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 #undef NO_NULL_COLORS
  18 #undef NO_NULL_COLORS_EXTRA_CSTS
  19 #undef NO_NULL_COLORS_WITH_COSTS
  20
  21 #if (defined(NO_NULL_COLORS_EXTRA_CSTS) || defined(NO_NULL_COLORS_WITH_COSTS)) && !defined(NO_NULL_COLORS)
  22 #error Chose your weapon!
  23 #endif
  24
  25 #undef PRECOLOR_MAX_CLIQUE
  26 #define PATH_CONSTRAINTS_FOR_CLASSES
  27
  28 #include "irprog.h"
  29
  30 #include <lpp/lpp.h>
  31 #include <lpp/lpp_net.h>
  32 //#include <lpp/lpp_cplex.h>
  33 #include <lpp/lpp_remote.h>
  34 #include "xmalloc.h"
  35 #include "pset.h"
  36 #include "irdom_t.h"
  37 #include "iredges_t.h"
  38 #include "bechordal_t.h"
  39 #include "becopyopt.h"
  40 #include "becopystat.h"
  41 #include "besched_t.h"
  42 #include "phiclass.h"
  43
  44 #define LPP_HOST "i44pc52"
  45 #define LPP_SOLVER "cplex"
  46
  47 #undef DUMP_MPS
  48 static firm_dbg_module_t *dbg = NULL;
  49
  50 #define MAX(a,b) ((a<b)?(b):(a))
  51 #define MIN(a,b) ((a<b)?(a):(b))
  52 #define EPSILON 0.00001
  53 #define SLOTS_LIVING 32
  54
  55 typedef struct _simpl_t {
  56         struct list_head chain;
  57         if_node_t *ifn;
  58 } simpl_t;
  59
  60 typedef struct _problem_instance_t {
  61         const copy_opt_t *co;                   /** the copy_opt problem */
  62         /* problem size reduction removing simple nodes */
  63         struct list_head simplicials;   /**< holds all simpl_t's in right order to color*/
  64         pset *removed;                                  /**< holds all removed simplicial irns */
  65         /* lp problem */
  66         lpp_t *curr_lp;                                 /**< points to the problem currently used */
  67         lpp_t *dilp;                                    /**< problem formulation directly as milp */
  68 #ifdef NO_NULL_COLORS_EXTRA_CSTS
  69         int first_nnc_cst_idx;                  /**< the first index of a constraint belonging to no-null-colors stuff*/
  70 #endif
  71         int first_nnc_var_idx;                  /**< the first index of a constraint belonging to no-null-colors stuff*/
  72
  73         int cst_counter, first_x_var, last_x_var;
  74         char buf[32];
  75         int all_simplicial;
  76         pset *done;
  77 } problem_instance_t;
  78
  79 #define is_removed(irn) pset_find_ptr(pi->removed, irn)
  80
  81 #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))
  82
  83 /*
  84  * Some stuff for variable name handling.
  85  */
  86 #define mangle_cst(buf, prefix, nr) \
  87                         snprintf((buf), sizeof(buf), "%c%d", (prefix), (nr))
  88
  89 #define mangle_var1(buf, prefix, color) \
  90                         snprintf((buf), sizeof(buf), "%c%d", (prefix), (color))
  91
  92 #define mangle_var2(buf, prefix, node_nr, color) \
  93                         snprintf((buf), sizeof(buf), "%c%d_%d", (prefix), (node_nr), (color))
  94
  95 #define mangle_var3(buf, prefix, n1, n2, col) \
  96                         snprintf((buf), sizeof(buf), "%c%d_%d_%d", (prefix), (n1), (n2), (col))
  97
  98 #define mangle_var_irn(buf, prefix, irn, color) \
  99                         mangle_var2((buf), (prefix), get_irn_graph_nr(irn), (color))
 100
 101 #define split_var(var, nnr, col) \
 102                         sscanf(var, "x%d_%d", (nnr), (col))
 103
 104
 105 /**
 106  * Checks if a node is simplicial in the graph
 107  * heeding the already removed nodes.
 108  */
 109 static INLINE int pi_is_simplicial(problem_instance_t *pi, const if_node_t *ifn) {
 110         int i, o, size = 0;
 111         if_node_t **all, *curr;
 112         all = alloca(ifn_get_degree(ifn) * sizeof(*all));
 113
 114         /* get all non-removed neighbors */
 115         foreach_neighb(ifn, curr)
 116                 if (!is_removed(curr))
 117                         all[size++] = curr;
 118
 119         /* check if these form a clique */
 120         for (i=0; i<size; ++i)
 121                 for (o=i+1; o<size; ++o)
 122                         if (!ifg_has_edge(pi->co->chordal_env, all[i], all[o]))
 123                                 return 0;
 124
 125         /* all edges exist so this is a clique */
 126         return 1;
 127 }
 128
 129 /**
 130  * Iterative finds and 'removes' from the graph all nodes which are
 131  * simplicial AND not member of a equal-color-wish
 132  */
 133 static void pi_find_simplicials(problem_instance_t *pi) {
 134         set *if_nodes;
 135         if_node_t *ifn;
 136         int redo = 1;
 137
 138         DBG((dbg, LEVEL_2, "Find simlicials...\n"));
 139
 140         if_nodes = be_ra_get_ifg_nodes(pi->co->chordal_env);
 141         while (redo) {
 142                 redo = 0;
 143                 for (ifn = set_first(if_nodes); ifn; ifn = set_next(if_nodes)) {
 144                         ir_node *irn = get_irn_for_graph_nr(get_irg(pi->co), ifn->nnr);
 145                         if (!is_removed(irn) && !is_optimizable(get_arch_env(pi->co), irn) && !is_optimizable_arg(pi->co, irn)) {
 146                         if (pi_is_simplicial(pi, ifn)) {
 147                                         simpl_t *s = xmalloc(sizeof(*s));
 148                                         s->ifn = ifn;
 149                                         list_add(&s->chain, &pi->simplicials);
 150                                         pset_insert_ptr(pi->removed, irn);
 151                                         redo = 1;
 152                                         DBG((dbg, LEVEL_2, " Removed %n %d\n", irn, get_irn_graph_nr(irn)));
 153                         }
 154                         }
 155                 }
 156         }
 157         if (set_count(be_ra_get_ifg_nodes(pi->co->chordal_env)) == pset_count(pi->removed))
 158                 pi->all_simplicial = 1;
 159 }
 160
 161 #ifdef NO_NULL_COLORS
 162 static void pi_add_constr_no_null_colors(problem_instance_t *pi) {
 163         int cst_counter=0, col, var_idx, cst_idx;
 164         int n_colors = pi->co->chordal_env->cls->n_regs;
 165         char buf[40];
 166
 167         for (col = 0; col < n_colors; ++col) {
 168                 mangle_var1(buf, 'u', col);
 169 #ifdef NO_NULL_COLORS_WITH_COSTS
 170                 var_idx = lpp_add_var(pi->curr_lp, buf, lpp_binary, 1.0 / (double) (1 << (col+1)) );
 171 #else
 172                 var_idx = lpp_add_var(pi->curr_lp, buf, lpp_binary, 1.0 / (2.0 * n_colors) );
 173 #endif
 174                 if (!pi->first_nnc_var_idx)
 175                         pi->first_nnc_var_idx = var_idx;
 176         }
 177
 178 #ifdef NO_NULL_COLORS_EXTRA_CSTS
 179         for (col = 0; col < n_colors; ++col) {
 180                 mangle_cst(buf, 'U', cst_counter++);
 181                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_greater, 0);
 182                 if (!pi->first_nnc_cst_idx)
 183                         pi->first_nnc_cst_idx = cst_idx;
 184                 lpp_set_factor_fast(pi->curr_lp, cst_idx, pi->first_nnc_var_idx+col, -1);
 185         }
 186 #endif
 187
 188 #ifndef NO_NULL_COLORS_WITH_COSTS
 189         for (col = 0; col < n_colors - 1; ++col) {
 190                 mangle_cst(buf, 'U', cst_counter++);
 191                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_greater, 0);
 192                 lpp_set_factor_fast(pi->curr_lp, cst_idx, pi->first_nnc_var_idx+col  ,  1);
 193                 lpp_set_factor_fast(pi->curr_lp, cst_idx, pi->first_nnc_var_idx+col+1, -1);
 194         }
 195 #endif
 196
 197 }
 198 #endif
 199
 200 /**
 201  * Add coloring-force conditions
 202  * Matrix A: knapsack constraint for each node
 203  */
 204 static void pi_add_constr_A(problem_instance_t *pi) {
 205         pmap_entry *pme;
 206
 207         DBG((dbg, LEVEL_2, "Add A constraints...\n"));
 208         /* iterate over all blocks */
 209         pmap_foreach(pi->co->chordal_env->border_heads, pme) {
 210                 struct list_head *head = pme->value;
 211                 border_t *curr;
 212                 bitset_t *pos_regs = bitset_alloca(pi->co->chordal_env->cls->n_regs);
 213
 214                 list_for_each_entry_reverse(border_t, curr, head, list)
 215                         if (curr->is_def && curr->is_real && !is_removed(curr->irn)) {
 216                                 int cst_idx, nnr, col;
 217
 218                                 nnr = get_irn_graph_nr(curr->irn);
 219                                 mangle_cst(pi->buf, 'A', nnr);
 220                                 cst_idx = lpp_add_cst(pi->curr_lp, pi->buf, lpp_equal, 1);
 221
 222                                 /* iterate over all possible colors in order */
 223                                 bitset_clear_all(pos_regs);
 224                                 arch_get_allocatable_regs(get_arch_env(pi->co), curr->irn, arch_pos_make_out(0), pi->co->chordal_env->cls, pos_regs);
 225                                 bitset_foreach(pos_regs, col) {
 226                                         int var_idx;
 227                                         mangle_var2(pi->buf, 'x', nnr, col);
 228                                         var_idx = lpp_add_var(pi->curr_lp, pi->buf, lpp_binary, 0);
 229                                         if (!pi->first_x_var)
 230                                                 pi->first_x_var = var_idx;
 231                                         pi->last_x_var = var_idx;
 232                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, var_idx, 1);
 233 #ifdef NO_NULL_COLORS_EXTRA_CSTS
 234                                         lpp_set_factor_fast(pi->curr_lp, pi->first_nnc_cst_idx+col, var_idx, 1);
 235 #endif
 236                                 }
 237                         }
 238         }
 239 }
 240
 241 /**
 242  * Checks if all nodes in @p living are live in in block @p block.
 243  * @return 1 if all are live in
 244  *         0 else
 245  */
 246 static INLINE int all_live_in(ir_node *block, pset *living) {
 247         ir_node *n;
 248         for (n = pset_first(living); n; n = pset_next(living))
 249                 if (!is_live_in(block, n)) {
 250                         pset_break(living);
 251                         return 0;
 252                 }
 253         return 1;
 254 }
 255
 256 /**
 257  * Finds cliques in the interference graph, considering only nodes
 258  * for which the color @p color is possible. Finds only 'maximal-cliques',
 259  * viz cliques which are not contained in another one.
 260  * Matrix B: interference constraints using cliques
 261  */
 262 static void pi_add_constr_B(problem_instance_t *pi, int color) {
 263         enum phase_t {growing, shrinking} phase = growing;
 264         border_t *b;
 265         pmap_entry *pme;
 266         pset *living = pset_new_ptr(SLOTS_LIVING);
 267
 268         DBG((dbg, LEVEL_2, "Add B constraints (col = %d)...\n", color));
 269         /* iterate over all blocks */
 270         pmap_foreach(pi->co->chordal_env->border_heads, pme) {
 271                 ir_node *block = pme->key;
 272                 struct list_head *head = pme->value;
 273
 274                 list_for_each_entry_reverse(border_t, b, head, list) {
 275                         const ir_node *irn = b->irn;
 276                         if (is_removed(irn) || !is_color_possible(irn, color))
 277                                 continue;
 278
 279                         if (b->is_def) {
 280                                 DBG((dbg, LEVEL_2, "Def %n\n", irn));
 281                                 pset_insert_ptr(living, irn);
 282                                 phase = growing;
 283                         } else { /* is_use */
 284                                 DBG((dbg, LEVEL_2, "Use %n\n", irn));
 285
 286                                 /* before shrinking the set, store the current 'maximum' clique;
 287                                  * do NOT if clique is a single node
 288                                  * do NOT if all values are live_in (in this case they were contained in a live-out clique elsewhere) */
 289                                 if (phase == growing && pset_count(living) >= 2 && !all_live_in(block, living)) {
 290                                         int cst_idx;
 291                                         ir_node *n;
 292                                         mangle_cst(pi->buf, 'B', pi->cst_counter);
 293 #ifdef NO_NULL_COLORS
 294                                         cst_idx = lpp_add_cst(pi->curr_lp, pi->buf, lpp_less, 0);
 295 #else
 296                                         cst_idx = lpp_add_cst(pi->curr_lp, pi->buf, lpp_less, 1);
 297 #endif
 298                                         for (n = pset_first(living); n; n = pset_next(living)) {
 299                                                 int var_idx;
 300                                                 mangle_var_irn(pi->buf, 'x', n, color);
 301                                                 var_idx = lpp_get_var_idx(pi->curr_lp, pi->buf);
 302                                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, var_idx, 1);
 303                                         }
 304 #ifdef NO_NULL_COLORS
 305                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, pi->first_nnc_var_idx+color, -1.0);
 306 #endif
 307                                         pi->cst_counter++;
 308                                 }
 309                                 pset_remove_ptr(living, irn);
 310                                 phase = shrinking;
 311                         }
 312                 }
 313         }
 314         assert(0 == pset_count(living));
 315         del_pset(living);
 316 }
 317
 318 /**
 319  * Generates constraints which interrelate x with y variables.
 320  * x1 and x2 have the different colors ==> y_12 = 1
 321  */
 322 static void pi_add_constr_E(problem_instance_t *pi) {
 323         unit_t *curr;
 324         bitset_t *root_regs, *arg_regs, *work_regs;
 325         int cst_counter = 0;
 326         unsigned nregs = pi->co->chordal_env->cls->n_regs;
 327         root_regs = bitset_alloca(nregs);
 328         arg_regs = bitset_alloca(nregs);
 329         work_regs = bitset_alloca(nregs);
 330
 331         DBG((dbg, LEVEL_2, "Add E constraints...\n"));
 332         /* for all roots of optimization units */
 333         list_for_each_entry(unit_t, curr, &pi->co->units, units) {
 334                 ir_node *root, *arg;
 335                 int rootnr, argnr, color;
 336                 int y_idx, i;
 337                 char buf[32];
 338
 339                 root = curr->nodes[0];
 340                 rootnr = get_irn_graph_nr(root);
 341                 bitset_clear_all(root_regs);
 342                 arch_get_allocatable_regs(get_arch_env(pi->co), root, arch_pos_make_out(0), pi->co->chordal_env->cls, root_regs);
 343
 344                 /* for all arguments of root */
 345                 for (i = 1; i < curr->node_count; ++i) {
 346                         arg = curr->nodes[i];
 347                         argnr = get_irn_graph_nr(arg);
 348                         bitset_clear_all(arg_regs);
 349                         arch_get_allocatable_regs(get_arch_env(pi->co), arg, arch_pos_make_out(0), pi->co->chordal_env->cls, arg_regs);
 350
 351                         /* Introduce new variable and set factor in objective function */
 352                         mangle_var2(buf, 'y', rootnr, argnr);
 353                         y_idx = lpp_add_var(pi->curr_lp, buf, lpp_binary, curr->costs[i]);
 354
 355                         /* set starting value */
 356                         lpp_set_start_value(pi->curr_lp, y_idx, (get_irn_col(pi->co, root) != get_irn_col(pi->co, arg)));
 357
 358                         /* For all colors root and arg have in common, add 2 constraints to E */
 359                         bitset_copy(work_regs, root_regs);
 360                         bitset_and(work_regs, arg_regs);
 361                         bitset_foreach(work_regs, color) {
 362                                 int root_idx, arg_idx, cst_idx;
 363                                 mangle_var2(buf, 'x', rootnr, color);
 364                                 root_idx = lpp_get_var_idx(pi->curr_lp, buf);
 365                                 mangle_var2(buf, 'x', argnr, color);
 366                                 arg_idx = lpp_get_var_idx(pi->curr_lp, buf);
 367
 368                                 /* add root-arg-y <= 0 */
 369                                 mangle_cst(buf, 'E', cst_counter++);
 370                                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_less, 0);
 371                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, root_idx, 1);
 372                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, arg_idx, -1);
 373                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, -1);
 374
 375                                 /* add arg-root-y <= 0 */
 376                                 mangle_cst(buf, 'E', cst_counter++);
 377                                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_less, 0);
 378                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, root_idx, -1);
 379                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, arg_idx, 1);
 380                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, -1);
 381                         }
 382                         /* For all colors root and arg have "disjunct", add 1 constraints to E.
 383                          * If root gets a color the arg is not possible to get then they will
 384                          * definetly get different colors. So y has to be 1.
 385                          * Vice versa for arg.
 386                          */
 387                         bitset_copy(work_regs, root_regs);
 388                         bitset_xor(work_regs, arg_regs);
 389                         bitset_foreach(work_regs, color) {
 390                                 int root_idx, arg_idx, cst_idx;
 391                                 mangle_var2(buf, 'x', rootnr, color);
 392                                 root_idx = lpp_get_var_idx(pi->curr_lp, buf);
 393                                 mangle_var2(buf, 'x', argnr, color);
 394                                 arg_idx = lpp_get_var_idx(pi->curr_lp, buf);
 395
 396                                 mangle_cst(buf, 'E', cst_counter++);
 397                                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_less, 0);
 398                                 if (bitset_is_set(root_regs, color)) {
 399                                         /* add root-y <= 0 */
 400                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, root_idx, 1);
 401                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, -1);
 402                                 } else {
 403                                         assert(bitset_is_set(arg_regs, color) && "bitset_xor is buggy");
 404                                         /* add arg-y <= 0 */
 405                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, arg_idx, 1);
 406                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, -1);
 407                                 }
 408                         }
 409                 }
 410         }
 411 }
 412
 413 static INLINE int get_costs(problem_instance_t *pi, ir_node *phi, ir_node *irn) {
 414         int i;
 415         unit_t *curr;
 416         /* search optimization unit for phi */
 417         list_for_each_entry(unit_t, curr, &pi->co->units, units)
 418                 if (curr->nodes[0] == phi) {
 419                         for (i=1; i<curr->node_count; ++i)
 420                                 if (curr->nodes[i] == irn)
 421                                         return curr->costs[i];
 422                         assert(0 && "irn must occur in this ou");
 423                 }
 424         assert(0 && "phi must be found in a ou");
 425         return 0;
 426 }
 427
 428 static void clique_path_walker(ir_node *block, void *env) {
 429         problem_instance_t *pi = env;
 430         int count, arity, row, col, other_row, *costs;
 431         ir_node **phis, *phi, *irn, **phi_matrix;
 432         pset *done;
 433         bitset_t *candidates;
 434
 435         /* Count all phi nodes of this block */
 436         for (count=0, irn = sched_first(block); is_Phi(irn); irn = sched_next(irn))
 437                 count++;
 438
 439         /* We at least 2 phi nodes for this class of inequalities */
 440         if (count < 2)
 441                 return;
 442
 443         /* Build the \Phi-Matrix */
 444         arity = get_irn_arity(sched_first(block));
 445         phis = alloca(count * sizeof(*phis));
 446         costs = alloca(count * sizeof(costs));
 447         phi_matrix = alloca(count*arity * sizeof(*phi_matrix));
 448         candidates = bitset_alloca(count);
 449
 450         phi = sched_first(block);
 451         for (row=0; row<count; ++row) {
 452                 phis[row] = phi;
 453                 for (col=0; col<arity; ++col) {
 454                         ir_node *arg = get_irn_n(phi, col);
 455                         /* Sort out all arguments interfering with its phi */
 456                         if (nodes_interfere(pi->co->chordal_env, phi, arg)) {
 457                                 phi_matrix[row*arity + col] =  NULL;
 458                         } else
 459                                 phi_matrix[row*arity + col] =  arg;
 460                 }
 461                 phi = sched_next(phi);
 462         }
 463
 464         /* Now find the interesting patterns in the matrix:
 465          * All nodes which are used at least twice in a column. */
 466         /* columnwise ... */
 467         for (col=0; col<arity; ++col) {
 468                 done = pset_new_ptr_default();
 469                 for (row=0; row<count; ++row) {
 470                         irn = phi_matrix[row*arity + col];
 471                         /*
 472                          * is this an interfering arg (NULL)
 473                          * or has the irn already been processed in this col?
 474                          */
 475                         if (!irn || pset_find_ptr(done, irn))
 476                                 continue;
 477                         else
 478                                 pset_insert_ptr(done, irn);
 479
 480                         /* insert irn in candidates */
 481                         bitset_clear_all(candidates);
 482                         bitset_set(candidates, row);
 483                         /* search the irn in the rows below */
 484                         for (other_row = row+1; other_row<count; ++other_row)
 485                                 if (irn == phi_matrix[other_row*arity + col]) {
 486                                         /* found the irn in the same col in another row */
 487                                         bitset_set(candidates, other_row);
 488                                 }
 489
 490                         /* now we know all occurences of irn in this col */
 491                         if (bitset_popcnt(candidates) < 2)
 492                                 continue;
 493
 494                         /* generate an unequation finally.
 495                          * phis are indexed in the bitset,
 496                          * shared argument is irn
 497                          * rhs is phi_count - 1 */
 498                         {
 499                                 char buf[32];
 500                                 ir_node *root;
 501                                 int pos, irnnr, rootnr, cst_idx, y_idx, cst_counter = 0;
 502                                 int minimal_unequal_count = bitset_popcnt(candidates)-1;
 503
 504                                 irnnr = get_irn_graph_nr(irn);
 505                                 mangle_cst(buf, 'M', cst_counter++);
 506                                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_greater, minimal_unequal_count);
 507
 508                                 /* for all phis */
 509                                 bitset_foreach(candidates, pos) {
 510                                         root = phis[pos];
 511                                         rootnr = get_irn_graph_nr(root);
 512                                         mangle_var2(buf, 'y', rootnr, irnnr);
 513                                         y_idx = lpp_get_var_idx(pi->curr_lp, buf);
 514                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, 1);
 515                                 }
 516                         }
 517                 }
 518                 del_pset(done); /* clear set for next row */
 519         } /*next col*/
 520 }
 521
 522 /**
 523  * Matrix M: Multi-Arg-Use. Interrelates different \phi-functions
 524  * in the same block, iff they use the same arg at the same pos.
 525  * Only one of the phis can get the arg.
 526  */
 527 static void pi_add_clique_path_cstr(problem_instance_t *pi) {
 528         DBG((dbg, LEVEL_2, "Adding clique path constraints...\n"));
 529         dom_tree_walk_irg(get_irg(pi->co), clique_path_walker, NULL, pi);
 530 }
 531
 532 #ifndef PATH_CONSTRAINTS_FOR_CLASSES
 533 /**
 534  * Matrix P: Path contraints.
 535  * If 2 nodes interfere and there is a path of equal-color-edges
 536  * connecting them, then at least one of those equal-color-edges
 537  * will break and cause some costs.
 538  */
 539 static void pi_add_path_cstr(problem_instance_t *pi) {
 540         unit_t *curr;
 541         int cst_counter = 0;
 542         DBG((dbg, LEVEL_2, "Adding path constraints...\n"));
 543
 544         /* for all optimization units (only phis) */
 545         list_for_each_entry(unit_t, curr, &pi->co->units, units) {
 546                 int i, o, rootnr;
 547
 548                 if (curr->min_nodes_costs == 0)
 549                         continue;
 550
 551                 rootnr = get_irn_graph_nr(curr->nodes[0]);
 552                 /* check all argument pairs for interference */
 553                 for (i=1; i<curr->node_count; ++i) {
 554                         const ir_node *arg1 = curr->nodes[i];
 555                         int arg1nr = get_irn_graph_nr(arg1);
 556                         for (o=i+1; o<curr->node_count; ++o) {
 557                                 const ir_node *arg2 = curr->nodes[o];
 558                                 int arg2nr = get_irn_graph_nr(arg2);
 559                                 if (nodes_interfere(pi->co->chordal_env, arg1, arg2)) {
 560                                         int cst_idx, y_idx;
 561                                         char buf[32];
 562
 563                                         mangle_cst(buf, 'P', cst_counter++);
 564                                         cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_greater, 1);
 565
 566                                         mangle_var2(buf, 'y', rootnr, arg1nr);
 567                                         y_idx = lpp_get_var_idx(pi->curr_lp, buf);
 568                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, 1);
 569
 570                                         mangle_var2(buf, 'y', rootnr, arg2nr);
 571                                         y_idx = lpp_get_var_idx(pi->curr_lp, buf);
 572                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, 1);
 573                                 }
 574                         }
 575                 }
 576         }
 577 }
 578 #endif
 579
 580 #ifdef PATH_CONSTRAINTS_FOR_CLASSES
 581 static INLINE int get_y_var_idx(problem_instance_t *pi, int nnr1, int nnr2) {
 582         int res;
 583         char buf[30];
 584
 585         mangle_var2(buf, 'y', nnr1, nnr2);
 586         if ((res = lpp_get_var_idx(pi->curr_lp, buf)) != -1)
 587                 return res;
 588
 589         mangle_var2(buf, 'y', nnr2, nnr1);
 590         if ((res = lpp_get_var_idx(pi->curr_lp, buf)) != -1)
 591                 return res;
 592
 593         assert(0 && "One of them must work");
 594 }
 595
 596 static void check_ecc_and_add_cut(problem_instance_t *pi, ir_node **path, int length, pset *remain, ir_node *tgt) {
 597         if (path[length-1] == tgt) { /* we found a path */
 598                 int cst_idx, var_idx, i, nnr1, nnr2;
 599                 char buf[30];
 600
 601                 /* add cut to ilp */
 602                 mangle_cst(buf, 'Q', pi->cst_counter++);
 603                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_greater, 1);
 604
 605                 /* add all vars along the path */
 606                 nnr2 = get_irn_graph_nr(path[0]);
 607                 for (i=1; i<length; ++i) {
 608                         nnr1 = nnr2;
 609                         nnr2 = get_irn_graph_nr(path[i]);
 610                         var_idx = get_y_var_idx(pi, nnr1, nnr2);
 611                         lpp_set_factor_fast(pi->curr_lp, cst_idx, var_idx, 1);
 612                 }
 613         } else { /* try to extend the path */
 614                 be_chordal_env_t *cenv = pi->co->chordal_env;
 615                 const ir_edge_t *edge;
 616                 ir_node *end = path[length-1];
 617                 ir_node **next = alloca(pset_count(remain) * sizeof(*next));
 618                 int i, o, max, next_pos = 0;
 619                 pset *done = pset_new_ptr_default();
 620
 621                 /* find all potential next nodes on path */
 622                 /*  args of phis */
 623                 if (is_Phi(end))
 624                         for(i=0, max=get_irn_arity(end); i<max; ++i) {
 625                                 ir_node *arg = get_irn_n(end, i);
 626                                 if (!pset_find_ptr(done, arg) && pset_find_ptr(remain, arg)) {
 627                                         next[next_pos++] = arg;
 628                                         pset_insert_ptr(done, arg);
 629                                 }
 630                         }
 631                 /*  outs of phis and other nodes */
 632                 foreach_out_edge(end, edge) {
 633                         ir_node *user = edge->src;
 634                         if (is_Phi(user) && !pset_find_ptr(done, user) && pset_find_ptr(remain, user)) {
 635                                 next[next_pos++] = user;
 636                                 pset_insert_ptr(done, user);
 637                         }
 638                 }
 639                 del_pset(done);
 640
 641
 642                 /* delete all potential nodes with interferences to other nodes in the path */
 643                 for (i=0; i<next_pos; ++i) {
 644                         ir_node *nn = next[i];
 645
 646                         /* if next is the tgt, it may interfere with path[0],
 647                          * so skip the first check */
 648                         o = (nn == tgt && length > 1) ? 1 : 0;
 649
 650                         for(; o<length; ++o)
 651                                 if (nodes_interfere(cenv, nn, path[o])) {
 652                                         next[i] = NULL;
 653                                         break;
 654                                 }
 655                 }
 656                 /* now we have all possible nodes in next; impossibles are NULL */
 657
 658                 /* try to finish path with all possible nodes */
 659                 for (i=0; i<next_pos; ++i) {
 660                         if (!next[i]) /* this was an impossible node */
 661                                 continue;
 662
 663                         path[length] = next[i];
 664                         pset_remove_ptr(remain, next[i]);
 665                         check_ecc_and_add_cut(pi, path, length+1, remain, tgt);
 666                         pset_insert_ptr(remain, next[i]);
 667                 }
 668         }
 669 }
 670
 671 static void path_cstr_for_classes_walker(ir_node *irn, void *env) {
 672         problem_instance_t *pi = env;
 673         be_chordal_env_t *cenv;
 674         int i, o, max;
 675         ir_node *m;
 676         pset *class = get_phi_class(irn);
 677         if (!class || pset_find_ptr(pi->done, class))
 678                 return;
 679
 680         pset_insert_ptr(pi->done, class);
 681
 682         /* pset to array */
 683         max = pset_count(class);
 684         ir_node **cls = alloca(max * sizeof(*cls));
 685         for(i=0, m = pset_first(class); m; i++, m = pset_next(class)) {
 686                 DBG((dbg, LEVEL_1, " class member: %+F\n", m));
 687                 cls[i] = m;
 688         }
 689
 690         cenv = pi->co->chordal_env;
 691         for(i=0; i<max; ++i) {
 692                 ir_node **path = alloca(max * sizeof(*path));
 693                 pset *remain = pset_new_ptr(8);
 694                 pset_insert_pset_ptr(remain, class);
 695
 696                 /* add cls[i] to path and remove it from remainder */
 697                 path[0] = cls[i];
 698                 pset_remove_ptr(remain, cls[i]);
 699
 700                 for(o=i+1; o<max; ++o)
 701                         if (nodes_interfere(cenv, cls[i], cls[o]))
 702                                 check_ecc_and_add_cut(pi, path, 1, remain, cls[o]);
 703
 704                 /* insert back into remainder */
 705                 pset_insert_ptr(remain, cls[i]);
 706         }
 707 }
 708
 709
 710 /**
 711  * Matrix P: Path contraints.
 712  * If 2 nodes interfere and there is a path of equal-color-edges
 713  * connecting them, then at least one of those equal-color-edges
 714  * will break and cause some costs.
 715  */
 716 static void pi_add_path_cstr_for_classes(problem_instance_t *pi) {
 717         DBG((dbg, LEVEL_2, "Adding path constraints for phi classes...\n"));
 718         pi->cst_counter = 0;
 719         pi->done = pset_new_ptr_default();
 720         irg_walk_graph(get_irg(pi->co), path_cstr_for_classes_walker, NULL, pi);
 721         del_pset(pi->done);
 722 }
 723 #endif
 724
 725 #ifdef PRECOLOR_MAX_CLIQUE
 726 struct pre_col {
 727         problem_instance_t *pi;
 728         pset **clique;
 729 };
 730
 731 #define has_reg_class(pi,irn) \
 732   (arch_get_irn_reg_class(pi->co->chordal_env->session_env->main_env->arch_env, \
 733                           irn, arch_pos_make_out(0)) == pi->co->chordal_env->cls)
 734
 735 static void preColoringWalker(ir_node *bl, void *env) {
 736         struct pre_col *e = env;
 737         pset **clique = e->clique;
 738         pset *max_clique = clique ? *clique : NULL;
 739         int max = max_clique ? pset_count(max_clique) : 0;
 740         problem_instance_t *pi = e->pi;
 741
 742         int i, n;
 743         pset *live       = pset_new_ptr_default();
 744         ir_node *irn;
 745         irn_live_t *li;
 746
 747         /* as always, bring the live end nodes to life here */
 748         live_foreach(bl, li) {
 749           if(live_is_end(li) && has_reg_class(pi, li->irn)) {
 750             pset_insert_ptr(live, irn);
 751           }
 752         }
 753
 754         sched_foreach_reverse(bl, irn) {
 755                 int pres = pset_count(live);
 756
 757                 if(pres > max) {
 758                         max = pres;
 759                         if(max_clique)
 760                                 del_pset(max_clique);
 761
 762                         max_clique = pset_new_ptr_default();
 763                         pset_insert_pset_ptr(max_clique, live);
 764                 }
 765
 766
 767
 768                 if(has_reg_class(pi, irn))
 769                         pset_remove_ptr(live, irn);
 770
 771                 for(i = 0, n = get_irn_arity(irn); i < n; ++i) {
 772                         ir_node *op = get_irn_n(irn, i);
 773                         if(has_reg_class(pi, op) && !is_Phi(irn))
 774                                 pset_insert_ptr(live, op);
 775                 }
 776         }
 777
 778   del_pset(live);
 779   *clique = max_clique;
 780 }
 781
 782 static void pi_add_constr_preColoring(problem_instance_t *pi) {
 783         ir_node *irn;
 784         int cst_counter, color;
 785         struct pre_col pre_col;
 786
 787         pre_col.clique = NULL;
 788         pre_col.pi = pi;
 789
 790         dom_tree_walk_irg(get_irg(pi->co), preColoringWalker, NULL, &pre_col);
 791
 792         color = 0;
 793         for (irn = pset_first(*pre_col.clique); irn; irn = pset_next(*pre_col.clique)) {
 794                 int cst_idx, var_idx, nnr = get_irn_graph_nr(irn);
 795                 char buf[100];
 796
 797                 mangle_cst(buf, 'K', cst_counter++);
 798                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_equal, 1);
 799
 800                 mangle_var2(buf, 'x', nnr, color++);
 801                 var_idx = lpp_get_var_idx(pi->curr_lp, buf);
 802                 lpp_set_factor_fast(pi->curr_lp, cst_idx, var_idx, 1);
 803         }
 804 }
 805 #endif
 806
 807 /**
 808  * Generate the initial problem matrices and vectors.
 809  */
 810 static problem_instance_t *new_pi(const copy_opt_t *co) {
 811         problem_instance_t *pi;
 812         int col;
 813
 814         DBG((dbg, LEVEL_2, "Generating new instance...\n"));
 815         pi = xcalloc(1, sizeof(*pi));
 816         pi->co = co;
 817         pi->removed = pset_new_ptr_default();
 818         INIT_LIST_HEAD(&pi->simplicials);
 819         pi->dilp     = new_lpp(co->name, lpp_minimize);
 820
 821         /* problem size reduction */
 822         pi_find_simplicials(pi);
 823         if (pi->all_simplicial)
 824                 return pi;
 825
 826         /* built objective and constraints */
 827         pi->curr_lp = pi->dilp;
 828 #ifdef NO_NULL_COLORS
 829         pi_add_constr_no_null_colors(pi);
 830 #endif
 831         pi_add_constr_A(pi);
 832         for (col = 0; col < pi->co->chordal_env->cls->n_regs; ++col)
 833                 pi_add_constr_B(pi, col);
 834         pi_add_constr_E(pi);
 835
 836 #ifdef PATH_CONSTRAINTS_FOR_CLASSES
 837         pi_add_path_cstr_for_classes(pi);
 838 #else
 839         pi_add_path_cstr(pi);
 840 #endif
 841         pi_add_clique_path_cstr(pi);
 842 #ifdef PRECOLOR_MAX_CLIQUE
 843         pi_add_constr_preColoring(pi);
 844 #endif
 845
 846         return pi;
 847 }
 848
 849 /**
 850  * Clean the problem instance
 851  */
 852 static void free_pi(problem_instance_t *pi) {
 853         simpl_t *simpl, *tmp;
 854
 855         DBG((dbg, LEVEL_2, "Free instance...\n"));
 856         free_lpp(pi->dilp);
 857         list_for_each_entry_safe(simpl_t, simpl, tmp, &pi->simplicials, chain)
 858                 free(simpl);
 859         del_pset(pi->removed);
 860         free(pi);
 861 }
 862
 863 /**
 864  * Set starting values for the mip problem according
 865  * to the current coloring of the graph.
 866  */
 867 static void pi_set_start_sol(problem_instance_t *pi) {
 868         int i;
 869         char var_name[64];
 870         DBG((dbg, LEVEL_2, "Set start solution...\n"));
 871         for (i=pi->first_x_var; i<=pi->last_x_var; ++i) {
 872                 int nnr, col;
 873                 double val;
 874                 /* get variable name */
 875                 lpp_get_var_name(pi->curr_lp, i, var_name, sizeof(var_name));
 876                 /* split into components */
 877                 if (split_var(var_name, &nnr, &col) == 2) {
 878                         assert(get_irn_col(pi->co, get_irn_for_graph_nr(get_irg(pi->co), nnr)) != -1);
 879                         val = (get_irn_col(pi->co, get_irn_for_graph_nr(get_irg(pi->co), nnr)) == col) ? 1 : 0;
 880                         lpp_set_start_value(pi->curr_lp, i, val);
 881                 } else {
 882                         fprintf(stderr, "Variable name is: %s\n", var_name);
 883                         assert(0 && "x vars always look like this 'x123_45'");
 884                 }
 885         }
 886 }
 887
 888 /**
 889  * Invoke a solver
 890  */
 891 static void pi_solve_ilp(problem_instance_t *pi) {
 892         pi_set_start_sol(pi);
 893         lpp_solve_net(pi->curr_lp, LPP_HOST, LPP_SOLVER);
 894 //      lpp_solve_cplex(pi->curr_lp);
 895         DBG((dbg, LEVEL_1, "Solution time: %.2f\n", pi->curr_lp->sol_time));
 896 }
 897
 898 /**
 899  * Set the color of all simplicial nodes removed form
 900  * the graph before transforming it to an ilp.
 901  */
 902 static void pi_set_simplicials(problem_instance_t *pi) {
 903         simpl_t *simpl, *tmp;
 904         bitset_t *used_cols = bitset_alloca(arch_register_class_n_regs(pi->co->chordal_env->cls));
 905
 906         DBG((dbg, LEVEL_2, "Set simplicials...\n"));
 907         /* color the simplicial nodes in right order */
 908         list_for_each_entry_safe(simpl_t, simpl, tmp, &pi->simplicials, chain) {
 909                 int free_col;
 910                 ir_node *other_irn, *irn;
 911                 if_node_t *other, *ifn;
 912
 913                 /* get free color by inspecting all neighbors */
 914                 ifn = simpl->ifn;
 915                 irn = get_irn_for_graph_nr(get_irg(pi->co), ifn->nnr);
 916                 bitset_clear_all(used_cols);
 917                 foreach_neighb(ifn, other) {
 918                         other_irn = get_irn_for_graph_nr(get_irg(pi->co), other->nnr);
 919                         if (!is_removed(other_irn)) /* only inspect nodes which are in graph right now */
 920                                 bitset_set(used_cols, get_irn_col(pi->co, other_irn));
 921                 }
 922
 923                 /* now all bits not set are possible colors */
 924                 free_col = bitset_next_clear(used_cols, 0);
 925                 assert(free_col != -1 && "No free color found. This can not be.");
 926                 set_irn_col(pi->co, irn, free_col);
 927                 pset_remove_ptr(pi->removed, irn); /* irn is back in graph again */
 928         }
 929 }
 930
 931 /**
 932  * Sets the colors of irns according to the values of variables
 933  * provided by the solution of the solver.
 934  */
 935 static void pi_apply_solution(problem_instance_t *pi) {
 936         int i;
 937         double *sol;
 938         lpp_sol_state_t state;
 939         DBG((dbg, LEVEL_2, "Applying solution...\n"));
 940
 941 #ifdef DO_STAT
 942         copystat_add_ilp_time((int)(1000.0*lpp_get_sol_time(pi->curr_lp)));  //now we have ms
 943         copystat_add_ilp_vars(lpp_get_var_count(pi->curr_lp));
 944         copystat_add_ilp_csts(lpp_get_cst_count(pi->curr_lp));
 945         copystat_add_ilp_iter(lpp_get_iter_cnt(pi->curr_lp));
 946 #endif
 947
 948         sol = xmalloc((pi->last_x_var - pi->first_x_var + 1) * sizeof(*sol));
 949         state = lpp_get_solution(pi->curr_lp, sol, pi->first_x_var, pi->last_x_var);
 950         if (state != lpp_optimal) {
 951                 printf("Solution state is not 'optimal': %d\n", state);
 952                 assert(state >= lpp_feasible && "The solution should at least be feasible!");
 953         }
 954         for (i=0; i<pi->last_x_var - pi->first_x_var + 1; ++i) {
 955                 int nnr, col;
 956                 char var_name[64];
 957
 958                 if (sol[i] > 1-EPSILON) { /* split varibale name into components */
 959                         lpp_get_var_name(pi->curr_lp, pi->first_x_var+i, var_name, sizeof(var_name));
 960                         if (split_var(var_name, &nnr, &col) == 2) {
 961                                 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]));
 962                                 DBG((dbg, LEVEL_2, "x%d = %d\n", nnr, col));
 963                                 set_irn_col(pi->co, get_irn_for_graph_nr(get_irg(pi->co), nnr), col);
 964                         } else
 965                                 assert(0 && "This should be a x-var");
 966                 }
 967         }
 968 }
 969
 970 void co_ilp_opt(copy_opt_t *co) {
 971         problem_instance_t *pi;
 972
 973         dbg = firm_dbg_register("ir.be.copyoptilp");
 974         if (!strcmp(co->name, DEBUG_IRG))
 975                 firm_dbg_set_mask(dbg, DEBUG_IRG_LVL_ILP);
 976         else
 977                 firm_dbg_set_mask(dbg, DEBUG_LVL_ILP);
 978
 979         pi = new_pi(co);
 980         if (!pi->all_simplicial) {
 981 #ifdef DUMP_MPS
 982                 char buf[512];
 983                 snprintf(buf, sizeof(buf), "%s.mps", co->name);
 984                 lpp_dump(pi->curr_lp, buf);
 985 #endif
 986                 pi_solve_ilp(pi);
 987                 pi_apply_solution(pi);
 988                 pi_set_simplicials(pi);
 989         }
 990         free_pi(pi);
 991 }