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

   1 /**
   2  * Author:      Daniel Grund
   3  * Date:                17.05.2005
   4  * Copyright:   (c) Universitaet Karlsruhe
   5  * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
   6  */
   7 #ifdef HAVE_CONFIG_H
   8 #include "config.h"
   9 #endif
  10 #ifdef HAVE_ALLOCA_H
  11 #include <alloca.h>
  12 #endif
  13 #ifdef HAVE_MALLOC_H
  14 #include <malloc.h>
  15 #endif
  16
  17 #define PATH_CONSTRAINTS_FOR_CLASSES
  18 #undef PRECOLOR_MAX_CLIQUE
  19 #undef NO_NULL_COLORS
  20 #undef NO_NULL_COLORS_EXTRA_CSTS
  21 #undef NO_NULL_COLORS_WITH_COSTS
  22 #if (defined(NO_NULL_COLORS_EXTRA_CSTS) || defined(NO_NULL_COLORS_WITH_COSTS)) && !defined(NO_NULL_COLORS)
  23 #error Chose your weapon!
  24 #endif
  25
  26 #include "irprog.h"
  27
  28 #include <lpp/lpp.h>
  29 #include <lpp/lpp_net.h>
  30 #include <lpp/lpp_cplex.h>
  31 #include <lpp/lpp_remote.h>
  32 #include "xmalloc.h"
  33 #include "pset.h"
  34 #include "irdom_t.h"
  35 #include "iredges_t.h"
  36 #include "bechordal_t.h"
  37 #include "becopyopt.h"
  38 #include "becopystat.h"
  39 #include "besched_t.h"
  40 #include "phiclass.h"
  41
  42 #define LPP_HOST "i44pc52"
  43 #define LPP_SOLVER "cplex"
  44
  45 #undef DUMP_MPS
  46 static firm_dbg_module_t *dbg = NULL;
  47
  48 #define MAX(a,b) ((a<b)?(b):(a))
  49 #define MIN(a,b) ((a<b)?(a):(b))
  50 #define EPSILON 0.00001
  51 #define SLOTS_LIVING 32
  52
  53 typedef struct _simpl_t {
  54         struct list_head chain;
  55         if_node_t *ifn;
  56 } simpl_t;
  57
  58 typedef struct _problem_instance_t {
  59         const copy_opt_t *co;                   /** the copy_opt problem */
  60         /* problem size reduction removing simple nodes */
  61         struct list_head simplicials;   /**< holds all simpl_t's in right order to color*/
  62         pset *removed;                                  /**< holds all removed simplicial irns */
  63         /* lp problem */
  64         lpp_t *curr_lp;                                 /**< points to the problem currently used */
  65         lpp_t *dilp;                                    /**< problem formulation directly as milp */
  66 #ifdef NO_NULL_COLORS_EXTRA_CSTS
  67         int first_nnc_cst_idx;                  /**< the first index of a constraint belonging to no-null-colors stuff*/
  68 #endif
  69         int first_nnc_var_idx;                  /**< the first index of a constraint belonging to no-null-colors stuff*/
  70
  71         int cst_counter, first_x_var, last_x_var;
  72         char buf[32];
  73         int all_simplicial;
  74         pset *done;
  75 } problem_instance_t;
  76
  77 #define is_removed(irn) pset_find_ptr(pi->removed, irn)
  78
  79 #define is_color_possible(irn,color) arch_reg_is_allocatable(get_arch_env(pi->co), irn, arch_pos_make_out(0), arch_register_for_index(pi->co->chordal_env->cls, color))
  80
  81 /*
  82  * Some stuff for variable name handling.
  83  */
  84 #define mangle_cst(buf, prefix, nr) \
  85                         snprintf((buf), sizeof(buf), "%c%d", (prefix), (nr))
  86
  87 #define mangle_var1(buf, prefix, color) \
  88                         snprintf((buf), sizeof(buf), "%c%d", (prefix), (color))
  89
  90 #define mangle_var2(buf, prefix, node_nr, color) \
  91                         snprintf((buf), sizeof(buf), "%c%d_%d", (prefix), (node_nr), (color))
  92
  93 #define mangle_var3(buf, prefix, n1, n2, col) \
  94                         snprintf((buf), sizeof(buf), "%c%d_%d_%d", (prefix), (n1), (n2), (col))
  95
  96 #define mangle_var_irn(buf, prefix, irn, color) \
  97                         mangle_var2((buf), (prefix), get_irn_graph_nr(irn), (color))
  98
  99 #define split_var(var, nnr, col) \
 100                         sscanf(var, "x%d_%d", (nnr), (col))
 101
 102
 103 /**
 104  * Checks if a node is simplicial in the graph
 105  * heeding the already removed nodes.
 106  */
 107 static INLINE int pi_is_simplicial(problem_instance_t *pi, const if_node_t *ifn) {
 108         int i, o, size = 0;
 109         if_node_t **all, *curr;
 110         all = alloca(ifn_get_degree(ifn) * sizeof(*all));
 111
 112         /* get all non-removed neighbors */
 113         foreach_neighb(ifn, curr)
 114                 if (!is_removed(curr))
 115                         all[size++] = curr;
 116
 117         /* check if these form a clique */
 118         for (i=0; i<size; ++i)
 119                 for (o=i+1; o<size; ++o)
 120                         if (!ifg_has_edge(pi->co->chordal_env, all[i], all[o]))
 121                                 return 0;
 122
 123         /* all edges exist so this is a clique */
 124         return 1;
 125 }
 126
 127 /**
 128  * Iterative finds and 'removes' from the graph all nodes which are
 129  * simplicial AND not member of a equal-color-wish
 130  */
 131 static void pi_find_simplicials(problem_instance_t *pi) {
 132         set *if_nodes;
 133         if_node_t *ifn;
 134         int redo = 1;
 135
 136         DBG((dbg, LEVEL_2, "Find simlicials...\n"));
 137
 138         if_nodes = be_ra_get_ifg_nodes(pi->co->chordal_env);
 139         while (redo) {
 140                 redo = 0;
 141                 for (ifn = set_first(if_nodes); ifn; ifn = set_next(if_nodes)) {
 142                         ir_node *irn = get_irn_for_graph_nr(get_irg(pi->co), ifn->nnr);
 143                         if (!is_removed(irn) && !is_optimizable(get_arch_env(pi->co), irn) && !is_optimizable_arg(pi->co, irn)) {
 144                         if (pi_is_simplicial(pi, ifn)) {
 145                                         simpl_t *s = xmalloc(sizeof(*s));
 146                                         s->ifn = ifn;
 147                                         list_add(&s->chain, &pi->simplicials);
 148                                         pset_insert_ptr(pi->removed, irn);
 149                                         redo = 1;
 150                                         DBG((dbg, LEVEL_2, " Removed %n %d\n", irn, get_irn_graph_nr(irn)));
 151                         }
 152                         }
 153                 }
 154         }
 155         if (set_count(be_ra_get_ifg_nodes(pi->co->chordal_env)) == pset_count(pi->removed))
 156                 pi->all_simplicial = 1;
 157 }
 158
 159 #ifdef NO_NULL_COLORS
 160 static void pi_add_constr_no_null_colors(problem_instance_t *pi) {
 161         int cst_counter=0, col, var_idx, cst_idx;
 162         int n_colors = pi->co->chordal_env->cls->n_regs;
 163         char buf[40];
 164
 165         for (col = 0; col < n_colors; ++col) {
 166                 mangle_var1(buf, 'u', col);
 167 #ifdef NO_NULL_COLORS_WITH_COSTS
 168                 var_idx = lpp_add_var(pi->curr_lp, buf, lpp_binary, 1.0 / (double) (1 << (col+1)) );
 169 #else
 170                 var_idx = lpp_add_var(pi->curr_lp, buf, lpp_binary, 1.0 / (2.0 * n_colors) );
 171 #endif
 172                 if (!pi->first_nnc_var_idx)
 173                         pi->first_nnc_var_idx = var_idx;
 174         }
 175
 176 #ifdef NO_NULL_COLORS_EXTRA_CSTS
 177         for (col = 0; col < n_colors; ++col) {
 178                 mangle_cst(buf, 'U', cst_counter++);
 179                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_greater, 0);
 180                 if (!pi->first_nnc_cst_idx)
 181                         pi->first_nnc_cst_idx = cst_idx;
 182                 lpp_set_factor_fast(pi->curr_lp, cst_idx, pi->first_nnc_var_idx+col, -1);
 183         }
 184 #endif
 185
 186 #ifndef NO_NULL_COLORS_WITH_COSTS
 187         for (col = 0; col < n_colors - 1; ++col) {
 188                 mangle_cst(buf, 'U', cst_counter++);
 189                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_greater, 0);
 190                 lpp_set_factor_fast(pi->curr_lp, cst_idx, pi->first_nnc_var_idx+col  ,  1);
 191                 lpp_set_factor_fast(pi->curr_lp, cst_idx, pi->first_nnc_var_idx+col+1, -1);
 192         }
 193 #endif
 194
 195 }
 196 #endif
 197
 198 /**
 199  * Add coloring-force conditions
 200  * Matrix A: knapsack constraint for each node
 201  */
 202 static void pi_add_constr_A(problem_instance_t *pi) {
 203         pmap_entry *pme;
 204
 205         DBG((dbg, LEVEL_2, "Add A constraints...\n"));
 206         /* iterate over all blocks */
 207         pmap_foreach(pi->co->chordal_env->border_heads, pme) {
 208                 struct list_head *head = pme->value;
 209                 border_t *curr;
 210                 bitset_t *pos_regs = bitset_alloca(pi->co->chordal_env->cls->n_regs);
 211
 212                 list_for_each_entry_reverse(border_t, curr, head, list)
 213                         if (curr->is_def && curr->is_real && !is_removed(curr->irn)) {
 214                                 int cst_idx, nnr, col;
 215
 216                                 nnr = get_irn_graph_nr(curr->irn);
 217                                 mangle_cst(pi->buf, 'A', nnr);
 218                                 cst_idx = lpp_add_cst(pi->curr_lp, pi->buf, lpp_equal, 1);
 219
 220                                 /* iterate over all possible colors in order */
 221                                 bitset_clear_all(pos_regs);
 222                                 arch_get_allocatable_regs(get_arch_env(pi->co), curr->irn, arch_pos_make_out(0), pi->co->chordal_env->cls, pos_regs);
 223                                 bitset_foreach(pos_regs, col) {
 224                                         int var_idx;
 225                                         mangle_var2(pi->buf, 'x', nnr, col);
 226                                         var_idx = lpp_add_var(pi->curr_lp, pi->buf, lpp_binary, 0);
 227                                         if (!pi->first_x_var)
 228                                                 pi->first_x_var = var_idx;
 229                                         pi->last_x_var = var_idx;
 230                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, var_idx, 1);
 231 #ifdef NO_NULL_COLORS_EXTRA_CSTS
 232                                         lpp_set_factor_fast(pi->curr_lp, pi->first_nnc_cst_idx+col, var_idx, 1);
 233 #endif
 234                                 }
 235                         }
 236         }
 237 }
 238
 239 /**
 240  * Checks if all nodes in @p living are live in in block @p block.
 241  * @return 1 if all are live in
 242  *         0 else
 243  */
 244 static INLINE int all_live_in(ir_node *block, pset *living) {
 245         ir_node *n;
 246         for (n = pset_first(living); n; n = pset_next(living))
 247                 if (!is_live_in(block, n)) {
 248                         pset_break(living);
 249                         return 0;
 250                 }
 251         return 1;
 252 }
 253
 254 /**
 255  * Finds cliques in the interference graph, considering only nodes
 256  * for which the color @p color is possible. Finds only 'maximal-cliques',
 257  * viz cliques which are not contained in another one.
 258  * Matrix B: interference constraints using cliques
 259  */
 260 static void pi_add_constr_B(problem_instance_t *pi, int color) {
 261         enum phase_t {growing, shrinking} phase = growing;
 262         border_t *b;
 263         pmap_entry *pme;
 264         pset *living = pset_new_ptr(SLOTS_LIVING);
 265
 266         DBG((dbg, LEVEL_2, "Add B constraints (col = %d)...\n", color));
 267         /* iterate over all blocks */
 268         pmap_foreach(pi->co->chordal_env->border_heads, pme) {
 269                 ir_node *block = pme->key;
 270                 struct list_head *head = pme->value;
 271
 272                 list_for_each_entry_reverse(border_t, b, head, list) {
 273                         const ir_node *irn = b->irn;
 274                         if (is_removed(irn) || !is_color_possible(irn, color))
 275                                 continue;
 276
 277                         if (b->is_def) {
 278                                 DBG((dbg, LEVEL_2, "Def %n\n", irn));
 279                                 pset_insert_ptr(living, irn);
 280                                 phase = growing;
 281                         } else { /* is_use */
 282                                 DBG((dbg, LEVEL_2, "Use %n\n", irn));
 283
 284                                 /* before shrinking the set, store the current 'maximum' clique;
 285                                  * do NOT if clique is a single node
 286                                  * do NOT if all values are live_in (in this case they were contained in a live-out clique elsewhere) */
 287                                 if (phase == growing && pset_count(living) >= 2 && !all_live_in(block, living)) {
 288                                         int cst_idx;
 289                                         ir_node *n;
 290                                         mangle_cst(pi->buf, 'B', pi->cst_counter);
 291 #ifdef NO_NULL_COLORS
 292                                         cst_idx = lpp_add_cst(pi->curr_lp, pi->buf, lpp_less, 0);
 293 #else
 294                                         cst_idx = lpp_add_cst(pi->curr_lp, pi->buf, lpp_less, 1);
 295 #endif
 296                                         for (n = pset_first(living); n; n = pset_next(living)) {
 297                                                 int var_idx;
 298                                                 mangle_var_irn(pi->buf, 'x', n, color);
 299                                                 var_idx = lpp_get_var_idx(pi->curr_lp, pi->buf);
 300                                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, var_idx, 1);
 301                                         }
 302 #ifdef NO_NULL_COLORS
 303                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, pi->first_nnc_var_idx+color, -1.0);
 304 #endif
 305                                         pi->cst_counter++;
 306                                 }
 307                                 pset_remove_ptr(living, irn);
 308                                 phase = shrinking;
 309                         }
 310                 }
 311         }
 312         assert(0 == pset_count(living));
 313         del_pset(living);
 314 }
 315
 316 /**
 317  * Generates constraints which interrelate x with y variables.
 318  * x1 and x2 have the different colors ==> y_12 = 1
 319  */
 320 static void pi_add_constr_E(problem_instance_t *pi) {
 321         unit_t *curr;
 322         bitset_t *root_regs, *arg_regs, *work_regs;
 323         int cst_counter = 0;
 324         unsigned nregs = pi->co->chordal_env->cls->n_regs;
 325         root_regs = bitset_alloca(nregs);
 326         arg_regs = bitset_alloca(nregs);
 327         work_regs = bitset_alloca(nregs);
 328
 329         DBG((dbg, LEVEL_2, "Add E constraints...\n"));
 330         /* for all roots of optimization units */
 331         list_for_each_entry(unit_t, curr, &pi->co->units, units) {
 332                 ir_node *root, *arg;
 333                 int rootnr, argnr, color;
 334                 int y_idx, i;
 335                 char buf[32];
 336
 337                 root = curr->nodes[0];
 338                 rootnr = get_irn_graph_nr(root);
 339                 bitset_clear_all(root_regs);
 340                 arch_get_allocatable_regs(get_arch_env(pi->co), root, arch_pos_make_out(0), pi->co->chordal_env->cls, root_regs);
 341
 342                 /* for all arguments of root */
 343                 for (i = 1; i < curr->node_count; ++i) {
 344                         arg = curr->nodes[i];
 345                         argnr = get_irn_graph_nr(arg);
 346                         bitset_clear_all(arg_regs);
 347                         arch_get_allocatable_regs(get_arch_env(pi->co), arg, arch_pos_make_out(0), pi->co->chordal_env->cls, arg_regs);
 348
 349                         /* Introduce new variable and set factor in objective function */
 350                         mangle_var2(buf, 'y', rootnr, argnr);
 351                         y_idx = lpp_add_var(pi->curr_lp, buf, lpp_binary, curr->costs[i]);
 352
 353                         /* set starting value */
 354                         lpp_set_start_value(pi->curr_lp, y_idx, (get_irn_col(pi->co, root) != get_irn_col(pi->co, arg)));
 355
 356                         /* For all colors root and arg have in common, add 2 constraints to E */
 357                         bitset_copy(work_regs, root_regs);
 358                         bitset_and(work_regs, arg_regs);
 359                         bitset_foreach(work_regs, color) {
 360                                 int root_idx, arg_idx, cst_idx;
 361                                 mangle_var2(buf, 'x', rootnr, color);
 362                                 root_idx = lpp_get_var_idx(pi->curr_lp, buf);
 363                                 mangle_var2(buf, 'x', argnr, color);
 364                                 arg_idx = lpp_get_var_idx(pi->curr_lp, buf);
 365
 366                                 /* add root-arg-y <= 0 */
 367                                 mangle_cst(buf, 'E', cst_counter++);
 368                                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_less, 0);
 369                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, root_idx, 1);
 370                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, arg_idx, -1);
 371                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, -1);
 372
 373                                 /* add arg-root-y <= 0 */
 374                                 mangle_cst(buf, 'E', cst_counter++);
 375                                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_less, 0);
 376                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, root_idx, -1);
 377                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, arg_idx, 1);
 378                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, -1);
 379                         }
 380                         /* For all colors root and arg have "disjunct", add 1 constraints to E.
 381                          * If root gets a color the arg is not possible to get then they will
 382                          * definetly get different colors. So y has to be 1.
 383                          * Vice versa for arg.
 384                          */
 385                         bitset_copy(work_regs, root_regs);
 386                         bitset_xor(work_regs, arg_regs);
 387                         bitset_foreach(work_regs, color) {
 388                                 int root_idx, arg_idx, cst_idx;
 389                                 mangle_var2(buf, 'x', rootnr, color);
 390                                 root_idx = lpp_get_var_idx(pi->curr_lp, buf);
 391                                 mangle_var2(buf, 'x', argnr, color);
 392                                 arg_idx = lpp_get_var_idx(pi->curr_lp, buf);
 393
 394                                 mangle_cst(buf, 'E', cst_counter++);
 395                                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_less, 0);
 396                                 if (bitset_is_set(root_regs, color)) {
 397                                         /* add root-y <= 0 */
 398                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, root_idx, 1);
 399                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, -1);
 400                                 } else {
 401                                         assert(bitset_is_set(arg_regs, color) && "bitset_xor is buggy");
 402                                         /* add arg-y <= 0 */
 403                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, arg_idx, 1);
 404                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, -1);
 405                                 }
 406                         }
 407                 }
 408         }
 409 }
 410
 411 static INLINE int get_costs(problem_instance_t *pi, ir_node *phi, ir_node *irn) {
 412         int i;
 413         unit_t *curr;
 414         /* search optimization unit for phi */
 415         list_for_each_entry(unit_t, curr, &pi->co->units, units)
 416                 if (curr->nodes[0] == phi) {
 417                         for (i=1; i<curr->node_count; ++i)
 418                                 if (curr->nodes[i] == irn)
 419                                         return curr->costs[i];
 420                         assert(0 && "irn must occur in this ou");
 421                 }
 422         assert(0 && "phi must be found in a ou");
 423         return 0;
 424 }
 425
 426 static void clique_path_walker(ir_node *block, void *env) {
 427         problem_instance_t *pi = env;
 428         int count, arity, row, col, other_row, *costs;
 429         ir_node **phis, *phi, *irn, **phi_matrix;
 430         pset *done;
 431         bitset_t *candidates;
 432
 433         /* Count all phi nodes of this block */
 434         for (count=0, irn = sched_first(block); is_Phi(irn); irn = sched_next(irn))
 435                 count++;
 436
 437         /* We at least 2 phi nodes for this class of inequalities */
 438         if (count < 2)
 439                 return;
 440
 441         /* Build the \Phi-Matrix */
 442         arity = get_irn_arity(sched_first(block));
 443         phis = alloca(count * sizeof(*phis));
 444         costs = alloca(count * sizeof(costs));
 445         phi_matrix = alloca(count*arity * sizeof(*phi_matrix));
 446         candidates = bitset_alloca(count);
 447
 448         phi = sched_first(block);
 449         for (row=0; row<count; ++row) {
 450                 phis[row] = phi;
 451                 for (col=0; col<arity; ++col) {
 452                         ir_node *arg = get_irn_n(phi, col);
 453                         /* Sort out all arguments interfering with its phi */
 454                         if (nodes_interfere(pi->co->chordal_env, phi, arg)) {
 455                                 phi_matrix[row*arity + col] =  NULL;
 456                         } else
 457                                 phi_matrix[row*arity + col] =  arg;
 458                 }
 459                 phi = sched_next(phi);
 460         }
 461
 462         /* Now find the interesting patterns in the matrix:
 463          * All nodes which are used at least twice in a column. */
 464         /* columnwise ... */
 465         for (col=0; col<arity; ++col) {
 466                 done = pset_new_ptr_default();
 467                 for (row=0; row<count; ++row) {
 468                         irn = phi_matrix[row*arity + col];
 469                         /*
 470                          * is this an interfering arg (NULL)
 471                          * or has the irn already been processed in this col?
 472                          */
 473                         if (!irn || pset_find_ptr(done, irn))
 474                                 continue;
 475                         else
 476                                 pset_insert_ptr(done, irn);
 477
 478                         /* insert irn in candidates */
 479                         bitset_clear_all(candidates);
 480                         bitset_set(candidates, row);
 481                         /* search the irn in the rows below */
 482                         for (other_row = row+1; other_row<count; ++other_row)
 483                                 if (irn == phi_matrix[other_row*arity + col]) {
 484                                         /* found the irn in the same col in another row */
 485                                         bitset_set(candidates, other_row);
 486                                 }
 487
 488                         /* now we know all occurences of irn in this col */
 489                         if (bitset_popcnt(candidates) < 2)
 490                                 continue;
 491
 492                         /* generate an unequation finally.
 493                          * phis are indexed in the bitset,
 494                          * shared argument is irn
 495                          * rhs is phi_count - 1 */
 496                         {
 497                                 char buf[32];
 498                                 ir_node *root;
 499                                 int pos, irnnr, rootnr, cst_idx, y_idx, cst_counter = 0;
 500                                 int minimal_unequal_count = bitset_popcnt(candidates)-1;
 501
 502                                 irnnr = get_irn_graph_nr(irn);
 503                                 mangle_cst(buf, 'M', cst_counter++);
 504                                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_greater, minimal_unequal_count);
 505
 506                                 /* for all phis */
 507                                 bitset_foreach(candidates, pos) {
 508                                         root = phis[pos];
 509                                         rootnr = get_irn_graph_nr(root);
 510                                         mangle_var2(buf, 'y', rootnr, irnnr);
 511                                         y_idx = lpp_get_var_idx(pi->curr_lp, buf);
 512                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, 1);
 513                                 }
 514                         }
 515                 }
 516                 del_pset(done); /* clear set for next row */
 517         } /*next col*/
 518 }
 519
 520 /**
 521  * Matrix M: Multi-Arg-Use. Interrelates different \phi-functions
 522  * in the same block, iff they use the same arg at the same pos.
 523  * Only one of the phis can get the arg.
 524  */
 525 static void pi_add_clique_path_cstr(problem_instance_t *pi) {
 526         DBG((dbg, LEVEL_2, "Adding clique path constraints...\n"));
 527         dom_tree_walk_irg(get_irg(pi->co), clique_path_walker, NULL, pi);
 528 }
 529
 530 #ifndef PATH_CONSTRAINTS_FOR_CLASSES
 531 /**
 532  * Matrix P: Path contraints.
 533  * If 2 nodes interfere and there is a path of equal-color-edges
 534  * connecting them, then at least one of those equal-color-edges
 535  * will break and cause some costs.
 536  */
 537 static void pi_add_path_cstr(problem_instance_t *pi) {
 538         unit_t *curr;
 539         int cst_counter = 0;
 540         DBG((dbg, LEVEL_2, "Adding path constraints...\n"));
 541
 542         /* for all optimization units (only phis) */
 543         list_for_each_entry(unit_t, curr, &pi->co->units, units) {
 544                 int i, o, rootnr;
 545
 546                 if (curr->min_nodes_costs == 0)
 547                         continue;
 548
 549                 rootnr = get_irn_graph_nr(curr->nodes[0]);
 550                 /* check all argument pairs for interference */
 551                 for (i=1; i<curr->node_count; ++i) {
 552                         const ir_node *arg1 = curr->nodes[i];
 553                         int arg1nr = get_irn_graph_nr(arg1);
 554                         for (o=i+1; o<curr->node_count; ++o) {
 555                                 const ir_node *arg2 = curr->nodes[o];
 556                                 int arg2nr = get_irn_graph_nr(arg2);
 557                                 if (nodes_interfere(pi->co->chordal_env, arg1, arg2)) {
 558                                         int cst_idx, y_idx;
 559                                         char buf[32];
 560
 561                                         mangle_cst(buf, 'P', cst_counter++);
 562                                         cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_greater, 1);
 563
 564                                         mangle_var2(buf, 'y', rootnr, arg1nr);
 565                                         y_idx = lpp_get_var_idx(pi->curr_lp, buf);
 566                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, 1);
 567
 568                                         mangle_var2(buf, 'y', rootnr, arg2nr);
 569                                         y_idx = lpp_get_var_idx(pi->curr_lp, buf);
 570                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, 1);
 571                                 }
 572                         }
 573                 }
 574         }
 575 }
 576 #endif
 577
 578 #ifdef PATH_CONSTRAINTS_FOR_CLASSES
 579 static INLINE int get_y_var_idx(problem_instance_t *pi, int nnr1, int nnr2) {
 580         int res;
 581         char buf[30];
 582
 583         mangle_var2(buf, 'y', nnr1, nnr2);
 584         if ((res = lpp_get_var_idx(pi->curr_lp, buf)) != -1)
 585                 return res;
 586
 587         mangle_var2(buf, 'y', nnr2, nnr1);
 588         if ((res = lpp_get_var_idx(pi->curr_lp, buf)) != -1)
 589                 return res;
 590
 591         assert(0 && "One of them must work");
 592 }
 593
 594 static void check_ecc_and_add_cut(problem_instance_t *pi, ir_node **path, int length, pset *remain, ir_node *tgt) {
 595         if (path[length-1] == tgt) { /* we found a path */
 596                 int cst_idx, var_idx, i, nnr1, nnr2;
 597                 char buf[30];
 598
 599                 /* add cut to ilp */
 600                 mangle_cst(buf, 'Q', pi->cst_counter++);
 601                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_greater, 1);
 602
 603                 /* add all vars along the path */
 604                 nnr2 = get_irn_graph_nr(path[0]);
 605                 for (i=1; i<length; ++i) {
 606                         nnr1 = nnr2;
 607                         nnr2 = get_irn_graph_nr(path[i]);
 608                         var_idx = get_y_var_idx(pi, nnr1, nnr2);
 609                         lpp_set_factor_fast(pi->curr_lp, cst_idx, var_idx, 1);
 610                 }
 611         } else { /* try to extend the path */
 612                 be_chordal_env_t *cenv = pi->co->chordal_env;
 613                 const ir_edge_t *edge;
 614                 ir_node *end = path[length-1];
 615                 ir_node **next = alloca(pset_count(remain) * sizeof(*next));
 616                 int i, o, max, next_pos = 0;
 617                 pset *done = pset_new_ptr_default();
 618
 619                 /* find all potential next nodes on path */
 620                 /*  args of phis */
 621                 if (is_Phi(end))
 622                         for(i=0, max=get_irn_arity(end); i<max; ++i) {
 623                                 ir_node *arg = get_irn_n(end, i);
 624                                 if (!pset_find_ptr(done, arg) && pset_find_ptr(remain, arg)) {
 625                                         next[next_pos++] = arg;
 626                                         pset_insert_ptr(done, arg);
 627                                 }
 628                         }
 629                 /*  outs of phis and other nodes */
 630                 foreach_out_edge(end, edge) {
 631                         ir_node *user = edge->src;
 632                         if (is_Phi(user) && !pset_find_ptr(done, user) && pset_find_ptr(remain, user)) {
 633                                 next[next_pos++] = user;
 634                                 pset_insert_ptr(done, user);
 635                         }
 636                 }
 637                 del_pset(done);
 638
 639
 640                 /* delete all potential nodes with interferences to other nodes in the path */
 641                 for (i=0; i<next_pos; ++i) {
 642                         ir_node *nn = next[i];
 643
 644                         /* if next is the tgt, it may interfere with path[0],
 645                          * so skip the first check */
 646                         o = (nn == tgt && length > 1) ? 1 : 0;
 647
 648                         for(; o<length; ++o)
 649                                 if (nodes_interfere(cenv, nn, path[o])) {
 650                                         next[i] = NULL;
 651                                         break;
 652                                 }
 653                 }
 654                 /* now we have all possible nodes in next; impossibles are NULL */
 655
 656                 /* try to finish path with all possible nodes */
 657                 for (i=0; i<next_pos; ++i) {
 658                         if (!next[i]) /* this was an impossible node */
 659                                 continue;
 660
 661                         path[length] = next[i];
 662                         pset_remove_ptr(remain, next[i]);
 663                         check_ecc_and_add_cut(pi, path, length+1, remain, tgt);
 664                         pset_insert_ptr(remain, next[i]);
 665                 }
 666         }
 667 }
 668
 669 static void path_cstr_for_classes_walker(ir_node *irn, void *env) {
 670         problem_instance_t *pi = env;
 671         be_chordal_env_t *cenv;
 672         int i, o, max;
 673         ir_node *m;
 674         pset *class = get_phi_class(irn);
 675         if (!class || pset_find_ptr(pi->done, class))
 676                 return;
 677
 678         pset_insert_ptr(pi->done, class);
 679
 680         /* pset to array */
 681         max = pset_count(class);
 682         ir_node **cls = alloca(max * sizeof(*cls));
 683         for(i=0, m = pset_first(class); m; i++, m = pset_next(class)) {
 684                 DBG((dbg, LEVEL_1, " class member: %+F\n", m));
 685                 cls[i] = m;
 686         }
 687
 688         cenv = pi->co->chordal_env;
 689         for(i=0; i<max; ++i) {
 690                 ir_node **path = alloca(max * sizeof(*path));
 691                 pset *remain = pset_new_ptr(8);
 692                 pset_insert_pset_ptr(remain, class);
 693
 694                 /* add cls[i] to path and remove it from remainder */
 695                 path[0] = cls[i];
 696                 pset_remove_ptr(remain, cls[i]);
 697
 698                 for(o=i+1; o<max; ++o)
 699                         if (nodes_interfere(cenv, cls[i], cls[o]))
 700                                 check_ecc_and_add_cut(pi, path, 1, remain, cls[o]);
 701
 702                 /* insert back into remainder */
 703                 pset_insert_ptr(remain, cls[i]);
 704         }
 705 }
 706
 707
 708 /**
 709  * Matrix P: Path contraints.
 710  * If 2 nodes interfere and there is a path of equal-color-edges
 711  * connecting them, then at least one of those equal-color-edges
 712  * will break and cause some costs.
 713  */
 714 static void pi_add_path_cstr_for_classes(problem_instance_t *pi) {
 715         DBG((dbg, LEVEL_2, "Adding path constraints for phi classes...\n"));
 716         pi->cst_counter = 0;
 717         pi->done = pset_new_ptr_default();
 718         irg_walk_graph(get_irg(pi->co), path_cstr_for_classes_walker, NULL, pi);
 719         del_pset(pi->done);
 720 }
 721 #endif
 722
 723 #ifdef PRECOLOR_MAX_CLIQUE
 724 struct pre_col {
 725         problem_instance_t *pi;
 726         pset **clique;
 727 };
 728
 729 #define has_reg_class(pi,irn) \
 730   (arch_get_irn_reg_class(pi->co->chordal_env->session_env->main_env->arch_env, \
 731                           irn, arch_pos_make_out(0)) == pi->co->chordal_env->cls)
 732
 733 static void preColoringWalker(ir_node *bl, void *env) {
 734         struct pre_col *e = env;
 735         pset **clique = e->clique;
 736         pset *max_clique = clique ? *clique : NULL;
 737         int max = max_clique ? pset_count(max_clique) : 0;
 738         problem_instance_t *pi = e->pi;
 739
 740         int i, n;
 741         pset *live       = pset_new_ptr_default();
 742         ir_node *irn;
 743         irn_live_t *li;
 744
 745         /* as always, bring the live end nodes to life here */
 746         live_foreach(bl, li) {
 747           if(live_is_end(li) && has_reg_class(pi, li->irn)) {
 748             pset_insert_ptr(live, irn);
 749           }
 750         }
 751
 752         sched_foreach_reverse(bl, irn) {
 753                 int pres = pset_count(live);
 754
 755                 if(pres > max) {
 756                         max = pres;
 757                         if(max_clique)
 758                                 del_pset(max_clique);
 759
 760                         max_clique = pset_new_ptr_default();
 761                         pset_insert_pset_ptr(max_clique, live);
 762                 }
 763
 764
 765
 766                 if(has_reg_class(pi, irn))
 767                         pset_remove_ptr(live, irn);
 768
 769                 for(i = 0, n = get_irn_arity(irn); i < n; ++i) {
 770                         ir_node *op = get_irn_n(irn, i);
 771                         if(has_reg_class(pi, op) && !is_Phi(irn))
 772                                 pset_insert_ptr(live, op);
 773                 }
 774         }
 775
 776   del_pset(live);
 777   *clique = max_clique;
 778 }
 779
 780 static void pi_add_constr_preColoring(problem_instance_t *pi) {
 781         ir_node *irn;
 782         int cst_counter, color;
 783         struct pre_col pre_col;
 784
 785         pre_col.clique = NULL;
 786         pre_col.pi = pi;
 787
 788         dom_tree_walk_irg(get_irg(pi->co), preColoringWalker, NULL, &pre_col);
 789
 790         color = 0;
 791         for (irn = pset_first(*pre_col.clique); irn; irn = pset_next(*pre_col.clique)) {
 792                 int cst_idx, var_idx, nnr = get_irn_graph_nr(irn);
 793                 char buf[100];
 794
 795                 mangle_cst(buf, 'K', cst_counter++);
 796                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_equal, 1);
 797
 798                 mangle_var2(buf, 'x', nnr, color++);
 799                 var_idx = lpp_get_var_idx(pi->curr_lp, buf);
 800                 lpp_set_factor_fast(pi->curr_lp, cst_idx, var_idx, 1);
 801         }
 802 }
 803 #endif
 804
 805 /**
 806  * Generate the initial problem matrices and vectors.
 807  */
 808 static problem_instance_t *new_pi(const copy_opt_t *co) {
 809         problem_instance_t *pi;
 810         int col;
 811
 812         DBG((dbg, LEVEL_2, "Generating new instance...\n"));
 813         pi = xcalloc(1, sizeof(*pi));
 814         pi->co = co;
 815         pi->removed = pset_new_ptr_default();
 816         INIT_LIST_HEAD(&pi->simplicials);
 817         pi->dilp     = new_lpp(co->name, lpp_minimize);
 818
 819         /* problem size reduction */
 820         pi_find_simplicials(pi);
 821         if (pi->all_simplicial)
 822                 return pi;
 823
 824         /* built objective and constraints */
 825         pi->curr_lp = pi->dilp;
 826 #ifdef NO_NULL_COLORS
 827         pi_add_constr_no_null_colors(pi);
 828 #endif
 829         pi_add_constr_A(pi);
 830         for (col = 0; col < pi->co->chordal_env->cls->n_regs; ++col)
 831                 pi_add_constr_B(pi, col);
 832         pi_add_constr_E(pi);
 833
 834 #ifdef PATH_CONSTRAINTS_FOR_CLASSES
 835         pi_add_path_cstr_for_classes(pi);
 836 #else
 837         pi_add_path_cstr(pi);
 838 #endif
 839         pi_add_clique_path_cstr(pi);
 840 #ifdef PRECOLOR_MAX_CLIQUE
 841         pi_add_constr_preColoring(pi);
 842 #endif
 843
 844         return pi;
 845 }
 846
 847 /**
 848  * Clean the problem instance
 849  */
 850 static void free_pi(problem_instance_t *pi) {
 851         simpl_t *simpl, *tmp;
 852
 853         DBG((dbg, LEVEL_2, "Free instance...\n"));
 854         free_lpp(pi->dilp);
 855         list_for_each_entry_safe(simpl_t, simpl, tmp, &pi->simplicials, chain)
 856                 free(simpl);
 857         del_pset(pi->removed);
 858         free(pi);
 859 }
 860
 861 /**
 862  * Set starting values for the mip problem according
 863  * to the current coloring of the graph.
 864  */
 865 static void pi_set_start_sol(problem_instance_t *pi) {
 866         int i;
 867         char var_name[64];
 868         DBG((dbg, LEVEL_2, "Set start solution...\n"));
 869         for (i=pi->first_x_var; i<=pi->last_x_var; ++i) {
 870                 int nnr, col;
 871                 double val;
 872                 /* get variable name */
 873                 lpp_get_var_name(pi->curr_lp, i, var_name, sizeof(var_name));
 874                 /* split into components */
 875                 if (split_var(var_name, &nnr, &col) == 2) {
 876                         assert(get_irn_col(pi->co, get_irn_for_graph_nr(get_irg(pi->co), nnr)) != -1);
 877                         val = (get_irn_col(pi->co, get_irn_for_graph_nr(get_irg(pi->co), nnr)) == col) ? 1 : 0;
 878                         lpp_set_start_value(pi->curr_lp, i, val);
 879                 } else {
 880                         fprintf(stderr, "Variable name is: %s\n", var_name);
 881                         assert(0 && "x vars always look like this 'x123_45'");
 882                 }
 883         }
 884 }
 885
 886 /**
 887  * Invoke a solver
 888  */
 889 static void pi_solve_ilp(problem_instance_t *pi) {
 890         pi_set_start_sol(pi);
 891 //      lpp_solve_net(pi->curr_lp, LPP_HOST, LPP_SOLVER);
 892         lpp_solve_cplex(pi->curr_lp);
 893         DBG((dbg, LEVEL_1, "Solution time: %.2f\n", pi->curr_lp->sol_time));
 894 }
 895
 896 /**
 897  * Set the color of all simplicial nodes removed form
 898  * the graph before transforming it to an ilp.
 899  */
 900 static void pi_set_simplicials(problem_instance_t *pi) {
 901         simpl_t *simpl, *tmp;
 902         bitset_t *used_cols = bitset_alloca(arch_register_class_n_regs(pi->co->chordal_env->cls));
 903
 904         DBG((dbg, LEVEL_2, "Set simplicials...\n"));
 905         /* color the simplicial nodes in right order */
 906         list_for_each_entry_safe(simpl_t, simpl, tmp, &pi->simplicials, chain) {
 907                 int free_col;
 908                 ir_node *other_irn, *irn;
 909                 if_node_t *other, *ifn;
 910
 911                 /* get free color by inspecting all neighbors */
 912                 ifn = simpl->ifn;
 913                 irn = get_irn_for_graph_nr(get_irg(pi->co), ifn->nnr);
 914                 bitset_clear_all(used_cols);
 915                 foreach_neighb(ifn, other) {
 916                         other_irn = get_irn_for_graph_nr(get_irg(pi->co), other->nnr);
 917                         if (!is_removed(other_irn)) /* only inspect nodes which are in graph right now */
 918                                 bitset_set(used_cols, get_irn_col(pi->co, other_irn));
 919                 }
 920
 921                 /* now all bits not set are possible colors */
 922                 free_col = bitset_next_clear(used_cols, 0);
 923                 assert(free_col != -1 && "No free color found. This can not be.");
 924                 set_irn_col(pi->co, irn, free_col);
 925                 pset_remove_ptr(pi->removed, irn); /* irn is back in graph again */
 926         }
 927 }
 928
 929 /**
 930  * Sets the colors of irns according to the values of variables
 931  * provided by the solution of the solver.
 932  */
 933 static void pi_apply_solution(problem_instance_t *pi) {
 934         int i;
 935         double *sol;
 936         lpp_sol_state_t state;
 937         DBG((dbg, LEVEL_2, "Applying solution...\n"));
 938
 939 #ifdef DO_STAT
 940         copystat_add_ilp_time((int)(1000.0*lpp_get_sol_time(pi->curr_lp)));  //now we have ms
 941         copystat_add_ilp_vars(lpp_get_var_count(pi->curr_lp));
 942         copystat_add_ilp_csts(lpp_get_cst_count(pi->curr_lp));
 943         copystat_add_ilp_iter(lpp_get_iter_cnt(pi->curr_lp));
 944 #endif
 945
 946         sol = xmalloc((pi->last_x_var - pi->first_x_var + 1) * sizeof(*sol));
 947         state = lpp_get_solution(pi->curr_lp, sol, pi->first_x_var, pi->last_x_var);
 948         if (state != lpp_optimal) {
 949                 printf("WARNING %s: Solution state is not 'optimal': %d\n", pi->co->name, state);
 950                 assert(state >= lpp_feasible && "The solution should at least be feasible!");
 951         }
 952         for (i=0; i<pi->last_x_var - pi->first_x_var + 1; ++i) {
 953                 int nnr, col;
 954                 char var_name[64];
 955
 956                 if (sol[i] > 1-EPSILON) { /* split varibale name into components */
 957                         lpp_get_var_name(pi->curr_lp, pi->first_x_var+i, var_name, sizeof(var_name));
 958                         if (split_var(var_name, &nnr, &col) == 2) {
 959                                 DBG((dbg, LEVEL_2, "Irn %n  Idx %d  Var %s  Val %f\n", get_irn_for_graph_nr(get_irg(pi->co), nnr), i, var_name, sol[i]));
 960                                 DBG((dbg, LEVEL_2, "x%d = %d\n", nnr, col));
 961                                 set_irn_col(pi->co, get_irn_for_graph_nr(get_irg(pi->co), nnr), col);
 962                         } else
 963                                 assert(0 && "This should be a x-var");
 964                 }
 965         }
 966 }
 967
 968 void co_ilp_opt(copy_opt_t *co, double time_limit) {
 969         problem_instance_t *pi;
 970
 971         dbg = firm_dbg_register("ir.be.copyoptilp");
 972         if (!strcmp(co->name, DEBUG_IRG))
 973                 firm_dbg_set_mask(dbg, DEBUG_IRG_LVL_ILP);
 974         else
 975                 firm_dbg_set_mask(dbg, DEBUG_LVL_ILP);
 976
 977         pi = new_pi(co);
 978         if (!pi->all_simplicial) {
 979 #ifdef DUMP_MPS
 980                 char buf[512];
 981                 snprintf(buf, sizeof(buf), "%s.mps", co->name);
 982                 lpp_dump(pi->curr_lp, buf);
 983 #endif
 984                 lpp_set_time_limit(pi->curr_lp, time_limit);
 985                 pi_solve_ilp(pi);
 986                 pi_apply_solution(pi);
 987                 pi_set_simplicials(pi);
 988         }
 989         free_pi(pi);
 990 }