nsz Git - libfirm/blob - ir/be/becopyilp2.c

   1 /*
   2  * Copyright (C) 1995-2008 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 #include "firm_config.h"
  52
  53 #ifdef WITH_ILP
  54
  55 #include "bitset.h"
  56 #include "raw_bitset.h"
  57 #include "pdeq.h"
  58
  59 #include "irtools.h"
  60 #include "irgwalk.h"
  61 #include "becopyilp_t.h"
  62 #include "beifg_t.h"
  63 #include "besched_t.h"
  64
  65 #define DEBUG_LVL 1
  66
  67 typedef struct _local_env_t {
  68         double time_limit;
  69         int first_x_var, last_x_var;
  70         int n_colors;
  71         bitset_t *normal_colors;
  72         pmap *nr_2_irn;
  73         DEBUG_ONLY(firm_dbg_module_t *dbg;)
  74 } local_env_t;
  75
  76 static void build_coloring_cstr(ilp_env_t *ienv) {
  77         be_ifg_t *ifg     = ienv->co->cenv->ifg;
  78         void *iter        = be_ifg_nodes_iter_alloca(ifg);
  79         bitset_t *colors;
  80         ir_node *irn;
  81         char buf[16];
  82
  83         colors = bitset_alloca(arch_register_class_n_regs(ienv->co->cls));
  84
  85         be_ifg_foreach_node(ifg, iter, irn)
  86                 if (!sr_is_removed(ienv->sr, irn)) {
  87                         bitset_pos_t col;
  88                         int cst_idx;
  89                         const arch_register_req_t *req;
  90                         int curr_node_color = get_irn_col(irn);
  91                         int node_nr = (int)get_irn_idx(irn);
  92                         local_env_t *lenv = ienv->env;
  93
  94                         pmap_insert(lenv->nr_2_irn, INT_TO_PTR(node_nr), irn);
  95
  96                         req = arch_get_register_req(irn, -1);
  97
  98                         bitset_clear_all(colors);
  99
 100                         /* get assignable colors */
 101                         if (arch_register_req_is(req, limited)) {
 102                                 rbitset_copy_to_bitset(req->limited, colors);
 103                         } else {
 104                                 bitset_copy(colors, lenv->normal_colors);
 105                         }
 106
 107                         /* add the coloring constraint */
 108                         cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 1.0);
 109
 110                         bitset_foreach(colors, col) {
 111                                 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, col), lpp_binary, 0.0);
 112                                 lpp_set_start_value(ienv->lp, var_idx, (col == (unsigned) curr_node_color) ? 1.0 : 0.0);
 113                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
 114
 115                                 lenv->last_x_var = var_idx;
 116                                 if (lenv->first_x_var == -1)
 117                                         lenv->first_x_var = var_idx;
 118                         }
 119
 120                         /* add register constraint constraints */
 121                         bitset_foreach_clear(colors, col) {
 122                                 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 0.0);
 123                                 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, col), lpp_binary, 0.0);
 124                                 lpp_set_start_value(ienv->lp, var_idx, 0.0);
 125                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
 126
 127                                 lenv->last_x_var = var_idx;
 128                         }
 129                 }
 130 }
 131
 132 static void build_interference_cstr(ilp_env_t *ienv) {
 133         lpp_t *lpp        = ienv->lp;
 134         local_env_t *lenv = ienv->env;
 135         be_ifg_t *ifg     = ienv->co->cenv->ifg;
 136         int n_colors      = lenv->n_colors;
 137         int i, col;
 138
 139         void *iter = be_ifg_cliques_iter_alloca(ifg);
 140         ir_node **clique = alloca(sizeof(*clique) * n_colors);
 141         int size;
 142
 143         char buf[16];
 144
 145         /* for each maximal clique */
 146         be_ifg_foreach_clique(ifg, iter, clique, &size) {
 147                 int realsize = 0;
 148
 149                 for (i=0; i<size; ++i)
 150                         if (!sr_is_removed(ienv->sr, clique[i]))
 151                                 ++realsize;
 152
 153                 if (realsize < 2)
 154                         continue;
 155
 156                 /* for all colors */
 157                 for (col=0; col<n_colors; ++col) {
 158                         int cst_idx = lpp_add_cst(lpp, NULL, lpp_less, 1.0);
 159
 160                         /* for each member of this clique */
 161                         for (i=0; i<size; ++i) {
 162                                 ir_node *irn = clique[i];
 163
 164                                 if (!sr_is_removed(ienv->sr, irn)) {
 165                                         int var_idx = lpp_get_var_idx(lpp, name_cdd(buf, 'x', (int)get_irn_idx(irn), col));
 166                                         lpp_set_factor_fast(lpp, cst_idx, var_idx, 1);
 167                                 }
 168                         }
 169                 }
 170         }
 171 }
 172
 173
 174 /**
 175  * TODO: Remove the dependency of the opt-units data structure
 176  *       by walking over all affinity edges. Graph structure
 177  *       does not provide this walker, yet.
 178  */
 179 static void build_affinity_cstr(ilp_env_t *ienv) {
 180         local_env_t *lenv = ienv->env;
 181         int n_colors      = lenv->n_colors;
 182         unit_t *curr;
 183
 184         /* for all optimization units */
 185         list_for_each_entry(unit_t, curr, &ienv->co->units, units) {
 186                 ir_node *root, *arg;
 187                 int root_nr, arg_nr, i, col, y_idx, root_idx, arg_idx;
 188                 char buf[16];
 189                 int root_col, arg_col;
 190
 191                 root = curr->nodes[0];
 192                 root_nr = (int) get_irn_idx(root);
 193                 root_col = get_irn_col(root);
 194
 195                 for (i = 1; i < curr->node_count; ++i) {
 196                         arg = curr->nodes[i];
 197                         arg_nr = (int) get_irn_idx(arg);
 198                         arg_col = get_irn_col(arg);
 199
 200                         /* add a new affinity variable */
 201                         y_idx = lpp_add_var(ienv->lp, name_cdd_sorted(buf, 'y', root_nr, arg_nr), lpp_binary, curr->costs[i]);
 202                         lpp_set_start_value(ienv->lp, y_idx, (root_col==arg_col) ? 0.0 : 1.0);
 203
 204                         /* add constraints relating the affinity var to the color vars */
 205                         for (col=0; col<n_colors; ++col) {
 206                                 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_less, 0.0);
 207                                 root_idx = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', root_nr, col));
 208                                 arg_idx  = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', arg_nr,  col));
 209
 210                                 lpp_set_factor_fast(ienv->lp, cst_idx, root_idx,  1.0);
 211                                 lpp_set_factor_fast(ienv->lp, cst_idx, arg_idx,  -1.0);
 212                                 lpp_set_factor_fast(ienv->lp, cst_idx, y_idx, -1.0);
 213                         }
 214                 }
 215         }
 216 }
 217
 218 /**
 219  * Helping stuff for build_clique_star_cstr
 220  */
 221 typedef struct _edge_t {
 222         ir_node *n1, *n2;
 223 } edge_t;
 224
 225 static int compare_edge_t(const void *k1, const void *k2, size_t size) {
 226         const edge_t *e1 = k1;
 227         const edge_t *e2 = k2;
 228         (void) size;
 229
 230         return ! (e1->n1 == e2->n1   &&   e1->n2 == e2->n2);
 231 }
 232
 233 #define HASH_EDGE(e) (hash_irn((e)->n1) ^ hash_irn((e)->n2))
 234
 235 static INLINE edge_t *add_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
 236         edge_t new_edge;
 237
 238         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 239                 new_edge.n1 = n1;
 240                 new_edge.n2 = n2;
 241         } else {
 242                 new_edge.n1 = n2;
 243                 new_edge.n2 = n1;
 244         }
 245         (*counter)++;
 246         return set_insert(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 247 }
 248
 249 static INLINE edge_t *find_edge(set *edges, ir_node *n1, ir_node *n2) {
 250         edge_t new_edge;
 251
 252         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 253                 new_edge.n1 = n1;
 254                 new_edge.n2 = n2;
 255         } else {
 256                 new_edge.n1 = n2;
 257                 new_edge.n2 = n1;
 258         }
 259         return set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 260 }
 261
 262 static INLINE void remove_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
 263         edge_t new_edge, *e;
 264
 265         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 266                 new_edge.n1 = n1;
 267                 new_edge.n2 = n2;
 268         } else {
 269                 new_edge.n1 = n2;
 270                 new_edge.n2 = n1;
 271         }
 272         e = set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 273         if (e) {
 274                 e->n1 = NULL;
 275                 e->n2 = NULL;
 276                 (*counter)--;
 277         }
 278 }
 279
 280 #define pset_foreach(pset, irn)  for(irn=pset_first(pset); irn; irn=pset_next(pset))
 281
 282 /**
 283  * Search for an interference clique and an external node
 284  * with affinity edges to all nodes of the clique.
 285  * At most 1 node of the clique can be colored equally with the external node.
 286  */
 287 static void build_clique_star_cstr(ilp_env_t *ienv) {
 288         affinity_node_t *aff;
 289
 290         /* for each node with affinity edges */
 291         co_gs_foreach_aff_node(ienv->co, aff) {
 292                 struct obstack ob;
 293                 neighb_t *nbr;
 294                 const ir_node *center = aff->irn;
 295                 ir_node **nodes;
 296                 set *edges;
 297                 int i, o, n_nodes, n_edges;
 298
 299                 if (arch_irn_is(ienv->co->aenv, aff->irn, ignore))
 300                         continue;
 301
 302                 obstack_init(&ob);
 303                 edges = new_set(compare_edge_t, 8);
 304
 305                 /* get all affinity neighbours */
 306                 n_nodes = 0;
 307                 co_gs_foreach_neighb(aff, nbr) {
 308                         if (!arch_irn_is(ienv->co->aenv, nbr->irn, ignore)) {
 309                                 obstack_ptr_grow(&ob, nbr->irn);
 310                                 ++n_nodes;
 311                         }
 312                 }
 313                 nodes = obstack_finish(&ob);
 314
 315                 /* get all interference edges between these */
 316                 n_edges = 0;
 317                 for (i=0; i<n_nodes; ++i)
 318                         for (o=0; o<i; ++o)
 319                                 if (be_ifg_connected(ienv->co->cenv->ifg, nodes[i], nodes[o]))
 320                                         add_edge(edges, nodes[i], nodes[o], &n_edges);
 321
 322                 /* cover all these interference edges with maximal cliques */
 323                 while (n_edges) {
 324                         edge_t *e;
 325                         pset *clique = pset_new_ptr(8);
 326                         int growed;
 327
 328                         /* get 2 starting nodes to form a clique */
 329                         for (e=set_first(edges); !e->n1; e=set_next(edges))
 330                                 /*nothing*/ ;
 331
 332                         /* we could be stepped out of the loop before the set iterated to the end */
 333                         set_break(edges);
 334
 335                         pset_insert_ptr(clique, e->n1);
 336                         pset_insert_ptr(clique, e->n2);
 337                         remove_edge(edges, e->n1, e->n2, &n_edges);
 338
 339                         /* while the clique is growing */
 340                         do {
 341                                 growed = 0;
 342
 343                                 /* search for a candidate to extend the clique */
 344                                 for (i=0; i<n_nodes; ++i) {
 345                                         ir_node *member, *cand = nodes[i];
 346                                         int is_cand;
 347
 348                                         /* if its already in the clique try the next */
 349                                         if (pset_find_ptr(clique, cand))
 350                                                 continue;
 351
 352                                         /* are there all necessary interferences? */
 353                                         is_cand = 1;
 354                                         pset_foreach(clique, member) {
 355                                                 if (!find_edge(edges, cand, member)) {
 356                                                         is_cand = 0;
 357                                                         pset_break(clique);
 358                                                         break;
 359                                                 }
 360                                         }
 361
 362                                         /* now we know if we have a clique extender */
 363                                         if (is_cand) {
 364                                                 /* first remove all covered edges */
 365                                                 pset_foreach(clique, member)
 366                                                         remove_edge(edges, cand, member, &n_edges);
 367
 368                                                 /* insert into clique */
 369                                                 pset_insert_ptr(clique, cand);
 370                                                 growed = 1;
 371                                                 break;
 372                                         }
 373                                 }
 374                         } while (growed);
 375
 376                         /* now the clique is maximal. Finally add the constraint */
 377                         {
 378                                 ir_node *member;
 379                                 int var_idx, cst_idx, center_nr, member_nr;
 380                                 char buf[16];
 381
 382                                 cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, pset_count(clique)-1);
 383                                 center_nr = get_irn_idx(center);
 384
 385                                 pset_foreach(clique, member) {
 386                                         member_nr = get_irn_idx(member);
 387                                         var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', center_nr, member_nr));
 388                                         lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 389                                 }
 390                         }
 391
 392                         del_pset(clique);
 393                 }
 394
 395                 del_set(edges);
 396                 obstack_free(&ob, NULL);
 397         }
 398 }
 399
 400
 401 static void extend_path(ilp_env_t *ienv, pdeq *path, const ir_node *irn) {
 402         be_ifg_t *ifg = ienv->co->cenv->ifg;
 403         int i, len;
 404         ir_node **curr_path;
 405         affinity_node_t *aff;
 406         neighb_t *nbr;
 407
 408         /* do not walk backwards or in circles */
 409         if (pdeq_contains(path, irn))
 410                 return;
 411
 412         if (arch_irn_is(ienv->co->aenv, irn, ignore))
 413                 return;
 414
 415         /* insert the new irn */
 416         pdeq_putr(path, irn);
 417
 418
 419
 420         /* check for forbidden interferences */
 421         len = pdeq_len(path);
 422         curr_path = alloca(len * sizeof(*curr_path));
 423         pdeq_copyl(path, (const void **)curr_path);
 424
 425         for (i=1; i<len; ++i)
 426                 if (be_ifg_connected(ifg, irn, curr_path[i]))
 427                         goto end;
 428
 429
 430
 431         /* check for terminating interference */
 432         if (be_ifg_connected(ifg, irn, curr_path[0])) {
 433
 434                 /* One node is not a path. */
 435                 /* And a path of length 2 is covered by a clique star constraint. */
 436                 if (len > 2) {
 437                         /* finally build the constraint */
 438                         int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, 1.0);
 439                         for (i=1; i<len; ++i) {
 440                                 char buf[16];
 441                                 int nr_1    = get_irn_idx(curr_path[i-1]);
 442                                 int nr_2    = get_irn_idx(curr_path[i]);
 443                                 int var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', nr_1, nr_2));
 444                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 445                         }
 446                 }
 447
 448                 /* this path cannot be extended anymore */
 449                 goto end;
 450         }
 451
 452
 453
 454         /* recursively extend the path */
 455         aff = get_affinity_info(ienv->co, irn);
 456         co_gs_foreach_neighb(aff, nbr)
 457                 extend_path(ienv, path, nbr->irn);
 458
 459
 460 end:
 461         /* remove the irn */
 462         pdeq_getr(path);
 463
 464 }
 465
 466 /**
 467  *  Search a path of affinity edges, whose ends are connected
 468  *  by an interference edge and there are no other interference
 469  *  edges in between.
 470  *  Then at least one of these affinity edges must break.
 471  */
 472 static void build_path_cstr(ilp_env_t *ienv) {
 473         affinity_node_t *aff_info;
 474
 475         /* for each node with affinity edges */
 476         co_gs_foreach_aff_node(ienv->co, aff_info) {
 477                 pdeq *path = new_pdeq();
 478
 479                 extend_path(ienv, path, aff_info->irn);
 480
 481                 del_pdeq(path);
 482         }
 483 }
 484
 485 static void ilp2_build(ilp_env_t *ienv) {
 486         local_env_t *lenv = ienv->env;
 487         int lower_bound;
 488
 489         ienv->lp = new_lpp(ienv->co->name, lpp_minimize);
 490         build_coloring_cstr(ienv);
 491         build_interference_cstr(ienv);
 492         build_affinity_cstr(ienv);
 493         build_clique_star_cstr(ienv);
 494         build_path_cstr(ienv);
 495
 496         lower_bound = co_get_lower_bound(ienv->co) - co_get_inevit_copy_costs(ienv->co);
 497         lpp_set_bound(ienv->lp, lower_bound);
 498         lpp_set_time_limit(ienv->lp, lenv->time_limit);
 499 }
 500
 501 static void ilp2_apply(ilp_env_t *ienv) {
 502         local_env_t *lenv = ienv->env;
 503         int i;
 504
 505         /* first check if there was sth. to optimize */
 506         if (lenv->first_x_var >= 0) {
 507                 int              count = lenv->last_x_var - lenv->first_x_var + 1;
 508                 double          *sol   = XMALLOCN(double, count);
 509                 lpp_sol_state_t  state = lpp_get_solution(ienv->lp, sol, lenv->first_x_var, lenv->last_x_var);
 510
 511                 if (state != lpp_optimal) {
 512                         printf("WARNING %s: Solution state is not 'optimal': %d\n", ienv->co->name, state);
 513                         assert(state >= lpp_feasible && "The solution should at least be feasible!");
 514                 }
 515
 516                 for (i=0; i<count; ++i) {
 517                         int nodenr, color;
 518                         char var_name[16];
 519
 520                         if (sol[i] > 1-EPSILON) { /* split variable name into components */
 521                                 lpp_get_var_name(ienv->lp, lenv->first_x_var+i, var_name, sizeof(var_name));
 522
 523                                 if (sscanf(var_name, "x_%d_%d", &nodenr, &color) == 2) {
 524                                         ir_node *irn = pmap_get(lenv->nr_2_irn, INT_TO_PTR(nodenr));
 525                                         assert(irn && "This node number must be present in the map");
 526
 527                                         set_irn_col(ienv->co, irn, color);
 528                                 } else
 529                                         assert(0 && "This should be a x-var");
 530                         }
 531                 }
 532
 533                 xfree(sol);
 534         }
 535
 536 #ifdef COPYOPT_STAT
 537         /* TODO adapt to multiple possible ILPs */
 538         copystat_add_ilp_time((int)(1000.0*lpp_get_sol_time(pi->curr_lp)));  //now we have ms
 539         copystat_add_ilp_vars(lpp_get_var_count(pi->curr_lp));
 540         copystat_add_ilp_csts(lpp_get_cst_count(pi->curr_lp));
 541         copystat_add_ilp_iter(lpp_get_iter_cnt(pi->curr_lp));
 542 #endif
 543 }
 544
 545 int co_solve_ilp2(copy_opt_t *co) {
 546         lpp_sol_state_t sol_state;
 547         ilp_env_t *ienv;
 548         local_env_t my;
 549
 550         ASSERT_OU_AVAIL(co); //See build_clique_st
 551         ASSERT_GS_AVAIL(co);
 552
 553         my.time_limit  = 0;
 554         my.first_x_var = -1;
 555         my.last_x_var  = -1;
 556         my.nr_2_irn    = pmap_create();
 557         FIRM_DBG_REGISTER(my.dbg, "firm.be.coilp2");
 558
 559         my.normal_colors = bitset_alloca(arch_register_class_n_regs(co->cls));
 560         bitset_clear_all(my.normal_colors);
 561         arch_put_non_ignore_regs(co->cls, my.normal_colors);
 562         my.n_colors = bitset_popcnt(my.normal_colors);
 563
 564         ienv = new_ilp_env(co, ilp2_build, ilp2_apply, &my);
 565
 566         sol_state = ilp_go(ienv);
 567
 568         pmap_destroy(my.nr_2_irn);
 569         free_ilp_env(ienv);
 570
 571         return sol_state == lpp_optimal;
 572 }
 573
 574 #else /* WITH_ILP */
 575
 576 static INLINE void only_that_you_can_compile_without_WITH_ILP_defined(void) {
 577 }
 578
 579 #endif /* WITH_ILP */