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  * CVS-ID:      $Id$
   7  */
   8 #ifdef HAVE_CONFIG_H
   9 #include "config.h"
  10 #endif
  11 #ifdef HAVE_ALLOCA_H
  12 #include <alloca.h>
  13 #endif
  14 #ifdef HAVE_MALLOC_H
  15 #include <malloc.h>
  16 #endif
  17
  18 #define PATH_CONSTRAINTS_FOR_CLASSES
  19 #undef PRECOLOR_MAX_CLIQUE
  20 #undef NO_NULL_COLORS
  21 #undef NO_NULL_COLORS_EXTRA_CSTS
  22 #undef NO_NULL_COLORS_WITH_COSTS
  23 #if (defined(NO_NULL_COLORS_EXTRA_CSTS) || defined(NO_NULL_COLORS_WITH_COSTS)) && !defined(NO_NULL_COLORS)
  24 #error Chose your weapon!
  25 #endif
  26
  27 #include "irprog.h"
  28
  29 #include <lpp/lpp.h>
  30 #include <lpp/lpp_net.h>
  31 #include <lpp/lpp_cplex.h>
  32 #include <lpp/lpp_remote.h>
  33 #include "xmalloc.h"
  34 #include "pset.h"
  35 #include "irdom_t.h"
  36 #include "iredges_t.h"
  37 #include "bechordal_t.h"
  38 #include "becopyopt.h"
  39 #include "becopystat.h"
  40 #include "besched_t.h"
  41 #include "phiclass.h"
  42
  43 #define LPP_HOST "i44pc52"
  44 #define LPP_SOLVER "cplex"
  45
  46 #undef DUMP_MPS
  47 static firm_dbg_module_t *dbg = NULL;
  48
  49 #define MAX(a,b) ((a<b)?(b):(a))
  50 #define MIN(a,b) ((a<b)?(a):(b))
  51 #define EPSILON 0.00001
  52 #define SLOTS_LIVING 32
  53
  54 typedef struct _simpl_t {
  55         struct list_head chain;
  56         ir_node *irn;
  57 } simpl_t;
  58
  59 typedef struct _problem_instance_t {
  60         const copy_opt_t *co;                   /** the copy_opt problem */
  61         /* problem size reduction removing simple nodes */
  62         struct list_head simplicials;   /**< holds all simpl_t's in right order to color*/
  63         pset *removed;                                  /**< holds all removed simplicial irns */
  64         /* lp problem */
  65         lpp_t *curr_lp;                                 /**< points to the problem currently used */
  66         lpp_t *dilp;                                    /**< problem formulation directly as milp */
  67 #ifdef NO_NULL_COLORS_EXTRA_CSTS
  68         int first_nnc_cst_idx;                  /**< the first index of a constraint belonging to no-null-colors stuff*/
  69 #endif
  70         int first_nnc_var_idx;                  /**< the first index of a constraint belonging to no-null-colors stuff*/
  71
  72         int cst_counter, first_x_var, last_x_var;
  73         char buf[32];
  74         int all_simplicial;
  75         pset *done;
  76 } problem_instance_t;
  77
  78 #define is_removed(irn) pset_find_ptr(pi->removed, irn)
  79
  80 #define is_color_possible(irn,color) arch_reg_is_allocatable(get_arch_env(pi->co), irn, -1, arch_register_for_index(pi->co->chordal_env->cls, color))
  81
  82 /*
  83  * Some stuff for variable name handling.
  84  */
  85 #define mangle_cst(buf, prefix, nr) \
  86                         snprintf((buf), sizeof(buf), "%c%d", (prefix), (nr))
  87
  88 #define mangle_var1(buf, prefix, color) \
  89                         snprintf((buf), sizeof(buf), "%c%d", (prefix), (color))
  90
  91 #define mangle_var2(buf, prefix, node_nr, color) \
  92                         snprintf((buf), sizeof(buf), "%c%d_%d", (prefix), (node_nr), (color))
  93
  94 #define mangle_var3(buf, prefix, n1, n2, col) \
  95                         snprintf((buf), sizeof(buf), "%c%d_%d_%d", (prefix), (n1), (n2), (col))
  96
  97 #define mangle_var_irn(buf, prefix, irn, color) \
  98                         mangle_var2((buf), (prefix), get_irn_graph_nr(irn), (color))
  99
 100 #define split_var(var, nnr, col) \
 101                         sscanf(var, "x%d_%d", (nnr), (col))
 102
 103
 104 /**
 105  * Checks if a node is simplicial in the graph
 106  * heeding the already removed nodes.
 107  */
 108 static INLINE int pi_is_simplicial(problem_instance_t *pi, const ir_node *ifn) {
 109         int i, o, size = 0;
 110         ir_node **all, *curr;
 111         be_ifg_t *ifg = pi->co->chordal_env->ifg;
 112         void *iter = be_ifg_iter_alloca(ifg);
 113
 114         all = alloca(be_ifg_degree(ifg, ifn) * sizeof(*all));
 115
 116         /* get all non-removed neighbors */
 117         be_ifg_foreach_neighbour(ifg, iter, ifn, curr)
 118                 if (!is_removed(curr))
 119                         all[size++] = curr;
 120
 121         /* check if these form a clique */
 122         for (i=0; i<size; ++i)
 123                 for (o=i+1; o<size; ++o)
 124                         if (!be_ifg_connected(ifg, all[i], all[o]))
 125                                 return 0;
 126
 127         /* all edges exist so this is a clique */
 128         return 1;
 129 }
 130
 131 static int irn_cmp(const void *a, const void *b, size_t n)
 132 {
 133         return a != b;
 134 }
 135
 136 /**
 137  * Iterative finds and 'removes' from the graph all nodes which are
 138  * simplicial AND not member of a equal-color-wish
 139  */
 140 static void pi_find_simplicials(problem_instance_t *pi) {
 141         ir_node *irn;
 142         int redo = 1;
 143         int n_nodes = 0;
 144         const be_ifg_t *ifg = pi->co->chordal_env->ifg;
 145         void *iter = be_ifg_iter_alloca(ifg);
 146
 147         DBG((dbg, LEVEL_2, "Find simlicials...\n"));
 148
 149         while (redo) {
 150                 redo = 0;
 151                 be_ifg_foreach_node(ifg, iter, irn) {
 152                         if (!is_removed(irn) && !is_optimizable(get_arch_env(pi->co), irn) && !is_optimizable_arg(pi->co, irn)) {
 153                         if (pi_is_simplicial(pi, irn)) {
 154                                         simpl_t *s = xmalloc(sizeof(*s));
 155                                         s->irn = irn;
 156                                         list_add(&s->chain, &pi->simplicials);
 157                                         pset_insert_ptr(pi->removed, irn);
 158                                         redo = 1;
 159                                         DBG((dbg, LEVEL_2, " Removed %+F\n", irn));
 160                         }
 161                         }
 162                 }
 163         }
 164
 165         /* TODO: Count inside the last look */
 166         be_ifg_foreach_node(ifg, iter, irn) {
 167                 n_nodes++;
 168         }
 169
 170         if (n_nodes == pset_count(pi->removed))
 171                 pi->all_simplicial = 1;
 172 }
 173
 174 #ifdef NO_NULL_COLORS
 175 static void pi_add_constr_no_null_colors(problem_instance_t *pi) {
 176         int cst_counter=0, col, var_idx, cst_idx;
 177         int n_colors = pi->co->chordal_env->cls->n_regs;
 178         char buf[40];
 179
 180         for (col = 0; col < n_colors; ++col) {
 181                 mangle_var1(buf, 'u', col);
 182 #ifdef NO_NULL_COLORS_WITH_COSTS
 183                 var_idx = lpp_add_var(pi->curr_lp, buf, lpp_binary, 1.0 / (double) (1 << (col+1)) );
 184 #else
 185                 var_idx = lpp_add_var(pi->curr_lp, buf, lpp_binary, 1.0 / (2.0 * n_colors) );
 186 #endif
 187                 if (!pi->first_nnc_var_idx)
 188                         pi->first_nnc_var_idx = var_idx;
 189         }
 190
 191 #ifdef NO_NULL_COLORS_EXTRA_CSTS
 192         for (col = 0; col < n_colors; ++col) {
 193                 mangle_cst(buf, 'U', cst_counter++);
 194                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_greater, 0);
 195                 if (!pi->first_nnc_cst_idx)
 196                         pi->first_nnc_cst_idx = cst_idx;
 197                 lpp_set_factor_fast(pi->curr_lp, cst_idx, pi->first_nnc_var_idx+col, -1);
 198         }
 199 #endif
 200
 201 #ifndef NO_NULL_COLORS_WITH_COSTS
 202         for (col = 0; col < n_colors - 1; ++col) {
 203                 mangle_cst(buf, 'U', cst_counter++);
 204                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_greater, 0);
 205                 lpp_set_factor_fast(pi->curr_lp, cst_idx, pi->first_nnc_var_idx+col  ,  1);
 206                 lpp_set_factor_fast(pi->curr_lp, cst_idx, pi->first_nnc_var_idx+col+1, -1);
 207         }
 208 #endif
 209
 210 }
 211 #endif
 212
 213 /**
 214  * Add coloring-force conditions
 215  * Matrix A: knapsack constraint for each node
 216  */
 217 static void pi_add_constr_A(problem_instance_t *pi) {
 218         pmap_entry *pme;
 219
 220         DBG((dbg, LEVEL_2, "Add A constraints...\n"));
 221         /* iterate over all blocks */
 222         pmap_foreach(pi->co->chordal_env->border_heads, pme) {
 223                 struct list_head *head = pme->value;
 224                 border_t *curr;
 225                 bitset_t *pos_regs = bitset_alloca(pi->co->chordal_env->cls->n_regs);
 226
 227                 list_for_each_entry_reverse(border_t, curr, head, list)
 228                         if (curr->is_def && curr->is_real && !is_removed(curr->irn)) {
 229                                 int cst_idx, nnr, col;
 230
 231                                 nnr = get_irn_graph_nr(curr->irn);
 232                                 mangle_cst(pi->buf, 'A', nnr);
 233                                 cst_idx = lpp_add_cst(pi->curr_lp, pi->buf, lpp_equal, 1);
 234
 235                                 /* iterate over all possible colors in order */
 236                                 bitset_clear_all(pos_regs);
 237                                 arch_get_allocatable_regs(get_arch_env(pi->co), curr->irn, -1, pi->co->chordal_env->cls, pos_regs);
 238                                 bitset_foreach(pos_regs, col) {
 239                                         int var_idx;
 240                                         mangle_var2(pi->buf, 'x', nnr, col);
 241                                         var_idx = lpp_add_var(pi->curr_lp, pi->buf, lpp_binary, 0);
 242                                         if (!pi->first_x_var)
 243                                                 pi->first_x_var = var_idx;
 244                                         pi->last_x_var = var_idx;
 245                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, var_idx, 1);
 246 #ifdef NO_NULL_COLORS_EXTRA_CSTS
 247                                         lpp_set_factor_fast(pi->curr_lp, pi->first_nnc_cst_idx+col, var_idx, 1);
 248 #endif
 249                                 }
 250                         }
 251         }
 252 }
 253
 254 /**
 255  * Checks if all nodes in @p living are live in in block @p block.
 256  * @return 1 if all are live in
 257  *         0 else
 258  */
 259 static INLINE int all_live_in(ir_node *block, pset *living) {
 260         ir_node *n;
 261         for (n = pset_first(living); n; n = pset_next(living))
 262                 if (!is_live_in(block, n)) {
 263                         pset_break(living);
 264                         return 0;
 265                 }
 266         return 1;
 267 }
 268
 269 /**
 270  * Finds cliques in the interference graph, considering only nodes
 271  * for which the color @p color is possible. Finds only 'maximal-cliques',
 272  * viz cliques which are not contained in another one.
 273  * Matrix B: interference constraints using cliques
 274  */
 275 static void pi_add_constr_B(problem_instance_t *pi, int color) {
 276         enum phase_t {growing, shrinking} phase = growing;
 277         border_t *b;
 278         pmap_entry *pme;
 279         pset *living = pset_new_ptr(SLOTS_LIVING);
 280
 281         DBG((dbg, LEVEL_2, "Add B constraints (col = %d)...\n", color));
 282         /* iterate over all blocks */
 283         pmap_foreach(pi->co->chordal_env->border_heads, pme) {
 284                 ir_node *block = pme->key;
 285                 struct list_head *head = pme->value;
 286
 287                 list_for_each_entry_reverse(border_t, b, head, list) {
 288                         const ir_node *irn = b->irn;
 289                         if (is_removed(irn) || !is_color_possible(irn, color))
 290                                 continue;
 291
 292                         if (b->is_def) {
 293                                 DBG((dbg, LEVEL_2, "Def %n\n", irn));
 294                                 pset_insert_ptr(living, irn);
 295                                 phase = growing;
 296                         } else { /* is_use */
 297                                 DBG((dbg, LEVEL_2, "Use %n\n", irn));
 298
 299                                 /* before shrinking the set, store the current 'maximum' clique;
 300                                  * do NOT if clique is a single node
 301                                  * do NOT if all values are live_in (in this case they were contained in a live-out clique elsewhere) */
 302                                 if (phase == growing && pset_count(living) >= 2 && !all_live_in(block, living)) {
 303                                         int cst_idx;
 304                                         ir_node *n;
 305                                         mangle_cst(pi->buf, 'B', pi->cst_counter);
 306 #ifdef NO_NULL_COLORS
 307                                         cst_idx = lpp_add_cst(pi->curr_lp, pi->buf, lpp_less, 0);
 308 #else
 309                                         cst_idx = lpp_add_cst(pi->curr_lp, pi->buf, lpp_less, 1);
 310 #endif
 311                                         for (n = pset_first(living); n; n = pset_next(living)) {
 312                                                 int var_idx;
 313                                                 mangle_var_irn(pi->buf, 'x', n, color);
 314                                                 var_idx = lpp_get_var_idx(pi->curr_lp, pi->buf);
 315                                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, var_idx, 1);
 316                                         }
 317 #ifdef NO_NULL_COLORS
 318                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, pi->first_nnc_var_idx+color, -1.0);
 319 #endif
 320                                         pi->cst_counter++;
 321                                 }
 322                                 pset_remove_ptr(living, irn);
 323                                 phase = shrinking;
 324                         }
 325                 }
 326         }
 327         assert(0 == pset_count(living));
 328         del_pset(living);
 329 }
 330
 331 /**
 332  * Generates constraints which interrelate x with y variables.
 333  * x1 and x2 have the different colors ==> y_12 = 1
 334  */
 335 static void pi_add_constr_E(problem_instance_t *pi) {
 336         unit_t *curr;
 337         bitset_t *root_regs, *arg_regs, *work_regs;
 338         int cst_counter = 0;
 339         unsigned nregs = pi->co->chordal_env->cls->n_regs;
 340         root_regs = bitset_alloca(nregs);
 341         arg_regs = bitset_alloca(nregs);
 342         work_regs = bitset_alloca(nregs);
 343
 344         DBG((dbg, LEVEL_2, "Add E constraints...\n"));
 345         /* for all roots of optimization units */
 346         list_for_each_entry(unit_t, curr, &pi->co->units, units) {
 347                 ir_node *root, *arg;
 348                 int rootnr, argnr, color;
 349                 int y_idx, i;
 350                 char buf[32];
 351
 352                 root = curr->nodes[0];
 353                 rootnr = get_irn_graph_nr(root);
 354                 bitset_clear_all(root_regs);
 355                 arch_get_allocatable_regs(get_arch_env(pi->co), root, -1, pi->co->chordal_env->cls, root_regs);
 356
 357                 /* for all arguments of root */
 358                 for (i = 1; i < curr->node_count; ++i) {
 359                         arg = curr->nodes[i];
 360                         argnr = get_irn_graph_nr(arg);
 361                         bitset_clear_all(arg_regs);
 362                         arch_get_allocatable_regs(get_arch_env(pi->co), arg, -1, pi->co->chordal_env->cls, arg_regs);
 363
 364                         /* Introduce new variable and set factor in objective function */
 365                         mangle_var2(buf, 'y', rootnr, argnr);
 366                         y_idx = lpp_add_var(pi->curr_lp, buf, lpp_binary, curr->costs[i]);
 367
 368                         /* set starting value */
 369                         lpp_set_start_value(pi->curr_lp, y_idx, (get_irn_col(pi->co, root) != get_irn_col(pi->co, arg)));
 370
 371                         /* For all colors root and arg have in common, add 2 constraints to E */
 372                         bitset_copy(work_regs, root_regs);
 373                         bitset_and(work_regs, arg_regs);
 374                         bitset_foreach(work_regs, color) {
 375                                 int root_idx, arg_idx, cst_idx;
 376                                 mangle_var2(buf, 'x', rootnr, color);
 377                                 root_idx = lpp_get_var_idx(pi->curr_lp, buf);
 378                                 mangle_var2(buf, 'x', argnr, color);
 379                                 arg_idx = lpp_get_var_idx(pi->curr_lp, buf);
 380
 381                                 /* add root-arg-y <= 0 */
 382                                 mangle_cst(buf, 'E', cst_counter++);
 383                                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_less, 0);
 384                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, root_idx, 1);
 385                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, arg_idx, -1);
 386                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, -1);
 387
 388                                 /* add arg-root-y <= 0 */
 389                                 mangle_cst(buf, 'E', cst_counter++);
 390                                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_less, 0);
 391                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, root_idx, -1);
 392                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, arg_idx, 1);
 393                                 lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, -1);
 394                         }
 395                         /* For all colors root and arg have "disjunct", add 1 constraints to E.
 396                          * If root gets a color the arg is not possible to get then they will
 397                          * definetly get different colors. So y has to be 1.
 398                          * Vice versa for arg.
 399                          */
 400                         bitset_copy(work_regs, root_regs);
 401                         bitset_xor(work_regs, arg_regs);
 402                         bitset_foreach(work_regs, color) {
 403                                 int root_idx, arg_idx, cst_idx;
 404                                 mangle_var2(buf, 'x', rootnr, color);
 405                                 root_idx = lpp_get_var_idx(pi->curr_lp, buf);
 406                                 mangle_var2(buf, 'x', argnr, color);
 407                                 arg_idx = lpp_get_var_idx(pi->curr_lp, buf);
 408
 409                                 mangle_cst(buf, 'E', cst_counter++);
 410                                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_less, 0);
 411                                 if (bitset_is_set(root_regs, color)) {
 412                                         /* add root-y <= 0 */
 413                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, root_idx, 1);
 414                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, -1);
 415                                 } else {
 416                                         assert(bitset_is_set(arg_regs, color) && "bitset_xor is buggy");
 417                                         /* add arg-y <= 0 */
 418                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, arg_idx, 1);
 419                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, -1);
 420                                 }
 421                         }
 422                 }
 423         }
 424 }
 425
 426 static INLINE int get_costs(problem_instance_t *pi, ir_node *phi, ir_node *irn) {
 427         int i;
 428         unit_t *curr;
 429         /* search optimization unit for phi */
 430         list_for_each_entry(unit_t, curr, &pi->co->units, units)
 431                 if (curr->nodes[0] == phi) {
 432                         for (i=1; i<curr->node_count; ++i)
 433                                 if (curr->nodes[i] == irn)
 434                                         return curr->costs[i];
 435                         assert(0 && "irn must occur in this ou");
 436                 }
 437         assert(0 && "phi must be found in a ou");
 438         return 0;
 439 }
 440
 441 static void clique_path_walker(ir_node *block, void *env) {
 442         problem_instance_t *pi = env;
 443         int count, arity, row, col, other_row, *costs;
 444         ir_node **phis, *phi, *irn, **phi_matrix;
 445         pset *done;
 446         bitset_t *candidates;
 447
 448         /* Count all phi nodes of this block */
 449         for (count=0, irn = sched_first(block); is_Phi(irn); irn = sched_next(irn))
 450                 count++;
 451
 452         /* We at least 2 phi nodes for this class of inequalities */
 453         if (count < 2)
 454                 return;
 455
 456         /* Build the \Phi-Matrix */
 457         arity = get_irn_arity(sched_first(block));
 458         phis = alloca(count * sizeof(*phis));
 459         costs = alloca(count * sizeof(costs));
 460         phi_matrix = alloca(count*arity * sizeof(*phi_matrix));
 461         candidates = bitset_alloca(count);
 462
 463         phi = sched_first(block);
 464         for (row=0; row<count; ++row) {
 465                 phis[row] = phi;
 466                 for (col=0; col<arity; ++col) {
 467                         ir_node *arg = get_irn_n(phi, col);
 468                         /* Sort out all arguments interfering with its phi */
 469                         if (nodes_interfere(pi->co->chordal_env, phi, arg)) {
 470                                 phi_matrix[row*arity + col] =  NULL;
 471                         } else
 472                                 phi_matrix[row*arity + col] =  arg;
 473                 }
 474                 phi = sched_next(phi);
 475         }
 476
 477         /* Now find the interesting patterns in the matrix:
 478          * All nodes which are used at least twice in a column. */
 479         /* columnwise ... */
 480         for (col=0; col<arity; ++col) {
 481                 done = pset_new_ptr_default();
 482                 for (row=0; row<count; ++row) {
 483                         irn = phi_matrix[row*arity + col];
 484                         /*
 485                          * is this an interfering arg (NULL)
 486                          * or has the irn already been processed in this col?
 487                          */
 488                         if (!irn || pset_find_ptr(done, irn))
 489                                 continue;
 490                         else
 491                                 pset_insert_ptr(done, irn);
 492
 493                         /* insert irn in candidates */
 494                         bitset_clear_all(candidates);
 495                         bitset_set(candidates, row);
 496                         /* search the irn in the rows below */
 497                         for (other_row = row+1; other_row<count; ++other_row)
 498                                 if (irn == phi_matrix[other_row*arity + col]) {
 499                                         /* found the irn in the same col in another row */
 500                                         bitset_set(candidates, other_row);
 501                                 }
 502
 503                         /* now we know all occurences of irn in this col */
 504                         if (bitset_popcnt(candidates) < 2)
 505                                 continue;
 506
 507                         /* generate an unequation finally.
 508                          * phis are indexed in the bitset,
 509                          * shared argument is irn
 510                          * rhs is phi_count - 1 */
 511                         {
 512                                 char buf[32];
 513                                 ir_node *root;
 514                                 int pos, irnnr, rootnr, cst_idx, y_idx, cst_counter = 0;
 515                                 int minimal_unequal_count = bitset_popcnt(candidates)-1;
 516
 517                                 irnnr = get_irn_graph_nr(irn);
 518                                 mangle_cst(buf, 'M', cst_counter++);
 519                                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_greater, minimal_unequal_count);
 520
 521                                 /* for all phis */
 522                                 bitset_foreach(candidates, pos) {
 523                                         root = phis[pos];
 524                                         rootnr = get_irn_graph_nr(root);
 525                                         mangle_var2(buf, 'y', rootnr, irnnr);
 526                                         y_idx = lpp_get_var_idx(pi->curr_lp, buf);
 527                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, 1);
 528                                 }
 529                         }
 530                 }
 531                 del_pset(done); /* clear set for next row */
 532         } /*next col*/
 533 }
 534
 535 /**
 536  * Matrix M: Multi-Arg-Use. Interrelates different \phi-functions
 537  * in the same block, iff they use the same arg at the same pos.
 538  * Only one of the phis can get the arg.
 539  */
 540 static void pi_add_clique_path_cstr(problem_instance_t *pi) {
 541         DBG((dbg, LEVEL_2, "Adding clique path constraints...\n"));
 542         dom_tree_walk_irg(get_irg(pi->co), clique_path_walker, NULL, pi);
 543 }
 544
 545 #ifndef PATH_CONSTRAINTS_FOR_CLASSES
 546 /**
 547  * Matrix P: Path contraints.
 548  * If 2 nodes interfere and there is a path of equal-color-edges
 549  * connecting them, then at least one of those equal-color-edges
 550  * will break and cause some costs.
 551  */
 552 static void pi_add_path_cstr(problem_instance_t *pi) {
 553         unit_t *curr;
 554         int cst_counter = 0;
 555         DBG((dbg, LEVEL_2, "Adding path constraints...\n"));
 556
 557         /* for all optimization units (only phis) */
 558         list_for_each_entry(unit_t, curr, &pi->co->units, units) {
 559                 int i, o, rootnr;
 560
 561                 if (curr->min_nodes_costs == 0)
 562                         continue;
 563
 564                 rootnr = get_irn_graph_nr(curr->nodes[0]);
 565                 /* check all argument pairs for interference */
 566                 for (i=1; i<curr->node_count; ++i) {
 567                         const ir_node *arg1 = curr->nodes[i];
 568                         int arg1nr = get_irn_graph_nr(arg1);
 569                         for (o=i+1; o<curr->node_count; ++o) {
 570                                 const ir_node *arg2 = curr->nodes[o];
 571                                 int arg2nr = get_irn_graph_nr(arg2);
 572                                 if (nodes_interfere(pi->co->chordal_env, arg1, arg2)) {
 573                                         int cst_idx, y_idx;
 574                                         char buf[32];
 575
 576                                         mangle_cst(buf, 'P', cst_counter++);
 577                                         cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_greater, 1);
 578
 579                                         mangle_var2(buf, 'y', rootnr, arg1nr);
 580                                         y_idx = lpp_get_var_idx(pi->curr_lp, buf);
 581                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, 1);
 582
 583                                         mangle_var2(buf, 'y', rootnr, arg2nr);
 584                                         y_idx = lpp_get_var_idx(pi->curr_lp, buf);
 585                                         lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, 1);
 586                                 }
 587                         }
 588                 }
 589         }
 590 }
 591 #endif
 592
 593 #ifdef PATH_CONSTRAINTS_FOR_CLASSES
 594 static INLINE int get_y_var_idx(problem_instance_t *pi, int nnr1, int nnr2) {
 595         int res;
 596         char buf[30];
 597
 598         mangle_var2(buf, 'y', nnr1, nnr2);
 599         if ((res = lpp_get_var_idx(pi->curr_lp, buf)) != -1)
 600                 return res;
 601
 602         mangle_var2(buf, 'y', nnr2, nnr1);
 603         if ((res = lpp_get_var_idx(pi->curr_lp, buf)) != -1)
 604                 return res;
 605
 606         assert(0 && "One of them must work");
 607   return -1;
 608 }
 609
 610 static void check_ecc_and_add_cut(problem_instance_t *pi, ir_node **path, int length, pset *remain, ir_node *tgt) {
 611         if (path[length-1] == tgt) { /* we found a path */
 612                 int cst_idx, var_idx, i, nnr1, nnr2;
 613                 char buf[30];
 614
 615                 /* add cut to ilp */
 616                 mangle_cst(buf, 'Q', pi->cst_counter++);
 617                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_greater, 1);
 618
 619                 /* add all vars along the path */
 620                 nnr2 = get_irn_graph_nr(path[0]);
 621                 for (i=1; i<length; ++i) {
 622                         nnr1 = nnr2;
 623                         nnr2 = get_irn_graph_nr(path[i]);
 624                         var_idx = get_y_var_idx(pi, nnr1, nnr2);
 625                         lpp_set_factor_fast(pi->curr_lp, cst_idx, var_idx, 1);
 626                 }
 627         } else { /* try to extend the path */
 628                 be_chordal_env_t *cenv = pi->co->chordal_env;
 629                 const ir_edge_t *edge;
 630                 ir_node *end = path[length-1];
 631                 ir_node **next = alloca(pset_count(remain) * sizeof(*next));
 632                 int i, o, max, next_pos = 0;
 633                 pset *done = pset_new_ptr_default();
 634
 635                 /* find all potential next nodes on path */
 636                 /*  args of phis */
 637                 if (is_Phi(end))
 638                         for(i=0, max=get_irn_arity(end); i<max; ++i) {
 639                                 ir_node *arg = get_irn_n(end, i);
 640                                 if (!pset_find_ptr(done, arg) && pset_find_ptr(remain, arg)) {
 641                                         next[next_pos++] = arg;
 642                                         pset_insert_ptr(done, arg);
 643                                 }
 644                         }
 645                 /*  outs of phis and other nodes */
 646                 foreach_out_edge(end, edge) {
 647                         ir_node *user = edge->src;
 648                         if (is_Phi(user) && !pset_find_ptr(done, user) && pset_find_ptr(remain, user)) {
 649                                 next[next_pos++] = user;
 650                                 pset_insert_ptr(done, user);
 651                         }
 652                 }
 653                 del_pset(done);
 654
 655
 656                 /* delete all potential nodes with interferences to other nodes in the path */
 657                 for (i=0; i<next_pos; ++i) {
 658                         ir_node *nn = next[i];
 659
 660                         /* if next is the tgt, it may interfere with path[0],
 661                          * so skip the first check */
 662                         o = (nn == tgt && length > 1) ? 1 : 0;
 663
 664                         for(; o<length; ++o)
 665                                 if (nodes_interfere(cenv, nn, path[o])) {
 666                                         next[i] = NULL;
 667                                         break;
 668                                 }
 669                 }
 670                 /* now we have all possible nodes in next; impossibles are NULL */
 671
 672                 /* try to finish path with all possible nodes */
 673                 for (i=0; i<next_pos; ++i) {
 674                         if (!next[i]) /* this was an impossible node */
 675                                 continue;
 676
 677                         path[length] = next[i];
 678                         pset_remove_ptr(remain, next[i]);
 679                         check_ecc_and_add_cut(pi, path, length+1, remain, tgt);
 680                         pset_insert_ptr(remain, next[i]);
 681                 }
 682         }
 683 }
 684
 685 static void path_cstr_for_classes_walker(ir_node *irn, void *env) {
 686         problem_instance_t *pi = env;
 687         be_chordal_env_t *cenv;
 688         int i, o, max;
 689         ir_node *m, **cls;
 690         pset *class = get_phi_class(irn);
 691         if (!class || pset_find_ptr(pi->done, class))
 692                 return;
 693
 694         pset_insert_ptr(pi->done, class);
 695
 696         /* pset to array */
 697         max = pset_count(class);
 698         cls = alloca(max * sizeof(*cls));
 699         for(i=0, m = pset_first(class); m; i++, m = pset_next(class)) {
 700                 DBG((dbg, LEVEL_1, " class member: %+F\n", m));
 701                 cls[i] = m;
 702         }
 703
 704         cenv = pi->co->chordal_env;
 705         for(i=0; i<max; ++i) {
 706                 ir_node **path = alloca(max * sizeof(*path));
 707                 pset *remain = pset_new_ptr(8);
 708                 pset_insert_pset_ptr(remain, class);
 709
 710                 /* add cls[i] to path and remove it from remainder */
 711                 path[0] = cls[i];
 712                 pset_remove_ptr(remain, cls[i]);
 713
 714                 for(o=i+1; o<max; ++o)
 715                         if (nodes_interfere(cenv, cls[i], cls[o]))
 716                                 check_ecc_and_add_cut(pi, path, 1, remain, cls[o]);
 717
 718                 /* insert back into remainder */
 719                 pset_insert_ptr(remain, cls[i]);
 720         }
 721 }
 722
 723
 724 /**
 725  * Matrix P: Path contraints.
 726  * If 2 nodes interfere and there is a path of equal-color-edges
 727  * connecting them, then at least one of those equal-color-edges
 728  * will break and cause some costs.
 729  */
 730 static void pi_add_path_cstr_for_classes(problem_instance_t *pi) {
 731         DBG((dbg, LEVEL_2, "Adding path constraints for phi classes...\n"));
 732         pi->cst_counter = 0;
 733         pi->done = pset_new_ptr_default();
 734         irg_walk_graph(get_irg(pi->co), path_cstr_for_classes_walker, NULL, pi);
 735         del_pset(pi->done);
 736 }
 737 #endif
 738
 739 #ifdef PRECOLOR_MAX_CLIQUE
 740 struct pre_col {
 741         problem_instance_t *pi;
 742         pset **clique;
 743 };
 744
 745 #define has_reg_class(pi,irn) \
 746   (arch_get_irn_reg_class(pi->co->chordal_env->session_env->main_env->arch_env, \
 747                           irn, -1) == pi->co->chordal_env->cls)
 748
 749 static void preColoringWalker(ir_node *bl, void *env) {
 750         struct pre_col *e = env;
 751         pset **clique = e->clique;
 752         pset *max_clique = clique ? *clique : NULL;
 753         int max = max_clique ? pset_count(max_clique) : 0;
 754         problem_instance_t *pi = e->pi;
 755
 756         int i, n;
 757         pset *live       = pset_new_ptr_default();
 758         ir_node *irn;
 759         irn_live_t *li;
 760
 761         /* as always, bring the live end nodes to life here */
 762         live_foreach(bl, li) {
 763           if(live_is_end(li) && has_reg_class(pi, li->irn)) {
 764             pset_insert_ptr(live, irn);
 765           }
 766         }
 767
 768         sched_foreach_reverse(bl, irn) {
 769                 int pres = pset_count(live);
 770
 771                 if(pres > max) {
 772                         max = pres;
 773                         if(max_clique)
 774                                 del_pset(max_clique);
 775
 776                         max_clique = pset_new_ptr_default();
 777                         pset_insert_pset_ptr(max_clique, live);
 778                 }
 779
 780
 781
 782                 if(has_reg_class(pi, irn))
 783                         pset_remove_ptr(live, irn);
 784
 785                 for(i = 0, n = get_irn_arity(irn); i < n; ++i) {
 786                         ir_node *op = get_irn_n(irn, i);
 787                         if(has_reg_class(pi, op) && !is_Phi(irn))
 788                                 pset_insert_ptr(live, op);
 789                 }
 790         }
 791
 792   del_pset(live);
 793   *clique = max_clique;
 794 }
 795
 796 static void pi_add_constr_preColoring(problem_instance_t *pi) {
 797         ir_node *irn;
 798         int cst_counter, color;
 799         struct pre_col pre_col;
 800
 801         pre_col.clique = NULL;
 802         pre_col.pi = pi;
 803
 804         dom_tree_walk_irg(get_irg(pi->co), preColoringWalker, NULL, &pre_col);
 805
 806         color = 0;
 807         for (irn = pset_first(*pre_col.clique); irn; irn = pset_next(*pre_col.clique)) {
 808                 int cst_idx, var_idx, nnr = get_irn_graph_nr(irn);
 809                 char buf[100];
 810
 811                 mangle_cst(buf, 'K', cst_counter++);
 812                 cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_equal, 1);
 813
 814                 mangle_var2(buf, 'x', nnr, color++);
 815                 var_idx = lpp_get_var_idx(pi->curr_lp, buf);
 816                 lpp_set_factor_fast(pi->curr_lp, cst_idx, var_idx, 1);
 817         }
 818 }
 819 #endif
 820
 821 /**
 822  * Generate the initial problem matrices and vectors.
 823  */
 824 static problem_instance_t *new_pi(const copy_opt_t *co) {
 825         problem_instance_t *pi;
 826         int col;
 827
 828         DBG((dbg, LEVEL_2, "Generating new instance...\n"));
 829         pi = xcalloc(1, sizeof(*pi));
 830         pi->co = co;
 831         pi->removed = pset_new_ptr_default();
 832         INIT_LIST_HEAD(&pi->simplicials);
 833         pi->dilp     = new_lpp(co->name, lpp_minimize);
 834
 835         /* problem size reduction */
 836         pi_find_simplicials(pi);
 837         if (pi->all_simplicial)
 838                 return pi;
 839
 840         /* built objective and constraints */
 841         pi->curr_lp = pi->dilp;
 842 #ifdef NO_NULL_COLORS
 843         pi_add_constr_no_null_colors(pi);
 844 #endif
 845         pi_add_constr_A(pi);
 846         for (col = 0; col < pi->co->chordal_env->cls->n_regs; ++col)
 847                 pi_add_constr_B(pi, col);
 848         pi_add_constr_E(pi);
 849
 850 #ifdef PATH_CONSTRAINTS_FOR_CLASSES
 851         pi_add_path_cstr_for_classes(pi);
 852 #else
 853         pi_add_path_cstr(pi);
 854 #endif
 855         pi_add_clique_path_cstr(pi);
 856 #ifdef PRECOLOR_MAX_CLIQUE
 857         pi_add_constr_preColoring(pi);
 858 #endif
 859
 860         return pi;
 861 }
 862
 863 /**
 864  * Clean the problem instance
 865  */
 866 static void free_pi(problem_instance_t *pi) {
 867         simpl_t *simpl, *tmp;
 868
 869         DBG((dbg, LEVEL_2, "Free instance...\n"));
 870         free_lpp(pi->dilp);
 871         list_for_each_entry_safe(simpl_t, simpl, tmp, &pi->simplicials, chain)
 872                 free(simpl);
 873         del_pset(pi->removed);
 874         free(pi);
 875 }
 876
 877 /**
 878  * Set starting values for the mip problem according
 879  * to the current coloring of the graph.
 880  */
 881 static void pi_set_start_sol(problem_instance_t *pi) {
 882         int i;
 883         char var_name[64];
 884         DBG((dbg, LEVEL_2, "Set start solution...\n"));
 885         for (i=pi->first_x_var; i<=pi->last_x_var; ++i) {
 886                 int nnr, col;
 887                 double val;
 888                 /* get variable name */
 889                 lpp_get_var_name(pi->curr_lp, i, var_name, sizeof(var_name));
 890                 /* split into components */
 891                 if (split_var(var_name, &nnr, &col) == 2) {
 892                         assert(get_irn_col(pi->co, get_irn_for_graph_nr(get_irg(pi->co), nnr)) != -1);
 893                         val = (get_irn_col(pi->co, get_irn_for_graph_nr(get_irg(pi->co), nnr)) == col) ? 1 : 0;
 894                         lpp_set_start_value(pi->curr_lp, i, val);
 895                 } else {
 896                         fprintf(stderr, "Variable name is: %s\n", var_name);
 897                         assert(0 && "x vars always look like this 'x123_45'");
 898                 }
 899         }
 900 }
 901
 902 /**
 903  * Invoke a solver
 904  */
 905 static void pi_solve_ilp(problem_instance_t *pi) {
 906   double lower_bound;
 907
 908         pi_set_start_sol(pi);
 909         lower_bound = co_get_lower_bound(pi->co) - co_get_inevit_copy_costs(pi->co);
 910         lpp_set_bound(pi->curr_lp, lower_bound);
 911         lpp_solve_net(pi->curr_lp, LPP_HOST, LPP_SOLVER);
 912 //      lpp_solve_cplex(pi->curr_lp);
 913         DBG((dbg, LEVEL_1, "Solution time: %.2f\n", pi->curr_lp->sol_time));
 914 }
 915
 916 /**
 917  * Set the color of all simplicial nodes removed form
 918  * the graph before transforming it to an ilp.
 919  */
 920 static void pi_set_simplicials(problem_instance_t *pi) {
 921         simpl_t *simpl, *tmp;
 922         be_ifg_t *ifg        = pi->co->chordal_env->ifg;
 923         bitset_t *used_cols  = bitset_alloca(arch_register_class_n_regs(pi->co->chordal_env->cls));
 924         void *iter           = be_ifg_iter_alloca(ifg);
 925
 926         DBG((dbg, LEVEL_2, "Set simplicials...\n"));
 927         /* color the simplicial nodes in right order */
 928         list_for_each_entry_safe(simpl_t, simpl, tmp, &pi->simplicials, chain) {
 929                 int free_col;
 930                 ir_node *other, *irn;
 931
 932                 /* get free color by inspecting all neighbors */
 933                 irn = simpl->irn;
 934                 bitset_clear_all(used_cols);
 935
 936                 be_ifg_foreach_neighbour(ifg, iter, irn, other) {
 937                         if (!is_removed(other)) /* only inspect nodes which are in graph right now */
 938                                 bitset_set(used_cols, get_irn_col(pi->co, other));
 939                 }
 940
 941                 /* now all bits not set are possible colors */
 942                 free_col = bitset_next_clear(used_cols, 0);
 943                 assert(free_col != -1 && "No free color found. This can not be.");
 944                 set_irn_col(pi->co, irn, free_col);
 945                 pset_remove_ptr(pi->removed, irn); /* irn is back in graph again */
 946         }
 947 }
 948
 949 /**
 950  * Sets the colors of irns according to the values of variables
 951  * provided by the solution of the solver.
 952  */
 953 static int pi_apply_solution(problem_instance_t *pi) {
 954         int res = 1, i;
 955         double *sol;
 956         lpp_sol_state_t state;
 957         DBG((dbg, LEVEL_2, "Applying solution...\n"));
 958
 959 #ifdef DO_STAT
 960         copystat_add_ilp_time((int)(1000.0*lpp_get_sol_time(pi->curr_lp)));  //now we have ms
 961         copystat_add_ilp_vars(lpp_get_var_count(pi->curr_lp));
 962         copystat_add_ilp_csts(lpp_get_cst_count(pi->curr_lp));
 963         copystat_add_ilp_iter(lpp_get_iter_cnt(pi->curr_lp));
 964 #endif
 965
 966         sol = xmalloc((pi->last_x_var - pi->first_x_var + 1) * sizeof(*sol));
 967         state = lpp_get_solution(pi->curr_lp, sol, pi->first_x_var, pi->last_x_var);
 968         if (state != lpp_optimal) {
 969                 printf("WARNING %s: Solution state is not 'optimal': %d\n", pi->co->name, state);
 970                 assert(state >= lpp_feasible && "The solution should at least be feasible!");
 971                 res = 0;
 972         }
 973         for (i=0; i<pi->last_x_var - pi->first_x_var + 1; ++i) {
 974                 int nnr, col;
 975                 char var_name[64];
 976
 977                 if (sol[i] > 1-EPSILON) { /* split varibale name into components */
 978                         lpp_get_var_name(pi->curr_lp, pi->first_x_var+i, var_name, sizeof(var_name));
 979                         if (split_var(var_name, &nnr, &col) == 2) {
 980                                 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]));
 981                                 DBG((dbg, LEVEL_2, "x%d = %d\n", nnr, col));
 982                                 set_irn_col(pi->co, get_irn_for_graph_nr(get_irg(pi->co), nnr), col);
 983                         } else
 984                                 assert(0 && "This should be a x-var");
 985                 }
 986         }
 987         return res;
 988 }
 989
 990 int co_ilp_opt(copy_opt_t *co, double time_limit) {
 991         int res = 1;
 992         problem_instance_t *pi;
 993
 994         dbg = firm_dbg_register("ir.be.copyoptilp");
 995         if (!strcmp(co->name, DEBUG_IRG))
 996                 firm_dbg_set_mask(dbg, DEBUG_IRG_LVL_ILP);
 997         else
 998                 firm_dbg_set_mask(dbg, DEBUG_LVL_ILP);
 999
1000         pi = new_pi(co);
1001         if (!pi->all_simplicial) {
1002 #ifdef DUMP_MPS
1003                 char buf[512];
1004                 snprintf(buf, sizeof(buf), "%s.mps", co->name);
1005                 lpp_dump(pi->curr_lp, buf);
1006 #endif
1007                 lpp_set_time_limit(pi->curr_lp, time_limit);
1008                 pi_solve_ilp(pi);
1009                 res = pi_apply_solution(pi);
1010                 pi_set_simplicials(pi);
1011         }
1012         free_pi(pi);
1013         return res;
1014 }