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.h"
  60 #include "bemodule.h"
  61
  62 #define DEBUG_LVL 1
  63
  64 typedef struct _local_env_t {
  65         double time_limit;
  66         int first_x_var, last_x_var;
  67         int n_colors;
  68         bitset_t *normal_colors;
  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(irn);
  88                         int node_nr = (int)get_irn_idx(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_out(irn);
  94
  95                         bitset_clear_all(colors);
  96
  97                         /* get assignable colors */
  98                         if (arch_register_req_is(req, limited)) {
  99                                 rbitset_copy_to_bitset(req->limited, colors);
 100                         } else {
 101                                 bitset_copy(colors, lenv->normal_colors);
 102                         }
 103
 104                         /* add the coloring constraint */
 105                         cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 1.0);
 106
 107                         bitset_foreach(colors, col) {
 108                                 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, col), lpp_binary, 0.0);
 109                                 lpp_set_start_value(ienv->lp, var_idx, (col == (unsigned) curr_node_color) ? 1.0 : 0.0);
 110                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
 111
 112                                 lenv->last_x_var = var_idx;
 113                                 if (lenv->first_x_var == -1)
 114                                         lenv->first_x_var = var_idx;
 115                         }
 116
 117                         /* add register constraint constraints */
 118                         bitset_foreach_clear(colors, col) {
 119                                 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 0.0);
 120                                 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, col), lpp_binary, 0.0);
 121                                 lpp_set_start_value(ienv->lp, var_idx, 0.0);
 122                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
 123
 124                                 lenv->last_x_var = var_idx;
 125                         }
 126                 }
 127 }
 128
 129 static void build_interference_cstr(ilp_env_t *ienv)
 130 {
 131         lpp_t       *lpp      = ienv->lp;
 132         local_env_t *lenv     = ienv->env;
 133         be_ifg_t    *ifg      = ienv->co->cenv->ifg;
 134         int          n_colors = lenv->n_colors;
 135         void        *iter     = be_ifg_cliques_iter_alloca(ifg);
 136         ir_node    **clique   = ALLOCAN(ir_node*, n_colors);
 137         int          size;
 138         int          col;
 139         int          i;
 140
 141         char buf[16];
 142
 143         /* for each maximal clique */
 144         be_ifg_foreach_clique(ifg, iter, clique, &size) {
 145                 int realsize = 0;
 146
 147                 for (i=0; i<size; ++i)
 148                         if (!sr_is_removed(ienv->sr, clique[i]))
 149                                 ++realsize;
 150
 151                 if (realsize < 2)
 152                         continue;
 153
 154                 /* for all colors */
 155                 for (col=0; col<n_colors; ++col) {
 156                         int cst_idx = lpp_add_cst(lpp, NULL, lpp_less, 1.0);
 157
 158                         /* for each member of this clique */
 159                         for (i=0; i<size; ++i) {
 160                                 ir_node *irn = clique[i];
 161
 162                                 if (!sr_is_removed(ienv->sr, irn)) {
 163                                         int var_idx = lpp_get_var_idx(lpp, name_cdd(buf, 'x', (int)get_irn_idx(irn), col));
 164                                         lpp_set_factor_fast(lpp, cst_idx, var_idx, 1);
 165                                 }
 166                         }
 167                 }
 168         }
 169 }
 170
 171
 172 /**
 173  * TODO: Remove the dependency of the opt-units data structure
 174  *       by walking over all affinity edges. Graph structure
 175  *       does not provide this walker, yet.
 176  */
 177 static void build_affinity_cstr(ilp_env_t *ienv) {
 178         local_env_t *lenv = ienv->env;
 179         int n_colors      = lenv->n_colors;
 180         unit_t *curr;
 181
 182         /* for all optimization units */
 183         list_for_each_entry(unit_t, curr, &ienv->co->units, units) {
 184                 ir_node *root, *arg;
 185                 int root_nr, arg_nr, i, col, y_idx, root_idx, arg_idx;
 186                 char buf[16];
 187                 int root_col, arg_col;
 188
 189                 root = curr->nodes[0];
 190                 root_nr = (int) get_irn_idx(root);
 191                 root_col = get_irn_col(root);
 192
 193                 for (i = 1; i < curr->node_count; ++i) {
 194                         arg = curr->nodes[i];
 195                         arg_nr = (int) get_irn_idx(arg);
 196                         arg_col = get_irn_col(arg);
 197
 198                         /* add a new affinity variable */
 199                         y_idx = lpp_add_var(ienv->lp, name_cdd_sorted(buf, 'y', root_nr, arg_nr), lpp_binary, curr->costs[i]);
 200                         lpp_set_start_value(ienv->lp, y_idx, (root_col==arg_col) ? 0.0 : 1.0);
 201
 202                         /* add constraints relating the affinity var to the color vars */
 203                         for (col=0; col<n_colors; ++col) {
 204                                 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_less, 0.0);
 205                                 root_idx = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', root_nr, col));
 206                                 arg_idx  = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', arg_nr,  col));
 207
 208                                 lpp_set_factor_fast(ienv->lp, cst_idx, root_idx,  1.0);
 209                                 lpp_set_factor_fast(ienv->lp, cst_idx, arg_idx,  -1.0);
 210                                 lpp_set_factor_fast(ienv->lp, cst_idx, y_idx, -1.0);
 211                         }
 212                 }
 213         }
 214 }
 215
 216 /**
 217  * Helping stuff for build_clique_star_cstr
 218  */
 219 typedef struct _edge_t {
 220         ir_node *n1, *n2;
 221 } edge_t;
 222
 223 static int compare_edge_t(const void *k1, const void *k2, size_t size) {
 224         const edge_t *e1 = k1;
 225         const edge_t *e2 = k2;
 226         (void) size;
 227
 228         return ! (e1->n1 == e2->n1   &&   e1->n2 == e2->n2);
 229 }
 230
 231 #define HASH_EDGE(e) (hash_irn((e)->n1) ^ hash_irn((e)->n2))
 232
 233 static inline edge_t *add_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
 234         edge_t new_edge;
 235
 236         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 237                 new_edge.n1 = n1;
 238                 new_edge.n2 = n2;
 239         } else {
 240                 new_edge.n1 = n2;
 241                 new_edge.n2 = n1;
 242         }
 243         (*counter)++;
 244         return set_insert(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 245 }
 246
 247 static inline edge_t *find_edge(set *edges, ir_node *n1, ir_node *n2) {
 248         edge_t new_edge;
 249
 250         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 251                 new_edge.n1 = n1;
 252                 new_edge.n2 = n2;
 253         } else {
 254                 new_edge.n1 = n2;
 255                 new_edge.n2 = n1;
 256         }
 257         return set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 258 }
 259
 260 static inline void remove_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
 261         edge_t new_edge, *e;
 262
 263         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 264                 new_edge.n1 = n1;
 265                 new_edge.n2 = n2;
 266         } else {
 267                 new_edge.n1 = n2;
 268                 new_edge.n2 = n1;
 269         }
 270         e = set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 271         if (e) {
 272                 e->n1 = NULL;
 273                 e->n2 = NULL;
 274                 (*counter)--;
 275         }
 276 }
 277
 278 #define pset_foreach(pset, irn)  for(irn=pset_first(pset); irn; irn=pset_next(pset))
 279
 280 /**
 281  * Search for an interference clique and an external node
 282  * with affinity edges to all nodes of the clique.
 283  * At most 1 node of the clique can be colored equally with the external node.
 284  */
 285 static void build_clique_star_cstr(ilp_env_t *ienv) {
 286         affinity_node_t *aff;
 287
 288         /* for each node with affinity edges */
 289         co_gs_foreach_aff_node(ienv->co, aff) {
 290                 struct obstack ob;
 291                 neighb_t *nbr;
 292                 const ir_node *center = aff->irn;
 293                 ir_node **nodes;
 294                 set *edges;
 295                 int i, o, n_nodes, n_edges;
 296
 297                 if (arch_irn_is_ignore(aff->irn))
 298                         continue;
 299
 300                 obstack_init(&ob);
 301                 edges = new_set(compare_edge_t, 8);
 302
 303                 /* get all affinity neighbours */
 304                 n_nodes = 0;
 305                 co_gs_foreach_neighb(aff, nbr) {
 306                         if (!arch_irn_is_ignore(nbr->irn)) {
 307                                 obstack_ptr_grow(&ob, nbr->irn);
 308                                 ++n_nodes;
 309                         }
 310                 }
 311                 nodes = obstack_finish(&ob);
 312
 313                 /* get all interference edges between these */
 314                 n_edges = 0;
 315                 for (i=0; i<n_nodes; ++i)
 316                         for (o=0; o<i; ++o)
 317                                 if (be_ifg_connected(ienv->co->cenv->ifg, nodes[i], nodes[o]))
 318                                         add_edge(edges, nodes[i], nodes[o], &n_edges);
 319
 320                 /* cover all these interference edges with maximal cliques */
 321                 while (n_edges) {
 322                         edge_t *e;
 323                         pset *clique = pset_new_ptr(8);
 324                         int growed;
 325
 326                         /* get 2 starting nodes to form a clique */
 327                         for (e=set_first(edges); !e->n1; e=set_next(edges))
 328                                 /*nothing*/ ;
 329
 330                         /* we could be stepped out of the loop before the set iterated to the end */
 331                         set_break(edges);
 332
 333                         pset_insert_ptr(clique, e->n1);
 334                         pset_insert_ptr(clique, e->n2);
 335                         remove_edge(edges, e->n1, e->n2, &n_edges);
 336
 337                         /* while the clique is growing */
 338                         do {
 339                                 growed = 0;
 340
 341                                 /* search for a candidate to extend the clique */
 342                                 for (i=0; i<n_nodes; ++i) {
 343                                         ir_node *member, *cand = nodes[i];
 344                                         int is_cand;
 345
 346                                         /* if its already in the clique try the next */
 347                                         if (pset_find_ptr(clique, cand))
 348                                                 continue;
 349
 350                                         /* are there all necessary interferences? */
 351                                         is_cand = 1;
 352                                         pset_foreach(clique, member) {
 353                                                 if (!find_edge(edges, cand, member)) {
 354                                                         is_cand = 0;
 355                                                         pset_break(clique);
 356                                                         break;
 357                                                 }
 358                                         }
 359
 360                                         /* now we know if we have a clique extender */
 361                                         if (is_cand) {
 362                                                 /* first remove all covered edges */
 363                                                 pset_foreach(clique, member)
 364                                                         remove_edge(edges, cand, member, &n_edges);
 365
 366                                                 /* insert into clique */
 367                                                 pset_insert_ptr(clique, cand);
 368                                                 growed = 1;
 369                                                 break;
 370                                         }
 371                                 }
 372                         } while (growed);
 373
 374                         /* now the clique is maximal. Finally add the constraint */
 375                         {
 376                                 ir_node *member;
 377                                 int var_idx, cst_idx, center_nr, member_nr;
 378                                 char buf[16];
 379
 380                                 cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, pset_count(clique)-1);
 381                                 center_nr = get_irn_idx(center);
 382
 383                                 pset_foreach(clique, member) {
 384                                         member_nr = get_irn_idx(member);
 385                                         var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', center_nr, member_nr));
 386                                         lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 387                                 }
 388                         }
 389
 390                         del_pset(clique);
 391                 }
 392
 393                 del_set(edges);
 394                 obstack_free(&ob, NULL);
 395         }
 396 }
 397
 398
 399 static void extend_path(ilp_env_t *ienv, pdeq *path, const ir_node *irn) {
 400         be_ifg_t *ifg = ienv->co->cenv->ifg;
 401         int i, len;
 402         ir_node **curr_path;
 403         affinity_node_t *aff;
 404         neighb_t *nbr;
 405
 406         /* do not walk backwards or in circles */
 407         if (pdeq_contains(path, irn))
 408                 return;
 409
 410         if (arch_irn_is_ignore(irn))
 411                 return;
 412
 413         /* insert the new irn */
 414         pdeq_putr(path, irn);
 415
 416
 417
 418         /* check for forbidden interferences */
 419         len       = pdeq_len(path);
 420         curr_path = ALLOCAN(ir_node*, len);
 421         pdeq_copyl(path, (const void **)curr_path);
 422
 423         for (i=1; i<len; ++i)
 424                 if (be_ifg_connected(ifg, irn, curr_path[i]))
 425                         goto end;
 426
 427
 428
 429         /* check for terminating interference */
 430         if (be_ifg_connected(ifg, irn, curr_path[0])) {
 431
 432                 /* One node is not a path. */
 433                 /* And a path of length 2 is covered by a clique star constraint. */
 434                 if (len > 2) {
 435                         /* finally build the constraint */
 436                         int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, 1.0);
 437                         for (i=1; i<len; ++i) {
 438                                 char buf[16];
 439                                 int nr_1    = get_irn_idx(curr_path[i-1]);
 440                                 int nr_2    = get_irn_idx(curr_path[i]);
 441                                 int var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', nr_1, nr_2));
 442                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 443                         }
 444                 }
 445
 446                 /* this path cannot be extended anymore */
 447                 goto end;
 448         }
 449
 450
 451
 452         /* recursively extend the path */
 453         aff = get_affinity_info(ienv->co, irn);
 454         co_gs_foreach_neighb(aff, nbr)
 455                 extend_path(ienv, path, nbr->irn);
 456
 457
 458 end:
 459         /* remove the irn */
 460         pdeq_getr(path);
 461
 462 }
 463
 464 /**
 465  *  Search a path of affinity edges, whose ends are connected
 466  *  by an interference edge and there are no other interference
 467  *  edges in between.
 468  *  Then at least one of these affinity edges must break.
 469  */
 470 static void build_path_cstr(ilp_env_t *ienv) {
 471         affinity_node_t *aff_info;
 472
 473         /* for each node with affinity edges */
 474         co_gs_foreach_aff_node(ienv->co, aff_info) {
 475                 pdeq *path = new_pdeq();
 476
 477                 extend_path(ienv, path, aff_info->irn);
 478
 479                 del_pdeq(path);
 480         }
 481 }
 482
 483 static void ilp2_build(ilp_env_t *ienv) {
 484         local_env_t *lenv = ienv->env;
 485         int lower_bound;
 486
 487         ienv->lp = new_lpp(ienv->co->name, lpp_minimize);
 488         build_coloring_cstr(ienv);
 489         build_interference_cstr(ienv);
 490         build_affinity_cstr(ienv);
 491         build_clique_star_cstr(ienv);
 492         build_path_cstr(ienv);
 493
 494         lower_bound = co_get_lower_bound(ienv->co) - co_get_inevit_copy_costs(ienv->co);
 495         lpp_set_bound(ienv->lp, lower_bound);
 496         lpp_set_time_limit(ienv->lp, lenv->time_limit);
 497 }
 498
 499 static void ilp2_apply(ilp_env_t *ienv) {
 500         local_env_t *lenv = ienv->env;
 501         int i;
 502
 503         /* first check if there was sth. to optimize */
 504         if (lenv->first_x_var >= 0) {
 505                 int              count = lenv->last_x_var - lenv->first_x_var + 1;
 506                 double          *sol   = XMALLOCN(double, count);
 507                 lpp_sol_state_t  state = lpp_get_solution(ienv->lp, sol, lenv->first_x_var, lenv->last_x_var);
 508
 509                 if (state != lpp_optimal) {
 510                         printf("WARNING %s: Solution state is not 'optimal': %d\n", ienv->co->name, state);
 511                         assert(state >= lpp_feasible && "The solution should at least be feasible!");
 512                 }
 513
 514                 for (i=0; i<count; ++i) {
 515                         int nodenr, color;
 516                         char var_name[16];
 517
 518                         if (sol[i] > 1-EPSILON) { /* split variable name into components */
 519                                 lpp_get_var_name(ienv->lp, lenv->first_x_var+i, var_name, sizeof(var_name));
 520
 521                                 if (sscanf(var_name, "x_%d_%d", &nodenr, &color) == 2) {
 522                                         ir_node *irn = pmap_get(lenv->nr_2_irn, INT_TO_PTR(nodenr));
 523                                         assert(irn && "This node number must be present in the map");
 524
 525                                         set_irn_col(ienv->co, irn, color);
 526                                 } else
 527                                         assert(0 && "This should be a x-var");
 528                         }
 529                 }
 530
 531                 xfree(sol);
 532         }
 533
 534 #ifdef COPYOPT_STAT
 535         /* TODO adapt to multiple possible ILPs */
 536         copystat_add_ilp_time((int)(1000.0*lpp_get_sol_time(pi->curr_lp)));  //now we have ms
 537         copystat_add_ilp_vars(lpp_get_var_count(pi->curr_lp));
 538         copystat_add_ilp_csts(lpp_get_cst_count(pi->curr_lp));
 539         copystat_add_ilp_iter(lpp_get_iter_cnt(pi->curr_lp));
 540 #endif
 541 }
 542
 543 void be_init_copyilp2(void)
 544 {
 545         static co_algo_info copyheur = {
 546                 co_solve_ilp2, 1
 547         };
 548
 549         be_register_copyopt("ilp", &copyheur);
 550 }
 551
 552 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyilp2);
 553
 554 int co_solve_ilp2(copy_opt_t *co) {
 555         lpp_sol_state_t sol_state;
 556         ilp_env_t *ienv;
 557         local_env_t my;
 558
 559         ASSERT_OU_AVAIL(co); //See build_clique_st
 560         ASSERT_GS_AVAIL(co);
 561
 562         my.time_limit  = 0;
 563         my.first_x_var = -1;
 564         my.last_x_var  = -1;
 565         my.nr_2_irn    = pmap_create();
 566         FIRM_DBG_REGISTER(my.dbg, "firm.be.coilp2");
 567
 568         my.normal_colors = bitset_alloca(arch_register_class_n_regs(co->cls));
 569         bitset_clear_all(my.normal_colors);
 570         arch_put_non_ignore_regs(co->cls, my.normal_colors);
 571         my.n_colors = bitset_popcnt(my.normal_colors);
 572
 573         ienv = new_ilp_env(co, ilp2_build, ilp2_apply, &my);
 574
 575         sol_state = ilp_go(ienv);
 576
 577         pmap_destroy(my.nr_2_irn);
 578         free_ilp_env(ienv);
 579
 580         return sol_state == lpp_optimal;
 581 }
 582
 583 #else /* WITH_ILP */
 584
 585 static inline void only_that_you_can_compile_without_WITH_ILP_defined(void) {
 586 }
 587
 588 #endif /* WITH_ILP */