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  *
  26  * ILP formalization using G=(V, E, Q):
  27  *  - 2 class of variables: coloring vars x_ic   and   equal color vars y_ij
  28  *  - Path constraints
  29  *  - Clique-star constraints
  30  *
  31  *
  32  * \min \sum_{ (i,j) \in Q }  w_ij y_ij
  33  *
  34  *     \sum_c x_nc           =  1           n \in N, c \in C
  35  *
  36  *     x_nc                  =  0           n \in N, c \not\in C(n)
  37  *
  38  *     \sum x_nc            <=  1           x_nc \in Clique \in AllCliques,  c \in C
  39  *
  40  *     \sum_{e \in p} y_e   >=  1           p \in P      path constraints
  41  *
  42  *     \sum_{e \in cs} y_e  >= |cs| - 1     cs \in CP    clique-star constraints
  43  *
  44  *     x_nc, y_ij \in N,   w_ij \in R^+
  45  */
  46 #include "config.h"
  47
  48 #include "bitset.h"
  49 #include "raw_bitset.h"
  50 #include "pdeq.h"
  51
  52 #include "util.h"
  53 #include "irgwalk.h"
  54 #include "becopyilp_t.h"
  55 #include "beifg.h"
  56 #include "besched.h"
  57 #include "bemodule.h"
  58
  59 #define DEBUG_LVL 1
  60
  61 typedef struct local_env_t {
  62         double time_limit;
  63         int first_x_var, last_x_var;
  64         int n_colors;
  65         bitset_t *normal_colors;
  66         pmap *nr_2_irn;
  67         DEBUG_ONLY(firm_dbg_module_t *dbg;)
  68 } local_env_t;
  69
  70 static void build_coloring_cstr(ilp_env_t *ienv)
  71 {
  72         be_ifg_t *ifg     = ienv->co->cenv->ifg;
  73         nodes_iter_t iter;
  74         bitset_t *colors;
  75         ir_node *irn;
  76         char buf[16];
  77
  78         colors = bitset_alloca(arch_register_class_n_regs(ienv->co->cls));
  79
  80         be_ifg_foreach_node(ifg, &iter, irn)
  81                 if (!sr_is_removed(ienv->sr, irn)) {
  82                         size_t col;
  83                         int cst_idx;
  84                         const arch_register_req_t *req;
  85                         int curr_node_color = get_irn_col(irn);
  86                         int node_nr = (int)get_irn_idx(irn);
  87                         local_env_t *lenv = ienv->env;
  88
  89                         pmap_insert(lenv->nr_2_irn, INT_TO_PTR(node_nr), irn);
  90
  91                         req = arch_get_irn_register_req(irn);
  92
  93                         bitset_clear_all(colors);
  94
  95                         /* get assignable colors */
  96                         if (arch_register_req_is(req, limited)) {
  97                                 rbitset_copy_to_bitset(req->limited, colors);
  98                         } else {
  99                                 bitset_copy(colors, lenv->normal_colors);
 100                         }
 101
 102                         /* add the coloring constraint */
 103                         cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 1.0);
 104
 105                         bitset_foreach(colors, col) {
 106                                 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, (int)col), lpp_binary, 0.0);
 107                                 lpp_set_start_value(ienv->lp, var_idx, (col == (unsigned) curr_node_color) ? 1.0 : 0.0);
 108                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
 109
 110                                 lenv->last_x_var = var_idx;
 111                                 if (lenv->first_x_var == -1)
 112                                         lenv->first_x_var = var_idx;
 113                         }
 114
 115                         /* add register constraint constraints */
 116                         bitset_foreach_clear(colors, col) {
 117                                 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 0.0);
 118                                 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, (int)col), lpp_binary, 0.0);
 119                                 lpp_set_start_value(ienv->lp, var_idx, 0.0);
 120                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
 121
 122                                 lenv->last_x_var = var_idx;
 123                         }
 124                 }
 125 }
 126
 127 static void build_interference_cstr(ilp_env_t *ienv)
 128 {
 129         lpp_t       *lpp      = ienv->lp;
 130         local_env_t *lenv     = ienv->env;
 131         be_ifg_t    *ifg      = ienv->co->cenv->ifg;
 132         int          n_colors = lenv->n_colors;
 133         cliques_iter_t iter;
 134         ir_node    **clique   = ALLOCAN(ir_node*, n_colors);
 135         int          size;
 136         int          col;
 137         int          i;
 138
 139         char buf[16];
 140
 141         /* for each maximal clique */
 142         be_ifg_foreach_clique(ifg, &iter, clique, &size) {
 143                 int realsize = 0;
 144
 145                 for (i=0; i<size; ++i)
 146                         if (!sr_is_removed(ienv->sr, clique[i]))
 147                                 ++realsize;
 148
 149                 if (realsize < 2)
 150                         continue;
 151
 152                 /* for all colors */
 153                 for (col=0; col<n_colors; ++col) {
 154                         int cst_idx = lpp_add_cst(lpp, NULL, lpp_less_equal, 1.0);
 155
 156                         /* for each member of this clique */
 157                         for (i=0; i<size; ++i) {
 158                                 ir_node *irn = clique[i];
 159
 160                                 if (!sr_is_removed(ienv->sr, irn)) {
 161                                         int var_idx = lpp_get_var_idx(lpp, name_cdd(buf, 'x', (int)get_irn_idx(irn), col));
 162                                         lpp_set_factor_fast(lpp, cst_idx, var_idx, 1);
 163                                 }
 164                         }
 165                 }
 166         }
 167 }
 168
 169
 170 /**
 171  * TODO: Remove the dependency of the opt-units data structure
 172  *       by walking over all affinity edges. Graph structure
 173  *       does not provide this walker, yet.
 174  */
 175 static void build_affinity_cstr(ilp_env_t *ienv)
 176 {
 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_equal, 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 {
 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 {
 235         edge_t new_edge;
 236
 237         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 238                 new_edge.n1 = n1;
 239                 new_edge.n2 = n2;
 240         } else {
 241                 new_edge.n1 = n2;
 242                 new_edge.n2 = n1;
 243         }
 244         (*counter)++;
 245         return set_insert(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 246 }
 247
 248 static inline edge_t *find_edge(set *edges, ir_node *n1, ir_node *n2)
 249 {
 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 {
 264         edge_t new_edge, *e;
 265
 266         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 267                 new_edge.n1 = n1;
 268                 new_edge.n2 = n2;
 269         } else {
 270                 new_edge.n1 = n2;
 271                 new_edge.n2 = n1;
 272         }
 273         e = set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 274         if (e) {
 275                 e->n1 = NULL;
 276                 e->n2 = NULL;
 277                 (*counter)--;
 278         }
 279 }
 280
 281 #define pset_foreach(pset, irn)  for (irn=pset_first(pset); irn; irn=pset_next(pset))
 282
 283 /**
 284  * Search for an interference clique and an external node
 285  * with affinity edges to all nodes of the clique.
 286  * At most 1 node of the clique can be colored equally with the external node.
 287  */
 288 static void build_clique_star_cstr(ilp_env_t *ienv)
 289 {
 290         affinity_node_t *aff;
 291
 292         /* for each node with affinity edges */
 293         co_gs_foreach_aff_node(ienv->co, aff) {
 294                 struct obstack ob;
 295                 neighb_t *nbr;
 296                 const ir_node *center = aff->irn;
 297                 ir_node **nodes;
 298                 set *edges;
 299                 int i, o, n_nodes, n_edges;
 300
 301                 if (arch_irn_is_ignore(aff->irn))
 302                         continue;
 303
 304                 obstack_init(&ob);
 305                 edges = new_set(compare_edge_t, 8);
 306
 307                 /* get all affinity neighbours */
 308                 n_nodes = 0;
 309                 co_gs_foreach_neighb(aff, nbr) {
 310                         if (!arch_irn_is_ignore(nbr->irn)) {
 311                                 obstack_ptr_grow(&ob, nbr->irn);
 312                                 ++n_nodes;
 313                         }
 314                 }
 315                 nodes = obstack_finish(&ob);
 316
 317                 /* get all interference edges between these */
 318                 n_edges = 0;
 319                 for (i=0; i<n_nodes; ++i)
 320                         for (o=0; o<i; ++o)
 321                                 if (be_ifg_connected(ienv->co->cenv->ifg, nodes[i], nodes[o]))
 322                                         add_edge(edges, nodes[i], nodes[o], &n_edges);
 323
 324                 /* cover all these interference edges with maximal cliques */
 325                 while (n_edges) {
 326                         edge_t *e;
 327                         pset *clique = pset_new_ptr(8);
 328                         int growed;
 329
 330                         /* get 2 starting nodes to form a clique */
 331                         for (e=set_first(edges); !e->n1; e=set_next(edges)) {
 332                         }
 333
 334                         /* we could be stepped out of the loop before the set iterated to the end */
 335                         set_break(edges);
 336
 337                         pset_insert_ptr(clique, e->n1);
 338                         pset_insert_ptr(clique, e->n2);
 339                         remove_edge(edges, e->n1, e->n2, &n_edges);
 340
 341                         /* while the clique is growing */
 342                         do {
 343                                 growed = 0;
 344
 345                                 /* search for a candidate to extend the clique */
 346                                 for (i=0; i<n_nodes; ++i) {
 347                                         ir_node *member, *cand = nodes[i];
 348                                         int is_cand;
 349
 350                                         /* if its already in the clique try the next */
 351                                         if (pset_find_ptr(clique, cand))
 352                                                 continue;
 353
 354                                         /* are there all necessary interferences? */
 355                                         is_cand = 1;
 356                                         pset_foreach(clique, member) {
 357                                                 if (!find_edge(edges, cand, member)) {
 358                                                         is_cand = 0;
 359                                                         pset_break(clique);
 360                                                         break;
 361                                                 }
 362                                         }
 363
 364                                         /* now we know if we have a clique extender */
 365                                         if (is_cand) {
 366                                                 /* first remove all covered edges */
 367                                                 pset_foreach(clique, member)
 368                                                         remove_edge(edges, cand, member, &n_edges);
 369
 370                                                 /* insert into clique */
 371                                                 pset_insert_ptr(clique, cand);
 372                                                 growed = 1;
 373                                                 break;
 374                                         }
 375                                 }
 376                         } while (growed);
 377
 378                         /* now the clique is maximal. Finally add the constraint */
 379                         {
 380                                 ir_node *member;
 381                                 int var_idx, cst_idx, center_nr, member_nr;
 382                                 char buf[16];
 383
 384                                 cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater_equal, pset_count(clique)-1);
 385                                 center_nr = get_irn_idx(center);
 386
 387                                 pset_foreach(clique, member) {
 388                                         member_nr = get_irn_idx(member);
 389                                         var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', center_nr, member_nr));
 390                                         lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 391                                 }
 392                         }
 393
 394                         del_pset(clique);
 395                 }
 396
 397                 del_set(edges);
 398                 obstack_free(&ob, NULL);
 399         }
 400 }
 401
 402
 403 static void extend_path(ilp_env_t *ienv, pdeq *path, const ir_node *irn)
 404 {
 405         be_ifg_t *ifg = ienv->co->cenv->ifg;
 406         int i, len;
 407         ir_node **curr_path;
 408         affinity_node_t *aff;
 409         neighb_t *nbr;
 410
 411         /* do not walk backwards or in circles */
 412         if (pdeq_contains(path, irn))
 413                 return;
 414
 415         if (arch_irn_is_ignore(irn))
 416                 return;
 417
 418         /* insert the new irn */
 419         pdeq_putr(path, irn);
 420
 421
 422
 423         /* check for forbidden interferences */
 424         len       = pdeq_len(path);
 425         curr_path = ALLOCAN(ir_node*, len);
 426         pdeq_copyl(path, (const void **)curr_path);
 427
 428         for (i=1; i<len; ++i)
 429                 if (be_ifg_connected(ifg, irn, curr_path[i]))
 430                         goto end;
 431
 432
 433
 434         /* check for terminating interference */
 435         if (be_ifg_connected(ifg, irn, curr_path[0])) {
 436
 437                 /* One node is not a path. */
 438                 /* And a path of length 2 is covered by a clique star constraint. */
 439                 if (len > 2) {
 440                         /* finally build the constraint */
 441                         int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater_equal, 1.0);
 442                         for (i=1; i<len; ++i) {
 443                                 char buf[16];
 444                                 int nr_1    = get_irn_idx(curr_path[i-1]);
 445                                 int nr_2    = get_irn_idx(curr_path[i]);
 446                                 int var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', nr_1, nr_2));
 447                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 448                         }
 449                 }
 450
 451                 /* this path cannot be extended anymore */
 452                 goto end;
 453         }
 454
 455
 456
 457         /* recursively extend the path */
 458         aff = get_affinity_info(ienv->co, irn);
 459         co_gs_foreach_neighb(aff, nbr)
 460                 extend_path(ienv, path, nbr->irn);
 461
 462
 463 end:
 464         /* remove the irn */
 465         pdeq_getr(path);
 466
 467 }
 468
 469 /**
 470  *  Search a path of affinity edges, whose ends are connected
 471  *  by an interference edge and there are no other interference
 472  *  edges in between.
 473  *  Then at least one of these affinity edges must break.
 474  */
 475 static void build_path_cstr(ilp_env_t *ienv)
 476 {
 477         affinity_node_t *aff_info;
 478
 479         /* for each node with affinity edges */
 480         co_gs_foreach_aff_node(ienv->co, aff_info) {
 481                 pdeq *path = new_pdeq();
 482
 483                 extend_path(ienv, path, aff_info->irn);
 484
 485                 del_pdeq(path);
 486         }
 487 }
 488
 489 static void ilp2_build(ilp_env_t *ienv)
 490 {
 491         local_env_t *lenv = ienv->env;
 492         int lower_bound;
 493
 494         ienv->lp = lpp_new(ienv->co->name, lpp_minimize);
 495         build_coloring_cstr(ienv);
 496         build_interference_cstr(ienv);
 497         build_affinity_cstr(ienv);
 498         build_clique_star_cstr(ienv);
 499         build_path_cstr(ienv);
 500
 501         lower_bound = co_get_lower_bound(ienv->co) - co_get_inevit_copy_costs(ienv->co);
 502         lpp_set_bound(ienv->lp, lower_bound);
 503         lpp_set_time_limit(ienv->lp, lenv->time_limit);
 504 }
 505
 506 static void ilp2_apply(ilp_env_t *ienv)
 507 {
 508         local_env_t *lenv = ienv->env;
 509         int i;
 510
 511         /* first check if there was sth. to optimize */
 512         if (lenv->first_x_var >= 0) {
 513                 int              count = lenv->last_x_var - lenv->first_x_var + 1;
 514                 double          *sol   = XMALLOCN(double, count);
 515                 lpp_sol_state_t  state = lpp_get_solution(ienv->lp, sol, lenv->first_x_var, lenv->last_x_var);
 516
 517                 if (state != lpp_optimal) {
 518                         printf("WARNING %s: Solution state is not 'optimal': %d\n", ienv->co->name, (int)state);
 519                         assert(state >= lpp_feasible && "The solution should at least be feasible!");
 520                 }
 521
 522                 for (i=0; i<count; ++i) {
 523                         int nodenr, color;
 524                         char var_name[16];
 525
 526                         if (sol[i] > 1-EPSILON) { /* split variable name into components */
 527                                 lpp_get_var_name(ienv->lp, lenv->first_x_var+i, var_name, sizeof(var_name));
 528
 529                                 if (sscanf(var_name, "x_%d_%d", &nodenr, &color) == 2) {
 530                                         ir_node *irn = pmap_get(lenv->nr_2_irn, INT_TO_PTR(nodenr));
 531                                         assert(irn && "This node number must be present in the map");
 532
 533                                         set_irn_col(ienv->co, irn, color);
 534                                 } else
 535                                         assert(0 && "This should be a x-var");
 536                         }
 537                 }
 538
 539                 xfree(sol);
 540         }
 541
 542 #ifdef COPYOPT_STAT
 543         /* TODO adapt to multiple possible ILPs */
 544         copystat_add_ilp_time((int)(1000.0*lpp_get_sol_time(pi->curr_lp)));  //now we have ms
 545         copystat_add_ilp_vars(lpp_get_var_count(pi->curr_lp));
 546         copystat_add_ilp_csts(lpp_get_cst_count(pi->curr_lp));
 547         copystat_add_ilp_iter(lpp_get_iter_cnt(pi->curr_lp));
 548 #endif
 549 }
 550
 551 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyilp2)
 552 void be_init_copyilp2(void)
 553 {
 554         static co_algo_info copyheur = {
 555                 co_solve_ilp2, 1
 556         };
 557
 558         be_register_copyopt("ilp", &copyheur);
 559 }
 560
 561 int co_solve_ilp2(copy_opt_t *co)
 562 {
 563         lpp_sol_state_t sol_state;
 564         ilp_env_t *ienv;
 565         local_env_t my;
 566
 567         ASSERT_OU_AVAIL(co); //See build_clique_st
 568         ASSERT_GS_AVAIL(co);
 569
 570         my.time_limit  = 0;
 571         my.first_x_var = -1;
 572         my.last_x_var  = -1;
 573         my.nr_2_irn    = pmap_create();
 574         FIRM_DBG_REGISTER(my.dbg, "firm.be.coilp2");
 575
 576         my.normal_colors = bitset_alloca(arch_register_class_n_regs(co->cls));
 577         bitset_clear_all(my.normal_colors);
 578         be_put_allocatable_regs(co->irg, co->cls, my.normal_colors);
 579         my.n_colors = bitset_popcount(my.normal_colors);
 580
 581         ienv = new_ilp_env(co, ilp2_build, ilp2_apply, &my);
 582
 583         sol_state = ilp_go(ienv);
 584
 585         pmap_destroy(my.nr_2_irn);
 586         free_ilp_env(ienv);
 587
 588         return sol_state == lpp_optimal;
 589 }