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.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 {
  75         be_ifg_t *ifg     = ienv->co->cenv->ifg;
  76         nodes_iter_t iter;
  77         bitset_t *colors;
  78         ir_node *irn;
  79         char buf[16];
  80
  81         colors = bitset_alloca(arch_register_class_n_regs(ienv->co->cls));
  82
  83         be_ifg_foreach_node(ifg, &iter, irn)
  84                 if (!sr_is_removed(ienv->sr, irn)) {
  85                         unsigned col;
  86                         int cst_idx;
  87                         const arch_register_req_t *req;
  88                         int curr_node_color = get_irn_col(irn);
  89                         int node_nr = (int)get_irn_idx(irn);
  90                         local_env_t *lenv = ienv->env;
  91
  92                         pmap_insert(lenv->nr_2_irn, INT_TO_PTR(node_nr), irn);
  93
  94                         req = arch_get_register_req_out(irn);
  95
  96                         bitset_clear_all(colors);
  97
  98                         /* get assignable colors */
  99                         if (arch_register_req_is(req, limited)) {
 100                                 rbitset_copy_to_bitset(req->limited, colors);
 101                         } else {
 102                                 bitset_copy(colors, lenv->normal_colors);
 103                         }
 104
 105                         /* add the coloring constraint */
 106                         cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 1.0);
 107
 108                         bitset_foreach(colors, col) {
 109                                 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, col), lpp_binary, 0.0);
 110                                 lpp_set_start_value(ienv->lp, var_idx, (col == (unsigned) curr_node_color) ? 1.0 : 0.0);
 111                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
 112
 113                                 lenv->last_x_var = var_idx;
 114                                 if (lenv->first_x_var == -1)
 115                                         lenv->first_x_var = var_idx;
 116                         }
 117
 118                         /* add register constraint constraints */
 119                         bitset_foreach_clear(colors, col) {
 120                                 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 0.0);
 121                                 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, col), lpp_binary, 0.0);
 122                                 lpp_set_start_value(ienv->lp, var_idx, 0.0);
 123                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
 124
 125                                 lenv->last_x_var = var_idx;
 126                         }
 127                 }
 128 }
 129
 130 static void build_interference_cstr(ilp_env_t *ienv)
 131 {
 132         lpp_t       *lpp      = ienv->lp;
 133         local_env_t *lenv     = ienv->env;
 134         be_ifg_t    *ifg      = ienv->co->cenv->ifg;
 135         int          n_colors = lenv->n_colors;
 136         cliques_iter_t iter;
 137         ir_node    **clique   = ALLOCAN(ir_node*, n_colors);
 138         int          size;
 139         int          col;
 140         int          i;
 141
 142         char buf[16];
 143
 144         /* for each maximal clique */
 145         be_ifg_foreach_clique(ifg, &iter, clique, &size) {
 146                 int realsize = 0;
 147
 148                 for (i=0; i<size; ++i)
 149                         if (!sr_is_removed(ienv->sr, clique[i]))
 150                                 ++realsize;
 151
 152                 if (realsize < 2)
 153                         continue;
 154
 155                 /* for all colors */
 156                 for (col=0; col<n_colors; ++col) {
 157                         int cst_idx = lpp_add_cst(lpp, NULL, lpp_less, 1.0);
 158
 159                         /* for each member of this clique */
 160                         for (i=0; i<size; ++i) {
 161                                 ir_node *irn = clique[i];
 162
 163                                 if (!sr_is_removed(ienv->sr, irn)) {
 164                                         int var_idx = lpp_get_var_idx(lpp, name_cdd(buf, 'x', (int)get_irn_idx(irn), col));
 165                                         lpp_set_factor_fast(lpp, cst_idx, var_idx, 1);
 166                                 }
 167                         }
 168                 }
 169         }
 170 }
 171
 172
 173 /**
 174  * TODO: Remove the dependency of the opt-units data structure
 175  *       by walking over all affinity edges. Graph structure
 176  *       does not provide this walker, yet.
 177  */
 178 static void build_affinity_cstr(ilp_env_t *ienv)
 179 {
 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 {
 227         const edge_t *e1 = k1;
 228         const edge_t *e2 = k2;
 229         (void) size;
 230
 231         return ! (e1->n1 == e2->n1   &&   e1->n2 == e2->n2);
 232 }
 233
 234 #define HASH_EDGE(e) (hash_irn((e)->n1) ^ hash_irn((e)->n2))
 235
 236 static inline edge_t *add_edge(set *edges, ir_node *n1, ir_node *n2, int *counter)
 237 {
 238         edge_t new_edge;
 239
 240         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 241                 new_edge.n1 = n1;
 242                 new_edge.n2 = n2;
 243         } else {
 244                 new_edge.n1 = n2;
 245                 new_edge.n2 = n1;
 246         }
 247         (*counter)++;
 248         return set_insert(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 249 }
 250
 251 static inline edge_t *find_edge(set *edges, ir_node *n1, ir_node *n2)
 252 {
 253         edge_t new_edge;
 254
 255         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 256                 new_edge.n1 = n1;
 257                 new_edge.n2 = n2;
 258         } else {
 259                 new_edge.n1 = n2;
 260                 new_edge.n2 = n1;
 261         }
 262         return set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 263 }
 264
 265 static inline void remove_edge(set *edges, ir_node *n1, ir_node *n2, int *counter)
 266 {
 267         edge_t new_edge, *e;
 268
 269         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 270                 new_edge.n1 = n1;
 271                 new_edge.n2 = n2;
 272         } else {
 273                 new_edge.n1 = n2;
 274                 new_edge.n2 = n1;
 275         }
 276         e = set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 277         if (e) {
 278                 e->n1 = NULL;
 279                 e->n2 = NULL;
 280                 (*counter)--;
 281         }
 282 }
 283
 284 #define pset_foreach(pset, irn)  for (irn=pset_first(pset); irn; irn=pset_next(pset))
 285
 286 /**
 287  * Search for an interference clique and an external node
 288  * with affinity edges to all nodes of the clique.
 289  * At most 1 node of the clique can be colored equally with the external node.
 290  */
 291 static void build_clique_star_cstr(ilp_env_t *ienv)
 292 {
 293         affinity_node_t *aff;
 294
 295         /* for each node with affinity edges */
 296         co_gs_foreach_aff_node(ienv->co, aff) {
 297                 struct obstack ob;
 298                 neighb_t *nbr;
 299                 const ir_node *center = aff->irn;
 300                 ir_node **nodes;
 301                 set *edges;
 302                 int i, o, n_nodes, n_edges;
 303
 304                 if (arch_irn_is_ignore(aff->irn))
 305                         continue;
 306
 307                 obstack_init(&ob);
 308                 edges = new_set(compare_edge_t, 8);
 309
 310                 /* get all affinity neighbours */
 311                 n_nodes = 0;
 312                 co_gs_foreach_neighb(aff, nbr) {
 313                         if (!arch_irn_is_ignore(nbr->irn)) {
 314                                 obstack_ptr_grow(&ob, nbr->irn);
 315                                 ++n_nodes;
 316                         }
 317                 }
 318                 nodes = obstack_finish(&ob);
 319
 320                 /* get all interference edges between these */
 321                 n_edges = 0;
 322                 for (i=0; i<n_nodes; ++i)
 323                         for (o=0; o<i; ++o)
 324                                 if (be_ifg_connected(ienv->co->cenv->ifg, nodes[i], nodes[o]))
 325                                         add_edge(edges, nodes[i], nodes[o], &n_edges);
 326
 327                 /* cover all these interference edges with maximal cliques */
 328                 while (n_edges) {
 329                         edge_t *e;
 330                         pset *clique = pset_new_ptr(8);
 331                         int growed;
 332
 333                         /* get 2 starting nodes to form a clique */
 334                         for (e=set_first(edges); !e->n1; e=set_next(edges))
 335                                 /*nothing*/ ;
 336
 337                         /* we could be stepped out of the loop before the set iterated to the end */
 338                         set_break(edges);
 339
 340                         pset_insert_ptr(clique, e->n1);
 341                         pset_insert_ptr(clique, e->n2);
 342                         remove_edge(edges, e->n1, e->n2, &n_edges);
 343
 344                         /* while the clique is growing */
 345                         do {
 346                                 growed = 0;
 347
 348                                 /* search for a candidate to extend the clique */
 349                                 for (i=0; i<n_nodes; ++i) {
 350                                         ir_node *member, *cand = nodes[i];
 351                                         int is_cand;
 352
 353                                         /* if its already in the clique try the next */
 354                                         if (pset_find_ptr(clique, cand))
 355                                                 continue;
 356
 357                                         /* are there all necessary interferences? */
 358                                         is_cand = 1;
 359                                         pset_foreach(clique, member) {
 360                                                 if (!find_edge(edges, cand, member)) {
 361                                                         is_cand = 0;
 362                                                         pset_break(clique);
 363                                                         break;
 364                                                 }
 365                                         }
 366
 367                                         /* now we know if we have a clique extender */
 368                                         if (is_cand) {
 369                                                 /* first remove all covered edges */
 370                                                 pset_foreach(clique, member)
 371                                                         remove_edge(edges, cand, member, &n_edges);
 372
 373                                                 /* insert into clique */
 374                                                 pset_insert_ptr(clique, cand);
 375                                                 growed = 1;
 376                                                 break;
 377                                         }
 378                                 }
 379                         } while (growed);
 380
 381                         /* now the clique is maximal. Finally add the constraint */
 382                         {
 383                                 ir_node *member;
 384                                 int var_idx, cst_idx, center_nr, member_nr;
 385                                 char buf[16];
 386
 387                                 cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, pset_count(clique)-1);
 388                                 center_nr = get_irn_idx(center);
 389
 390                                 pset_foreach(clique, member) {
 391                                         member_nr = get_irn_idx(member);
 392                                         var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', center_nr, member_nr));
 393                                         lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 394                                 }
 395                         }
 396
 397                         del_pset(clique);
 398                 }
 399
 400                 del_set(edges);
 401                 obstack_free(&ob, NULL);
 402         }
 403 }
 404
 405
 406 static void extend_path(ilp_env_t *ienv, pdeq *path, const ir_node *irn)
 407 {
 408         be_ifg_t *ifg = ienv->co->cenv->ifg;
 409         int i, len;
 410         ir_node **curr_path;
 411         affinity_node_t *aff;
 412         neighb_t *nbr;
 413
 414         /* do not walk backwards or in circles */
 415         if (pdeq_contains(path, irn))
 416                 return;
 417
 418         if (arch_irn_is_ignore(irn))
 419                 return;
 420
 421         /* insert the new irn */
 422         pdeq_putr(path, irn);
 423
 424
 425
 426         /* check for forbidden interferences */
 427         len       = pdeq_len(path);
 428         curr_path = ALLOCAN(ir_node*, len);
 429         pdeq_copyl(path, (const void **)curr_path);
 430
 431         for (i=1; i<len; ++i)
 432                 if (be_ifg_connected(ifg, irn, curr_path[i]))
 433                         goto end;
 434
 435
 436
 437         /* check for terminating interference */
 438         if (be_ifg_connected(ifg, irn, curr_path[0])) {
 439
 440                 /* One node is not a path. */
 441                 /* And a path of length 2 is covered by a clique star constraint. */
 442                 if (len > 2) {
 443                         /* finally build the constraint */
 444                         int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, 1.0);
 445                         for (i=1; i<len; ++i) {
 446                                 char buf[16];
 447                                 int nr_1    = get_irn_idx(curr_path[i-1]);
 448                                 int nr_2    = get_irn_idx(curr_path[i]);
 449                                 int var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', nr_1, nr_2));
 450                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 451                         }
 452                 }
 453
 454                 /* this path cannot be extended anymore */
 455                 goto end;
 456         }
 457
 458
 459
 460         /* recursively extend the path */
 461         aff = get_affinity_info(ienv->co, irn);
 462         co_gs_foreach_neighb(aff, nbr)
 463                 extend_path(ienv, path, nbr->irn);
 464
 465
 466 end:
 467         /* remove the irn */
 468         pdeq_getr(path);
 469
 470 }
 471
 472 /**
 473  *  Search a path of affinity edges, whose ends are connected
 474  *  by an interference edge and there are no other interference
 475  *  edges in between.
 476  *  Then at least one of these affinity edges must break.
 477  */
 478 static void build_path_cstr(ilp_env_t *ienv)
 479 {
 480         affinity_node_t *aff_info;
 481
 482         /* for each node with affinity edges */
 483         co_gs_foreach_aff_node(ienv->co, aff_info) {
 484                 pdeq *path = new_pdeq();
 485
 486                 extend_path(ienv, path, aff_info->irn);
 487
 488                 del_pdeq(path);
 489         }
 490 }
 491
 492 static void ilp2_build(ilp_env_t *ienv)
 493 {
 494         local_env_t *lenv = ienv->env;
 495         int lower_bound;
 496
 497         ienv->lp = new_lpp(ienv->co->name, lpp_minimize);
 498         build_coloring_cstr(ienv);
 499         build_interference_cstr(ienv);
 500         build_affinity_cstr(ienv);
 501         build_clique_star_cstr(ienv);
 502         build_path_cstr(ienv);
 503
 504         lower_bound = co_get_lower_bound(ienv->co) - co_get_inevit_copy_costs(ienv->co);
 505         lpp_set_bound(ienv->lp, lower_bound);
 506         lpp_set_time_limit(ienv->lp, lenv->time_limit);
 507 }
 508
 509 static void ilp2_apply(ilp_env_t *ienv)
 510 {
 511         local_env_t *lenv = ienv->env;
 512         int i;
 513
 514         /* first check if there was sth. to optimize */
 515         if (lenv->first_x_var >= 0) {
 516                 int              count = lenv->last_x_var - lenv->first_x_var + 1;
 517                 double          *sol   = XMALLOCN(double, count);
 518                 lpp_sol_state_t  state = lpp_get_solution(ienv->lp, sol, lenv->first_x_var, lenv->last_x_var);
 519
 520                 if (state != lpp_optimal) {
 521                         printf("WARNING %s: Solution state is not 'optimal': %d\n", ienv->co->name, state);
 522                         assert(state >= lpp_feasible && "The solution should at least be feasible!");
 523                 }
 524
 525                 for (i=0; i<count; ++i) {
 526                         int nodenr, color;
 527                         char var_name[16];
 528
 529                         if (sol[i] > 1-EPSILON) { /* split variable name into components */
 530                                 lpp_get_var_name(ienv->lp, lenv->first_x_var+i, var_name, sizeof(var_name));
 531
 532                                 if (sscanf(var_name, "x_%d_%d", &nodenr, &color) == 2) {
 533                                         ir_node *irn = pmap_get(lenv->nr_2_irn, INT_TO_PTR(nodenr));
 534                                         assert(irn && "This node number must be present in the map");
 535
 536                                         set_irn_col(ienv->co, irn, color);
 537                                 } else
 538                                         assert(0 && "This should be a x-var");
 539                         }
 540                 }
 541
 542                 xfree(sol);
 543         }
 544
 545 #ifdef COPYOPT_STAT
 546         /* TODO adapt to multiple possible ILPs */
 547         copystat_add_ilp_time((int)(1000.0*lpp_get_sol_time(pi->curr_lp)));  //now we have ms
 548         copystat_add_ilp_vars(lpp_get_var_count(pi->curr_lp));
 549         copystat_add_ilp_csts(lpp_get_cst_count(pi->curr_lp));
 550         copystat_add_ilp_iter(lpp_get_iter_cnt(pi->curr_lp));
 551 #endif
 552 }
 553
 554 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyilp2);
 555 void be_init_copyilp2(void)
 556 {
 557         static co_algo_info copyheur = {
 558                 co_solve_ilp2, 1
 559         };
 560
 561         be_register_copyopt("ilp", &copyheur);
 562 }
 563
 564 int co_solve_ilp2(copy_opt_t *co)
 565 {
 566         lpp_sol_state_t sol_state;
 567         ilp_env_t *ienv;
 568         local_env_t my;
 569
 570         ASSERT_OU_AVAIL(co); //See build_clique_st
 571         ASSERT_GS_AVAIL(co);
 572
 573         my.time_limit  = 0;
 574         my.first_x_var = -1;
 575         my.last_x_var  = -1;
 576         my.nr_2_irn    = pmap_create();
 577         FIRM_DBG_REGISTER(my.dbg, "firm.be.coilp2");
 578
 579         my.normal_colors = bitset_alloca(arch_register_class_n_regs(co->cls));
 580         bitset_clear_all(my.normal_colors);
 581         arch_put_non_ignore_regs(co->cls, my.normal_colors);
 582         my.n_colors = bitset_popcount(my.normal_colors);
 583
 584         ienv = new_ilp_env(co, ilp2_build, ilp2_apply, &my);
 585
 586         sol_state = ilp_go(ienv);
 587
 588         pmap_destroy(my.nr_2_irn);
 589         free_ilp_env(ienv);
 590
 591         return sol_state == lpp_optimal;
 592 }
 593
 594 #else /* WITH_ILP */
 595
 596 static inline void only_that_you_can_compile_without_WITH_ILP_defined(void)
 597 {
 598 }
 599
 600 #endif /* WITH_ILP */