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 #include "config.h"
  48
  49 #ifdef WITH_ILP
  50
  51 #include "bitset.h"
  52 #include "raw_bitset.h"
  53 #include "pdeq.h"
  54
  55 #include "irtools.h"
  56 #include "irgwalk.h"
  57 #include "becopyilp_t.h"
  58 #include "beifg_t.h"
  59 #include "besched_t.h"
  60
  61 #define DEBUG_LVL 1
  62
  63 typedef struct _local_env_t {
  64         double time_limit;
  65         int first_x_var, last_x_var;
  66         int n_colors;
  67         bitset_t *normal_colors;
  68         pmap *nr_2_irn;
  69         DEBUG_ONLY(firm_dbg_module_t *dbg;)
  70 } local_env_t;
  71
  72 static void build_coloring_cstr(ilp_env_t *ienv) {
  73         be_ifg_t *ifg     = ienv->co->cenv->ifg;
  74         void *iter        = be_ifg_nodes_iter_alloca(ifg);
  75         bitset_t *colors;
  76         ir_node *irn;
  77         char buf[16];
  78
  79         colors = bitset_alloca(arch_register_class_n_regs(ienv->co->cls));
  80
  81         be_ifg_foreach_node(ifg, iter, irn)
  82                 if (!sr_is_removed(ienv->sr, irn)) {
  83                         bitset_pos_t col;
  84                         int cst_idx;
  85                         const arch_register_req_t *req;
  86                         int curr_node_color = get_irn_col(irn);
  87                         int node_nr = (int)get_irn_idx(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_out(irn);
  93
  94                         bitset_clear_all(colors);
  95
  96                         /* get assignable colors */
  97                         if (arch_register_req_is(req, limited)) {
  98                                 rbitset_copy_to_bitset(req->limited, colors);
  99                         } else {
 100                                 bitset_copy(colors, lenv->normal_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 {
 130         lpp_t       *lpp      = ienv->lp;
 131         local_env_t *lenv     = ienv->env;
 132         be_ifg_t    *ifg      = ienv->co->cenv->ifg;
 133         int          n_colors = lenv->n_colors;
 134         void        *iter     = be_ifg_cliques_iter_alloca(ifg);
 135         ir_node    **clique   = ALLOCAN(ir_node*, n_colors);
 136         int          size;
 137         int          col;
 138         int          i;
 139
 140         char buf[16];
 141
 142         /* for each maximal clique */
 143         be_ifg_foreach_clique(ifg, iter, clique, &size) {
 144                 int realsize = 0;
 145
 146                 for (i=0; i<size; ++i)
 147                         if (!sr_is_removed(ienv->sr, clique[i]))
 148                                 ++realsize;
 149
 150                 if (realsize < 2)
 151                         continue;
 152
 153                 /* for all colors */
 154                 for (col=0; col<n_colors; ++col) {
 155                         int cst_idx = lpp_add_cst(lpp, NULL, lpp_less, 1.0);
 156
 157                         /* for each member of this clique */
 158                         for (i=0; i<size; ++i) {
 159                                 ir_node *irn = clique[i];
 160
 161                                 if (!sr_is_removed(ienv->sr, irn)) {
 162                                         int var_idx = lpp_get_var_idx(lpp, name_cdd(buf, 'x', (int)get_irn_idx(irn), col));
 163                                         lpp_set_factor_fast(lpp, cst_idx, var_idx, 1);
 164                                 }
 165                         }
 166                 }
 167         }
 168 }
 169
 170
 171 /**
 172  * TODO: Remove the dependency of the opt-units data structure
 173  *       by walking over all affinity edges. Graph structure
 174  *       does not provide this walker, yet.
 175  */
 176 static void build_affinity_cstr(ilp_env_t *ienv) {
 177         local_env_t *lenv = ienv->env;
 178         int n_colors      = lenv->n_colors;
 179         unit_t *curr;
 180
 181         /* for all optimization units */
 182         list_for_each_entry(unit_t, curr, &ienv->co->units, units) {
 183                 ir_node *root, *arg;
 184                 int root_nr, arg_nr, i, col, y_idx, root_idx, arg_idx;
 185                 char buf[16];
 186                 int root_col, arg_col;
 187
 188                 root = curr->nodes[0];
 189                 root_nr = (int) get_irn_idx(root);
 190                 root_col = get_irn_col(root);
 191
 192                 for (i = 1; i < curr->node_count; ++i) {
 193                         arg = curr->nodes[i];
 194                         arg_nr = (int) get_irn_idx(arg);
 195                         arg_col = get_irn_col(arg);
 196
 197                         /* add a new affinity variable */
 198                         y_idx = lpp_add_var(ienv->lp, name_cdd_sorted(buf, 'y', root_nr, arg_nr), lpp_binary, curr->costs[i]);
 199                         lpp_set_start_value(ienv->lp, y_idx, (root_col==arg_col) ? 0.0 : 1.0);
 200
 201                         /* add constraints relating the affinity var to the color vars */
 202                         for (col=0; col<n_colors; ++col) {
 203                                 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_less, 0.0);
 204                                 root_idx = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', root_nr, col));
 205                                 arg_idx  = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', arg_nr,  col));
 206
 207                                 lpp_set_factor_fast(ienv->lp, cst_idx, root_idx,  1.0);
 208                                 lpp_set_factor_fast(ienv->lp, cst_idx, arg_idx,  -1.0);
 209                                 lpp_set_factor_fast(ienv->lp, cst_idx, y_idx, -1.0);
 210                         }
 211                 }
 212         }
 213 }
 214
 215 /**
 216  * Helping stuff for build_clique_star_cstr
 217  */
 218 typedef struct _edge_t {
 219         ir_node *n1, *n2;
 220 } edge_t;
 221
 222 static int compare_edge_t(const void *k1, const void *k2, size_t size) {
 223         const edge_t *e1 = k1;
 224         const edge_t *e2 = k2;
 225         (void) size;
 226
 227         return ! (e1->n1 == e2->n1   &&   e1->n2 == e2->n2);
 228 }
 229
 230 #define HASH_EDGE(e) (hash_irn((e)->n1) ^ hash_irn((e)->n2))
 231
 232 static inline edge_t *add_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
 233         edge_t new_edge;
 234
 235         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 236                 new_edge.n1 = n1;
 237                 new_edge.n2 = n2;
 238         } else {
 239                 new_edge.n1 = n2;
 240                 new_edge.n2 = n1;
 241         }
 242         (*counter)++;
 243         return set_insert(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 244 }
 245
 246 static inline edge_t *find_edge(set *edges, ir_node *n1, ir_node *n2) {
 247         edge_t new_edge;
 248
 249         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 250                 new_edge.n1 = n1;
 251                 new_edge.n2 = n2;
 252         } else {
 253                 new_edge.n1 = n2;
 254                 new_edge.n2 = n1;
 255         }
 256         return set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 257 }
 258
 259 static inline void remove_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
 260         edge_t new_edge, *e;
 261
 262         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 263                 new_edge.n1 = n1;
 264                 new_edge.n2 = n2;
 265         } else {
 266                 new_edge.n1 = n2;
 267                 new_edge.n2 = n1;
 268         }
 269         e = set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 270         if (e) {
 271                 e->n1 = NULL;
 272                 e->n2 = NULL;
 273                 (*counter)--;
 274         }
 275 }
 276
 277 #define pset_foreach(pset, irn)  for(irn=pset_first(pset); irn; irn=pset_next(pset))
 278
 279 /**
 280  * Search for an interference clique and an external node
 281  * with affinity edges to all nodes of the clique.
 282  * At most 1 node of the clique can be colored equally with the external node.
 283  */
 284 static void build_clique_star_cstr(ilp_env_t *ienv) {
 285         affinity_node_t *aff;
 286
 287         /* for each node with affinity edges */
 288         co_gs_foreach_aff_node(ienv->co, aff) {
 289                 struct obstack ob;
 290                 neighb_t *nbr;
 291                 const ir_node *center = aff->irn;
 292                 ir_node **nodes;
 293                 set *edges;
 294                 int i, o, n_nodes, n_edges;
 295
 296                 if (arch_irn_is_ignore(aff->irn))
 297                         continue;
 298
 299                 obstack_init(&ob);
 300                 edges = new_set(compare_edge_t, 8);
 301
 302                 /* get all affinity neighbours */
 303                 n_nodes = 0;
 304                 co_gs_foreach_neighb(aff, nbr) {
 305                         if (!arch_irn_is_ignore(nbr->irn)) {
 306                                 obstack_ptr_grow(&ob, nbr->irn);
 307                                 ++n_nodes;
 308                         }
 309                 }
 310                 nodes = obstack_finish(&ob);
 311
 312                 /* get all interference edges between these */
 313                 n_edges = 0;
 314                 for (i=0; i<n_nodes; ++i)
 315                         for (o=0; o<i; ++o)
 316                                 if (be_ifg_connected(ienv->co->cenv->ifg, nodes[i], nodes[o]))
 317                                         add_edge(edges, nodes[i], nodes[o], &n_edges);
 318
 319                 /* cover all these interference edges with maximal cliques */
 320                 while (n_edges) {
 321                         edge_t *e;
 322                         pset *clique = pset_new_ptr(8);
 323                         int growed;
 324
 325                         /* get 2 starting nodes to form a clique */
 326                         for (e=set_first(edges); !e->n1; e=set_next(edges))
 327                                 /*nothing*/ ;
 328
 329                         /* we could be stepped out of the loop before the set iterated to the end */
 330                         set_break(edges);
 331
 332                         pset_insert_ptr(clique, e->n1);
 333                         pset_insert_ptr(clique, e->n2);
 334                         remove_edge(edges, e->n1, e->n2, &n_edges);
 335
 336                         /* while the clique is growing */
 337                         do {
 338                                 growed = 0;
 339
 340                                 /* search for a candidate to extend the clique */
 341                                 for (i=0; i<n_nodes; ++i) {
 342                                         ir_node *member, *cand = nodes[i];
 343                                         int is_cand;
 344
 345                                         /* if its already in the clique try the next */
 346                                         if (pset_find_ptr(clique, cand))
 347                                                 continue;
 348
 349                                         /* are there all necessary interferences? */
 350                                         is_cand = 1;
 351                                         pset_foreach(clique, member) {
 352                                                 if (!find_edge(edges, cand, member)) {
 353                                                         is_cand = 0;
 354                                                         pset_break(clique);
 355                                                         break;
 356                                                 }
 357                                         }
 358
 359                                         /* now we know if we have a clique extender */
 360                                         if (is_cand) {
 361                                                 /* first remove all covered edges */
 362                                                 pset_foreach(clique, member)
 363                                                         remove_edge(edges, cand, member, &n_edges);
 364
 365                                                 /* insert into clique */
 366                                                 pset_insert_ptr(clique, cand);
 367                                                 growed = 1;
 368                                                 break;
 369                                         }
 370                                 }
 371                         } while (growed);
 372
 373                         /* now the clique is maximal. Finally add the constraint */
 374                         {
 375                                 ir_node *member;
 376                                 int var_idx, cst_idx, center_nr, member_nr;
 377                                 char buf[16];
 378
 379                                 cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, pset_count(clique)-1);
 380                                 center_nr = get_irn_idx(center);
 381
 382                                 pset_foreach(clique, member) {
 383                                         member_nr = get_irn_idx(member);
 384                                         var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', center_nr, member_nr));
 385                                         lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 386                                 }
 387                         }
 388
 389                         del_pset(clique);
 390                 }
 391
 392                 del_set(edges);
 393                 obstack_free(&ob, NULL);
 394         }
 395 }
 396
 397
 398 static void extend_path(ilp_env_t *ienv, pdeq *path, const ir_node *irn) {
 399         be_ifg_t *ifg = ienv->co->cenv->ifg;
 400         int i, len;
 401         ir_node **curr_path;
 402         affinity_node_t *aff;
 403         neighb_t *nbr;
 404
 405         /* do not walk backwards or in circles */
 406         if (pdeq_contains(path, irn))
 407                 return;
 408
 409         if (arch_irn_is_ignore(irn))
 410                 return;
 411
 412         /* insert the new irn */
 413         pdeq_putr(path, irn);
 414
 415
 416
 417         /* check for forbidden interferences */
 418         len       = pdeq_len(path);
 419         curr_path = ALLOCAN(ir_node*, len);
 420         pdeq_copyl(path, (const void **)curr_path);
 421
 422         for (i=1; i<len; ++i)
 423                 if (be_ifg_connected(ifg, irn, curr_path[i]))
 424                         goto end;
 425
 426
 427
 428         /* check for terminating interference */
 429         if (be_ifg_connected(ifg, irn, curr_path[0])) {
 430
 431                 /* One node is not a path. */
 432                 /* And a path of length 2 is covered by a clique star constraint. */
 433                 if (len > 2) {
 434                         /* finally build the constraint */
 435                         int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, 1.0);
 436                         for (i=1; i<len; ++i) {
 437                                 char buf[16];
 438                                 int nr_1    = get_irn_idx(curr_path[i-1]);
 439                                 int nr_2    = get_irn_idx(curr_path[i]);
 440                                 int var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', nr_1, nr_2));
 441                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 442                         }
 443                 }
 444
 445                 /* this path cannot be extended anymore */
 446                 goto end;
 447         }
 448
 449
 450
 451         /* recursively extend the path */
 452         aff = get_affinity_info(ienv->co, irn);
 453         co_gs_foreach_neighb(aff, nbr)
 454                 extend_path(ienv, path, nbr->irn);
 455
 456
 457 end:
 458         /* remove the irn */
 459         pdeq_getr(path);
 460
 461 }
 462
 463 /**
 464  *  Search a path of affinity edges, whose ends are connected
 465  *  by an interference edge and there are no other interference
 466  *  edges in between.
 467  *  Then at least one of these affinity edges must break.
 468  */
 469 static void build_path_cstr(ilp_env_t *ienv) {
 470         affinity_node_t *aff_info;
 471
 472         /* for each node with affinity edges */
 473         co_gs_foreach_aff_node(ienv->co, aff_info) {
 474                 pdeq *path = new_pdeq();
 475
 476                 extend_path(ienv, path, aff_info->irn);
 477
 478                 del_pdeq(path);
 479         }
 480 }
 481
 482 static void ilp2_build(ilp_env_t *ienv) {
 483         local_env_t *lenv = ienv->env;
 484         int lower_bound;
 485
 486         ienv->lp = new_lpp(ienv->co->name, lpp_minimize);
 487         build_coloring_cstr(ienv);
 488         build_interference_cstr(ienv);
 489         build_affinity_cstr(ienv);
 490         build_clique_star_cstr(ienv);
 491         build_path_cstr(ienv);
 492
 493         lower_bound = co_get_lower_bound(ienv->co) - co_get_inevit_copy_costs(ienv->co);
 494         lpp_set_bound(ienv->lp, lower_bound);
 495         lpp_set_time_limit(ienv->lp, lenv->time_limit);
 496 }
 497
 498 static void ilp2_apply(ilp_env_t *ienv) {
 499         local_env_t *lenv = ienv->env;
 500         int i;
 501
 502         /* first check if there was sth. to optimize */
 503         if (lenv->first_x_var >= 0) {
 504                 int              count = lenv->last_x_var - lenv->first_x_var + 1;
 505                 double          *sol   = XMALLOCN(double, count);
 506                 lpp_sol_state_t  state = lpp_get_solution(ienv->lp, sol, lenv->first_x_var, lenv->last_x_var);
 507
 508                 if (state != lpp_optimal) {
 509                         printf("WARNING %s: Solution state is not 'optimal': %d\n", ienv->co->name, state);
 510                         assert(state >= lpp_feasible && "The solution should at least be feasible!");
 511                 }
 512
 513                 for (i=0; i<count; ++i) {
 514                         int nodenr, color;
 515                         char var_name[16];
 516
 517                         if (sol[i] > 1-EPSILON) { /* split variable name into components */
 518                                 lpp_get_var_name(ienv->lp, lenv->first_x_var+i, var_name, sizeof(var_name));
 519
 520                                 if (sscanf(var_name, "x_%d_%d", &nodenr, &color) == 2) {
 521                                         ir_node *irn = pmap_get(lenv->nr_2_irn, INT_TO_PTR(nodenr));
 522                                         assert(irn && "This node number must be present in the map");
 523
 524                                         set_irn_col(ienv->co, irn, color);
 525                                 } else
 526                                         assert(0 && "This should be a x-var");
 527                         }
 528                 }
 529
 530                 xfree(sol);
 531         }
 532
 533 #ifdef COPYOPT_STAT
 534         /* TODO adapt to multiple possible ILPs */
 535         copystat_add_ilp_time((int)(1000.0*lpp_get_sol_time(pi->curr_lp)));  //now we have ms
 536         copystat_add_ilp_vars(lpp_get_var_count(pi->curr_lp));
 537         copystat_add_ilp_csts(lpp_get_cst_count(pi->curr_lp));
 538         copystat_add_ilp_iter(lpp_get_iter_cnt(pi->curr_lp));
 539 #endif
 540 }
 541
 542 int co_solve_ilp2(copy_opt_t *co) {
 543         lpp_sol_state_t sol_state;
 544         ilp_env_t *ienv;
 545         local_env_t my;
 546
 547         ASSERT_OU_AVAIL(co); //See build_clique_st
 548         ASSERT_GS_AVAIL(co);
 549
 550         my.time_limit  = 0;
 551         my.first_x_var = -1;
 552         my.last_x_var  = -1;
 553         my.nr_2_irn    = pmap_create();
 554         FIRM_DBG_REGISTER(my.dbg, "firm.be.coilp2");
 555
 556         my.normal_colors = bitset_alloca(arch_register_class_n_regs(co->cls));
 557         bitset_clear_all(my.normal_colors);
 558         arch_put_non_ignore_regs(co->cls, my.normal_colors);
 559         my.n_colors = bitset_popcnt(my.normal_colors);
 560
 561         ienv = new_ilp_env(co, ilp2_build, ilp2_apply, &my);
 562
 563         sol_state = ilp_go(ienv);
 564
 565         pmap_destroy(my.nr_2_irn);
 566         free_ilp_env(ienv);
 567
 568         return sol_state == lpp_optimal;
 569 }
 570
 571 #else /* WITH_ILP */
 572
 573 static inline void only_that_you_can_compile_without_WITH_ILP_defined(void) {
 574 }
 575
 576 #endif /* WITH_ILP */