nothing*/ ;
 320
 321                         /* we could be stepped out of the loop before the set iterated to the end */
 322                         set_break(edges);
 323
 324                         pset_insert_ptr(clique, e->n1);
 325                         pset_insert_ptr(clique, e->n2);
 326                         remove_edge(edges, e->n1, e->n2, &n_edges);
 327
 328                         /* while the clique is growing */
 329                         do {
 330                                 growed = 0;
 331
 332                                 /* search for a candidate to extend the clique */
 333                                 for (i=0; i<n_nodes; ++i) {
 334                                         ir_node *member, *cand = nodes[i];
 335                                         int is_cand;
 336
 337                                         /* if its already in the clique try the next */
 338                                         if (pset_find_ptr(clique, cand))
 339                                                 continue;
 340
 341                                         /* are there all necessary interferences? */
 342                                         is_cand = 1;
 343                                         pset_foreach(clique, member) {
 344                                                 if (!find_edge(edges, cand, member)) {
 345                                                         is_cand = 0;
 346                                                         pset_break(clique);
 347                                                         break;
 348                                                 }
 349                                         }
 350
 351                                         /* now we know if we have a clique extender */
 352                                         if (is_cand) {
 353                                                 /* first remove all covered edges */
 354                                                 pset_foreach(clique, member)
 355                                                         remove_edge(edges, cand, member, &n_edges);
 356
 357                                                 /* insert into clique */
 358                                                 pset_insert_ptr(clique, cand);
 359                                                 growed = 1;
 360                                                 break;
 361                                         }
 362                                 }
 363                         } while (growed);
 364
 365                         /* now the clique is maximal. Finally add the constraint */
 366                         {
 367                                 ir_node *member;
 368                                 int var_idx, cst_idx, center_nr, member_nr;
 369                                 char buf[16];
 370
 371                                 cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, pset_count(clique)-1);
 372                                 center_nr = get_irn_node_nr(center);
 373
 374                                 pset_foreach(clique, member) {
 375                                         member_nr = get_irn_node_nr(member);
 376                                         var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', center_nr, member_nr));
 377                                         lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 378                                 }
 379                         }
 380
 381                         del_pset(clique);
 382                 }
 383
 384                 del_set(edges);
 385                 obstack_free(&ob, NULL);
 386         }
 387 }
 388
 389
 390 static void extend_path(ilp_env_t *ienv, pdeq *path, ir_node *irn) {
 391         be_ifg_t *ifg = ienv->co->cenv->ifg;
 392         int i, len;
 393         ir_node **curr_path;
 394         affinity_node_t *aff;
 395         neighb_t *nbr;
 396
 397         /* do not walk backwards or in circles */
 398         if (pdeq_contains(path, irn))
 399                 return;
 400
 401         /* insert the new irn */
 402         pdeq_putr(path, irn);
 403
 404
 405
 406         /* check for forbidden interferences */
 407         len = pdeq_len(path);
 408         curr_path = alloca(len * sizeof(*curr_path));
 409         pdeq_copyl(path, (const void **)curr_path);
 410
 411         for (i=1; i<len; ++i)
 412                 if (be_ifg_connected(ifg, irn, curr_path[i]))
 413                         goto end;
 414
 415
 416
 417         /* check for terminating interference */
 418         if (be_ifg_connected(ifg, irn, curr_path[0])) {
 419
 420                 /* One node is not a path. */
 421                 /* And a path of length 2 is covered by a clique star constraint. */
 422                 if (len > 2) {
 423                         /* finally build the constraint */
 424                         int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, 1.0);
 425                         for (i=1; i<len; ++i) {
 426                                 char buf[16];
 427                                 int nr_1    = get_irn_node_nr(curr_path[i-1]);
 428                                 int nr_2    = get_irn_node_nr(curr_path[i]);
 429                                 int var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', nr_1, nr_2));
 430                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 431                         }
 432                 }
 433
 434                 /* this path cannot be extended anymore */
 435                 goto end;
 436         }
 437
 438
 439
 440         /* recursively extend the path */
 441         aff = get_affinity_info(ienv->co, irn);
 442         co_gs_foreach_neighb(aff, nbr)
 443                 extend_path(ienv, path, nbr->irn);
 444
 445
 446 end:
 447         /* remove the irn */
 448         pdeq_getr(path);
 449
 450 }
 451
 452 /**
 453  *  Search a path of affinity edges, whose ends are connected
 454  *  by an interference edge and there are no other interference
 455  *  edges in between.
 456  *  Then at least one of these affinity edges must break.
 457  */
 458 static void build_path_cstr(ilp_env_t *ienv) {
 459         affinity_node_t *aff_info;
 460
 461         /* for each node with affinity edges */
 462         co_gs_foreach_aff_node(ienv->co, aff_info) {
 463                 pdeq *path = new_pdeq();
 464
 465                 extend_path(ienv, path, aff_info->irn);
 466
 467                 del_pdeq(path);
 468         }
 469 }
 470
 471 static void ilp2_build(ilp_env_t *ienv) {
 472         local_env_t *lenv = ienv->env;
 473         int lower_bound;
 474
 475         ienv->lp = new_lpp(ienv->co->name, lpp_minimize);
 476         build_coloring_cstr(ienv);
 477         build_interference_cstr(ienv);
 478         build_affinity_cstr(ienv);
 479         build_clique_star_cstr(ienv);
 480         build_path_cstr(ienv);
 481
 482         lower_bound = co_get_lower_bound(ienv->co) - co_get_inevit_copy_costs(ienv->co);
 483         lpp_set_bound(ienv->lp, lower_bound);
 484         lpp_set_time_limit(ienv->lp, lenv->time_limit);
 485 }
 486
 487 static void ilp2_apply(ilp_env_t *ienv) {
 488         local_env_t *lenv = ienv->env;
 489         double *sol;
 490         lpp_sol_state_t state;
 491         int i, count;
 492
 493         /* first check if there was sth. to optimize */
 494         if (lenv->first_x_var >= 0) {
 495
 496                 count = lenv->last_x_var - lenv->first_x_var + 1;
 497                 sol = xmalloc(count * sizeof(sol[0]));
 498                 state = lpp_get_solution(ienv->lp, sol, lenv->first_x_var, lenv->last_x_var);
 499                 if (state != lpp_optimal) {
 500                         printf("WARNING %s: Solution state is not 'optimal': %d\n", ienv->co->name, state);
 501                         assert(state >= lpp_feasible && "The solution should at least be feasible!");
 502                 }
 503
 504                 for (i=0; i<count; ++i) {
 505                         int nodenr, color;
 506                         char var_name[16];
 507
 508                         if (sol[i] > 1-EPSILON) { /* split variable name into components */
 509                                 lpp_get_var_name(ienv->lp, lenv->first_x_var+i, var_name, sizeof(var_name));
 510
 511                                 if (sscanf(var_name, "x_%d_%d", &nodenr, &color) == 2) {
 512                                         ir_node *irn = pmap_get(lenv->nr_2_irn, INT_TO_PTR(nodenr));
 513                                         assert(irn && "This node number must be present in the map");
 514
 515                                         set_irn_col(ienv->co, irn, color);
 516                                 } else
 517                                         assert(0 && "This should be a x-var");
 518                         }
 519                 }
 520         }
 521
 522 #ifdef COPYOPT_STAT
 523         /* TODO adapt to multiple possible ILPs */
 524         copystat_add_ilp_time((int)(1000.0*lpp_get_sol_time(pi->curr_lp)));  //now we have ms
 525         copystat_add_ilp_vars(lpp_get_var_count(pi->curr_lp));
 526         copystat_add_ilp_csts(lpp_get_cst_count(pi->curr_lp));
 527         copystat_add_ilp_iter(lpp_get_iter_cnt(pi->curr_lp));
 528 #endif
 529 }
 530
 531 int co_solve_ilp2(copy_opt_t *co) {
 532         lpp_sol_state_t sol_state;
 533         ilp_env_t *ienv;
 534         local_env_t my;
 535
 536         ASSERT_OU_AVAIL(co); //See build_clique_st
 537         ASSERT_GS_AVAIL(co);
 538
 539         my.time_limit  = 0;
 540         my.first_x_var = -1;
 541         my.last_x_var  = -1;
 542         my.nr_2_irn    = pmap_create();
 543         FIRM_DBG_REGISTER(my.dbg, "firm.be.coilp2");
 544
 545         ienv = new_ilp_env(co, ilp2_build, ilp2_apply, &my);
 546
 547         sol_state = ilp_go(ienv);
 548
 549         pmap_destroy(my.nr_2_irn);
 550         free_ilp_env(ienv);
 551
 552         return sol_state == lpp_optimal;
 553 }
 554
 555 #else /* WITH_ILP */
 556
 557 static INLINE void only_that_you_can_compile_without_WITH_ILP_defined(void) {
 558 }
 559
 560 #endif /* WITH_ILP */