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